mirror of
				https://github.com/paboyle/Grid.git
				synced 2025-10-24 17:54:47 +01:00 
			
		
		
		
	Compare commits
	
		
			5 Commits
		
	
	
		
			da81a73b4a
			...
			sycl-linki
		
	
	| Author | SHA1 | Date | |
|---|---|---|---|
|  | 6d25badce5 | ||
|  | adcc97cbbb | ||
|  | bc8c5fb16c | ||
|  | 845d757bb0 | ||
|  | 0d5470c363 | 
							
								
								
									
										54
									
								
								.github/ISSUE_TEMPLATE/bug-report.yml
									
									
									
									
										vendored
									
									
								
							
							
						
						
									
										54
									
								
								.github/ISSUE_TEMPLATE/bug-report.yml
									
									
									
									
										vendored
									
									
								
							| @@ -1,54 +0,0 @@ | |||||||
| name: Bug report |  | ||||||
| description: Report a bug. |  | ||||||
| title: "<insert title>" |  | ||||||
| labels: [bug] |  | ||||||
|  |  | ||||||
| body: |  | ||||||
|   - type: markdown |  | ||||||
|     attributes: |  | ||||||
|       value: > |  | ||||||
|         Thank you for taking the time to file a bug report. |  | ||||||
|         Please check that the code is pointing to the HEAD of develop |  | ||||||
|         or any commit in master which is tagged with a version number. |  | ||||||
|  |  | ||||||
|   - type: textarea |  | ||||||
|     attributes: |  | ||||||
|       label: "Describe the issue:" |  | ||||||
|       description: > |  | ||||||
|         Describe the issue and any previous attempt to solve it. |  | ||||||
|     validations: |  | ||||||
|       required: true |  | ||||||
|  |  | ||||||
|   - type: textarea |  | ||||||
|     attributes: |  | ||||||
|       label: "Code example:" |  | ||||||
|       description: > |  | ||||||
|         If relevant, show how to reproduce the issue using a minimal working |  | ||||||
|         example. |  | ||||||
|       placeholder: | |  | ||||||
|         << your code here >> |  | ||||||
|       render: shell |  | ||||||
|     validations: |  | ||||||
|       required: false |  | ||||||
|  |  | ||||||
|   - type: textarea |  | ||||||
|     attributes: |  | ||||||
|       label: "Target platform:" |  | ||||||
|       description: > |  | ||||||
|         Give a description of the target platform (CPU, network, compiler). |  | ||||||
|         Please give the full CPU part description, using for example |  | ||||||
|         `cat /proc/cpuinfo | grep 'model name' | uniq` (Linux) |  | ||||||
|         or `sysctl machdep.cpu.brand_string` (macOS) and the full output |  | ||||||
|         the `--version` option of your compiler. |  | ||||||
|     validations: |  | ||||||
|       required: true |  | ||||||
|  |  | ||||||
|   - type: textarea |  | ||||||
|     attributes: |  | ||||||
|       label: "Configure options:" |  | ||||||
|       description: > |  | ||||||
|         Please give the exact configure command used and attach |  | ||||||
|         `config.log`, `grid.config.summary` and the output of `make V=1`. |  | ||||||
|       render: shell |  | ||||||
|     validations: |  | ||||||
|       required: true |  | ||||||
							
								
								
									
										5
									
								
								.gitignore
									
									
									
									
										vendored
									
									
								
							
							
						
						
									
										5
									
								
								.gitignore
									
									
									
									
										vendored
									
									
								
							| @@ -1,7 +1,3 @@ | |||||||
| # Doxygen stuff |  | ||||||
| html/* |  | ||||||
| latex/* |  | ||||||
|  |  | ||||||
| # Compiled Object files # | # Compiled Object files # | ||||||
| ######################### | ######################### | ||||||
| *.slo | *.slo | ||||||
| @@ -92,7 +88,6 @@ Thumbs.db | |||||||
| # build directory # | # build directory # | ||||||
| ################### | ################### | ||||||
| build*/* | build*/* | ||||||
| Documentation/_build |  | ||||||
|  |  | ||||||
| # IDE related files # | # IDE related files # | ||||||
| ##################### | ##################### | ||||||
|   | |||||||
							
								
								
									
										56
									
								
								.travis.yml
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										56
									
								
								.travis.yml
									
									
									
									
									
										Normal file
									
								
							| @@ -0,0 +1,56 @@ | |||||||
|  | language: cpp | ||||||
|  |  | ||||||
|  | cache: | ||||||
|  |   directories: | ||||||
|  |     - clang | ||||||
|  |  | ||||||
|  | matrix: | ||||||
|  |   include: | ||||||
|  |     - os:        osx | ||||||
|  |       osx_image: xcode8.3 | ||||||
|  |       compiler: clang | ||||||
|  |        | ||||||
|  | before_install: | ||||||
|  |     - export GRIDDIR=`pwd` | ||||||
|  |     - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]] && [ ! -e clang/bin ]; then wget $CLANG_LINK; tar -xf `basename $CLANG_LINK`; mkdir clang; mv clang+*/* clang/; fi | ||||||
|  |     - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export PATH="${GRIDDIR}/clang/bin:${PATH}"; fi | ||||||
|  |     - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export LD_LIBRARY_PATH="${GRIDDIR}/clang/lib:${LD_LIBRARY_PATH}"; fi | ||||||
|  |     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew update; fi | ||||||
|  |     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc openssl; fi | ||||||
|  |      | ||||||
|  | install: | ||||||
|  |     - export CWD=`pwd` | ||||||
|  |     - echo $CWD | ||||||
|  |     - export CC=$CC$VERSION | ||||||
|  |     - export CXX=$CXX$VERSION | ||||||
|  |     - echo $PATH | ||||||
|  |     - which autoconf | ||||||
|  |     - autoconf  --version | ||||||
|  |     - which automake | ||||||
|  |     - automake  --version | ||||||
|  |     - which $CC | ||||||
|  |     - $CC  --version | ||||||
|  |     - which $CXX | ||||||
|  |     - $CXX --version | ||||||
|  |     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export LDFLAGS='-L/usr/local/lib'; fi | ||||||
|  |     - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then export EXTRACONF='--with-openssl=/usr/local/opt/openssl'; fi | ||||||
|  |      | ||||||
|  | script: | ||||||
|  |     - ./bootstrap.sh | ||||||
|  |     - mkdir build | ||||||
|  |     - cd build | ||||||
|  |     - mkdir lime | ||||||
|  |     - cd lime | ||||||
|  |     - mkdir build | ||||||
|  |     - cd build | ||||||
|  |     - wget http://usqcd-software.github.io/downloads/c-lime/lime-1.3.2.tar.gz | ||||||
|  |     - tar xf lime-1.3.2.tar.gz | ||||||
|  |     - cd lime-1.3.2 | ||||||
|  |     - ./configure --prefix=$CWD/build/lime/install | ||||||
|  |     - make -j4 | ||||||
|  |     - make install | ||||||
|  |     - cd $CWD/build | ||||||
|  |     - ../configure --enable-simd=SSE4 --enable-comms=none --with-lime=$CWD/build/lime/install ${EXTRACONF} | ||||||
|  |     - make -j4  | ||||||
|  |     - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals | ||||||
|  |     - make check | ||||||
| @@ -37,29 +37,19 @@ directory | |||||||
| #endif | #endif | ||||||
|  |  | ||||||
|  //disables and intel compiler specific warning (in json.hpp) |  //disables and intel compiler specific warning (in json.hpp) | ||||||
| #ifdef __ICC |  | ||||||
| #pragma warning disable 488   | #pragma warning disable 488   | ||||||
| #endif |  | ||||||
|  |  | ||||||
| #ifdef __NVCC__ | #ifdef __NVCC__ | ||||||
|  //disables nvcc specific warning in json.hpp |  //disables nvcc specific warning in json.hpp | ||||||
| #pragma clang diagnostic ignored "-Wdeprecated-register" | #pragma clang diagnostic ignored "-Wdeprecated-register" | ||||||
|  |  | ||||||
| #ifdef __NVCC_DIAG_PRAGMA_SUPPORT__ |  | ||||||
|  //disables nvcc specific warning in json.hpp |  | ||||||
| #pragma nv_diag_suppress unsigned_compare_with_zero |  | ||||||
| #pragma nv_diag_suppress cast_to_qualified_type |  | ||||||
|  //disables nvcc specific warning in many files |  | ||||||
| #pragma nv_diag_suppress esa_on_defaulted_function_ignored |  | ||||||
| #pragma nv_diag_suppress extra_semicolon |  | ||||||
| #else |  | ||||||
|  //disables nvcc specific warning in json.hpp |  | ||||||
| #pragma diag_suppress unsigned_compare_with_zero | #pragma diag_suppress unsigned_compare_with_zero | ||||||
| #pragma diag_suppress cast_to_qualified_type | #pragma diag_suppress cast_to_qualified_type | ||||||
|  |  | ||||||
|  //disables nvcc specific warning in many files |  //disables nvcc specific warning in many files | ||||||
| #pragma diag_suppress esa_on_defaulted_function_ignored | #pragma diag_suppress esa_on_defaulted_function_ignored | ||||||
| #pragma diag_suppress extra_semicolon | #pragma diag_suppress extra_semicolon | ||||||
| #endif |  | ||||||
|  | //Eigen only | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
| // Disable vectorisation in Eigen on the Power8/9 and PowerPC | // Disable vectorisation in Eigen on the Power8/9 and PowerPC | ||||||
|   | |||||||
| @@ -44,10 +44,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk> | |||||||
| #include <Grid/GridStd.h> | #include <Grid/GridStd.h> | ||||||
| #include <Grid/threads/Pragmas.h> | #include <Grid/threads/Pragmas.h> | ||||||
| #include <Grid/perfmon/Timer.h> | #include <Grid/perfmon/Timer.h> | ||||||
| //#include <Grid/perfmon/PerfCount.h> | #include <Grid/perfmon/PerfCount.h> | ||||||
| #include <Grid/util/Util.h> | #include <Grid/util/Util.h> | ||||||
| #include <Grid/log/Log.h> | #include <Grid/log/Log.h> | ||||||
| #include <Grid/perfmon/Tracing.h> |  | ||||||
| #include <Grid/allocator/Allocator.h> | #include <Grid/allocator/Allocator.h> | ||||||
| #include <Grid/simd/Simd.h> | #include <Grid/simd/Simd.h> | ||||||
| #include <Grid/threads/ThreadReduction.h> | #include <Grid/threads/ThreadReduction.h> | ||||||
|   | |||||||
| @@ -36,7 +36,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk> | |||||||
| #include <Grid/GridCore.h> | #include <Grid/GridCore.h> | ||||||
| #include <Grid/qcd/QCD.h> | #include <Grid/qcd/QCD.h> | ||||||
| #include <Grid/qcd/spin/Spin.h> | #include <Grid/qcd/spin/Spin.h> | ||||||
| #include <Grid/qcd/gparity/Gparity.h> |  | ||||||
| #include <Grid/qcd/utils/Utils.h> | #include <Grid/qcd/utils/Utils.h> | ||||||
| #include <Grid/qcd/representations/Representations.h> | #include <Grid/qcd/representations/Representations.h> | ||||||
| NAMESPACE_CHECK(GridQCDCore); | NAMESPACE_CHECK(GridQCDCore); | ||||||
|   | |||||||
| @@ -16,7 +16,6 @@ | |||||||
| #include <functional> | #include <functional> | ||||||
| #include <stdio.h> | #include <stdio.h> | ||||||
| #include <stdlib.h> | #include <stdlib.h> | ||||||
| #include <strings.h> |  | ||||||
| #include <stdio.h> | #include <stdio.h> | ||||||
| #include <signal.h> | #include <signal.h> | ||||||
| #include <ctime> | #include <ctime> | ||||||
|   | |||||||
| @@ -14,11 +14,7 @@ | |||||||
| /* NVCC save and restore compile environment*/ | /* NVCC save and restore compile environment*/ | ||||||
| #ifdef __NVCC__ | #ifdef __NVCC__ | ||||||
| #pragma push | #pragma push | ||||||
| #ifdef __NVCC_DIAG_PRAGMA_SUPPORT__ |  | ||||||
| #pragma nv_diag_suppress code_is_unreachable |  | ||||||
| #else |  | ||||||
| #pragma diag_suppress code_is_unreachable | #pragma diag_suppress code_is_unreachable | ||||||
| #endif |  | ||||||
| #pragma push_macro("__CUDA_ARCH__") | #pragma push_macro("__CUDA_ARCH__") | ||||||
| #pragma push_macro("__NVCC__") | #pragma push_macro("__NVCC__") | ||||||
| #pragma push_macro("__CUDACC__") | #pragma push_macro("__CUDACC__") | ||||||
| @@ -34,7 +30,7 @@ | |||||||
| #pragma push_macro("__SYCL_DEVICE_ONLY__") | #pragma push_macro("__SYCL_DEVICE_ONLY__") | ||||||
| #undef __SYCL_DEVICE_ONLY__ | #undef __SYCL_DEVICE_ONLY__ | ||||||
| #define EIGEN_DONT_VECTORIZE | #define EIGEN_DONT_VECTORIZE | ||||||
| #undef EIGEN_USE_SYCL | //#undef EIGEN_USE_SYCL | ||||||
| #define __SYCL__REDEFINE__ | #define __SYCL__REDEFINE__ | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
|   | |||||||
| @@ -54,11 +54,9 @@ Version.h: version-cache | |||||||
| include Make.inc | include Make.inc | ||||||
| include Eigen.inc | include Eigen.inc | ||||||
|  |  | ||||||
|  | extra_sources+=$(ZWILS_FERMION_FILES) | ||||||
| extra_sources+=$(WILS_FERMION_FILES) | extra_sources+=$(WILS_FERMION_FILES) | ||||||
| extra_sources+=$(STAG_FERMION_FILES) | extra_sources+=$(STAG_FERMION_FILES) | ||||||
| if BUILD_ZMOBIUS |  | ||||||
|   extra_sources+=$(ZWILS_FERMION_FILES) |  | ||||||
| endif |  | ||||||
| if BUILD_GPARITY | if BUILD_GPARITY | ||||||
|   extra_sources+=$(GP_FERMION_FILES) |   extra_sources+=$(GP_FERMION_FILES) | ||||||
| endif | endif | ||||||
| @@ -66,10 +64,6 @@ if BUILD_FERMION_REPS | |||||||
|   extra_sources+=$(ADJ_FERMION_FILES) |   extra_sources+=$(ADJ_FERMION_FILES) | ||||||
|   extra_sources+=$(TWOIND_FERMION_FILES) |   extra_sources+=$(TWOIND_FERMION_FILES) | ||||||
| endif | endif | ||||||
| if BUILD_SP |  | ||||||
|     extra_sources+=$(SP_FERMION_FILES) |  | ||||||
|     extra_sources+=$(SP_TWOIND_FERMION_FILES) |  | ||||||
| endif |  | ||||||
|  |  | ||||||
| lib_LIBRARIES = libGrid.a | lib_LIBRARIES = libGrid.a | ||||||
|  |  | ||||||
|   | |||||||
| @@ -54,8 +54,6 @@ NAMESPACE_CHECK(BiCGSTAB); | |||||||
| #include <Grid/algorithms/iterative/SchurRedBlack.h> | #include <Grid/algorithms/iterative/SchurRedBlack.h> | ||||||
| #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h> | #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h> | ||||||
| #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h> | #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h> | ||||||
| #include <Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h> |  | ||||||
| #include <Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h> |  | ||||||
| #include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h> | #include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h> | ||||||
| #include <Grid/algorithms/iterative/BlockConjugateGradient.h> | #include <Grid/algorithms/iterative/BlockConjugateGradient.h> | ||||||
| #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h> | #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h> | ||||||
|   | |||||||
| @@ -31,7 +31,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk> | |||||||
| #ifndef  GRID_ALGORITHM_COARSENED_MATRIX_H | #ifndef  GRID_ALGORITHM_COARSENED_MATRIX_H | ||||||
| #define  GRID_ALGORITHM_COARSENED_MATRIX_H | #define  GRID_ALGORITHM_COARSENED_MATRIX_H | ||||||
|  |  | ||||||
| #include <Grid/qcd/QCD.h> // needed for Dagger(Yes|No), Inverse(Yes|No) |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| @@ -60,14 +59,12 @@ inline void blockMaskedInnerProduct(Lattice<CComplex> &CoarseInner, | |||||||
| class Geometry { | class Geometry { | ||||||
| public: | public: | ||||||
|   int npoint; |   int npoint; | ||||||
|   int base; |  | ||||||
|   std::vector<int> directions   ; |   std::vector<int> directions   ; | ||||||
|   std::vector<int> displacements; |   std::vector<int> displacements; | ||||||
|   std::vector<int> points_dagger; |  | ||||||
|  |  | ||||||
|   Geometry(int _d)  { |   Geometry(int _d)  { | ||||||
|      |      | ||||||
|     base = (_d==5) ? 1:0; |     int base = (_d==5) ? 1:0; | ||||||
|  |  | ||||||
|     // make coarse grid stencil for 4d , not 5d |     // make coarse grid stencil for 4d , not 5d | ||||||
|     if ( _d==5 ) _d=4; |     if ( _d==5 ) _d=4; | ||||||
| @@ -75,51 +72,16 @@ public: | |||||||
|     npoint = 2*_d+1; |     npoint = 2*_d+1; | ||||||
|     directions.resize(npoint); |     directions.resize(npoint); | ||||||
|     displacements.resize(npoint); |     displacements.resize(npoint); | ||||||
|     points_dagger.resize(npoint); |  | ||||||
|     for(int d=0;d<_d;d++){ |     for(int d=0;d<_d;d++){ | ||||||
|       directions[d   ] = d+base; |       directions[d   ] = d+base; | ||||||
|       directions[d+_d] = d+base; |       directions[d+_d] = d+base; | ||||||
|       displacements[d  ] = +1; |       displacements[d  ] = +1; | ||||||
|       displacements[d+_d]= -1; |       displacements[d+_d]= -1; | ||||||
|       points_dagger[d   ] = d+_d; |  | ||||||
|       points_dagger[d+_d] = d; |  | ||||||
|     } |     } | ||||||
|     directions   [2*_d]=0; |     directions   [2*_d]=0; | ||||||
|     displacements[2*_d]=0; |     displacements[2*_d]=0; | ||||||
|     points_dagger[2*_d]=2*_d; |  | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   int point(int dir, int disp) { |  | ||||||
|     assert(disp == -1 || disp == 0 || disp == 1); |  | ||||||
|     assert(base+0 <= dir && dir < base+4); |  | ||||||
|  |  | ||||||
|     // directions faster index = new indexing |  | ||||||
|     // 4d (base = 0): |  | ||||||
|     // point 0  1  2  3  4  5  6  7  8 |  | ||||||
|     // dir   0  1  2  3  0  1  2  3  0 |  | ||||||
|     // disp +1 +1 +1 +1 -1 -1 -1 -1  0 |  | ||||||
|     // 5d (base = 1): |  | ||||||
|     // point 0  1  2  3  4  5  6  7  8 |  | ||||||
|     // dir   1  2  3  4  1  2  3  4  0 |  | ||||||
|     // disp +1 +1 +1 +1 -1 -1 -1 -1  0 |  | ||||||
|  |  | ||||||
|     // displacements faster index = old indexing |  | ||||||
|     // 4d (base = 0): |  | ||||||
|     // point 0  1  2  3  4  5  6  7  8 |  | ||||||
|     // dir   0  0  1  1  2  2  3  3  0 |  | ||||||
|     // disp +1 -1 +1 -1 +1 -1 +1 -1  0 |  | ||||||
|     // 5d (base = 1): |  | ||||||
|     // point 0  1  2  3  4  5  6  7  8 |  | ||||||
|     // dir   1  1  2  2  3  3  4  4  0 |  | ||||||
|     // disp +1 -1 +1 -1 +1 -1 +1 -1  0 |  | ||||||
|  |  | ||||||
|     if(dir == 0 and disp == 0) |  | ||||||
|       return 8; |  | ||||||
|     else // New indexing |  | ||||||
|       return (1 - disp) / 2 * 4 + dir - base; |  | ||||||
|     // else // Old indexing |  | ||||||
|     //   return (4 * (dir - base) + 1 - disp) / 2; |  | ||||||
|   } |  | ||||||
| }; | }; | ||||||
|    |    | ||||||
| template<class Fobj,class CComplex,int nbasis> | template<class Fobj,class CComplex,int nbasis> | ||||||
| @@ -262,7 +224,7 @@ public: | |||||||
| 	autoView( Tnp_v , (*Tnp), AcceleratorWrite); | 	autoView( Tnp_v , (*Tnp), AcceleratorWrite); | ||||||
| 	autoView( Tnm_v , (*Tnm), AcceleratorWrite); | 	autoView( Tnm_v , (*Tnm), AcceleratorWrite); | ||||||
| 	const int Nsimd = CComplex::Nsimd(); | 	const int Nsimd = CComplex::Nsimd(); | ||||||
| 	accelerator_for(ss, FineGrid->oSites(), Nsimd, { | 	accelerator_forNB(ss, FineGrid->oSites(), Nsimd, { | ||||||
| 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss)); | 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss)); | ||||||
| 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss)); | 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss)); | ||||||
|         }); |         }); | ||||||
| @@ -296,7 +258,7 @@ public: | |||||||
| // Fine Object == (per site) type of fine field | // Fine Object == (per site) type of fine field | ||||||
| // nbasis      == number of deflation vectors | // nbasis      == number of deflation vectors | ||||||
| template<class Fobj,class CComplex,int nbasis> | template<class Fobj,class CComplex,int nbasis> | ||||||
| class CoarsenedMatrix : public CheckerBoardedSparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  { | class CoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > >  { | ||||||
| public: | public: | ||||||
|      |      | ||||||
|   typedef iVector<CComplex,nbasis >           siteVector; |   typedef iVector<CComplex,nbasis >           siteVector; | ||||||
| @@ -306,59 +268,33 @@ public: | |||||||
|   typedef iMatrix<CComplex,nbasis >  Cobj; |   typedef iMatrix<CComplex,nbasis >  Cobj; | ||||||
|   typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field |   typedef Lattice< CComplex >   CoarseScalar; // used for inner products on fine field | ||||||
|   typedef Lattice<Fobj >        FineField; |   typedef Lattice<Fobj >        FineField; | ||||||
|   typedef CoarseVector FermionField; |  | ||||||
|  |  | ||||||
|   // enrich interface, use default implementation as in FermionOperator /////// |  | ||||||
|   void Dminus(CoarseVector const& in, CoarseVector& out) { out = in; } |  | ||||||
|   void DminusDag(CoarseVector const& in, CoarseVector& out) { out = in; } |  | ||||||
|   void ImportPhysicalFermionSource(CoarseVector const& input, CoarseVector& imported) { imported = input; } |  | ||||||
|   void ImportUnphysicalFermion(CoarseVector const& input, CoarseVector& imported) { imported = input; } |  | ||||||
|   void ExportPhysicalFermionSolution(CoarseVector const& solution, CoarseVector& exported) { exported = solution; }; |  | ||||||
|   void ExportPhysicalFermionSource(CoarseVector const& solution, CoarseVector& exported) { exported = solution; }; |  | ||||||
|  |  | ||||||
|   //////////////////// |   //////////////////// | ||||||
|   // Data members |   // Data members | ||||||
|   //////////////////// |   //////////////////// | ||||||
|   Geometry         geom; |   Geometry         geom; | ||||||
|   GridBase *       _grid;  |   GridBase *       _grid;  | ||||||
|   GridBase*        _cbgrid; |  | ||||||
|   int hermitian; |   int hermitian; | ||||||
|  |  | ||||||
|   CartesianStencil<siteVector,siteVector,DefaultImplParams> Stencil;  |   CartesianStencil<siteVector,siteVector,int> Stencil;  | ||||||
|   CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilEven; |  | ||||||
|   CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilOdd; |  | ||||||
|  |  | ||||||
|   std::vector<CoarseMatrix> A; |   std::vector<CoarseMatrix> A; | ||||||
|   std::vector<CoarseMatrix> Aeven; |      | ||||||
|   std::vector<CoarseMatrix> Aodd; |  | ||||||
|  |  | ||||||
|   CoarseMatrix AselfInv; |  | ||||||
|   CoarseMatrix AselfInvEven; |  | ||||||
|   CoarseMatrix AselfInvOdd; |  | ||||||
|  |  | ||||||
|   Vector<RealD> dag_factor; |  | ||||||
|  |  | ||||||
|   /////////////////////// |   /////////////////////// | ||||||
|   // Interface |   // Interface | ||||||
|   /////////////////////// |   /////////////////////// | ||||||
|   GridBase * Grid(void)         { return _grid; };   // this is all the linalg routines need to know |   GridBase * Grid(void)         { return _grid; };   // this is all the linalg routines need to know | ||||||
|   GridBase * RedBlackGrid()     { return _cbgrid; }; |  | ||||||
|  |  | ||||||
|   int ConstEE() { return 0; } |  | ||||||
|  |  | ||||||
|   void M (const CoarseVector &in, CoarseVector &out) |   void M (const CoarseVector &in, CoarseVector &out) | ||||||
|   { |   { | ||||||
|     conformable(_grid,in.Grid()); |     conformable(_grid,in.Grid()); | ||||||
|     conformable(in.Grid(),out.Grid()); |     conformable(in.Grid(),out.Grid()); | ||||||
|     out.Checkerboard() = in.Checkerboard(); |  | ||||||
|  |  | ||||||
|     SimpleCompressor<siteVector> compressor; |     SimpleCompressor<siteVector> compressor; | ||||||
|  |  | ||||||
|     Stencil.HaloExchange(in,compressor); |     Stencil.HaloExchange(in,compressor); | ||||||
|     autoView( in_v , in, AcceleratorRead); |     autoView( in_v , in, AcceleratorRead); | ||||||
|     autoView( out_v , out, AcceleratorWrite); |     autoView( out_v , out, AcceleratorWrite); | ||||||
|     autoView( Stencil_v  , Stencil, AcceleratorRead); |  | ||||||
|     int npoint = geom.npoint; |  | ||||||
|     typedef LatticeView<Cobj> Aview; |     typedef LatticeView<Cobj> Aview; | ||||||
|        |        | ||||||
|     Vector<Aview> AcceleratorViewContainer; |     Vector<Aview> AcceleratorViewContainer; | ||||||
| @@ -380,14 +316,14 @@ public: | |||||||
|       int ptype; |       int ptype; | ||||||
|       StencilEntry *SE; |       StencilEntry *SE; | ||||||
|  |  | ||||||
|       for(int point=0;point<npoint;point++){ |       for(int point=0;point<geom.npoint;point++){ | ||||||
|  |  | ||||||
| 	SE=Stencil_v.GetEntry(ptype,point,ss); | 	SE=Stencil.GetEntry(ptype,point,ss); | ||||||
| 	   | 	   | ||||||
| 	if(SE->_is_local) {  | 	if(SE->_is_local) {  | ||||||
| 	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); | 	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); | ||||||
| 	} else { | 	} else { | ||||||
| 	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]); | 	  nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]); | ||||||
| 	} | 	} | ||||||
| 	acceleratorSynchronise(); | 	acceleratorSynchronise(); | ||||||
|  |  | ||||||
| @@ -408,74 +344,12 @@ public: | |||||||
|       return M(in,out); |       return M(in,out); | ||||||
|     } else { |     } else { | ||||||
|       // corresponds to Galerkin coarsening |       // corresponds to Galerkin coarsening | ||||||
|       return MdagNonHermitian(in, out); |       CoarseVector tmp(Grid()); | ||||||
|  |       G5C(tmp, in);  | ||||||
|  |       M(tmp, out); | ||||||
|  |       G5C(out, out); | ||||||
|     } |     } | ||||||
|   }; |   }; | ||||||
|  |  | ||||||
|   void MdagNonHermitian(const CoarseVector &in, CoarseVector &out) |  | ||||||
|   { |  | ||||||
|     conformable(_grid,in.Grid()); |  | ||||||
|     conformable(in.Grid(),out.Grid()); |  | ||||||
|     out.Checkerboard() = in.Checkerboard(); |  | ||||||
|  |  | ||||||
|     SimpleCompressor<siteVector> compressor; |  | ||||||
|  |  | ||||||
|     Stencil.HaloExchange(in,compressor); |  | ||||||
|     autoView( in_v , in, AcceleratorRead); |  | ||||||
|     autoView( out_v , out, AcceleratorWrite); |  | ||||||
|     autoView( Stencil_v  , Stencil, AcceleratorRead); |  | ||||||
|     int npoint = geom.npoint; |  | ||||||
|     typedef LatticeView<Cobj> Aview; |  | ||||||
|  |  | ||||||
|     Vector<Aview> AcceleratorViewContainer; |  | ||||||
|  |  | ||||||
|     for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead)); |  | ||||||
|     Aview *Aview_p = & AcceleratorViewContainer[0]; |  | ||||||
|  |  | ||||||
|     const int Nsimd = CComplex::Nsimd(); |  | ||||||
|     typedef decltype(coalescedRead(in_v[0])) calcVector; |  | ||||||
|     typedef decltype(coalescedRead(in_v[0](0))) calcComplex; |  | ||||||
|  |  | ||||||
|     int osites=Grid()->oSites(); |  | ||||||
|  |  | ||||||
|     Vector<int> points(geom.npoint, 0); |  | ||||||
|     for(int p=0; p<geom.npoint; p++) |  | ||||||
|       points[p] = geom.points_dagger[p]; |  | ||||||
|  |  | ||||||
|     auto points_p = &points[0]; |  | ||||||
|  |  | ||||||
|     RealD* dag_factor_p = &dag_factor[0]; |  | ||||||
|  |  | ||||||
|     accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, { |  | ||||||
|       int ss = sss/nbasis; |  | ||||||
|       int b  = sss%nbasis; |  | ||||||
|       calcComplex res = Zero(); |  | ||||||
|       calcVector nbr; |  | ||||||
|       int ptype; |  | ||||||
|       StencilEntry *SE; |  | ||||||
|  |  | ||||||
|       for(int p=0;p<npoint;p++){ |  | ||||||
|         int point = points_p[p]; |  | ||||||
|  |  | ||||||
| 	SE=Stencil_v.GetEntry(ptype,point,ss); |  | ||||||
|  |  | ||||||
| 	if(SE->_is_local) { |  | ||||||
| 	  nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); |  | ||||||
| 	} else { |  | ||||||
| 	  nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]); |  | ||||||
| 	} |  | ||||||
| 	acceleratorSynchronise(); |  | ||||||
|  |  | ||||||
| 	for(int bb=0;bb<nbasis;bb++) { |  | ||||||
| 	  res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb); |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|       coalescedWrite(out_v[ss](b),res); |  | ||||||
|       }); |  | ||||||
|  |  | ||||||
|     for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose(); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void MdirComms(const CoarseVector &in) |   void MdirComms(const CoarseVector &in) | ||||||
|   { |   { | ||||||
|     SimpleCompressor<siteVector> compressor; |     SimpleCompressor<siteVector> compressor; | ||||||
| @@ -485,7 +359,6 @@ public: | |||||||
|   { |   { | ||||||
|     conformable(_grid,in.Grid()); |     conformable(_grid,in.Grid()); | ||||||
|     conformable(_grid,out.Grid()); |     conformable(_grid,out.Grid()); | ||||||
|     out.Checkerboard() = in.Checkerboard(); |  | ||||||
|  |  | ||||||
|     typedef LatticeView<Cobj> Aview; |     typedef LatticeView<Cobj> Aview; | ||||||
|     Vector<Aview> AcceleratorViewContainer; |     Vector<Aview> AcceleratorViewContainer; | ||||||
| @@ -494,7 +367,6 @@ public: | |||||||
|  |  | ||||||
|     autoView( out_v , out, AcceleratorWrite); |     autoView( out_v , out, AcceleratorWrite); | ||||||
|     autoView( in_v  , in, AcceleratorRead); |     autoView( in_v  , in, AcceleratorRead); | ||||||
|     autoView( Stencil_v  , Stencil, AcceleratorRead); |  | ||||||
|  |  | ||||||
|     const int Nsimd = CComplex::Nsimd(); |     const int Nsimd = CComplex::Nsimd(); | ||||||
|     typedef decltype(coalescedRead(in_v[0])) calcVector; |     typedef decltype(coalescedRead(in_v[0])) calcVector; | ||||||
| @@ -508,12 +380,12 @@ public: | |||||||
|       int ptype; |       int ptype; | ||||||
|       StencilEntry *SE; |       StencilEntry *SE; | ||||||
|  |  | ||||||
|       SE=Stencil_v.GetEntry(ptype,point,ss); |       SE=Stencil.GetEntry(ptype,point,ss); | ||||||
| 	   | 	   | ||||||
|       if(SE->_is_local) {  |       if(SE->_is_local) {  | ||||||
| 	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); | 	nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); | ||||||
|       } else { |       } else { | ||||||
| 	nbr = coalescedRead(Stencil_v.CommBuf()[SE->_offset]); | 	nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]); | ||||||
|       } |       } | ||||||
|       acceleratorSynchronise(); |       acceleratorSynchronise(); | ||||||
|  |  | ||||||
| @@ -541,7 +413,34 @@ public: | |||||||
|  |  | ||||||
|     this->MdirComms(in); |     this->MdirComms(in); | ||||||
|  |  | ||||||
|     MdirCalc(in,out,geom.point(dir,disp)); |     int ndim = in.Grid()->Nd(); | ||||||
|  |  | ||||||
|  |     ////////////// | ||||||
|  |     // 4D action like wilson | ||||||
|  |     // 0+ => 0  | ||||||
|  |     // 0- => 1 | ||||||
|  |     // 1+ => 2  | ||||||
|  |     // 1- => 3 | ||||||
|  |     // etc.. | ||||||
|  |     ////////////// | ||||||
|  |     // 5D action like DWF | ||||||
|  |     // 1+ => 0  | ||||||
|  |     // 1- => 1 | ||||||
|  |     // 2+ => 2  | ||||||
|  |     // 2- => 3 | ||||||
|  |     // etc.. | ||||||
|  |     auto point = [dir, disp, ndim](){ | ||||||
|  |       if(dir == 0 and disp == 0) | ||||||
|  | 	return 8; | ||||||
|  |       else if ( ndim==4 ) {  | ||||||
|  | 	return (4 * dir + 1 - disp) / 2; | ||||||
|  |       } else {  | ||||||
|  | 	return (4 * (dir-1) + 1 - disp) / 2; | ||||||
|  |       } | ||||||
|  |     }(); | ||||||
|  |  | ||||||
|  |     MdirCalc(in,out,point); | ||||||
|  |  | ||||||
|   }; |   }; | ||||||
|  |  | ||||||
|   void Mdiag(const CoarseVector &in, CoarseVector &out) |   void Mdiag(const CoarseVector &in, CoarseVector &out) | ||||||
| @@ -550,298 +449,23 @@ public: | |||||||
|     MdirCalc(in, out, point); // No comms |     MdirCalc(in, out, point); // No comms | ||||||
|   }; |   }; | ||||||
|  |  | ||||||
|   void Mooee(const CoarseVector &in, CoarseVector &out) { |  | ||||||
|     MooeeInternal(in, out, DaggerNo, InverseNo); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void MooeeInv(const CoarseVector &in, CoarseVector &out) { |  | ||||||
|     MooeeInternal(in, out, DaggerNo, InverseYes); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void MooeeDag(const CoarseVector &in, CoarseVector &out) { |  | ||||||
|     MooeeInternal(in, out, DaggerYes, InverseNo); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void MooeeInvDag(const CoarseVector &in, CoarseVector &out) { |  | ||||||
|     MooeeInternal(in, out, DaggerYes, InverseYes); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void Meooe(const CoarseVector &in, CoarseVector &out) { |  | ||||||
|     if(in.Checkerboard() == Odd) { |  | ||||||
|       DhopEO(in, out, DaggerNo); |  | ||||||
|     } else { |  | ||||||
|       DhopOE(in, out, DaggerNo); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void MeooeDag(const CoarseVector &in, CoarseVector &out) { |  | ||||||
|     if(in.Checkerboard() == Odd) { |  | ||||||
|       DhopEO(in, out, DaggerYes); |  | ||||||
|     } else { |  | ||||||
|       DhopOE(in, out, DaggerYes); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void Dhop(const CoarseVector &in, CoarseVector &out, int dag) { |  | ||||||
|     conformable(in.Grid(), _grid); // verifies full grid |  | ||||||
|     conformable(in.Grid(), out.Grid()); |  | ||||||
|  |  | ||||||
|     out.Checkerboard() = in.Checkerboard(); |  | ||||||
|  |  | ||||||
|     DhopInternal(Stencil, A, in, out, dag); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void DhopOE(const CoarseVector &in, CoarseVector &out, int dag) { |  | ||||||
|     conformable(in.Grid(), _cbgrid);    // verifies half grid |  | ||||||
|     conformable(in.Grid(), out.Grid()); // drops the cb check |  | ||||||
|  |  | ||||||
|     assert(in.Checkerboard() == Even); |  | ||||||
|     out.Checkerboard() = Odd; |  | ||||||
|  |  | ||||||
|     DhopInternal(StencilEven, Aodd, in, out, dag); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void DhopEO(const CoarseVector &in, CoarseVector &out, int dag) { |  | ||||||
|     conformable(in.Grid(), _cbgrid);    // verifies half grid |  | ||||||
|     conformable(in.Grid(), out.Grid()); // drops the cb check |  | ||||||
|  |  | ||||||
|     assert(in.Checkerboard() == Odd); |  | ||||||
|     out.Checkerboard() = Even; |  | ||||||
|  |  | ||||||
|     DhopInternal(StencilOdd, Aeven, in, out, dag); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void MooeeInternal(const CoarseVector &in, CoarseVector &out, int dag, int inv) { |  | ||||||
|     out.Checkerboard() = in.Checkerboard(); |  | ||||||
|     assert(in.Checkerboard() == Odd || in.Checkerboard() == Even); |  | ||||||
|  |  | ||||||
|     CoarseMatrix *Aself = nullptr; |  | ||||||
|     if(in.Grid()->_isCheckerBoarded) { |  | ||||||
|       if(in.Checkerboard() == Odd) { |  | ||||||
|         Aself = (inv) ? &AselfInvOdd : &Aodd[geom.npoint-1]; |  | ||||||
|         DselfInternal(StencilOdd, *Aself, in, out, dag); |  | ||||||
|       } else { |  | ||||||
|         Aself = (inv) ? &AselfInvEven : &Aeven[geom.npoint-1]; |  | ||||||
|         DselfInternal(StencilEven, *Aself, in, out, dag); |  | ||||||
|       } |  | ||||||
|     } else { |  | ||||||
|       Aself = (inv) ? &AselfInv : &A[geom.npoint-1]; |  | ||||||
|       DselfInternal(Stencil, *Aself, in, out, dag); |  | ||||||
|     } |  | ||||||
|     assert(Aself != nullptr); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void DselfInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, CoarseMatrix &a, |  | ||||||
|                        const CoarseVector &in, CoarseVector &out, int dag) { |  | ||||||
|     int point = geom.npoint-1; |  | ||||||
|     autoView( out_v, out, AcceleratorWrite); |  | ||||||
|     autoView( in_v,  in,  AcceleratorRead); |  | ||||||
|     autoView( st_v,  st,  AcceleratorRead); |  | ||||||
|     autoView( a_v,   a,   AcceleratorRead); |  | ||||||
|  |  | ||||||
|     const int Nsimd = CComplex::Nsimd(); |  | ||||||
|     typedef decltype(coalescedRead(in_v[0])) calcVector; |  | ||||||
|     typedef decltype(coalescedRead(in_v[0](0))) calcComplex; |  | ||||||
|  |  | ||||||
|     RealD* dag_factor_p = &dag_factor[0]; |  | ||||||
|  |  | ||||||
|     if(dag) { |  | ||||||
|       accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, { |  | ||||||
|         int ss = sss/nbasis; |  | ||||||
|         int b  = sss%nbasis; |  | ||||||
|         calcComplex res = Zero(); |  | ||||||
|         calcVector nbr; |  | ||||||
|         int ptype; |  | ||||||
|         StencilEntry *SE; |  | ||||||
|  |  | ||||||
|         SE=st_v.GetEntry(ptype,point,ss); |  | ||||||
|  |  | ||||||
|         if(SE->_is_local) { |  | ||||||
|           nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); |  | ||||||
|         } else { |  | ||||||
|           nbr = coalescedRead(st_v.CommBuf()[SE->_offset]); |  | ||||||
|         } |  | ||||||
|         acceleratorSynchronise(); |  | ||||||
|  |  | ||||||
|         for(int bb=0;bb<nbasis;bb++) { |  | ||||||
|           res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(a_v[ss](b,bb))*nbr(bb); |  | ||||||
|         } |  | ||||||
|         coalescedWrite(out_v[ss](b),res); |  | ||||||
|       }); |  | ||||||
|     } else { |  | ||||||
|       accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, { |  | ||||||
|         int ss = sss/nbasis; |  | ||||||
|         int b  = sss%nbasis; |  | ||||||
|         calcComplex res = Zero(); |  | ||||||
|         calcVector nbr; |  | ||||||
|         int ptype; |  | ||||||
|         StencilEntry *SE; |  | ||||||
|  |  | ||||||
|         SE=st_v.GetEntry(ptype,point,ss); |  | ||||||
|  |  | ||||||
|         if(SE->_is_local) { |  | ||||||
|           nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); |  | ||||||
|         } else { |  | ||||||
|           nbr = coalescedRead(st_v.CommBuf()[SE->_offset]); |  | ||||||
|         } |  | ||||||
|         acceleratorSynchronise(); |  | ||||||
|  |  | ||||||
|         for(int bb=0;bb<nbasis;bb++) { |  | ||||||
|           res = res + coalescedRead(a_v[ss](b,bb))*nbr(bb); |  | ||||||
|         } |  | ||||||
|         coalescedWrite(out_v[ss](b),res); |  | ||||||
|       }); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void DhopInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, std::vector<CoarseMatrix> &a, |  | ||||||
|                     const CoarseVector &in, CoarseVector &out, int dag) { |  | ||||||
|     SimpleCompressor<siteVector> compressor; |  | ||||||
|  |  | ||||||
|     st.HaloExchange(in,compressor); |  | ||||||
|     autoView( in_v,  in,  AcceleratorRead); |  | ||||||
|     autoView( out_v, out, AcceleratorWrite); |  | ||||||
|     autoView( st_v , st,  AcceleratorRead); |  | ||||||
|     typedef LatticeView<Cobj> Aview; |  | ||||||
|  |  | ||||||
|     // determine in what order we need the points |  | ||||||
|     int npoint = geom.npoint-1; |  | ||||||
|     Vector<int> points(npoint, 0); |  | ||||||
|     for(int p=0; p<npoint; p++) |  | ||||||
|       points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p; |  | ||||||
|  |  | ||||||
|     auto points_p = &points[0]; |  | ||||||
|  |  | ||||||
|     Vector<Aview> AcceleratorViewContainer; |  | ||||||
|     for(int p=0;p<npoint;p++) AcceleratorViewContainer.push_back(a[p].View(AcceleratorRead)); |  | ||||||
|     Aview *Aview_p = & AcceleratorViewContainer[0]; |  | ||||||
|  |  | ||||||
|     const int Nsimd = CComplex::Nsimd(); |  | ||||||
|     typedef decltype(coalescedRead(in_v[0])) calcVector; |  | ||||||
|     typedef decltype(coalescedRead(in_v[0](0))) calcComplex; |  | ||||||
|  |  | ||||||
|     RealD* dag_factor_p = &dag_factor[0]; |  | ||||||
|  |  | ||||||
|     if(dag) { |  | ||||||
|       accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, { |  | ||||||
|         int ss = sss/nbasis; |  | ||||||
|         int b  = sss%nbasis; |  | ||||||
|         calcComplex res = Zero(); |  | ||||||
|         calcVector nbr; |  | ||||||
|         int ptype; |  | ||||||
|         StencilEntry *SE; |  | ||||||
|  |  | ||||||
|         for(int p=0;p<npoint;p++){ |  | ||||||
|           int point = points_p[p]; |  | ||||||
|           SE=st_v.GetEntry(ptype,point,ss); |  | ||||||
|  |  | ||||||
|           if(SE->_is_local) { |  | ||||||
|             nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); |  | ||||||
|           } else { |  | ||||||
|             nbr = coalescedRead(st_v.CommBuf()[SE->_offset]); |  | ||||||
|           } |  | ||||||
|           acceleratorSynchronise(); |  | ||||||
|  |  | ||||||
|           for(int bb=0;bb<nbasis;bb++) { |  | ||||||
|             res = res + dag_factor_p[b*nbasis+bb]*coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb); |  | ||||||
|           } |  | ||||||
|         } |  | ||||||
|         coalescedWrite(out_v[ss](b),res); |  | ||||||
|       }); |  | ||||||
|     } else { |  | ||||||
|       accelerator_for(sss, in.Grid()->oSites()*nbasis, Nsimd, { |  | ||||||
|         int ss = sss/nbasis; |  | ||||||
|         int b  = sss%nbasis; |  | ||||||
|         calcComplex res = Zero(); |  | ||||||
|         calcVector nbr; |  | ||||||
|         int ptype; |  | ||||||
|         StencilEntry *SE; |  | ||||||
|  |  | ||||||
|         for(int p=0;p<npoint;p++){ |  | ||||||
|           int point = points_p[p]; |  | ||||||
|           SE=st_v.GetEntry(ptype,point,ss); |  | ||||||
|  |  | ||||||
|           if(SE->_is_local) { |  | ||||||
|             nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute); |  | ||||||
|           } else { |  | ||||||
|             nbr = coalescedRead(st_v.CommBuf()[SE->_offset]); |  | ||||||
|           } |  | ||||||
|           acceleratorSynchronise(); |  | ||||||
|  |  | ||||||
|           for(int bb=0;bb<nbasis;bb++) { |  | ||||||
|             res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb); |  | ||||||
|           } |  | ||||||
|         } |  | ||||||
|         coalescedWrite(out_v[ss](b),res); |  | ||||||
|       }); |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     for(int p=0;p<npoint;p++) AcceleratorViewContainer[p].ViewClose(); |  | ||||||
|   } |  | ||||||
|    |    | ||||||
|   CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	: |  CoarsenedMatrix(GridCartesian &CoarseGrid, int hermitian_=0) 	:  | ||||||
|     _grid(&CoarseGrid), |  | ||||||
|     _cbgrid(new GridRedBlackCartesian(&CoarseGrid)), |  | ||||||
|     geom(CoarseGrid._ndimension), |  | ||||||
|     hermitian(hermitian_), |  | ||||||
|     Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements), |  | ||||||
|     StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements), |  | ||||||
|     StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements), |  | ||||||
|     A(geom.npoint,&CoarseGrid), |  | ||||||
|     Aeven(geom.npoint,_cbgrid), |  | ||||||
|     Aodd(geom.npoint,_cbgrid), |  | ||||||
|     AselfInv(&CoarseGrid), |  | ||||||
|     AselfInvEven(_cbgrid), |  | ||||||
|     AselfInvOdd(_cbgrid), |  | ||||||
|     dag_factor(nbasis*nbasis) |  | ||||||
|   { |  | ||||||
|     fillFactor(); |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   CoarsenedMatrix(GridCartesian &CoarseGrid, GridRedBlackCartesian &CoarseRBGrid, int hermitian_=0) 	: |  | ||||||
|  |  | ||||||
|     _grid(&CoarseGrid), |     _grid(&CoarseGrid), | ||||||
|     _cbgrid(&CoarseRBGrid), |  | ||||||
|     geom(CoarseGrid._ndimension), |     geom(CoarseGrid._ndimension), | ||||||
|     hermitian(hermitian_), |     hermitian(hermitian_), | ||||||
|     Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements), |     Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0), | ||||||
|     StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements), |       A(geom.npoint,&CoarseGrid) | ||||||
|     StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements), |  | ||||||
|     A(geom.npoint,&CoarseGrid), |  | ||||||
|     Aeven(geom.npoint,&CoarseRBGrid), |  | ||||||
|     Aodd(geom.npoint,&CoarseRBGrid), |  | ||||||
|     AselfInv(&CoarseGrid), |  | ||||||
|     AselfInvEven(&CoarseRBGrid), |  | ||||||
|     AselfInvOdd(&CoarseRBGrid), |  | ||||||
|     dag_factor(nbasis*nbasis) |  | ||||||
|   { |   { | ||||||
|     fillFactor(); |  | ||||||
|   }; |   }; | ||||||
|  |  | ||||||
|   void fillFactor() { |  | ||||||
|     Eigen::MatrixXd dag_factor_eigen = Eigen::MatrixXd::Ones(nbasis, nbasis); |  | ||||||
|     if(!hermitian) { |  | ||||||
|       const int nb = nbasis/2; |  | ||||||
|       dag_factor_eigen.block(0,nb,nb,nb) *= -1.0; |  | ||||||
|       dag_factor_eigen.block(nb,0,nb,nb) *= -1.0; |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     // GPU readable prefactor |  | ||||||
|     thread_for(i, nbasis*nbasis, { |  | ||||||
|       int j = i/nbasis; |  | ||||||
|       int k = i%nbasis; |  | ||||||
|       dag_factor[i] = dag_factor_eigen(j, k); |  | ||||||
|     }); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop, |   void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop, | ||||||
| 		       Aggregation<Fobj,CComplex,nbasis> & Subspace) | 		       Aggregation<Fobj,CComplex,nbasis> & Subspace) | ||||||
|   { |   { | ||||||
|     typedef Lattice<typename Fobj::tensor_reduced> FineComplexField; |     typedef Lattice<typename Fobj::tensor_reduced> FineComplexField; | ||||||
|     typedef typename Fobj::scalar_type scalar_type; |     typedef typename Fobj::scalar_type scalar_type; | ||||||
|  |  | ||||||
|     std::cout << GridLogMessage<< "CoarsenMatrix "<< std::endl; |  | ||||||
|  |  | ||||||
|     FineComplexField one(FineGrid); one=scalar_type(1.0,0.0); |     FineComplexField one(FineGrid); one=scalar_type(1.0,0.0); | ||||||
|     FineComplexField zero(FineGrid); zero=scalar_type(0.0,0.0); |     FineComplexField zero(FineGrid); zero=scalar_type(0.0,0.0); | ||||||
|  |  | ||||||
| @@ -872,13 +496,11 @@ public: | |||||||
|  |  | ||||||
|     CoarseScalar InnerProd(Grid());  |     CoarseScalar InnerProd(Grid());  | ||||||
|  |  | ||||||
|     std::cout << GridLogMessage<< "CoarsenMatrix Orthog "<< std::endl; |  | ||||||
|     // Orthogonalise the subblocks over the basis |     // Orthogonalise the subblocks over the basis | ||||||
|     blockOrthogonalise(InnerProd,Subspace.subspace); |     blockOrthogonalise(InnerProd,Subspace.subspace); | ||||||
|  |  | ||||||
|     // Compute the matrix elements of linop between this orthonormal |     // Compute the matrix elements of linop between this orthonormal | ||||||
|     // set of vectors. |     // set of vectors. | ||||||
|     std::cout << GridLogMessage<< "CoarsenMatrix masks "<< std::endl; |  | ||||||
|     int self_stencil=-1; |     int self_stencil=-1; | ||||||
|     for(int p=0;p<geom.npoint;p++) |     for(int p=0;p<geom.npoint;p++) | ||||||
|     {  |     {  | ||||||
| @@ -917,7 +539,7 @@ public: | |||||||
|  |  | ||||||
|       phi=Subspace.subspace[i]; |       phi=Subspace.subspace[i]; | ||||||
|  |  | ||||||
|       std::cout << GridLogMessage<< "CoarsenMatrix vector "<<i << std::endl; |       //      std::cout << GridLogMessage<< "CoarsenMatrix vector "<<i << std::endl; | ||||||
|       linop.OpDirAll(phi,Mphi_p); |       linop.OpDirAll(phi,Mphi_p); | ||||||
|       linop.OpDiag  (phi,Mphi_p[geom.npoint-1]); |       linop.OpDiag  (phi,Mphi_p[geom.npoint-1]); | ||||||
|  |  | ||||||
| @@ -946,18 +568,6 @@ public: | |||||||
| 	    autoView( A_self  , A[self_stencil], AcceleratorWrite); | 	    autoView( A_self  , A[self_stencil], AcceleratorWrite); | ||||||
|  |  | ||||||
| 	    accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); }); | 	    accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); }); | ||||||
| 	    if ( hermitian && (disp==-1) ) { |  | ||||||
| 	      for(int pp=0;pp<geom.npoint;pp++){// Find the opposite link and set <j|A|i> = <i|A|j>* |  | ||||||
| 		int dirp   = geom.directions[pp]; |  | ||||||
| 		int dispp  = geom.displacements[pp]; |  | ||||||
| 		if ( (dirp==dir) && (dispp==1) ){ |  | ||||||
| 		  auto sft = conjugate(Cshift(oZProj,dir,1)); |  | ||||||
| 		  autoView( sft_v    ,  sft  , AcceleratorWrite); |  | ||||||
| 		  autoView( A_pp     ,  A[pp], AcceleratorWrite); |  | ||||||
| 		  accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_pp[ss](i,j),sft_v(ss)); }); |  | ||||||
| 		} |  | ||||||
| 	      } |  | ||||||
| 	    } |  | ||||||
|  |  | ||||||
| 	  } | 	  } | ||||||
| 	} | 	} | ||||||
| @@ -996,54 +606,28 @@ public: | |||||||
|     } |     } | ||||||
|     if(hermitian) { |     if(hermitian) { | ||||||
|       std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl; |       std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl; | ||||||
|  |       ForceHermitian(); | ||||||
|     } |     } | ||||||
|  |  | ||||||
|     InvertSelfStencilLink(); std::cout << GridLogMessage << "Coarse self link inverted" << std::endl; |  | ||||||
|     FillHalfCbs(); std::cout << GridLogMessage << "Coarse half checkerboards filled" << std::endl; |  | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   void InvertSelfStencilLink() { |   void ForceHermitian(void) { | ||||||
|     std::cout << GridLogDebug << "CoarsenedMatrix::InvertSelfStencilLink" << std::endl; |     CoarseMatrix Diff  (Grid()); | ||||||
|     int localVolume = Grid()->lSites(); |     for(int p=0;p<geom.npoint;p++){ | ||||||
|  |       int dir   = geom.directions[p]; | ||||||
|     typedef typename Cobj::scalar_object scalar_object; |       int disp  = geom.displacements[p]; | ||||||
|  |       if(disp==-1) { | ||||||
|     autoView(Aself_v,    A[geom.npoint-1], CpuRead); | 	// Find the opposite link | ||||||
|     autoView(AselfInv_v, AselfInv,         CpuWrite); | 	for(int pp=0;pp<geom.npoint;pp++){ | ||||||
|     thread_for(site, localVolume, { // NOTE: Not able to bring this to GPU because of Eigen + peek/poke | 	  int dirp   = geom.directions[pp]; | ||||||
|       Eigen::MatrixXcd selfLinkEigen    = Eigen::MatrixXcd::Zero(nbasis, nbasis); | 	  int dispp  = geom.displacements[pp]; | ||||||
|       Eigen::MatrixXcd selfLinkInvEigen = Eigen::MatrixXcd::Zero(nbasis, nbasis); | 	  if ( (dirp==dir) && (dispp==1) ){ | ||||||
|  | 	    //	    Diff = adj(Cshift(A[p],dir,1)) - A[pp];  | ||||||
|       scalar_object selfLink    = Zero(); | 	    //	    std::cout << GridLogMessage<<" Replacing stencil leg "<<pp<<" with leg "<<p<< " diff "<<norm2(Diff) <<std::endl; | ||||||
|       scalar_object selfLinkInv = Zero(); | 	    A[pp] = adj(Cshift(A[p],dir,1)); | ||||||
|  | 	  } | ||||||
|       Coordinate lcoor; | 	} | ||||||
|  |       } | ||||||
|       Grid()->LocalIndexToLocalCoor(site, lcoor); |  | ||||||
|       peekLocalSite(selfLink, Aself_v, lcoor); |  | ||||||
|  |  | ||||||
|       for (int i = 0; i < nbasis; ++i) |  | ||||||
|         for (int j = 0; j < nbasis; ++j) |  | ||||||
|           selfLinkEigen(i, j) = static_cast<ComplexD>(TensorRemove(selfLink(i, j))); |  | ||||||
|  |  | ||||||
|       selfLinkInvEigen = selfLinkEigen.inverse(); |  | ||||||
|  |  | ||||||
|       for(int i = 0; i < nbasis; ++i) |  | ||||||
|         for(int j = 0; j < nbasis; ++j) |  | ||||||
|           selfLinkInv(i, j) = selfLinkInvEigen(i, j); |  | ||||||
|  |  | ||||||
|       pokeLocalSite(selfLinkInv, AselfInv_v, lcoor); |  | ||||||
|     }); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void FillHalfCbs() { |  | ||||||
|     std::cout << GridLogDebug << "CoarsenedMatrix::FillHalfCbs" << std::endl; |  | ||||||
|     for(int p = 0; p < geom.npoint; ++p) { |  | ||||||
|       pickCheckerboard(Even, Aeven[p], A[p]); |  | ||||||
|       pickCheckerboard(Odd, Aodd[p], A[p]); |  | ||||||
|     } |     } | ||||||
|     pickCheckerboard(Even, AselfInvEven, AselfInv); |  | ||||||
|     pickCheckerboard(Odd, AselfInvOdd, AselfInv); |  | ||||||
|   } |   } | ||||||
| }; | }; | ||||||
|  |  | ||||||
|   | |||||||
| @@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk> | |||||||
| #define _GRID_FFT_H_ | #define _GRID_FFT_H_ | ||||||
|  |  | ||||||
| #ifdef HAVE_FFTW | #ifdef HAVE_FFTW | ||||||
| #if defined(USE_MKL) || defined(GRID_SYCL) | #ifdef USE_MKL | ||||||
| #include <fftw/fftw3.h> | #include <fftw/fftw3.h> | ||||||
| #else | #else | ||||||
| #include <fftw3.h> | #include <fftw3.h> | ||||||
| @@ -136,7 +136,7 @@ public: | |||||||
|     flops=0; |     flops=0; | ||||||
|     usec =0; |     usec =0; | ||||||
|     Coordinate layout(Nd,1); |     Coordinate layout(Nd,1); | ||||||
|     sgrid = new GridCartesian(dimensions,layout,processors,*grid); |     sgrid = new GridCartesian(dimensions,layout,processors); | ||||||
|   }; |   }; | ||||||
|      |      | ||||||
|   ~FFT ( void)  { |   ~FFT ( void)  { | ||||||
| @@ -182,7 +182,7 @@ public: | |||||||
|     pencil_gd[dim] = G*processors[dim]; |     pencil_gd[dim] = G*processors[dim]; | ||||||
|        |        | ||||||
|     // Pencil global vol LxLxGxLxL per node |     // Pencil global vol LxLxGxLxL per node | ||||||
|     GridCartesian pencil_g(pencil_gd,layout,processors,*vgrid); |     GridCartesian pencil_g(pencil_gd,layout,processors); | ||||||
|        |        | ||||||
|     // Construct pencils |     // Construct pencils | ||||||
|     typedef typename vobj::scalar_object sobj; |     typedef typename vobj::scalar_object sobj; | ||||||
|   | |||||||
| @@ -52,7 +52,6 @@ public: | |||||||
|   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base |   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base | ||||||
|   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0; |   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0; | ||||||
|   virtual void HermOp(const Field &in, Field &out)=0; |   virtual void HermOp(const Field &in, Field &out)=0; | ||||||
|   virtual ~LinearOperatorBase(){}; |  | ||||||
| }; | }; | ||||||
|  |  | ||||||
|  |  | ||||||
| @@ -508,7 +507,7 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> { | |||||||
|   virtual  void MpcDag   (const Field &in, Field &out){ |   virtual  void MpcDag   (const Field &in, Field &out){ | ||||||
|     Mpc(in,out); |     Mpc(in,out); | ||||||
|   } |   } | ||||||
|   virtual void MpcDagMpc(const Field &in, Field &out) { |   virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) { | ||||||
|     assert(0);// Never need with staggered |     assert(0);// Never need with staggered | ||||||
|   } |   } | ||||||
| }; | }; | ||||||
| @@ -526,23 +525,11 @@ public: | |||||||
|       (*this)(Linop,in[k],out[k]); |       (*this)(Linop,in[k],out[k]); | ||||||
|     } |     } | ||||||
|   }; |   }; | ||||||
|   virtual ~OperatorFunction(){}; |  | ||||||
| }; | }; | ||||||
|  |  | ||||||
| template<class Field> class LinearFunction { | template<class Field> class LinearFunction { | ||||||
| public: | public: | ||||||
|   virtual void operator() (const Field &in, Field &out) = 0; |   virtual void operator() (const Field &in, Field &out) = 0; | ||||||
|  |  | ||||||
|   virtual void operator() (const std::vector<Field> &in, std::vector<Field> &out) |  | ||||||
|   { |  | ||||||
|     assert(in.size() == out.size()); |  | ||||||
|  |  | ||||||
|     for (unsigned int i = 0; i < in.size(); ++i) |  | ||||||
|     { |  | ||||||
|       (*this)(in[i], out[i]); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|   virtual ~LinearFunction(){}; |  | ||||||
| }; | }; | ||||||
|  |  | ||||||
| template<class Field> class IdentityLinearFunction : public LinearFunction<Field> { | template<class Field> class IdentityLinearFunction : public LinearFunction<Field> { | ||||||
| @@ -588,7 +575,6 @@ class HermOpOperatorFunction : public OperatorFunction<Field> { | |||||||
| template<typename Field> | template<typename Field> | ||||||
| class PlainHermOp : public LinearFunction<Field> { | class PlainHermOp : public LinearFunction<Field> { | ||||||
| public: | public: | ||||||
|   using LinearFunction<Field>::operator(); |  | ||||||
|   LinearOperatorBase<Field> &_Linop; |   LinearOperatorBase<Field> &_Linop; | ||||||
|        |        | ||||||
|   PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop)  |   PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop)  | ||||||
| @@ -602,7 +588,6 @@ public: | |||||||
| template<typename Field> | template<typename Field> | ||||||
| class FunctionHermOp : public LinearFunction<Field> { | class FunctionHermOp : public LinearFunction<Field> { | ||||||
| public: | public: | ||||||
|   using LinearFunction<Field>::operator();  |  | ||||||
|   OperatorFunction<Field>   & _poly; |   OperatorFunction<Field>   & _poly; | ||||||
|   LinearOperatorBase<Field> &_Linop; |   LinearOperatorBase<Field> &_Linop; | ||||||
|        |        | ||||||
|   | |||||||
| @@ -30,19 +30,13 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk> | |||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| template<class Field> using Preconditioner =  LinearFunction<Field> ; | template<class Field> class Preconditioner :  public LinearFunction<Field> {  | ||||||
|  |  | ||||||
| /* |  | ||||||
| template<class Field> class Preconditioner :  public LinearFunction<Field> { |  | ||||||
|   using LinearFunction<Field>::operator(); |  | ||||||
|   virtual void operator()(const Field &src, Field & psi)=0; |   virtual void operator()(const Field &src, Field & psi)=0; | ||||||
| }; | }; | ||||||
| */ |  | ||||||
|  |  | ||||||
| template<class Field> class TrivialPrecon :  public Preconditioner<Field> {  | template<class Field> class TrivialPrecon :  public Preconditioner<Field> {  | ||||||
| public: | public: | ||||||
|   using Preconditioner<Field>::operator(); |   void operator()(const Field &src, Field & psi){ | ||||||
|   virtual void operator()(const Field &src, Field & psi){ |  | ||||||
|     psi = src; |     psi = src; | ||||||
|   } |   } | ||||||
|   TrivialPrecon(void){}; |   TrivialPrecon(void){}; | ||||||
|   | |||||||
| @@ -48,7 +48,6 @@ public: | |||||||
|   virtual  void Mdiag    (const Field &in, Field &out)=0; |   virtual  void Mdiag    (const Field &in, Field &out)=0; | ||||||
|   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0; |   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0; | ||||||
|   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0; |   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0; | ||||||
|   virtual ~SparseMatrixBase() {}; |  | ||||||
| }; | }; | ||||||
|  |  | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| @@ -73,7 +72,7 @@ public: | |||||||
|   virtual  void MeooeDag    (const Field &in, Field &out)=0; |   virtual  void MeooeDag    (const Field &in, Field &out)=0; | ||||||
|   virtual  void MooeeDag    (const Field &in, Field &out)=0; |   virtual  void MooeeDag    (const Field &in, Field &out)=0; | ||||||
|   virtual  void MooeeInvDag (const Field &in, Field &out)=0; |   virtual  void MooeeInvDag (const Field &in, Field &out)=0; | ||||||
|   virtual ~CheckerBoardedSparseMatrixBase() {}; |  | ||||||
| }; | }; | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|   | |||||||
| @@ -258,12 +258,26 @@ public: | |||||||
|     for(int n=2;n<order;n++){ |     for(int n=2;n<order;n++){ | ||||||
|  |  | ||||||
|       Linop.HermOp(*Tn,y); |       Linop.HermOp(*Tn,y); | ||||||
|  | #if 0 | ||||||
|  |       auto y_v = y.View(); | ||||||
|  |       auto Tn_v = Tn->View(); | ||||||
|  |       auto Tnp_v = Tnp->View(); | ||||||
|  |       auto Tnm_v = Tnm->View(); | ||||||
|  |       constexpr int Nsimd = vector_type::Nsimd(); | ||||||
|  |       accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, { | ||||||
|  | 	  coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss)); | ||||||
|  | 	  coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss)); | ||||||
|  |       }); | ||||||
|  |       if ( Coeffs[n] != 0.0) { | ||||||
|  | 	axpy(out,Coeffs[n],*Tnp,out); | ||||||
|  |       } | ||||||
|  | #else | ||||||
|       axpby(y,xscale,mscale,y,(*Tn)); |       axpby(y,xscale,mscale,y,(*Tn)); | ||||||
|       axpby(*Tnp,2.0,-1.0,y,(*Tnm)); |       axpby(*Tnp,2.0,-1.0,y,(*Tnm)); | ||||||
|       if ( Coeffs[n] != 0.0) { |       if ( Coeffs[n] != 0.0) { | ||||||
| 	axpy(out,Coeffs[n],*Tnp,out); | 	axpy(out,Coeffs[n],*Tnp,out); | ||||||
|       } |       } | ||||||
|  | #endif | ||||||
|       // Cycle pointers to avoid copies |       // Cycle pointers to avoid copies | ||||||
|       Field *swizzle = Tnm; |       Field *swizzle = Tnm; | ||||||
|       Tnm    =Tn; |       Tnm    =Tn; | ||||||
|   | |||||||
| @@ -293,7 +293,7 @@ static void sncndnFK(INTERNAL_PRECISION u, INTERNAL_PRECISION k, | |||||||
|  * Set type = 0 for the Zolotarev approximation, which is zero at x = 0, and |  * Set type = 0 for the Zolotarev approximation, which is zero at x = 0, and | ||||||
|  * type = 1 for the approximation which is infinite at x = 0. */ |  * type = 1 for the approximation which is infinite at x = 0. */ | ||||||
|  |  | ||||||
| zolotarev_data* zolotarev(ZOLO_PRECISION epsilon, int n, int type) { | zolotarev_data* zolotarev(PRECISION epsilon, int n, int type) { | ||||||
|   INTERNAL_PRECISION A, c, cp, kp, ksq, sn, cn, dn, Kp, Kj, z, z0, t, M, F, |   INTERNAL_PRECISION A, c, cp, kp, ksq, sn, cn, dn, Kp, Kj, z, z0, t, M, F, | ||||||
|     l, invlambda, xi, xisq, *tv, s, opl; |     l, invlambda, xi, xisq, *tv, s, opl; | ||||||
|   int m, czero, ts; |   int m, czero, ts; | ||||||
| @@ -375,12 +375,12 @@ zolotarev_data* zolotarev(ZOLO_PRECISION epsilon, int n, int type) { | |||||||
|   construct_partfrac(d); |   construct_partfrac(d); | ||||||
|   construct_contfrac(d); |   construct_contfrac(d); | ||||||
|  |  | ||||||
|   /* Converting everything to ZOLO_PRECISION for external use only */ |   /* Converting everything to PRECISION for external use only */ | ||||||
|  |  | ||||||
|   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data)); |   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data)); | ||||||
|   zd -> A = (ZOLO_PRECISION) d -> A; |   zd -> A = (PRECISION) d -> A; | ||||||
|   zd -> Delta = (ZOLO_PRECISION) d -> Delta; |   zd -> Delta = (PRECISION) d -> Delta; | ||||||
|   zd -> epsilon = (ZOLO_PRECISION) d -> epsilon; |   zd -> epsilon = (PRECISION) d -> epsilon; | ||||||
|   zd -> n = d -> n; |   zd -> n = d -> n; | ||||||
|   zd -> type = d -> type; |   zd -> type = d -> type; | ||||||
|   zd -> dn = d -> dn; |   zd -> dn = d -> dn; | ||||||
| @@ -390,24 +390,24 @@ zolotarev_data* zolotarev(ZOLO_PRECISION epsilon, int n, int type) { | |||||||
|   zd -> deg_num = d -> deg_num; |   zd -> deg_num = d -> deg_num; | ||||||
|   zd -> deg_denom = d -> deg_denom; |   zd -> deg_denom = d -> deg_denom; | ||||||
|  |  | ||||||
|   zd -> a = (ZOLO_PRECISION*) malloc(zd -> dn * sizeof(ZOLO_PRECISION)); |   zd -> a = (PRECISION*) malloc(zd -> dn * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> dn; m++) zd -> a[m] = (ZOLO_PRECISION) d -> a[m]; |   for (m = 0; m < zd -> dn; m++) zd -> a[m] = (PRECISION) d -> a[m]; | ||||||
|   free(d -> a); |   free(d -> a); | ||||||
|  |  | ||||||
|   zd -> ap = (ZOLO_PRECISION*) malloc(zd -> dd * sizeof(ZOLO_PRECISION)); |   zd -> ap = (PRECISION*) malloc(zd -> dd * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (ZOLO_PRECISION) d -> ap[m]; |   for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (PRECISION) d -> ap[m]; | ||||||
|   free(d -> ap); |   free(d -> ap); | ||||||
|  |  | ||||||
|   zd -> alpha = (ZOLO_PRECISION*) malloc(zd -> da * sizeof(ZOLO_PRECISION)); |   zd -> alpha = (PRECISION*) malloc(zd -> da * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (ZOLO_PRECISION) d -> alpha[m]; |   for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (PRECISION) d -> alpha[m]; | ||||||
|   free(d -> alpha); |   free(d -> alpha); | ||||||
|  |  | ||||||
|   zd -> beta = (ZOLO_PRECISION*) malloc(zd -> db * sizeof(ZOLO_PRECISION)); |   zd -> beta = (PRECISION*) malloc(zd -> db * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> db; m++) zd -> beta[m] = (ZOLO_PRECISION) d -> beta[m]; |   for (m = 0; m < zd -> db; m++) zd -> beta[m] = (PRECISION) d -> beta[m]; | ||||||
|   free(d -> beta); |   free(d -> beta); | ||||||
|  |  | ||||||
|   zd -> gamma = (ZOLO_PRECISION*) malloc(zd -> n * sizeof(ZOLO_PRECISION)); |   zd -> gamma = (PRECISION*) malloc(zd -> n * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (ZOLO_PRECISION) d -> gamma[m]; |   for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (PRECISION) d -> gamma[m]; | ||||||
|   free(d -> gamma); |   free(d -> gamma); | ||||||
|  |  | ||||||
|   free(d); |   free(d); | ||||||
| @@ -426,7 +426,7 @@ void zolotarev_free(zolotarev_data *zdata) | |||||||
| } | } | ||||||
|  |  | ||||||
|  |  | ||||||
| zolotarev_data* higham(ZOLO_PRECISION epsilon, int n) { | zolotarev_data* higham(PRECISION epsilon, int n) { | ||||||
|   INTERNAL_PRECISION A, M, c, cp, z, z0, t, epssq; |   INTERNAL_PRECISION A, M, c, cp, z, z0, t, epssq; | ||||||
|   int m, czero; |   int m, czero; | ||||||
|   zolotarev_data *zd; |   zolotarev_data *zd; | ||||||
| @@ -481,9 +481,9 @@ zolotarev_data* higham(ZOLO_PRECISION epsilon, int n) { | |||||||
|   /* Converting everything to PRECISION for external use only */ |   /* Converting everything to PRECISION for external use only */ | ||||||
|  |  | ||||||
|   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data)); |   zd = (zolotarev_data*) malloc(sizeof(zolotarev_data)); | ||||||
|   zd -> A = (ZOLO_PRECISION) d -> A; |   zd -> A = (PRECISION) d -> A; | ||||||
|   zd -> Delta = (ZOLO_PRECISION) d -> Delta; |   zd -> Delta = (PRECISION) d -> Delta; | ||||||
|   zd -> epsilon = (ZOLO_PRECISION) d -> epsilon; |   zd -> epsilon = (PRECISION) d -> epsilon; | ||||||
|   zd -> n = d -> n; |   zd -> n = d -> n; | ||||||
|   zd -> type = d -> type; |   zd -> type = d -> type; | ||||||
|   zd -> dn = d -> dn; |   zd -> dn = d -> dn; | ||||||
| @@ -493,24 +493,24 @@ zolotarev_data* higham(ZOLO_PRECISION epsilon, int n) { | |||||||
|   zd -> deg_num = d -> deg_num; |   zd -> deg_num = d -> deg_num; | ||||||
|   zd -> deg_denom = d -> deg_denom; |   zd -> deg_denom = d -> deg_denom; | ||||||
|  |  | ||||||
|   zd -> a = (ZOLO_PRECISION*) malloc(zd -> dn * sizeof(ZOLO_PRECISION)); |   zd -> a = (PRECISION*) malloc(zd -> dn * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> dn; m++) zd -> a[m] = (ZOLO_PRECISION) d -> a[m]; |   for (m = 0; m < zd -> dn; m++) zd -> a[m] = (PRECISION) d -> a[m]; | ||||||
|   free(d -> a); |   free(d -> a); | ||||||
|  |  | ||||||
|   zd -> ap = (ZOLO_PRECISION*) malloc(zd -> dd * sizeof(ZOLO_PRECISION)); |   zd -> ap = (PRECISION*) malloc(zd -> dd * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (ZOLO_PRECISION) d -> ap[m]; |   for (m = 0; m < zd -> dd; m++) zd -> ap[m] = (PRECISION) d -> ap[m]; | ||||||
|   free(d -> ap); |   free(d -> ap); | ||||||
|  |  | ||||||
|   zd -> alpha = (ZOLO_PRECISION*) malloc(zd -> da * sizeof(ZOLO_PRECISION)); |   zd -> alpha = (PRECISION*) malloc(zd -> da * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (ZOLO_PRECISION) d -> alpha[m]; |   for (m = 0; m < zd -> da; m++) zd -> alpha[m] = (PRECISION) d -> alpha[m]; | ||||||
|   free(d -> alpha); |   free(d -> alpha); | ||||||
|  |  | ||||||
|   zd -> beta = (ZOLO_PRECISION*) malloc(zd -> db * sizeof(ZOLO_PRECISION)); |   zd -> beta = (PRECISION*) malloc(zd -> db * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> db; m++) zd -> beta[m] = (ZOLO_PRECISION) d -> beta[m]; |   for (m = 0; m < zd -> db; m++) zd -> beta[m] = (PRECISION) d -> beta[m]; | ||||||
|   free(d -> beta); |   free(d -> beta); | ||||||
|  |  | ||||||
|   zd -> gamma = (ZOLO_PRECISION*) malloc(zd -> n * sizeof(ZOLO_PRECISION)); |   zd -> gamma = (PRECISION*) malloc(zd -> n * sizeof(PRECISION)); | ||||||
|   for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (ZOLO_PRECISION) d -> gamma[m]; |   for (m = 0; m < zd -> n; m++) zd -> gamma[m] = (PRECISION) d -> gamma[m]; | ||||||
|   free(d -> gamma); |   free(d -> gamma); | ||||||
|  |  | ||||||
|   free(d); |   free(d); | ||||||
| @@ -523,17 +523,17 @@ NAMESPACE_END(Grid); | |||||||
| #ifdef TEST | #ifdef TEST | ||||||
|  |  | ||||||
| #undef ZERO | #undef ZERO | ||||||
| #define ZERO ((ZOLO_PRECISION) 0) | #define ZERO ((PRECISION) 0) | ||||||
| #undef ONE | #undef ONE | ||||||
| #define ONE ((ZOLO_PRECISION) 1) | #define ONE ((PRECISION) 1) | ||||||
| #undef TWO | #undef TWO | ||||||
| #define TWO ((ZOLO_PRECISION) 2) | #define TWO ((PRECISION) 2) | ||||||
|  |  | ||||||
| /* Evaluate the rational approximation R(x) using the factored form */ | /* Evaluate the rational approximation R(x) using the factored form */ | ||||||
|  |  | ||||||
| static ZOLO_PRECISION zolotarev_eval(ZOLO_PRECISION x, zolotarev_data* rdata) { | static PRECISION zolotarev_eval(PRECISION x, zolotarev_data* rdata) { | ||||||
|   int m; |   int m; | ||||||
|   ZOLO_PRECISION R; |   PRECISION R; | ||||||
|  |  | ||||||
|   if (rdata -> type == 0) { |   if (rdata -> type == 0) { | ||||||
|     R = rdata -> A * x; |     R = rdata -> A * x; | ||||||
| @@ -551,9 +551,9 @@ static ZOLO_PRECISION zolotarev_eval(ZOLO_PRECISION x, zolotarev_data* rdata) { | |||||||
|  |  | ||||||
| /* Evaluate the rational approximation R(x) using the partial fraction form */ | /* Evaluate the rational approximation R(x) using the partial fraction form */ | ||||||
|  |  | ||||||
| static ZOLO_PRECISION zolotarev_partfrac_eval(ZOLO_PRECISION x, zolotarev_data* rdata) { | static PRECISION zolotarev_partfrac_eval(PRECISION x, zolotarev_data* rdata) { | ||||||
|   int m; |   int m; | ||||||
|   ZOLO_PRECISION R = rdata -> alpha[rdata -> da - 1]; |   PRECISION R = rdata -> alpha[rdata -> da - 1]; | ||||||
|   for (m = 0; m < rdata -> dd; m++) |   for (m = 0; m < rdata -> dd; m++) | ||||||
|     R += rdata -> alpha[m] / (x * x - rdata -> ap[m]); |     R += rdata -> alpha[m] / (x * x - rdata -> ap[m]); | ||||||
|   if (rdata -> type == 1) R += rdata -> alpha[rdata -> dd] / (x * x); |   if (rdata -> type == 1) R += rdata -> alpha[rdata -> dd] / (x * x); | ||||||
| @@ -568,18 +568,18 @@ static ZOLO_PRECISION zolotarev_partfrac_eval(ZOLO_PRECISION x, zolotarev_data* | |||||||
|  * non-signalling overflow this will work correctly since 1/(1/0) = 1/INF = 0, |  * non-signalling overflow this will work correctly since 1/(1/0) = 1/INF = 0, | ||||||
|  * but with signalling overflow you will get an error message. */ |  * but with signalling overflow you will get an error message. */ | ||||||
|  |  | ||||||
| static ZOLO_PRECISION zolotarev_contfrac_eval(ZOLO_PRECISION x, zolotarev_data* rdata) { | static PRECISION zolotarev_contfrac_eval(PRECISION x, zolotarev_data* rdata) { | ||||||
|   int m; |   int m; | ||||||
|   ZOLO_PRECISION R = rdata -> beta[0] * x; |   PRECISION R = rdata -> beta[0] * x; | ||||||
|   for (m = 1; m < rdata -> db; m++) R = rdata -> beta[m] * x + ONE / R; |   for (m = 1; m < rdata -> db; m++) R = rdata -> beta[m] * x + ONE / R; | ||||||
|   return R; |   return R; | ||||||
| }     | }     | ||||||
|  |  | ||||||
| /* Evaluate the rational approximation R(x) using Cayley form */ | /* Evaluate the rational approximation R(x) using Cayley form */ | ||||||
|  |  | ||||||
| static ZOLO_PRECISION zolotarev_cayley_eval(ZOLO_PRECISION x, zolotarev_data* rdata) { | static PRECISION zolotarev_cayley_eval(PRECISION x, zolotarev_data* rdata) { | ||||||
|   int m; |   int m; | ||||||
|   ZOLO_PRECISION T; |   PRECISION T; | ||||||
|  |  | ||||||
|   T = rdata -> type == 0 ? ONE : -ONE; |   T = rdata -> type == 0 ? ONE : -ONE; | ||||||
|   for (m = 0; m < rdata -> n; m++) |   for (m = 0; m < rdata -> n; m++) | ||||||
| @@ -607,7 +607,7 @@ int main(int argc, char** argv) { | |||||||
|   int m, n, plotpts = 5000, type = 0; |   int m, n, plotpts = 5000, type = 0; | ||||||
|   float eps, x, ypferr, ycferr, ycaylerr, maxypferr, maxycferr, maxycaylerr; |   float eps, x, ypferr, ycferr, ycaylerr, maxypferr, maxycferr, maxycaylerr; | ||||||
|   zolotarev_data *rdata; |   zolotarev_data *rdata; | ||||||
|   ZOLO_PRECISION y; |   PRECISION y; | ||||||
|   FILE *plot_function, *plot_error,  |   FILE *plot_function, *plot_error,  | ||||||
|     *plot_partfrac, *plot_contfrac, *plot_cayley; |     *plot_partfrac, *plot_contfrac, *plot_cayley; | ||||||
|  |  | ||||||
| @@ -626,13 +626,13 @@ int main(int argc, char** argv) { | |||||||
|   } |   } | ||||||
|  |  | ||||||
|   rdata = type == 2  |   rdata = type == 2  | ||||||
|     ? higham((ZOLO_PRECISION) eps, n)  |     ? higham((PRECISION) eps, n)  | ||||||
|     : zolotarev((ZOLO_PRECISION) eps, n, type); |     : zolotarev((PRECISION) eps, n, type); | ||||||
|  |  | ||||||
|   printf("Zolotarev Test: R(epsilon = %g, n = %d, type = %d)\n\t"  |   printf("Zolotarev Test: R(epsilon = %g, n = %d, type = %d)\n\t"  | ||||||
| 	 STRINGIFY(VERSION) "\n\t" STRINGIFY(HVERSION) | 	 STRINGIFY(VERSION) "\n\t" STRINGIFY(HVERSION) | ||||||
| 	 "\n\tINTERNAL_PRECISION = " STRINGIFY(INTERNAL_PRECISION) | 	 "\n\tINTERNAL_PRECISION = " STRINGIFY(INTERNAL_PRECISION) | ||||||
| 	 "\tZOLO_PRECISION = " STRINGIFY(ZOLO_PRECISION) | 	 "\tPRECISION = " STRINGIFY(PRECISION) | ||||||
| 	 "\n\n\tRational approximation of degree (%d,%d), %s at x = 0\n" | 	 "\n\n\tRational approximation of degree (%d,%d), %s at x = 0\n" | ||||||
| 	 "\tDelta = %g (maximum error)\n\n" | 	 "\tDelta = %g (maximum error)\n\n" | ||||||
| 	 "\tA = %g (overall factor)\n", | 	 "\tA = %g (overall factor)\n", | ||||||
| @@ -681,15 +681,15 @@ int main(int argc, char** argv) { | |||||||
|     x = 2.4 * (float) m / plotpts - 1.2; |     x = 2.4 * (float) m / plotpts - 1.2; | ||||||
|     if (rdata -> type == 0 || fabs(x) * (float) plotpts > 1.0) { |     if (rdata -> type == 0 || fabs(x) * (float) plotpts > 1.0) { | ||||||
|       /* skip x = 0 for type 1, as R(0) is singular */ |       /* skip x = 0 for type 1, as R(0) is singular */ | ||||||
|       y = zolotarev_eval((ZOLO_PRECISION) x, rdata); |       y = zolotarev_eval((PRECISION) x, rdata); | ||||||
|       fprintf(plot_function, "%g %g\n", x, (float) y); |       fprintf(plot_function, "%g %g\n", x, (float) y); | ||||||
|       fprintf(plot_error, "%g %g\n", |       fprintf(plot_error, "%g %g\n", | ||||||
| 	      x, (float)((y - ((x > 0.0 ? ONE : -ONE))) / rdata -> Delta)); | 	      x, (float)((y - ((x > 0.0 ? ONE : -ONE))) / rdata -> Delta)); | ||||||
|       ypferr = (float)((zolotarev_partfrac_eval((ZOLO_PRECISION) x, rdata) - y) |       ypferr = (float)((zolotarev_partfrac_eval((PRECISION) x, rdata) - y) | ||||||
| 		       / rdata -> Delta); | 		       / rdata -> Delta); | ||||||
|       ycferr = (float)((zolotarev_contfrac_eval((ZOLO_PRECISION) x, rdata) - y) |       ycferr = (float)((zolotarev_contfrac_eval((PRECISION) x, rdata) - y) | ||||||
| 		       / rdata -> Delta); | 		       / rdata -> Delta); | ||||||
|       ycaylerr = (float)((zolotarev_cayley_eval((ZOLO_PRECISION) x, rdata) - y) |       ycaylerr = (float)((zolotarev_cayley_eval((PRECISION) x, rdata) - y) | ||||||
| 		       / rdata -> Delta); | 		       / rdata -> Delta); | ||||||
|       if (fabs(x) < 1.0 && fabs(x) > rdata -> epsilon) { |       if (fabs(x) < 1.0 && fabs(x) > rdata -> epsilon) { | ||||||
| 	maxypferr = MAX(maxypferr, fabs(ypferr)); | 	maxypferr = MAX(maxypferr, fabs(ypferr)); | ||||||
|   | |||||||
| @@ -9,10 +9,10 @@ NAMESPACE_BEGIN(Approx); | |||||||
| #define HVERSION Header Time-stamp: <14-OCT-2004 09:26:51.00 adk@MISSCONTRARY> | #define HVERSION Header Time-stamp: <14-OCT-2004 09:26:51.00 adk@MISSCONTRARY> | ||||||
|  |  | ||||||
| #ifndef ZOLOTAREV_INTERNAL | #ifndef ZOLOTAREV_INTERNAL | ||||||
| #ifndef ZOLO_PRECISION | #ifndef PRECISION | ||||||
| #define ZOLO_PRECISION double | #define PRECISION double | ||||||
| #endif | #endif | ||||||
| #define ZPRECISION ZOLO_PRECISION | #define ZPRECISION PRECISION | ||||||
| #define ZOLOTAREV_DATA zolotarev_data | #define ZOLOTAREV_DATA zolotarev_data | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
| @@ -77,8 +77,8 @@ typedef struct { | |||||||
|  * zolotarev_data structure. The arguments must satisfy the constraints that |  * zolotarev_data structure. The arguments must satisfy the constraints that | ||||||
|  * epsilon > 0, n > 0, and type = 0 or 1. */ |  * epsilon > 0, n > 0, and type = 0 or 1. */ | ||||||
|  |  | ||||||
| ZOLOTAREV_DATA* higham(ZOLO_PRECISION epsilon, int n) ; | ZOLOTAREV_DATA* higham(PRECISION epsilon, int n) ; | ||||||
| ZOLOTAREV_DATA* zolotarev(ZOLO_PRECISION epsilon, int n, int type); | ZOLOTAREV_DATA* zolotarev(PRECISION epsilon, int n, int type); | ||||||
| void zolotarev_free(zolotarev_data *zdata); | void zolotarev_free(zolotarev_data *zdata); | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
| @@ -86,4 +86,3 @@ void zolotarev_free(zolotarev_data *zdata); | |||||||
| NAMESPACE_END(Approx); | NAMESPACE_END(Approx); | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
|   | |||||||
| @@ -1,34 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid  |  | ||||||
|  |  | ||||||
|     Source file: BatchedBlas.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2023 |  | ||||||
|  |  | ||||||
| Author: Peter Boyle <pboyle@bnl.gov> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
| *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
| #include <Grid/GridCore.h> |  | ||||||
| #include <Grid/algorithms/blas/BatchedBlas.h> |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
| gridblasHandle_t GridBLAS::gridblasHandle; |  | ||||||
| int              GridBLAS::gridblasInit; |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
|  |  | ||||||
| @@ -1,727 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid  |  | ||||||
|  |  | ||||||
|     Source file: BatchedBlas.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2023 |  | ||||||
|  |  | ||||||
| Author: Peter Boyle <pboyle@bnl.gov> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
| *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
| #pragma once |  | ||||||
|  |  | ||||||
| #ifdef GRID_HIP |  | ||||||
| #include <hipblas/hipblas.h> |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
| #include <cublas_v2.h> |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
| #include <oneapi/mkl.hpp> |  | ||||||
| #endif |  | ||||||
| #if 0 |  | ||||||
| #define GRID_ONE_MKL |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_ONE_MKL |  | ||||||
| #include <oneapi/mkl.hpp> |  | ||||||
| #endif |  | ||||||
| ///////////////////////////////////////////////////////////////////////	   |  | ||||||
| // Need to rearrange lattice data to be in the right format for a |  | ||||||
| // batched multiply. Might as well make these static, dense packed |  | ||||||
| /////////////////////////////////////////////////////////////////////// |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
| #ifdef GRID_HIP |  | ||||||
|   typedef hipblasHandle_t gridblasHandle_t; |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|   typedef cublasHandle_t gridblasHandle_t; |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
|   typedef cl::sycl::queue *gridblasHandle_t; |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_ONE_MKL |  | ||||||
|   typedef cl::sycl::queue *gridblasHandle_t; |  | ||||||
| #endif |  | ||||||
| #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL) |  | ||||||
|   typedef int32_t gridblasHandle_t; |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
| enum GridBLASOperation_t { GridBLAS_OP_N, GridBLAS_OP_T, GridBLAS_OP_C } ; |  | ||||||
|  |  | ||||||
| class GridBLAS { |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|    |  | ||||||
|   static gridblasHandle_t gridblasHandle; |  | ||||||
|   static int            gridblasInit; |  | ||||||
|    |  | ||||||
|   static void Init(void) |  | ||||||
|   { |  | ||||||
|     if ( ! gridblasInit ) { |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|       std::cout << "cublasCreate"<<std::endl; |  | ||||||
|       cublasCreate(&gridblasHandle); |  | ||||||
|       cublasSetPointerMode(gridblasHandle, CUBLAS_POINTER_MODE_DEVICE); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_HIP |  | ||||||
|       std::cout << "hipblasCreate"<<std::endl; |  | ||||||
|       hipblasCreate(&gridblasHandle); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
|       gridblasHandle = theGridAccelerator; |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_ONE_MKL |  | ||||||
|       cl::sycl::cpu_selector selector; |  | ||||||
|       cl::sycl::device selectedDevice { selector }; |  | ||||||
|       gridblasHandle =new sycl::queue (selectedDevice); |  | ||||||
| #endif |  | ||||||
|       gridblasInit=1; |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|   // Force construct once |  | ||||||
|   GridBLAS() { Init(); }; |  | ||||||
|   ~GridBLAS() { }; |  | ||||||
|    |  | ||||||
|   ///////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|   // BLAS GEMM conventions: |  | ||||||
|   ///////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|   // - C = alpha A * B + beta C |  | ||||||
|   // Dimensions: |  | ||||||
|   // - C_m.n |  | ||||||
|   // - A_m.k |  | ||||||
|   // - B_k.n |  | ||||||
|   // - Flops = 8 M N K |  | ||||||
|   // - Bytes = 2*sizeof(word) * (MN+MK+KN) |  | ||||||
|   // M=60, N=12 |  | ||||||
|   // Flop/Byte = 8 . 60.60.12 / (60.12+60.60+60.12)/16 = 4 so expect about 4 TF/s on a GCD |  | ||||||
|   ///////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|   void synchronise(void) |  | ||||||
|   { |  | ||||||
| #ifdef GRID_HIP |  | ||||||
|     auto err = hipDeviceSynchronize(); |  | ||||||
|     assert(err==hipSuccess); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|     auto err = cudaDeviceSynchronize(); |  | ||||||
|     assert(err==cudaSuccess); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
|     accelerator_barrier(); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_ONE_MKL |  | ||||||
|     gridblasHandle->wait(); |  | ||||||
| #endif |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|   void gemmBatched(int m,int n, int k, |  | ||||||
| 		   ComplexD alpha, |  | ||||||
| 		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices |  | ||||||
| 		   deviceVector<ComplexD*> &Bkn, |  | ||||||
| 		   ComplexD beta, |  | ||||||
| 		   deviceVector<ComplexD*> &Cmn) |  | ||||||
|   { |  | ||||||
|     gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, |  | ||||||
| 		m,n,k, |  | ||||||
| 		alpha, |  | ||||||
| 		Amk, |  | ||||||
| 		Bkn, |  | ||||||
| 		beta, |  | ||||||
| 		Cmn); |  | ||||||
|   } |  | ||||||
|   void gemmBatched(int m,int n, int k, |  | ||||||
| 		   ComplexF alpha, |  | ||||||
| 		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices |  | ||||||
| 		   deviceVector<ComplexF*> &Bkn, |  | ||||||
| 		   ComplexF beta, |  | ||||||
| 		   deviceVector<ComplexF*> &Cmn) |  | ||||||
|   { |  | ||||||
|     gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, |  | ||||||
| 		m,n,k, |  | ||||||
| 		alpha, |  | ||||||
| 		Amk, |  | ||||||
| 		Bkn, |  | ||||||
| 		beta, |  | ||||||
| 		Cmn); |  | ||||||
|   } |  | ||||||
|   void gemmBatched(int m,int n, int k, |  | ||||||
| 		   RealD alpha, |  | ||||||
| 		   deviceVector<RealD*> &Amk,  // pointer list to matrices |  | ||||||
| 		   deviceVector<RealD*> &Bkn, |  | ||||||
| 		   RealD beta, |  | ||||||
| 		   deviceVector<RealD*> &Cmn) |  | ||||||
|   { |  | ||||||
|     gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, |  | ||||||
| 		m,n,k, |  | ||||||
| 		alpha, |  | ||||||
| 		Amk, |  | ||||||
| 		Bkn, |  | ||||||
| 		beta, |  | ||||||
| 		Cmn); |  | ||||||
|   } |  | ||||||
|   void gemmBatched(int m,int n, int k, |  | ||||||
| 		   RealF alpha, |  | ||||||
| 		   deviceVector<RealF*> &Amk,  // pointer list to matrices |  | ||||||
| 		   deviceVector<RealF*> &Bkn, |  | ||||||
| 		   RealF beta, |  | ||||||
| 		   deviceVector<RealF*> &Cmn) |  | ||||||
|   { |  | ||||||
|     gemmBatched(GridBLAS_OP_N,GridBLAS_OP_N, |  | ||||||
| 		m,n,k, |  | ||||||
| 		alpha, |  | ||||||
| 		Amk, |  | ||||||
| 		Bkn, |  | ||||||
| 		beta, |  | ||||||
| 		Cmn); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void gemmBatched(GridBLASOperation_t OpA, |  | ||||||
| 		   GridBLASOperation_t OpB, |  | ||||||
| 		   int m,int n, int k, |  | ||||||
| 		   ComplexD alpha, |  | ||||||
| 		   deviceVector<ComplexD*> &Amk,  // pointer list to matrices |  | ||||||
| 		   deviceVector<ComplexD*> &Bkn, |  | ||||||
| 		   ComplexD beta, |  | ||||||
| 		   deviceVector<ComplexD*> &Cmn) |  | ||||||
|   { |  | ||||||
|     RealD t2=usecond(); |  | ||||||
|     int32_t batchCount = Amk.size(); |  | ||||||
|     assert(Bkn.size()==batchCount); |  | ||||||
|     assert(Cmn.size()==batchCount); |  | ||||||
|  |  | ||||||
|     int lda = m; // m x k column major |  | ||||||
|     int ldb = k; // k x n column major |  | ||||||
|     int ldc = m; // m x b column major |  | ||||||
|     if(OpA!=GridBLAS_OP_N) |  | ||||||
|       lda = k; |  | ||||||
|     if(OpB!=GridBLAS_OP_N) |  | ||||||
|       ldb = n; |  | ||||||
|      |  | ||||||
|     static deviceVector<ComplexD> alpha_p(1); |  | ||||||
|     static deviceVector<ComplexD> beta_p(1); |  | ||||||
|     // can prestore the 1 and the zero on device |  | ||||||
|     acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD)); |  | ||||||
|     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD)); |  | ||||||
|     RealD t0=usecond(); |  | ||||||
|     //    std::cout << "ZgemmBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl; |  | ||||||
| #ifdef GRID_HIP |  | ||||||
|     hipblasOperation_t hOpA; |  | ||||||
|     hipblasOperation_t hOpB; |  | ||||||
|     if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N; |  | ||||||
|     if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T; |  | ||||||
|     if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C; |  | ||||||
|     if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N; |  | ||||||
|     if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T; |  | ||||||
|     if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C; |  | ||||||
|     auto err = hipblasZgemmBatched(gridblasHandle, |  | ||||||
| 				   hOpA, |  | ||||||
| 				   hOpB, |  | ||||||
| 				   m,n,k, |  | ||||||
| 				   (hipblasDoubleComplex *) &alpha_p[0], |  | ||||||
| 				   (hipblasDoubleComplex **)&Amk[0], lda, |  | ||||||
| 				   (hipblasDoubleComplex **)&Bkn[0], ldb, |  | ||||||
| 				   (hipblasDoubleComplex *) &beta_p[0], |  | ||||||
| 				   (hipblasDoubleComplex **)&Cmn[0], ldc, |  | ||||||
| 				   batchCount); |  | ||||||
|     //	 std::cout << " hipblas return code " <<(int)err<<std::endl; |  | ||||||
|     assert(err==HIPBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|     cublasOperation_t hOpA; |  | ||||||
|     cublasOperation_t hOpB; |  | ||||||
|     if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N; |  | ||||||
|     if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T; |  | ||||||
|     if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C; |  | ||||||
|     if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N; |  | ||||||
|     if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T; |  | ||||||
|     if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C; |  | ||||||
|     auto err = cublasZgemmBatched(gridblasHandle, |  | ||||||
| 				  hOpA, |  | ||||||
| 				  hOpB, |  | ||||||
| 				  m,n,k, |  | ||||||
| 				  (cuDoubleComplex *) &alpha_p[0], |  | ||||||
| 				  (cuDoubleComplex **)&Amk[0], lda, |  | ||||||
| 				  (cuDoubleComplex **)&Bkn[0], ldb, |  | ||||||
| 				  (cuDoubleComplex *) &beta_p[0], |  | ||||||
| 				  (cuDoubleComplex **)&Cmn[0], ldc, |  | ||||||
| 				  batchCount); |  | ||||||
|     assert(err==CUBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
|     //MKL’s cblas_<T>gemm_batch & OneAPI |  | ||||||
| #warning "oneMKL implementation not built " |  | ||||||
| #endif |  | ||||||
| #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) |  | ||||||
|     // Need a default/reference implementation |  | ||||||
|     int sda = lda*k; |  | ||||||
|     int sdb = ldb*k; |  | ||||||
|     int sdc = ldc*n; |  | ||||||
|     for (int p = 0; p < batchCount; ++p) { |  | ||||||
|       for (int mm = 0; mm < m; ++mm) { |  | ||||||
| 	for (int nn = 0; nn < n; ++nn) { |  | ||||||
| 	  ComplexD c_mn(0.0); |  | ||||||
| 	  for (int kk = 0; kk < k; ++kk) |  | ||||||
| 	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb]; |  | ||||||
| 	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ]; |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
| #endif |  | ||||||
|     //    synchronise(); |  | ||||||
|      RealD t1=usecond(); |  | ||||||
|      RealD flops = 8.0*m*n*k*batchCount; |  | ||||||
|      RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n)*batchCount; |  | ||||||
|      //     std::cout <<GridLogMessage<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl; |  | ||||||
|      //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; |  | ||||||
|      //     std::cout <<GridLogMessage<< " batched Blas zGemm call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void gemmBatched(GridBLASOperation_t OpA, |  | ||||||
| 		   GridBLASOperation_t OpB, |  | ||||||
| 		   int m,int n, int k, |  | ||||||
| 		   ComplexF alpha, |  | ||||||
| 		   deviceVector<ComplexF*> &Amk,  // pointer list to matrices |  | ||||||
| 		   deviceVector<ComplexF*> &Bkn, |  | ||||||
| 		   ComplexF beta, |  | ||||||
| 		   deviceVector<ComplexF*> &Cmn) |  | ||||||
|   { |  | ||||||
|     RealD t2=usecond(); |  | ||||||
|     int32_t batchCount = Amk.size(); |  | ||||||
|  |  | ||||||
|     int lda = m; // m x k column major |  | ||||||
|     int ldb = k; // k x n column major |  | ||||||
|     int ldc = m; // m x b column major |  | ||||||
|     if(OpA!=GridBLAS_OP_N) |  | ||||||
|       lda = k; |  | ||||||
|     if(OpB!=GridBLAS_OP_N) |  | ||||||
|       ldb = n; |  | ||||||
|     static deviceVector<ComplexF> alpha_p(1); |  | ||||||
|     static deviceVector<ComplexF> beta_p(1); |  | ||||||
|     // can prestore the 1 and the zero on device |  | ||||||
|     acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexF)); |  | ||||||
|     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexF)); |  | ||||||
|     RealD t0=usecond(); |  | ||||||
|  |  | ||||||
|     assert(Bkn.size()==batchCount); |  | ||||||
|     assert(Cmn.size()==batchCount); |  | ||||||
| #ifdef GRID_HIP |  | ||||||
|     hipblasOperation_t hOpA; |  | ||||||
|     hipblasOperation_t hOpB; |  | ||||||
|     if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N; |  | ||||||
|     if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T; |  | ||||||
|     if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C; |  | ||||||
|     if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N; |  | ||||||
|     if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T; |  | ||||||
|     if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C; |  | ||||||
|     auto err = hipblasCgemmBatched(gridblasHandle, |  | ||||||
| 				   hOpA, |  | ||||||
| 				   hOpB, |  | ||||||
| 				   m,n,k, |  | ||||||
| 				   (hipblasComplex *) &alpha_p[0], |  | ||||||
| 				   (hipblasComplex **)&Amk[0], lda, |  | ||||||
| 				   (hipblasComplex **)&Bkn[0], ldb, |  | ||||||
| 				   (hipblasComplex *) &beta_p[0], |  | ||||||
| 				   (hipblasComplex **)&Cmn[0], ldc, |  | ||||||
| 				   batchCount); |  | ||||||
|  |  | ||||||
|     assert(err==HIPBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|     cublasOperation_t hOpA; |  | ||||||
|     cublasOperation_t hOpB; |  | ||||||
|     if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N; |  | ||||||
|     if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T; |  | ||||||
|     if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C; |  | ||||||
|     if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N; |  | ||||||
|     if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T; |  | ||||||
|     if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C; |  | ||||||
|     auto err = cublasCgemmBatched(gridblasHandle, |  | ||||||
| 				  hOpA, |  | ||||||
| 				  hOpB, |  | ||||||
| 				  m,n,k, |  | ||||||
| 				  (cuComplex *) &alpha_p[0], |  | ||||||
| 				  (cuComplex **)&Amk[0], lda, |  | ||||||
| 				  (cuComplex **)&Bkn[0], ldb, |  | ||||||
| 				  (cuComplex *) &beta_p[0], |  | ||||||
| 				  (cuComplex **)&Cmn[0], ldc, |  | ||||||
| 				  batchCount); |  | ||||||
|     assert(err==CUBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
|     //MKL’s cblas_<T>gemm_batch & OneAPI |  | ||||||
| #warning "oneMKL implementation not built " |  | ||||||
| #endif |  | ||||||
| #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) |  | ||||||
|     int sda = lda*k; |  | ||||||
|     int sdb = ldb*k; |  | ||||||
|     int sdc = ldc*n; |  | ||||||
|     ComplexF alphaf(real(alpha),imag(alpha)); |  | ||||||
|     ComplexF betaf(real(beta),imag(beta)); |  | ||||||
|     // Need a default/reference implementation |  | ||||||
|     for (int p = 0; p < batchCount; ++p) { |  | ||||||
|       for (int mm = 0; mm < m; ++mm) { |  | ||||||
| 	for (int nn = 0; nn < n; ++nn) { |  | ||||||
| 	  ComplexF c_mn(0.0); |  | ||||||
| 	  for (int kk = 0; kk < k; ++kk) |  | ||||||
| 	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb]; |  | ||||||
| 	  Cmn[p][mm + nn*ldc] =  (alphaf)*c_mn + (betaf)*Cmn[p][mm + nn*ldc ]; |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
| #endif |  | ||||||
|      RealD t1=usecond(); |  | ||||||
|      RealD flops = 8.0*m*n*k*batchCount; |  | ||||||
|      RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount; |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|   /////////////////////////////////////////////////////////////////////////// |  | ||||||
|   // Single precision real GEMM |  | ||||||
|   /////////////////////////////////////////////////////////////////////////// |  | ||||||
|  |  | ||||||
|   void gemmBatched(GridBLASOperation_t OpA, |  | ||||||
| 		   GridBLASOperation_t OpB, |  | ||||||
| 		   int m,int n, int k, |  | ||||||
| 		   RealF alpha, |  | ||||||
| 		   deviceVector<RealF*> &Amk,  // pointer list to matrices |  | ||||||
| 		   deviceVector<RealF*> &Bkn, |  | ||||||
| 		   RealF beta, |  | ||||||
| 		   deviceVector<RealF*> &Cmn) |  | ||||||
|   { |  | ||||||
|     RealD t2=usecond(); |  | ||||||
|     int32_t batchCount = Amk.size(); |  | ||||||
|  |  | ||||||
|     int lda = m; // m x k column major |  | ||||||
|     int ldb = k; // k x n column major |  | ||||||
|     int ldc = m; // m x b column major |  | ||||||
|     if(OpA!=GridBLAS_OP_N) |  | ||||||
|       lda = k; |  | ||||||
|     if(OpB!=GridBLAS_OP_N) |  | ||||||
|       ldb = n; |  | ||||||
|     static deviceVector<RealF> alpha_p(1); |  | ||||||
|     static deviceVector<RealF> beta_p(1); |  | ||||||
|     // can prestore the 1 and the zero on device |  | ||||||
|     acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealF)); |  | ||||||
|     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealF)); |  | ||||||
|     RealD t0=usecond(); |  | ||||||
|  |  | ||||||
|     assert(Bkn.size()==batchCount); |  | ||||||
|     assert(Cmn.size()==batchCount); |  | ||||||
| #ifdef GRID_HIP |  | ||||||
|     hipblasOperation_t hOpA; |  | ||||||
|     hipblasOperation_t hOpB; |  | ||||||
|     if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N; |  | ||||||
|     if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T; |  | ||||||
|     if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C; |  | ||||||
|     if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N; |  | ||||||
|     if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T; |  | ||||||
|     if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C; |  | ||||||
|     auto err = hipblasSgemmBatched(gridblasHandle, |  | ||||||
| 				   hOpA, |  | ||||||
| 				   hOpB, |  | ||||||
| 				   m,n,k, |  | ||||||
| 				   (float *) &alpha_p[0], |  | ||||||
| 				   (float **)&Amk[0], lda, |  | ||||||
| 				   (float **)&Bkn[0], ldb, |  | ||||||
| 				   (float *) &beta_p[0], |  | ||||||
| 				   (float **)&Cmn[0], ldc, |  | ||||||
| 				   batchCount); |  | ||||||
|     assert(err==HIPBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|     cublasOperation_t hOpA; |  | ||||||
|     cublasOperation_t hOpB; |  | ||||||
|     if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N; |  | ||||||
|     if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T; |  | ||||||
|     if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C; |  | ||||||
|     if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N; |  | ||||||
|     if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T; |  | ||||||
|     if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C; |  | ||||||
|     auto err = cublasSgemmBatched(gridblasHandle, |  | ||||||
| 				  hOpA, |  | ||||||
| 				  hOpB, |  | ||||||
| 				  m,n,k, |  | ||||||
| 				  (float *) &alpha_p[0], |  | ||||||
| 				  (float **)&Amk[0], lda, |  | ||||||
| 				  (float **)&Bkn[0], ldb, |  | ||||||
| 				  (float *) &beta_p[0], |  | ||||||
| 				  (float **)&Cmn[0], ldc, |  | ||||||
| 				  batchCount); |  | ||||||
|     assert(err==CUBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
|     //MKL’s cblas_<T>gemm_batch & OneAPI |  | ||||||
| #warning "oneMKL implementation not built " |  | ||||||
| #endif |  | ||||||
| #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) |  | ||||||
|     int sda = lda*k; |  | ||||||
|     int sdb = ldb*k; |  | ||||||
|     int sdc = ldc*n; |  | ||||||
|     // Need a default/reference implementation |  | ||||||
|     for (int p = 0; p < batchCount; ++p) { |  | ||||||
|       for (int mm = 0; mm < m; ++mm) { |  | ||||||
| 	for (int nn = 0; nn < n; ++nn) { |  | ||||||
| 	  RealD c_mn(0.0); |  | ||||||
| 	  for (int kk = 0; kk < k; ++kk) |  | ||||||
| 	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb]; |  | ||||||
| 	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ]; |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
| #endif |  | ||||||
|      RealD t1=usecond(); |  | ||||||
|      RealD flops = 2.0*m*n*k*batchCount; |  | ||||||
|      RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount; |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|    |  | ||||||
|   /////////////////////////////////////////////////////////////////////////// |  | ||||||
|   // Double precision real GEMM |  | ||||||
|   /////////////////////////////////////////////////////////////////////////// |  | ||||||
|  |  | ||||||
|   void gemmBatched(GridBLASOperation_t OpA, |  | ||||||
| 		   GridBLASOperation_t OpB, |  | ||||||
| 		   int m,int n, int k, |  | ||||||
| 		   RealD alpha, |  | ||||||
| 		   deviceVector<RealD*> &Amk,  // pointer list to matrices |  | ||||||
| 		   deviceVector<RealD*> &Bkn, |  | ||||||
| 		   RealD beta, |  | ||||||
| 		   deviceVector<RealD*> &Cmn) |  | ||||||
|   { |  | ||||||
|     RealD t2=usecond(); |  | ||||||
|     int32_t batchCount = Amk.size(); |  | ||||||
|  |  | ||||||
|     int lda = m; // m x k column major |  | ||||||
|     int ldb = k; // k x n column major |  | ||||||
|     int ldc = m; // m x b column major |  | ||||||
|     if(OpA!=GridBLAS_OP_N) |  | ||||||
|       lda = k; |  | ||||||
|     if(OpB!=GridBLAS_OP_N) |  | ||||||
|       ldb = n; |  | ||||||
|      |  | ||||||
|     static deviceVector<RealD> alpha_p(1); |  | ||||||
|     static deviceVector<RealD> beta_p(1); |  | ||||||
|     // can prestore the 1 and the zero on device |  | ||||||
|     acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(RealD)); |  | ||||||
|     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(RealD)); |  | ||||||
|     RealD t0=usecond(); |  | ||||||
|  |  | ||||||
|     assert(Bkn.size()==batchCount); |  | ||||||
|     assert(Cmn.size()==batchCount); |  | ||||||
| #ifdef GRID_HIP |  | ||||||
|     hipblasOperation_t hOpA; |  | ||||||
|     hipblasOperation_t hOpB; |  | ||||||
|     if ( OpA == GridBLAS_OP_N ) hOpA = HIPBLAS_OP_N; |  | ||||||
|     if ( OpA == GridBLAS_OP_T ) hOpA = HIPBLAS_OP_T; |  | ||||||
|     if ( OpA == GridBLAS_OP_C ) hOpA = HIPBLAS_OP_C; |  | ||||||
|     if ( OpB == GridBLAS_OP_N ) hOpB = HIPBLAS_OP_N; |  | ||||||
|     if ( OpB == GridBLAS_OP_T ) hOpB = HIPBLAS_OP_T; |  | ||||||
|     if ( OpB == GridBLAS_OP_C ) hOpB = HIPBLAS_OP_C; |  | ||||||
|     auto err = hipblasDgemmBatched(gridblasHandle, |  | ||||||
| 				   HIPBLAS_OP_N, |  | ||||||
| 				   HIPBLAS_OP_N, |  | ||||||
| 				   m,n,k, |  | ||||||
| 				   (double *) &alpha_p[0], |  | ||||||
| 				   (double **)&Amk[0], lda, |  | ||||||
| 				   (double **)&Bkn[0], ldb, |  | ||||||
| 				   (double *) &beta_p[0], |  | ||||||
| 				   (double **)&Cmn[0], ldc, |  | ||||||
| 				   batchCount); |  | ||||||
|     assert(err==HIPBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|     cublasOperation_t hOpA; |  | ||||||
|     cublasOperation_t hOpB; |  | ||||||
|     if ( OpA == GridBLAS_OP_N ) hOpA = CUBLAS_OP_N; |  | ||||||
|     if ( OpA == GridBLAS_OP_T ) hOpA = CUBLAS_OP_T; |  | ||||||
|     if ( OpA == GridBLAS_OP_C ) hOpA = CUBLAS_OP_C; |  | ||||||
|     if ( OpB == GridBLAS_OP_N ) hOpB = CUBLAS_OP_N; |  | ||||||
|     if ( OpB == GridBLAS_OP_T ) hOpB = CUBLAS_OP_T; |  | ||||||
|     if ( OpB == GridBLAS_OP_C ) hOpB = CUBLAS_OP_C; |  | ||||||
|     auto err = cublasDgemmBatched(gridblasHandle, |  | ||||||
| 				  hOpA, |  | ||||||
| 				  hOpB, |  | ||||||
| 				  m,n,k, |  | ||||||
| 				  (double *) &alpha_p[0], |  | ||||||
| 				  (double **)&Amk[0], lda, |  | ||||||
| 				  (double **)&Bkn[0], ldb, |  | ||||||
| 				  (double *) &beta_p[0], |  | ||||||
| 				  (double **)&Cmn[0], ldc, |  | ||||||
| 				  batchCount); |  | ||||||
|     assert(err==CUBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
|     /* |  | ||||||
|       int64_t m64=m; |  | ||||||
|       int64_t n64=n; |  | ||||||
|       int64_t k64=k; |  | ||||||
|       int64_t batchCount64=batchCount; |  | ||||||
|       oneapi::mkl::blas::column_major::gemm_batch(*theGridAccelerator, |  | ||||||
|       onemkl::transpose::N, |  | ||||||
|       onemkl::transpose::N, |  | ||||||
|       &m64,&n64,&k64, |  | ||||||
|       (double *) &alpha_p[0], |  | ||||||
|       (double **)&Amk[0], lda, |  | ||||||
|       (double **)&Bkn[0], ldb, |  | ||||||
|       (double *) &beta_p[0], |  | ||||||
|       (double **)&Cmn[0], ldc, |  | ||||||
|       1,&batchCount64); |  | ||||||
|      */ |  | ||||||
|     //MKL’s cblas_<T>gemm_batch & OneAPI |  | ||||||
| #warning "oneMKL implementation not built " |  | ||||||
| #endif |  | ||||||
| #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) |  | ||||||
|     int sda = lda*k; |  | ||||||
|     int sdb = ldb*k; |  | ||||||
|     int sdc = ldc*n; |  | ||||||
|     // Need a default/reference implementation |  | ||||||
|     for (int p = 0; p < batchCount; ++p) { |  | ||||||
|       for (int mm = 0; mm < m; ++mm) { |  | ||||||
| 	for (int nn = 0; nn < n; ++nn) { |  | ||||||
| 	  RealD c_mn(0.0); |  | ||||||
| 	  for (int kk = 0; kk < k; ++kk) |  | ||||||
| 	    c_mn += Amk[p][mm + kk*lda ] * Bkn[p][kk + nn*ldb]; |  | ||||||
| 	  Cmn[p][mm + nn*ldc] =  (alpha)*c_mn + (beta)*Cmn[p][mm + nn*ldc ]; |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
| #endif |  | ||||||
|      RealD t1=usecond(); |  | ||||||
|      RealD flops = 2.0*m*n*k*batchCount; |  | ||||||
|      RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount; |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|  |  | ||||||
|    |  | ||||||
|   //////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|   // Strided case used by benchmark, but generally unused in Grid |  | ||||||
|   // Keep a code example in double complex, but don't generate the single and real variants for now |  | ||||||
|   //////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|    |  | ||||||
|   void gemmStridedBatched(int m,int n, int k, |  | ||||||
| 			  ComplexD alpha, |  | ||||||
| 			  ComplexD* Amk,  // pointer list to matrices |  | ||||||
| 			  ComplexD* Bkn, |  | ||||||
| 			  ComplexD beta, |  | ||||||
| 			  ComplexD* Cmn, |  | ||||||
| 			  int batchCount) |  | ||||||
|   { |  | ||||||
|     // Use C-row major storage, so transpose calls |  | ||||||
|     int lda = m; // m x k column major |  | ||||||
|     int ldb = k; // k x n column major |  | ||||||
|     int ldc = m; // m x b column major |  | ||||||
|     int sda = m*k; |  | ||||||
|     int sdb = k*n; |  | ||||||
|     int sdc = m*n; |  | ||||||
|     deviceVector<ComplexD> alpha_p(1); |  | ||||||
|     deviceVector<ComplexD> beta_p(1); |  | ||||||
|     acceleratorCopyToDevice((void *)&alpha,(void *)&alpha_p[0],sizeof(ComplexD)); |  | ||||||
|     acceleratorCopyToDevice((void *)&beta ,(void *)&beta_p[0],sizeof(ComplexD)); |  | ||||||
|  |  | ||||||
|     //    std::cout << "blasZgemmStridedBatched mnk  "<<m<<","<<n<<","<<k<<" count "<<batchCount<<std::endl; |  | ||||||
|     //    std::cout << "blasZgemmStridedBatched ld   "<<lda<<","<<ldb<<","<<ldc<<std::endl; |  | ||||||
|     //    std::cout << "blasZgemmStridedBatched sd   "<<sda<<","<<sdb<<","<<sdc<<std::endl; |  | ||||||
| #ifdef GRID_HIP |  | ||||||
|     auto err = hipblasZgemmStridedBatched(gridblasHandle, |  | ||||||
| 					  HIPBLAS_OP_N, |  | ||||||
| 					  HIPBLAS_OP_N, |  | ||||||
| 					  m,n,k, |  | ||||||
| 					  (hipblasDoubleComplex *) &alpha_p[0], |  | ||||||
| 					  (hipblasDoubleComplex *) Amk, lda, sda, |  | ||||||
| 					  (hipblasDoubleComplex *) Bkn, ldb, sdb, |  | ||||||
| 					  (hipblasDoubleComplex *) &beta_p[0], |  | ||||||
| 					  (hipblasDoubleComplex *) Cmn, ldc, sdc, |  | ||||||
| 					  batchCount); |  | ||||||
|     assert(err==HIPBLAS_STATUS_SUCCESS); |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|     cublasZgemmStridedBatched(gridblasHandle, |  | ||||||
| 			      CUBLAS_OP_N, |  | ||||||
| 			      CUBLAS_OP_N, |  | ||||||
| 			      m,n,k, |  | ||||||
| 			      (cuDoubleComplex *) &alpha_p[0], |  | ||||||
| 			      (cuDoubleComplex *) Amk, lda, sda, |  | ||||||
| 			      (cuDoubleComplex *) Bkn, ldb, sdb, |  | ||||||
| 			      (cuDoubleComplex *) &beta_p[0], |  | ||||||
| 			      (cuDoubleComplex *) Cmn, ldc, sdc, |  | ||||||
| 			      batchCount); |  | ||||||
| #endif |  | ||||||
| #if defined(GRID_SYCL) || defined(GRID_ONE_MKL) |  | ||||||
|     oneapi::mkl::blas::column_major::gemm_batch(*gridblasHandle, |  | ||||||
| 						oneapi::mkl::transpose::N, |  | ||||||
| 						oneapi::mkl::transpose::N, |  | ||||||
| 						m,n,k, |  | ||||||
| 						alpha, |  | ||||||
| 						(const ComplexD *)Amk,lda,sda, |  | ||||||
| 						(const ComplexD *)Bkn,ldb,sdb, |  | ||||||
| 						beta, |  | ||||||
| 						(ComplexD *)Cmn,ldc,sdc, |  | ||||||
| 						batchCount); |  | ||||||
| #endif |  | ||||||
| #if !defined(GRID_SYCL) && !defined(GRID_CUDA) && !defined(GRID_HIP) && !defined(GRID_ONE_MKL) |  | ||||||
|      // Need a default/reference implementation |  | ||||||
|      for (int p = 0; p < batchCount; ++p) { |  | ||||||
|        for (int mm = 0; mm < m; ++mm) { |  | ||||||
| 	 for (int nn = 0; nn < n; ++nn) { |  | ||||||
| 	   ComplexD c_mn(0.0); |  | ||||||
| 	   for (int kk = 0; kk < k; ++kk) |  | ||||||
| 	     c_mn += Amk[mm + kk*lda + p*sda] * Bkn[kk + nn*ldb + p*sdb]; |  | ||||||
| 	   Cmn[mm + nn*ldc + p*sdc] =  (alpha)*c_mn + (beta)*Cmn[mm + nn*ldc + p*sdc]; |  | ||||||
| 	 } |  | ||||||
|        } |  | ||||||
|      } |  | ||||||
| #endif |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   double benchmark(int M, int N, int K, int BATCH) |  | ||||||
|   { |  | ||||||
|     int32_t N_A = M*K*BATCH; |  | ||||||
|     int32_t N_B = K*N*BATCH; |  | ||||||
|     int32_t N_C = M*N*BATCH; |  | ||||||
|     deviceVector<ComplexD> A(N_A); acceleratorMemSet(&A[0],0,N_A*sizeof(ComplexD)); |  | ||||||
|     deviceVector<ComplexD> B(N_B); acceleratorMemSet(&B[0],0,N_B*sizeof(ComplexD)); |  | ||||||
|     deviceVector<ComplexD> C(N_C); acceleratorMemSet(&C[0],0,N_C*sizeof(ComplexD)); |  | ||||||
|     ComplexD alpha(1.0); |  | ||||||
|     ComplexD beta (1.0); |  | ||||||
|     RealD flops = 8.0*M*N*K*BATCH; |  | ||||||
|     int ncall=10; |  | ||||||
|     RealD t0 = usecond(); |  | ||||||
|     for(int i=0;i<ncall;i++){ |  | ||||||
|       gemmStridedBatched(M,N,K, |  | ||||||
| 			 alpha, |  | ||||||
| 			 &A[0], // m x k  |  | ||||||
| 			 &B[0], // k x n |  | ||||||
| 			 beta,  |  | ||||||
| 			 &C[0], // m x n |  | ||||||
| 			 BATCH); |  | ||||||
|     } |  | ||||||
|     synchronise(); |  | ||||||
|     RealD t1 = usecond(); |  | ||||||
|     RealD bytes = 1.0*sizeof(ComplexD)*(M*N*2+N*K+M*K)*BATCH; |  | ||||||
|     flops = 8.0*M*N*K*BATCH*ncall; |  | ||||||
|     flops = flops/(t1-t0)/1.e3; |  | ||||||
|     return flops; // Returns gigaflops |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|  |  | ||||||
|  |  | ||||||
|  |  | ||||||
| }; |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
| @@ -36,8 +36,7 @@ NAMESPACE_BEGIN(Grid); | |||||||
| template<class FieldD, class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>  | template<class FieldD, class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>  | ||||||
| class MixedPrecisionBiCGSTAB : public LinearFunction<FieldD>  | class MixedPrecisionBiCGSTAB : public LinearFunction<FieldD>  | ||||||
| { | { | ||||||
|   public: |   public:                                                 | ||||||
|     using LinearFunction<FieldD>::operator(); |  | ||||||
|     RealD   Tolerance; |     RealD   Tolerance; | ||||||
|     RealD   InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed |     RealD   InnerTolerance; // Initial tolerance for inner CG. Defaults to Tolerance but can be changed | ||||||
|     Integer MaxInnerIterations; |     Integer MaxInnerIterations; | ||||||
|   | |||||||
| @@ -58,7 +58,6 @@ public: | |||||||
|  |  | ||||||
|   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) { |   void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) { | ||||||
|  |  | ||||||
|     GRID_TRACE("ConjugateGradient"); |  | ||||||
|     psi.Checkerboard() = src.Checkerboard(); |     psi.Checkerboard() = src.Checkerboard(); | ||||||
|  |  | ||||||
|     conformable(psi, src); |     conformable(psi, src); | ||||||
| @@ -118,13 +117,9 @@ public: | |||||||
|     GridStopWatch MatrixTimer; |     GridStopWatch MatrixTimer; | ||||||
|     GridStopWatch SolverTimer; |     GridStopWatch SolverTimer; | ||||||
|  |  | ||||||
|     RealD usecs = -usecond(); |  | ||||||
|     SolverTimer.Start(); |     SolverTimer.Start(); | ||||||
|     int k; |     int k; | ||||||
|     for (k = 1; k <= MaxIterations; k++) { |     for (k = 1; k <= MaxIterations; k++) { | ||||||
|  |  | ||||||
|       GridStopWatch IterationTimer; |  | ||||||
|       IterationTimer.Start(); |  | ||||||
|       c = cp; |       c = cp; | ||||||
|  |  | ||||||
|       MatrixTimer.Start(); |       MatrixTimer.Start(); | ||||||
| @@ -157,41 +152,31 @@ public: | |||||||
|       LinearCombTimer.Stop(); |       LinearCombTimer.Stop(); | ||||||
|       LinalgTimer.Stop(); |       LinalgTimer.Stop(); | ||||||
|  |  | ||||||
|       IterationTimer.Stop(); |       std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k | ||||||
|       if ( (k % 500) == 0 ) { |  | ||||||
| 	std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k |  | ||||||
|                 << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl; |                 << " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl; | ||||||
|       } else {  |  | ||||||
| 	std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k |  | ||||||
| 		  << " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl; |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|       // Stopping condition |       // Stopping condition | ||||||
|       if (cp <= rsq) { |       if (cp <= rsq) { | ||||||
| 	usecs +=usecond(); |  | ||||||
|         SolverTimer.Stop(); |         SolverTimer.Stop(); | ||||||
|         Linop.HermOpAndNorm(psi, mmp, d, qq); |         Linop.HermOpAndNorm(psi, mmp, d, qq); | ||||||
|         p = mmp - src; |         p = mmp - src; | ||||||
| 	GridBase *grid = src.Grid(); |  | ||||||
| 	RealD DwfFlops = (1452. )*grid->gSites()*4*k |  | ||||||
|    	               + (8+4+8+4+4)*12*grid->gSites()*k; // CG linear algebra |  | ||||||
|         RealD srcnorm = std::sqrt(norm2(src)); |         RealD srcnorm = std::sqrt(norm2(src)); | ||||||
|         RealD resnorm = std::sqrt(norm2(p)); |         RealD resnorm = std::sqrt(norm2(p)); | ||||||
|         RealD true_residual = resnorm / srcnorm; |         RealD true_residual = resnorm / srcnorm; | ||||||
|  |  | ||||||
|         std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k  |         std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k  | ||||||
| 		  << "\tComputed residual " << std::sqrt(cp / ssq) | 		  << "\tComputed residual " << std::sqrt(cp / ssq) | ||||||
| 		  << "\tTrue residual " << true_residual | 		  << "\tTrue residual " << true_residual | ||||||
| 		  << "\tTarget " << Tolerance << std::endl; | 		  << "\tTarget " << Tolerance << std::endl; | ||||||
|  |  | ||||||
|         std::cout << GridLogMessage << "Time breakdown "<<std::endl; |         std::cout << GridLogIterative << "Time breakdown "<<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl; | 	std::cout << GridLogIterative << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl; | 	std::cout << GridLogIterative << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl; | 	std::cout << GridLogIterative << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tInner      " << InnerTimer.Elapsed() <<std::endl; | 	std::cout << GridLogIterative << "\tInner      " << InnerTimer.Elapsed() <<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl; | 	std::cout << GridLogIterative << "\tAxpyNorm   " << AxpyNormTimer.Elapsed() <<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl; | 	std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl; | ||||||
|  |  | ||||||
| 	std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl; |  | ||||||
|  |  | ||||||
|         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0); |         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0); | ||||||
|  |  | ||||||
|   | |||||||
| @@ -35,8 +35,7 @@ NAMESPACE_BEGIN(Grid); | |||||||
|     typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, |     typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, | ||||||
|     typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>  |     typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>  | ||||||
|   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> { |   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> { | ||||||
|   public: |   public:                                                 | ||||||
|     using LinearFunction<FieldD>::operator(); |  | ||||||
|     RealD   Tolerance; |     RealD   Tolerance; | ||||||
|     RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed |     RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed | ||||||
|     Integer MaxInnerIterations; |     Integer MaxInnerIterations; | ||||||
| @@ -49,7 +48,6 @@ NAMESPACE_BEGIN(Grid); | |||||||
|     Integer TotalInnerIterations; //Number of inner CG iterations |     Integer TotalInnerIterations; //Number of inner CG iterations | ||||||
|     Integer TotalOuterIterations; //Number of restarts |     Integer TotalOuterIterations; //Number of restarts | ||||||
|     Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step |     Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step | ||||||
|     RealD TrueResidual; |  | ||||||
|  |  | ||||||
|     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess |     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess | ||||||
|     LinearFunction<FieldF> *guesser; |     LinearFunction<FieldF> *guesser; | ||||||
| @@ -69,7 +67,6 @@ NAMESPACE_BEGIN(Grid); | |||||||
|     } |     } | ||||||
|    |    | ||||||
|   void operator() (const FieldD &src_d_in, FieldD &sol_d){ |   void operator() (const FieldD &src_d_in, FieldD &sol_d){ | ||||||
|     std::cout << GridLogMessage << "MixedPrecisionConjugateGradient: Starting mixed precision CG with outer tolerance " << Tolerance << " and inner tolerance " << InnerTolerance << std::endl; |  | ||||||
|     TotalInnerIterations = 0; |     TotalInnerIterations = 0; | ||||||
| 	 | 	 | ||||||
|     GridStopWatch TotalTimer; |     GridStopWatch TotalTimer; | ||||||
| @@ -99,7 +96,6 @@ NAMESPACE_BEGIN(Grid); | |||||||
|     FieldF sol_f(SinglePrecGrid); |     FieldF sol_f(SinglePrecGrid); | ||||||
|     sol_f.Checkerboard() = cb; |     sol_f.Checkerboard() = cb; | ||||||
|      |      | ||||||
|     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting initial inner CG with tolerance " << inner_tol << std::endl; |  | ||||||
|     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations); |     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations); | ||||||
|     CG_f.ErrorOnNoConverge = false; |     CG_f.ErrorOnNoConverge = false; | ||||||
|  |  | ||||||
| @@ -108,10 +104,7 @@ NAMESPACE_BEGIN(Grid); | |||||||
|     GridStopWatch PrecChangeTimer; |     GridStopWatch PrecChangeTimer; | ||||||
|      |      | ||||||
|     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count |     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count | ||||||
|  |        | ||||||
|     precisionChangeWorkspace pc_wk_sp_to_dp(DoublePrecGrid, SinglePrecGrid); |  | ||||||
|     precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, DoublePrecGrid); |  | ||||||
|      |  | ||||||
|     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){ |     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){ | ||||||
|       //Compute double precision rsd and also new RHS vector. |       //Compute double precision rsd and also new RHS vector. | ||||||
|       Linop_d.HermOp(sol_d, tmp_d); |       Linop_d.HermOp(sol_d, tmp_d); | ||||||
| @@ -126,7 +119,7 @@ NAMESPACE_BEGIN(Grid); | |||||||
|       while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ?? |       while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ?? | ||||||
|  |  | ||||||
|       PrecChangeTimer.Start(); |       PrecChangeTimer.Start(); | ||||||
|       precisionChange(src_f, src_d, pc_wk_dp_to_sp); |       precisionChange(src_f, src_d); | ||||||
|       PrecChangeTimer.Stop(); |       PrecChangeTimer.Stop(); | ||||||
|        |        | ||||||
|       sol_f = Zero(); |       sol_f = Zero(); | ||||||
| @@ -136,7 +129,6 @@ NAMESPACE_BEGIN(Grid); | |||||||
| 	(*guesser)(src_f, sol_f); | 	(*guesser)(src_f, sol_f); | ||||||
|  |  | ||||||
|       //Inner CG |       //Inner CG | ||||||
|       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " << outer_iter << " starting inner CG with tolerance " << inner_tol << std::endl; |  | ||||||
|       CG_f.Tolerance = inner_tol; |       CG_f.Tolerance = inner_tol; | ||||||
|       InnerCGtimer.Start(); |       InnerCGtimer.Start(); | ||||||
|       CG_f(Linop_f, src_f, sol_f); |       CG_f(Linop_f, src_f, sol_f); | ||||||
| @@ -145,7 +137,7 @@ NAMESPACE_BEGIN(Grid); | |||||||
|        |        | ||||||
|       //Convert sol back to double and add to double prec solution |       //Convert sol back to double and add to double prec solution | ||||||
|       PrecChangeTimer.Start(); |       PrecChangeTimer.Start(); | ||||||
|       precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp); |       precisionChange(tmp_d, sol_f); | ||||||
|       PrecChangeTimer.Stop(); |       PrecChangeTimer.Stop(); | ||||||
|        |        | ||||||
|       axpy(sol_d, 1.0, tmp_d, sol_d); |       axpy(sol_d, 1.0, tmp_d, sol_d); | ||||||
| @@ -157,7 +149,6 @@ NAMESPACE_BEGIN(Grid); | |||||||
|     ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations); |     ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations); | ||||||
|     CG_d(Linop_d, src_d_in, sol_d); |     CG_d(Linop_d, src_d_in, sol_d); | ||||||
|     TotalFinalStepIterations = CG_d.IterationsToComplete; |     TotalFinalStepIterations = CG_d.IterationsToComplete; | ||||||
|     TrueResidual = CG_d.TrueResidual; |  | ||||||
|  |  | ||||||
|     TotalTimer.Stop(); |     TotalTimer.Stop(); | ||||||
|     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl; |     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl; | ||||||
|   | |||||||
| @@ -1,213 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid  |  | ||||||
|  |  | ||||||
|     Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2015 |  | ||||||
|  |  | ||||||
|     Author: Raoul Hodgson <raoul.hodgson@ed.ac.uk> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
| *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
| #ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H |  | ||||||
| #define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| //Mixed precision restarted defect correction CG |  | ||||||
| template<class FieldD,class FieldF,  |  | ||||||
|   typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, |  | ||||||
|   typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>  |  | ||||||
| class MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> { |  | ||||||
| public: |  | ||||||
|   using LinearFunction<FieldD>::operator(); |  | ||||||
|   RealD   Tolerance; |  | ||||||
|   RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed |  | ||||||
|   Integer MaxInnerIterations; |  | ||||||
|   Integer MaxOuterIterations; |  | ||||||
|   Integer MaxPatchupIterations; |  | ||||||
|   GridBase* SinglePrecGrid; //Grid for single-precision fields |  | ||||||
|   RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance |  | ||||||
|   LinearOperatorBase<FieldF> &Linop_f; |  | ||||||
|   LinearOperatorBase<FieldD> &Linop_d; |  | ||||||
|  |  | ||||||
|   //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess |  | ||||||
|   LinearFunction<FieldF> *guesser; |  | ||||||
|   bool updateResidual; |  | ||||||
|    |  | ||||||
|   MixedPrecisionConjugateGradientBatched(RealD tol,  |  | ||||||
|           Integer maxinnerit,  |  | ||||||
|           Integer maxouterit,  |  | ||||||
|           Integer maxpatchit, |  | ||||||
|           GridBase* _sp_grid,  |  | ||||||
|           LinearOperatorBase<FieldF> &_Linop_f,  |  | ||||||
|           LinearOperatorBase<FieldD> &_Linop_d, |  | ||||||
|           bool _updateResidual=true) : |  | ||||||
|     Linop_f(_Linop_f), Linop_d(_Linop_d), |  | ||||||
|     Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid), |  | ||||||
|     OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { }; |  | ||||||
|  |  | ||||||
|   void useGuesser(LinearFunction<FieldF> &g){ |  | ||||||
|     guesser = &g; |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|   void operator() (const FieldD &src_d_in, FieldD &sol_d){ |  | ||||||
|     std::vector<FieldD> srcs_d_in{src_d_in}; |  | ||||||
|     std::vector<FieldD> sols_d{sol_d}; |  | ||||||
|  |  | ||||||
|     (*this)(srcs_d_in,sols_d); |  | ||||||
|  |  | ||||||
|     sol_d = sols_d[0]; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){ |  | ||||||
|     assert(src_d_in.size() == sol_d.size()); |  | ||||||
|     int NBatch = src_d_in.size(); |  | ||||||
|  |  | ||||||
|     std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl; |  | ||||||
|  |  | ||||||
|     Integer TotalOuterIterations = 0; //Number of restarts |  | ||||||
|     std::vector<Integer> TotalInnerIterations(NBatch,0);     //Number of inner CG iterations |  | ||||||
|     std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step |  | ||||||
|    |  | ||||||
|     GridStopWatch TotalTimer; |  | ||||||
|     TotalTimer.Start(); |  | ||||||
|  |  | ||||||
|     GridStopWatch InnerCGtimer; |  | ||||||
|     GridStopWatch PrecChangeTimer; |  | ||||||
|      |  | ||||||
|     int cb = src_d_in[0].Checkerboard(); |  | ||||||
|      |  | ||||||
|     std::vector<RealD> src_norm; |  | ||||||
|     std::vector<RealD> norm; |  | ||||||
|     std::vector<RealD> stop; |  | ||||||
|      |  | ||||||
|     GridBase* DoublePrecGrid = src_d_in[0].Grid(); |  | ||||||
|     FieldD tmp_d(DoublePrecGrid); |  | ||||||
|     tmp_d.Checkerboard() = cb; |  | ||||||
|      |  | ||||||
|     FieldD tmp2_d(DoublePrecGrid); |  | ||||||
|     tmp2_d.Checkerboard() = cb; |  | ||||||
|  |  | ||||||
|     std::vector<FieldD> src_d; |  | ||||||
|     std::vector<FieldF> src_f; |  | ||||||
|     std::vector<FieldF> sol_f; |  | ||||||
|  |  | ||||||
|     for (int i=0; i<NBatch; i++) { |  | ||||||
|       sol_d[i].Checkerboard() = cb; |  | ||||||
|  |  | ||||||
|       src_norm.push_back(norm2(src_d_in[i])); |  | ||||||
|       norm.push_back(0.); |  | ||||||
|       stop.push_back(src_norm[i] * Tolerance*Tolerance); |  | ||||||
|  |  | ||||||
|       src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation |  | ||||||
|  |  | ||||||
|       src_f.push_back(SinglePrecGrid); |  | ||||||
|       src_f[i].Checkerboard() = cb; |  | ||||||
|  |  | ||||||
|       sol_f.push_back(SinglePrecGrid); |  | ||||||
|       sol_f[i].Checkerboard() = cb; |  | ||||||
|     } |  | ||||||
|      |  | ||||||
|     RealD inner_tol = InnerTolerance; |  | ||||||
|      |  | ||||||
|     ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations); |  | ||||||
|     CG_f.ErrorOnNoConverge = false; |  | ||||||
|      |  | ||||||
|     Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count |  | ||||||
|        |  | ||||||
|     for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){ |  | ||||||
|       std::cout << GridLogMessage << std::endl; |  | ||||||
|       std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl; |  | ||||||
|        |  | ||||||
|       bool allConverged = true; |  | ||||||
|        |  | ||||||
|       for (int i=0; i<NBatch; i++) { |  | ||||||
|         //Compute double precision rsd and also new RHS vector. |  | ||||||
|         Linop_d.HermOp(sol_d[i], tmp_d); |  | ||||||
|         norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector |  | ||||||
|          |  | ||||||
|         std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl; |  | ||||||
|  |  | ||||||
|         PrecChangeTimer.Start(); |  | ||||||
|         precisionChange(src_f[i], src_d[i]); |  | ||||||
|         PrecChangeTimer.Stop(); |  | ||||||
|          |  | ||||||
|         sol_f[i] = Zero(); |  | ||||||
|        |  | ||||||
|         if(norm[i] > OuterLoopNormMult * stop[i]) { |  | ||||||
|           allConverged = false; |  | ||||||
|         } |  | ||||||
|       } |  | ||||||
|       if (allConverged) break; |  | ||||||
|  |  | ||||||
|       if (updateResidual) { |  | ||||||
|         RealD normMax = *std::max_element(std::begin(norm), std::end(norm)); |  | ||||||
|         RealD stopMax = *std::max_element(std::begin(stop), std::end(stop)); |  | ||||||
|         while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ?? |  | ||||||
|         CG_f.Tolerance = inner_tol; |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|       //Optionally improve inner solver guess (eg using known eigenvectors) |  | ||||||
|       if(guesser != NULL) { |  | ||||||
|         (*guesser)(src_f, sol_f); |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|       for (int i=0; i<NBatch; i++) { |  | ||||||
|         //Inner CG |  | ||||||
|         InnerCGtimer.Start(); |  | ||||||
|         CG_f(Linop_f, src_f[i], sol_f[i]); |  | ||||||
|         InnerCGtimer.Stop(); |  | ||||||
|         TotalInnerIterations[i] += CG_f.IterationsToComplete; |  | ||||||
|          |  | ||||||
|         //Convert sol back to double and add to double prec solution |  | ||||||
|         PrecChangeTimer.Start(); |  | ||||||
|         precisionChange(tmp_d, sol_f[i]); |  | ||||||
|         PrecChangeTimer.Stop(); |  | ||||||
|          |  | ||||||
|         axpy(sol_d[i], 1.0, tmp_d, sol_d[i]); |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|     } |  | ||||||
|      |  | ||||||
|     //Final trial CG |  | ||||||
|     std::cout << GridLogMessage << std::endl; |  | ||||||
|     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl; |  | ||||||
|      |  | ||||||
|     for (int i=0; i<NBatch; i++) { |  | ||||||
|       ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations); |  | ||||||
|       CG_d(Linop_d, src_d_in[i], sol_d[i]); |  | ||||||
|       TotalFinalStepIterations[i] += CG_d.IterationsToComplete; |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     TotalTimer.Stop(); |  | ||||||
|  |  | ||||||
|     std::cout << GridLogMessage << std::endl; |  | ||||||
|     for (int i=0; i<NBatch; i++) { |  | ||||||
|       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl; |  | ||||||
|     } |  | ||||||
|     std::cout << GridLogMessage << std::endl; |  | ||||||
|     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl; |  | ||||||
|      |  | ||||||
|   } |  | ||||||
| }; |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
|  |  | ||||||
| #endif |  | ||||||
| @@ -44,7 +44,7 @@ public: | |||||||
|  |  | ||||||
|   using OperatorFunction<Field>::operator(); |   using OperatorFunction<Field>::operator(); | ||||||
|  |  | ||||||
|   //  RealD   Tolerance; |   RealD   Tolerance; | ||||||
|   Integer MaxIterations; |   Integer MaxIterations; | ||||||
|   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion |   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion | ||||||
|   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift |   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift | ||||||
| @@ -52,7 +52,7 @@ public: | |||||||
|   MultiShiftFunction shifts; |   MultiShiftFunction shifts; | ||||||
|   std::vector<RealD> TrueResidualShift; |   std::vector<RealD> TrueResidualShift; | ||||||
|  |  | ||||||
|   ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) :  |   ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) :  | ||||||
|     MaxIterations(maxit), |     MaxIterations(maxit), | ||||||
|     shifts(_shifts) |     shifts(_shifts) | ||||||
|   {  |   {  | ||||||
| @@ -84,7 +84,6 @@ public: | |||||||
|  |  | ||||||
|   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi) |   void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi) | ||||||
|   { |   { | ||||||
|     GRID_TRACE("ConjugateGradientMultiShift"); |  | ||||||
|    |    | ||||||
|     GridBase *grid = src.Grid(); |     GridBase *grid = src.Grid(); | ||||||
|    |    | ||||||
| @@ -183,9 +182,6 @@ public: | |||||||
|     for(int s=0;s<nshift;s++) { |     for(int s=0;s<nshift;s++) { | ||||||
|       axpby(psi[s],0.,-bs[s]*alpha[s],src,src); |       axpby(psi[s],0.,-bs[s]*alpha[s],src,src); | ||||||
|     } |     } | ||||||
|  |  | ||||||
|     std::cout << GridLogIterative << "ConjugateGradientMultiShift: initial rn (|src|^2) =" << rn << " qq (|MdagM src|^2) =" << qq << " d ( dot(src, [MdagM + m_0]src) ) =" << d << " c=" << c << std::endl; |  | ||||||
|      |  | ||||||
|    |    | ||||||
|   /////////////////////////////////////// |   /////////////////////////////////////// | ||||||
|   // Timers |   // Timers | ||||||
| @@ -325,8 +321,8 @@ public: | |||||||
|  |  | ||||||
|       std::cout << GridLogMessage << "Time Breakdown "<<std::endl; |       std::cout << GridLogMessage << "Time Breakdown "<<std::endl; | ||||||
|       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl; |       std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl; | ||||||
|       std::cout << GridLogMessage << "\tAXPY     " << AXPYTimer.Elapsed()     <<std::endl; |       std::cout << GridLogMessage << "\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl; | ||||||
|       std::cout << GridLogMessage << "\tMatrix   " << MatrixTimer.Elapsed()     <<std::endl; |       std::cout << GridLogMessage << "\tMarix    " << MatrixTimer.Elapsed()     <<std::endl; | ||||||
|       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl; |       std::cout << GridLogMessage << "\tShift    " << ShiftTimer.Elapsed()     <<std::endl; | ||||||
|  |  | ||||||
|       IterationsToComplete = k;	 |       IterationsToComplete = k;	 | ||||||
|   | |||||||
| @@ -1,373 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid  |  | ||||||
|  |  | ||||||
|     Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2015 |  | ||||||
|  |  | ||||||
| Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk> |  | ||||||
| Author: Peter Boyle <paboyle@ph.ed.ac.uk> |  | ||||||
| Author: Christopher Kelly <ckelly@bnl.gov> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
| *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
| #pragma once |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| //CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision.  |  | ||||||
| //The residual is stored in single precision, but the search directions and solution are stored in double precision.  |  | ||||||
| //Every update_freq iterations the residual is corrected in double precision.  |  | ||||||
| //For safety the a final regular CG is applied to clean up if necessary |  | ||||||
|  |  | ||||||
| //PB Pure single, then double fixup |  | ||||||
|  |  | ||||||
| template<class FieldD, class FieldF, |  | ||||||
| 	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, |  | ||||||
| 	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>  |  | ||||||
| class ConjugateGradientMultiShiftMixedPrecCleanup : public OperatorMultiFunction<FieldD>, |  | ||||||
| 					     public OperatorFunction<FieldD> |  | ||||||
| { |  | ||||||
| public:                                                 |  | ||||||
|  |  | ||||||
|   using OperatorFunction<FieldD>::operator(); |  | ||||||
|  |  | ||||||
|   RealD   Tolerance; |  | ||||||
|   Integer MaxIterationsMshift; |  | ||||||
|   Integer MaxIterations; |  | ||||||
|   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion |  | ||||||
|   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift |  | ||||||
|   int verbose; |  | ||||||
|   MultiShiftFunction shifts; |  | ||||||
|   std::vector<RealD> TrueResidualShift; |  | ||||||
|  |  | ||||||
|   int ReliableUpdateFreq; //number of iterations between reliable updates |  | ||||||
|  |  | ||||||
|   GridBase* SinglePrecGrid; //Grid for single-precision fields |  | ||||||
|   LinearOperatorBase<FieldF> &Linop_f; //single precision |  | ||||||
|  |  | ||||||
|   ConjugateGradientMultiShiftMixedPrecCleanup(Integer maxit, const MultiShiftFunction &_shifts, |  | ||||||
| 				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f, |  | ||||||
| 				       int _ReliableUpdateFreq) :  |  | ||||||
|     MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq), |  | ||||||
|     MaxIterations(20000) |  | ||||||
|   {  |  | ||||||
|     verbose=1; |  | ||||||
|     IterationsToCompleteShift.resize(_shifts.order); |  | ||||||
|     TrueResidualShift.resize(_shifts.order); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi) |  | ||||||
|   { |  | ||||||
|     GridBase *grid = src.Grid(); |  | ||||||
|     int nshift = shifts.order; |  | ||||||
|     std::vector<FieldD> results(nshift,grid); |  | ||||||
|     (*this)(Linop,src,results,psi); |  | ||||||
|   } |  | ||||||
|   void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi) |  | ||||||
|   { |  | ||||||
|     int nshift = shifts.order; |  | ||||||
|  |  | ||||||
|     (*this)(Linop,src,results); |  | ||||||
|    |  | ||||||
|     psi = shifts.norm*src; |  | ||||||
|     for(int i=0;i<nshift;i++){ |  | ||||||
|       psi = psi + shifts.residues[i]*results[i]; |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     return; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d) |  | ||||||
|   {  |  | ||||||
|     GRID_TRACE("ConjugateGradientMultiShiftMixedPrecCleanup"); |  | ||||||
|     GridBase *DoublePrecGrid = src_d.Grid(); |  | ||||||
|  |  | ||||||
|     //////////////////////////////////////////////////////////////////////// |  | ||||||
|     // Convenience references to the info stored in "MultiShiftFunction" |  | ||||||
|     //////////////////////////////////////////////////////////////////////// |  | ||||||
|     int nshift = shifts.order; |  | ||||||
|  |  | ||||||
|     std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts" |  | ||||||
|     std::vector<RealD> &mresidual(shifts.tolerances); |  | ||||||
|     std::vector<RealD> alpha(nshift,1.0); |  | ||||||
|  |  | ||||||
|     //Double precision search directions |  | ||||||
|     FieldD p_d(DoublePrecGrid); |  | ||||||
|     std::vector<FieldF> ps_f (nshift, SinglePrecGrid);// Search directions (single precision) |  | ||||||
|     std::vector<FieldF> psi_f(nshift, SinglePrecGrid);// solutions (single precision) |  | ||||||
|  |  | ||||||
|     FieldD tmp_d(DoublePrecGrid); |  | ||||||
|     FieldD r_d(DoublePrecGrid); |  | ||||||
|     FieldF r_f(SinglePrecGrid); |  | ||||||
|     FieldD mmp_d(DoublePrecGrid); |  | ||||||
|  |  | ||||||
|     assert(psi_d.size()==nshift); |  | ||||||
|     assert(mass.size()==nshift); |  | ||||||
|     assert(mresidual.size()==nshift); |  | ||||||
|    |  | ||||||
|     // dynamic sized arrays on stack; 2d is a pain with vector |  | ||||||
|     RealD  bs[nshift]; |  | ||||||
|     RealD  rsq[nshift]; |  | ||||||
|     RealD  rsqf[nshift]; |  | ||||||
|     RealD  z[nshift][2]; |  | ||||||
|     int     converged[nshift]; |  | ||||||
|    |  | ||||||
|     const int       primary =0; |  | ||||||
|    |  | ||||||
|     //Primary shift fields CG iteration |  | ||||||
|     RealD a,b,c,d; |  | ||||||
|     RealD cp,bp,qq; //prev |  | ||||||
|    |  | ||||||
|     // Matrix mult fields |  | ||||||
|     FieldF p_f(SinglePrecGrid); |  | ||||||
|     FieldF mmp_f(SinglePrecGrid); |  | ||||||
|  |  | ||||||
|     // Check lightest mass |  | ||||||
|     for(int s=0;s<nshift;s++){ |  | ||||||
|       assert( mass[s]>= mass[primary] ); |  | ||||||
|       converged[s]=0; |  | ||||||
|     } |  | ||||||
|    |  | ||||||
|     // Wire guess to zero |  | ||||||
|     // Residuals "r" are src |  | ||||||
|     // First search direction "p" is also src |  | ||||||
|     cp = norm2(src_d); |  | ||||||
|  |  | ||||||
|     // Handle trivial case of zero src. |  | ||||||
|     if( cp == 0. ){ |  | ||||||
|       for(int s=0;s<nshift;s++){ |  | ||||||
| 	psi_d[s] = Zero(); |  | ||||||
| 	psi_f[s] = Zero(); |  | ||||||
| 	IterationsToCompleteShift[s] = 1; |  | ||||||
| 	TrueResidualShift[s] = 0.; |  | ||||||
|       } |  | ||||||
|       return; |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     for(int s=0;s<nshift;s++){ |  | ||||||
|       rsq[s] = cp * mresidual[s] * mresidual[s]; |  | ||||||
|       rsqf[s] =rsq[s]; |  | ||||||
|       std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift "<< s <<" target resid "<<rsq[s]<<std::endl; |  | ||||||
|       //      ps_d[s] = src_d; |  | ||||||
|       precisionChange(ps_f[s],src_d); |  | ||||||
|     } |  | ||||||
|     // r and p for primary |  | ||||||
|     p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys |  | ||||||
|     r_d = p_d; |  | ||||||
|      |  | ||||||
|     //MdagM+m[0] |  | ||||||
|     precisionChange(p_f,p_d); |  | ||||||
|     Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp) |  | ||||||
|     precisionChange(tmp_d,mmp_f); |  | ||||||
|     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp) |  | ||||||
|     tmp_d = tmp_d - mmp_d; |  | ||||||
|     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl; |  | ||||||
|     //    assert(norm2(tmp_d)< 1.0e-4); |  | ||||||
|  |  | ||||||
|     axpy(mmp_d,mass[0],p_d,mmp_d); |  | ||||||
|     RealD rn = norm2(p_d); |  | ||||||
|     d += rn*mass[0]; |  | ||||||
|  |  | ||||||
|     b = -cp /d; |  | ||||||
|    |  | ||||||
|     // Set up the various shift variables |  | ||||||
|     int       iz=0; |  | ||||||
|     z[0][1-iz] = 1.0; |  | ||||||
|     z[0][iz]   = 1.0; |  | ||||||
|     bs[0]      = b; |  | ||||||
|     for(int s=1;s<nshift;s++){ |  | ||||||
|       z[s][1-iz] = 1.0; |  | ||||||
|       z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0])); |  | ||||||
|       bs[s]      = b*z[s][iz];  |  | ||||||
|     } |  | ||||||
|    |  | ||||||
|     // r += b[0] A.p[0] |  | ||||||
|     // c= norm(r) |  | ||||||
|     c=axpy_norm(r_d,b,mmp_d,r_d); |  | ||||||
|    |  | ||||||
|     for(int s=0;s<nshift;s++) { |  | ||||||
|       axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d); |  | ||||||
|       precisionChange(psi_f[s],psi_d[s]); |  | ||||||
|     } |  | ||||||
|    |  | ||||||
|     /////////////////////////////////////// |  | ||||||
|     // Timers |  | ||||||
|     /////////////////////////////////////// |  | ||||||
|     GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer; |  | ||||||
|  |  | ||||||
|     SolverTimer.Start(); |  | ||||||
|    |  | ||||||
|     // Iteration loop |  | ||||||
|     int k; |  | ||||||
|    |  | ||||||
|     for (k=1;k<=MaxIterationsMshift;k++){     |  | ||||||
|  |  | ||||||
|       a = c /cp; |  | ||||||
|       AXPYTimer.Start(); |  | ||||||
|       axpy(p_d,a,p_d,r_d);  |  | ||||||
|       AXPYTimer.Stop(); |  | ||||||
|  |  | ||||||
|       PrecChangeTimer.Start(); |  | ||||||
|       precisionChange(r_f, r_d); |  | ||||||
|       PrecChangeTimer.Stop(); |  | ||||||
|  |  | ||||||
|       AXPYTimer.Start(); |  | ||||||
|       for(int s=0;s<nshift;s++){ |  | ||||||
| 	if ( ! converged[s] ) {  |  | ||||||
| 	  if (s==0){ |  | ||||||
| 	    axpy(ps_f[s],a,ps_f[s],r_f); |  | ||||||
| 	  } else{ |  | ||||||
| 	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b); |  | ||||||
| 	    axpby(ps_f[s],z[s][iz],as,r_f,ps_f[s]); |  | ||||||
| 	  } |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|       AXPYTimer.Stop(); |  | ||||||
|  |  | ||||||
|       cp=c; |  | ||||||
|       PrecChangeTimer.Start(); |  | ||||||
|       precisionChange(p_f, p_d); //get back single prec search direction for linop |  | ||||||
|       PrecChangeTimer.Stop(); |  | ||||||
|       MatrixTimer.Start();   |  | ||||||
|       Linop_f.HermOp(p_f,mmp_f); |  | ||||||
|       MatrixTimer.Stop();   |  | ||||||
|       PrecChangeTimer.Start(); |  | ||||||
|       precisionChange(mmp_d, mmp_f); // From Float to Double |  | ||||||
|       PrecChangeTimer.Stop(); |  | ||||||
|  |  | ||||||
|       d=real(innerProduct(p_d,mmp_d));     |  | ||||||
|       axpy(mmp_d,mass[0],p_d,mmp_d); |  | ||||||
|       RealD rn = norm2(p_d); |  | ||||||
|       d += rn*mass[0]; |  | ||||||
|      |  | ||||||
|       bp=b; |  | ||||||
|       b=-cp/d; |  | ||||||
|  |  | ||||||
|       // Toggle the recurrence history |  | ||||||
|       bs[0] = b; |  | ||||||
|       iz = 1-iz; |  | ||||||
|       ShiftTimer.Start(); |  | ||||||
|       for(int s=1;s<nshift;s++){ |  | ||||||
| 	if((!converged[s])){ |  | ||||||
| 	  RealD z0 = z[s][1-iz]; |  | ||||||
| 	  RealD z1 = z[s][iz]; |  | ||||||
| 	  z[s][iz] = z0*z1*bp |  | ||||||
| 	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b));  |  | ||||||
| 	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|       ShiftTimer.Stop(); |  | ||||||
|  |  | ||||||
|       //Update single precision solutions |  | ||||||
|       AXPYTimer.Start(); |  | ||||||
|       for(int s=0;s<nshift;s++){ |  | ||||||
| 	int ss = s; |  | ||||||
| 	if( (!converged[s]) ) {  |  | ||||||
| 	  axpy(psi_f[ss],-bs[s]*alpha[s],ps_f[s],psi_f[ss]); |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|       c = axpy_norm(r_d,b,mmp_d,r_d); |  | ||||||
|       AXPYTimer.Stop(); |  | ||||||
|      |  | ||||||
|       // Convergence checks |  | ||||||
|       int all_converged = 1; |  | ||||||
|       for(int s=0;s<nshift;s++){ |  | ||||||
|        |  | ||||||
| 	if ( (!converged[s]) ){ |  | ||||||
| 	  IterationsToCompleteShift[s] = k; |  | ||||||
| 	 |  | ||||||
| 	  RealD css  = c * z[s][iz]* z[s][iz]; |  | ||||||
| 	 |  | ||||||
| 	  if(css<rsqf[s]){ |  | ||||||
| 	    if ( ! converged[s] ) |  | ||||||
| 	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup k="<<k<<" Shift "<<s<<" has converged"<<std::endl; |  | ||||||
| 	    converged[s]=1; |  | ||||||
| 	  } else { |  | ||||||
| 	    all_converged=0; |  | ||||||
| 	  } |  | ||||||
|  |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|       if ( all_converged || k == MaxIterationsMshift-1){ |  | ||||||
|  |  | ||||||
| 	SolverTimer.Stop(); |  | ||||||
|  |  | ||||||
| 	for(int s=0;s<nshift;s++){ |  | ||||||
| 	  precisionChange(psi_d[s],psi_f[s]); |  | ||||||
| 	} |  | ||||||
|  |  | ||||||
| 	 |  | ||||||
| 	if ( all_converged ){ |  | ||||||
| 	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: All shifts have converged iteration "<<k<<std::endl; |  | ||||||
| 	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Checking solutions"<<std::endl; |  | ||||||
| 	} else { |  | ||||||
| 	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Not all shifts have converged iteration "<<k<<std::endl; |  | ||||||
| 	} |  | ||||||
| 	 |  | ||||||
| 	// Check answers  |  | ||||||
| 	for(int s=0; s < nshift; s++) {  |  | ||||||
| 	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq); |  | ||||||
| 	  axpy(tmp_d,mass[s],psi_d[s],mmp_d); |  | ||||||
| 	  axpy(r_d,-alpha[s],src_d,tmp_d); |  | ||||||
| 	  RealD rn = norm2(r_d); |  | ||||||
| 	  RealD cn = norm2(src_d); |  | ||||||
| 	  TrueResidualShift[s] = std::sqrt(rn/cn); |  | ||||||
| 	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl; |  | ||||||
|  |  | ||||||
| 	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup |  | ||||||
| 	  if(rn >= rsq[s]){ |  | ||||||
| 	    CleanupTimer.Start(); |  | ||||||
| 	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: performing cleanup step for shift " << s << std::endl; |  | ||||||
|  |  | ||||||
| 	    //Setup linear operators for final cleanup |  | ||||||
| 	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]); |  | ||||||
| 	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]); |  | ||||||
| 					        |  | ||||||
| 	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d);  |  | ||||||
| 	    cg(src_d, psi_d[s]); |  | ||||||
| 	     |  | ||||||
| 	    TrueResidualShift[s] = cg.TrueResidual; |  | ||||||
| 	    CleanupTimer.Stop(); |  | ||||||
| 	  } |  | ||||||
| 	} |  | ||||||
|  |  | ||||||
| 	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrecCleanup: Time Breakdown for body"<<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl; |  | ||||||
|  |  | ||||||
| 	IterationsToComplete = k;	 |  | ||||||
|  |  | ||||||
| 	return; |  | ||||||
|       } |  | ||||||
|     |  | ||||||
|     } |  | ||||||
|     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl; |  | ||||||
|     assert(0); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
| }; |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
|  |  | ||||||
| @@ -1,416 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid  |  | ||||||
|  |  | ||||||
|     Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2015 |  | ||||||
|  |  | ||||||
| Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk> |  | ||||||
| Author: Peter Boyle <paboyle@ph.ed.ac.uk> |  | ||||||
| Author: Christopher Kelly <ckelly@bnl.gov> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
| *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
| #ifndef GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H |  | ||||||
| #define GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| //CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision.  |  | ||||||
| //The residual is stored in single precision, but the search directions and solution are stored in double precision.  |  | ||||||
| //Every update_freq iterations the residual is corrected in double precision.  |  | ||||||
|      |  | ||||||
| //For safety the a final regular CG is applied to clean up if necessary |  | ||||||
|  |  | ||||||
| //Linop to add shift to input linop, used in cleanup CG |  | ||||||
| namespace ConjugateGradientMultiShiftMixedPrecSupport{ |  | ||||||
| template<typename Field> |  | ||||||
| class ShiftedLinop: public LinearOperatorBase<Field>{ |  | ||||||
| public: |  | ||||||
|   LinearOperatorBase<Field> &linop_base; |  | ||||||
|   RealD shift; |  | ||||||
|  |  | ||||||
|   ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){} |  | ||||||
|  |  | ||||||
|   void OpDiag (const Field &in, Field &out){ assert(0); } |  | ||||||
|   void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); } |  | ||||||
|   void OpDirAll  (const Field &in, std::vector<Field> &out){ assert(0); } |  | ||||||
|    |  | ||||||
|   void Op     (const Field &in, Field &out){ assert(0); } |  | ||||||
|   void AdjOp  (const Field &in, Field &out){ assert(0); } |  | ||||||
|  |  | ||||||
|   void HermOp(const Field &in, Field &out){ |  | ||||||
|     linop_base.HermOp(in, out); |  | ||||||
|     axpy(out, shift, in, out); |  | ||||||
|   }     |  | ||||||
|  |  | ||||||
|   void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ |  | ||||||
|     HermOp(in,out); |  | ||||||
|     ComplexD dot = innerProduct(in,out); |  | ||||||
|     n1=real(dot); |  | ||||||
|     n2=norm2(out); |  | ||||||
|   } |  | ||||||
| }; |  | ||||||
| }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
| template<class FieldD, class FieldF, |  | ||||||
| 	 typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0, |  | ||||||
| 	 typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>  |  | ||||||
| class ConjugateGradientMultiShiftMixedPrec : public OperatorMultiFunction<FieldD>, |  | ||||||
| 					     public OperatorFunction<FieldD> |  | ||||||
| { |  | ||||||
| public:                                                 |  | ||||||
|  |  | ||||||
|   using OperatorFunction<FieldD>::operator(); |  | ||||||
|  |  | ||||||
|   RealD   Tolerance; |  | ||||||
|   Integer MaxIterationsMshift; |  | ||||||
|   Integer MaxIterations; |  | ||||||
|   Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion |  | ||||||
|   std::vector<int> IterationsToCompleteShift;  // Iterations for this shift |  | ||||||
|   int verbose; |  | ||||||
|   MultiShiftFunction shifts; |  | ||||||
|   std::vector<RealD> TrueResidualShift; |  | ||||||
|  |  | ||||||
|   int ReliableUpdateFreq; //number of iterations between reliable updates |  | ||||||
|  |  | ||||||
|   GridBase* SinglePrecGrid; //Grid for single-precision fields |  | ||||||
|   LinearOperatorBase<FieldF> &Linop_f; //single precision |  | ||||||
|  |  | ||||||
|   ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts, |  | ||||||
| 				       GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f, |  | ||||||
| 				       int _ReliableUpdateFreq) :  |  | ||||||
|     MaxIterationsMshift(maxit),  shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq), |  | ||||||
|     MaxIterations(20000) |  | ||||||
|   {  |  | ||||||
|     verbose=1; |  | ||||||
|     IterationsToCompleteShift.resize(_shifts.order); |  | ||||||
|     TrueResidualShift.resize(_shifts.order); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi) |  | ||||||
|   { |  | ||||||
|     GridBase *grid = src.Grid(); |  | ||||||
|     int nshift = shifts.order; |  | ||||||
|     std::vector<FieldD> results(nshift,grid); |  | ||||||
|     (*this)(Linop,src,results,psi); |  | ||||||
|   } |  | ||||||
|   void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi) |  | ||||||
|   { |  | ||||||
|     int nshift = shifts.order; |  | ||||||
|  |  | ||||||
|     (*this)(Linop,src,results); |  | ||||||
|    |  | ||||||
|     psi = shifts.norm*src; |  | ||||||
|     for(int i=0;i<nshift;i++){ |  | ||||||
|       psi = psi + shifts.residues[i]*results[i]; |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     return; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d) |  | ||||||
|   {  |  | ||||||
|     GRID_TRACE("ConjugateGradientMultiShiftMixedPrec"); |  | ||||||
|     GridBase *DoublePrecGrid = src_d.Grid(); |  | ||||||
|  |  | ||||||
|     precisionChangeWorkspace pc_wk_s_to_d(DoublePrecGrid,SinglePrecGrid); |  | ||||||
|     precisionChangeWorkspace pc_wk_d_to_s(SinglePrecGrid,DoublePrecGrid); |  | ||||||
|      |  | ||||||
|     //////////////////////////////////////////////////////////////////////// |  | ||||||
|     // Convenience references to the info stored in "MultiShiftFunction" |  | ||||||
|     //////////////////////////////////////////////////////////////////////// |  | ||||||
|     int nshift = shifts.order; |  | ||||||
|  |  | ||||||
|     std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts" |  | ||||||
|     std::vector<RealD> &mresidual(shifts.tolerances); |  | ||||||
|     std::vector<RealD> alpha(nshift,1.0); |  | ||||||
|  |  | ||||||
|     //Double precision search directions |  | ||||||
|     FieldD p_d(DoublePrecGrid); |  | ||||||
|     std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision) |  | ||||||
|  |  | ||||||
|     FieldD tmp_d(DoublePrecGrid); |  | ||||||
|     FieldD r_d(DoublePrecGrid); |  | ||||||
|     FieldD mmp_d(DoublePrecGrid); |  | ||||||
|  |  | ||||||
|     assert(psi_d.size()==nshift); |  | ||||||
|     assert(mass.size()==nshift); |  | ||||||
|     assert(mresidual.size()==nshift); |  | ||||||
|    |  | ||||||
|     // dynamic sized arrays on stack; 2d is a pain with vector |  | ||||||
|     RealD  bs[nshift]; |  | ||||||
|     RealD  rsq[nshift]; |  | ||||||
|     RealD  rsqf[nshift]; |  | ||||||
|     RealD  z[nshift][2]; |  | ||||||
|     int     converged[nshift]; |  | ||||||
|    |  | ||||||
|     const int       primary =0; |  | ||||||
|    |  | ||||||
|     //Primary shift fields CG iteration |  | ||||||
|     RealD a,b,c,d; |  | ||||||
|     RealD cp,bp,qq; //prev |  | ||||||
|    |  | ||||||
|     // Matrix mult fields |  | ||||||
|     FieldF p_f(SinglePrecGrid); |  | ||||||
|     FieldF mmp_f(SinglePrecGrid); |  | ||||||
|  |  | ||||||
|     // Check lightest mass |  | ||||||
|     for(int s=0;s<nshift;s++){ |  | ||||||
|       assert( mass[s]>= mass[primary] ); |  | ||||||
|       converged[s]=0; |  | ||||||
|     } |  | ||||||
|    |  | ||||||
|     // Wire guess to zero |  | ||||||
|     // Residuals "r" are src |  | ||||||
|     // First search direction "p" is also src |  | ||||||
|     cp = norm2(src_d); |  | ||||||
|  |  | ||||||
|     // Handle trivial case of zero src. |  | ||||||
|     if( cp == 0. ){ |  | ||||||
|       for(int s=0;s<nshift;s++){ |  | ||||||
| 	psi_d[s] = Zero(); |  | ||||||
| 	IterationsToCompleteShift[s] = 1; |  | ||||||
| 	TrueResidualShift[s] = 0.; |  | ||||||
|       } |  | ||||||
|       return; |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     for(int s=0;s<nshift;s++){ |  | ||||||
|       rsq[s] = cp * mresidual[s] * mresidual[s]; |  | ||||||
|       rsqf[s] =rsq[s]; |  | ||||||
|       std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl; |  | ||||||
|       ps_d[s] = src_d; |  | ||||||
|     } |  | ||||||
|     // r and p for primary |  | ||||||
|     p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys |  | ||||||
|     r_d = p_d; |  | ||||||
|      |  | ||||||
|     //MdagM+m[0] |  | ||||||
|     precisionChange(p_f, p_d, pc_wk_d_to_s); |  | ||||||
|  |  | ||||||
|     Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp) |  | ||||||
|     precisionChange(tmp_d, mmp_f, pc_wk_s_to_d); |  | ||||||
|     Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p        d=real(dot(p, mmp)),  qq=norm2(mmp) |  | ||||||
|     tmp_d = tmp_d - mmp_d; |  | ||||||
|     std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl; |  | ||||||
|     assert(norm2(tmp_d)< 1.0); |  | ||||||
|  |  | ||||||
|     axpy(mmp_d,mass[0],p_d,mmp_d); |  | ||||||
|     RealD rn = norm2(p_d); |  | ||||||
|     d += rn*mass[0]; |  | ||||||
|  |  | ||||||
|     b = -cp /d; |  | ||||||
|    |  | ||||||
|     // Set up the various shift variables |  | ||||||
|     int       iz=0; |  | ||||||
|     z[0][1-iz] = 1.0; |  | ||||||
|     z[0][iz]   = 1.0; |  | ||||||
|     bs[0]      = b; |  | ||||||
|     for(int s=1;s<nshift;s++){ |  | ||||||
|       z[s][1-iz] = 1.0; |  | ||||||
|       z[s][iz]   = 1.0/( 1.0 - b*(mass[s]-mass[0])); |  | ||||||
|       bs[s]      = b*z[s][iz];  |  | ||||||
|     } |  | ||||||
|    |  | ||||||
|     // r += b[0] A.p[0] |  | ||||||
|     // c= norm(r) |  | ||||||
|     c=axpy_norm(r_d,b,mmp_d,r_d); |  | ||||||
|    |  | ||||||
|     for(int s=0;s<nshift;s++) { |  | ||||||
|       axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d); |  | ||||||
|     } |  | ||||||
|    |  | ||||||
|     /////////////////////////////////////// |  | ||||||
|     // Timers |  | ||||||
|     /////////////////////////////////////// |  | ||||||
|     GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer; |  | ||||||
|  |  | ||||||
|     SolverTimer.Start(); |  | ||||||
|    |  | ||||||
|     // Iteration loop |  | ||||||
|     int k; |  | ||||||
|    |  | ||||||
|     for (k=1;k<=MaxIterationsMshift;k++){     |  | ||||||
|  |  | ||||||
|       a = c /cp; |  | ||||||
|       AXPYTimer.Start(); |  | ||||||
|       axpy(p_d,a,p_d,r_d);  |  | ||||||
|  |  | ||||||
|       for(int s=0;s<nshift;s++){ |  | ||||||
| 	if ( ! converged[s] ) {  |  | ||||||
| 	  if (s==0){ |  | ||||||
| 	    axpy(ps_d[s],a,ps_d[s],r_d); |  | ||||||
| 	  } else{ |  | ||||||
| 	    RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b); |  | ||||||
| 	    axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]); |  | ||||||
| 	  } |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|       AXPYTimer.Stop(); |  | ||||||
|  |  | ||||||
|       PrecChangeTimer.Start(); |  | ||||||
|       precisionChange(p_f, p_d, pc_wk_d_to_s); //get back single prec search direction for linop |  | ||||||
|       PrecChangeTimer.Stop(); |  | ||||||
|  |  | ||||||
|       cp=c; |  | ||||||
|       MatrixTimer.Start();   |  | ||||||
|       Linop_f.HermOp(p_f,mmp_f); |  | ||||||
|       MatrixTimer.Stop();   |  | ||||||
|  |  | ||||||
|       PrecChangeTimer.Start(); |  | ||||||
|       precisionChange(mmp_d, mmp_f, pc_wk_s_to_d); // From Float to Double |  | ||||||
|       PrecChangeTimer.Stop(); |  | ||||||
|  |  | ||||||
|       AXPYTimer.Start(); |  | ||||||
|       d=real(innerProduct(p_d,mmp_d));     |  | ||||||
|       axpy(mmp_d,mass[0],p_d,mmp_d); |  | ||||||
|       AXPYTimer.Stop(); |  | ||||||
|       RealD rn = norm2(p_d); |  | ||||||
|       d += rn*mass[0]; |  | ||||||
|      |  | ||||||
|       bp=b; |  | ||||||
|       b=-cp/d; |  | ||||||
|  |  | ||||||
|       // Toggle the recurrence history |  | ||||||
|       bs[0] = b; |  | ||||||
|       iz = 1-iz; |  | ||||||
|       ShiftTimer.Start(); |  | ||||||
|       for(int s=1;s<nshift;s++){ |  | ||||||
| 	if((!converged[s])){ |  | ||||||
| 	  RealD z0 = z[s][1-iz]; |  | ||||||
| 	  RealD z1 = z[s][iz]; |  | ||||||
| 	  z[s][iz] = z0*z1*bp |  | ||||||
| 	    / (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b));  |  | ||||||
| 	  bs[s] = b*z[s][iz]/z0; // NB sign  rel to Mike |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|       ShiftTimer.Stop(); |  | ||||||
|  |  | ||||||
|       //Update double precision solutions |  | ||||||
|       AXPYTimer.Start(); |  | ||||||
|       for(int s=0;s<nshift;s++){ |  | ||||||
| 	int ss = s; |  | ||||||
| 	if( (!converged[s]) ) {  |  | ||||||
| 	  axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]); |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|       //Perform reliable update if necessary; otherwise update residual from single-prec mmp |  | ||||||
|       c = axpy_norm(r_d,b,mmp_d,r_d); |  | ||||||
|  |  | ||||||
|       AXPYTimer.Stop(); |  | ||||||
|  |  | ||||||
|       if(k % ReliableUpdateFreq == 0){ |  | ||||||
| 	RealD c_old = c; |  | ||||||
| 	//Replace r with true residual |  | ||||||
| 	MatrixTimer.Start();   |  | ||||||
| 	Linop_d.HermOp(psi_d[0],mmp_d);  |  | ||||||
| 	MatrixTimer.Stop();   |  | ||||||
|  |  | ||||||
| 	AXPYTimer.Start(); |  | ||||||
| 	axpy(mmp_d,mass[0],psi_d[0],mmp_d); |  | ||||||
|  |  | ||||||
| 	c = axpy_norm(r_d, -1.0, mmp_d, src_d); |  | ||||||
| 	AXPYTimer.Stop(); |  | ||||||
|  |  | ||||||
| 	std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_old <<" with |r|^2 = "<<c<<std::endl; |  | ||||||
|       } |  | ||||||
|      |  | ||||||
|       // Convergence checks |  | ||||||
|       int all_converged = 1; |  | ||||||
|       for(int s=0;s<nshift;s++){ |  | ||||||
|        |  | ||||||
| 	if ( (!converged[s]) ){ |  | ||||||
| 	  IterationsToCompleteShift[s] = k; |  | ||||||
| 	 |  | ||||||
| 	  RealD css  = c * z[s][iz]* z[s][iz]; |  | ||||||
| 	 |  | ||||||
| 	  if(css<rsqf[s]){ |  | ||||||
| 	    if ( ! converged[s] ) |  | ||||||
| 	      std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl; |  | ||||||
| 	    converged[s]=1; |  | ||||||
| 	  } else { |  | ||||||
| 	    all_converged=0; |  | ||||||
| 	  } |  | ||||||
|  |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|       if ( all_converged || k == MaxIterationsMshift-1){ |  | ||||||
|  |  | ||||||
| 	SolverTimer.Stop(); |  | ||||||
|  |  | ||||||
| 	if ( all_converged ){ |  | ||||||
| 	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl; |  | ||||||
| 	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl; |  | ||||||
| 	} else { |  | ||||||
| 	  std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Not all shifts have converged iteration "<<k<<std::endl; |  | ||||||
| 	} |  | ||||||
| 	 |  | ||||||
| 	// Check answers  |  | ||||||
| 	for(int s=0; s < nshift; s++) {  |  | ||||||
| 	  Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq); |  | ||||||
| 	  axpy(tmp_d,mass[s],psi_d[s],mmp_d); |  | ||||||
| 	  axpy(r_d,-alpha[s],src_d,tmp_d); |  | ||||||
| 	  RealD rn = norm2(r_d); |  | ||||||
| 	  RealD cn = norm2(src_d); |  | ||||||
| 	  TrueResidualShift[s] = std::sqrt(rn/cn); |  | ||||||
| 	  std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl; |  | ||||||
|  |  | ||||||
| 	  //If we have not reached the desired tolerance, do a (mixed precision) CG cleanup |  | ||||||
| 	  if(rn >= rsq[s]){ |  | ||||||
| 	    CleanupTimer.Start(); |  | ||||||
| 	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl; |  | ||||||
|  |  | ||||||
| 	    //Setup linear operators for final cleanup |  | ||||||
| 	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]); |  | ||||||
| 	    ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]); |  | ||||||
| 					        |  | ||||||
| 	    MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d);  |  | ||||||
| 	    cg(src_d, psi_d[s]); |  | ||||||
| 	     |  | ||||||
| 	    TrueResidualShift[s] = cg.TrueResidual; |  | ||||||
| 	    CleanupTimer.Stop(); |  | ||||||
| 	  } |  | ||||||
| 	} |  | ||||||
|  |  | ||||||
| 	std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\tSolver    " << SolverTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\t\tAXPY    " << AXPYTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\t\tMatrix    " << MatrixTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\t\tShift    " << ShiftTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed()     <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl; |  | ||||||
|  |  | ||||||
| 	IterationsToComplete = k;	 |  | ||||||
|  |  | ||||||
| 	return; |  | ||||||
|       } |  | ||||||
|     |  | ||||||
|     } |  | ||||||
|     std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl; |  | ||||||
|     assert(0); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
| }; |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
| #endif |  | ||||||
| @@ -48,7 +48,7 @@ public: | |||||||
|   LinearOperatorBase<FieldF> &Linop_f; |   LinearOperatorBase<FieldF> &Linop_f; | ||||||
|   LinearOperatorBase<FieldD> &Linop_d; |   LinearOperatorBase<FieldD> &Linop_d; | ||||||
|   GridBase* SinglePrecGrid; |   GridBase* SinglePrecGrid; | ||||||
|   RealD Delta; //reliable update parameter. A reliable update is performed when the residual drops by a factor of Delta relative to its value at the last update |   RealD Delta; //reliable update parameter | ||||||
|  |  | ||||||
|   //Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single |   //Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single | ||||||
|   LinearOperatorBase<FieldF> *Linop_fallback; |   LinearOperatorBase<FieldF> *Linop_fallback; | ||||||
| @@ -65,9 +65,7 @@ public: | |||||||
|       ErrorOnNoConverge(err_on_no_conv), |       ErrorOnNoConverge(err_on_no_conv), | ||||||
|       DoFinalCleanup(true), |       DoFinalCleanup(true), | ||||||
|       Linop_fallback(NULL) |       Linop_fallback(NULL) | ||||||
|   { |   {}; | ||||||
|     assert(Delta > 0. && Delta < 1. && "Expect  0 < Delta < 1"); |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){ |   void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){ | ||||||
|     Linop_fallback = &_Linop_fallback; |     Linop_fallback = &_Linop_fallback; | ||||||
| @@ -75,7 +73,6 @@ public: | |||||||
|   } |   } | ||||||
|      |      | ||||||
|   void operator()(const FieldD &src, FieldD &psi) { |   void operator()(const FieldD &src, FieldD &psi) { | ||||||
|     GRID_TRACE("ConjugateGradientReliableUpdate"); |  | ||||||
|     LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f; |     LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f; | ||||||
|     bool using_fallback = false; |     bool using_fallback = false; | ||||||
|        |        | ||||||
| @@ -118,12 +115,9 @@ public: | |||||||
|     } |     } | ||||||
|  |  | ||||||
|     //Single prec initialization |     //Single prec initialization | ||||||
|     precisionChangeWorkspace pc_wk_sp_to_dp(src.Grid(), SinglePrecGrid); |  | ||||||
|     precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, src.Grid()); |  | ||||||
|      |  | ||||||
|     FieldF r_f(SinglePrecGrid); |     FieldF r_f(SinglePrecGrid); | ||||||
|     r_f.Checkerboard() = r.Checkerboard(); |     r_f.Checkerboard() = r.Checkerboard(); | ||||||
|     precisionChange(r_f, r, pc_wk_dp_to_sp); |     precisionChange(r_f, r); | ||||||
|  |  | ||||||
|     FieldF psi_f(r_f); |     FieldF psi_f(r_f); | ||||||
|     psi_f = Zero(); |     psi_f = Zero(); | ||||||
| @@ -139,8 +133,7 @@ public: | |||||||
|     GridStopWatch LinalgTimer; |     GridStopWatch LinalgTimer; | ||||||
|     GridStopWatch MatrixTimer; |     GridStopWatch MatrixTimer; | ||||||
|     GridStopWatch SolverTimer; |     GridStopWatch SolverTimer; | ||||||
|     GridStopWatch PrecChangeTimer; |  | ||||||
|      |  | ||||||
|     SolverTimer.Start(); |     SolverTimer.Start(); | ||||||
|     int k = 0; |     int k = 0; | ||||||
|     int l = 0; |     int l = 0; | ||||||
| @@ -179,9 +172,7 @@ public: | |||||||
|       // Stopping condition |       // Stopping condition | ||||||
|       if (cp <= rsq) { |       if (cp <= rsq) { | ||||||
| 	//Although not written in the paper, I assume that I have to add on the final solution | 	//Although not written in the paper, I assume that I have to add on the final solution | ||||||
| 	PrecChangeTimer.Start(); | 	precisionChange(mmp, psi_f); | ||||||
| 	precisionChange(mmp, psi_f, pc_wk_sp_to_dp); |  | ||||||
| 	PrecChangeTimer.Stop(); |  | ||||||
| 	psi = psi + mmp; | 	psi = psi + mmp; | ||||||
| 	 | 	 | ||||||
| 	 | 	 | ||||||
| @@ -202,10 +193,7 @@ public: | |||||||
| 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl; | 	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl; | 	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl; | 	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl; | ||||||
| 	std::cout << GridLogMessage << "\tPrecChange " << PrecChangeTimer.Elapsed() <<std::endl; |  | ||||||
| 	std::cout << GridLogMessage << "\tPrecChange avg time " << PrecChangeTimer.Elapsed()/(2*l+1) <<std::endl; |  | ||||||
|  |  | ||||||
| 	 |  | ||||||
| 	IterationsToComplete = k;	 | 	IterationsToComplete = k;	 | ||||||
| 	ReliableUpdatesPerformed = l; | 	ReliableUpdatesPerformed = l; | ||||||
| 	   | 	   | ||||||
| @@ -225,21 +213,14 @@ public: | |||||||
|       else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update |       else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update | ||||||
| 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate " | 	std::cout << GridLogMessage << "ConjugateGradientReliableUpdate " | ||||||
| 		  << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n"; | 		  << cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n"; | ||||||
| 	PrecChangeTimer.Start(); | 	precisionChange(mmp, psi_f); | ||||||
| 	precisionChange(mmp, psi_f, pc_wk_sp_to_dp); |  | ||||||
| 	PrecChangeTimer.Stop(); |  | ||||||
| 	psi = psi + mmp; | 	psi = psi + mmp; | ||||||
|  |  | ||||||
| 	MatrixTimer.Start(); |  | ||||||
| 	Linop_d.HermOpAndNorm(psi, mmp, d, qq); | 	Linop_d.HermOpAndNorm(psi, mmp, d, qq); | ||||||
| 	MatrixTimer.Stop(); |  | ||||||
| 	 |  | ||||||
| 	r = src - mmp; | 	r = src - mmp; | ||||||
|  |  | ||||||
| 	psi_f = Zero(); | 	psi_f = Zero(); | ||||||
| 	PrecChangeTimer.Start(); | 	precisionChange(r_f, r); | ||||||
| 	precisionChange(r_f, r, pc_wk_dp_to_sp); |  | ||||||
| 	PrecChangeTimer.Stop(); |  | ||||||
| 	cp = norm2(r); | 	cp = norm2(r); | ||||||
| 	MaxResidSinceLastRelUp = cp; | 	MaxResidSinceLastRelUp = cp; | ||||||
|  |  | ||||||
|   | |||||||
| @@ -33,19 +33,16 @@ namespace Grid { | |||||||
| template<class Field> | template<class Field> | ||||||
| class ZeroGuesser: public LinearFunction<Field> { | class ZeroGuesser: public LinearFunction<Field> { | ||||||
| public: | public: | ||||||
|   using LinearFunction<Field>::operator(); |  | ||||||
|     virtual void operator()(const Field &src, Field &guess) { guess = Zero(); }; |     virtual void operator()(const Field &src, Field &guess) { guess = Zero(); }; | ||||||
| }; | }; | ||||||
| template<class Field> | template<class Field> | ||||||
| class DoNothingGuesser: public LinearFunction<Field> { | class DoNothingGuesser: public LinearFunction<Field> { | ||||||
| public: | public: | ||||||
|   using LinearFunction<Field>::operator(); |  | ||||||
|   virtual void operator()(const Field &src, Field &guess) {  }; |   virtual void operator()(const Field &src, Field &guess) {  }; | ||||||
| }; | }; | ||||||
| template<class Field> | template<class Field> | ||||||
| class SourceGuesser: public LinearFunction<Field> { | class SourceGuesser: public LinearFunction<Field> { | ||||||
| public: | public: | ||||||
|   using LinearFunction<Field>::operator(); |  | ||||||
|   virtual void operator()(const Field &src, Field &guess) { guess = src; }; |   virtual void operator()(const Field &src, Field &guess) { guess = src; }; | ||||||
| }; | }; | ||||||
|  |  | ||||||
| @@ -57,24 +54,15 @@ class DeflatedGuesser: public LinearFunction<Field> { | |||||||
| private: | private: | ||||||
|   const std::vector<Field> &evec; |   const std::vector<Field> &evec; | ||||||
|   const std::vector<RealD> &eval; |   const std::vector<RealD> &eval; | ||||||
|   const unsigned int       N; |  | ||||||
|  |  | ||||||
| public: | public: | ||||||
|   using LinearFunction<Field>::operator(); |  | ||||||
|  |  | ||||||
|   DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) |   DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {}; | ||||||
|   : DeflatedGuesser(_evec, _eval, _evec.size()) |  | ||||||
|   {} |  | ||||||
|  |  | ||||||
|   DeflatedGuesser(const std::vector<Field> & _evec, const std::vector<RealD> & _eval, const unsigned int _N) |  | ||||||
|   : evec(_evec), eval(_eval), N(_N) |  | ||||||
|   { |  | ||||||
|     assert(evec.size()==eval.size()); |  | ||||||
|     assert(N <= evec.size()); |  | ||||||
|   }  |  | ||||||
|  |  | ||||||
|   virtual void operator()(const Field &src,Field &guess) { |   virtual void operator()(const Field &src,Field &guess) { | ||||||
|     guess = Zero(); |     guess = Zero(); | ||||||
|  |     assert(evec.size()==eval.size()); | ||||||
|  |     auto N = evec.size(); | ||||||
|     for (int i=0;i<N;i++) { |     for (int i=0;i<N;i++) { | ||||||
|       const Field& tmp = evec[i]; |       const Field& tmp = evec[i]; | ||||||
|       axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess); |       axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess); | ||||||
| @@ -91,7 +79,6 @@ private: | |||||||
|   const std::vector<RealD>       &eval_coarse; |   const std::vector<RealD>       &eval_coarse; | ||||||
| public: | public: | ||||||
|    |    | ||||||
|   using LinearFunction<FineField>::operator(); |  | ||||||
|   LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace, |   LocalCoherenceDeflatedGuesser(const std::vector<FineField>   &_subspace, | ||||||
| 				const std::vector<CoarseField> &_evec_coarse, | 				const std::vector<CoarseField> &_evec_coarse, | ||||||
| 				const std::vector<RealD>       &_eval_coarse) | 				const std::vector<RealD>       &_eval_coarse) | ||||||
| @@ -113,43 +100,7 @@ public: | |||||||
|     blockPromote(guess_coarse,guess,subspace); |     blockPromote(guess_coarse,guess,subspace); | ||||||
|     guess.Checkerboard() = src.Checkerboard(); |     guess.Checkerboard() = src.Checkerboard(); | ||||||
|   }; |   }; | ||||||
|  | }; | ||||||
|   void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) { |  | ||||||
|     int Nevec = (int)evec_coarse.size(); |  | ||||||
|     int Nsrc = (int)src.size(); |  | ||||||
|     // make temp variables |  | ||||||
|     std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid()); |  | ||||||
|     std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());     |  | ||||||
|     //Preporcessing |  | ||||||
|     std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl; |  | ||||||
|     for (int j=0;j<Nsrc;j++) |  | ||||||
|     { |  | ||||||
|     guess_coarse[j] = Zero(); |  | ||||||
|     std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl; |  | ||||||
|     blockProject(src_coarse[j],src[j],subspace); |  | ||||||
|     } |  | ||||||
|     //deflation set up for eigen vector batchsize 1 and source batch size equal number of sources |  | ||||||
|     std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl; |  | ||||||
|     for (int i=0;i<Nevec;i++) |  | ||||||
|     { |  | ||||||
|       std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl; |  | ||||||
|       const CoarseField & tmp = evec_coarse[i]; |  | ||||||
|       for (int j=0;j<Nsrc;j++) |  | ||||||
|       { |  | ||||||
|         axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]); |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
|     //postprocessing |  | ||||||
|     std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl; |  | ||||||
|     for (int j=0;j<Nsrc;j++) |  | ||||||
|     { |  | ||||||
|     std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl; |  | ||||||
|     blockPromote(guess_coarse[j],guess[j],subspace); |  | ||||||
|     guess[j].Checkerboard() = src[j].Checkerboard(); |  | ||||||
|     } |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
|  |  | ||||||
|   | |||||||
										
											
												File diff suppressed because it is too large
												Load Diff
											
										
									
								
							| @@ -419,15 +419,14 @@ until convergence | |||||||
| 	} | 	} | ||||||
|       } |       } | ||||||
|  |  | ||||||
|       if ( Nconv < Nstop ) { |       if ( Nconv < Nstop ) | ||||||
| 	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl; | 	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl; | ||||||
| 	std::cout << GridLogIRL << "returning Nstop vectors, the last "<< Nstop-Nconv << "of which might meet convergence criterion only approximately" <<std::endl; |  | ||||||
|       } |  | ||||||
|       eval=eval2; |       eval=eval2; | ||||||
|        |        | ||||||
|       //Keep only converged |       //Keep only converged | ||||||
|       eval.resize(Nstop);// was Nconv |       eval.resize(Nconv);// Nstop? | ||||||
|       evec.resize(Nstop,grid);// was Nconv |       evec.resize(Nconv,grid);// Nstop? | ||||||
|       basisSortInPlace(evec,eval,reverse); |       basisSortInPlace(evec,eval,reverse); | ||||||
|        |        | ||||||
|     } |     } | ||||||
|   | |||||||
| @@ -44,7 +44,6 @@ public: | |||||||
| 				  int, MinRes);    // Must restart | 				  int, MinRes);    // Must restart | ||||||
| }; | }; | ||||||
|  |  | ||||||
| //This class is the input parameter class for some testing programs |  | ||||||
| struct LocalCoherenceLanczosParams : Serializable { | struct LocalCoherenceLanczosParams : Serializable { | ||||||
| public: | public: | ||||||
|   GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams, |   GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams, | ||||||
| @@ -68,7 +67,6 @@ public: | |||||||
| template<class Fobj,class CComplex,int nbasis> | template<class Fobj,class CComplex,int nbasis> | ||||||
| class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > { | class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > { | ||||||
| public: | public: | ||||||
|   using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator(); |  | ||||||
|   typedef iVector<CComplex,nbasis >           CoarseSiteVector; |   typedef iVector<CComplex,nbasis >           CoarseSiteVector; | ||||||
|   typedef Lattice<CoarseSiteVector>           CoarseField; |   typedef Lattice<CoarseSiteVector>           CoarseField; | ||||||
|   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field |   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field | ||||||
| @@ -99,7 +97,6 @@ public: | |||||||
| template<class Fobj,class CComplex,int nbasis> | template<class Fobj,class CComplex,int nbasis> | ||||||
| class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > { | class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > { | ||||||
| public: | public: | ||||||
|   using LinearFunction<Lattice<iVector<CComplex,nbasis > > >::operator(); |  | ||||||
|   typedef iVector<CComplex,nbasis >           CoarseSiteVector; |   typedef iVector<CComplex,nbasis >           CoarseSiteVector; | ||||||
|   typedef Lattice<CoarseSiteVector>           CoarseField; |   typedef Lattice<CoarseSiteVector>           CoarseField; | ||||||
|   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field |   typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field | ||||||
| @@ -146,24 +143,16 @@ public: | |||||||
|   LinearOperatorBase<FineField> &_Linop; |   LinearOperatorBase<FineField> &_Linop; | ||||||
|   RealD                             _coarse_relax_tol; |   RealD                             _coarse_relax_tol; | ||||||
|   std::vector<FineField>        &_subspace; |   std::vector<FineField>        &_subspace; | ||||||
|  |  | ||||||
|   int _largestEvalIdxForReport; //The convergence of the LCL is based on the evals of the coarse grid operator, not those of the underlying fine grid operator |  | ||||||
|                                 //As a result we do not know what the eval range of the fine operator is until the very end, making tuning the Cheby bounds very difficult |  | ||||||
|                                 //To work around this issue, every restart we separately reconstruct the fine operator eval for the lowest and highest evec and print these |  | ||||||
|                                 //out alongside the evals of the coarse operator. To do so we need to know the index of the largest eval (i.e. Nstop-1) |  | ||||||
|                                 //NOTE: If largestEvalIdxForReport=-1 (default) then this is not performed |  | ||||||
|    |    | ||||||
|   ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly, |   ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly, | ||||||
| 					   OperatorFunction<FineField>   &smoother, | 					   OperatorFunction<FineField>   &smoother, | ||||||
| 					   LinearOperatorBase<FineField> &Linop, | 					   LinearOperatorBase<FineField> &Linop, | ||||||
| 					   std::vector<FineField>        &subspace, | 					   std::vector<FineField>        &subspace, | ||||||
| 					   RealD coarse_relax_tol=5.0e3, | 					   RealD coarse_relax_tol=5.0e3)  | ||||||
| 					   int largestEvalIdxForReport=-1)  |  | ||||||
|     : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace), |     : _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace), | ||||||
|       _coarse_relax_tol(coarse_relax_tol), _largestEvalIdxForReport(largestEvalIdxForReport) |       _coarse_relax_tol(coarse_relax_tol)   | ||||||
|   {    }; |   {    }; | ||||||
|  |  | ||||||
|   //evalMaxApprox: approximation of largest eval of the fine Chebyshev operator (suitably wrapped by block projection) |  | ||||||
|   int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox) |   int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox) | ||||||
|   { |   { | ||||||
|     CoarseField v(B); |     CoarseField v(B); | ||||||
| @@ -186,26 +175,12 @@ public: | |||||||
| 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv | 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv | ||||||
| 	     <<std::endl; | 	     <<std::endl; | ||||||
|  |  | ||||||
|     if(_largestEvalIdxForReport != -1 && (j==0 || j==_largestEvalIdxForReport)){ |  | ||||||
|       std::cout<<GridLogIRL << "Estimating true eval of fine grid operator for eval idx " << j << std::endl; |  | ||||||
|       RealD tmp_eval; |  | ||||||
|       ReconstructEval(j,eresid,B,tmp_eval,1.0); //don't use evalMaxApprox of coarse operator! (cf below) |  | ||||||
|     } |  | ||||||
|      |  | ||||||
|     int conv=0; |     int conv=0; | ||||||
|     if( (vv<eresid*eresid) ) conv = 1; |     if( (vv<eresid*eresid) ) conv = 1; | ||||||
|     return conv; |     return conv; | ||||||
|   } |   } | ||||||
|  |   int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox) | ||||||
|   //This function is called at the end of the coarse grid Lanczos. It promotes the coarse eigenvector 'B' to the fine grid, |  | ||||||
|   //applies a smoother to the result then computes the computes the *fine grid* eigenvalue (output as 'eval'). |  | ||||||
|  |  | ||||||
|   //evalMaxApprox should be the approximation of the largest eval of the fine Hermop. However when this function is called by IRL it actually passes the largest eval of the *Chebyshev* operator (as this is the max approx used for the TestConvergence above) |  | ||||||
|   //As the largest eval of the Chebyshev is typically several orders of magnitude larger this makes the convergence test pass even when it should not. |  | ||||||
|   //We therefore ignore evalMaxApprox here and use a value of 1.0 (note this value is already used by TestCoarse) |  | ||||||
|   int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)   |  | ||||||
|   { |   { | ||||||
|     evalMaxApprox = 1.0; //cf above |  | ||||||
|     GridBase *FineGrid = _subspace[0].Grid();     |     GridBase *FineGrid = _subspace[0].Grid();     | ||||||
|     int checkerboard   = _subspace[0].Checkerboard(); |     int checkerboard   = _subspace[0].Checkerboard(); | ||||||
|     FineField fB(FineGrid);fB.Checkerboard() =checkerboard; |     FineField fB(FineGrid);fB.Checkerboard() =checkerboard; | ||||||
| @@ -224,13 +199,13 @@ public: | |||||||
|     eval   = vnum/vden; |     eval   = vnum/vden; | ||||||
|     fv -= eval*fB; |     fv -= eval*fB; | ||||||
|     RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0); |     RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0); | ||||||
|     if ( j > nbasis ) eresid = eresid*_coarse_relax_tol; |  | ||||||
|      |  | ||||||
|     std::cout.precision(13); |     std::cout.precision(13); | ||||||
|     std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] " |     std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] " | ||||||
| 	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")" | 	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")" | ||||||
| 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv << " target " << eresid*eresid | 	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv | ||||||
| 	     <<std::endl; | 	     <<std::endl; | ||||||
|  |     if ( j > nbasis ) eresid = eresid*_coarse_relax_tol; | ||||||
|     if( (vv<eresid*eresid) ) return 1; |     if( (vv<eresid*eresid) ) return 1; | ||||||
|     return 0; |     return 0; | ||||||
|   } |   } | ||||||
| @@ -308,10 +283,6 @@ public: | |||||||
|     evals_coarse.resize(0); |     evals_coarse.resize(0); | ||||||
|   }; |   }; | ||||||
|  |  | ||||||
|   //The block inner product is the inner product on the fine grid locally summed over the blocks |  | ||||||
|   //to give a Lattice<Scalar> on the coarse grid. This function orthnormalizes the fine-grid subspace |  | ||||||
|   //vectors under the block inner product. This step must be performed after computing the fine grid |  | ||||||
|   //eigenvectors and before computing the coarse grid eigenvectors.     |  | ||||||
|   void Orthogonalise(void ) { |   void Orthogonalise(void ) { | ||||||
|     CoarseScalar InnerProd(_CoarseGrid); |     CoarseScalar InnerProd(_CoarseGrid); | ||||||
|     std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl; |     std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl; | ||||||
| @@ -355,8 +326,6 @@ public: | |||||||
|     } |     } | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   //While this method serves to check the coarse eigenvectors, it also recomputes the eigenvalues from the smoothed reconstructed eigenvectors |  | ||||||
|   //hence the smoother can be tuned after running the coarse Lanczos by using a different smoother here |  | ||||||
|   void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax)  |   void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax)  | ||||||
|   { |   { | ||||||
|     assert(evals_fine.size() == nbasis); |     assert(evals_fine.size() == nbasis); | ||||||
| @@ -405,31 +374,25 @@ public: | |||||||
|     evals_fine.resize(nbasis); |     evals_fine.resize(nbasis); | ||||||
|     subspace.resize(nbasis,_FineGrid); |     subspace.resize(nbasis,_FineGrid); | ||||||
|   } |   } | ||||||
|  |  | ||||||
|  |  | ||||||
|   //cheby_op: Parameters of the fine grid Chebyshev polynomial used for the Lanczos acceleration |  | ||||||
|   //cheby_smooth: Parameters of a separate Chebyshev polynomial used after the Lanczos has completed to smooth out high frequency noise in the reconstructed fine grid eigenvectors prior to computing the eigenvalue |  | ||||||
|   //relax: Reconstructed eigenvectors (post smoothing) are naturally not as precise as true eigenvectors. This factor acts as a multiplier on the stopping condition when determining whether the results satisfy the user provided stopping condition |  | ||||||
|   void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax, |   void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax, | ||||||
| 		  int Nstop, int Nk, int Nm,RealD resid,  | 		  int Nstop, int Nk, int Nm,RealD resid,  | ||||||
| 		  RealD MaxIt, RealD betastp, int MinRes) | 		  RealD MaxIt, RealD betastp, int MinRes) | ||||||
|   { |   { | ||||||
|     Chebyshev<FineField>                          Cheby(cheby_op); //Chebyshev of fine operator on fine grid |     Chebyshev<FineField>                          Cheby(cheby_op); | ||||||
|     ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace); //Fine operator on coarse grid with intermediate fine grid conversion |     ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,subspace); | ||||||
|     ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace); //Chebyshev of fine operator on coarse grid with intermediate fine grid conversion |     ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace); | ||||||
|     ////////////////////////////////////////////////////////////////////////////////////////////////// |     ////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
|     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL |     // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL | ||||||
|     ////////////////////////////////////////////////////////////////////////////////////////////////// |     ////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
|  |  | ||||||
|     Chebyshev<FineField>                                           ChebySmooth(cheby_smooth); //lower order Chebyshev of fine operator on fine grid used to smooth regenerated eigenvectors |     Chebyshev<FineField>                                           ChebySmooth(cheby_smooth); | ||||||
|     ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1);  |     ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax); | ||||||
|  |  | ||||||
|     evals_coarse.resize(Nm); |     evals_coarse.resize(Nm); | ||||||
|     evec_coarse.resize(Nm,_CoarseGrid); |     evec_coarse.resize(Nm,_CoarseGrid); | ||||||
|  |  | ||||||
|     CoarseField src(_CoarseGrid);     src=1.0;  |     CoarseField src(_CoarseGrid);     src=1.0;  | ||||||
|  |  | ||||||
|     //Note the "tester" here is also responsible for generating the fine grid eigenvalues which are output into the "evals_coarse" array |  | ||||||
|     ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes); |     ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes); | ||||||
|     int Nconv=0; |     int Nconv=0; | ||||||
|     IRL.calc(evals_coarse,evec_coarse,src,Nconv,false); |     IRL.calc(evals_coarse,evec_coarse,src,Nconv,false); | ||||||
| @@ -440,14 +403,6 @@ public: | |||||||
|       std::cout << i << " Coarse eval = " << evals_coarse[i]  << std::endl; |       std::cout << i << " Coarse eval = " << evals_coarse[i]  << std::endl; | ||||||
|     } |     } | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   //Get the fine eigenvector 'i' by reconstruction |  | ||||||
|   void getFineEvecEval(FineField &evec, RealD &eval, const int i) const{ |  | ||||||
|     blockPromote(evec_coarse[i],evec,subspace);   |  | ||||||
|     eval = evals_coarse[i]; |  | ||||||
|   } |  | ||||||
|      |  | ||||||
|      |  | ||||||
| }; | }; | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|   | |||||||
| @@ -29,8 +29,6 @@ template<class Field> class PowerMethod | |||||||
|       RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.  |       RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.  | ||||||
|       RealD vden = norm2(src_n);  |       RealD vden = norm2(src_n);  | ||||||
|       RealD na = vnum/vden;  |       RealD na = vnum/vden;  | ||||||
|  |  | ||||||
|       std::cout << GridLogIterative << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl; |  | ||||||
|        |        | ||||||
|       if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) {  |       if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) {  | ||||||
|  	evalMaxApprox = na;  |  	evalMaxApprox = na;  | ||||||
|   | |||||||
| @@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid); | |||||||
| template<class Field> | template<class Field> | ||||||
| class PrecGeneralisedConjugateResidual : public LinearFunction<Field> { | class PrecGeneralisedConjugateResidual : public LinearFunction<Field> { | ||||||
| public:                                                 | public:                                                 | ||||||
|   using LinearFunction<Field>::operator(); |  | ||||||
|   RealD   Tolerance; |   RealD   Tolerance; | ||||||
|   Integer MaxIterations; |   Integer MaxIterations; | ||||||
|   int verbose; |   int verbose; | ||||||
|   | |||||||
| @@ -43,7 +43,7 @@ NAMESPACE_BEGIN(Grid); | |||||||
| template<class Field> | template<class Field> | ||||||
| class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> { | class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> { | ||||||
| public:                                                 | public:                                                 | ||||||
|   using LinearFunction<Field>::operator(); |  | ||||||
|   RealD   Tolerance; |   RealD   Tolerance; | ||||||
|   Integer MaxIterations; |   Integer MaxIterations; | ||||||
|   int verbose; |   int verbose; | ||||||
| @@ -119,8 +119,7 @@ public: | |||||||
|   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){ |   RealD GCRnStep(const Field &src, Field &psi,RealD rsq){ | ||||||
|  |  | ||||||
|     RealD cp; |     RealD cp; | ||||||
|     ComplexD a, b; |     ComplexD a, b, zAz; | ||||||
|     //    ComplexD zAz; |  | ||||||
|     RealD zAAz; |     RealD zAAz; | ||||||
|     ComplexD rq; |     ComplexD rq; | ||||||
|  |  | ||||||
| @@ -147,7 +146,7 @@ public: | |||||||
|     ////////////////////////////////// |     ////////////////////////////////// | ||||||
|     MatTimer.Start(); |     MatTimer.Start(); | ||||||
|     Linop.Op(psi,Az); |     Linop.Op(psi,Az); | ||||||
|     //    zAz = innerProduct(Az,psi); |     zAz = innerProduct(Az,psi); | ||||||
|     zAAz= norm2(Az); |     zAAz= norm2(Az); | ||||||
|     MatTimer.Stop(); |     MatTimer.Stop(); | ||||||
|      |      | ||||||
| @@ -171,7 +170,7 @@ public: | |||||||
|  |  | ||||||
|     LinalgTimer.Start(); |     LinalgTimer.Start(); | ||||||
|  |  | ||||||
|     //    zAz = innerProduct(Az,psi); |     zAz = innerProduct(Az,psi); | ||||||
|     zAAz= norm2(Az); |     zAAz= norm2(Az); | ||||||
|  |  | ||||||
|     //p[0],q[0],qq[0]  |     //p[0],q[0],qq[0]  | ||||||
| @@ -213,7 +212,7 @@ public: | |||||||
|       MatTimer.Start(); |       MatTimer.Start(); | ||||||
|       Linop.Op(z,Az); |       Linop.Op(z,Az); | ||||||
|       MatTimer.Stop(); |       MatTimer.Stop(); | ||||||
|       //      zAz = innerProduct(Az,psi); |       zAz = innerProduct(Az,psi); | ||||||
|       zAAz= norm2(Az); |       zAAz= norm2(Az); | ||||||
|  |  | ||||||
|       LinalgTimer.Start(); |       LinalgTimer.Start(); | ||||||
|   | |||||||
| @@ -132,31 +132,6 @@ namespace Grid { | |||||||
|       (*this)(_Matrix,in,out,guess); |       (*this)(_Matrix,in,out,guess); | ||||||
|     } |     } | ||||||
|  |  | ||||||
|     void RedBlackSource(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &src_o)  |  | ||||||
|     { |  | ||||||
|       GridBase *grid = _Matrix.RedBlackGrid(); |  | ||||||
|       Field tmp(grid); |  | ||||||
|       int nblock = in.size(); |  | ||||||
|       for(int b=0;b<nblock;b++){ |  | ||||||
| 	RedBlackSource(_Matrix,in[b],tmp,src_o[b]); |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
|     // James can write his own deflated guesser |  | ||||||
|     // with optimised code for the inner products |  | ||||||
|     //    RedBlackSolveSplitGrid(); |  | ||||||
|     //    RedBlackSolve(_Matrix,src_o,sol_o);  |  | ||||||
|  |  | ||||||
|     void RedBlackSolution(Matrix &_Matrix, const std::vector<Field> &in, const std::vector<Field> &sol_o, std::vector<Field> &out) |  | ||||||
|     { |  | ||||||
|       GridBase *grid = _Matrix.RedBlackGrid(); |  | ||||||
|       Field tmp(grid); |  | ||||||
|       int nblock = in.size(); |  | ||||||
|       for(int b=0;b<nblock;b++) { |  | ||||||
| 	pickCheckerboard(Even,tmp,in[b]); |  | ||||||
| 	RedBlackSolution(_Matrix,sol_o[b],tmp,out[b]); |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     template<class Guesser> |     template<class Guesser> | ||||||
|     void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out,Guesser &guess)  |     void operator()(Matrix &_Matrix, const std::vector<Field> &in, std::vector<Field> &out,Guesser &guess)  | ||||||
|     { |     { | ||||||
| @@ -175,29 +150,24 @@ namespace Grid { | |||||||
|       //////////////////////////////////////////////// |       //////////////////////////////////////////////// | ||||||
|       // Prepare RedBlack source |       // Prepare RedBlack source | ||||||
|       //////////////////////////////////////////////// |       //////////////////////////////////////////////// | ||||||
|       RedBlackSource(_Matrix,in,src_o); |       for(int b=0;b<nblock;b++){ | ||||||
| 	//      for(int b=0;b<nblock;b++){ | 	RedBlackSource(_Matrix,in[b],tmp,src_o[b]); | ||||||
| 	//	RedBlackSource(_Matrix,in[b],tmp,src_o[b]); |       } | ||||||
| 	//      } |  | ||||||
|        |  | ||||||
|       //////////////////////////////////////////////// |       //////////////////////////////////////////////// | ||||||
|       // Make the guesses |       // Make the guesses | ||||||
|       //////////////////////////////////////////////// |       //////////////////////////////////////////////// | ||||||
|       if ( subGuess ) guess_save.resize(nblock,grid); |       if ( subGuess ) guess_save.resize(nblock,grid); | ||||||
|  |  | ||||||
|        |       for(int b=0;b<nblock;b++){ | ||||||
|       if(useSolnAsInitGuess) { |         if(useSolnAsInitGuess) { | ||||||
|         for(int b=0;b<nblock;b++){ |  | ||||||
|           pickCheckerboard(Odd, sol_o[b], out[b]); |           pickCheckerboard(Odd, sol_o[b], out[b]); | ||||||
|  |         } else { | ||||||
|  |           guess(src_o[b],sol_o[b]);  | ||||||
|         } |         } | ||||||
|       } else { |  | ||||||
|         guess(src_o, sol_o);  |  | ||||||
|       } |  | ||||||
|  |  | ||||||
| 	    if ( subGuess ) {  | 	if ( subGuess ) {  | ||||||
|         for(int b=0;b<nblock;b++){ | 	  guess_save[b] = sol_o[b]; | ||||||
|           guess_save[b] = sol_o[b]; | 	} | ||||||
|         } |  | ||||||
|       } |       } | ||||||
|       ////////////////////////////////////////////////////////////// |       ////////////////////////////////////////////////////////////// | ||||||
|       // Call the block solver |       // Call the block solver | ||||||
|   | |||||||
| @@ -173,10 +173,8 @@ template<class T> using cshiftAllocator = devAllocator<T>; | |||||||
| template<class T> using cshiftAllocator = std::allocator<T>; | template<class T> using cshiftAllocator = std::allocator<T>; | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
| template<class T> using Vector        = std::vector<T,uvmAllocator<T> >;            | template<class T> using Vector     = std::vector<T,uvmAllocator<T> >;            | ||||||
| template<class T> using stencilVector = std::vector<T,alignedAllocator<T> >;            |  | ||||||
| template<class T> using commVector = std::vector<T,devAllocator<T> >; | template<class T> using commVector = std::vector<T,devAllocator<T> >; | ||||||
| template<class T> using deviceVector  = std::vector<T,devAllocator<T> >; |  | ||||||
| template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >; | template<class T> using cshiftVector = std::vector<T,cshiftAllocator<T> >; | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|   | |||||||
| @@ -4,156 +4,106 @@ NAMESPACE_BEGIN(Grid); | |||||||
|  |  | ||||||
| /*Allocation types, saying which pointer cache should be used*/ | /*Allocation types, saying which pointer cache should be used*/ | ||||||
| #define Cpu      (0) | #define Cpu      (0) | ||||||
| #define CpuHuge  (1) | #define CpuSmall (1) | ||||||
| #define CpuSmall (2) | #define Acc      (2) | ||||||
| #define Acc      (3) | #define AccSmall (3) | ||||||
| #define AccHuge  (4) | #define Shared   (4) | ||||||
| #define AccSmall (5) | #define SharedSmall (5) | ||||||
| #define Shared   (6) |  | ||||||
| #define SharedHuge  (7) |  | ||||||
| #define SharedSmall (8) |  | ||||||
| #undef GRID_MM_VERBOSE  |  | ||||||
| uint64_t total_shared; | uint64_t total_shared; | ||||||
| uint64_t total_device; | uint64_t total_device; | ||||||
| uint64_t total_host;; | uint64_t total_host;; | ||||||
| void MemoryManager::PrintBytes(void) | void MemoryManager::PrintBytes(void) | ||||||
| { | { | ||||||
|   std::cout << " MemoryManager : ------------------------------------ "<<std::endl; |   std::cout << " MemoryManager : "<<total_shared<<" shared      bytes "<<std::endl; | ||||||
|   std::cout << " MemoryManager : PrintBytes "<<std::endl; |   std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl; | ||||||
|   std::cout << " MemoryManager : ------------------------------------ "<<std::endl; |   std::cout << " MemoryManager : "<<total_host  <<" cpu         bytes "<<std::endl; | ||||||
|   std::cout << " MemoryManager : "<<(total_shared>>20)<<" shared      Mbytes "<<std::endl; |  | ||||||
|   std::cout << " MemoryManager : "<<(total_device>>20)<<" accelerator Mbytes "<<std::endl; |  | ||||||
|   std::cout << " MemoryManager : "<<(total_host>>20)  <<" cpu         Mbytes "<<std::endl; |  | ||||||
|   uint64_t cacheBytes; |  | ||||||
|   cacheBytes = CacheBytes[Cpu]; |  | ||||||
|   std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu cache Mbytes "<<std::endl; |  | ||||||
|   cacheBytes = CacheBytes[Acc]; |  | ||||||
|   std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc cache Mbytes "<<std::endl; |  | ||||||
|   cacheBytes = CacheBytes[Shared]; |  | ||||||
|   std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" shared cache Mbytes "<<std::endl; |  | ||||||
|    |  | ||||||
| #ifdef GRID_CUDA |  | ||||||
|   cuda_mem(); |  | ||||||
| #endif |  | ||||||
|    |  | ||||||
| } | } | ||||||
|  |  | ||||||
| uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; } |  | ||||||
| uint64_t MemoryManager::HostCacheBytes()   { return CacheBytes[Cpu] + CacheBytes[CpuHuge] + CacheBytes[CpuSmall]; } |  | ||||||
|  |  | ||||||
| ////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////// | ||||||
| // Data tables for recently freed pooiniter caches | // Data tables for recently freed pooiniter caches | ||||||
| ////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////// | ||||||
| MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax]; | MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax]; | ||||||
| int MemoryManager::Victim[MemoryManager::NallocType]; | int MemoryManager::Victim[MemoryManager::NallocType]; | ||||||
| int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 0, 8, 8, 0, 16, 8, 0, 16 }; | int MemoryManager::Ncache[MemoryManager::NallocType] = { 8, 32, 8, 32, 8, 32 }; | ||||||
| uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType]; |  | ||||||
| ////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////// | ||||||
| // Actual allocation and deallocation utils | // Actual allocation and deallocation utils | ||||||
| ////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////// | ||||||
| void *MemoryManager::AcceleratorAllocate(size_t bytes) | void *MemoryManager::AcceleratorAllocate(size_t bytes) | ||||||
| { | { | ||||||
|   total_device+=bytes; |  | ||||||
|   void *ptr = (void *) Lookup(bytes,Acc); |   void *ptr = (void *) Lookup(bytes,Acc); | ||||||
|   if ( ptr == (void *) NULL ) { |   if ( ptr == (void *) NULL ) { | ||||||
|     ptr = (void *) acceleratorAllocDevice(bytes); |     ptr = (void *) acceleratorAllocDevice(bytes); | ||||||
|  |     total_device+=bytes; | ||||||
|   } |   } | ||||||
| #ifdef GRID_MM_VERBOSE |  | ||||||
|   std::cout <<"AcceleratorAllocate "<<std::endl; |  | ||||||
|   PrintBytes(); |  | ||||||
| #endif |  | ||||||
|   return ptr; |   return ptr; | ||||||
| } | } | ||||||
| void  MemoryManager::AcceleratorFree    (void *ptr,size_t bytes) | void  MemoryManager::AcceleratorFree    (void *ptr,size_t bytes) | ||||||
| { | { | ||||||
|   total_device-=bytes; |  | ||||||
|   void *__freeme = Insert(ptr,bytes,Acc); |   void *__freeme = Insert(ptr,bytes,Acc); | ||||||
|   if ( __freeme ) { |   if ( __freeme ) { | ||||||
|     acceleratorFreeDevice(__freeme); |     acceleratorFreeDevice(__freeme); | ||||||
|  |     total_device-=bytes; | ||||||
|  |     //    PrintBytes(); | ||||||
|   } |   } | ||||||
| #ifdef GRID_MM_VERBOSE |  | ||||||
|   std::cout <<"AcceleratorFree "<<std::endl; |  | ||||||
|   PrintBytes(); |  | ||||||
| #endif |  | ||||||
| } | } | ||||||
| void *MemoryManager::SharedAllocate(size_t bytes) | void *MemoryManager::SharedAllocate(size_t bytes) | ||||||
| { | { | ||||||
|   total_shared+=bytes; |  | ||||||
|   void *ptr = (void *) Lookup(bytes,Shared); |   void *ptr = (void *) Lookup(bytes,Shared); | ||||||
|   if ( ptr == (void *) NULL ) { |   if ( ptr == (void *) NULL ) { | ||||||
|     ptr = (void *) acceleratorAllocShared(bytes); |     ptr = (void *) acceleratorAllocShared(bytes); | ||||||
|  |     total_shared+=bytes; | ||||||
|  |     //    std::cout <<"AcceleratorAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl; | ||||||
|  |     //    PrintBytes(); | ||||||
|   } |   } | ||||||
| #ifdef GRID_MM_VERBOSE |  | ||||||
|   std::cout <<"SharedAllocate "<<std::endl; |  | ||||||
|   PrintBytes(); |  | ||||||
| #endif |  | ||||||
|   return ptr; |   return ptr; | ||||||
| } | } | ||||||
| void  MemoryManager::SharedFree    (void *ptr,size_t bytes) | void  MemoryManager::SharedFree    (void *ptr,size_t bytes) | ||||||
| { | { | ||||||
|   total_shared-=bytes; |  | ||||||
|   void *__freeme = Insert(ptr,bytes,Shared); |   void *__freeme = Insert(ptr,bytes,Shared); | ||||||
|   if ( __freeme ) { |   if ( __freeme ) { | ||||||
|     acceleratorFreeShared(__freeme); |     acceleratorFreeShared(__freeme); | ||||||
|  |     total_shared-=bytes; | ||||||
|  |     //    PrintBytes(); | ||||||
|   } |   } | ||||||
| #ifdef GRID_MM_VERBOSE |  | ||||||
|   std::cout <<"SharedFree "<<std::endl; |  | ||||||
|   PrintBytes(); |  | ||||||
| #endif |  | ||||||
| } | } | ||||||
| #ifdef GRID_UVM | #ifdef GRID_UVM | ||||||
| void *MemoryManager::CpuAllocate(size_t bytes) | void *MemoryManager::CpuAllocate(size_t bytes) | ||||||
| { | { | ||||||
|   total_host+=bytes; |  | ||||||
|   void *ptr = (void *) Lookup(bytes,Cpu); |   void *ptr = (void *) Lookup(bytes,Cpu); | ||||||
|   if ( ptr == (void *) NULL ) { |   if ( ptr == (void *) NULL ) { | ||||||
|     ptr = (void *) acceleratorAllocShared(bytes); |     ptr = (void *) acceleratorAllocShared(bytes); | ||||||
|  |     total_host+=bytes; | ||||||
|   } |   } | ||||||
| #ifdef GRID_MM_VERBOSE |  | ||||||
|   std::cout <<"CpuAllocate "<<std::endl; |  | ||||||
|   PrintBytes(); |  | ||||||
| #endif |  | ||||||
|   return ptr; |   return ptr; | ||||||
| } | } | ||||||
| void  MemoryManager::CpuFree    (void *_ptr,size_t bytes) | void  MemoryManager::CpuFree    (void *_ptr,size_t bytes) | ||||||
| { | { | ||||||
|   total_host-=bytes; |  | ||||||
|   NotifyDeletion(_ptr); |   NotifyDeletion(_ptr); | ||||||
|   void *__freeme = Insert(_ptr,bytes,Cpu); |   void *__freeme = Insert(_ptr,bytes,Cpu); | ||||||
|   if ( __freeme ) {  |   if ( __freeme ) {  | ||||||
|     acceleratorFreeShared(__freeme); |     acceleratorFreeShared(__freeme); | ||||||
|  |     total_host-=bytes; | ||||||
|   } |   } | ||||||
| #ifdef GRID_MM_VERBOSE |  | ||||||
|   std::cout <<"CpuFree "<<std::endl; |  | ||||||
|   PrintBytes(); |  | ||||||
| #endif |  | ||||||
| } | } | ||||||
| #else | #else | ||||||
| void *MemoryManager::CpuAllocate(size_t bytes) | void *MemoryManager::CpuAllocate(size_t bytes) | ||||||
| { | { | ||||||
|   total_host+=bytes; |  | ||||||
|   void *ptr = (void *) Lookup(bytes,Cpu); |   void *ptr = (void *) Lookup(bytes,Cpu); | ||||||
|   if ( ptr == (void *) NULL ) { |   if ( ptr == (void *) NULL ) { | ||||||
|     ptr = (void *) acceleratorAllocCpu(bytes); |     ptr = (void *) acceleratorAllocCpu(bytes); | ||||||
|  |     total_host+=bytes; | ||||||
|   } |   } | ||||||
| #ifdef GRID_MM_VERBOSE |  | ||||||
|   std::cout <<"CpuAllocate "<<std::endl; |  | ||||||
|   PrintBytes(); |  | ||||||
| #endif |  | ||||||
|   return ptr; |   return ptr; | ||||||
| } | } | ||||||
| void  MemoryManager::CpuFree    (void *_ptr,size_t bytes) | void  MemoryManager::CpuFree    (void *_ptr,size_t bytes) | ||||||
| { | { | ||||||
|   total_host-=bytes; |  | ||||||
|   NotifyDeletion(_ptr); |   NotifyDeletion(_ptr); | ||||||
|   void *__freeme = Insert(_ptr,bytes,Cpu); |   void *__freeme = Insert(_ptr,bytes,Cpu); | ||||||
|   if ( __freeme ) {  |   if ( __freeme ) {  | ||||||
|     acceleratorFreeCpu(__freeme); |     acceleratorFreeCpu(__freeme); | ||||||
|  |     total_host-=bytes; | ||||||
|   } |   } | ||||||
| #ifdef GRID_MM_VERBOSE |  | ||||||
|   std::cout <<"CpuFree "<<std::endl; |  | ||||||
|   PrintBytes(); |  | ||||||
| #endif |  | ||||||
| } | } | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
| @@ -165,6 +115,7 @@ void MemoryManager::Init(void) | |||||||
|  |  | ||||||
|   char * str; |   char * str; | ||||||
|   int Nc; |   int Nc; | ||||||
|  |   int NcS; | ||||||
|    |    | ||||||
|   str= getenv("GRID_ALLOC_NCACHE_LARGE"); |   str= getenv("GRID_ALLOC_NCACHE_LARGE"); | ||||||
|   if ( str ) { |   if ( str ) { | ||||||
| @@ -176,16 +127,6 @@ void MemoryManager::Init(void) | |||||||
|     } |     } | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   str= getenv("GRID_ALLOC_NCACHE_HUGE"); |  | ||||||
|   if ( str ) { |  | ||||||
|     Nc = atoi(str); |  | ||||||
|     if ( (Nc>=0) && (Nc < NallocCacheMax)) { |  | ||||||
|       Ncache[CpuHuge]=Nc; |  | ||||||
|       Ncache[AccHuge]=Nc; |  | ||||||
|       Ncache[SharedHuge]=Nc; |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   str= getenv("GRID_ALLOC_NCACHE_SMALL"); |   str= getenv("GRID_ALLOC_NCACHE_SMALL"); | ||||||
|   if ( str ) { |   if ( str ) { | ||||||
|     Nc = atoi(str); |     Nc = atoi(str); | ||||||
| @@ -206,9 +147,7 @@ void MemoryManager::InitMessage(void) { | |||||||
|    |    | ||||||
|   std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl; |   std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl; | ||||||
| #ifdef ALLOCATION_CACHE | #ifdef ALLOCATION_CACHE | ||||||
|   std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent host   allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<" HUGE "<<Ncache[CpuHuge]<<std::endl; |   std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl; | ||||||
|   std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent device allocations: SMALL "<<Ncache[AccSmall]<<" LARGE "<<Ncache[Acc]<<" Huge "<<Ncache[AccHuge]<<std::endl; |  | ||||||
|   std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent shared allocations: SMALL "<<Ncache[SharedSmall]<<" LARGE "<<Ncache[Shared]<<" Huge "<<Ncache[SharedHuge]<<std::endl; |  | ||||||
| #endif | #endif | ||||||
|    |    | ||||||
| #ifdef GRID_UVM | #ifdef GRID_UVM | ||||||
| @@ -240,25 +179,21 @@ void MemoryManager::InitMessage(void) { | |||||||
| void *MemoryManager::Insert(void *ptr,size_t bytes,int type)  | void *MemoryManager::Insert(void *ptr,size_t bytes,int type)  | ||||||
| { | { | ||||||
| #ifdef ALLOCATION_CACHE | #ifdef ALLOCATION_CACHE | ||||||
|   int cache; |   bool small = (bytes < GRID_ALLOC_SMALL_LIMIT); | ||||||
|   if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2; |   int cache = type + small; | ||||||
|   else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1; |   return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache]);   | ||||||
|   else                                     cache = type; |  | ||||||
|  |  | ||||||
|   return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);   |  | ||||||
| #else | #else | ||||||
|   return ptr; |   return ptr; | ||||||
| #endif | #endif | ||||||
| } | } | ||||||
|  |  | ||||||
| void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes)  | void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim)  | ||||||
| { | { | ||||||
|  |   assert(ncache>0); | ||||||
| #ifdef GRID_OMP | #ifdef GRID_OMP | ||||||
|   assert(omp_in_parallel()==0); |   assert(omp_in_parallel()==0); | ||||||
| #endif  | #endif  | ||||||
|  |  | ||||||
|   if (ncache == 0) return ptr; |  | ||||||
|  |  | ||||||
|   void * ret = NULL; |   void * ret = NULL; | ||||||
|   int v = -1; |   int v = -1; | ||||||
|  |  | ||||||
| @@ -276,7 +211,6 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries | |||||||
|  |  | ||||||
|   if ( entries[v].valid ) { |   if ( entries[v].valid ) { | ||||||
|     ret = entries[v].address; |     ret = entries[v].address; | ||||||
|     cacheBytes -= entries[v].bytes; |  | ||||||
|     entries[v].valid = 0; |     entries[v].valid = 0; | ||||||
|     entries[v].address = NULL; |     entries[v].address = NULL; | ||||||
|     entries[v].bytes = 0; |     entries[v].bytes = 0; | ||||||
| @@ -285,7 +219,6 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries | |||||||
|   entries[v].address=ptr; |   entries[v].address=ptr; | ||||||
|   entries[v].bytes  =bytes; |   entries[v].bytes  =bytes; | ||||||
|   entries[v].valid  =1; |   entries[v].valid  =1; | ||||||
|   cacheBytes += bytes; |  | ||||||
|  |  | ||||||
|   return ret; |   return ret; | ||||||
| } | } | ||||||
| @@ -293,26 +226,23 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries | |||||||
| void *MemoryManager::Lookup(size_t bytes,int type) | void *MemoryManager::Lookup(size_t bytes,int type) | ||||||
| { | { | ||||||
| #ifdef ALLOCATION_CACHE | #ifdef ALLOCATION_CACHE | ||||||
|   int cache; |   bool small = (bytes < GRID_ALLOC_SMALL_LIMIT); | ||||||
|   if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2; |   int cache = type+small; | ||||||
|   else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1; |   return Lookup(bytes,Entries[cache],Ncache[cache]); | ||||||
|   else                                     cache = type; |  | ||||||
|  |  | ||||||
|   return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]); |  | ||||||
| #else | #else | ||||||
|   return NULL; |   return NULL; | ||||||
| #endif | #endif | ||||||
| } | } | ||||||
|  |  | ||||||
| void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes)  | void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache)  | ||||||
| { | { | ||||||
|  |   assert(ncache>0); | ||||||
| #ifdef GRID_OMP | #ifdef GRID_OMP | ||||||
|   assert(omp_in_parallel()==0); |   assert(omp_in_parallel()==0); | ||||||
| #endif  | #endif  | ||||||
|   for(int e=0;e<ncache;e++){ |   for(int e=0;e<ncache;e++){ | ||||||
|     if ( entries[e].valid && ( entries[e].bytes == bytes ) ) { |     if ( entries[e].valid && ( entries[e].bytes == bytes ) ) { | ||||||
|       entries[e].valid = 0; |       entries[e].valid = 0; | ||||||
|       cacheBytes -= entries[e].bytes; |  | ||||||
|       return entries[e].address; |       return entries[e].address; | ||||||
|     } |     } | ||||||
|   } |   } | ||||||
|   | |||||||
| @@ -34,13 +34,9 @@ NAMESPACE_BEGIN(Grid); | |||||||
|  |  | ||||||
| // Move control to configure.ac and Config.h? | // Move control to configure.ac and Config.h? | ||||||
|  |  | ||||||
|  | #define ALLOCATION_CACHE | ||||||
|  | #define GRID_ALLOC_ALIGN (2*1024*1024) | ||||||
| #define GRID_ALLOC_SMALL_LIMIT (4096) | #define GRID_ALLOC_SMALL_LIMIT (4096) | ||||||
| #define GRID_ALLOC_HUGE_LIMIT  (2147483648) |  | ||||||
|  |  | ||||||
| #define STRINGIFY(x) #x |  | ||||||
| #define TOSTRING(x) STRINGIFY(x) |  | ||||||
| #define FILE_LINE __FILE__ ":" TOSTRING(__LINE__) |  | ||||||
| #define AUDIT(a) MemoryManager::Audit(FILE_LINE) |  | ||||||
|  |  | ||||||
| /*Pinning pages is costly*/ | /*Pinning pages is costly*/ | ||||||
| //////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////// | ||||||
| @@ -71,21 +67,6 @@ enum ViewMode { | |||||||
|   CpuWriteDiscard = 0x10 // same for now |   CpuWriteDiscard = 0x10 // same for now | ||||||
| }; | }; | ||||||
|  |  | ||||||
| struct MemoryStatus { |  | ||||||
|   uint64_t     DeviceBytes; |  | ||||||
|   uint64_t     DeviceLRUBytes; |  | ||||||
|   uint64_t     DeviceMaxBytes; |  | ||||||
|   uint64_t     HostToDeviceBytes; |  | ||||||
|   uint64_t     DeviceToHostBytes; |  | ||||||
|   uint64_t     HostToDeviceXfer; |  | ||||||
|   uint64_t     DeviceToHostXfer; |  | ||||||
|   uint64_t     DeviceEvictions; |  | ||||||
|   uint64_t     DeviceDestroy; |  | ||||||
|   uint64_t     DeviceAllocCacheBytes; |  | ||||||
|   uint64_t     HostAllocCacheBytes; |  | ||||||
| }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
| class MemoryManager { | class MemoryManager { | ||||||
| private: | private: | ||||||
|  |  | ||||||
| @@ -99,23 +80,21 @@ private: | |||||||
|   } AllocationCacheEntry; |   } AllocationCacheEntry; | ||||||
|  |  | ||||||
|   static const int NallocCacheMax=128;  |   static const int NallocCacheMax=128;  | ||||||
|   static const int NallocType=9; |   static const int NallocType=6; | ||||||
|   static AllocationCacheEntry Entries[NallocType][NallocCacheMax]; |   static AllocationCacheEntry Entries[NallocType][NallocCacheMax]; | ||||||
|   static int Victim[NallocType]; |   static int Victim[NallocType]; | ||||||
|   static int Ncache[NallocType]; |   static int Ncache[NallocType]; | ||||||
|   static uint64_t CacheBytes[NallocType]; |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////////// |   ///////////////////////////////////////////////// | ||||||
|   // Free pool |   // Free pool | ||||||
|   ///////////////////////////////////////////////// |   ///////////////////////////////////////////////// | ||||||
|   static void *Insert(void *ptr,size_t bytes,int type) ; |   static void *Insert(void *ptr,size_t bytes,int type) ; | ||||||
|   static void *Lookup(size_t bytes,int type) ; |   static void *Lookup(size_t bytes,int type) ; | ||||||
|   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ; |   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) ; | ||||||
|   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ; |   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ; | ||||||
|  |  | ||||||
|  public: |  | ||||||
|   static void PrintBytes(void); |   static void PrintBytes(void); | ||||||
|   static void Audit(std::string s); |  public: | ||||||
|   static void Init(void); |   static void Init(void); | ||||||
|   static void InitMessage(void); |   static void InitMessage(void); | ||||||
|   static void *AcceleratorAllocate(size_t bytes); |   static void *AcceleratorAllocate(size_t bytes); | ||||||
| @@ -135,28 +114,7 @@ private: | |||||||
|   static uint64_t     DeviceToHostBytes; |   static uint64_t     DeviceToHostBytes; | ||||||
|   static uint64_t     HostToDeviceXfer; |   static uint64_t     HostToDeviceXfer; | ||||||
|   static uint64_t     DeviceToHostXfer; |   static uint64_t     DeviceToHostXfer; | ||||||
|   static uint64_t     DeviceEvictions; |   | ||||||
|   static uint64_t     DeviceDestroy; |  | ||||||
|    |  | ||||||
|   static uint64_t     DeviceCacheBytes(); |  | ||||||
|   static uint64_t     HostCacheBytes(); |  | ||||||
|  |  | ||||||
|   static MemoryStatus GetFootprint(void) { |  | ||||||
|     MemoryStatus stat; |  | ||||||
|     stat.DeviceBytes       = DeviceBytes; |  | ||||||
|     stat.DeviceLRUBytes    = DeviceLRUBytes; |  | ||||||
|     stat.DeviceMaxBytes    = DeviceMaxBytes; |  | ||||||
|     stat.HostToDeviceBytes = HostToDeviceBytes; |  | ||||||
|     stat.DeviceToHostBytes = DeviceToHostBytes; |  | ||||||
|     stat.HostToDeviceXfer  = HostToDeviceXfer; |  | ||||||
|     stat.DeviceToHostXfer  = DeviceToHostXfer; |  | ||||||
|     stat.DeviceEvictions   = DeviceEvictions; |  | ||||||
|     stat.DeviceDestroy     = DeviceDestroy; |  | ||||||
|     stat.DeviceAllocCacheBytes = DeviceCacheBytes(); |  | ||||||
|     stat.HostAllocCacheBytes   = HostCacheBytes(); |  | ||||||
|     return stat; |  | ||||||
|   }; |  | ||||||
|    |  | ||||||
|  private: |  private: | ||||||
| #ifndef GRID_UVM | #ifndef GRID_UVM | ||||||
|   ////////////////////////////////////////////////////////////////////// |   ////////////////////////////////////////////////////////////////////// | ||||||
| @@ -213,8 +171,6 @@ private: | |||||||
|  |  | ||||||
|  public: |  public: | ||||||
|   static void Print(void); |   static void Print(void); | ||||||
|   static void PrintAll(void); |  | ||||||
|   static void PrintState( void* CpuPtr); |  | ||||||
|   static int   isOpen   (void* CpuPtr); |   static int   isOpen   (void* CpuPtr); | ||||||
|   static void  ViewClose(void* CpuPtr,ViewMode mode); |   static void  ViewClose(void* CpuPtr,ViewMode mode); | ||||||
|   static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint); |   static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint); | ||||||
|   | |||||||
| @@ -1,16 +1,10 @@ | |||||||
| #include <Grid/GridCore.h> | #include <Grid/GridCore.h> | ||||||
|  |  | ||||||
| #ifndef GRID_UVM | #ifndef GRID_UVM | ||||||
|  |  | ||||||
| #warning "Using explicit device memory copies" | #warning "Using explicit device memory copies" | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  | #define dprintf(...) | ||||||
| #define MAXLINE 512 |  | ||||||
| static char print_buffer [ MAXLINE ]; |  | ||||||
|  |  | ||||||
| #define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer; |  | ||||||
| #define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer; |  | ||||||
| //#define dprintf(...)  |  | ||||||
|  |  | ||||||
|  |  | ||||||
| //////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////// | ||||||
| // For caching copies of data on device | // For caching copies of data on device | ||||||
| @@ -28,8 +22,6 @@ uint64_t  MemoryManager::HostToDeviceBytes; | |||||||
| uint64_t  MemoryManager::DeviceToHostBytes; | uint64_t  MemoryManager::DeviceToHostBytes; | ||||||
| uint64_t  MemoryManager::HostToDeviceXfer; | uint64_t  MemoryManager::HostToDeviceXfer; | ||||||
| uint64_t  MemoryManager::DeviceToHostXfer; | uint64_t  MemoryManager::DeviceToHostXfer; | ||||||
| uint64_t  MemoryManager::DeviceEvictions; |  | ||||||
| uint64_t  MemoryManager::DeviceDestroy; |  | ||||||
|  |  | ||||||
| //////////////////////////////////// | //////////////////////////////////// | ||||||
| // Priority ordering for unlocked entries | // Priority ordering for unlocked entries | ||||||
| @@ -111,17 +103,15 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache) | |||||||
|   /////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////// | ||||||
|   assert(AccCache.state!=Empty); |   assert(AccCache.state!=Empty); | ||||||
|    |    | ||||||
|   mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);  |   //  dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);  | ||||||
|   assert(AccCache.accLock==0); |   assert(AccCache.accLock==0); | ||||||
|   assert(AccCache.cpuLock==0); |   assert(AccCache.cpuLock==0); | ||||||
|   assert(AccCache.CpuPtr!=(uint64_t)NULL); |   assert(AccCache.CpuPtr!=(uint64_t)NULL); | ||||||
|   if(AccCache.AccPtr) { |   if(AccCache.AccPtr) { | ||||||
|     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes); |     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes); | ||||||
|     DeviceDestroy++; |  | ||||||
|     DeviceBytes   -=AccCache.bytes; |     DeviceBytes   -=AccCache.bytes; | ||||||
|     LRUremove(AccCache); |     LRUremove(AccCache); | ||||||
|     AccCache.AccPtr=(uint64_t) NULL; |     //    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);   | ||||||
|     dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);   |  | ||||||
|   } |   } | ||||||
|   uint64_t CpuPtr = AccCache.CpuPtr; |   uint64_t CpuPtr = AccCache.CpuPtr; | ||||||
|   EntryErase(CpuPtr); |   EntryErase(CpuPtr); | ||||||
| @@ -130,36 +120,26 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache) | |||||||
| void MemoryManager::Evict(AcceleratorViewEntry &AccCache) | void MemoryManager::Evict(AcceleratorViewEntry &AccCache) | ||||||
| { | { | ||||||
|   /////////////////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////////////////// | ||||||
|   // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry |   // Make CPU consistent, remove from Accelerator, remove entry | ||||||
|   // Cannot be acclocked. If allocated must be in LRU pool. |   // Cannot be locked. If allocated must be in LRU pool. | ||||||
|   // |  | ||||||
|   // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock. |  | ||||||
|   //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen |  | ||||||
|   //                          but there is a weakness where CpuLock entries are attempted for erase |  | ||||||
|   //                          Take these OUT LRU queue when CPU locked? |  | ||||||
|   //                          Cannot take out the table as cpuLock data is important. |  | ||||||
|   /////////////////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////////////////// | ||||||
|   assert(AccCache.state!=Empty); |   assert(AccCache.state!=Empty); | ||||||
|    |    | ||||||
|   mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n", |   //  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);  | ||||||
| 	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr, |   assert(AccCache.accLock==0); | ||||||
| 	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock);  |   assert(AccCache.cpuLock==0); | ||||||
|   if (AccCache.accLock!=0) return; |  | ||||||
|   if (AccCache.cpuLock!=0) return; |  | ||||||
|   if(AccCache.state==AccDirty) { |   if(AccCache.state==AccDirty) { | ||||||
|     Flush(AccCache); |     Flush(AccCache); | ||||||
|   } |   } | ||||||
|  |   assert(AccCache.CpuPtr!=(uint64_t)NULL); | ||||||
|   if(AccCache.AccPtr) { |   if(AccCache.AccPtr) { | ||||||
|     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes); |     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes); | ||||||
|     LRUremove(AccCache); |  | ||||||
|     AccCache.AccPtr=(uint64_t)NULL; |  | ||||||
|     AccCache.state=CpuDirty; // CPU primary now |  | ||||||
|     DeviceBytes   -=AccCache.bytes; |     DeviceBytes   -=AccCache.bytes; | ||||||
|     dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);   |     LRUremove(AccCache); | ||||||
|  |     //    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);   | ||||||
|   } |   } | ||||||
|   //  uint64_t CpuPtr = AccCache.CpuPtr; |   uint64_t CpuPtr = AccCache.CpuPtr; | ||||||
|   DeviceEvictions++; |   EntryErase(CpuPtr); | ||||||
|   //  EntryErase(CpuPtr); |  | ||||||
| } | } | ||||||
| void MemoryManager::Flush(AcceleratorViewEntry &AccCache) | void MemoryManager::Flush(AcceleratorViewEntry &AccCache) | ||||||
| { | { | ||||||
| @@ -169,7 +149,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache) | |||||||
|   assert(AccCache.AccPtr!=(uint64_t)NULL); |   assert(AccCache.AccPtr!=(uint64_t)NULL); | ||||||
|   assert(AccCache.CpuPtr!=(uint64_t)NULL); |   assert(AccCache.CpuPtr!=(uint64_t)NULL); | ||||||
|   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes); |   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes); | ||||||
|   mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout); |   //  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout); | ||||||
|   DeviceToHostBytes+=AccCache.bytes; |   DeviceToHostBytes+=AccCache.bytes; | ||||||
|   DeviceToHostXfer++; |   DeviceToHostXfer++; | ||||||
|   AccCache.state=Consistent; |   AccCache.state=Consistent; | ||||||
| @@ -184,7 +164,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache) | |||||||
|     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes); |     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes); | ||||||
|     DeviceBytes+=AccCache.bytes; |     DeviceBytes+=AccCache.bytes; | ||||||
|   } |   } | ||||||
|   mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout); |   //  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout); | ||||||
|   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes); |   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes); | ||||||
|   HostToDeviceBytes+=AccCache.bytes; |   HostToDeviceBytes+=AccCache.bytes; | ||||||
|   HostToDeviceXfer++; |   HostToDeviceXfer++; | ||||||
| @@ -210,7 +190,6 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache) | |||||||
| void MemoryManager::ViewClose(void* Ptr,ViewMode mode) | void MemoryManager::ViewClose(void* Ptr,ViewMode mode) | ||||||
| { | { | ||||||
|   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){ |   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){ | ||||||
|     dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr); |  | ||||||
|     AcceleratorViewClose((uint64_t)Ptr); |     AcceleratorViewClose((uint64_t)Ptr); | ||||||
|   } else if( (mode==CpuRead)||(mode==CpuWrite)){ |   } else if( (mode==CpuRead)||(mode==CpuWrite)){ | ||||||
|     CpuViewClose((uint64_t)Ptr); |     CpuViewClose((uint64_t)Ptr); | ||||||
| @@ -222,7 +201,6 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis | |||||||
| { | { | ||||||
|   uint64_t CpuPtr = (uint64_t)_CpuPtr; |   uint64_t CpuPtr = (uint64_t)_CpuPtr; | ||||||
|   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){ |   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){ | ||||||
|     dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr); |  | ||||||
|     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint); |     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint); | ||||||
|   } else if( (mode==CpuRead)||(mode==CpuWrite)){ |   } else if( (mode==CpuRead)||(mode==CpuWrite)){ | ||||||
|     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint); |     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint); | ||||||
| @@ -233,16 +211,13 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis | |||||||
| } | } | ||||||
| void  MemoryManager::EvictVictims(uint64_t bytes) | void  MemoryManager::EvictVictims(uint64_t bytes) | ||||||
| { | { | ||||||
|   assert(bytes<DeviceMaxBytes); |  | ||||||
|   while(bytes+DeviceLRUBytes > DeviceMaxBytes){ |   while(bytes+DeviceLRUBytes > DeviceMaxBytes){ | ||||||
|     if ( DeviceLRUBytes > 0){ |     if ( DeviceLRUBytes > 0){ | ||||||
|       assert(LRU.size()>0); |       assert(LRU.size()>0); | ||||||
|       uint64_t victim = LRU.back(); // From the LRU |       uint64_t victim = LRU.back(); | ||||||
|       auto AccCacheIterator = EntryLookup(victim); |       auto AccCacheIterator = EntryLookup(victim); | ||||||
|       auto & AccCache = AccCacheIterator->second; |       auto & AccCache = AccCacheIterator->second; | ||||||
|       Evict(AccCache); |       Evict(AccCache); | ||||||
|     } else { |  | ||||||
|       return; |  | ||||||
|     } |     } | ||||||
|   } |   } | ||||||
| } | } | ||||||
| @@ -252,25 +227,18 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod | |||||||
|   // Find if present, otherwise get or force an empty |   // Find if present, otherwise get or force an empty | ||||||
|   //////////////////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////////////////// | ||||||
|   if ( EntryPresent(CpuPtr)==0 ){ |   if ( EntryPresent(CpuPtr)==0 ){ | ||||||
|  |     EvictVictims(bytes); | ||||||
|     EntryCreate(CpuPtr,bytes,mode,hint); |     EntryCreate(CpuPtr,bytes,mode,hint); | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   auto AccCacheIterator = EntryLookup(CpuPtr); |   auto AccCacheIterator = EntryLookup(CpuPtr); | ||||||
|   auto & AccCache = AccCacheIterator->second; |   auto & AccCache = AccCacheIterator->second; | ||||||
|   if (!AccCache.AccPtr) { |    | ||||||
|     EvictVictims(bytes);  |  | ||||||
|   }  |  | ||||||
|   assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard)); |   assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard)); | ||||||
|  |  | ||||||
|   assert(AccCache.cpuLock==0);  // Programming error |   assert(AccCache.cpuLock==0);  // Programming error | ||||||
|  |  | ||||||
|   if(AccCache.state!=Empty) { |   if(AccCache.state!=Empty) { | ||||||
|     dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n", |  | ||||||
| 		    (uint64_t)AccCache.CpuPtr, |  | ||||||
| 		    (uint64_t)CpuPtr, |  | ||||||
| 		    (uint64_t)AccCache.bytes, |  | ||||||
| 	            (uint64_t)bytes, |  | ||||||
| 		    (uint64_t)AccCache.accLock); |  | ||||||
|     assert(AccCache.CpuPtr == CpuPtr); |     assert(AccCache.CpuPtr == CpuPtr); | ||||||
|     assert(AccCache.bytes  ==bytes); |     assert(AccCache.bytes  ==bytes); | ||||||
|   } |   } | ||||||
| @@ -305,7 +273,6 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod | |||||||
|       AccCache.state  = Consistent; // Empty + AccRead => Consistent |       AccCache.state  = Consistent; // Empty + AccRead => Consistent | ||||||
|     } |     } | ||||||
|     AccCache.accLock= 1; |     AccCache.accLock= 1; | ||||||
|     dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock); |  | ||||||
|   } else if(AccCache.state==CpuDirty ){ |   } else if(AccCache.state==CpuDirty ){ | ||||||
|     if(mode==AcceleratorWriteDiscard) { |     if(mode==AcceleratorWriteDiscard) { | ||||||
|       CpuDiscard(AccCache); |       CpuDiscard(AccCache); | ||||||
| @@ -318,30 +285,28 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod | |||||||
|       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent |       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent | ||||||
|     } |     } | ||||||
|     AccCache.accLock++; |     AccCache.accLock++; | ||||||
|     dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock); |     //    printf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock); | ||||||
|   } else if(AccCache.state==Consistent) { |   } else if(AccCache.state==Consistent) { | ||||||
|     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard)) |     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard)) | ||||||
|       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty |       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty | ||||||
|     else |     else | ||||||
|       AccCache.state  = Consistent; // Consistent + AccRead => Consistent |       AccCache.state  = Consistent; // Consistent + AccRead => Consistent | ||||||
|     AccCache.accLock++; |     AccCache.accLock++; | ||||||
|     dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock); |     //    printf("Consistent entry into device accLock %d\n",AccCache.accLock); | ||||||
|   } else if(AccCache.state==AccDirty) { |   } else if(AccCache.state==AccDirty) { | ||||||
|     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard)) |     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard)) | ||||||
|       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty |       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty | ||||||
|     else |     else | ||||||
|       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty |       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty | ||||||
|     AccCache.accLock++; |     AccCache.accLock++; | ||||||
|     dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock); |     //    printf("AccDirty entry into device accLock %d\n",AccCache.accLock); | ||||||
|   } else { |   } else { | ||||||
|     assert(0); |     assert(0); | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   assert(AccCache.accLock>0); |   // If view is opened on device remove from LRU | ||||||
|   // If view is opened on device must remove from LRU |  | ||||||
|   if(AccCache.LRU_valid==1){ |   if(AccCache.LRU_valid==1){ | ||||||
|     // must possibly remove from LRU as now locked on GPU |     // must possibly remove from LRU as now locked on GPU | ||||||
|     dprintf("AccCache entry removed from LRU \n"); |  | ||||||
|     LRUremove(AccCache); |     LRUremove(AccCache); | ||||||
|   } |   } | ||||||
|  |  | ||||||
| @@ -362,12 +327,10 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr) | |||||||
|   assert(AccCache.accLock>0); |   assert(AccCache.accLock>0); | ||||||
|  |  | ||||||
|   AccCache.accLock--; |   AccCache.accLock--; | ||||||
|  |  | ||||||
|   // Move to LRU queue if not locked and close on device |   // Move to LRU queue if not locked and close on device | ||||||
|   if(AccCache.accLock==0) { |   if(AccCache.accLock==0) { | ||||||
|     dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock); |  | ||||||
|     LRUinsert(AccCache); |     LRUinsert(AccCache); | ||||||
|   } else { |  | ||||||
|     dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock); |  | ||||||
|   } |   } | ||||||
| } | } | ||||||
| void MemoryManager::CpuViewClose(uint64_t CpuPtr) | void MemoryManager::CpuViewClose(uint64_t CpuPtr) | ||||||
| @@ -398,17 +361,13 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V | |||||||
|   // Find if present, otherwise get or force an empty |   // Find if present, otherwise get or force an empty | ||||||
|   //////////////////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////////////////// | ||||||
|   if ( EntryPresent(CpuPtr)==0 ){ |   if ( EntryPresent(CpuPtr)==0 ){ | ||||||
|  |     EvictVictims(bytes); | ||||||
|     EntryCreate(CpuPtr,bytes,mode,transient); |     EntryCreate(CpuPtr,bytes,mode,transient); | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   auto AccCacheIterator = EntryLookup(CpuPtr); |   auto AccCacheIterator = EntryLookup(CpuPtr); | ||||||
|   auto & AccCache = AccCacheIterator->second; |   auto & AccCache = AccCacheIterator->second; | ||||||
|  |    | ||||||
|   // CPU doesn't need to free space |  | ||||||
|   //  if (!AccCache.AccPtr) { |  | ||||||
|   //    EvictVictims(bytes); |  | ||||||
|   //  } |  | ||||||
|  |  | ||||||
|   assert((mode==CpuRead)||(mode==CpuWrite)); |   assert((mode==CpuRead)||(mode==CpuWrite)); | ||||||
|   assert(AccCache.accLock==0);  // Programming error |   assert(AccCache.accLock==0);  // Programming error | ||||||
|  |  | ||||||
| @@ -460,29 +419,20 @@ void  MemoryManager::NotifyDeletion(void *_ptr) | |||||||
| } | } | ||||||
| void  MemoryManager::Print(void) | void  MemoryManager::Print(void) | ||||||
| { | { | ||||||
|   PrintBytes(); |   std::cout << GridLogDebug << "--------------------------------------------" << std::endl; | ||||||
|   std::cout << GridLogMessage << "--------------------------------------------" << std::endl; |   std::cout << GridLogDebug << "Memory Manager                             " << std::endl; | ||||||
|   std::cout << GridLogMessage << "Memory Manager                             " << std::endl; |   std::cout << GridLogDebug << "--------------------------------------------" << std::endl; | ||||||
|   std::cout << GridLogMessage << "--------------------------------------------" << std::endl; |   std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl; | ||||||
|   std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl; |   std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl; | ||||||
|   std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl; |   std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl; | ||||||
|   std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl; |   std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl; | ||||||
|   std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl; |   std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl; | ||||||
|   std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl; |   std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl; | ||||||
|   std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl; |   std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl; | ||||||
|   std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl; |   std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl; | ||||||
|   std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl; |   std::cout << GridLogDebug << "--------------------------------------------" << std::endl; | ||||||
|   std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl; |   std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl; | ||||||
|   std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl; |   std::cout << GridLogDebug << "--------------------------------------------" << std::endl; | ||||||
|   std::cout << GridLogMessage << "--------------------------------------------" << std::endl; |  | ||||||
| } |  | ||||||
| void  MemoryManager::PrintAll(void) |  | ||||||
| { |  | ||||||
|   Print(); |  | ||||||
|   std::cout << GridLogMessage << std::endl; |  | ||||||
|   std::cout << GridLogMessage << "--------------------------------------------" << std::endl; |  | ||||||
|   std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl; |  | ||||||
|   std::cout << GridLogMessage << "--------------------------------------------" << std::endl; |  | ||||||
|   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){ |   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){ | ||||||
|     auto &AccCache = it->second; |     auto &AccCache = it->second; | ||||||
|      |      | ||||||
| @@ -492,13 +442,13 @@ void  MemoryManager::PrintAll(void) | |||||||
|     if ( AccCache.state==AccDirty ) str = std::string("AccDirty"); |     if ( AccCache.state==AccDirty ) str = std::string("AccDirty"); | ||||||
|     if ( AccCache.state==Consistent)str = std::string("Consistent"); |     if ( AccCache.state==Consistent)str = std::string("Consistent"); | ||||||
|  |  | ||||||
|     std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec |     std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec | ||||||
| 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str | 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str | ||||||
| 	      << "\t" << AccCache.cpuLock | 	      << "\t" << AccCache.cpuLock | ||||||
| 	      << "\t" << AccCache.accLock | 	      << "\t" << AccCache.accLock | ||||||
| 	      << "\t" << AccCache.LRU_valid<<std::endl; | 	      << "\t" << AccCache.LRU_valid<<std::endl; | ||||||
|   } |   } | ||||||
|   std::cout << GridLogMessage << "--------------------------------------------" << std::endl; |   std::cout << GridLogDebug << "--------------------------------------------" << std::endl; | ||||||
|  |  | ||||||
| }; | }; | ||||||
| int   MemoryManager::isOpen   (void* _CpuPtr)  | int   MemoryManager::isOpen   (void* _CpuPtr)  | ||||||
| @@ -512,89 +462,6 @@ int   MemoryManager::isOpen   (void* _CpuPtr) | |||||||
|     return 0; |     return 0; | ||||||
|   } |   } | ||||||
| } | } | ||||||
| void MemoryManager::Audit(std::string s) |  | ||||||
| { |  | ||||||
|   uint64_t CpuBytes=0; |  | ||||||
|   uint64_t AccBytes=0; |  | ||||||
|   uint64_t LruBytes1=0; |  | ||||||
|   uint64_t LruBytes2=0; |  | ||||||
|   uint64_t LruCnt=0; |  | ||||||
|    |  | ||||||
|   std::cout << " Memory Manager::Audit() from "<<s<<std::endl; |  | ||||||
|   for(auto it=LRU.begin();it!=LRU.end();it++){ |  | ||||||
|     uint64_t cpuPtr = *it; |  | ||||||
|     assert(EntryPresent(cpuPtr)); |  | ||||||
|     auto AccCacheIterator = EntryLookup(cpuPtr); |  | ||||||
|     auto & AccCache = AccCacheIterator->second; |  | ||||||
|     LruBytes2+=AccCache.bytes; |  | ||||||
|     assert(AccCache.LRU_valid==1); |  | ||||||
|     assert(AccCache.LRU_entry==it); |  | ||||||
|   } |  | ||||||
|   std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl; |  | ||||||
|  |  | ||||||
|   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){ |  | ||||||
|     auto &AccCache = it->second; |  | ||||||
|      |  | ||||||
|     std::string str; |  | ||||||
|     if ( AccCache.state==Empty    ) str = std::string("Empty"); |  | ||||||
|     if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty"); |  | ||||||
|     if ( AccCache.state==AccDirty ) str = std::string("AccDirty"); |  | ||||||
|     if ( AccCache.state==Consistent)str = std::string("Consistent"); |  | ||||||
|  |  | ||||||
|     CpuBytes+=AccCache.bytes; |  | ||||||
|     if( AccCache.AccPtr )    AccBytes+=AccCache.bytes; |  | ||||||
|     if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes; |  | ||||||
|     if( AccCache.LRU_valid ) LruCnt++; |  | ||||||
|      |  | ||||||
|     if ( AccCache.cpuLock || AccCache.accLock ) { |  | ||||||
|       assert(AccCache.LRU_valid==0); |  | ||||||
|  |  | ||||||
|       std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec |  | ||||||
| 		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str |  | ||||||
| 		<< "\t cpuLock  " << AccCache.cpuLock |  | ||||||
| 		<< "\t accLock  " << AccCache.accLock |  | ||||||
| 		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl; |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|     assert( AccCache.cpuLock== 0 ) ; |  | ||||||
|     assert( AccCache.accLock== 0 ) ; |  | ||||||
|   } |  | ||||||
|   std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl; |  | ||||||
|   assert(LruBytes1==LruBytes2); |  | ||||||
|   assert(LruBytes1==DeviceLRUBytes); |  | ||||||
|   std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl; |  | ||||||
|   assert(AccBytes==DeviceBytes); |  | ||||||
|   std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl; |  | ||||||
|   assert(LruCnt == LRU.size()); |  | ||||||
|   std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl; |  | ||||||
|  |  | ||||||
| } |  | ||||||
|  |  | ||||||
| void MemoryManager::PrintState(void* _CpuPtr) |  | ||||||
| { |  | ||||||
|   uint64_t CpuPtr = (uint64_t)_CpuPtr; |  | ||||||
|  |  | ||||||
|   if ( EntryPresent(CpuPtr) ){ |  | ||||||
|     auto AccCacheIterator = EntryLookup(CpuPtr); |  | ||||||
|     auto & AccCache = AccCacheIterator->second; |  | ||||||
|     std::string str; |  | ||||||
|     if ( AccCache.state==Empty    ) str = std::string("Empty"); |  | ||||||
|     if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty"); |  | ||||||
|     if ( AccCache.state==AccDirty ) str = std::string("AccDirty"); |  | ||||||
|     if ( AccCache.state==Consistent)str = std::string("Consistent"); |  | ||||||
|     if ( AccCache.state==EvictNext) str = std::string("EvictNext"); |  | ||||||
|  |  | ||||||
|     std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl; |  | ||||||
|     std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec |  | ||||||
|     << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str |  | ||||||
|     << "\t" << AccCache.cpuLock |  | ||||||
|     << "\t" << AccCache.accLock |  | ||||||
|     << "\t" << AccCache.LRU_valid<<std::endl; |  | ||||||
|  |  | ||||||
|   } else { |  | ||||||
|     std::cout << GridLogMessage << "No Entry in AccCache table." << std::endl;  |  | ||||||
|   } |  | ||||||
| } |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|  |  | ||||||
|   | |||||||
| @@ -1,6 +1,7 @@ | |||||||
| #include <Grid/GridCore.h> | #include <Grid/GridCore.h> | ||||||
| #ifdef GRID_UVM | #ifdef GRID_UVM | ||||||
|  |  | ||||||
|  | #warning "Grid is assuming unified virtual memory address space" | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
| ///////////////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////////////// | ||||||
| // View management is 1:1 address space mapping | // View management is 1:1 address space mapping | ||||||
| @@ -12,19 +13,11 @@ uint64_t  MemoryManager::HostToDeviceBytes; | |||||||
| uint64_t  MemoryManager::DeviceToHostBytes; | uint64_t  MemoryManager::DeviceToHostBytes; | ||||||
| uint64_t  MemoryManager::HostToDeviceXfer; | uint64_t  MemoryManager::HostToDeviceXfer; | ||||||
| uint64_t  MemoryManager::DeviceToHostXfer; | uint64_t  MemoryManager::DeviceToHostXfer; | ||||||
| uint64_t  MemoryManager::DeviceEvictions; |  | ||||||
| uint64_t  MemoryManager::DeviceDestroy; |  | ||||||
|  |  | ||||||
| void  MemoryManager::Audit(std::string s){}; |  | ||||||
| void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){}; | void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){}; | ||||||
| void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; }; | void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; }; | ||||||
| int   MemoryManager::isOpen   (void* CpuPtr) { return 0;} | int   MemoryManager::isOpen   (void* CpuPtr) { return 0;} | ||||||
| void  MemoryManager::PrintState(void* CpuPtr) |  | ||||||
| { |  | ||||||
| std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl; |  | ||||||
| }; |  | ||||||
| void  MemoryManager::Print(void){}; | void  MemoryManager::Print(void){}; | ||||||
| void  MemoryManager::PrintAll(void){}; |  | ||||||
| void  MemoryManager::NotifyDeletion(void *ptr){}; | void  MemoryManager::NotifyDeletion(void *ptr){}; | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|   | |||||||
| @@ -36,7 +36,7 @@ static const int CbBlack=1; | |||||||
| static const int Even   =CbRed; | static const int Even   =CbRed; | ||||||
| static const int Odd    =CbBlack; | static const int Odd    =CbBlack; | ||||||
|  |  | ||||||
| accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex,const Coordinate &rdim,const Coordinate &chk_dim_msk) | accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex, Coordinate &rdim, Coordinate &chk_dim_msk) | ||||||
| { | { | ||||||
|   int nd=rdim.size(); |   int nd=rdim.size(); | ||||||
|   Coordinate coor(nd); |   Coordinate coor(nd); | ||||||
|   | |||||||
| @@ -33,8 +33,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk> | |||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| bool Stencil_force_mpi = true; |  | ||||||
|  |  | ||||||
| /////////////////////////////////////////////////////////////// | /////////////////////////////////////////////////////////////// | ||||||
| // Info that is setup once and indept of cartesian layout | // Info that is setup once and indept of cartesian layout | ||||||
| /////////////////////////////////////////////////////////////// | /////////////////////////////////////////////////////////////// | ||||||
|   | |||||||
| @@ -1,3 +1,4 @@ | |||||||
|  |  | ||||||
| /************************************************************************************* | /************************************************************************************* | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid  |     Grid physics library, www.github.com/paboyle/Grid  | ||||||
| @@ -35,8 +36,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk> | |||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| extern bool Stencil_force_mpi ; |  | ||||||
|  |  | ||||||
| class CartesianCommunicator : public SharedMemory { | class CartesianCommunicator : public SharedMemory { | ||||||
|  |  | ||||||
| public:     | public:     | ||||||
| @@ -53,11 +52,10 @@ public: | |||||||
|   // Communicator should know nothing of the physics grid, only processor grid. |   // Communicator should know nothing of the physics grid, only processor grid. | ||||||
|   //////////////////////////////////////////// |   //////////////////////////////////////////// | ||||||
|   int              _Nprocessors;     // How many in all |   int              _Nprocessors;     // How many in all | ||||||
|   int              _processor;       // linear processor rank |  | ||||||
|   unsigned long    _ndimension; |  | ||||||
|   Coordinate _shm_processors;  // Which dimensions get relayed out over processors lanes. |  | ||||||
|   Coordinate _processors;      // Which dimensions get relayed out over processors lanes. |   Coordinate _processors;      // Which dimensions get relayed out over processors lanes. | ||||||
|  |   int              _processor;       // linear processor rank | ||||||
|   Coordinate _processor_coor;  // linear processor coordinate |   Coordinate _processor_coor;  // linear processor coordinate | ||||||
|  |   unsigned long    _ndimension; | ||||||
|   static Grid_MPI_Comm      communicator_world; |   static Grid_MPI_Comm      communicator_world; | ||||||
|   Grid_MPI_Comm             communicator; |   Grid_MPI_Comm             communicator; | ||||||
|   std::vector<Grid_MPI_Comm> communicator_halo; |   std::vector<Grid_MPI_Comm> communicator_halo; | ||||||
| @@ -98,22 +96,18 @@ public: | |||||||
|   int                      BossRank(void)          ; |   int                      BossRank(void)          ; | ||||||
|   int                      ThisRank(void)          ; |   int                      ThisRank(void)          ; | ||||||
|   const Coordinate & ThisProcessorCoor(void) ; |   const Coordinate & ThisProcessorCoor(void) ; | ||||||
|   const Coordinate & ShmGrid(void)  { return _shm_processors; }  ; |  | ||||||
|   const Coordinate & ProcessorGrid(void)     ; |   const Coordinate & ProcessorGrid(void)     ; | ||||||
|   int                ProcessorCount(void)    ; |   int                      ProcessorCount(void)    ; | ||||||
|  |  | ||||||
|   //////////////////////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////////////////////// | ||||||
|   // very VERY rarely (Log, serial RNG) we need world without a grid |   // very VERY rarely (Log, serial RNG) we need world without a grid | ||||||
|   //////////////////////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////////////////////// | ||||||
|   static int  RankWorld(void) ; |   static int  RankWorld(void) ; | ||||||
|   static void BroadcastWorld(int root,void* data, int bytes); |   static void BroadcastWorld(int root,void* data, int bytes); | ||||||
|   static void BarrierWorld(void); |  | ||||||
|    |    | ||||||
|   //////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////// | ||||||
|   // Reduction |   // Reduction | ||||||
|   //////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////// | ||||||
|   void GlobalMax(RealD &); |  | ||||||
|   void GlobalMax(RealF &); |  | ||||||
|   void GlobalSum(RealF &); |   void GlobalSum(RealF &); | ||||||
|   void GlobalSumVector(RealF *,int N); |   void GlobalSumVector(RealF *,int N); | ||||||
|   void GlobalSum(RealD &); |   void GlobalSum(RealD &); | ||||||
| @@ -131,7 +125,7 @@ public: | |||||||
|   template<class obj> void GlobalSum(obj &o){ |   template<class obj> void GlobalSum(obj &o){ | ||||||
|     typedef typename obj::scalar_type scalar_type; |     typedef typename obj::scalar_type scalar_type; | ||||||
|     int words = sizeof(obj)/sizeof(scalar_type); |     int words = sizeof(obj)/sizeof(scalar_type); | ||||||
|     scalar_type * ptr = (scalar_type *)& o; // Safe alias  |     scalar_type * ptr = (scalar_type *)& o; | ||||||
|     GlobalSumVector(ptr,words); |     GlobalSumVector(ptr,words); | ||||||
|   } |   } | ||||||
|    |    | ||||||
| @@ -145,17 +139,17 @@ public: | |||||||
| 		      int bytes); | 		      int bytes); | ||||||
|    |    | ||||||
|   double StencilSendToRecvFrom(void *xmit, |   double StencilSendToRecvFrom(void *xmit, | ||||||
| 			       int xmit_to_rank,int do_xmit, | 			       int xmit_to_rank, | ||||||
| 			       void *recv, | 			       void *recv, | ||||||
| 			       int recv_from_rank,int do_recv, | 			       int recv_from_rank, | ||||||
| 			       int bytes,int dir); | 			       int bytes,int dir); | ||||||
|  |  | ||||||
|   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list, |   double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list, | ||||||
| 				    void *xmit, | 				    void *xmit, | ||||||
| 				    int xmit_to_rank,int do_xmit, | 				    int xmit_to_rank, | ||||||
| 				    void *recv, | 				    void *recv, | ||||||
| 				    int recv_from_rank,int do_recv, | 				    int recv_from_rank, | ||||||
| 				    int xbytes,int rbytes,int dir); | 				    int bytes,int dir); | ||||||
|    |    | ||||||
|    |    | ||||||
|   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i); |   void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i); | ||||||
|   | |||||||
| @@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) | |||||||
|   // Remap using the shared memory optimising routine |   // Remap using the shared memory optimising routine | ||||||
|   // The remap creates a comm which must be freed |   // The remap creates a comm which must be freed | ||||||
|   //////////////////////////////////////////////////// |   //////////////////////////////////////////////////// | ||||||
|   GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm,_shm_processors); |   GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm); | ||||||
|   InitFromMPICommunicator(processors,optimal_comm); |   InitFromMPICommunicator(processors,optimal_comm); | ||||||
|   SetCommunicator(optimal_comm); |   SetCommunicator(optimal_comm); | ||||||
|   /////////////////////////////////////////////////// |   /////////////////////////////////////////////////// | ||||||
| @@ -124,13 +124,12 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const | |||||||
|   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension); |   int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension); | ||||||
|   Coordinate parent_processor_coor(_ndimension,0); |   Coordinate parent_processor_coor(_ndimension,0); | ||||||
|   Coordinate parent_processors    (_ndimension,1); |   Coordinate parent_processors    (_ndimension,1); | ||||||
|   Coordinate shm_processors       (_ndimension,1); |  | ||||||
|   // Can make 5d grid from 4d etc... |   // Can make 5d grid from 4d etc... | ||||||
|   int pad = _ndimension-parent_ndimension; |   int pad = _ndimension-parent_ndimension; | ||||||
|   for(int d=0;d<parent_ndimension;d++){ |   for(int d=0;d<parent_ndimension;d++){ | ||||||
|     parent_processor_coor[pad+d]=parent._processor_coor[d]; |     parent_processor_coor[pad+d]=parent._processor_coor[d]; | ||||||
|     parent_processors    [pad+d]=parent._processors[d]; |     parent_processors    [pad+d]=parent._processors[d]; | ||||||
|     shm_processors       [pad+d]=parent._shm_processors[d]; |  | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   ////////////////////////////////////////////////////////////////////////////////////////////////////// |   ////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| @@ -155,7 +154,6 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const | |||||||
|     ccoor[d] = parent_processor_coor[d] % processors[d]; |     ccoor[d] = parent_processor_coor[d] % processors[d]; | ||||||
|     scoor[d] = parent_processor_coor[d] / processors[d]; |     scoor[d] = parent_processor_coor[d] / processors[d]; | ||||||
|     ssize[d] = parent_processors[d]     / processors[d]; |     ssize[d] = parent_processors[d]     / processors[d]; | ||||||
|     if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting. |  | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms |   // rank within subcomm ; srank is rank of subcomm within blocks of subcomms | ||||||
| @@ -277,16 +275,6 @@ void CartesianCommunicator::GlobalXOR(uint64_t &u){ | |||||||
|   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator); |   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator); | ||||||
|   assert(ierr==0); |   assert(ierr==0); | ||||||
| } | } | ||||||
| void CartesianCommunicator::GlobalMax(float &f) |  | ||||||
| { |  | ||||||
|   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_MAX,communicator); |  | ||||||
|   assert(ierr==0); |  | ||||||
| } |  | ||||||
| void CartesianCommunicator::GlobalMax(double &d) |  | ||||||
| { |  | ||||||
|   int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_MAX,communicator); |  | ||||||
|   assert(ierr==0); |  | ||||||
| } |  | ||||||
| void CartesianCommunicator::GlobalSum(float &f){ | void CartesianCommunicator::GlobalSum(float &f){ | ||||||
|   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator); |   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator); | ||||||
|   assert(ierr==0); |   assert(ierr==0); | ||||||
| @@ -337,24 +325,23 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit, | |||||||
| } | } | ||||||
| // Basic Halo comms primitive | // Basic Halo comms primitive | ||||||
| double CartesianCommunicator::StencilSendToRecvFrom( void *xmit, | double CartesianCommunicator::StencilSendToRecvFrom( void *xmit, | ||||||
| 						     int dest, int dox, | 						     int dest, | ||||||
| 						     void *recv, | 						     void *recv, | ||||||
| 						     int from, int dor, | 						     int from, | ||||||
| 						     int bytes,int dir) | 						     int bytes,int dir) | ||||||
| { | { | ||||||
|   std::vector<CommsRequest_t> list; |   std::vector<CommsRequest_t> list; | ||||||
|   double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir); |   double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir); | ||||||
|   StencilSendToRecvFromComplete(list,dir); |   StencilSendToRecvFromComplete(list,dir); | ||||||
|   return offbytes; |   return offbytes; | ||||||
| } | } | ||||||
|  |  | ||||||
| #undef NVLINK_GET // Define to use get instead of put DMA |  | ||||||
| double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list, | double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list, | ||||||
| 							 void *xmit, | 							 void *xmit, | ||||||
| 							 int dest,int dox, | 							 int dest, | ||||||
| 							 void *recv, | 							 void *recv, | ||||||
| 							 int from,int dor, | 							 int from, | ||||||
| 							 int xbytes,int rbytes,int dir) | 							 int bytes,int dir) | ||||||
| { | { | ||||||
|   int ncomm  =communicator_halo.size(); |   int ncomm  =communicator_halo.size(); | ||||||
|   int commdir=dir%ncomm; |   int commdir=dir%ncomm; | ||||||
| @@ -373,37 +360,24 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques | |||||||
|   double off_node_bytes=0.0; |   double off_node_bytes=0.0; | ||||||
|   int tag; |   int tag; | ||||||
|  |  | ||||||
|   if ( dor ) { |   if ( gfrom ==MPI_UNDEFINED) { | ||||||
|     if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) { |     tag= dir+from*32; | ||||||
|       tag= dir+from*32; |     ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq); | ||||||
|       ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq); |     assert(ierr==0); | ||||||
|       assert(ierr==0); |     list.push_back(rrq); | ||||||
|       list.push_back(rrq); |     off_node_bytes+=bytes; | ||||||
|       off_node_bytes+=rbytes; |  | ||||||
|     } |  | ||||||
| #ifdef NVLINK_GET |  | ||||||
|       void *shm = (void *) this->ShmBufferTranslate(from,xmit); |  | ||||||
|       assert(shm!=NULL); |  | ||||||
|       acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes); |  | ||||||
| #endif |  | ||||||
|   } |   } | ||||||
|    |  | ||||||
|   if (dox) { |   if ( gdest == MPI_UNDEFINED ) { | ||||||
|     //  rcrc = crc32(rcrc,(unsigned char *)recv,bytes); |     tag= dir+_processor*32; | ||||||
|     if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) { |     ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq); | ||||||
|       tag= dir+_processor*32; |     assert(ierr==0); | ||||||
|       ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq); |     list.push_back(xrq); | ||||||
|       assert(ierr==0); |     off_node_bytes+=bytes; | ||||||
|       list.push_back(xrq); |   } | ||||||
|       off_node_bytes+=xbytes; |  | ||||||
|     } else { |   if ( CommunicatorPolicy == CommunicatorPolicySequential ) { | ||||||
| #ifndef NVLINK_GET |     this->StencilSendToRecvFromComplete(list,dir); | ||||||
|       void *shm = (void *) this->ShmBufferTranslate(dest,recv); |  | ||||||
|       assert(shm!=NULL); |  | ||||||
|       acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes); |  | ||||||
| #endif |  | ||||||
|        |  | ||||||
|     } |  | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   return off_node_bytes; |   return off_node_bytes; | ||||||
| @@ -412,8 +386,6 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque | |||||||
| { | { | ||||||
|   int nreq=list.size(); |   int nreq=list.size(); | ||||||
|  |  | ||||||
|   acceleratorCopySynchronise(); |  | ||||||
|  |  | ||||||
|   if (nreq==0) return; |   if (nreq==0) return; | ||||||
|  |  | ||||||
|   std::vector<MPI_Status> status(nreq); |   std::vector<MPI_Status> status(nreq); | ||||||
| @@ -447,10 +419,6 @@ int CartesianCommunicator::RankWorld(void){ | |||||||
|   MPI_Comm_rank(communicator_world,&r); |   MPI_Comm_rank(communicator_world,&r); | ||||||
|   return r; |   return r; | ||||||
| } | } | ||||||
| void CartesianCommunicator::BarrierWorld(void){ |  | ||||||
|   int ierr = MPI_Barrier(communicator_world); |  | ||||||
|   assert(ierr==0); |  | ||||||
| } |  | ||||||
| void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) | void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) | ||||||
| { | { | ||||||
|   int ierr= MPI_Bcast(data, |   int ierr= MPI_Bcast(data, | ||||||
|   | |||||||
| @@ -45,14 +45,12 @@ void CartesianCommunicator::Init(int *argc, char *** arv) | |||||||
| CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)  | CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)  | ||||||
|   : CartesianCommunicator(processors)  |   : CartesianCommunicator(processors)  | ||||||
| { | { | ||||||
|   _shm_processors = Coordinate(processors.size(),1); |  | ||||||
|   srank=0; |   srank=0; | ||||||
|   SetCommunicator(communicator_world); |   SetCommunicator(communicator_world); | ||||||
| } | } | ||||||
|  |  | ||||||
| CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) | CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) | ||||||
| { | { | ||||||
|   _shm_processors = Coordinate(processors.size(),1); |  | ||||||
|   _processors = processors; |   _processors = processors; | ||||||
|   _ndimension = processors.size();  assert(_ndimension>=1); |   _ndimension = processors.size();  assert(_ndimension>=1); | ||||||
|   _processor_coor.resize(_ndimension); |   _processor_coor.resize(_ndimension); | ||||||
| @@ -69,8 +67,6 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors) | |||||||
|  |  | ||||||
| CartesianCommunicator::~CartesianCommunicator(){} | CartesianCommunicator::~CartesianCommunicator(){} | ||||||
|  |  | ||||||
| void CartesianCommunicator::GlobalMax(float &){} |  | ||||||
| void CartesianCommunicator::GlobalMax(double &){} |  | ||||||
| void CartesianCommunicator::GlobalSum(float &){} | void CartesianCommunicator::GlobalSum(float &){} | ||||||
| void CartesianCommunicator::GlobalSumVector(float *,int N){} | void CartesianCommunicator::GlobalSumVector(float *,int N){} | ||||||
| void CartesianCommunicator::GlobalSum(double &){} | void CartesianCommunicator::GlobalSum(double &){} | ||||||
| @@ -104,7 +100,6 @@ int  CartesianCommunicator::RankWorld(void){return 0;} | |||||||
| void CartesianCommunicator::Barrier(void){} | void CartesianCommunicator::Barrier(void){} | ||||||
| void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {} | void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {} | ||||||
| void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { } | void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { } | ||||||
| void CartesianCommunicator::BarrierWorld(void) { } |  | ||||||
| int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;} | int  CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) {  return 0;} | ||||||
| void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; } | void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){  coor = _processor_coor; } | ||||||
| void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest) | void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest) | ||||||
| @@ -114,21 +109,21 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest | |||||||
| } | } | ||||||
|  |  | ||||||
| double CartesianCommunicator::StencilSendToRecvFrom( void *xmit, | double CartesianCommunicator::StencilSendToRecvFrom( void *xmit, | ||||||
| 						     int xmit_to_rank,int dox, | 						     int xmit_to_rank, | ||||||
| 						     void *recv, | 						     void *recv, | ||||||
| 						     int recv_from_rank,int dor, | 						     int recv_from_rank, | ||||||
| 						     int bytes, int dir) | 						     int bytes, int dir) | ||||||
| { | { | ||||||
|   return 2.0*bytes; |   return 2.0*bytes; | ||||||
| } | } | ||||||
| double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list, | double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list, | ||||||
| 							 void *xmit, | 							 void *xmit, | ||||||
| 							 int xmit_to_rank,int dox, | 							 int xmit_to_rank, | ||||||
| 							 void *recv, | 							 void *recv, | ||||||
| 							 int recv_from_rank,int dor, | 							 int recv_from_rank, | ||||||
| 							 int xbytes,int rbytes, int dir) | 							 int bytes, int dir) | ||||||
| { | { | ||||||
|   return xbytes+rbytes; |   return 2.0*bytes; | ||||||
| } | } | ||||||
| void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir) | void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir) | ||||||
| { | { | ||||||
|   | |||||||
| @@ -40,9 +40,6 @@ int                 GlobalSharedMemory::_ShmAlloc; | |||||||
| uint64_t            GlobalSharedMemory::_ShmAllocBytes; | uint64_t            GlobalSharedMemory::_ShmAllocBytes; | ||||||
|  |  | ||||||
| std::vector<void *> GlobalSharedMemory::WorldShmCommBufs; | std::vector<void *> GlobalSharedMemory::WorldShmCommBufs; | ||||||
| #ifndef ACCELERATOR_AWARE_MPI |  | ||||||
| void * GlobalSharedMemory::HostCommBuf; |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
| Grid_MPI_Comm       GlobalSharedMemory::WorldShmComm; | Grid_MPI_Comm       GlobalSharedMemory::WorldShmComm; | ||||||
| int                 GlobalSharedMemory::WorldShmRank; | int                 GlobalSharedMemory::WorldShmRank; | ||||||
| @@ -69,26 +66,6 @@ void GlobalSharedMemory::SharedMemoryFree(void) | |||||||
| ///////////////////////////////// | ///////////////////////////////// | ||||||
| // Alloc, free shmem region | // Alloc, free shmem region | ||||||
| ///////////////////////////////// | ///////////////////////////////// | ||||||
| #ifndef ACCELERATOR_AWARE_MPI |  | ||||||
| void *SharedMemory::HostBufferMalloc(size_t bytes){ |  | ||||||
|   void *ptr = (void *)host_heap_top; |  | ||||||
|   host_heap_top  += bytes; |  | ||||||
|   host_heap_bytes+= bytes; |  | ||||||
|   if (host_heap_bytes >= host_heap_size) { |  | ||||||
|     std::cout<< " HostBufferMalloc exceeded heap size -- try increasing with --shm <MB> flag" <<std::endl; |  | ||||||
|     std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl; |  | ||||||
|     std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl; |  | ||||||
|     std::cout<< " Current bytes is " << (host_heap_bytes/(1024*1024)) <<"MB"<<std::endl; |  | ||||||
|     std::cout<< " Current heap  is " << (host_heap_size/(1024*1024)) <<"MB"<<std::endl; |  | ||||||
|     assert(host_heap_bytes<host_heap_size); |  | ||||||
|   } |  | ||||||
|   return ptr; |  | ||||||
| } |  | ||||||
| void SharedMemory::HostBufferFreeAll(void) {  |  | ||||||
|   host_heap_top  =(size_t)HostCommBuf; |  | ||||||
|   host_heap_bytes=0; |  | ||||||
| } |  | ||||||
| #endif |  | ||||||
| void *SharedMemory::ShmBufferMalloc(size_t bytes){ | void *SharedMemory::ShmBufferMalloc(size_t bytes){ | ||||||
|   //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes |   //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes | ||||||
|   void *ptr = (void *)heap_top; |   void *ptr = (void *)heap_top; | ||||||
| @@ -114,59 +91,6 @@ void *SharedMemory::ShmBufferSelf(void) | |||||||
|   //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl; |   //std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl; | ||||||
|   return ShmCommBufs[ShmRank]; |   return ShmCommBufs[ShmRank]; | ||||||
| } | } | ||||||
| static inline int divides(int a,int b) |  | ||||||
| { |  | ||||||
|   return ( b == ( (b/a)*a ) ); |  | ||||||
| } |  | ||||||
| void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims) |  | ||||||
| { |  | ||||||
|   //////////////////////////////////////////////////////////////// |  | ||||||
|   // Allow user to configure through environment variable |  | ||||||
|   //////////////////////////////////////////////////////////////// |  | ||||||
|   char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str()); |  | ||||||
|   if ( str ) { |  | ||||||
|     std::vector<int> IntShmDims; |  | ||||||
|     GridCmdOptionIntVector(std::string(str),IntShmDims); |  | ||||||
|     assert(IntShmDims.size() == WorldDims.size()); |  | ||||||
|     long ShmSize = 1; |  | ||||||
|     for (int dim=0;dim<WorldDims.size();dim++) { |  | ||||||
|       ShmSize *= (ShmDims[dim] = IntShmDims[dim]); |  | ||||||
|       assert(divides(ShmDims[dim],WorldDims[dim])); |  | ||||||
|     } |  | ||||||
|     assert(ShmSize == WorldShmSize); |  | ||||||
|     return; |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|   //////////////////////////////////////////////////////////////// |  | ||||||
|   // Powers of 2,3,5 only in prime decomposition for now |  | ||||||
|   //////////////////////////////////////////////////////////////// |  | ||||||
|   int ndimension = WorldDims.size(); |  | ||||||
|   ShmDims=Coordinate(ndimension,1); |  | ||||||
|  |  | ||||||
|   std::vector<int> primes({2,3,5}); |  | ||||||
|  |  | ||||||
|   int dim = 0; |  | ||||||
|   int last_dim = ndimension - 1; |  | ||||||
|   int AutoShmSize = 1; |  | ||||||
|   while(AutoShmSize != WorldShmSize) { |  | ||||||
|     int p; |  | ||||||
|     for(p=0;p<primes.size();p++) { |  | ||||||
|       int prime=primes[p]; |  | ||||||
|       if ( divides(prime,WorldDims[dim]/ShmDims[dim]) |  | ||||||
|         && divides(prime,WorldShmSize/AutoShmSize)  ) { |  | ||||||
|   AutoShmSize*=prime; |  | ||||||
|   ShmDims[dim]*=prime; |  | ||||||
|   last_dim = dim; |  | ||||||
|   break; |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
|     if (p == primes.size() && last_dim == dim) { |  | ||||||
|       std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl; |  | ||||||
|       exit(EXIT_FAILURE); |  | ||||||
|     } |  | ||||||
|     dim=(dim+1) %ndimension; |  | ||||||
|   } |  | ||||||
| } |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid);  | NAMESPACE_END(Grid);  | ||||||
|  |  | ||||||
|   | |||||||
| @@ -75,9 +75,7 @@ public: | |||||||
|   static int           Hugepages; |   static int           Hugepages; | ||||||
|  |  | ||||||
|   static std::vector<void *> WorldShmCommBufs; |   static std::vector<void *> WorldShmCommBufs; | ||||||
| #ifndef ACCELERATOR_AWARE_MPI |  | ||||||
|   static void *HostCommBuf; |  | ||||||
| #endif |  | ||||||
|   static Grid_MPI_Comm WorldComm; |   static Grid_MPI_Comm WorldComm; | ||||||
|   static int           WorldRank; |   static int           WorldRank; | ||||||
|   static int           WorldSize; |   static int           WorldSize; | ||||||
| @@ -95,17 +93,16 @@ public: | |||||||
|   // Create an optimal reordered communicator that makes MPI_Cart_create get it right |   // Create an optimal reordered communicator that makes MPI_Cart_create get it right | ||||||
|   ////////////////////////////////////////////////////////////////////////////////////// |   ////////////////////////////////////////////////////////////////////////////////////// | ||||||
|   static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD |   static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD | ||||||
|   // Turns MPI_COMM_WORLD into right layout for Cartesian |   static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian | ||||||
|   static void OptimalCommunicator            (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);  |   static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian | ||||||
|   static void OptimalCommunicatorHypercube   (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);  |   static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian | ||||||
|   static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);  |  | ||||||
|   static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims); |   static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims); | ||||||
|   /////////////////////////////////////////////////// |   /////////////////////////////////////////////////// | ||||||
|   // Provide shared memory facilities off comm world |   // Provide shared memory facilities off comm world | ||||||
|   /////////////////////////////////////////////////// |   /////////////////////////////////////////////////// | ||||||
|   static void SharedMemoryAllocate(uint64_t bytes, int flags); |   static void SharedMemoryAllocate(uint64_t bytes, int flags); | ||||||
|   static void SharedMemoryFree(void); |   static void SharedMemoryFree(void); | ||||||
|   static void SharedMemoryCopy(void *dest,void *src,size_t bytes); |   static void SharedMemoryCopy(void *dest,const void *src,size_t bytes); | ||||||
|   static void SharedMemoryZero(void *dest,size_t bytes); |   static void SharedMemoryZero(void *dest,size_t bytes); | ||||||
|  |  | ||||||
| }; | }; | ||||||
| @@ -122,13 +119,6 @@ private: | |||||||
|   size_t heap_bytes; |   size_t heap_bytes; | ||||||
|   size_t heap_size; |   size_t heap_size; | ||||||
|  |  | ||||||
| #ifndef ACCELERATOR_AWARE_MPI |  | ||||||
|   size_t host_heap_top;  // set in free all |  | ||||||
|   size_t host_heap_bytes;// set in free all |  | ||||||
|   void *HostCommBuf;     // set in SetCommunicator |  | ||||||
|   size_t host_heap_size; // set in SetCommunicator |  | ||||||
| #endif |  | ||||||
|    |  | ||||||
| protected: | protected: | ||||||
|  |  | ||||||
|   Grid_MPI_Comm    ShmComm; // for barriers |   Grid_MPI_Comm    ShmComm; // for barriers | ||||||
| @@ -160,10 +150,7 @@ public: | |||||||
|   void *ShmBufferTranslate(int rank,void * local_p); |   void *ShmBufferTranslate(int rank,void * local_p); | ||||||
|   void *ShmBufferMalloc(size_t bytes); |   void *ShmBufferMalloc(size_t bytes); | ||||||
|   void  ShmBufferFreeAll(void) ; |   void  ShmBufferFreeAll(void) ; | ||||||
| #ifndef ACCELERATOR_AWARE_MPI |    | ||||||
|   void *HostBufferMalloc(size_t bytes); |  | ||||||
|   void HostBufferFreeAll(void); |  | ||||||
| #endif   |  | ||||||
|   ////////////////////////////////////////////////////////////////////////// |   ////////////////////////////////////////////////////////////////////////// | ||||||
|   // Make info on Nodes & ranks and Shared memory available |   // Make info on Nodes & ranks and Shared memory available | ||||||
|   ////////////////////////////////////////////////////////////////////////// |   ////////////////////////////////////////////////////////////////////////// | ||||||
|   | |||||||
| @@ -7,7 +7,6 @@ | |||||||
|     Copyright (C) 2015 |     Copyright (C) 2015 | ||||||
|  |  | ||||||
| Author: Peter Boyle <paboyle@ph.ed.ac.uk> | Author: Peter Boyle <paboyle@ph.ed.ac.uk> | ||||||
| Author: Christoph Lehner <christoph@lhnr.de> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |     This program is free software; you can redistribute it and/or modify | ||||||
|     it under the terms of the GNU General Public License as published by |     it under the terms of the GNU General Public License as published by | ||||||
| @@ -27,8 +26,6 @@ Author: Christoph Lehner <christoph@lhnr.de> | |||||||
| *************************************************************************************/ | *************************************************************************************/ | ||||||
| /*  END LEGAL */ | /*  END LEGAL */ | ||||||
|  |  | ||||||
| #define Mheader "SharedMemoryMpi: " |  | ||||||
|  |  | ||||||
| #include <Grid/GridCore.h> | #include <Grid/GridCore.h> | ||||||
| #include <pwd.h> | #include <pwd.h> | ||||||
|  |  | ||||||
| @@ -38,122 +35,9 @@ Author: Christoph Lehner <christoph@lhnr.de> | |||||||
| #ifdef GRID_HIP | #ifdef GRID_HIP | ||||||
| #include <hip/hip_runtime_api.h> | #include <hip/hip_runtime_api.h> | ||||||
| #endif | #endif | ||||||
| #ifdef GRID_SYCL |  | ||||||
| #ifdef ACCELERATOR_AWARE_MPI |  | ||||||
| #define GRID_SYCL_LEVEL_ZERO_IPC |  | ||||||
| #define SHM_SOCKETS |  | ||||||
| #endif  |  | ||||||
| #include <syscall.h> |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
| #include <sys/socket.h> |  | ||||||
| #include <sys/un.h> |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid);  | NAMESPACE_BEGIN(Grid);  | ||||||
|  | #define header "SharedMemoryMpi: " | ||||||
| #ifdef SHM_SOCKETS |  | ||||||
|  |  | ||||||
| /* |  | ||||||
|  * Barbaric extra intranode communication route in case we need sockets to pass FDs |  | ||||||
|  * Forced by level_zero not being nicely designed |  | ||||||
|  */ |  | ||||||
| static int sock; |  | ||||||
| static const char *sock_path_fmt = "/tmp/GridUnixSocket.%d"; |  | ||||||
| static char sock_path[256]; |  | ||||||
| class UnixSockets { |  | ||||||
| public: |  | ||||||
|   static void Open(int rank) |  | ||||||
|   { |  | ||||||
|     int errnum; |  | ||||||
|  |  | ||||||
|     sock = socket(AF_UNIX, SOCK_DGRAM, 0);  assert(sock>0); |  | ||||||
|  |  | ||||||
|     struct sockaddr_un sa_un = { 0 }; |  | ||||||
|     sa_un.sun_family = AF_UNIX; |  | ||||||
|     snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,rank); |  | ||||||
|     unlink(sa_un.sun_path); |  | ||||||
|     if (bind(sock, (struct sockaddr *)&sa_un, sizeof(sa_un))) { |  | ||||||
|       perror("bind failure"); |  | ||||||
|       exit(EXIT_FAILURE); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static int RecvFileDescriptor(void) |  | ||||||
|   { |  | ||||||
|     int n; |  | ||||||
|     int fd; |  | ||||||
|     char buf[1]; |  | ||||||
|     struct iovec iov; |  | ||||||
|     struct msghdr msg; |  | ||||||
|     struct cmsghdr *cmsg; |  | ||||||
|     char cms[CMSG_SPACE(sizeof(int))]; |  | ||||||
|  |  | ||||||
|     iov.iov_base = buf; |  | ||||||
|     iov.iov_len = 1; |  | ||||||
|  |  | ||||||
|     memset(&msg, 0, sizeof msg); |  | ||||||
|     msg.msg_name = 0; |  | ||||||
|     msg.msg_namelen = 0; |  | ||||||
|     msg.msg_iov = &iov; |  | ||||||
|     msg.msg_iovlen = 1; |  | ||||||
|  |  | ||||||
|     msg.msg_control = (caddr_t)cms; |  | ||||||
|     msg.msg_controllen = sizeof cms; |  | ||||||
|  |  | ||||||
|     if((n=recvmsg(sock, &msg, 0)) < 0) { |  | ||||||
|       perror("recvmsg failed"); |  | ||||||
|       return -1; |  | ||||||
|     } |  | ||||||
|     if(n == 0){ |  | ||||||
|       perror("recvmsg returned 0"); |  | ||||||
|       return -1; |  | ||||||
|     } |  | ||||||
|     cmsg = CMSG_FIRSTHDR(&msg); |  | ||||||
|  |  | ||||||
|     memmove(&fd, CMSG_DATA(cmsg), sizeof(int)); |  | ||||||
|  |  | ||||||
|     return fd; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static void SendFileDescriptor(int fildes,int xmit_to_rank) |  | ||||||
|   { |  | ||||||
|     struct msghdr msg; |  | ||||||
|     struct iovec iov; |  | ||||||
|     struct cmsghdr *cmsg = NULL; |  | ||||||
|     char ctrl[CMSG_SPACE(sizeof(int))]; |  | ||||||
|     char data = ' '; |  | ||||||
|  |  | ||||||
|     memset(&msg, 0, sizeof(struct msghdr)); |  | ||||||
|     memset(ctrl, 0, CMSG_SPACE(sizeof(int))); |  | ||||||
|     iov.iov_base = &data; |  | ||||||
|     iov.iov_len = sizeof(data); |  | ||||||
|      |  | ||||||
|     sprintf(sock_path,sock_path_fmt,xmit_to_rank); |  | ||||||
|      |  | ||||||
|     struct sockaddr_un sa_un = { 0 }; |  | ||||||
|     sa_un.sun_family = AF_UNIX; |  | ||||||
|     snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,xmit_to_rank); |  | ||||||
|  |  | ||||||
|     msg.msg_name = (void *)&sa_un; |  | ||||||
|     msg.msg_namelen = sizeof(sa_un); |  | ||||||
|     msg.msg_iov = &iov; |  | ||||||
|     msg.msg_iovlen = 1; |  | ||||||
|     msg.msg_controllen =  CMSG_SPACE(sizeof(int)); |  | ||||||
|     msg.msg_control = ctrl; |  | ||||||
|  |  | ||||||
|     cmsg = CMSG_FIRSTHDR(&msg); |  | ||||||
|     cmsg->cmsg_level = SOL_SOCKET; |  | ||||||
|     cmsg->cmsg_type = SCM_RIGHTS; |  | ||||||
|     cmsg->cmsg_len = CMSG_LEN(sizeof(int)); |  | ||||||
|  |  | ||||||
|     *((int *) CMSG_DATA(cmsg)) = fildes; |  | ||||||
|  |  | ||||||
|     sendmsg(sock, &msg, 0); |  | ||||||
|   }; |  | ||||||
| }; |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
|  |  | ||||||
| /*Construct from an MPI communicator*/ | /*Construct from an MPI communicator*/ | ||||||
| void GlobalSharedMemory::Init(Grid_MPI_Comm comm) | void GlobalSharedMemory::Init(Grid_MPI_Comm comm) | ||||||
| { | { | ||||||
| @@ -176,8 +60,8 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm) | |||||||
|   MPI_Comm_size(WorldShmComm     ,&WorldShmSize); |   MPI_Comm_size(WorldShmComm     ,&WorldShmSize); | ||||||
|  |  | ||||||
|   if ( WorldRank == 0) { |   if ( WorldRank == 0) { | ||||||
|     std::cout << Mheader " World communicator of size " <<WorldSize << std::endl;   |     std::cout << header " World communicator of size " <<WorldSize << std::endl;   | ||||||
|     std::cout << Mheader " Node  communicator of size " <<WorldShmSize << std::endl; |     std::cout << header " Node  communicator of size " <<WorldShmSize << std::endl; | ||||||
|   } |   } | ||||||
|   // WorldShmComm, WorldShmSize, WorldShmRank |   // WorldShmComm, WorldShmSize, WorldShmRank | ||||||
|  |  | ||||||
| @@ -185,7 +69,6 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm) | |||||||
|   WorldNodes = WorldSize/WorldShmSize; |   WorldNodes = WorldSize/WorldShmSize; | ||||||
|   assert( (WorldNodes * WorldShmSize) == WorldSize ); |   assert( (WorldNodes * WorldShmSize) == WorldSize ); | ||||||
|  |  | ||||||
|  |  | ||||||
|   // FIXME: Check all WorldShmSize are the same ? |   // FIXME: Check all WorldShmSize are the same ? | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////////////////////////////// |   ///////////////////////////////////////////////////////////////////// | ||||||
| @@ -264,7 +147,7 @@ int Log2Size(int TwoToPower,int MAXLOG2) | |||||||
|   } |   } | ||||||
|   return log2size; |   return log2size; | ||||||
| } | } | ||||||
| void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM) | void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm) | ||||||
| { | { | ||||||
|   ////////////////////////////////////////////////////////////////////////////// |   ////////////////////////////////////////////////////////////////////////////// | ||||||
|   // Look and see if it looks like an HPE 8600 based on hostname conventions |   // Look and see if it looks like an HPE 8600 based on hostname conventions | ||||||
| @@ -277,11 +160,46 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M | |||||||
|   gethostname(name,namelen); |   gethostname(name,namelen); | ||||||
|   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ; |   int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ; | ||||||
|  |  | ||||||
|   if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM); |   if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm); | ||||||
|   else                          OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM); |   else                          OptimalCommunicatorSharedMemory(processors,optimal_comm); | ||||||
| } | } | ||||||
|  | static inline int divides(int a,int b) | ||||||
|  | { | ||||||
|  |   return ( b == ( (b/a)*a ) ); | ||||||
|  | } | ||||||
|  | void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims) | ||||||
|  | { | ||||||
|  |   //////////////////////////////////////////////////////////////// | ||||||
|  |   // Powers of 2,3,5 only in prime decomposition for now | ||||||
|  |   //////////////////////////////////////////////////////////////// | ||||||
|  |   int ndimension = WorldDims.size(); | ||||||
|  |   ShmDims=Coordinate(ndimension,1); | ||||||
|  |  | ||||||
| void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM) |   std::vector<int> primes({2,3,5}); | ||||||
|  |  | ||||||
|  |   int dim = 0; | ||||||
|  |   int last_dim = ndimension - 1; | ||||||
|  |   int AutoShmSize = 1; | ||||||
|  |   while(AutoShmSize != WorldShmSize) { | ||||||
|  |     int p; | ||||||
|  |     for(p=0;p<primes.size();p++) { | ||||||
|  |       int prime=primes[p]; | ||||||
|  |       if ( divides(prime,WorldDims[dim]/ShmDims[dim]) | ||||||
|  |         && divides(prime,WorldShmSize/AutoShmSize)  ) { | ||||||
|  | 	AutoShmSize*=prime; | ||||||
|  | 	ShmDims[dim]*=prime; | ||||||
|  | 	last_dim = dim; | ||||||
|  | 	break; | ||||||
|  |       } | ||||||
|  |     } | ||||||
|  |     if (p == primes.size() && last_dim == dim) { | ||||||
|  |       std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl; | ||||||
|  |       exit(EXIT_FAILURE); | ||||||
|  |     } | ||||||
|  |     dim=(dim+1) %ndimension; | ||||||
|  |   } | ||||||
|  | } | ||||||
|  | void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm) | ||||||
| { | { | ||||||
|   //////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////// | ||||||
|   // Assert power of two shm_size. |   // Assert power of two shm_size. | ||||||
| @@ -354,8 +272,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo | |||||||
|   Coordinate HyperCoor(ndimension); |   Coordinate HyperCoor(ndimension); | ||||||
|  |  | ||||||
|   GetShmDims(WorldDims,ShmDims); |   GetShmDims(WorldDims,ShmDims); | ||||||
|   SHM = ShmDims; |  | ||||||
|    |  | ||||||
|   //////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////// | ||||||
|   // Establish torus of processes and nodes with sub-blockings |   // Establish torus of processes and nodes with sub-blockings | ||||||
|   //////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////// | ||||||
| @@ -402,7 +319,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo | |||||||
|   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm); |   int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm); | ||||||
|   assert(ierr==0); |   assert(ierr==0); | ||||||
| } | } | ||||||
| void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM) | void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm) | ||||||
| { | { | ||||||
|   //////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////// | ||||||
|   // Identify subblock of ranks on node spreading across dims |   // Identify subblock of ranks on node spreading across dims | ||||||
| @@ -414,8 +331,6 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce | |||||||
|   Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension); |   Coordinate ShmCoor(ndimension);    Coordinate NodeCoor(ndimension);   Coordinate WorldCoor(ndimension); | ||||||
|  |  | ||||||
|   GetShmDims(WorldDims,ShmDims); |   GetShmDims(WorldDims,ShmDims); | ||||||
|   SHM=ShmDims; |  | ||||||
|  |  | ||||||
|   //////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////// | ||||||
|   // Establish torus of processes and nodes with sub-blockings |   // Establish torus of processes and nodes with sub-blockings | ||||||
|   //////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////// | ||||||
| @@ -454,7 +369,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce | |||||||
| #ifdef GRID_MPI3_SHMGET | #ifdef GRID_MPI3_SHMGET | ||||||
| void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | ||||||
| { | { | ||||||
|   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl; |   std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl; | ||||||
|   assert(_ShmSetup==1); |   assert(_ShmSetup==1); | ||||||
|   assert(_ShmAlloc==0); |   assert(_ShmAlloc==0); | ||||||
|  |  | ||||||
| @@ -513,7 +428,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
| //////////////////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| // Hugetlbfs mapping intended | // Hugetlbfs mapping intended | ||||||
| //////////////////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| #if defined(GRID_CUDA) ||defined(GRID_HIP)  || defined(GRID_SYCL) | #if defined(GRID_CUDA) ||defined(GRID_HIP) | ||||||
| void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | ||||||
| { | { | ||||||
|   void * ShmCommBuf ;  |   void * ShmCommBuf ;  | ||||||
| @@ -536,65 +451,28 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
|   /////////////////////////////////////////////////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
|   // Each MPI rank should allocate our own buffer |   // Each MPI rank should allocate our own buffer | ||||||
|   /////////////////////////////////////////////////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| #ifndef ACCELERATOR_AWARE_MPI |  | ||||||
|   HostCommBuf= malloc(bytes); |  | ||||||
| #endif   |  | ||||||
|   ShmCommBuf = acceleratorAllocDevice(bytes); |   ShmCommBuf = acceleratorAllocDevice(bytes); | ||||||
|  |  | ||||||
|   if (ShmCommBuf == (void *)NULL ) { |   if (ShmCommBuf == (void *)NULL ) { | ||||||
|     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl; |     std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl; | ||||||
|     exit(EXIT_FAILURE);   |     exit(EXIT_FAILURE);   | ||||||
|   } |   } | ||||||
|   if ( WorldRank == 0 ){ |   //  if ( WorldRank == 0 ){ | ||||||
|     std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes  |   if ( 1 ){ | ||||||
| 	      << "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl; |     std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes  | ||||||
|  | 	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl; | ||||||
|   } |   } | ||||||
|   SharedMemoryZero(ShmCommBuf,bytes); |   SharedMemoryZero(ShmCommBuf,bytes); | ||||||
|   std::cout<< "Setting up IPC"<<std::endl; |  | ||||||
|   /////////////////////////////////////////////////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
|   // Loop over ranks/gpu's on our node |   // Loop over ranks/gpu's on our node | ||||||
|   /////////////////////////////////////////////////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| #ifdef SHM_SOCKETS |  | ||||||
|   UnixSockets::Open(WorldShmRank); |  | ||||||
| #endif |  | ||||||
|   for(int r=0;r<WorldShmSize;r++){ |   for(int r=0;r<WorldShmSize;r++){ | ||||||
|  |  | ||||||
|     MPI_Barrier(WorldShmComm); |  | ||||||
|  |  | ||||||
| #ifndef GRID_MPI3_SHM_NONE | #ifndef GRID_MPI3_SHM_NONE | ||||||
|     ////////////////////////////////////////////////// |     ////////////////////////////////////////////////// | ||||||
|     // If it is me, pass around the IPC access key |     // If it is me, pass around the IPC access key | ||||||
|     ////////////////////////////////////////////////// |     ////////////////////////////////////////////////// | ||||||
|     void * thisBuf = ShmCommBuf; |  | ||||||
|     if(!Stencil_force_mpi) { |  | ||||||
| #ifdef GRID_SYCL_LEVEL_ZERO_IPC |  | ||||||
|     typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t; |  | ||||||
|  |  | ||||||
|     auto zeDevice    = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device()); |  | ||||||
|     auto zeContext   = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context()); |  | ||||||
|        |  | ||||||
|     ze_ipc_mem_handle_t ihandle; |  | ||||||
|     clone_mem_t handle; |  | ||||||
|      |  | ||||||
|     if ( r==WorldShmRank ) {  |  | ||||||
|       auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle); |  | ||||||
|       if ( err != ZE_RESULT_SUCCESS ) { |  | ||||||
| 	std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; |  | ||||||
| 	exit(EXIT_FAILURE); |  | ||||||
|       } else { |  | ||||||
| 	std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl; |  | ||||||
|       } |  | ||||||
|       memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int)); |  | ||||||
|       handle.pid = getpid(); |  | ||||||
|       memcpy((void *)&handle.ze,(void *)&ihandle,sizeof(ihandle)); |  | ||||||
| #ifdef SHM_SOCKETS |  | ||||||
|       for(int rr=0;rr<WorldShmSize;rr++){ |  | ||||||
| 	if(rr!=r){ |  | ||||||
| 	  UnixSockets::SendFileDescriptor(handle.fd,rr); |  | ||||||
| 	} |  | ||||||
|       } |  | ||||||
| #endif |  | ||||||
|     } |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA | #ifdef GRID_CUDA | ||||||
|     cudaIpcMemHandle_t handle; |     cudaIpcMemHandle_t handle; | ||||||
|     if ( r==WorldShmRank ) {  |     if ( r==WorldShmRank ) {  | ||||||
| @@ -615,12 +493,10 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
|       } |       } | ||||||
|     } |     } | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
|     ////////////////////////////////////////////////// |     ////////////////////////////////////////////////// | ||||||
|     // Share this IPC handle across the Shm Comm |     // Share this IPC handle across the Shm Comm | ||||||
|     ////////////////////////////////////////////////// |     ////////////////////////////////////////////////// | ||||||
|     {  |     {  | ||||||
|       MPI_Barrier(WorldShmComm); |  | ||||||
|       int ierr=MPI_Bcast(&handle, |       int ierr=MPI_Bcast(&handle, | ||||||
| 			 sizeof(handle), | 			 sizeof(handle), | ||||||
| 			 MPI_BYTE, | 			 MPI_BYTE, | ||||||
| @@ -632,45 +508,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
|     /////////////////////////////////////////////////////////////// |     /////////////////////////////////////////////////////////////// | ||||||
|     // If I am not the source, overwrite thisBuf with remote buffer |     // If I am not the source, overwrite thisBuf with remote buffer | ||||||
|     /////////////////////////////////////////////////////////////// |     /////////////////////////////////////////////////////////////// | ||||||
|  |     void * thisBuf = ShmCommBuf; | ||||||
| #ifdef GRID_SYCL_LEVEL_ZERO_IPC |  | ||||||
|     if ( r!=WorldShmRank ) { |  | ||||||
|       thisBuf = nullptr; |  | ||||||
|       int myfd; |  | ||||||
| #ifdef SHM_SOCKETS |  | ||||||
|       myfd=UnixSockets::RecvFileDescriptor(); |  | ||||||
| #else |  | ||||||
|       std::cout<<"mapping seeking remote pid/fd " |  | ||||||
| 	       <<handle.pid<<"/" |  | ||||||
| 	       <<handle.fd<<std::endl; |  | ||||||
|  |  | ||||||
|       int pidfd = syscall(SYS_pidfd_open,handle.pid,0); |  | ||||||
|       std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n"; |  | ||||||
|       //      int myfd  = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0); |  | ||||||
|       myfd  = syscall(438,pidfd,handle.fd,0); |  | ||||||
|       int err_t = errno; |  | ||||||
|       if (myfd < 0) { |  | ||||||
|         fprintf(stderr,"pidfd_getfd returned %d errno was %d\n", myfd,err_t); fflush(stderr); |  | ||||||
| 	perror("pidfd_getfd failed "); |  | ||||||
| 	assert(0); |  | ||||||
|       } |  | ||||||
| #endif |  | ||||||
|       std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n"; |  | ||||||
|       memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle)); |  | ||||||
|       memcpy((void *)&ihandle,(void *)&myfd,sizeof(int)); |  | ||||||
|  |  | ||||||
|       auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf); |  | ||||||
|       if ( err != ZE_RESULT_SUCCESS ) { |  | ||||||
| 	std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl; |  | ||||||
| 	std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;  |  | ||||||
| 	exit(EXIT_FAILURE); |  | ||||||
|       } else { |  | ||||||
| 	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl; |  | ||||||
| 	std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle pointer is "<<std::hex<<thisBuf<<std::dec<<std::endl; |  | ||||||
|       } |  | ||||||
|       assert(thisBuf!=nullptr); |  | ||||||
|     } |  | ||||||
| #endif |  | ||||||
| #ifdef GRID_CUDA | #ifdef GRID_CUDA | ||||||
|     if ( r!=WorldShmRank ) {  |     if ( r!=WorldShmRank ) {  | ||||||
|       auto err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess); |       auto err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess); | ||||||
| @@ -692,23 +530,20 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
|     /////////////////////////////////////////////////////////////// |     /////////////////////////////////////////////////////////////// | ||||||
|     // Save a copy of the device buffers |     // Save a copy of the device buffers | ||||||
|     /////////////////////////////////////////////////////////////// |     /////////////////////////////////////////////////////////////// | ||||||
|     } |  | ||||||
|     WorldShmCommBufs[r] = thisBuf; |     WorldShmCommBufs[r] = thisBuf; | ||||||
| #else | #else | ||||||
|     WorldShmCommBufs[r] = ShmCommBuf; |     WorldShmCommBufs[r] = ShmCommBuf; | ||||||
| #endif | #endif | ||||||
|     MPI_Barrier(WorldShmComm); |  | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   _ShmAllocBytes=bytes; |   _ShmAllocBytes=bytes; | ||||||
|   _ShmAlloc=1; |   _ShmAlloc=1; | ||||||
| } | } | ||||||
|  |  | ||||||
| #else  | #else  | ||||||
| #ifdef GRID_MPI3_SHMMMAP | #ifdef GRID_MPI3_SHMMMAP | ||||||
| void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | ||||||
| { | { | ||||||
|   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl; |   std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl; | ||||||
|   assert(_ShmSetup==1); |   assert(_ShmSetup==1); | ||||||
|   assert(_ShmAlloc==0); |   assert(_ShmAlloc==0); | ||||||
|   ////////////////////////////////////////////////////////////////////////////////////////////////////////// |   ////////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| @@ -745,7 +580,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
|     assert(((uint64_t)ptr&0x3F)==0); |     assert(((uint64_t)ptr&0x3F)==0); | ||||||
|     close(fd); |     close(fd); | ||||||
|     WorldShmCommBufs[r] =ptr; |     WorldShmCommBufs[r] =ptr; | ||||||
|     //    std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl; |     //    std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl; | ||||||
|   } |   } | ||||||
|   _ShmAlloc=1; |   _ShmAlloc=1; | ||||||
|   _ShmAllocBytes  = bytes; |   _ShmAllocBytes  = bytes; | ||||||
| @@ -755,7 +590,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
| #ifdef GRID_MPI3_SHM_NONE | #ifdef GRID_MPI3_SHM_NONE | ||||||
| void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | ||||||
| { | { | ||||||
|   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl; |   std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl; | ||||||
|   assert(_ShmSetup==1); |   assert(_ShmSetup==1); | ||||||
|   assert(_ShmAlloc==0); |   assert(_ShmAlloc==0); | ||||||
|   ////////////////////////////////////////////////////////////////////////////////////////////////////////// |   ////////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| @@ -802,7 +637,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
| //////////////////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | ||||||
| {  | {  | ||||||
|   std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl; |   std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl; | ||||||
|   assert(_ShmSetup==1); |   assert(_ShmSetup==1); | ||||||
|   assert(_ShmAlloc==0);  |   assert(_ShmAlloc==0);  | ||||||
|   MPI_Barrier(WorldShmComm); |   MPI_Barrier(WorldShmComm); | ||||||
| @@ -874,16 +709,16 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
| ///////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////// | ||||||
| void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes) | void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes) | ||||||
| { | { | ||||||
| #if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL) | #ifdef GRID_CUDA | ||||||
|   acceleratorMemSet(dest,0,bytes); |   cudaMemset(dest,0,bytes); | ||||||
| #else | #else | ||||||
|   bzero(dest,bytes); |   bzero(dest,bytes); | ||||||
| #endif | #endif | ||||||
| } | } | ||||||
| void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes) | void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes) | ||||||
| { | { | ||||||
| #if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL) | #ifdef GRID_CUDA | ||||||
|   acceleratorCopyToDevice(src,dest,bytes); |   cudaMemcpy(dest,src,bytes,cudaMemcpyDefault); | ||||||
| #else    | #else    | ||||||
|   bcopy(src,dest,bytes); |   bcopy(src,dest,bytes); | ||||||
| #endif | #endif | ||||||
| @@ -926,12 +761,6 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm) | |||||||
|   } |   } | ||||||
|   ShmBufferFreeAll(); |   ShmBufferFreeAll(); | ||||||
|  |  | ||||||
| #ifndef ACCELERATOR_AWARE_MPI |  | ||||||
|   host_heap_size = heap_size; |  | ||||||
|   HostCommBuf= GlobalSharedMemory::HostCommBuf; |  | ||||||
|   HostBufferFreeAll(); |  | ||||||
| #endif   |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////////////////////////////// |   ///////////////////////////////////////////////////////////////////// | ||||||
|   // find comm ranks in our SHM group (i.e. which ranks are on our node) |   // find comm ranks in our SHM group (i.e. which ranks are on our node) | ||||||
|   ///////////////////////////////////////////////////////////////////// |   ///////////////////////////////////////////////////////////////////// | ||||||
| @@ -953,7 +782,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm) | |||||||
|   } |   } | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
|   //SharedMemoryTest(); |   SharedMemoryTest(); | ||||||
| } | } | ||||||
| ////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////// | ||||||
| // On node barrier | // On node barrier | ||||||
|   | |||||||
| @@ -29,7 +29,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk> | |||||||
| #include <Grid/GridCore.h> | #include <Grid/GridCore.h> | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid);  | NAMESPACE_BEGIN(Grid);  | ||||||
| #define header "SharedMemoryNone: " |  | ||||||
|  |  | ||||||
| /*Construct from an MPI communicator*/ | /*Construct from an MPI communicator*/ | ||||||
| void GlobalSharedMemory::Init(Grid_MPI_Comm comm) | void GlobalSharedMemory::Init(Grid_MPI_Comm comm) | ||||||
| @@ -48,47 +47,14 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm) | |||||||
|   _ShmSetup=1; |   _ShmSetup=1; | ||||||
| } | } | ||||||
|  |  | ||||||
| void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM) | void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm) | ||||||
| { | { | ||||||
|   optimal_comm = WorldComm; |   optimal_comm = WorldComm; | ||||||
|   SHM = Coordinate(processors.size(),1); |  | ||||||
| } | } | ||||||
|  |  | ||||||
| //////////////////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| // Hugetlbfs mapping intended, use anonymous mmap | // Hugetlbfs mapping intended, use anonymous mmap | ||||||
| //////////////////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| #if 1 |  | ||||||
| void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) |  | ||||||
| { |  | ||||||
|   std::cout << header "SharedMemoryAllocate "<< bytes<< " GPU implementation "<<std::endl; |  | ||||||
|   void * ShmCommBuf ;  |  | ||||||
|   assert(_ShmSetup==1); |  | ||||||
|   assert(_ShmAlloc==0); |  | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|   // Each MPI rank should allocate our own buffer |  | ||||||
|   /////////////////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|   ShmCommBuf = acceleratorAllocDevice(bytes); |  | ||||||
|  |  | ||||||
|   if (ShmCommBuf == (void *)NULL ) { |  | ||||||
|     std::cerr << " SharedMemoryNone.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl; |  | ||||||
|     exit(EXIT_FAILURE);   |  | ||||||
|   } |  | ||||||
|   if ( WorldRank == 0 ){ |  | ||||||
|     std::cout << WorldRank << header " SharedMemoryNone.cc acceleratorAllocDevice "<< bytes  |  | ||||||
| 	      << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl; |  | ||||||
|   } |  | ||||||
|   SharedMemoryZero(ShmCommBuf,bytes); |  | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|   // Loop over ranks/gpu's on our node |  | ||||||
|   /////////////////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|   WorldShmCommBufs[0] = ShmCommBuf; |  | ||||||
|  |  | ||||||
|   _ShmAllocBytes=bytes; |  | ||||||
|   _ShmAlloc=1; |  | ||||||
| } |  | ||||||
| #else |  | ||||||
| void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | ||||||
| { | { | ||||||
|   void * ShmCommBuf ;  |   void * ShmCommBuf ;  | ||||||
| @@ -117,15 +83,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) | |||||||
|   _ShmAllocBytes=bytes; |   _ShmAllocBytes=bytes; | ||||||
|   _ShmAlloc=1; |   _ShmAlloc=1; | ||||||
| }; | }; | ||||||
| #endif |  | ||||||
| void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes) |  | ||||||
| { |  | ||||||
|   acceleratorMemSet(dest,0,bytes); |  | ||||||
| } |  | ||||||
| void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes) |  | ||||||
| { |  | ||||||
|   acceleratorCopyToDevice(src,dest,bytes); |  | ||||||
| } |  | ||||||
| //////////////////////////////////////////////////////// | //////////////////////////////////////////////////////// | ||||||
| // Global shared functionality finished | // Global shared functionality finished | ||||||
| // Now move to per communicator functionality | // Now move to per communicator functionality | ||||||
|   | |||||||
| @@ -29,27 +29,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk> | |||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| extern std::vector<std::pair<int,int> > Cshift_table;  | extern Vector<std::pair<int,int> > Cshift_table;  | ||||||
| extern commVector<std::pair<int,int> > Cshift_table_device;  |  | ||||||
|  |  | ||||||
| inline std::pair<int,int> *MapCshiftTable(void) |  | ||||||
| { |  | ||||||
|   // GPU version |  | ||||||
| #ifdef ACCELERATOR_CSHIFT     |  | ||||||
|   uint64_t sz=Cshift_table.size(); |  | ||||||
|   if (Cshift_table_device.size()!=sz )    { |  | ||||||
|     Cshift_table_device.resize(sz); |  | ||||||
|   } |  | ||||||
|   acceleratorCopyToDevice((void *)&Cshift_table[0], |  | ||||||
| 			  (void *)&Cshift_table_device[0], |  | ||||||
| 			  sizeof(Cshift_table[0])*sz); |  | ||||||
|  |  | ||||||
|   return &Cshift_table_device[0]; |  | ||||||
| #else  |  | ||||||
|   return &Cshift_table[0]; |  | ||||||
| #endif |  | ||||||
|   // CPU version use identify map |  | ||||||
| } |  | ||||||
| /////////////////////////////////////////////////////////////////// | /////////////////////////////////////////////////////////////////// | ||||||
| // Gather for when there is no need to SIMD split  | // Gather for when there is no need to SIMD split  | ||||||
| /////////////////////////////////////////////////////////////////// | /////////////////////////////////////////////////////////////////// | ||||||
| @@ -93,8 +74,8 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim | |||||||
|   } |   } | ||||||
|   { |   { | ||||||
|     auto buffer_p = & buffer[0]; |     auto buffer_p = & buffer[0]; | ||||||
|     auto table = MapCshiftTable(); |     auto table = &Cshift_table[0]; | ||||||
| #ifdef ACCELERATOR_CSHIFT | #ifdef ACCELERATOR_CSHIFT     | ||||||
|     autoView(rhs_v , rhs, AcceleratorRead); |     autoView(rhs_v , rhs, AcceleratorRead); | ||||||
|     accelerator_for(i,ent,vobj::Nsimd(),{ |     accelerator_for(i,ent,vobj::Nsimd(),{ | ||||||
| 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second])); | 	coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second])); | ||||||
| @@ -129,11 +110,9 @@ Gather_plane_extract(const Lattice<vobj> &rhs, | |||||||
|   int n1=rhs.Grid()->_slice_stride[dimension]; |   int n1=rhs.Grid()->_slice_stride[dimension]; | ||||||
|  |  | ||||||
|   if ( cbmask ==0x3){ |   if ( cbmask ==0x3){ | ||||||
| #ifdef ACCELERATOR_CSHIFT | #ifdef ACCELERATOR_CSHIFT     | ||||||
|     autoView(rhs_v , rhs, AcceleratorRead); |     autoView(rhs_v , rhs, AcceleratorRead); | ||||||
|     accelerator_for(nn,e1*e2,1,{ |     accelerator_for2d(n,e1,b,e2,1,{ | ||||||
| 	int n = nn%e1; |  | ||||||
| 	int b = nn/e1; |  | ||||||
| 	int o      =   n*n1; | 	int o      =   n*n1; | ||||||
| 	int offset = b+n*e2; | 	int offset = b+n*e2; | ||||||
| 	 | 	 | ||||||
| @@ -156,9 +135,7 @@ Gather_plane_extract(const Lattice<vobj> &rhs, | |||||||
|     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb? |     std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb? | ||||||
| #ifdef ACCELERATOR_CSHIFT     | #ifdef ACCELERATOR_CSHIFT     | ||||||
|     autoView(rhs_v , rhs, AcceleratorRead); |     autoView(rhs_v , rhs, AcceleratorRead); | ||||||
|     accelerator_for(nn,e1*e2,1,{ |     accelerator_for2d(n,e1,b,e2,1,{ | ||||||
| 	int n = nn%e1; |  | ||||||
| 	int b = nn/e1; |  | ||||||
|  |  | ||||||
| 	Coordinate coor; | 	Coordinate coor; | ||||||
|  |  | ||||||
| @@ -244,7 +221,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector< | |||||||
|    |    | ||||||
|   { |   { | ||||||
|     auto buffer_p = & buffer[0]; |     auto buffer_p = & buffer[0]; | ||||||
|     auto table = MapCshiftTable(); |     auto table = &Cshift_table[0]; | ||||||
| #ifdef ACCELERATOR_CSHIFT     | #ifdef ACCELERATOR_CSHIFT     | ||||||
|     autoView( rhs_v, rhs, AcceleratorWrite); |     autoView( rhs_v, rhs, AcceleratorWrite); | ||||||
|     accelerator_for(i,ent,vobj::Nsimd(),{ |     accelerator_for(i,ent,vobj::Nsimd(),{ | ||||||
| @@ -280,9 +257,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA | |||||||
|     int _slice_block = rhs.Grid()->_slice_block[dimension]; |     int _slice_block = rhs.Grid()->_slice_block[dimension]; | ||||||
| #ifdef ACCELERATOR_CSHIFT     | #ifdef ACCELERATOR_CSHIFT     | ||||||
|     autoView( rhs_v , rhs, AcceleratorWrite); |     autoView( rhs_v , rhs, AcceleratorWrite); | ||||||
|     accelerator_for(nn,e1*e2,1,{ |     accelerator_for2d(n,e1,b,e2,1,{ | ||||||
| 	int n = nn%e1; |  | ||||||
| 	int b = nn/e1; |  | ||||||
| 	int o      = n*_slice_stride; | 	int o      = n*_slice_stride; | ||||||
| 	int offset = b+n*_slice_block; | 	int offset = b+n*_slice_block; | ||||||
| 	merge(rhs_v[so+o+b],pointers,offset); | 	merge(rhs_v[so+o+b],pointers,offset); | ||||||
| @@ -299,7 +274,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA | |||||||
|  |  | ||||||
|     // Case of SIMD split AND checker dim cannot currently be hit, except in  |     // Case of SIMD split AND checker dim cannot currently be hit, except in  | ||||||
|     // Test_cshift_red_black code. |     // Test_cshift_red_black code. | ||||||
|     std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME |     //    std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME | ||||||
|     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl; |     std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl; | ||||||
|     assert(0); // This will fail if hit on GPU |     assert(0); // This will fail if hit on GPU | ||||||
|     autoView( rhs_v, rhs, CpuWrite); |     autoView( rhs_v, rhs, CpuWrite); | ||||||
| @@ -359,7 +334,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs | |||||||
|   } |   } | ||||||
|  |  | ||||||
|   { |   { | ||||||
|     auto table = MapCshiftTable(); |     auto table = &Cshift_table[0]; | ||||||
| #ifdef ACCELERATOR_CSHIFT     | #ifdef ACCELERATOR_CSHIFT     | ||||||
|     autoView(rhs_v , rhs, AcceleratorRead); |     autoView(rhs_v , rhs, AcceleratorRead); | ||||||
|     autoView(lhs_v , lhs, AcceleratorWrite); |     autoView(lhs_v , lhs, AcceleratorWrite); | ||||||
| @@ -411,7 +386,7 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo | |||||||
|   } |   } | ||||||
|  |  | ||||||
|   { |   { | ||||||
|     auto table = MapCshiftTable(); |     auto table = &Cshift_table[0]; | ||||||
| #ifdef ACCELERATOR_CSHIFT     | #ifdef ACCELERATOR_CSHIFT     | ||||||
|     autoView( rhs_v, rhs, AcceleratorRead); |     autoView( rhs_v, rhs, AcceleratorRead); | ||||||
|     autoView( lhs_v, lhs, AcceleratorWrite); |     autoView( lhs_v, lhs, AcceleratorWrite); | ||||||
|   | |||||||
| @@ -52,8 +52,7 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension | |||||||
|   int comm_dim        = rhs.Grid()->_processors[dimension] >1 ; |   int comm_dim        = rhs.Grid()->_processors[dimension] >1 ; | ||||||
|   int splice_dim      = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim); |   int splice_dim      = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim); | ||||||
|  |  | ||||||
|   RealD t1,t0; |  | ||||||
|   t0=usecond(); |  | ||||||
|   if ( !comm_dim ) { |   if ( !comm_dim ) { | ||||||
|     //std::cout << "CSHIFT: Cshift_local" <<std::endl; |     //std::cout << "CSHIFT: Cshift_local" <<std::endl; | ||||||
|     Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding |     Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding | ||||||
| @@ -64,8 +63,6 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension | |||||||
|     //std::cout << "CSHIFT: Cshift_comms" <<std::endl; |     //std::cout << "CSHIFT: Cshift_comms" <<std::endl; | ||||||
|     Cshift_comms(ret,rhs,dimension,shift); |     Cshift_comms(ret,rhs,dimension,shift); | ||||||
|   } |   } | ||||||
|   t1=usecond(); |  | ||||||
|   //  std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl; |  | ||||||
|   return ret; |   return ret; | ||||||
| } | } | ||||||
|  |  | ||||||
| @@ -125,25 +122,21 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r | |||||||
|   assert(shift<fd); |   assert(shift<fd); | ||||||
|    |    | ||||||
|   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension]; |   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension]; | ||||||
|   static cshiftVector<vobj> send_buf; send_buf.resize(buffer_size); |   cshiftVector<vobj> send_buf(buffer_size); | ||||||
|   static cshiftVector<vobj> recv_buf; recv_buf.resize(buffer_size); |   cshiftVector<vobj> recv_buf(buffer_size); | ||||||
|      |      | ||||||
|   int cb= (cbmask==0x2)? Odd : Even; |   int cb= (cbmask==0x2)? Odd : Even; | ||||||
|   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb); |   int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb); | ||||||
|   RealD tcopy=0.0; |  | ||||||
|   RealD tgather=0.0; |  | ||||||
|   RealD tscatter=0.0; |  | ||||||
|   RealD tcomms=0.0; |  | ||||||
|   uint64_t xbytes=0; |  | ||||||
|   for(int x=0;x<rd;x++){        |   for(int x=0;x<rd;x++){        | ||||||
|  |  | ||||||
|     int sx        =  (x+sshift)%rd; |     int sx        =  (x+sshift)%rd; | ||||||
|     int comm_proc = ((x+sshift)/rd)%pd; |     int comm_proc = ((x+sshift)/rd)%pd; | ||||||
|      |      | ||||||
|     if (comm_proc==0) { |     if (comm_proc==0) { | ||||||
|       tcopy-=usecond(); |  | ||||||
|       Copy_plane(ret,rhs,dimension,x,sx,cbmask);  |       Copy_plane(ret,rhs,dimension,x,sx,cbmask);  | ||||||
|       tcopy+=usecond(); |  | ||||||
|     } else { |     } else { | ||||||
|  |  | ||||||
|       int words = buffer_size; |       int words = buffer_size; | ||||||
| @@ -151,39 +144,26 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r | |||||||
|  |  | ||||||
|       int bytes = words * sizeof(vobj); |       int bytes = words * sizeof(vobj); | ||||||
|  |  | ||||||
|       tgather-=usecond(); |  | ||||||
|       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask); |       Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask); | ||||||
|       tgather+=usecond(); |  | ||||||
|  |  | ||||||
|       //      int rank           = grid->_processor; |       //      int rank           = grid->_processor; | ||||||
|       int recv_from_rank; |       int recv_from_rank; | ||||||
|       int xmit_to_rank; |       int xmit_to_rank; | ||||||
|       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank); |       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank); | ||||||
|        |  | ||||||
|       tcomms-=usecond(); |       grid->Barrier(); | ||||||
|       //      grid->Barrier(); |  | ||||||
|  |  | ||||||
|       grid->SendToRecvFrom((void *)&send_buf[0], |       grid->SendToRecvFrom((void *)&send_buf[0], | ||||||
| 			   xmit_to_rank, | 			   xmit_to_rank, | ||||||
| 			   (void *)&recv_buf[0], | 			   (void *)&recv_buf[0], | ||||||
| 			   recv_from_rank, | 			   recv_from_rank, | ||||||
| 			   bytes); | 			   bytes); | ||||||
|       xbytes+=bytes; |  | ||||||
|       //      grid->Barrier(); |  | ||||||
|       tcomms+=usecond(); |  | ||||||
|  |  | ||||||
|       tscatter-=usecond(); |       grid->Barrier(); | ||||||
|  |  | ||||||
|       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask); |       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask); | ||||||
|       tscatter+=usecond(); |  | ||||||
|     } |     } | ||||||
|   } |   } | ||||||
|   /* |  | ||||||
|   std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl; |  | ||||||
|   */ |  | ||||||
| } | } | ||||||
|  |  | ||||||
| template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask) | template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask) | ||||||
| @@ -210,12 +190,6 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|   assert(shift>=0); |   assert(shift>=0); | ||||||
|   assert(shift<fd); |   assert(shift<fd); | ||||||
|  |  | ||||||
|   RealD tcopy=0.0; |  | ||||||
|   RealD tgather=0.0; |  | ||||||
|   RealD tscatter=0.0; |  | ||||||
|   RealD tcomms=0.0; |  | ||||||
|   uint64_t xbytes=0; |  | ||||||
|    |  | ||||||
|   int permute_type=grid->PermuteType(dimension); |   int permute_type=grid->PermuteType(dimension); | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////// |   /////////////////////////////////////////////// | ||||||
| @@ -224,8 +198,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension]; |   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension]; | ||||||
|   //  int words = sizeof(vobj)/sizeof(vector_type); |   //  int words = sizeof(vobj)/sizeof(vector_type); | ||||||
|  |  | ||||||
|   static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd); |   std::vector<cshiftVector<scalar_object> >  send_buf_extract(Nsimd); | ||||||
|   static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd); |   std::vector<cshiftVector<scalar_object> >  recv_buf_extract(Nsimd); | ||||||
|   scalar_object *  recv_buf_extract_mpi; |   scalar_object *  recv_buf_extract_mpi; | ||||||
|   scalar_object *  send_buf_extract_mpi; |   scalar_object *  send_buf_extract_mpi; | ||||||
|   |   | ||||||
| @@ -253,9 +227,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|       pointers[i] = &send_buf_extract[i][0]; |       pointers[i] = &send_buf_extract[i][0]; | ||||||
|     } |     } | ||||||
|     int sx   = (x+sshift)%rd; |     int sx   = (x+sshift)%rd; | ||||||
|     tgather-=usecond(); |  | ||||||
|     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask); |     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask); | ||||||
|     tgather+=usecond(); |  | ||||||
|  |  | ||||||
|     for(int i=0;i<Nsimd;i++){ |     for(int i=0;i<Nsimd;i++){ | ||||||
|        |        | ||||||
| @@ -280,8 +252,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|       if(nbr_proc){ |       if(nbr_proc){ | ||||||
| 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);  | 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);  | ||||||
|  |  | ||||||
| 	tcomms-=usecond(); | 	grid->Barrier(); | ||||||
| 	//	grid->Barrier(); |  | ||||||
|  |  | ||||||
| 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0]; | 	send_buf_extract_mpi = &send_buf_extract[nbr_lane][0]; | ||||||
| 	recv_buf_extract_mpi = &recv_buf_extract[i][0]; | 	recv_buf_extract_mpi = &recv_buf_extract[i][0]; | ||||||
| @@ -291,9 +262,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
| 			     recv_from_rank, | 			     recv_from_rank, | ||||||
| 			     bytes); | 			     bytes); | ||||||
|  |  | ||||||
| 	xbytes+=bytes; | 	grid->Barrier(); | ||||||
| 	//	grid->Barrier(); |  | ||||||
| 	tcomms+=usecond(); |  | ||||||
|  |  | ||||||
| 	rpointers[i] = &recv_buf_extract[i][0]; | 	rpointers[i] = &recv_buf_extract[i][0]; | ||||||
|       } else {  |       } else {  | ||||||
| @@ -301,17 +270,9 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|       } |       } | ||||||
|  |  | ||||||
|     } |     } | ||||||
|     tscatter-=usecond(); |  | ||||||
|     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask); |     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask); | ||||||
|     tscatter+=usecond(); |  | ||||||
|   } |   } | ||||||
|   /* |  | ||||||
|   std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl; |  | ||||||
|   */ |  | ||||||
| } | } | ||||||
| #else  | #else  | ||||||
| template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask) | template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask) | ||||||
| @@ -331,15 +292,10 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r | |||||||
|   assert(comm_dim==1); |   assert(comm_dim==1); | ||||||
|   assert(shift>=0); |   assert(shift>=0); | ||||||
|   assert(shift<fd); |   assert(shift<fd); | ||||||
|   RealD tcopy=0.0; |  | ||||||
|   RealD tgather=0.0; |  | ||||||
|   RealD tscatter=0.0; |  | ||||||
|   RealD tcomms=0.0; |  | ||||||
|   uint64_t xbytes=0; |  | ||||||
|    |    | ||||||
|   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension]; |   int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension]; | ||||||
|   static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size); |   cshiftVector<vobj> send_buf_v(buffer_size); | ||||||
|   static cshiftVector<vobj> recv_buf_v; recv_buf_v.resize(buffer_size); |   cshiftVector<vobj> recv_buf_v(buffer_size); | ||||||
|   vobj *send_buf; |   vobj *send_buf; | ||||||
|   vobj *recv_buf; |   vobj *recv_buf; | ||||||
|   { |   { | ||||||
| @@ -359,9 +315,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r | |||||||
|      |      | ||||||
|     if (comm_proc==0) { |     if (comm_proc==0) { | ||||||
|  |  | ||||||
|       tcopy-=usecond(); |  | ||||||
|       Copy_plane(ret,rhs,dimension,x,sx,cbmask);  |       Copy_plane(ret,rhs,dimension,x,sx,cbmask);  | ||||||
|       tcopy+=usecond(); |  | ||||||
|  |  | ||||||
|     } else { |     } else { | ||||||
|  |  | ||||||
| @@ -370,9 +324,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r | |||||||
|  |  | ||||||
|       int bytes = words * sizeof(vobj); |       int bytes = words * sizeof(vobj); | ||||||
|  |  | ||||||
|       tgather-=usecond(); |  | ||||||
|       Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask); |       Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask); | ||||||
|       tgather+=usecond(); |  | ||||||
|  |  | ||||||
|       //      int rank           = grid->_processor; |       //      int rank           = grid->_processor; | ||||||
|       int recv_from_rank; |       int recv_from_rank; | ||||||
| @@ -380,8 +332,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r | |||||||
|       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank); |       grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank); | ||||||
|  |  | ||||||
|  |  | ||||||
|       tcomms-=usecond(); |       grid->Barrier(); | ||||||
|       //      grid->Barrier(); |  | ||||||
|  |  | ||||||
|       acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes); |       acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes); | ||||||
|       grid->SendToRecvFrom((void *)&send_buf[0], |       grid->SendToRecvFrom((void *)&send_buf[0], | ||||||
| @@ -389,24 +340,13 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r | |||||||
| 			   (void *)&recv_buf[0], | 			   (void *)&recv_buf[0], | ||||||
| 			   recv_from_rank, | 			   recv_from_rank, | ||||||
| 			   bytes); | 			   bytes); | ||||||
|       xbytes+=bytes; |  | ||||||
|       acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes); |       acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes); | ||||||
|  |  | ||||||
|       //      grid->Barrier(); |       grid->Barrier(); | ||||||
|       tcomms+=usecond(); |  | ||||||
|  |  | ||||||
|       tscatter-=usecond(); |  | ||||||
|       Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask); |       Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask); | ||||||
|       tscatter+=usecond(); |  | ||||||
|     } |     } | ||||||
|   } |   } | ||||||
|   /* |  | ||||||
|   std::cout << GridLogPerformance << " Cshift copy    "<<tcopy/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift gather  "<<tgather/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift comm    "<<tcomms/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl; |  | ||||||
|   */ |  | ||||||
| } | } | ||||||
|  |  | ||||||
| template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask) | template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask) | ||||||
| @@ -432,11 +372,6 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|   assert(simd_layout==2); |   assert(simd_layout==2); | ||||||
|   assert(shift>=0); |   assert(shift>=0); | ||||||
|   assert(shift<fd); |   assert(shift<fd); | ||||||
|   RealD tcopy=0.0; |  | ||||||
|   RealD tgather=0.0; |  | ||||||
|   RealD tscatter=0.0; |  | ||||||
|   RealD tcomms=0.0; |  | ||||||
|   uint64_t xbytes=0; |  | ||||||
|  |  | ||||||
|   int permute_type=grid->PermuteType(dimension); |   int permute_type=grid->PermuteType(dimension); | ||||||
|  |  | ||||||
| @@ -446,8 +381,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension]; |   int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension]; | ||||||
|   //  int words = sizeof(vobj)/sizeof(vector_type); |   //  int words = sizeof(vobj)/sizeof(vector_type); | ||||||
|  |  | ||||||
|   static std::vector<cshiftVector<scalar_object> >  send_buf_extract; send_buf_extract.resize(Nsimd); |   std::vector<cshiftVector<scalar_object> >  send_buf_extract(Nsimd); | ||||||
|   static std::vector<cshiftVector<scalar_object> >  recv_buf_extract; recv_buf_extract.resize(Nsimd); |   std::vector<cshiftVector<scalar_object> >  recv_buf_extract(Nsimd); | ||||||
|   scalar_object *  recv_buf_extract_mpi; |   scalar_object *  recv_buf_extract_mpi; | ||||||
|   scalar_object *  send_buf_extract_mpi; |   scalar_object *  send_buf_extract_mpi; | ||||||
|   { |   { | ||||||
| @@ -479,10 +414,8 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|     for(int i=0;i<Nsimd;i++){        |     for(int i=0;i<Nsimd;i++){        | ||||||
|       pointers[i] = &send_buf_extract[i][0]; |       pointers[i] = &send_buf_extract[i][0]; | ||||||
|     } |     } | ||||||
|     tgather-=usecond(); |  | ||||||
|     int sx   = (x+sshift)%rd; |     int sx   = (x+sshift)%rd; | ||||||
|     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask); |     Gather_plane_extract(rhs,pointers,dimension,sx,cbmask); | ||||||
|     tgather+=usecond(); |  | ||||||
|  |  | ||||||
|     for(int i=0;i<Nsimd;i++){ |     for(int i=0;i<Nsimd;i++){ | ||||||
|        |        | ||||||
| @@ -507,8 +440,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
|       if(nbr_proc){ |       if(nbr_proc){ | ||||||
| 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);  | 	grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);  | ||||||
|  |  | ||||||
| 	tcomms-=usecond(); | 	grid->Barrier(); | ||||||
| 	//	grid->Barrier(); |  | ||||||
|  |  | ||||||
| 	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes); | 	acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes); | ||||||
| 	grid->SendToRecvFrom((void *)send_buf_extract_mpi, | 	grid->SendToRecvFrom((void *)send_buf_extract_mpi, | ||||||
| @@ -517,28 +449,17 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo | |||||||
| 			     recv_from_rank, | 			     recv_from_rank, | ||||||
| 			     bytes); | 			     bytes); | ||||||
| 	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes); | 	acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes); | ||||||
| 	xbytes+=bytes; |  | ||||||
|  |  | ||||||
| 	//	grid->Barrier(); | 	grid->Barrier(); | ||||||
| 	tcomms+=usecond(); |  | ||||||
| 	rpointers[i] = &recv_buf_extract[i][0]; | 	rpointers[i] = &recv_buf_extract[i][0]; | ||||||
|       } else {  |       } else {  | ||||||
| 	rpointers[i] = &send_buf_extract[nbr_lane][0]; | 	rpointers[i] = &send_buf_extract[nbr_lane][0]; | ||||||
|       } |       } | ||||||
|  |  | ||||||
|     } |     } | ||||||
|     tscatter-=usecond(); |  | ||||||
|     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask); |     Scatter_plane_merge(ret,rpointers,dimension,x,cbmask); | ||||||
|     tscatter+=usecond(); |  | ||||||
|  |  | ||||||
|   } |   } | ||||||
|   /* |  | ||||||
|   std::cout << GridLogPerformance << " Cshift (s) copy    "<<tcopy/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift (s) gather  "<<tgather/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift (s) comm    "<<tcomms/1e3<<" ms"<<std::endl; |  | ||||||
|   std::cout << GridLogPerformance << " Cshift BW      "<<(2.0*xbytes)/tcomms<<" MB/s"<<std::endl; |  | ||||||
|   */ |  | ||||||
| } | } | ||||||
| #endif | #endif | ||||||
| NAMESPACE_END(Grid);  | NAMESPACE_END(Grid);  | ||||||
|   | |||||||
| @@ -1,5 +1,4 @@ | |||||||
| #include <Grid/GridCore.h>        | #include <Grid/GridCore.h>        | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
| std::vector<std::pair<int,int> > Cshift_table;  | Vector<std::pair<int,int> > Cshift_table;  | ||||||
| commVector<std::pair<int,int> > Cshift_table_device;  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|   | |||||||
							
								
								
									
										24231
									
								
								Grid/json/json.hpp
									
									
									
									
									
								
							
							
						
						
									
										24231
									
								
								Grid/json/json.hpp
									
									
									
									
									
								
							
										
											
												File diff suppressed because it is too large
												Load Diff
											
										
									
								
							| @@ -35,7 +35,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk> | |||||||
| #include <Grid/lattice/Lattice_transpose.h> | #include <Grid/lattice/Lattice_transpose.h> | ||||||
| #include <Grid/lattice/Lattice_local.h> | #include <Grid/lattice/Lattice_local.h> | ||||||
| #include <Grid/lattice/Lattice_reduction.h> | #include <Grid/lattice/Lattice_reduction.h> | ||||||
| #include <Grid/lattice/Lattice_crc.h> |  | ||||||
| #include <Grid/lattice/Lattice_peekpoke.h> | #include <Grid/lattice/Lattice_peekpoke.h> | ||||||
| #include <Grid/lattice/Lattice_reality.h> | #include <Grid/lattice/Lattice_reality.h> | ||||||
| #include <Grid/lattice/Lattice_real_imag.h> | #include <Grid/lattice/Lattice_real_imag.h> | ||||||
| @@ -47,4 +46,3 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk> | |||||||
| #include <Grid/lattice/Lattice_unary.h> | #include <Grid/lattice/Lattice_unary.h> | ||||||
| #include <Grid/lattice/Lattice_transfer.h> | #include <Grid/lattice/Lattice_transfer.h> | ||||||
| #include <Grid/lattice/Lattice_basis.h> | #include <Grid/lattice/Lattice_basis.h> | ||||||
| #include <Grid/lattice/PaddedCell.h> |  | ||||||
|   | |||||||
| @@ -63,7 +63,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate, | |||||||
|   typename std::remove_const<vobj>::type ret; |   typename std::remove_const<vobj>::type ret; | ||||||
|  |  | ||||||
|   typedef typename vobj::scalar_object scalar_object; |   typedef typename vobj::scalar_object scalar_object; | ||||||
|   //  typedef typename vobj::scalar_type scalar_type; |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_type vector_type; |   typedef typename vobj::vector_type vector_type; | ||||||
|  |  | ||||||
|   const int Nsimd = vobj::vector_type::Nsimd(); |   const int Nsimd = vobj::vector_type::Nsimd(); | ||||||
| @@ -345,9 +345,7 @@ GridUnopClass(UnaryNot, Not(a)); | |||||||
| GridUnopClass(UnaryTrace, trace(a)); | GridUnopClass(UnaryTrace, trace(a)); | ||||||
| GridUnopClass(UnaryTranspose, transpose(a)); | GridUnopClass(UnaryTranspose, transpose(a)); | ||||||
| GridUnopClass(UnaryTa, Ta(a)); | GridUnopClass(UnaryTa, Ta(a)); | ||||||
| GridUnopClass(UnarySpTa, SpTa(a)); |  | ||||||
| GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a)); | GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a)); | ||||||
| GridUnopClass(UnaryProjectOnSpGroup, ProjectOnSpGroup(a)); |  | ||||||
| GridUnopClass(UnaryTimesI, timesI(a)); | GridUnopClass(UnaryTimesI, timesI(a)); | ||||||
| GridUnopClass(UnaryTimesMinusI, timesMinusI(a)); | GridUnopClass(UnaryTimesMinusI, timesMinusI(a)); | ||||||
| GridUnopClass(UnaryAbs, abs(a)); | GridUnopClass(UnaryAbs, abs(a)); | ||||||
| @@ -458,9 +456,7 @@ GRID_DEF_UNOP(operator!, UnaryNot); | |||||||
| GRID_DEF_UNOP(trace, UnaryTrace); | GRID_DEF_UNOP(trace, UnaryTrace); | ||||||
| GRID_DEF_UNOP(transpose, UnaryTranspose); | GRID_DEF_UNOP(transpose, UnaryTranspose); | ||||||
| GRID_DEF_UNOP(Ta, UnaryTa); | GRID_DEF_UNOP(Ta, UnaryTa); | ||||||
| GRID_DEF_UNOP(SpTa, UnarySpTa); |  | ||||||
| GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup); | GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup); | ||||||
| GRID_DEF_UNOP(ProjectOnSpGroup, UnaryProjectOnSpGroup); |  | ||||||
| GRID_DEF_UNOP(timesI, UnaryTimesI); | GRID_DEF_UNOP(timesI, UnaryTimesI); | ||||||
| GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI); | GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI); | ||||||
| GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the | GRID_DEF_UNOP(abs, UnaryAbs);  // abs overloaded in cmath C++98; DON'T do the | ||||||
|   | |||||||
| @@ -36,7 +36,6 @@ NAMESPACE_BEGIN(Grid); | |||||||
| ////////////////////////////////////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | ||||||
|   GRID_TRACE("mult"); |  | ||||||
|   ret.Checkerboard() = lhs.Checkerboard(); |   ret.Checkerboard() = lhs.Checkerboard(); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
|   autoView( lhs_v , lhs, AcceleratorRead); |   autoView( lhs_v , lhs, AcceleratorRead); | ||||||
| @@ -54,7 +53,6 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | |||||||
|    |    | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | ||||||
|   GRID_TRACE("mac"); |  | ||||||
|   ret.Checkerboard() = lhs.Checkerboard(); |   ret.Checkerboard() = lhs.Checkerboard(); | ||||||
|   conformable(ret,rhs); |   conformable(ret,rhs); | ||||||
|   conformable(lhs,rhs); |   conformable(lhs,rhs); | ||||||
| @@ -72,7 +70,6 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | |||||||
|    |    | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | ||||||
|   GRID_TRACE("sub"); |  | ||||||
|   ret.Checkerboard() = lhs.Checkerboard(); |   ret.Checkerboard() = lhs.Checkerboard(); | ||||||
|   conformable(ret,rhs); |   conformable(ret,rhs); | ||||||
|   conformable(lhs,rhs); |   conformable(lhs,rhs); | ||||||
| @@ -89,7 +86,6 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | |||||||
| } | } | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | ||||||
|   GRID_TRACE("add"); |  | ||||||
|   ret.Checkerboard() = lhs.Checkerboard(); |   ret.Checkerboard() = lhs.Checkerboard(); | ||||||
|   conformable(ret,rhs); |   conformable(ret,rhs); | ||||||
|   conformable(lhs,rhs); |   conformable(lhs,rhs); | ||||||
| @@ -110,7 +106,6 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){ | |||||||
| ////////////////////////////////////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | ||||||
|   GRID_TRACE("mult"); |  | ||||||
|   ret.Checkerboard() = lhs.Checkerboard(); |   ret.Checkerboard() = lhs.Checkerboard(); | ||||||
|   conformable(lhs,ret); |   conformable(lhs,ret); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
| @@ -124,7 +119,6 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | |||||||
|    |    | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | ||||||
|   GRID_TRACE("mac"); |  | ||||||
|   ret.Checkerboard() = lhs.Checkerboard(); |   ret.Checkerboard() = lhs.Checkerboard(); | ||||||
|   conformable(ret,lhs); |   conformable(ret,lhs); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
| @@ -139,7 +133,6 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | |||||||
|    |    | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | ||||||
|   GRID_TRACE("sub"); |  | ||||||
|   ret.Checkerboard() = lhs.Checkerboard(); |   ret.Checkerboard() = lhs.Checkerboard(); | ||||||
|   conformable(ret,lhs); |   conformable(ret,lhs); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
| @@ -153,7 +146,6 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | |||||||
| } | } | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | ||||||
|   GRID_TRACE("add"); |  | ||||||
|   ret.Checkerboard() = lhs.Checkerboard(); |   ret.Checkerboard() = lhs.Checkerboard(); | ||||||
|   conformable(lhs,ret); |   conformable(lhs,ret); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
| @@ -171,7 +163,6 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){ | |||||||
| ////////////////////////////////////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | ||||||
|   GRID_TRACE("mult"); |  | ||||||
|   ret.Checkerboard() = rhs.Checkerboard(); |   ret.Checkerboard() = rhs.Checkerboard(); | ||||||
|   conformable(ret,rhs); |   conformable(ret,rhs); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
| @@ -186,7 +177,6 @@ void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | |||||||
|    |    | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | ||||||
|   GRID_TRACE("mac"); |  | ||||||
|   ret.Checkerboard() = rhs.Checkerboard(); |   ret.Checkerboard() = rhs.Checkerboard(); | ||||||
|   conformable(ret,rhs); |   conformable(ret,rhs); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
| @@ -201,7 +191,6 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | |||||||
|    |    | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | ||||||
|   GRID_TRACE("sub"); |  | ||||||
|   ret.Checkerboard() = rhs.Checkerboard(); |   ret.Checkerboard() = rhs.Checkerboard(); | ||||||
|   conformable(ret,rhs); |   conformable(ret,rhs); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
| @@ -215,7 +204,6 @@ void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | |||||||
| } | } | ||||||
| template<class obj1,class obj2,class obj3> inline | template<class obj1,class obj2,class obj3> inline | ||||||
| void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | ||||||
|   GRID_TRACE("add"); |  | ||||||
|   ret.Checkerboard() = rhs.Checkerboard(); |   ret.Checkerboard() = rhs.Checkerboard(); | ||||||
|   conformable(ret,rhs); |   conformable(ret,rhs); | ||||||
|   autoView( ret_v , ret, AcceleratorWrite); |   autoView( ret_v , ret, AcceleratorWrite); | ||||||
| @@ -230,7 +218,6 @@ void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){ | |||||||
|    |    | ||||||
| template<class sobj,class vobj> inline | template<class sobj,class vobj> inline | ||||||
| void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){ | void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){ | ||||||
|   GRID_TRACE("axpy"); |  | ||||||
|   ret.Checkerboard() = x.Checkerboard(); |   ret.Checkerboard() = x.Checkerboard(); | ||||||
|   conformable(ret,x); |   conformable(ret,x); | ||||||
|   conformable(x,y); |   conformable(x,y); | ||||||
| @@ -238,13 +225,12 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> & | |||||||
|   autoView( x_v , x, AcceleratorRead); |   autoView( x_v , x, AcceleratorRead); | ||||||
|   autoView( y_v , y, AcceleratorRead); |   autoView( y_v , y, AcceleratorRead); | ||||||
|   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{ |   accelerator_for(ss,x_v.size(),vobj::Nsimd(),{ | ||||||
|     auto tmp = a*coalescedRead(x_v[ss])+coalescedRead(y_v[ss]); |     auto tmp = a*x_v(ss)+y_v(ss); | ||||||
|     coalescedWrite(ret_v[ss],tmp); |     coalescedWrite(ret_v[ss],tmp); | ||||||
|   }); |   }); | ||||||
| } | } | ||||||
| template<class sobj,class vobj> inline | template<class sobj,class vobj> inline | ||||||
| void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){ | void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){ | ||||||
|   GRID_TRACE("axpby"); |  | ||||||
|   ret.Checkerboard() = x.Checkerboard(); |   ret.Checkerboard() = x.Checkerboard(); | ||||||
|   conformable(ret,x); |   conformable(ret,x); | ||||||
|   conformable(x,y); |   conformable(x,y); | ||||||
| @@ -260,52 +246,13 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice | |||||||
| template<class sobj,class vobj> inline | template<class sobj,class vobj> inline | ||||||
| RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y) | RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y) | ||||||
| { | { | ||||||
|   GRID_TRACE("axpy_norm"); |  | ||||||
|     return axpy_norm_fast(ret,a,x,y); |     return axpy_norm_fast(ret,a,x,y); | ||||||
| } | } | ||||||
| template<class sobj,class vobj> inline | template<class sobj,class vobj> inline | ||||||
| RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y) | RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y) | ||||||
| { | { | ||||||
|   GRID_TRACE("axpby_norm"); |  | ||||||
|     return axpby_norm_fast(ret,a,b,x,y); |     return axpby_norm_fast(ret,a,b,x,y); | ||||||
| } | } | ||||||
|  |  | ||||||
| /// Trace product |  | ||||||
| template<class obj> auto traceProduct(const Lattice<obj> &rhs_1,const Lattice<obj> &rhs_2) |  | ||||||
|   -> Lattice<decltype(trace(obj()))> |  | ||||||
| { |  | ||||||
|   typedef decltype(trace(obj())) robj; |  | ||||||
|   Lattice<robj> ret_i(rhs_1.Grid()); |  | ||||||
|   autoView( rhs1 , rhs_1, AcceleratorRead); |  | ||||||
|   autoView( rhs2 , rhs_2, AcceleratorRead); |  | ||||||
|   autoView( ret , ret_i, AcceleratorWrite); |  | ||||||
|   ret.Checkerboard() = rhs_1.Checkerboard(); |  | ||||||
|   accelerator_for(ss,rhs1.size(),obj::Nsimd(),{ |  | ||||||
|       coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2(ss))); |  | ||||||
|   }); |  | ||||||
|   return ret_i; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class obj1,class obj2> auto traceProduct(const Lattice<obj1> &rhs_1,const obj2 &rhs2) |  | ||||||
|   -> Lattice<decltype(trace(obj1()))> |  | ||||||
| { |  | ||||||
|   typedef decltype(trace(obj1())) robj; |  | ||||||
|   Lattice<robj> ret_i(rhs_1.Grid()); |  | ||||||
|   autoView( rhs1 , rhs_1, AcceleratorRead); |  | ||||||
|   autoView( ret , ret_i, AcceleratorWrite); |  | ||||||
|   ret.Checkerboard() = rhs_1.Checkerboard(); |  | ||||||
|   accelerator_for(ss,rhs1.size(),obj1::Nsimd(),{ |  | ||||||
|       coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2)); |  | ||||||
|   }); |  | ||||||
|   return ret_i; |  | ||||||
| } |  | ||||||
| template<class obj1,class obj2> auto traceProduct(const obj2 &rhs_2,const Lattice<obj1> &rhs_1) |  | ||||||
|   -> Lattice<decltype(trace(obj1()))> |  | ||||||
| { |  | ||||||
|   return traceProduct(rhs_1,rhs_2); |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
| #endif | #endif | ||||||
|   | |||||||
| @@ -88,13 +88,6 @@ public: | |||||||
|     LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode); |     LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode); | ||||||
|     accessor.ViewClose(); |     accessor.ViewClose(); | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   // Helper function to print the state of this object in the AccCache |  | ||||||
|   void PrintCacheState(void) |  | ||||||
|   { |  | ||||||
|     MemoryManager::PrintState(this->_odata); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////////////////////////////////////////// |   ///////////////////////////////////////////////////////////////////////////////// | ||||||
|   // Return a view object that may be dereferenced in site loops. |   // Return a view object that may be dereferenced in site loops. | ||||||
|   // The view is trivially copy constructible and may be copied to an accelerator device |   // The view is trivially copy constructible and may be copied to an accelerator device | ||||||
| @@ -117,7 +110,6 @@ public: | |||||||
|   //////////////////////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////////////////////// | ||||||
|   template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr) |   template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr) | ||||||
|   { |   { | ||||||
|     GRID_TRACE("ExpressionTemplateEval"); |  | ||||||
|     GridBase *egrid(nullptr); |     GridBase *egrid(nullptr); | ||||||
|     GridFromExpression(egrid,expr); |     GridFromExpression(egrid,expr); | ||||||
|     assert(egrid!=nullptr); |     assert(egrid!=nullptr); | ||||||
| @@ -141,7 +133,6 @@ public: | |||||||
|   } |   } | ||||||
|   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr) |   template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr) | ||||||
|   { |   { | ||||||
|     GRID_TRACE("ExpressionTemplateEval"); |  | ||||||
|     GridBase *egrid(nullptr); |     GridBase *egrid(nullptr); | ||||||
|     GridFromExpression(egrid,expr); |     GridFromExpression(egrid,expr); | ||||||
|     assert(egrid!=nullptr); |     assert(egrid!=nullptr); | ||||||
| @@ -165,7 +156,6 @@ public: | |||||||
|   } |   } | ||||||
|   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr) |   template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr) | ||||||
|   { |   { | ||||||
|     GRID_TRACE("ExpressionTemplateEval"); |  | ||||||
|     GridBase *egrid(nullptr); |     GridBase *egrid(nullptr); | ||||||
|     GridFromExpression(egrid,expr); |     GridFromExpression(egrid,expr); | ||||||
|     assert(egrid!=nullptr); |     assert(egrid!=nullptr); | ||||||
| @@ -291,8 +281,8 @@ public: | |||||||
|     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0; |     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0; | ||||||
|     conformable(*this,r); |     conformable(*this,r); | ||||||
|     this->checkerboard = r.Checkerboard(); |     this->checkerboard = r.Checkerboard(); | ||||||
|     auto him= r.View(AcceleratorRead); |  | ||||||
|     auto me =   View(AcceleratorWriteDiscard); |     auto me =   View(AcceleratorWriteDiscard); | ||||||
|  |     auto him= r.View(AcceleratorRead); | ||||||
|     accelerator_for(ss,me.size(),vobj::Nsimd(),{ |     accelerator_for(ss,me.size(),vobj::Nsimd(),{ | ||||||
|       coalescedWrite(me[ss],him(ss)); |       coalescedWrite(me[ss],him(ss)); | ||||||
|     }); |     }); | ||||||
| @@ -306,8 +296,8 @@ public: | |||||||
|   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){ |   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){ | ||||||
|     this->checkerboard = r.Checkerboard(); |     this->checkerboard = r.Checkerboard(); | ||||||
|     conformable(*this,r); |     conformable(*this,r); | ||||||
|     auto him= r.View(AcceleratorRead); |  | ||||||
|     auto me =   View(AcceleratorWriteDiscard); |     auto me =   View(AcceleratorWriteDiscard); | ||||||
|  |     auto him= r.View(AcceleratorRead); | ||||||
|     accelerator_for(ss,me.size(),vobj::Nsimd(),{ |     accelerator_for(ss,me.size(),vobj::Nsimd(),{ | ||||||
|       coalescedWrite(me[ss],him(ss)); |       coalescedWrite(me[ss],him(ss)); | ||||||
|     }); |     }); | ||||||
|   | |||||||
| @@ -62,7 +62,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm) | |||||||
|     basis_v.push_back(basis[k].View(AcceleratorWrite)); |     basis_v.push_back(basis[k].View(AcceleratorWrite)); | ||||||
|   } |   } | ||||||
|  |  | ||||||
| #if ( !(defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)) ) | #if ( (!defined(GRID_SYCL)) && (!defined(GRID_CUDA)) ) | ||||||
|   int max_threads = thread_max(); |   int max_threads = thread_max(); | ||||||
|   Vector < vobj > Bt(Nm * max_threads); |   Vector < vobj > Bt(Nm * max_threads); | ||||||
|   thread_region |   thread_region | ||||||
| @@ -125,7 +125,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm) | |||||||
|  |  | ||||||
| 	for(int k=k0; k<k1; ++k){ | 	for(int k=k0; k<k1; ++k){ | ||||||
| 	  auto tmp = coalescedRead(Bp[ss*nrot+j]); | 	  auto tmp = coalescedRead(Bp[ss*nrot+j]); | ||||||
| 	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_vp[k][sss])); | 	  coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_v[k][sss])); | ||||||
| 	} | 	} | ||||||
|       }); |       }); | ||||||
|  |  | ||||||
| @@ -134,7 +134,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm) | |||||||
| 	int jj  =j0+j; | 	int jj  =j0+j; | ||||||
| 	int ss =sj/nrot; | 	int ss =sj/nrot; | ||||||
| 	int sss=ss+s; | 	int sss=ss+s; | ||||||
| 	coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j])); | 	coalescedWrite(basis_v[jj][sss],coalescedRead(Bp[ss*nrot+j])); | ||||||
|       }); |       }); | ||||||
|   } |   } | ||||||
| #endif | #endif | ||||||
| @@ -164,8 +164,7 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in | |||||||
|   auto basis_vp=& basis_v[0]; |   auto basis_vp=& basis_v[0]; | ||||||
|   autoView(result_v,result,AcceleratorWrite); |   autoView(result_v,result,AcceleratorWrite); | ||||||
|   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{ |   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{ | ||||||
|     vobj zzz=Zero(); |     auto B=coalescedRead(zz); | ||||||
|     auto B=coalescedRead(zzz); |  | ||||||
|     for(int k=k0; k<k1; ++k){ |     for(int k=k0; k<k1; ++k){ | ||||||
|       B +=Qt_j[k] * coalescedRead(basis_vp[k][ss]); |       B +=Qt_j[k] * coalescedRead(basis_vp[k][ss]); | ||||||
|     } |     } | ||||||
|   | |||||||
| @@ -1,55 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid  |  | ||||||
|  |  | ||||||
|     Source file: ./lib/lattice/Lattice_crc.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2021 |  | ||||||
|  |  | ||||||
| Author: Peter Boyle <paboyle@ph.ed.ac.uk> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
| *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
| #pragma once |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1) |  | ||||||
| { |  | ||||||
|   auto ff = localNorm2(f); |  | ||||||
|   if ( mu==-1 ) mu = f.Grid()->Nd()-1; |  | ||||||
|   typedef typename vobj::tensor_reduced normtype; |  | ||||||
|   typedef typename normtype::scalar_object scalar; |  | ||||||
|   std::vector<scalar> sff; |  | ||||||
|   sliceSum(ff,sff,mu); |  | ||||||
|   for(int t=0;t<sff.size();t++){ |  | ||||||
|     std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl; |  | ||||||
|   } |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class vobj> uint32_t crc(const Lattice<vobj> & buf) |  | ||||||
| { |  | ||||||
|   autoView( buf_v , buf, CpuRead); |  | ||||||
|   return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites()); |  | ||||||
| } |  | ||||||
|  |  | ||||||
| #define CRC(U) std::cerr << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl; |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
|  |  | ||||||
|  |  | ||||||
| @@ -32,6 +32,7 @@ template<class vobj> | |||||||
| static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)  | static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)  | ||||||
| {     | {     | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_type vector_type; |   typedef typename vobj::vector_type vector_type; | ||||||
|  |  | ||||||
|   int Nblock = X.Grid()->GlobalDimensions()[Orthog]; |   int Nblock = X.Grid()->GlobalDimensions()[Orthog]; | ||||||
| @@ -81,6 +82,7 @@ template<class vobj> | |||||||
| static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0)  | static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0)  | ||||||
| {     | {     | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_type vector_type; |   typedef typename vobj::vector_type vector_type; | ||||||
|  |  | ||||||
|   int Nblock = X.Grid()->GlobalDimensions()[Orthog]; |   int Nblock = X.Grid()->GlobalDimensions()[Orthog]; | ||||||
| @@ -128,6 +130,7 @@ template<class vobj> | |||||||
| static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)  | static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)  | ||||||
| { | { | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_type vector_type; |   typedef typename vobj::vector_type vector_type; | ||||||
|    |    | ||||||
|   GridBase *FullGrid  = lhs.Grid(); |   GridBase *FullGrid  = lhs.Grid(); | ||||||
|   | |||||||
| @@ -96,6 +96,9 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){ | |||||||
|  |  | ||||||
|   GridBase *grid=l.Grid(); |   GridBase *grid=l.Grid(); | ||||||
|  |  | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|  |   typedef typename vobj::vector_type vector_type; | ||||||
|  |  | ||||||
|   int Nsimd = grid->Nsimd(); |   int Nsimd = grid->Nsimd(); | ||||||
|  |  | ||||||
|   assert( l.Checkerboard()== l.Grid()->CheckerBoard(site)); |   assert( l.Checkerboard()== l.Grid()->CheckerBoard(site)); | ||||||
| @@ -122,17 +125,14 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){ | |||||||
| ////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////// | ||||||
| // Peek a scalar object from the SIMD array | // Peek a scalar object from the SIMD array | ||||||
| ////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////// | ||||||
| template<class vobj> |  | ||||||
| typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){ |  | ||||||
|   typename vobj::scalar_object s; |  | ||||||
|   peekSite(s,l,site); |  | ||||||
|   return s; |  | ||||||
| }         |  | ||||||
| template<class vobj,class sobj> | template<class vobj,class sobj> | ||||||
| void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){ | void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){ | ||||||
|          |          | ||||||
|   GridBase *grid=l.Grid(); |   GridBase *grid=l.Grid(); | ||||||
|  |  | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|  |   typedef typename vobj::vector_type vector_type; | ||||||
|  |  | ||||||
|   int Nsimd = grid->Nsimd(); |   int Nsimd = grid->Nsimd(); | ||||||
|  |  | ||||||
|   assert( l.Checkerboard() == l.Grid()->CheckerBoard(site)); |   assert( l.Checkerboard() == l.Grid()->CheckerBoard(site)); | ||||||
| @@ -173,11 +173,11 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site) | |||||||
|   idx= grid->iIndex(site); |   idx= grid->iIndex(site); | ||||||
|   odx= grid->oIndex(site); |   odx= grid->oIndex(site); | ||||||
|    |    | ||||||
|   const vector_type *vp = (const vector_type *) &l[odx]; |   scalar_type * vp = (scalar_type *)&l[odx]; | ||||||
|   scalar_type * pt = (scalar_type *)&s; |   scalar_type * pt = (scalar_type *)&s; | ||||||
|        |        | ||||||
|   for(int w=0;w<words;w++){ |   for(int w=0;w<words;w++){ | ||||||
|     pt[w] = getlane(vp[w],idx); |     pt[w] = vp[idx+w*Nsimd]; | ||||||
|   } |   } | ||||||
|        |        | ||||||
|   return; |   return; | ||||||
| @@ -210,10 +210,10 @@ inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site) | |||||||
|   idx= grid->iIndex(site); |   idx= grid->iIndex(site); | ||||||
|   odx= grid->oIndex(site); |   odx= grid->oIndex(site); | ||||||
|  |  | ||||||
|   vector_type * vp = (vector_type *)&l[odx]; |   scalar_type * vp = (scalar_type *)&l[odx]; | ||||||
|   scalar_type * pt = (scalar_type *)&s; |   scalar_type * pt = (scalar_type *)&s; | ||||||
|   for(int w=0;w<words;w++){ |   for(int w=0;w<words;w++){ | ||||||
|     putlane(vp[w],pt[w],idx); |     vp[idx+w*Nsimd] = pt[w]; | ||||||
|   } |   } | ||||||
|   return; |   return; | ||||||
| }; | }; | ||||||
|   | |||||||
| @@ -28,10 +28,6 @@ Author: Christoph Lehner <christoph@lhnr.de> | |||||||
| #if defined(GRID_CUDA)||defined(GRID_HIP) | #if defined(GRID_CUDA)||defined(GRID_HIP) | ||||||
| #include <Grid/lattice/Lattice_reduction_gpu.h> | #include <Grid/lattice/Lattice_reduction_gpu.h> | ||||||
| #endif | #endif | ||||||
| #if defined(GRID_SYCL) |  | ||||||
| #include <Grid/lattice/Lattice_reduction_sycl.h> |  | ||||||
| #endif |  | ||||||
| #include <Grid/lattice/Lattice_slicesum_core.h> |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| @@ -95,40 +91,17 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites) | |||||||
|   for(int i=0;i<nthread;i++){ |   for(int i=0;i<nthread;i++){ | ||||||
|     ssum = ssum+sumarray[i]; |     ssum = ssum+sumarray[i]; | ||||||
|   }  |   }  | ||||||
|   return ssum; |    | ||||||
|  |   typedef typename vobj::scalar_object ssobj; | ||||||
|  |   ssobj ret = ssum; | ||||||
|  |   return ret; | ||||||
| } | } | ||||||
| /* |  | ||||||
| Threaded max, don't use for now |  | ||||||
| template<class Double> |  | ||||||
| inline Double max(const Double *arg, Integer osites) |  | ||||||
| { |  | ||||||
|   //  const int Nsimd = vobj::Nsimd(); |  | ||||||
|   const int nthread = GridThread::GetThreads(); |  | ||||||
|  |  | ||||||
|   std::vector<Double> maxarray(nthread); |  | ||||||
|    |  | ||||||
|   thread_for(thr,nthread, { |  | ||||||
|     int nwork, mywork, myoff; |  | ||||||
|     nwork = osites; |  | ||||||
|     GridThread::GetWork(nwork,thr,mywork,myoff); |  | ||||||
|     Double max=arg[0]; |  | ||||||
|     for(int ss=myoff;ss<mywork+myoff; ss++){ |  | ||||||
|       if( arg[ss] > max ) max = arg[ss]; |  | ||||||
|     } |  | ||||||
|     maxarray[thr]=max; |  | ||||||
|   }); |  | ||||||
|    |  | ||||||
|   Double tmax=maxarray[0]; |  | ||||||
|   for(int i=0;i<nthread;i++){ |  | ||||||
|     if (maxarray[i]>tmax) tmax = maxarray[i]; |  | ||||||
|   }  |  | ||||||
|   return tmax; |  | ||||||
| } |  | ||||||
| */ |  | ||||||
| template<class vobj> | template<class vobj> | ||||||
| inline typename vobj::scalar_object sum(const vobj *arg, Integer osites) | inline typename vobj::scalar_object sum(const vobj *arg, Integer osites) | ||||||
| { | { | ||||||
| #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL) | #if defined(GRID_CUDA)||defined(GRID_HIP) | ||||||
|   return sum_gpu(arg,osites); |   return sum_gpu(arg,osites); | ||||||
| #else | #else | ||||||
|   return sum_cpu(arg,osites); |   return sum_cpu(arg,osites); | ||||||
| @@ -137,61 +110,25 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites) | |||||||
| template<class vobj> | template<class vobj> | ||||||
| inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites) | inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites) | ||||||
| { | { | ||||||
| #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL) | #if defined(GRID_CUDA)||defined(GRID_HIP) | ||||||
|   return sumD_gpu(arg,osites); |   return sumD_gpu(arg,osites); | ||||||
| #else | #else | ||||||
|   return sumD_cpu(arg,osites); |   return sumD_cpu(arg,osites); | ||||||
| #endif   | #endif   | ||||||
| } | } | ||||||
| template<class vobj> |  | ||||||
| inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites) |  | ||||||
| { |  | ||||||
| #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL) |  | ||||||
|   return sumD_gpu_large(arg,osites); |  | ||||||
| #else |  | ||||||
|   return sumD_cpu(arg,osites); |  | ||||||
| #endif   |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class vobj> |  | ||||||
| inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg) |  | ||||||
| { |  | ||||||
|   Integer osites = arg.Grid()->oSites(); |  | ||||||
| #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL) |  | ||||||
|   autoView( arg_v, arg, AcceleratorRead); |  | ||||||
|   return sum_gpu(&arg_v[0],osites); |  | ||||||
| #else |  | ||||||
|   autoView(arg_v, arg, CpuRead); |  | ||||||
|   return sum_cpu(&arg_v[0],osites); |  | ||||||
| #endif   |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class vobj> | template<class vobj> | ||||||
| inline typename vobj::scalar_object sum(const Lattice<vobj> &arg) | inline typename vobj::scalar_object sum(const Lattice<vobj> &arg) | ||||||
| { | { | ||||||
|   auto ssum = rankSum(arg); | #if defined(GRID_CUDA)||defined(GRID_HIP) | ||||||
|   arg.Grid()->GlobalSum(ssum); |  | ||||||
|   return ssum; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class vobj> |  | ||||||
| inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg) |  | ||||||
| { |  | ||||||
| #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL) |  | ||||||
|   autoView( arg_v, arg, AcceleratorRead); |   autoView( arg_v, arg, AcceleratorRead); | ||||||
|   Integer osites = arg.Grid()->oSites(); |   Integer osites = arg.Grid()->oSites(); | ||||||
|   return sum_gpu_large(&arg_v[0],osites); |   auto ssum= sum_gpu(&arg_v[0],osites); | ||||||
| #else | #else | ||||||
|   autoView(arg_v, arg, CpuRead); |   autoView(arg_v, arg, CpuRead); | ||||||
|   Integer osites = arg.Grid()->oSites(); |   Integer osites = arg.Grid()->oSites(); | ||||||
|   return sum_cpu(&arg_v[0],osites); |   auto ssum= sum_cpu(&arg_v[0],osites); | ||||||
| #endif | #endif   | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class vobj> |  | ||||||
| inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg) |  | ||||||
| { |  | ||||||
|   auto ssum = rankSumLarge(arg); |  | ||||||
|   arg.Grid()->GlobalSum(ssum); |   arg.Grid()->GlobalSum(ssum); | ||||||
|   return ssum; |   return ssum; | ||||||
| } | } | ||||||
| @@ -204,36 +141,11 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){ | |||||||
|   return real(nrm);  |   return real(nrm);  | ||||||
| } | } | ||||||
|  |  | ||||||
| //The global maximum of the site norm2 |  | ||||||
| template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg) |  | ||||||
| { |  | ||||||
|   typedef typename vobj::tensor_reduced vscalar;  //iScalar<iScalar<.... <vPODtype> > > |  | ||||||
|   typedef typename vscalar::scalar_object  scalar;   //iScalar<iScalar<.... <PODtype> > > |  | ||||||
|  |  | ||||||
|   Lattice<vscalar> inner = localNorm2(arg); |  | ||||||
|  |  | ||||||
|   auto grid = arg.Grid(); |  | ||||||
|  |  | ||||||
|   RealD max; |  | ||||||
|   for(int l=0;l<grid->lSites();l++){ |  | ||||||
|     Coordinate coor; |  | ||||||
|     scalar val; |  | ||||||
|     RealD r; |  | ||||||
|     grid->LocalIndexToLocalCoor(l,coor); |  | ||||||
|     peekLocalSite(val,inner,coor); |  | ||||||
|     r=real(TensorRemove(val)); |  | ||||||
|     if( (l==0) || (r>max)){ |  | ||||||
|       max=r; |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|   grid->GlobalMax(max); |  | ||||||
|   return max; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| // Double inner product | // Double inner product | ||||||
| template<class vobj> | template<class vobj> | ||||||
| inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) | inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) | ||||||
| { | { | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_typeD vector_type; |   typedef typename vobj::vector_typeD vector_type; | ||||||
|   ComplexD  nrm; |   ComplexD  nrm; | ||||||
|    |    | ||||||
| @@ -243,67 +155,33 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> & | |||||||
|   const uint64_t sites = grid->oSites(); |   const uint64_t sites = grid->oSites(); | ||||||
|    |    | ||||||
|   // Might make all code paths go this way. |   // Might make all code paths go this way. | ||||||
| #if 0 |  | ||||||
|   typedef decltype(innerProductD(vobj(),vobj())) inner_t; |   typedef decltype(innerProductD(vobj(),vobj())) inner_t; | ||||||
|   Vector<inner_t> inner_tmp(sites); |   Vector<inner_t> inner_tmp(sites); | ||||||
|   auto inner_tmp_v = &inner_tmp[0]; |   auto inner_tmp_v = &inner_tmp[0]; | ||||||
|   { |  | ||||||
|     autoView( left_v , left, AcceleratorRead); |  | ||||||
|     autoView( right_v,right, AcceleratorRead); |  | ||||||
|     // This code could read coalesce |  | ||||||
|     // GPU - SIMT lane compliance... |  | ||||||
|     accelerator_for( ss, sites, nsimd,{ |  | ||||||
| 	auto x_l = left_v(ss); |  | ||||||
| 	auto y_l = right_v(ss); |  | ||||||
| 	coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l)); |  | ||||||
|     }); |  | ||||||
|   } |  | ||||||
| #else |  | ||||||
|   typedef decltype(innerProduct(vobj(),vobj())) inner_t; |  | ||||||
|   Vector<inner_t> inner_tmp(sites); |  | ||||||
|   auto inner_tmp_v = &inner_tmp[0]; |  | ||||||
|      |      | ||||||
|   { |   { | ||||||
|     autoView( left_v , left, AcceleratorRead); |     autoView( left_v , left, AcceleratorRead); | ||||||
|     autoView( right_v,right, AcceleratorRead); |     autoView( right_v,right, AcceleratorRead); | ||||||
|  |  | ||||||
|     // GPU - SIMT lane compliance... |     // GPU - SIMT lane compliance... | ||||||
|     accelerator_for( ss, sites, nsimd,{ |     accelerator_for( ss, sites, 1,{ | ||||||
| 	auto x_l = left_v(ss); | 	auto x_l = left_v[ss]; | ||||||
| 	auto y_l = right_v(ss); | 	auto y_l = right_v[ss]; | ||||||
| 	coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l)); | 	inner_tmp_v[ss]=innerProductD(x_l,y_l); | ||||||
|     }); |     }); | ||||||
|   } |   } | ||||||
| #endif |  | ||||||
|   // This is in single precision and fails some tests |   // This is in single precision and fails some tests | ||||||
|   auto anrm = sumD(inner_tmp_v,sites);   |   auto anrm = sum(inner_tmp_v,sites);   | ||||||
|   nrm = anrm; |   nrm = anrm; | ||||||
|   return nrm; |   return nrm; | ||||||
| } | } | ||||||
|  |  | ||||||
|  |  | ||||||
| template<class vobj> | template<class vobj> | ||||||
| inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) { | inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) { | ||||||
|   GridBase *grid = left.Grid(); |   GridBase *grid = left.Grid(); | ||||||
|  |  | ||||||
| #ifdef GRID_SYCL |  | ||||||
|   uint64_t csum=0; |  | ||||||
|   if ( FlightRecorder::LoggingMode != FlightRecorder::LoggingModeNone) |  | ||||||
|   { |  | ||||||
|     // Hack |  | ||||||
|     // Fast integer xor checksum. Can also be used in comms now. |  | ||||||
|     autoView(l_v,left,AcceleratorRead); |  | ||||||
|     Integer words = left.Grid()->oSites()*sizeof(vobj)/sizeof(uint64_t); |  | ||||||
|     uint64_t *base= (uint64_t *)&l_v[0]; |  | ||||||
|     csum=svm_xor(base,words); |  | ||||||
|   } |  | ||||||
|   FlightRecorder::CsumLog(csum); |  | ||||||
| #endif |  | ||||||
|   ComplexD nrm = rankInnerProduct(left,right); |   ComplexD nrm = rankInnerProduct(left,right); | ||||||
|   RealD local = real(nrm); |  | ||||||
|   FlightRecorder::NormLog(real(nrm));  |  | ||||||
|   grid->GlobalSum(nrm); |   grid->GlobalSum(nrm); | ||||||
|   FlightRecorder::ReductionLog(local,real(nrm));  |  | ||||||
|   return nrm; |   return nrm; | ||||||
| } | } | ||||||
|  |  | ||||||
| @@ -327,7 +205,8 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt | |||||||
|   conformable(z,x); |   conformable(z,x); | ||||||
|   conformable(x,y); |   conformable(x,y); | ||||||
|  |  | ||||||
|   //  typedef typename vobj::vector_typeD vector_type; |   typedef typename vobj::scalar_type scalar_type; | ||||||
|  |   typedef typename vobj::vector_typeD vector_type; | ||||||
|   RealD  nrm; |   RealD  nrm; | ||||||
|    |    | ||||||
|   GridBase *grid = x.Grid(); |   GridBase *grid = x.Grid(); | ||||||
| @@ -339,29 +218,17 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt | |||||||
|   autoView( x_v, x, AcceleratorRead); |   autoView( x_v, x, AcceleratorRead); | ||||||
|   autoView( y_v, y, AcceleratorRead); |   autoView( y_v, y, AcceleratorRead); | ||||||
|   autoView( z_v, z, AcceleratorWrite); |   autoView( z_v, z, AcceleratorWrite); | ||||||
| #if 0 |  | ||||||
|   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t; |   typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t; | ||||||
|   Vector<inner_t> inner_tmp(sites); |   Vector<inner_t> inner_tmp(sites); | ||||||
|   auto inner_tmp_v = &inner_tmp[0]; |   auto inner_tmp_v = &inner_tmp[0]; | ||||||
|  |  | ||||||
|   accelerator_for( ss, sites, nsimd,{ |   accelerator_for( ss, sites, 1,{ | ||||||
|       auto tmp = a*x_v(ss)+b*y_v(ss); |       auto tmp = a*x_v[ss]+b*y_v[ss]; | ||||||
|       coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp)); |       inner_tmp_v[ss]=innerProductD(tmp,tmp); | ||||||
|       coalescedWrite(z_v[ss],tmp); |       z_v[ss]=tmp; | ||||||
|   }); |   }); | ||||||
|   nrm = real(TensorRemove(sum(inner_tmp_v,sites))); |   nrm = real(TensorRemove(sum(inner_tmp_v,sites))); | ||||||
| #else |  | ||||||
|   typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t; |  | ||||||
|   Vector<inner_t> inner_tmp(sites); |  | ||||||
|   auto inner_tmp_v = &inner_tmp[0]; |  | ||||||
|  |  | ||||||
|   accelerator_for( ss, sites, nsimd,{ |  | ||||||
|       auto tmp = a*x_v(ss)+b*y_v(ss); |  | ||||||
|       coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp)); |  | ||||||
|       coalescedWrite(z_v[ss],tmp); |  | ||||||
|   }); |  | ||||||
|   nrm = real(TensorRemove(sumD(inner_tmp_v,sites))); |  | ||||||
| #endif |  | ||||||
|   grid->GlobalSum(nrm); |   grid->GlobalSum(nrm); | ||||||
|   return nrm;  |   return nrm;  | ||||||
| } | } | ||||||
| @@ -371,6 +238,7 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti | |||||||
| { | { | ||||||
|   conformable(left,right); |   conformable(left,right); | ||||||
|  |  | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_typeD vector_type; |   typedef typename vobj::vector_typeD vector_type; | ||||||
|   Vector<ComplexD> tmp(2); |   Vector<ComplexD> tmp(2); | ||||||
|  |  | ||||||
| @@ -441,7 +309,6 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector< | |||||||
|   // But easily avoided by using double precision fields |   // But easily avoided by using double precision fields | ||||||
|   /////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////// | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|   typedef typename vobj::scalar_object::scalar_type scalar_type; |  | ||||||
|   GridBase  *grid = Data.Grid(); |   GridBase  *grid = Data.Grid(); | ||||||
|   assert(grid!=NULL); |   assert(grid!=NULL); | ||||||
|  |  | ||||||
| @@ -467,10 +334,19 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector< | |||||||
|   int e1=    grid->_slice_nblock[orthogdim]; |   int e1=    grid->_slice_nblock[orthogdim]; | ||||||
|   int e2=    grid->_slice_block [orthogdim]; |   int e2=    grid->_slice_block [orthogdim]; | ||||||
|   int stride=grid->_slice_stride[orthogdim]; |   int stride=grid->_slice_stride[orthogdim]; | ||||||
|   int ostride=grid->_ostride[orthogdim]; |  | ||||||
|    |   // sum over reduced dimension planes, breaking out orthog dir | ||||||
|   //Reduce Data down to lvSum |   // Parallel over orthog direction | ||||||
|   sliceSumReduction(Data,lvSum,rd, e1,e2,stride,ostride,Nsimd); |   autoView( Data_v, Data, CpuRead); | ||||||
|  |   thread_for( r,rd, { | ||||||
|  |     int so=r*grid->_ostride[orthogdim]; // base offset for start of plane  | ||||||
|  |     for(int n=0;n<e1;n++){ | ||||||
|  |       for(int b=0;b<e2;b++){ | ||||||
|  | 	int ss= so+n*stride+b; | ||||||
|  | 	lvSum[r]=lvSum[r]+Data_v[ss]; | ||||||
|  |       } | ||||||
|  |     } | ||||||
|  |   }); | ||||||
|  |  | ||||||
|   // Sum across simd lanes in the plane, breaking out orthog dir. |   // Sum across simd lanes in the plane, breaking out orthog dir. | ||||||
|   Coordinate icoor(Nd); |   Coordinate icoor(Nd); | ||||||
| @@ -491,29 +367,21 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector< | |||||||
|   } |   } | ||||||
|    |    | ||||||
|   // sum over nodes. |   // sum over nodes. | ||||||
|  |   sobj gsum; | ||||||
|   for(int t=0;t<fd;t++){ |   for(int t=0;t<fd;t++){ | ||||||
|     int pt = t/ld; // processor plane |     int pt = t/ld; // processor plane | ||||||
|     int lt = t%ld; |     int lt = t%ld; | ||||||
|     if ( pt == grid->_processor_coor[orthogdim] ) { |     if ( pt == grid->_processor_coor[orthogdim] ) { | ||||||
|       result[t]=lsSum[lt]; |       gsum=lsSum[lt]; | ||||||
|     } else { |     } else { | ||||||
|       result[t]=Zero(); |       gsum=Zero(); | ||||||
|     } |     } | ||||||
|  |  | ||||||
|   } |     grid->GlobalSum(gsum); | ||||||
|   scalar_type * ptr = (scalar_type *) &result[0]; |  | ||||||
|   int words = fd*sizeof(sobj)/sizeof(scalar_type); |  | ||||||
|   grid->GlobalSumVector(ptr, words); |  | ||||||
| } |  | ||||||
| template<class vobj> inline |  | ||||||
| std::vector<typename vobj::scalar_object>  |  | ||||||
| sliceSum(const Lattice<vobj> &Data,int orthogdim) |  | ||||||
| { |  | ||||||
|   std::vector<typename vobj::scalar_object> result; |  | ||||||
|   sliceSum(Data,result,orthogdim); |  | ||||||
|   return result; |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |     result[t]=gsum; | ||||||
|  |   } | ||||||
|  | } | ||||||
|  |  | ||||||
| template<class vobj> | template<class vobj> | ||||||
| static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim)  | static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim)  | ||||||
| @@ -618,8 +486,7 @@ static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Ortho | |||||||
| template<class vobj> | template<class vobj> | ||||||
| static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y, | static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y, | ||||||
| 			    int orthogdim,RealD scale=1.0)  | 			    int orthogdim,RealD scale=1.0)  | ||||||
| { | {     | ||||||
|   // perhaps easier to just promote A to a field and use regular madd |  | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|   typedef typename vobj::scalar_type scalar_type; |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_type vector_type; |   typedef typename vobj::vector_type vector_type; | ||||||
| @@ -650,7 +517,8 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice | |||||||
|     for(int l=0;l<Nsimd;l++){ |     for(int l=0;l<Nsimd;l++){ | ||||||
|       grid->iCoorFromIindex(icoor,l); |       grid->iCoorFromIindex(icoor,l); | ||||||
|       int ldx =r+icoor[orthogdim]*rd; |       int ldx =r+icoor[orthogdim]*rd; | ||||||
|       av.putlane(scalar_type(a[ldx])*zscale,l); |       scalar_type *as =(scalar_type *)&av; | ||||||
|  |       as[l] = scalar_type(a[ldx])*zscale; | ||||||
|     } |     } | ||||||
|  |  | ||||||
|     tensor_reduced at; at=av; |     tensor_reduced at; at=av; | ||||||
| @@ -690,6 +558,7 @@ template<class vobj> | |||||||
| static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)  | static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)  | ||||||
| {     | {     | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_type vector_type; |   typedef typename vobj::vector_type vector_type; | ||||||
|  |  | ||||||
|   int Nblock = X.Grid()->GlobalDimensions()[Orthog]; |   int Nblock = X.Grid()->GlobalDimensions()[Orthog]; | ||||||
| @@ -743,6 +612,7 @@ template<class vobj> | |||||||
| static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0)  | static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0)  | ||||||
| {     | {     | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_type vector_type; |   typedef typename vobj::vector_type vector_type; | ||||||
|  |  | ||||||
|   int Nblock = X.Grid()->GlobalDimensions()[Orthog]; |   int Nblock = X.Grid()->GlobalDimensions()[Orthog]; | ||||||
| @@ -796,6 +666,7 @@ template<class vobj> | |||||||
| static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)  | static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)  | ||||||
| { | { | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|  |   typedef typename vobj::scalar_type scalar_type; | ||||||
|   typedef typename vobj::vector_type vector_type; |   typedef typename vobj::vector_type vector_type; | ||||||
|    |    | ||||||
|   GridBase *FullGrid  = lhs.Grid(); |   GridBase *FullGrid  = lhs.Grid(); | ||||||
|   | |||||||
| @@ -23,27 +23,28 @@ unsigned int nextPow2(Iterator x) { | |||||||
| } | } | ||||||
|  |  | ||||||
| template <class Iterator> | template <class Iterator> | ||||||
| int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) { | void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) { | ||||||
|    |    | ||||||
|   int device; |   int device; | ||||||
| #ifdef GRID_CUDA | #ifdef GRID_CUDA | ||||||
|   cudaGetDevice(&device); |   cudaGetDevice(&device); | ||||||
| #endif | #endif | ||||||
| #ifdef GRID_HIP | #ifdef GRID_HIP | ||||||
|   auto r=hipGetDevice(&device); |   hipGetDevice(&device); | ||||||
| #endif | #endif | ||||||
|    |    | ||||||
|   Iterator warpSize            = gpu_props[device].warpSize; |   Iterator warpSize            = gpu_props[device].warpSize; | ||||||
|   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock; |   Iterator sharedMemPerBlock   = gpu_props[device].sharedMemPerBlock; | ||||||
|   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock; |   Iterator maxThreadsPerBlock  = gpu_props[device].maxThreadsPerBlock; | ||||||
|   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount; |   Iterator multiProcessorCount = gpu_props[device].multiProcessorCount; | ||||||
|   /*   |    | ||||||
|   std::cout << GridLogDebug << "GPU has:" << std::endl; |   std::cout << GridLogDebug << "GPU has:" << std::endl; | ||||||
|   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl; |   std::cout << GridLogDebug << "\twarpSize            = " << warpSize << std::endl; | ||||||
|   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl; |   std::cout << GridLogDebug << "\tsharedMemPerBlock   = " << sharedMemPerBlock << std::endl; | ||||||
|   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl; |   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << maxThreadsPerBlock << std::endl; | ||||||
|  |   std::cout << GridLogDebug << "\tmaxThreadsPerBlock  = " << warpSize << std::endl; | ||||||
|   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl; |   std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl; | ||||||
|   */   |    | ||||||
|   if (warpSize != WARP_SIZE) { |   if (warpSize != WARP_SIZE) { | ||||||
|     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl; |     std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl; | ||||||
|     exit(EXIT_FAILURE); |     exit(EXIT_FAILURE); | ||||||
| @@ -51,14 +52,10 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator & | |||||||
|    |    | ||||||
|   // let the number of threads in a block be a multiple of 2, starting from warpSize |   // let the number of threads in a block be a multiple of 2, starting from warpSize | ||||||
|   threads = warpSize; |   threads = warpSize; | ||||||
|   if ( threads*sizeofsobj > sharedMemPerBlock ) { |  | ||||||
|     std::cout << GridLogError << "The object is too large for the shared memory." << std::endl; |  | ||||||
|     return 0; |  | ||||||
|   } |  | ||||||
|   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2; |   while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2; | ||||||
|   // keep all the streaming multiprocessors busy |   // keep all the streaming multiprocessors busy | ||||||
|   blocks = nextPow2(multiProcessorCount); |   blocks = nextPow2(multiProcessorCount); | ||||||
|   return 1; |    | ||||||
| } | } | ||||||
|  |  | ||||||
| template <class sobj, class Iterator> | template <class sobj, class Iterator> | ||||||
| @@ -198,7 +195,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) { | |||||||
| // Possibly promote to double and sum | // Possibly promote to double and sum | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| template <class vobj> | template <class vobj> | ||||||
| inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)  | inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)  | ||||||
| { | { | ||||||
|   typedef typename vobj::scalar_objectD sobj; |   typedef typename vobj::scalar_objectD sobj; | ||||||
|   typedef decltype(lat) Iterator; |   typedef decltype(lat) Iterator; | ||||||
| @@ -207,77 +204,17 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi | |||||||
|   Integer size = osites*nsimd; |   Integer size = osites*nsimd; | ||||||
|  |  | ||||||
|   Integer numThreads, numBlocks; |   Integer numThreads, numBlocks; | ||||||
|   int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks); |   getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks); | ||||||
|   assert(ok); |  | ||||||
|  |  | ||||||
|   Integer smemSize = numThreads * sizeof(sobj); |   Integer smemSize = numThreads * sizeof(sobj); | ||||||
|   // Move out of UVM |  | ||||||
|   // Turns out I had messed up the synchronise after move to compute stream |  | ||||||
|   // as running this on the default stream fools the synchronise |  | ||||||
| #undef UVM_BLOCK_BUFFER   |  | ||||||
| #ifndef UVM_BLOCK_BUFFER   |  | ||||||
|   commVector<sobj> buffer(numBlocks); |  | ||||||
|   sobj *buffer_v = &buffer[0]; |  | ||||||
|   sobj result; |  | ||||||
|   reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size); |  | ||||||
|   accelerator_barrier(); |  | ||||||
|   acceleratorCopyFromDevice(buffer_v,&result,sizeof(result)); |  | ||||||
| #else |  | ||||||
|   Vector<sobj> buffer(numBlocks); |   Vector<sobj> buffer(numBlocks); | ||||||
|   sobj *buffer_v = &buffer[0]; |   sobj *buffer_v = &buffer[0]; | ||||||
|   sobj result; |    | ||||||
|   reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size); |   reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size); | ||||||
|   accelerator_barrier(); |   accelerator_barrier(); | ||||||
|   result = *buffer_v; |   auto result = buffer_v[0]; | ||||||
| #endif |  | ||||||
|   return result; |   return result; | ||||||
| } | } | ||||||
|  |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites) |  | ||||||
| { |  | ||||||
|   typedef typename vobj::vector_type  vector; |  | ||||||
|   typedef typename vobj::scalar_typeD scalarD; |  | ||||||
|   typedef typename vobj::scalar_objectD sobj; |  | ||||||
|   sobj ret; |  | ||||||
|   scalarD *ret_p = (scalarD *)&ret; |  | ||||||
|    |  | ||||||
|   const int words = sizeof(vobj)/sizeof(vector); |  | ||||||
|  |  | ||||||
|   Vector<vector> buffer(osites); |  | ||||||
|   vector *dat = (vector *)lat; |  | ||||||
|   vector *buf = &buffer[0]; |  | ||||||
|   iScalar<vector> *tbuf =(iScalar<vector> *)  &buffer[0]; |  | ||||||
|   for(int w=0;w<words;w++) { |  | ||||||
|  |  | ||||||
|     accelerator_for(ss,osites,1,{ |  | ||||||
| 	buf[ss] = dat[ss*words+w]; |  | ||||||
|       }); |  | ||||||
|        |  | ||||||
|     ret_p[w] = sumD_gpu_small(tbuf,osites); |  | ||||||
|   } |  | ||||||
|   return ret; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) |  | ||||||
| { |  | ||||||
|   typedef typename vobj::scalar_objectD sobj; |  | ||||||
|   sobj ret; |  | ||||||
|    |  | ||||||
|   Integer nsimd= vobj::Nsimd(); |  | ||||||
|   Integer size = osites*nsimd; |  | ||||||
|   Integer numThreads, numBlocks; |  | ||||||
|   int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks); |  | ||||||
|    |  | ||||||
|   if ( ok ) { |  | ||||||
|     ret = sumD_gpu_small(lat,osites); |  | ||||||
|   } else { |  | ||||||
|     ret = sumD_gpu_large(lat,osites); |  | ||||||
|   } |  | ||||||
|   return ret; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| // Return as same precision as input performing reduction in double precision though | // Return as same precision as input performing reduction in double precision though | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| @@ -290,13 +227,6 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites) | |||||||
|   return result; |   return result; | ||||||
| } | } | ||||||
|  |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites) |  | ||||||
| { |  | ||||||
|   typedef typename vobj::scalar_object sobj; |  | ||||||
|   sobj result; |  | ||||||
|   result = sumD_gpu_large(lat,osites); |  | ||||||
|   return result; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|   | |||||||
| @@ -1,126 +0,0 @@ | |||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
| // Possibly promote to double and sum |  | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites)  |  | ||||||
| { |  | ||||||
|   typedef typename vobj::scalar_object sobj; |  | ||||||
|   typedef typename vobj::scalar_objectD sobjD; |  | ||||||
|   sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator); |  | ||||||
|   sobj identity; zeroit(identity); |  | ||||||
|   sobj ret ;  |  | ||||||
|  |  | ||||||
|   Integer nsimd= vobj::Nsimd(); |  | ||||||
|    |  | ||||||
|   theGridAccelerator->submit([&](cl::sycl::handler &cgh) { |  | ||||||
|      auto Reduction = cl::sycl::reduction(mysum,identity,std::plus<>()); |  | ||||||
|      cgh.parallel_for(cl::sycl::range<1>{osites}, |  | ||||||
| 		      Reduction, |  | ||||||
| 		      [=] (cl::sycl::id<1> item, auto &sum) { |  | ||||||
|       auto osite   = item[0]; |  | ||||||
|       sum +=Reduce(lat[osite]); |  | ||||||
|      }); |  | ||||||
|    }); |  | ||||||
|   theGridAccelerator->wait(); |  | ||||||
|   ret = mysum[0]; |  | ||||||
|   free(mysum,*theGridAccelerator); |  | ||||||
|   sobjD dret; convertType(dret,ret); |  | ||||||
|   return dret; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites) |  | ||||||
| { |  | ||||||
|   return sumD_gpu_tensor(lat,osites); |  | ||||||
| } |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites) |  | ||||||
| { |  | ||||||
|   return sumD_gpu_large(lat,osites); |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites) |  | ||||||
| { |  | ||||||
|   return sumD_gpu_large(lat,osites); |  | ||||||
| } |  | ||||||
|  |  | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
| // Return as same precision as input performing reduction in double precision though |  | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////////////////////// |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)  |  | ||||||
| { |  | ||||||
|   typedef typename vobj::scalar_object sobj; |  | ||||||
|   sobj result; |  | ||||||
|   result = sumD_gpu(lat,osites); |  | ||||||
|   return result; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites) |  | ||||||
| { |  | ||||||
|   typedef typename vobj::scalar_object sobj; |  | ||||||
|   sobj result; |  | ||||||
|   result = sumD_gpu_large(lat,osites); |  | ||||||
|   return result; |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |  | ||||||
| template<class Word> Word svm_xor(Word *vec,uint64_t L) |  | ||||||
| { |  | ||||||
|   Word xorResult; xorResult = 0; |  | ||||||
|   Word *d_sum =(Word *)cl::sycl::malloc_shared(sizeof(Word),*theGridAccelerator); |  | ||||||
|   Word identity;  identity=0; |  | ||||||
|   theGridAccelerator->submit([&](cl::sycl::handler &cgh) { |  | ||||||
|      auto Reduction = cl::sycl::reduction(d_sum,identity,std::bit_xor<>()); |  | ||||||
|      cgh.parallel_for(cl::sycl::range<1>{L}, |  | ||||||
| 		      Reduction, |  | ||||||
| 		      [=] (cl::sycl::id<1> index, auto &sum) { |  | ||||||
| 	 sum ^=vec[index]; |  | ||||||
|      }); |  | ||||||
|    }); |  | ||||||
|   theGridAccelerator->wait(); |  | ||||||
|   Word ret = d_sum[0]; |  | ||||||
|   free(d_sum,*theGridAccelerator); |  | ||||||
|   return ret; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
|  |  | ||||||
| /* |  | ||||||
|  |  | ||||||
| template <class vobj> |  | ||||||
| inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites) |  | ||||||
| { |  | ||||||
|   typedef typename vobj::vector_type  vector; |  | ||||||
|   typedef typename vobj::scalar_type  scalar; |  | ||||||
|  |  | ||||||
|   typedef typename vobj::scalar_typeD scalarD; |  | ||||||
|   typedef typename vobj::scalar_objectD sobjD; |  | ||||||
|  |  | ||||||
|   sobjD ret; |  | ||||||
|   scalarD *ret_p = (scalarD *)&ret; |  | ||||||
|    |  | ||||||
|   const int nsimd = vobj::Nsimd(); |  | ||||||
|   const int words = sizeof(vobj)/sizeof(vector); |  | ||||||
|  |  | ||||||
|   Vector<scalar> buffer(osites*nsimd); |  | ||||||
|   scalar *buf = &buffer[0]; |  | ||||||
|   vector *dat = (vector *)lat; |  | ||||||
|  |  | ||||||
|   for(int w=0;w<words;w++) { |  | ||||||
|  |  | ||||||
|     accelerator_for(ss,osites,nsimd,{ |  | ||||||
| 	int lane = acceleratorSIMTlane(nsimd); |  | ||||||
| 	buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane); |  | ||||||
|     }); |  | ||||||
|     //Precision change at this point is to late to gain precision |  | ||||||
|     ret_p[w] = svm_reduce(buf,nsimd*osites); |  | ||||||
|   } |  | ||||||
|   return ret; |  | ||||||
| } |  | ||||||
| */ |  | ||||||
| @@ -152,7 +152,6 @@ public: | |||||||
| #ifdef RNG_FAST_DISCARD | #ifdef RNG_FAST_DISCARD | ||||||
|   static void Skip(RngEngine &eng,uint64_t site) |   static void Skip(RngEngine &eng,uint64_t site) | ||||||
|   { |   { | ||||||
| #if 0 |  | ||||||
|     ///////////////////////////////////////////////////////////////////////////////////// |     ///////////////////////////////////////////////////////////////////////////////////// | ||||||
|     // Skip by 2^40 elements between successive lattice sites |     // Skip by 2^40 elements between successive lattice sites | ||||||
|     // This goes by 10^12. |     // This goes by 10^12. | ||||||
| @@ -163,9 +162,9 @@ public: | |||||||
|     // tens of seconds per trajectory so this is clean in all reasonable cases, |     // tens of seconds per trajectory so this is clean in all reasonable cases, | ||||||
|     // and margin of safety is orders of magnitude. |     // and margin of safety is orders of magnitude. | ||||||
|     // We could hack Sitmo to skip in the higher order words of state if necessary |     // We could hack Sitmo to skip in the higher order words of state if necessary | ||||||
|     // |       // | ||||||
|     // Replace with 2^30 ; avoid problem on large volumes |       // Replace with 2^30 ; avoid problem on large volumes | ||||||
|     // |       // | ||||||
|     ///////////////////////////////////////////////////////////////////////////////////// |     ///////////////////////////////////////////////////////////////////////////////////// | ||||||
|     //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init |     //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init | ||||||
|     const int shift = 30; |     const int shift = 30; | ||||||
| @@ -180,9 +179,6 @@ public: | |||||||
|     assert((skip >> shift)==site); // check for overflow |     assert((skip >> shift)==site); // check for overflow | ||||||
|  |  | ||||||
|     eng.discard(skip); |     eng.discard(skip); | ||||||
| #else |  | ||||||
|     eng.discardhi(site); |  | ||||||
| #endif |  | ||||||
|     //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl; |     //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl; | ||||||
|   }  |   }  | ||||||
| #endif | #endif | ||||||
| @@ -411,7 +407,7 @@ public: | |||||||
|       std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl; |       std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl; | ||||||
|       SeedFixedIntegers(seeds); |       SeedFixedIntegers(seeds); | ||||||
|     } |     } | ||||||
|   void SeedFixedIntegers(const std::vector<int> &seeds, int britney=0){ |   void SeedFixedIntegers(const std::vector<int> &seeds){ | ||||||
|  |  | ||||||
|     // Everyone generates the same seed_seq based on input seeds |     // Everyone generates the same seed_seq based on input seeds | ||||||
|     CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size()); |     CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size()); | ||||||
| @@ -428,29 +424,22 @@ public: | |||||||
|     // MT implementation does not implement fast discard even though |     // MT implementation does not implement fast discard even though | ||||||
|     // in principle this is possible |     // in principle this is possible | ||||||
|     //////////////////////////////////////////////// |     //////////////////////////////////////////////// | ||||||
|     thread_for( lidx, _grid->lSites(), { |  | ||||||
|  |  | ||||||
| 	int gidx; |     // Everybody loops over global volume. | ||||||
|  |     thread_for( gidx, _grid->_gsites, { | ||||||
|  | 	// Where is it? | ||||||
|  | 	int rank; | ||||||
| 	int o_idx; | 	int o_idx; | ||||||
| 	int i_idx; | 	int i_idx; | ||||||
| 	int rank; |  | ||||||
| 	Coordinate pcoor; |  | ||||||
| 	Coordinate lcoor; |  | ||||||
| 	Coordinate gcoor; | 	Coordinate gcoor; | ||||||
| 	_grid->LocalIndexToLocalCoor(lidx,lcoor); | 	_grid->GlobalIndexToGlobalCoor(gidx,gcoor); | ||||||
| 	pcoor=_grid->ThisProcessorCoor(); |  | ||||||
| 	_grid->ProcessorCoorLocalCoorToGlobalCoor(pcoor,lcoor,gcoor); |  | ||||||
| 	_grid->GlobalCoorToGlobalIndex(gcoor,gidx); |  | ||||||
|  |  | ||||||
| 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor); | 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor); | ||||||
|  |  | ||||||
| 	assert(rank == _grid->ThisRank() ); |  | ||||||
| 	 | 	 | ||||||
| 	int l_idx=generator_idx(o_idx,i_idx); | 	// If this is one of mine we take it | ||||||
| 	_generators[l_idx] = master_engine; | 	if( rank == _grid->ThisRank() ){ | ||||||
| 	if ( britney ) {  | 	  int l_idx=generator_idx(o_idx,i_idx); | ||||||
| 	  Skip(_generators[l_idx],l_idx); // Skip to next RNG sequence | 	  _generators[l_idx] = master_engine; | ||||||
| 	} else { 	 |  | ||||||
| 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence | 	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence | ||||||
| 	} | 	} | ||||||
|     }); |     }); | ||||||
|   | |||||||
| @@ -1,224 +0,0 @@ | |||||||
| #pragma once |  | ||||||
|  |  | ||||||
| #if defined(GRID_CUDA) |  | ||||||
|  |  | ||||||
| #include <cub/cub.cuh> |  | ||||||
| #define gpucub cub |  | ||||||
| #define gpuError_t cudaError_t |  | ||||||
| #define gpuSuccess cudaSuccess |  | ||||||
|  |  | ||||||
| #elif defined(GRID_HIP) |  | ||||||
|  |  | ||||||
| #include <hipcub/hipcub.hpp> |  | ||||||
| #define gpucub hipcub |  | ||||||
| #define gpuError_t hipError_t |  | ||||||
| #define gpuSuccess hipSuccess |  | ||||||
|  |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
|  |  | ||||||
| #if defined(GRID_CUDA) || defined(GRID_HIP) |  | ||||||
| template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) { |  | ||||||
|   size_t subvol_size = e1*e2; |  | ||||||
|   commVector<vobj> reduction_buffer(rd*subvol_size); |  | ||||||
|   auto rb_p = &reduction_buffer[0]; |  | ||||||
|   vobj zero_init; |  | ||||||
|   zeroit(zero_init); |  | ||||||
|  |  | ||||||
|    |  | ||||||
|   void *temp_storage_array = NULL; |  | ||||||
|   size_t temp_storage_bytes = 0; |  | ||||||
|   vobj *d_out; |  | ||||||
|   int* d_offsets; |  | ||||||
|  |  | ||||||
|   std::vector<int> offsets(rd+1,0); |  | ||||||
|  |  | ||||||
|   for (int i = 0; i < offsets.size(); i++) { |  | ||||||
|     offsets[i] = i*subvol_size; |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|   //Allocate memory for output and offset arrays on device |  | ||||||
|   d_out = static_cast<vobj*>(acceleratorAllocDevice(rd*sizeof(vobj))); |  | ||||||
|    |  | ||||||
|   d_offsets = static_cast<int*>(acceleratorAllocDevice((rd+1)*sizeof(int))); |  | ||||||
|    |  | ||||||
|   //copy offsets to device |  | ||||||
|   acceleratorCopyToDeviceAsync(&offsets[0],d_offsets,sizeof(int)*(rd+1),computeStream); |  | ||||||
|    |  | ||||||
|    |  | ||||||
|   gpuError_t gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p,d_out, rd, d_offsets, d_offsets+1, ::gpucub::Sum(), zero_init, computeStream); |  | ||||||
|   if (gpuErr!=gpuSuccess) { |  | ||||||
|     std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce (setup)! Error: " << gpuErr <<std::endl; |  | ||||||
|     exit(EXIT_FAILURE); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   //allocate memory for temp_storage_array   |  | ||||||
|   temp_storage_array = acceleratorAllocDevice(temp_storage_bytes); |  | ||||||
|    |  | ||||||
|   //prepare buffer for reduction |  | ||||||
|   //use non-blocking accelerator_for to avoid syncs (ok because we submit to same computeStream) |  | ||||||
|   //use 2d accelerator_for to avoid launch latencies found when serially looping over rd  |  | ||||||
|   accelerator_for2dNB( s,subvol_size, r,rd, Nsimd,{  |  | ||||||
|    |  | ||||||
|     int n = s / e2; |  | ||||||
|     int b = s % e2; |  | ||||||
|     int so=r*ostride; // base offset for start of plane  |  | ||||||
|     int ss= so+n*stride+b; |  | ||||||
|  |  | ||||||
|     coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss])); |  | ||||||
|  |  | ||||||
|   }); |  | ||||||
|    |  | ||||||
|   //issue segmented reductions in computeStream |  | ||||||
|   gpuErr = gpucub::DeviceSegmentedReduce::Reduce(temp_storage_array, temp_storage_bytes, rb_p, d_out, rd, d_offsets, d_offsets+1,::gpucub::Sum(), zero_init, computeStream); |  | ||||||
|   if (gpuErr!=gpuSuccess) { |  | ||||||
|     std::cout << GridLogError << "Lattice_slicesum_gpu.h: Encountered error during gpucub::DeviceSegmentedReduce::Reduce! Error: " << gpuErr <<std::endl; |  | ||||||
|     exit(EXIT_FAILURE); |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|   acceleratorCopyFromDeviceAsync(d_out,&lvSum[0],rd*sizeof(vobj),computeStream); |  | ||||||
|    |  | ||||||
|   //sync after copy |  | ||||||
|   accelerator_barrier(); |  | ||||||
|   |  | ||||||
|   acceleratorFreeDevice(temp_storage_array); |  | ||||||
|   acceleratorFreeDevice(d_out); |  | ||||||
|   acceleratorFreeDevice(d_offsets); |  | ||||||
|    |  | ||||||
|  |  | ||||||
| } |  | ||||||
| #endif  |  | ||||||
|  |  | ||||||
|  |  | ||||||
| #if defined(GRID_SYCL) |  | ||||||
| template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data, Vector <vobj> &lvSum, const int  &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd) |  | ||||||
| { |  | ||||||
|   size_t subvol_size = e1*e2; |  | ||||||
|  |  | ||||||
|   vobj *mysum = (vobj *) malloc_shared(rd*sizeof(vobj),*theGridAccelerator); |  | ||||||
|   vobj vobj_zero; |  | ||||||
|   zeroit(vobj_zero); |  | ||||||
|   for (int r = 0; r<rd; r++) {  |  | ||||||
|     mysum[r] = vobj_zero;  |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   commVector<vobj> reduction_buffer(rd*subvol_size);     |  | ||||||
|  |  | ||||||
|   auto rb_p = &reduction_buffer[0]; |  | ||||||
|  |  | ||||||
|   // autoView(Data_v, Data, AcceleratorRead); |  | ||||||
|  |  | ||||||
|   //prepare reduction buffer  |  | ||||||
|   accelerator_for2d( s,subvol_size, r,rd, (size_t)Nsimd,{  |  | ||||||
|    |  | ||||||
|       int n = s / e2; |  | ||||||
|       int b = s % e2; |  | ||||||
|       int so=r*ostride; // base offset for start of plane  |  | ||||||
|       int ss= so+n*stride+b; |  | ||||||
|  |  | ||||||
|       coalescedWrite(rb_p[r*subvol_size+s], coalescedRead(Data[ss])); |  | ||||||
|  |  | ||||||
|   }); |  | ||||||
|  |  | ||||||
|   for (int r = 0; r < rd; r++) { |  | ||||||
|       theGridAccelerator->submit([&](cl::sycl::handler &cgh) { |  | ||||||
|           auto Reduction = cl::sycl::reduction(&mysum[r],std::plus<>()); |  | ||||||
|           cgh.parallel_for(cl::sycl::range<1>{subvol_size}, |  | ||||||
|           Reduction, |  | ||||||
|           [=](cl::sycl::id<1> item, auto &sum) { |  | ||||||
|               auto s = item[0]; |  | ||||||
|               sum += rb_p[r*subvol_size+s]; |  | ||||||
|           }); |  | ||||||
|       }); |  | ||||||
|        |  | ||||||
|       |  | ||||||
|   } |  | ||||||
|   theGridAccelerator->wait(); |  | ||||||
|   for (int r = 0; r < rd; r++) { |  | ||||||
|     lvSum[r] = mysum[r]; |  | ||||||
|   } |  | ||||||
|   free(mysum,*theGridAccelerator); |  | ||||||
| } |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
| template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) { |  | ||||||
|   typedef typename vobj::vector_type vector; |  | ||||||
|   const int words = sizeof(vobj)/sizeof(vector); |  | ||||||
|   const int osites = rd*e1*e2; |  | ||||||
|   commVector<vector>buffer(osites); |  | ||||||
|   vector *dat = (vector *)Data; |  | ||||||
|   vector *buf = &buffer[0]; |  | ||||||
|   Vector<vector> lvSum_small(rd); |  | ||||||
|   vector *lvSum_ptr = (vector *)&lvSum[0]; |  | ||||||
|  |  | ||||||
|   for (int w = 0; w < words; w++) { |  | ||||||
|     accelerator_for(ss,osites,1,{ |  | ||||||
| 	    buf[ss] = dat[ss*words+w]; |  | ||||||
|     }); |  | ||||||
|  |  | ||||||
|     #if defined(GRID_CUDA) || defined(GRID_HIP) |  | ||||||
|       sliceSumReduction_cub_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd); |  | ||||||
|     #elif defined(GRID_SYCL) |  | ||||||
|       sliceSumReduction_sycl_small(buf,lvSum_small,rd,e1,e2,stride, ostride,Nsimd); |  | ||||||
|     #endif |  | ||||||
|  |  | ||||||
|     for (int r = 0; r < rd; r++) { |  | ||||||
|       lvSum_ptr[w+words*r]=lvSum_small[r]; |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|    |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) |  | ||||||
| { |  | ||||||
|   autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case. |  | ||||||
|     if constexpr (sizeof(vobj) <= 256) {  |  | ||||||
|  |  | ||||||
|       #if defined(GRID_CUDA) || defined(GRID_HIP) |  | ||||||
|         sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd); |  | ||||||
|       #elif defined (GRID_SYCL) |  | ||||||
|         sliceSumReduction_sycl_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd); |  | ||||||
|       #endif |  | ||||||
|  |  | ||||||
|     } |  | ||||||
|     else { |  | ||||||
|       sliceSumReduction_large(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd); |  | ||||||
|     } |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |  | ||||||
| template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd) |  | ||||||
| { |  | ||||||
|   // sum over reduced dimension planes, breaking out orthog dir |  | ||||||
|   // Parallel over orthog direction |  | ||||||
|   autoView( Data_v, Data, CpuRead); |  | ||||||
|   thread_for( r,rd, { |  | ||||||
|     int so=r*ostride; // base offset for start of plane  |  | ||||||
|     for(int n=0;n<e1;n++){ |  | ||||||
|       for(int b=0;b<e2;b++){ |  | ||||||
|         int ss= so+n*stride+b; |  | ||||||
|         lvSum[r]=lvSum[r]+Data_v[ss]; |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
|   }); |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)  |  | ||||||
| { |  | ||||||
|   #if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL) |  | ||||||
|    |  | ||||||
|   sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd); |  | ||||||
|    |  | ||||||
|   #else |  | ||||||
|   sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd); |  | ||||||
|  |  | ||||||
|   #endif |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
| @@ -66,65 +66,6 @@ inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex< | |||||||
|   return ret; |   return ret; | ||||||
| }; | }; | ||||||
|  |  | ||||||
| template<int N, class Vec> |  | ||||||
| Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu) |  | ||||||
| { |  | ||||||
|   GridBase *grid=Umu.Grid(); |  | ||||||
|   auto lvol = grid->lSites(); |  | ||||||
|   Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid); |  | ||||||
|   typedef typename Vec::scalar_type scalar; |  | ||||||
|   autoView(Umu_v,Umu,CpuRead); |  | ||||||
|   autoView(ret_v,ret,CpuWrite); |  | ||||||
|   thread_for(site,lvol,{ |  | ||||||
|     Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N); |  | ||||||
|     Coordinate lcoor; |  | ||||||
|     grid->LocalIndexToLocalCoor(site, lcoor); |  | ||||||
|     iScalar<iScalar<iMatrix<scalar, N> > > Us; |  | ||||||
|     peekLocalSite(Us, Umu_v, lcoor); |  | ||||||
|     for(int i=0;i<N;i++){ |  | ||||||
|       for(int j=0;j<N;j++){ |  | ||||||
| 	scalar tmp= Us()()(i,j); |  | ||||||
| 	ComplexD ztmp(real(tmp),imag(tmp)); |  | ||||||
| 	EigenU(i,j)=ztmp; |  | ||||||
|       }} |  | ||||||
|     ComplexD detD  = EigenU.determinant(); |  | ||||||
|     typename Vec::scalar_type det(detD.real(),detD.imag()); |  | ||||||
|     pokeLocalSite(det,ret_v,lcoor); |  | ||||||
|   }); |  | ||||||
|   return ret; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<int N> |  | ||||||
| Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu) |  | ||||||
| { |  | ||||||
|   GridBase *grid=Umu.Grid(); |  | ||||||
|   auto lvol = grid->lSites(); |  | ||||||
|   Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid); |  | ||||||
|    |  | ||||||
|   autoView(Umu_v,Umu,CpuRead); |  | ||||||
|   autoView(ret_v,ret,CpuWrite); |  | ||||||
|   thread_for(site,lvol,{ |  | ||||||
|     Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N); |  | ||||||
|     Coordinate lcoor; |  | ||||||
|     grid->LocalIndexToLocalCoor(site, lcoor); |  | ||||||
|     iScalar<iScalar<iMatrix<ComplexD, N> > > Us; |  | ||||||
|     iScalar<iScalar<iMatrix<ComplexD, N> > > Ui; |  | ||||||
|     peekLocalSite(Us, Umu_v, lcoor); |  | ||||||
|     for(int i=0;i<N;i++){ |  | ||||||
|       for(int j=0;j<N;j++){ |  | ||||||
| 	EigenU(i,j) = Us()()(i,j); |  | ||||||
|       }} |  | ||||||
|     Eigen::MatrixXcd EigenUinv = EigenU.inverse(); |  | ||||||
|     for(int i=0;i<N;i++){ |  | ||||||
|       for(int j=0;j<N;j++){ |  | ||||||
| 	Ui()()(i,j) = EigenUinv(i,j); |  | ||||||
|       }} |  | ||||||
|     pokeLocalSite(Ui,ret_v,lcoor); |  | ||||||
|   }); |  | ||||||
|   return ret; |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
|   | |||||||
| @@ -85,76 +85,6 @@ template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Latti | |||||||
|   }); |   }); | ||||||
| } | } | ||||||
|  |  | ||||||
| template<class vobj> inline void acceleratorPickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full, int checker_dim_half=0) |  | ||||||
| { |  | ||||||
|   half.Checkerboard() = cb; |  | ||||||
|   autoView(half_v, half, AcceleratorWrite); |  | ||||||
|   autoView(full_v, full, AcceleratorRead); |  | ||||||
|   Coordinate rdim_full             = full.Grid()->_rdimensions; |  | ||||||
|   Coordinate rdim_half             = half.Grid()->_rdimensions; |  | ||||||
|   unsigned long ndim_half          = half.Grid()->_ndimension; |  | ||||||
|   Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask; |  | ||||||
|   Coordinate ostride_half          = half.Grid()->_ostride; |  | ||||||
|   accelerator_for(ss, full.Grid()->oSites(),full.Grid()->Nsimd(),{ |  | ||||||
|      |  | ||||||
|     Coordinate coor; |  | ||||||
|     int cbos; |  | ||||||
|     int linear=0; |  | ||||||
|  |  | ||||||
|     Lexicographic::CoorFromIndex(coor,ss,rdim_full); |  | ||||||
|     assert(coor.size()==ndim_half); |  | ||||||
|  |  | ||||||
|     for(int d=0;d<ndim_half;d++){  |  | ||||||
|       if(checker_dim_mask_half[d]) linear += coor[d]; |  | ||||||
|     } |  | ||||||
|     cbos = (linear&0x1); |  | ||||||
|  |  | ||||||
|     if (cbos==cb) { |  | ||||||
|       int ssh=0; |  | ||||||
|       for(int d=0;d<ndim_half;d++) { |  | ||||||
|         if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]); |  | ||||||
|         else ssh += ostride_half[d] * (coor[d] % rdim_half[d]); |  | ||||||
|       } |  | ||||||
|       coalescedWrite(half_v[ssh],full_v(ss)); |  | ||||||
|     } |  | ||||||
|   }); |  | ||||||
| } |  | ||||||
| template<class vobj> inline void acceleratorSetCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half, int checker_dim_half=0) |  | ||||||
| { |  | ||||||
|   int cb = half.Checkerboard(); |  | ||||||
|   autoView(half_v , half, AcceleratorRead); |  | ||||||
|   autoView(full_v , full, AcceleratorWrite); |  | ||||||
|   Coordinate rdim_full             = full.Grid()->_rdimensions; |  | ||||||
|   Coordinate rdim_half             = half.Grid()->_rdimensions; |  | ||||||
|   unsigned long ndim_half          = half.Grid()->_ndimension; |  | ||||||
|   Coordinate checker_dim_mask_half = half.Grid()->_checker_dim_mask; |  | ||||||
|   Coordinate ostride_half          = half.Grid()->_ostride; |  | ||||||
|   accelerator_for(ss,full.Grid()->oSites(),full.Grid()->Nsimd(),{ |  | ||||||
|  |  | ||||||
|     Coordinate coor; |  | ||||||
|     int cbos; |  | ||||||
|     int linear=0; |  | ||||||
|    |  | ||||||
|     Lexicographic::CoorFromIndex(coor,ss,rdim_full); |  | ||||||
|     assert(coor.size()==ndim_half); |  | ||||||
|  |  | ||||||
|     for(int d=0;d<ndim_half;d++){  |  | ||||||
|       if(checker_dim_mask_half[d]) linear += coor[d]; |  | ||||||
|     } |  | ||||||
|     cbos = (linear&0x1); |  | ||||||
|  |  | ||||||
|     if (cbos==cb) { |  | ||||||
|       int ssh=0; |  | ||||||
|       for(int d=0;d<ndim_half;d++){ |  | ||||||
|         if (d == checker_dim_half) ssh += ostride_half[d] * ((coor[d] / 2) % rdim_half[d]); |  | ||||||
|         else ssh += ostride_half[d] * (coor[d] % rdim_half[d]); |  | ||||||
|       } |  | ||||||
|       coalescedWrite(full_v[ss],half_v(ssh)); |  | ||||||
|     } |  | ||||||
|  |  | ||||||
|   }); |  | ||||||
| } |  | ||||||
|  |  | ||||||
| //////////////////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////////////////// | ||||||
| // Flexible Type Conversion for internal promotion to double as well as graceful | // Flexible Type Conversion for internal promotion to double as well as graceful | ||||||
| // treatment of scalar-compatible types | // treatment of scalar-compatible types | ||||||
| @@ -167,20 +97,6 @@ accelerator_inline void convertType(ComplexF & out, const std::complex<float> & | |||||||
|   out = in; |   out = in; | ||||||
| } | } | ||||||
|  |  | ||||||
| template<typename T> |  | ||||||
| accelerator_inline EnableIf<isGridFundamental<T>> convertType(T & out, const T & in) { |  | ||||||
|   out = in; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| // This would allow for conversions between GridFundamental types, but is not strictly needed as yet |  | ||||||
| /*template<typename T1, typename T2> |  | ||||||
| accelerator_inline typename std::enable_if<isGridFundamental<T1>::value && isGridFundamental<T2>::value>::type |  | ||||||
| // Or to make this very broad, conversions between anything that's not a GridTensor could be allowed |  | ||||||
| //accelerator_inline typename std::enable_if<!isGridTensor<T1>::value && !isGridTensor<T2>::value>::type |  | ||||||
| convertType(T1 & out, const T2 & in) { |  | ||||||
|   out = in; |  | ||||||
| }*/ |  | ||||||
|  |  | ||||||
| #ifdef GRID_SIMT | #ifdef GRID_SIMT | ||||||
| accelerator_inline void convertType(vComplexF & out, const ComplexF & in) { | accelerator_inline void convertType(vComplexF & out, const ComplexF & in) { | ||||||
|   ((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in; |   ((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in; | ||||||
| @@ -194,25 +110,25 @@ accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) { | |||||||
| #endif | #endif | ||||||
|  |  | ||||||
| accelerator_inline void convertType(vComplexF & out, const vComplexD2 & in) { | accelerator_inline void convertType(vComplexF & out, const vComplexD2 & in) { | ||||||
|   precisionChange(out,in); |   out.v = Optimization::PrecisionChange::DtoS(in._internal[0].v,in._internal[1].v); | ||||||
| } | } | ||||||
|  |  | ||||||
| accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) { | accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) { | ||||||
|   precisionChange(out,in); |   Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v); | ||||||
| } | } | ||||||
|  |  | ||||||
| template<typename T1,typename T2> | template<typename T1,typename T2,int N> | ||||||
| accelerator_inline void convertType(iScalar<T1> & out, const iScalar<T2> & in) { |   accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in); | ||||||
|   convertType(out._internal,in._internal); | template<typename T1,typename T2,int N> | ||||||
| } |   accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in); | ||||||
|  |  | ||||||
| template<typename T1,typename T2> | template<typename T1,typename T2, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr> | ||||||
| accelerator_inline NotEnableIf<isGridScalar<T1>> convertType(T1 & out, const iScalar<T2> & in) { | accelerator_inline void convertType(T1 & out, const iScalar<T2> & in) { | ||||||
|   convertType(out,in._internal); |   convertType(out,in._internal); | ||||||
| } | } | ||||||
|  |  | ||||||
| template<typename T1,typename T2> | template<typename T1,typename T2> | ||||||
| accelerator_inline NotEnableIf<isGridScalar<T2>> convertType(iScalar<T1> & out, const T2 & in) { | accelerator_inline void convertType(iScalar<T1> & out, const T2 & in) { | ||||||
|   convertType(out._internal,in); |   convertType(out._internal,in); | ||||||
| } | } | ||||||
|  |  | ||||||
| @@ -229,6 +145,11 @@ accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & i | |||||||
|     convertType(out._internal[i],in._internal[i]); |     convertType(out._internal[i],in._internal[i]); | ||||||
| } | } | ||||||
|  |  | ||||||
|  | template<typename T, typename std::enable_if<isGridFundamental<T>::value, T>::type* = nullptr> | ||||||
|  | accelerator_inline void convertType(T & out, const T & in) { | ||||||
|  |   out = in; | ||||||
|  | } | ||||||
|  |  | ||||||
| template<typename T1,typename T2> | template<typename T1,typename T2> | ||||||
| accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) { | accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) { | ||||||
|   autoView( out_v , out,AcceleratorWrite); |   autoView( out_v , out,AcceleratorWrite); | ||||||
| @@ -288,36 +209,7 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData, | |||||||
|     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed);  |     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed);  | ||||||
|   } |   } | ||||||
| } | } | ||||||
| template<class vobj,class CComplex,int nbasis,class VLattice> |  | ||||||
| inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData, |  | ||||||
|                                const std::vector<Lattice<vobj>> &fineData, |  | ||||||
|                                const VLattice &Basis) |  | ||||||
| { |  | ||||||
|   int NBatch = fineData.size(); |  | ||||||
|   assert(coarseData.size() == NBatch); |  | ||||||
|  |  | ||||||
|   GridBase * fine  = fineData[0].Grid(); |  | ||||||
|   GridBase * coarse= coarseData[0].Grid(); |  | ||||||
|  |  | ||||||
|   Lattice<iScalar<CComplex>> ip(coarse); |  | ||||||
|   std::vector<Lattice<vobj>> fineDataCopy = fineData; |  | ||||||
|  |  | ||||||
|   autoView(ip_, ip, AcceleratorWrite); |  | ||||||
|   for(int v=0;v<nbasis;v++) { |  | ||||||
|     for (int k=0; k<NBatch; k++) { |  | ||||||
|       autoView( coarseData_ , coarseData[k], AcceleratorWrite); |  | ||||||
|       blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // ip = <basis|fine> |  | ||||||
|       accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), { |  | ||||||
|         convertType(coarseData_[sc](v),ip_[sc]); |  | ||||||
|       }); |  | ||||||
|  |  | ||||||
|       // improve numerical stability of projection |  | ||||||
|       // |fine> = |fine> - <basis|fine> |basis> |  | ||||||
|       ip=-ip; |  | ||||||
|       blockZAXPY(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]);  |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
| } |  | ||||||
|  |  | ||||||
| template<class vobj,class vobj2,class CComplex> | template<class vobj,class vobj2,class CComplex> | ||||||
|   inline void blockZAXPY(Lattice<vobj> &fineZ, |   inline void blockZAXPY(Lattice<vobj> &fineZ, | ||||||
| @@ -463,20 +355,16 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData) | |||||||
|   autoView( coarseData_ , coarseData, AcceleratorWrite); |   autoView( coarseData_ , coarseData, AcceleratorWrite); | ||||||
|   autoView( fineData_   , fineData, AcceleratorRead); |   autoView( fineData_   , fineData, AcceleratorRead); | ||||||
|  |  | ||||||
|   auto coarseData_p = &coarseData_[0]; |  | ||||||
|   auto fineData_p = &fineData_[0]; |  | ||||||
|    |  | ||||||
|   Coordinate fine_rdimensions = fine->_rdimensions; |   Coordinate fine_rdimensions = fine->_rdimensions; | ||||||
|   Coordinate coarse_rdimensions = coarse->_rdimensions; |   Coordinate coarse_rdimensions = coarse->_rdimensions; | ||||||
|  |    | ||||||
|   accelerator_for(sc,coarse->oSites(),1,{ |   accelerator_for(sc,coarse->oSites(),1,{ | ||||||
|  |  | ||||||
|       // One thread per sub block |       // One thread per sub block | ||||||
|       Coordinate coor_c(_ndimension); |       Coordinate coor_c(_ndimension); | ||||||
|       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate |       Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions);  // Block coordinate | ||||||
|  |       coarseData_[sc]=Zero(); | ||||||
|  |  | ||||||
|       vobj cd = Zero(); |  | ||||||
|        |  | ||||||
|       for(int sb=0;sb<blockVol;sb++){ |       for(int sb=0;sb<blockVol;sb++){ | ||||||
|  |  | ||||||
| 	int sf; | 	int sf; | ||||||
| @@ -486,11 +374,9 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData) | |||||||
| 	for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d]; | 	for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d]; | ||||||
| 	Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions); | 	Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions); | ||||||
|  |  | ||||||
| 	cd=cd+fineData_p[sf]; | 	coarseData_[sc]=coarseData_[sc]+fineData_[sf]; | ||||||
|       } |       } | ||||||
|  |  | ||||||
|       coarseData_p[sc] = cd; |  | ||||||
|  |  | ||||||
|     }); |     }); | ||||||
|   return; |   return; | ||||||
| } | } | ||||||
| @@ -617,26 +503,6 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData, | |||||||
| } | } | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
| template<class vobj,class CComplex,int nbasis,class VLattice> |  | ||||||
| inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData, |  | ||||||
|                                std::vector<Lattice<vobj>> &fineData, |  | ||||||
|                                const VLattice &Basis) |  | ||||||
| { |  | ||||||
|   int NBatch = coarseData.size(); |  | ||||||
|   assert(fineData.size() == NBatch); |  | ||||||
|  |  | ||||||
|   GridBase * fine   = fineData[0].Grid(); |  | ||||||
|   GridBase * coarse = coarseData[0].Grid(); |  | ||||||
|   for (int k=0; k<NBatch; k++) |  | ||||||
|     fineData[k]=Zero(); |  | ||||||
|   for (int i=0;i<nbasis;i++) { |  | ||||||
|     for (int k=0; k<NBatch; k++) { |  | ||||||
|       Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i); |  | ||||||
|       blockZAXPY(fineData[k],ip,Basis[i],fineData[k]); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
| } |  | ||||||
|  |  | ||||||
| // Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars. | // Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars. | ||||||
| // Simd layouts need not match since we use peek/poke Local | // Simd layouts need not match since we use peek/poke Local | ||||||
| template<class vobj,class vvobj> | template<class vobj,class vvobj> | ||||||
| @@ -695,68 +561,8 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro | |||||||
|   for(int d=0;d<nd;d++){ |   for(int d=0;d<nd;d++){ | ||||||
|     assert(Fg->_processors[d]  == Tg->_processors[d]); |     assert(Fg->_processors[d]  == Tg->_processors[d]); | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   // the above should guarantee that the operations are local |   // the above should guarantee that the operations are local | ||||||
|    |  | ||||||
| #if 1 |  | ||||||
|  |  | ||||||
|   size_t nsite = 1; |  | ||||||
|   for(int i=0;i<nd;i++) nsite *= RegionSize[i]; |  | ||||||
|    |  | ||||||
|   size_t tbytes = 4*nsite*sizeof(int); |  | ||||||
|   int *table = (int*)malloc(tbytes); |  | ||||||
|   |  | ||||||
|   thread_for(idx, nsite, { |  | ||||||
|       Coordinate from_coor, to_coor; |  | ||||||
|       size_t rem = idx; |  | ||||||
|       for(int i=0;i<nd;i++){ |  | ||||||
| 	size_t base_i  = rem % RegionSize[i]; rem /= RegionSize[i]; |  | ||||||
| 	from_coor[i] = base_i + FromLowerLeft[i]; |  | ||||||
| 	to_coor[i] = base_i + ToLowerLeft[i]; |  | ||||||
|       } |  | ||||||
|        |  | ||||||
|       int foidx = Fg->oIndex(from_coor); |  | ||||||
|       int fiidx = Fg->iIndex(from_coor); |  | ||||||
|       int toidx = Tg->oIndex(to_coor); |  | ||||||
|       int tiidx = Tg->iIndex(to_coor); |  | ||||||
|       int* tt = table + 4*idx; |  | ||||||
|       tt[0] = foidx; |  | ||||||
|       tt[1] = fiidx; |  | ||||||
|       tt[2] = toidx; |  | ||||||
|       tt[3] = tiidx; |  | ||||||
|     }); |  | ||||||
|    |  | ||||||
|   int* table_d = (int*)acceleratorAllocDevice(tbytes); |  | ||||||
|   acceleratorCopyToDevice(table,table_d,tbytes); |  | ||||||
|  |  | ||||||
|   typedef typename vobj::vector_type vector_type; |  | ||||||
|   typedef typename vobj::scalar_type scalar_type; |  | ||||||
|  |  | ||||||
|   autoView(from_v,From,AcceleratorRead); |  | ||||||
|   autoView(to_v,To,AcceleratorWrite); |  | ||||||
|    |  | ||||||
|   accelerator_for(idx,nsite,1,{ |  | ||||||
|       static const int words=sizeof(vobj)/sizeof(vector_type); |  | ||||||
|       int* tt = table_d + 4*idx; |  | ||||||
|       int from_oidx = *tt++; |  | ||||||
|       int from_lane = *tt++; |  | ||||||
|       int to_oidx = *tt++; |  | ||||||
|       int to_lane = *tt; |  | ||||||
|  |  | ||||||
|       const vector_type* from = (const vector_type *)&from_v[from_oidx]; |  | ||||||
|       vector_type* to = (vector_type *)&to_v[to_oidx]; |  | ||||||
|        |  | ||||||
|       scalar_type stmp; |  | ||||||
|       for(int w=0;w<words;w++){ |  | ||||||
| 	stmp = getlane(from[w], from_lane); |  | ||||||
| 	putlane(to[w], stmp, to_lane); |  | ||||||
|       } |  | ||||||
|     }); |  | ||||||
|    |  | ||||||
|   acceleratorFreeDevice(table_d);     |  | ||||||
|   free(table); |  | ||||||
|    |  | ||||||
|  |  | ||||||
| #else   |  | ||||||
|   Coordinate ldf = Fg->_ldimensions; |   Coordinate ldf = Fg->_ldimensions; | ||||||
|   Coordinate rdf = Fg->_rdimensions; |   Coordinate rdf = Fg->_rdimensions; | ||||||
|   Coordinate isf = Fg->_istride; |   Coordinate isf = Fg->_istride; | ||||||
| @@ -765,9 +571,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro | |||||||
|   Coordinate ist = Tg->_istride; |   Coordinate ist = Tg->_istride; | ||||||
|   Coordinate ost = Tg->_ostride; |   Coordinate ost = Tg->_ostride; | ||||||
|  |  | ||||||
|   autoView( t_v , To, CpuWrite); |   autoView( t_v , To, AcceleratorWrite); | ||||||
|   autoView( f_v , From, CpuRead); |   autoView( f_v , From, AcceleratorRead); | ||||||
|   thread_for(idx,Fg->lSites(),{ |   accelerator_for(idx,Fg->lSites(),1,{ | ||||||
|     sobj s; |     sobj s; | ||||||
|     Coordinate Fcoor(nd); |     Coordinate Fcoor(nd); | ||||||
|     Coordinate Tcoor(nd); |     Coordinate Tcoor(nd); | ||||||
| @@ -780,24 +586,17 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro | |||||||
|       Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d]; |       Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d]; | ||||||
|     } |     } | ||||||
|     if (in_region) { |     if (in_region) { | ||||||
| #if 0       |       Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); | ||||||
|       Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); // inner index from |       Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); | ||||||
|       Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); // inner index to |       Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); | ||||||
|       Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); // outer index from |       Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); | ||||||
|       Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); // outer index to |  | ||||||
|       scalar_type * fp = (scalar_type *)&f_v[odx_f]; |       scalar_type * fp = (scalar_type *)&f_v[odx_f]; | ||||||
|       scalar_type * tp = (scalar_type *)&t_v[odx_t]; |       scalar_type * tp = (scalar_type *)&t_v[odx_t]; | ||||||
|       for(int w=0;w<words;w++){ |       for(int w=0;w<words;w++){ | ||||||
| 	tp[w].putlane(fp[w].getlane(idx_f),idx_t); | 	tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd];  // FIXME IF RRII layout, type pun no worke | ||||||
|       } |       } | ||||||
| #else |  | ||||||
|     peekLocalSite(s,f_v,Fcoor); |  | ||||||
|     pokeLocalSite(s,t_v,Tcoor); |  | ||||||
| #endif |  | ||||||
|     } |     } | ||||||
|   }); |   }); | ||||||
|  |  | ||||||
| #endif |  | ||||||
| } | } | ||||||
|  |  | ||||||
|  |  | ||||||
| @@ -890,8 +689,6 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic | |||||||
| } | } | ||||||
|  |  | ||||||
|  |  | ||||||
| //Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim |  | ||||||
| //The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same |  | ||||||
| template<class vobj> | template<class vobj> | ||||||
| void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog) | void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog) | ||||||
| { | { | ||||||
| @@ -908,70 +705,11 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int | |||||||
|  |  | ||||||
|   for(int d=0;d<nh;d++){ |   for(int d=0;d<nh;d++){ | ||||||
|     if ( d!=orthog ) { |     if ( d!=orthog ) { | ||||||
|       assert(lg->_processors[d]  == hg->_processors[d]); |     assert(lg->_processors[d]  == hg->_processors[d]); | ||||||
|       assert(lg->_ldimensions[d] == hg->_ldimensions[d]); |     assert(lg->_ldimensions[d] == hg->_ldimensions[d]); | ||||||
|     } |   } | ||||||
|   } |   } | ||||||
|  |  | ||||||
| #if 1 |  | ||||||
|   size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog]; |  | ||||||
|   size_t tbytes = 4*nsite*sizeof(int); |  | ||||||
|   int *table = (int*)malloc(tbytes); |  | ||||||
|    |  | ||||||
|   thread_for(idx,nsite,{ |  | ||||||
|     Coordinate lcoor(nl); |  | ||||||
|     Coordinate hcoor(nh); |  | ||||||
|     lcoor[orthog] = slice_lo; |  | ||||||
|     hcoor[orthog] = slice_hi; |  | ||||||
|     size_t rem = idx; |  | ||||||
|     for(int mu=0;mu<nl;mu++){ |  | ||||||
|       if(mu != orthog){ |  | ||||||
| 	int xmu = rem % lg->LocalDimensions()[mu];  rem /= lg->LocalDimensions()[mu]; |  | ||||||
| 	lcoor[mu] = hcoor[mu] = xmu; |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
|     int loidx = lg->oIndex(lcoor); |  | ||||||
|     int liidx = lg->iIndex(lcoor); |  | ||||||
|     int hoidx = hg->oIndex(hcoor); |  | ||||||
|     int hiidx = hg->iIndex(hcoor); |  | ||||||
|     int* tt = table + 4*idx; |  | ||||||
|     tt[0] = loidx; |  | ||||||
|     tt[1] = liidx; |  | ||||||
|     tt[2] = hoidx; |  | ||||||
|     tt[3] = hiidx; |  | ||||||
|     }); |  | ||||||
|     |  | ||||||
|   int* table_d = (int*)acceleratorAllocDevice(tbytes); |  | ||||||
|   acceleratorCopyToDevice(table,table_d,tbytes); |  | ||||||
|  |  | ||||||
|   typedef typename vobj::vector_type vector_type; |  | ||||||
|   typedef typename vobj::scalar_type scalar_type; |  | ||||||
|  |  | ||||||
|   autoView(lowDim_v,lowDim,AcceleratorRead); |  | ||||||
|   autoView(higherDim_v,higherDim,AcceleratorWrite); |  | ||||||
|    |  | ||||||
|   accelerator_for(idx,nsite,1,{ |  | ||||||
|       static const int words=sizeof(vobj)/sizeof(vector_type); |  | ||||||
|       int* tt = table_d + 4*idx; |  | ||||||
|       int from_oidx = *tt++; |  | ||||||
|       int from_lane = *tt++; |  | ||||||
|       int to_oidx = *tt++; |  | ||||||
|       int to_lane = *tt; |  | ||||||
|  |  | ||||||
|       const vector_type* from = (const vector_type *)&lowDim_v[from_oidx]; |  | ||||||
|       vector_type* to = (vector_type *)&higherDim_v[to_oidx]; |  | ||||||
|        |  | ||||||
|       scalar_type stmp; |  | ||||||
|       for(int w=0;w<words;w++){ |  | ||||||
| 	stmp = getlane(from[w], from_lane); |  | ||||||
| 	putlane(to[w], stmp, to_lane); |  | ||||||
|       } |  | ||||||
|     }); |  | ||||||
|    |  | ||||||
|   acceleratorFreeDevice(table_d);     |  | ||||||
|   free(table); |  | ||||||
|    |  | ||||||
| #else |  | ||||||
|   // the above should guarantee that the operations are local |   // the above should guarantee that the operations are local | ||||||
|   autoView(lowDimv,lowDim,CpuRead); |   autoView(lowDimv,lowDim,CpuRead); | ||||||
|   autoView(higherDimv,higherDim,CpuWrite); |   autoView(higherDimv,higherDim,CpuWrite); | ||||||
| @@ -987,7 +725,6 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int | |||||||
|       pokeLocalSite(s,higherDimv,hcoor); |       pokeLocalSite(s,higherDimv,hcoor); | ||||||
|     } |     } | ||||||
|   }); |   }); | ||||||
| #endif |  | ||||||
| } | } | ||||||
|  |  | ||||||
|  |  | ||||||
| @@ -1031,7 +768,7 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int | |||||||
|  |  | ||||||
|  |  | ||||||
| template<class vobj> | template<class vobj> | ||||||
| void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine) | void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine) | ||||||
| { | { | ||||||
|   typedef typename vobj::scalar_object sobj; |   typedef typename vobj::scalar_object sobj; | ||||||
|  |  | ||||||
| @@ -1256,27 +993,9 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out) | |||||||
|   }); |   }); | ||||||
| } | } | ||||||
|  |  | ||||||
| //Very fast precision change. Requires in/out objects to reside on same Grid (e.g. by using double2 for the double-precision field) | //Convert a Lattice from one precision to another | ||||||
| template<class VobjOut, class VobjIn> | template<class VobjOut, class VobjIn> | ||||||
| void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in) | void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in) | ||||||
| { |  | ||||||
|   typedef typename VobjOut::vector_type Vout; |  | ||||||
|   typedef typename VobjIn::vector_type Vin; |  | ||||||
|   const int N = sizeof(VobjOut)/sizeof(Vout); |  | ||||||
|   conformable(out.Grid(),in.Grid()); |  | ||||||
|   out.Checkerboard() = in.Checkerboard(); |  | ||||||
|   int nsimd = out.Grid()->Nsimd(); |  | ||||||
|   autoView( out_v  , out, AcceleratorWrite); |  | ||||||
|   autoView(  in_v ,   in, AcceleratorRead); |  | ||||||
|   accelerator_for(idx,out.Grid()->oSites(),1,{ |  | ||||||
|       Vout *vout = (Vout *)&out_v[idx]; |  | ||||||
|       Vin  *vin  = (Vin  *)&in_v[idx]; |  | ||||||
|       precisionChange(vout,vin,N); |  | ||||||
|   }); |  | ||||||
| } |  | ||||||
| //Convert a Lattice from one precision to another (original, slow implementation) |  | ||||||
| template<class VobjOut, class VobjIn> |  | ||||||
| void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in) |  | ||||||
| { | { | ||||||
|   assert(out.Grid()->Nd() == in.Grid()->Nd()); |   assert(out.Grid()->Nd() == in.Grid()->Nd()); | ||||||
|   for(int d=0;d<out.Grid()->Nd();d++){ |   for(int d=0;d<out.Grid()->Nd();d++){ | ||||||
| @@ -1291,7 +1010,7 @@ void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in) | |||||||
|  |  | ||||||
|   int ndim = out.Grid()->Nd(); |   int ndim = out.Grid()->Nd(); | ||||||
|   int out_nsimd = out_grid->Nsimd(); |   int out_nsimd = out_grid->Nsimd(); | ||||||
|   int in_nsimd = in_grid->Nsimd(); |      | ||||||
|   std::vector<Coordinate > out_icoor(out_nsimd); |   std::vector<Coordinate > out_icoor(out_nsimd); | ||||||
|        |        | ||||||
|   for(int lane=0; lane < out_nsimd; lane++){ |   for(int lane=0; lane < out_nsimd; lane++){ | ||||||
| @@ -1322,128 +1041,6 @@ void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in) | |||||||
|   }); |   }); | ||||||
| } | } | ||||||
|  |  | ||||||
| //The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls |  | ||||||
| class precisionChangeWorkspace{ |  | ||||||
|   std::pair<Integer,Integer>* fmap_device; //device pointer |  | ||||||
|   //maintain grids for checking |  | ||||||
|   GridBase* _out_grid; |  | ||||||
|   GridBase* _in_grid; |  | ||||||
| public: |  | ||||||
|   precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid): _out_grid(out_grid), _in_grid(in_grid){ |  | ||||||
|     //Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device |  | ||||||
|     assert(out_grid->Nd() == in_grid->Nd()); |  | ||||||
|     for(int d=0;d<out_grid->Nd();d++){ |  | ||||||
|       assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]); |  | ||||||
|     } |  | ||||||
|     int Nsimd_out = out_grid->Nsimd(); |  | ||||||
|  |  | ||||||
|     std::vector<Coordinate> out_icorrs(out_grid->Nsimd()); //reuse these |  | ||||||
|     for(int lane=0; lane < out_grid->Nsimd(); lane++) |  | ||||||
|       out_grid->iCoorFromIindex(out_icorrs[lane], lane); |  | ||||||
|    |  | ||||||
|     std::vector<std::pair<Integer,Integer> > fmap_host(out_grid->lSites()); //lsites = osites*Nsimd |  | ||||||
|     thread_for(out_oidx,out_grid->oSites(),{ |  | ||||||
| 	Coordinate out_ocorr;  |  | ||||||
| 	out_grid->oCoorFromOindex(out_ocorr, out_oidx); |  | ||||||
|        |  | ||||||
| 	Coordinate lcorr; //the local coordinate (common to both in and out as full coordinate) |  | ||||||
| 	for(int out_lane=0; out_lane < Nsimd_out; out_lane++){ |  | ||||||
| 	  out_grid->InOutCoorToLocalCoor(out_ocorr, out_icorrs[out_lane], lcorr); |  | ||||||
| 	 |  | ||||||
| 	  //int in_oidx = in_grid->oIndex(lcorr), in_lane = in_grid->iIndex(lcorr); |  | ||||||
| 	  //Note oIndex and OcorrFromOindex (and same for iIndex) are not inverse for checkerboarded lattice, the former coordinates being defined on the full lattice and the latter on the reduced lattice |  | ||||||
| 	  //Until this is fixed we need to circumvent the problem locally. Here I will use the coordinates defined on the reduced lattice for simplicity |  | ||||||
| 	  int in_oidx = 0, in_lane = 0; |  | ||||||
| 	  for(int d=0;d<in_grid->_ndimension;d++){ |  | ||||||
| 	    in_oidx += in_grid->_ostride[d] * ( lcorr[d] % in_grid->_rdimensions[d] ); |  | ||||||
| 	    in_lane += in_grid->_istride[d] * ( lcorr[d] / in_grid->_rdimensions[d] ); |  | ||||||
| 	  } |  | ||||||
| 	  fmap_host[out_lane + Nsimd_out*out_oidx] = std::pair<Integer,Integer>( in_oidx, in_lane ); |  | ||||||
| 	} |  | ||||||
|       }); |  | ||||||
|  |  | ||||||
|     //Copy the map to the device (if we had a way to tell if an accelerator is in use we could avoid this copy for CPU-only machines) |  | ||||||
|     size_t fmap_bytes = out_grid->lSites() * sizeof(std::pair<Integer,Integer>); |  | ||||||
|     fmap_device = (std::pair<Integer,Integer>*)acceleratorAllocDevice(fmap_bytes); |  | ||||||
|     acceleratorCopyToDevice(fmap_host.data(), fmap_device, fmap_bytes);  |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   //Prevent moving or copying |  | ||||||
|   precisionChangeWorkspace(const precisionChangeWorkspace &r) = delete; |  | ||||||
|   precisionChangeWorkspace(precisionChangeWorkspace &&r) = delete; |  | ||||||
|   precisionChangeWorkspace &operator=(const precisionChangeWorkspace &r) = delete; |  | ||||||
|   precisionChangeWorkspace &operator=(precisionChangeWorkspace &&r) = delete; |  | ||||||
|    |  | ||||||
|   std::pair<Integer,Integer> const* getMap() const{ return fmap_device; } |  | ||||||
|  |  | ||||||
|   void checkGrids(GridBase* out, GridBase* in) const{ |  | ||||||
|     conformable(out, _out_grid); |  | ||||||
|     conformable(in, _in_grid); |  | ||||||
|   } |  | ||||||
|    |  | ||||||
|   ~precisionChangeWorkspace(){ |  | ||||||
|     acceleratorFreeDevice(fmap_device); |  | ||||||
|   } |  | ||||||
| }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
| //We would like to use precisionChangeFast when possible. However usage of this requires the Grids to be the same (runtime check) |  | ||||||
| //*and* the precisionChange(VobjOut::vector_type, VobjIn, int) function to be defined for the types; this requires an extra compile-time check which we do using some SFINAE trickery |  | ||||||
| template<class VobjOut, class VobjIn> |  | ||||||
| auto _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, int dummy)->decltype( precisionChange( ((typename VobjOut::vector_type*)0), ((typename VobjIn::vector_type*)0), 1), int()){ |  | ||||||
|   if(out.Grid() == in.Grid()){ |  | ||||||
|     precisionChangeFast(out,in); |  | ||||||
|     return 1; |  | ||||||
|   }else{ |  | ||||||
|     return 0; |  | ||||||
|   } |  | ||||||
| } |  | ||||||
| template<class VobjOut, class VobjIn> |  | ||||||
| int _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, long dummy){ //note long here is intentional; it means the above is preferred if available |  | ||||||
|   return 0; |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |  | ||||||
| //Convert a lattice of one precision to another. Much faster than original implementation but requires a pregenerated workspace |  | ||||||
| //which contains the mapping data. |  | ||||||
| template<class VobjOut, class VobjIn> |  | ||||||
| void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){ |  | ||||||
|   if(_precisionChangeFastWrap(out,in,0)) return; |  | ||||||
|    |  | ||||||
|   static_assert( std::is_same<typename VobjOut::scalar_typeD, typename VobjIn::scalar_typeD>::value == 1, "precisionChange: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same |  | ||||||
|  |  | ||||||
|   out.Checkerboard() = in.Checkerboard(); |  | ||||||
|   constexpr int Nsimd_out = VobjOut::Nsimd(); |  | ||||||
|  |  | ||||||
|   workspace.checkGrids(out.Grid(),in.Grid()); |  | ||||||
|   std::pair<Integer,Integer> const* fmap_device = workspace.getMap(); |  | ||||||
|  |  | ||||||
|   //Do the copy/precision change |  | ||||||
|   autoView( out_v , out, AcceleratorWrite); |  | ||||||
|   autoView( in_v , in, AcceleratorRead); |  | ||||||
|  |  | ||||||
|   accelerator_for(out_oidx, out.Grid()->oSites(), 1,{ |  | ||||||
|       std::pair<Integer,Integer> const* fmap_osite = fmap_device + out_oidx*Nsimd_out; |  | ||||||
|       for(int out_lane=0; out_lane < Nsimd_out; out_lane++){       |  | ||||||
| 	int in_oidx = fmap_osite[out_lane].first; |  | ||||||
| 	int in_lane = fmap_osite[out_lane].second; |  | ||||||
| 	copyLane(out_v[out_oidx], out_lane, in_v[in_oidx], in_lane); |  | ||||||
|       } |  | ||||||
|     }); |  | ||||||
| } |  | ||||||
|  |  | ||||||
| //Convert a Lattice from one precision to another. Much faster than original implementation but slower than precisionChangeFast |  | ||||||
| //or precisionChange called with pregenerated workspace, as it needs to internally generate the workspace on the host and copy to device |  | ||||||
| template<class VobjOut, class VobjIn> |  | ||||||
| void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){ |  | ||||||
|   if(_precisionChangeFastWrap(out,in,0)) return;    |  | ||||||
|   precisionChangeWorkspace workspace(out.Grid(), in.Grid()); |  | ||||||
|   precisionChange(out, in, workspace); |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |  | ||||||
|  |  | ||||||
|  |  | ||||||
| //////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////// | ||||||
| // Communicate between grids | // Communicate between grids | ||||||
| //////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////// | ||||||
|   | |||||||
| @@ -45,7 +45,6 @@ public: | |||||||
|   }; |   }; | ||||||
|   // Host only |   // Host only | ||||||
|   GridBase * getGrid(void) const { return _grid; }; |   GridBase * getGrid(void) const { return _grid; }; | ||||||
|   vobj* getHostPointer(void) const { return _odata; }; |  | ||||||
| }; | }; | ||||||
|  |  | ||||||
| ///////////////////////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////////////////////// | ||||||
| @@ -68,14 +67,9 @@ public: | |||||||
|   accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; } |   accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; } | ||||||
| #endif | #endif | ||||||
|  |  | ||||||
| #if 1 |  | ||||||
|   //  accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; }; |  | ||||||
|   accelerator_inline vobj       & operator[](size_t i) const { return this->_odata[i]; }; |  | ||||||
| #else |  | ||||||
|   accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; }; |   accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; }; | ||||||
|   accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; }; |   accelerator_inline vobj       & operator[](size_t i)       { return this->_odata[i]; }; | ||||||
| #endif |  | ||||||
|    |  | ||||||
|   accelerator_inline uint64_t begin(void) const { return 0;}; |   accelerator_inline uint64_t begin(void) const { return 0;}; | ||||||
|   accelerator_inline uint64_t end(void)   const { return this->_odata_size; }; |   accelerator_inline uint64_t end(void)   const { return this->_odata_size; }; | ||||||
|   accelerator_inline uint64_t size(void)  const { return this->_odata_size; }; |   accelerator_inline uint64_t size(void)  const { return this->_odata_size; }; | ||||||
|   | |||||||
| @@ -43,7 +43,7 @@ inline void whereWolf(Lattice<vobj> &ret,const Lattice<iobj> &predicate,Lattice< | |||||||
|   conformable(iftrue,predicate); |   conformable(iftrue,predicate); | ||||||
|   conformable(iftrue,ret); |   conformable(iftrue,ret); | ||||||
|  |  | ||||||
|   GridBase *grid=iftrue.Grid(); |   GridBase *grid=iftrue._grid; | ||||||
|  |  | ||||||
|   typedef typename vobj::scalar_object scalar_object; |   typedef typename vobj::scalar_object scalar_object; | ||||||
|   typedef typename vobj::scalar_type scalar_type; |   typedef typename vobj::scalar_type scalar_type; | ||||||
| @@ -52,23 +52,22 @@ inline void whereWolf(Lattice<vobj> &ret,const Lattice<iobj> &predicate,Lattice< | |||||||
|  |  | ||||||
|   const int Nsimd = grid->Nsimd(); |   const int Nsimd = grid->Nsimd(); | ||||||
|  |  | ||||||
|   autoView(iftrue_v,iftrue,CpuRead); |   std::vector<Integer> mask(Nsimd); | ||||||
|   autoView(iffalse_v,iffalse,CpuRead); |   std::vector<scalar_object> truevals (Nsimd); | ||||||
|   autoView(predicate_v,predicate,CpuRead); |   std::vector<scalar_object> falsevals(Nsimd); | ||||||
|   autoView(ret_v,ret,CpuWrite); |  | ||||||
|   Integer NN= grid->oSites(); |   parallel_for(int ss=0;ss<iftrue._grid->oSites(); ss++){ | ||||||
|   thread_for(ss,NN,{ |  | ||||||
|     Integer mask; |     extract(iftrue._odata[ss]   ,truevals); | ||||||
|     scalar_object trueval; |     extract(iffalse._odata[ss]  ,falsevals); | ||||||
|     scalar_object falseval; |     extract<vInteger,Integer>(TensorRemove(predicate._odata[ss]),mask); | ||||||
|     for(int l=0;l<Nsimd;l++){ |  | ||||||
|       trueval =extractLane(l,iftrue_v[ss]); |     for(int s=0;s<Nsimd;s++){ | ||||||
|       falseval=extractLane(l,iffalse_v[ss]); |       if (mask[s]) falsevals[s]=truevals[s]; | ||||||
|       mask    =extractLane(l,predicate_v[ss]); |  | ||||||
|       if (mask) falseval=trueval; |  | ||||||
|       insertLane(l,ret_v[ss],falseval); |  | ||||||
|     } |     } | ||||||
|   }); |  | ||||||
|  |     merge(ret._odata[ss],falsevals); | ||||||
|  |   } | ||||||
| } | } | ||||||
|  |  | ||||||
| template<class vobj,class iobj> | template<class vobj,class iobj> | ||||||
| @@ -77,9 +76,9 @@ inline Lattice<vobj> whereWolf(const Lattice<iobj> &predicate,Lattice<vobj> &ift | |||||||
|   conformable(iftrue,iffalse); |   conformable(iftrue,iffalse); | ||||||
|   conformable(iftrue,predicate); |   conformable(iftrue,predicate); | ||||||
|  |  | ||||||
|   Lattice<vobj> ret(iftrue.Grid()); |   Lattice<vobj> ret(iftrue._grid); | ||||||
|  |  | ||||||
|   whereWolf(ret,predicate,iftrue,iffalse); |   where(ret,predicate,iftrue,iffalse); | ||||||
|  |  | ||||||
|   return ret; |   return ret; | ||||||
| } | } | ||||||
|   | |||||||
| @@ -1,174 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid  |  | ||||||
|  |  | ||||||
|     Source file: ./lib/lattice/PaddedCell.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2019 |  | ||||||
|  |  | ||||||
| Author: Peter Boyle pboyle@bnl.gov |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
| *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
| #pragma once |  | ||||||
|  |  | ||||||
| #include<Grid/cshift/Cshift.h> |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| //Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions |  | ||||||
| template<typename vobj> |  | ||||||
| struct CshiftImplBase{ |  | ||||||
|   virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0; |  | ||||||
|   virtual ~CshiftImplBase(){} |  | ||||||
| }; |  | ||||||
| template<typename vobj> |  | ||||||
| struct CshiftImplDefault: public CshiftImplBase<vobj>{ |  | ||||||
|   Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); } |  | ||||||
| }; |  | ||||||
| template<typename Gimpl> |  | ||||||
| struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{ |  | ||||||
|   typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); } |  | ||||||
| };   |  | ||||||
|  |  | ||||||
| class PaddedCell { |  | ||||||
| public: |  | ||||||
|   GridCartesian * unpadded_grid; |  | ||||||
|   int dims; |  | ||||||
|   int depth; |  | ||||||
|   std::vector<GridCartesian *> grids; |  | ||||||
|  |  | ||||||
|   ~PaddedCell() |  | ||||||
|   { |  | ||||||
|     DeleteGrids(); |  | ||||||
|   } |  | ||||||
|   PaddedCell(int _depth,GridCartesian *_grid) |  | ||||||
|   { |  | ||||||
|     unpadded_grid = _grid; |  | ||||||
|     depth=_depth; |  | ||||||
|     dims=_grid->Nd(); |  | ||||||
|     AllocateGrids(); |  | ||||||
|     Coordinate local     =unpadded_grid->LocalDimensions(); |  | ||||||
|     for(int d=0;d<dims;d++){ |  | ||||||
|       assert(local[d]>=depth); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|   void DeleteGrids(void) |  | ||||||
|   { |  | ||||||
|     for(int d=0;d<grids.size();d++){ |  | ||||||
|       delete grids[d]; |  | ||||||
|     } |  | ||||||
|     grids.resize(0); |  | ||||||
|   }; |  | ||||||
|   void AllocateGrids(void) |  | ||||||
|   { |  | ||||||
|     Coordinate local     =unpadded_grid->LocalDimensions(); |  | ||||||
|     Coordinate simd      =unpadded_grid->_simd_layout; |  | ||||||
|     Coordinate processors=unpadded_grid->_processors; |  | ||||||
|     Coordinate plocal    =unpadded_grid->LocalDimensions(); |  | ||||||
|     Coordinate global(dims); |  | ||||||
|  |  | ||||||
|     // expand up one dim at a time |  | ||||||
|     for(int d=0;d<dims;d++){ |  | ||||||
|  |  | ||||||
|       plocal[d] += 2*depth;  |  | ||||||
|  |  | ||||||
|       for(int d=0;d<dims;d++){ |  | ||||||
| 	global[d] = plocal[d]*processors[d]; |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|       grids.push_back(new GridCartesian(global,simd,processors)); |  | ||||||
|     } |  | ||||||
|   }; |  | ||||||
|   template<class vobj> |  | ||||||
|   inline Lattice<vobj> Extract(const Lattice<vobj> &in) const |  | ||||||
|   { |  | ||||||
|     Lattice<vobj> out(unpadded_grid); |  | ||||||
|  |  | ||||||
|     Coordinate local     =unpadded_grid->LocalDimensions(); |  | ||||||
|     Coordinate fll(dims,depth); // depends on the MPI spread |  | ||||||
|     Coordinate tll(dims,0); // depends on the MPI spread |  | ||||||
|     localCopyRegion(in,out,fll,tll,local); |  | ||||||
|     return out; |  | ||||||
|   } |  | ||||||
|   template<class vobj> |  | ||||||
|   inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const |  | ||||||
|   { |  | ||||||
|     GridBase *old_grid = in.Grid(); |  | ||||||
|     int dims = old_grid->Nd(); |  | ||||||
|     Lattice<vobj> tmp = in; |  | ||||||
|     for(int d=0;d<dims;d++){ |  | ||||||
|       tmp = Expand(d,tmp,cshift); // rvalue && assignment |  | ||||||
|     } |  | ||||||
|     return tmp; |  | ||||||
|   } |  | ||||||
|   // expand up one dim at a time |  | ||||||
|   template<class vobj> |  | ||||||
|   inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const |  | ||||||
|   { |  | ||||||
|     GridBase *old_grid = in.Grid(); |  | ||||||
|     GridCartesian *new_grid = grids[dim];//These are new grids |  | ||||||
|     Lattice<vobj>  padded(new_grid); |  | ||||||
|     Lattice<vobj> shifted(old_grid);     |  | ||||||
|     Coordinate local     =old_grid->LocalDimensions(); |  | ||||||
|     Coordinate plocal    =new_grid->LocalDimensions(); |  | ||||||
|     if(dim==0) conformable(old_grid,unpadded_grid); |  | ||||||
|     else       conformable(old_grid,grids[dim-1]); |  | ||||||
|  |  | ||||||
|     std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl; |  | ||||||
|  |  | ||||||
|     double tins=0, tshift=0; |  | ||||||
|      |  | ||||||
|     // Middle bit |  | ||||||
|     double t = usecond(); |  | ||||||
|     for(int x=0;x<local[dim];x++){ |  | ||||||
|       InsertSliceLocal(in,padded,x,depth+x,dim); |  | ||||||
|     } |  | ||||||
|     tins += usecond() - t; |  | ||||||
|      |  | ||||||
|     // High bit |  | ||||||
|     t = usecond(); |  | ||||||
|     shifted = cshift.Cshift(in,dim,depth); |  | ||||||
|     tshift += usecond() - t; |  | ||||||
|  |  | ||||||
|     t=usecond(); |  | ||||||
|     for(int x=0;x<depth;x++){ |  | ||||||
|       InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim); |  | ||||||
|     } |  | ||||||
|     tins += usecond() - t; |  | ||||||
|      |  | ||||||
|     // Low bit |  | ||||||
|     t = usecond(); |  | ||||||
|     shifted = cshift.Cshift(in,dim,-depth); |  | ||||||
|     tshift += usecond() - t; |  | ||||||
|      |  | ||||||
|     t = usecond(); |  | ||||||
|     for(int x=0;x<depth;x++){ |  | ||||||
|       InsertSliceLocal(shifted,padded,x,x,dim); |  | ||||||
|     } |  | ||||||
|     tins += usecond() - t; |  | ||||||
|  |  | ||||||
|     std::cout << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl; |  | ||||||
|      |  | ||||||
|     return padded; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
| }; |  | ||||||
|   |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
|  |  | ||||||
| @@ -65,40 +65,29 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL"); | |||||||
| GridLogger GridLogError  (1, "Error" , GridLogColours, "RED"); | GridLogger GridLogError  (1, "Error" , GridLogColours, "RED"); | ||||||
| GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW"); | GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW"); | ||||||
| GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL"); | GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL"); | ||||||
| GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL"); |  | ||||||
| GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL"); |  | ||||||
| GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE"); | GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE"); | ||||||
| GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN"); | GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN"); | ||||||
| GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE"); |  | ||||||
| GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE"); | GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE"); | ||||||
| GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE"); | GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE"); | ||||||
| GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE"); |  | ||||||
|  |  | ||||||
| void GridLogConfigure(std::vector<std::string> &logstreams) { | void GridLogConfigure(std::vector<std::string> &logstreams) { | ||||||
|   GridLogError.Active(1); |   GridLogError.Active(0); | ||||||
|   GridLogWarning.Active(0); |   GridLogWarning.Active(0); | ||||||
|   GridLogMessage.Active(1); // at least the messages should be always on |   GridLogMessage.Active(1); // at least the messages should be always on | ||||||
|   GridLogMemory.Active(0);  |  | ||||||
|   GridLogTracing.Active(0);  |  | ||||||
|   GridLogIterative.Active(0); |   GridLogIterative.Active(0); | ||||||
|   GridLogDebug.Active(0); |   GridLogDebug.Active(0); | ||||||
|   GridLogPerformance.Active(0); |   GridLogPerformance.Active(0); | ||||||
|   GridLogDslash.Active(0); |  | ||||||
|   GridLogIntegrator.Active(1); |   GridLogIntegrator.Active(1); | ||||||
|   GridLogColours.Active(0); |   GridLogColours.Active(0); | ||||||
|   GridLogHMC.Active(1); |  | ||||||
|  |  | ||||||
|   for (int i = 0; i < logstreams.size(); i++) { |   for (int i = 0; i < logstreams.size(); i++) { | ||||||
|     if (logstreams[i] == std::string("Tracing"))     GridLogTracing.Active(1); |     if (logstreams[i] == std::string("Error"))       GridLogError.Active(1); | ||||||
|     if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1); |  | ||||||
|     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1); |     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1); | ||||||
|     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0); |     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0); | ||||||
|     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1); |     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1); | ||||||
|     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1); |     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1); | ||||||
|     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1); |     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1); | ||||||
|     if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1); |     if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1); | ||||||
|     if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0); |  | ||||||
|     if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0); |  | ||||||
|     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1); |     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1); | ||||||
|   } |   } | ||||||
| } | } | ||||||
|   | |||||||
| @@ -138,8 +138,7 @@ public: | |||||||
|         stream << std::setw(log.topWidth); |         stream << std::setw(log.topWidth); | ||||||
|       } |       } | ||||||
|       stream << log.topName << log.background()<< " : "; |       stream << log.topName << log.background()<< " : "; | ||||||
|       //      stream << log.colour() <<  std::left; |       stream << log.colour() <<  std::left; | ||||||
|       stream <<  std::left; |  | ||||||
|       if (log.chanWidth > 0) |       if (log.chanWidth > 0) | ||||||
|       { |       { | ||||||
|         stream << std::setw(log.chanWidth); |         stream << std::setw(log.chanWidth); | ||||||
| @@ -154,9 +153,9 @@ public: | |||||||
| 	stream << log.evidence() | 	stream << log.evidence() | ||||||
| 	       << now	       << log.background() << " : " ; | 	       << now	       << log.background() << " : " ; | ||||||
|       } |       } | ||||||
|       //      stream << log.colour(); |       stream << log.colour(); | ||||||
|       stream <<  std::right; |  | ||||||
|       stream.flags(f); |       stream.flags(f); | ||||||
|  |  | ||||||
|       return stream; |       return stream; | ||||||
|     } else {  |     } else {  | ||||||
|       return devnull; |       return devnull; | ||||||
| @@ -179,53 +178,14 @@ extern GridLogger GridLogSolver; | |||||||
| extern GridLogger GridLogError; | extern GridLogger GridLogError; | ||||||
| extern GridLogger GridLogWarning; | extern GridLogger GridLogWarning; | ||||||
| extern GridLogger GridLogMessage; | extern GridLogger GridLogMessage; | ||||||
| extern GridLogger GridLogDebug; | extern GridLogger GridLogDebug  ; | ||||||
| extern GridLogger GridLogPerformance; | extern GridLogger GridLogPerformance; | ||||||
| extern GridLogger GridLogDslash; | extern GridLogger GridLogIterative  ; | ||||||
| extern GridLogger GridLogIterative; | extern GridLogger GridLogIntegrator  ; | ||||||
| extern GridLogger GridLogIntegrator; |  | ||||||
| extern GridLogger GridLogHMC; |  | ||||||
| extern GridLogger GridLogMemory; |  | ||||||
| extern GridLogger GridLogTracing; |  | ||||||
| extern Colours    GridLogColours; | extern Colours    GridLogColours; | ||||||
|  |  | ||||||
| std::string demangle(const char* name) ; | std::string demangle(const char* name) ; | ||||||
|  |  | ||||||
| template<typename... Args> |  | ||||||
| inline std::string sjoin(Args&&... args) noexcept { |  | ||||||
|     std::ostringstream msg; |  | ||||||
|     (msg << ... << args); |  | ||||||
|     return msg.str(); |  | ||||||
| } |  | ||||||
|  |  | ||||||
| /*!  @brief make log messages work like python print */ |  | ||||||
| template <typename... Args> |  | ||||||
| inline void Grid_log(Args&&... args) { |  | ||||||
|     std::string msg = sjoin(std::forward<Args>(args)...); |  | ||||||
|     std::cout << GridLogMessage << msg << std::endl; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| /*!  @brief make warning messages work like python print */ |  | ||||||
| template <typename... Args> |  | ||||||
| inline void Grid_warn(Args&&... args) { |  | ||||||
|     std::string msg = sjoin(std::forward<Args>(args)...); |  | ||||||
|     std::cout << "\033[33m" << GridLogWarning << msg << "\033[0m" << std::endl; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| /*!  @brief make error messages work like python print */ |  | ||||||
| template <typename... Args> |  | ||||||
| inline void Grid_error(Args&&... args) { |  | ||||||
|     std::string msg = sjoin(std::forward<Args>(args)...); |  | ||||||
|     std::cout << "\033[31m" << GridLogError << msg << "\033[0m" << std::endl; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| /*!  @brief make pass messages work like python print */ |  | ||||||
| template <typename... Args> |  | ||||||
| inline void Grid_pass(Args&&... args) { |  | ||||||
|     std::string msg = sjoin(std::forward<Args>(args)...); |  | ||||||
|     std::cout << "\033[32m" << GridLogMessage << msg << "\033[0m" << std::endl; |  | ||||||
| } |  | ||||||
|  |  | ||||||
| #define _NBACKTRACE (256) | #define _NBACKTRACE (256) | ||||||
| extern void * Grid_backtrace_buffer[_NBACKTRACE]; | extern void * Grid_backtrace_buffer[_NBACKTRACE]; | ||||||
|  |  | ||||||
|   | |||||||
| @@ -31,7 +31,6 @@ directory | |||||||
| #include <fstream> | #include <fstream> | ||||||
| #include <iomanip> | #include <iomanip> | ||||||
| #include <iostream> | #include <iostream> | ||||||
| #include <string> |  | ||||||
| #include <map> | #include <map> | ||||||
|  |  | ||||||
| #include <pwd.h> | #include <pwd.h> | ||||||
| @@ -124,7 +123,7 @@ assert(GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) < 1.0e-5); | |||||||
|  //////////////////////////////////////////////////////////// |  //////////////////////////////////////////////////////////// | ||||||
|  // Helper to fill out metadata |  // Helper to fill out metadata | ||||||
|  //////////////////////////////////////////////////////////// |  //////////////////////////////////////////////////////////// | ||||||
| template<class vobj> void ScidacMetaData(Lattice<vobj> & field, |  template<class vobj> void ScidacMetaData(Lattice<vobj> & field, | ||||||
| 					  FieldMetaData &header, | 					  FieldMetaData &header, | ||||||
| 					  scidacRecord & _scidacRecord, | 					  scidacRecord & _scidacRecord, | ||||||
| 					  scidacFile   & _scidacFile)  | 					  scidacFile   & _scidacFile)  | ||||||
| @@ -577,8 +576,6 @@ class ScidacReader : public GridLimeReader { | |||||||
|     std::string rec_name(ILDG_BINARY_DATA); |     std::string rec_name(ILDG_BINARY_DATA); | ||||||
|     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) {  |     while ( limeReaderNextRecord(LimeR) == LIME_SUCCESS ) {  | ||||||
|       if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) { |       if ( !strncmp(limeReaderType(LimeR), rec_name.c_str(),strlen(rec_name.c_str()) )  ) { | ||||||
|   // in principle should do the line below, but that breaks backard compatibility with old data |  | ||||||
|   // skipPastObjectRecord(std::string(GRID_FIELD_NORM)); |  | ||||||
| 	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM)); | 	skipPastObjectRecord(std::string(SCIDAC_CHECKSUM)); | ||||||
| 	return; | 	return; | ||||||
|       } |       } | ||||||
| @@ -622,12 +619,12 @@ class IldgWriter : public ScidacWriter { | |||||||
|   // Don't require scidac records EXCEPT checksum |   // Don't require scidac records EXCEPT checksum | ||||||
|   // Use Grid MetaData object if present. |   // Use Grid MetaData object if present. | ||||||
|   //////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////// | ||||||
|   template <class stats = PeriodicGaugeStatistics> |   template <class vsimd> | ||||||
|   void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,int sequence,std::string LFN,std::string description)  |   void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description)  | ||||||
|   { |   { | ||||||
|     GridBase * grid = Umu.Grid(); |     GridBase * grid = Umu.Grid(); | ||||||
|     typedef Lattice<vLorentzColourMatrixD> GaugeField; |     typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField; | ||||||
|     typedef vLorentzColourMatrixD vobj; |     typedef iLorentzColourMatrix<vsimd> vobj; | ||||||
|     typedef typename vobj::scalar_object sobj; |     typedef typename vobj::scalar_object sobj; | ||||||
|  |  | ||||||
|     //////////////////////////////////////// |     //////////////////////////////////////// | ||||||
| @@ -639,9 +636,6 @@ class IldgWriter : public ScidacWriter { | |||||||
|  |  | ||||||
|     ScidacMetaData(Umu,header,_scidacRecord,_scidacFile); |     ScidacMetaData(Umu,header,_scidacRecord,_scidacFile); | ||||||
|  |  | ||||||
|     stats Stats; |  | ||||||
|     Stats(Umu,header); |  | ||||||
|      |  | ||||||
|     std::string format = header.floating_point; |     std::string format = header.floating_point; | ||||||
|     header.ensemble_id    = description; |     header.ensemble_id    = description; | ||||||
|     header.ensemble_label = description; |     header.ensemble_label = description; | ||||||
| @@ -655,8 +649,7 @@ class IldgWriter : public ScidacWriter { | |||||||
|     // Fill ILDG header data struct |     // Fill ILDG header data struct | ||||||
|     ////////////////////////////////////////////////////// |     ////////////////////////////////////////////////////// | ||||||
|     ildgFormat ildgfmt ; |     ildgFormat ildgfmt ; | ||||||
|     const std::string stNC = std::to_string( Nc ) ; |     ildgfmt.field     = std::string("su3gauge"); | ||||||
|     ildgfmt.field          = std::string("su"+stNC+"gauge"); |  | ||||||
|  |  | ||||||
|     if ( format == std::string("IEEE32BIG") ) {  |     if ( format == std::string("IEEE32BIG") ) {  | ||||||
|       ildgfmt.precision = 32; |       ildgfmt.precision = 32; | ||||||
| @@ -712,10 +705,10 @@ class IldgReader : public GridLimeReader { | |||||||
|   // Else use ILDG MetaData object if present. |   // Else use ILDG MetaData object if present. | ||||||
|   // Else use SciDAC MetaData object if present. |   // Else use SciDAC MetaData object if present. | ||||||
|   //////////////////////////////////////////////////////////////// |   //////////////////////////////////////////////////////////////// | ||||||
|   template <class stats = PeriodicGaugeStatistics> |   template <class vsimd> | ||||||
|   void readConfiguration(Lattice<vLorentzColourMatrixD> &Umu, FieldMetaData &FieldMetaData_) { |   void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, FieldMetaData &FieldMetaData_) { | ||||||
|  |  | ||||||
|     typedef Lattice<vLorentzColourMatrixD > GaugeField; |     typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField; | ||||||
|     typedef typename GaugeField::vector_object  vobj; |     typedef typename GaugeField::vector_object  vobj; | ||||||
|     typedef typename vobj::scalar_object sobj; |     typedef typename vobj::scalar_object sobj; | ||||||
|  |  | ||||||
| @@ -873,8 +866,7 @@ class IldgReader : public GridLimeReader { | |||||||
|     } else {  |     } else {  | ||||||
|  |  | ||||||
|       assert(found_ildgFormat); |       assert(found_ildgFormat); | ||||||
|       const std::string stNC = std::to_string( Nc ) ; |       assert ( ildgFormat_.field == std::string("su3gauge") ); | ||||||
|       assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") ); |  | ||||||
|  |  | ||||||
|       /////////////////////////////////////////////////////////////////////////////////////// |       /////////////////////////////////////////////////////////////////////////////////////// | ||||||
|       // Populate our Grid metadata as best we can |       // Populate our Grid metadata as best we can | ||||||
| @@ -882,7 +874,7 @@ class IldgReader : public GridLimeReader { | |||||||
|  |  | ||||||
|       std::ostringstream vers; vers << ildgFormat_.version; |       std::ostringstream vers; vers << ildgFormat_.version; | ||||||
|       FieldMetaData_.hdr_version = vers.str(); |       FieldMetaData_.hdr_version = vers.str(); | ||||||
|       FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC); |       FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3"); | ||||||
|  |  | ||||||
|       FieldMetaData_.nd=4; |       FieldMetaData_.nd=4; | ||||||
|       FieldMetaData_.dimension.resize(4); |       FieldMetaData_.dimension.resize(4); | ||||||
| @@ -929,8 +921,7 @@ class IldgReader : public GridLimeReader { | |||||||
|  |  | ||||||
|     if ( found_FieldMetaData || found_usqcdInfo ) { |     if ( found_FieldMetaData || found_usqcdInfo ) { | ||||||
|       FieldMetaData checker; |       FieldMetaData checker; | ||||||
|       stats Stats; |       GaugeStatistics(Umu,checker); | ||||||
|       Stats(Umu,checker); |  | ||||||
|       assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5); |       assert(fabs(checker.plaquette  - FieldMetaData_.plaquette )<1.0e-5); | ||||||
|       assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5); |       assert(fabs(checker.link_trace - FieldMetaData_.link_trace)<1.0e-5); | ||||||
|       std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl; |       std::cout << GridLogMessage<<"Plaquette and link trace match " << std::endl; | ||||||
|   | |||||||
| @@ -6,8 +6,8 @@ | |||||||
|  |  | ||||||
|     Copyright (C) 2015 |     Copyright (C) 2015 | ||||||
|  |  | ||||||
|  |  | ||||||
|     Author: Peter Boyle <paboyle@ph.ed.ac.uk> |     Author: Peter Boyle <paboyle@ph.ed.ac.uk> | ||||||
|     Author: Jamie Hudspith <renwick.james.hudspth@gmail.com> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |     This program is free software; you can redistribute it and/or modify | ||||||
|     it under the terms of the GNU General Public License as published by |     it under the terms of the GNU General Public License as published by | ||||||
| @@ -128,7 +128,7 @@ inline void MachineCharacteristics(FieldMetaData &header) | |||||||
|   std::time_t t = std::time(nullptr); |   std::time_t t = std::time(nullptr); | ||||||
|   std::tm tm_ = *std::localtime(&t); |   std::tm tm_ = *std::localtime(&t); | ||||||
|   std::ostringstream oss;  |   std::ostringstream oss;  | ||||||
|   oss << std::put_time(&tm_, "%c %Z"); |   //      oss << std::put_time(&tm_, "%c %Z"); | ||||||
|   header.creation_date = oss.str(); |   header.creation_date = oss.str(); | ||||||
|   header.archive_date  = header.creation_date; |   header.archive_date  = header.creation_date; | ||||||
|  |  | ||||||
| @@ -176,18 +176,29 @@ template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMet | |||||||
|   GridMetaData(grid,header);  |   GridMetaData(grid,header);  | ||||||
|   MachineCharacteristics(header); |   MachineCharacteristics(header); | ||||||
| } | } | ||||||
| template<class Impl> | inline void GaugeStatistics(Lattice<vLorentzColourMatrixF> & data,FieldMetaData &header) | ||||||
| class GaugeStatistics |  | ||||||
| { | { | ||||||
| public: |   // How to convert data precision etc... | ||||||
|   void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header) |   header.link_trace=WilsonLoops<PeriodicGimplF>::linkTrace(data); | ||||||
|   { |   header.plaquette =WilsonLoops<PeriodicGimplF>::avgPlaquette(data); | ||||||
|     header.link_trace = WilsonLoops<Impl>::linkTrace(data); | } | ||||||
|     header.plaquette  = WilsonLoops<Impl>::avgPlaquette(data); | inline void GaugeStatistics(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header) | ||||||
|   } | { | ||||||
| }; |   // How to convert data precision etc... | ||||||
| typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics; |   header.link_trace=WilsonLoops<PeriodicGimplD>::linkTrace(data); | ||||||
| typedef GaugeStatistics<ConjugateGimplD> ConjugateGaugeStatistics; |   header.plaquette =WilsonLoops<PeriodicGimplD>::avgPlaquette(data); | ||||||
|  | } | ||||||
|  | template<> inline void PrepareMetaData<vLorentzColourMatrixF>(Lattice<vLorentzColourMatrixF> & field, FieldMetaData &header) | ||||||
|  | { | ||||||
|  |     | ||||||
|  |   GridBase *grid = field.Grid(); | ||||||
|  |   std::string format = getFormatString<vLorentzColourMatrixF>(); | ||||||
|  |   header.floating_point = format; | ||||||
|  |   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac | ||||||
|  |   GridMetaData(grid,header);  | ||||||
|  |   GaugeStatistics(field,header); | ||||||
|  |   MachineCharacteristics(header); | ||||||
|  | } | ||||||
| template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header) | template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header) | ||||||
| { | { | ||||||
|   GridBase *grid = field.Grid(); |   GridBase *grid = field.Grid(); | ||||||
| @@ -195,6 +206,7 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo | |||||||
|   header.floating_point = format; |   header.floating_point = format; | ||||||
|   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac |   header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac | ||||||
|   GridMetaData(grid,header);  |   GridMetaData(grid,header);  | ||||||
|  |   GaugeStatistics(field,header); | ||||||
|   MachineCharacteristics(header); |   MachineCharacteristics(header); | ||||||
| } | } | ||||||
|  |  | ||||||
| @@ -203,24 +215,20 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo | |||||||
| ////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////// | ||||||
| inline void reconstruct3(LorentzColourMatrix & cm) | inline void reconstruct3(LorentzColourMatrix & cm) | ||||||
| { | { | ||||||
|   assert( Nc < 4 && Nc > 1 ) ; |   const int x=0; | ||||||
|  |   const int y=1; | ||||||
|  |   const int z=2; | ||||||
|   for(int mu=0;mu<Nd;mu++){ |   for(int mu=0;mu<Nd;mu++){ | ||||||
|     #if Nc == 2 |     cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy | ||||||
|       cm(mu)()(1,0) = -adj(cm(mu)()(0,y)) ; |     cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz | ||||||
|       cm(mu)()(1,1) =  adj(cm(mu)()(0,x)) ; |     cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx | ||||||
|     #else |  | ||||||
|       const int x=0 , y=1 , z=2 ; // a little disinenuous labelling |  | ||||||
|       cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy |  | ||||||
|       cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz |  | ||||||
|       cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx |  | ||||||
|     #endif |  | ||||||
|   } |   } | ||||||
| } | } | ||||||
|  |  | ||||||
| //////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////// | ||||||
| // Some data types for intermediate storage | // Some data types for intermediate storage | ||||||
| //////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////// | ||||||
| template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, Nc-1>, Nd >; | template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >; | ||||||
|  |  | ||||||
| typedef iLorentzColour2x3<Complex>  LorentzColour2x3; | typedef iLorentzColour2x3<Complex>  LorentzColour2x3; | ||||||
| typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F; | typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F; | ||||||
| @@ -282,6 +290,7 @@ struct GaugeSimpleMunger{ | |||||||
|  |  | ||||||
| template <class fobj, class sobj> | template <class fobj, class sobj> | ||||||
| struct GaugeSimpleUnmunger { | struct GaugeSimpleUnmunger { | ||||||
|  |  | ||||||
|   void operator()(sobj &in, fobj &out) { |   void operator()(sobj &in, fobj &out) { | ||||||
|     for (int mu = 0; mu < Nd; mu++) { |     for (int mu = 0; mu < Nd; mu++) { | ||||||
|       for (int i = 0; i < Nc; i++) { |       for (int i = 0; i < Nc; i++) { | ||||||
| @@ -320,8 +329,8 @@ template<class fobj,class sobj> | |||||||
| struct Gauge3x2munger{ | struct Gauge3x2munger{ | ||||||
|   void operator() (fobj &in,sobj &out){ |   void operator() (fobj &in,sobj &out){ | ||||||
|     for(int mu=0;mu<Nd;mu++){ |     for(int mu=0;mu<Nd;mu++){ | ||||||
|       for(int i=0;i<Nc-1;i++){ |       for(int i=0;i<2;i++){ | ||||||
| 	for(int j=0;j<Nc;j++){ | 	for(int j=0;j<3;j++){ | ||||||
| 	  out(mu)()(i,j) = in(mu)(i)(j); | 	  out(mu)()(i,j) = in(mu)(i)(j); | ||||||
| 	}} | 	}} | ||||||
|     } |     } | ||||||
| @@ -333,8 +342,8 @@ template<class fobj,class sobj> | |||||||
| struct Gauge3x2unmunger{ | struct Gauge3x2unmunger{ | ||||||
|   void operator() (sobj &in,fobj &out){ |   void operator() (sobj &in,fobj &out){ | ||||||
|     for(int mu=0;mu<Nd;mu++){ |     for(int mu=0;mu<Nd;mu++){ | ||||||
|       for(int i=0;i<Nc-1;i++){ |       for(int i=0;i<2;i++){ | ||||||
| 	for(int j=0;j<Nc;j++){ | 	for(int j=0;j<3;j++){ | ||||||
| 	  out(mu)(i)(j) = in(mu)()(i,j); | 	  out(mu)(i)(j) = in(mu)()(i,j); | ||||||
| 	}} | 	}} | ||||||
|     } |     } | ||||||
|   | |||||||
| @@ -9,7 +9,6 @@ | |||||||
|     Author: Matt Spraggs <matthew.spraggs@gmail.com> |     Author: Matt Spraggs <matthew.spraggs@gmail.com> | ||||||
|     Author: Peter Boyle <paboyle@ph.ed.ac.uk> |     Author: Peter Boyle <paboyle@ph.ed.ac.uk> | ||||||
|     Author: paboyle <paboyle@ph.ed.ac.uk> |     Author: paboyle <paboyle@ph.ed.ac.uk> | ||||||
|     Author: Jamie Hudspith <renwick.james.hudspth@gmail.com> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |     This program is free software; you can redistribute it and/or modify | ||||||
|     it under the terms of the GNU General Public License as published by |     it under the terms of the GNU General Public License as published by | ||||||
| @@ -31,8 +30,6 @@ | |||||||
| #ifndef GRID_NERSC_IO_H | #ifndef GRID_NERSC_IO_H | ||||||
| #define GRID_NERSC_IO_H | #define GRID_NERSC_IO_H | ||||||
|  |  | ||||||
| #include <string> |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| using namespace Grid; | using namespace Grid; | ||||||
| @@ -42,10 +39,6 @@ using namespace Grid; | |||||||
| //////////////////////////////////////////////////////////////////////////////// | //////////////////////////////////////////////////////////////////////////////// | ||||||
| class NerscIO : public BinaryIO {  | class NerscIO : public BinaryIO {  | ||||||
| public: | public: | ||||||
|   typedef Lattice<vLorentzColourMatrixD> GaugeField; |  | ||||||
|  |  | ||||||
|   // Enable/disable exiting if the plaquette in the header does not match the value computed (default true) |  | ||||||
|   static bool & exitOnReadPlaquetteMismatch(){ static bool v=true; return v; } |  | ||||||
|  |  | ||||||
|   static inline void truncate(std::string file){ |   static inline void truncate(std::string file){ | ||||||
|     std::ofstream fout(file,std::ios::out); |     std::ofstream fout(file,std::ios::out); | ||||||
| @@ -136,12 +129,12 @@ public: | |||||||
|   // Now the meat: the object readers |   // Now the meat: the object readers | ||||||
|   ///////////////////////////////////////////////////////////////////////////////////////////////////////////////////// |   ///////////////////////////////////////////////////////////////////////////////////////////////////////////////////// | ||||||
|  |  | ||||||
|   template<class GaugeStats=PeriodicGaugeStatistics> |   template<class vsimd> | ||||||
|   static inline void readConfiguration(GaugeField &Umu, |   static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, | ||||||
| 				       FieldMetaData& header, | 				       FieldMetaData& header, | ||||||
| 				       std::string file, | 				       std::string file) | ||||||
| 				       GaugeStats GaugeStatisticsCalculator=GaugeStats()) |  | ||||||
|   { |   { | ||||||
|  |     typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField; | ||||||
|  |  | ||||||
|     GridBase *grid = Umu.Grid(); |     GridBase *grid = Umu.Grid(); | ||||||
|     uint64_t offset = readHeader(file,Umu.Grid(),header); |     uint64_t offset = readHeader(file,Umu.Grid(),header); | ||||||
| @@ -150,35 +143,33 @@ public: | |||||||
|  |  | ||||||
|     std::string format(header.floating_point); |     std::string format(header.floating_point); | ||||||
|  |  | ||||||
|     const int ieee32big = (format == std::string("IEEE32BIG")); |     int ieee32big = (format == std::string("IEEE32BIG")); | ||||||
|     const int ieee32    = (format == std::string("IEEE32")); |     int ieee32    = (format == std::string("IEEE32")); | ||||||
|     const int ieee64big = (format == std::string("IEEE64BIG")); |     int ieee64big = (format == std::string("IEEE64BIG")); | ||||||
|     const int ieee64    = (format == std::string("IEEE64") || \ |     int ieee64    = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE")); | ||||||
| 			   format == std::string("IEEE64LITTLE")); |  | ||||||
|  |  | ||||||
|     uint32_t nersc_csum,scidac_csuma,scidac_csumb; |     uint32_t nersc_csum,scidac_csuma,scidac_csumb; | ||||||
|     // depending on datatype, set up munger; |     // depending on datatype, set up munger; | ||||||
|     // munger is a function of <floating point, Real, data_type> |     // munger is a function of <floating point, Real, data_type> | ||||||
|     const std::string stNC = std::to_string( Nc ) ; |     if ( header.data_type == std::string("4D_SU3_GAUGE") ) { | ||||||
|     if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) { |  | ||||||
|       if ( ieee32 || ieee32big ) { |       if ( ieee32 || ieee32big ) { | ||||||
| 	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F>  | 	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F>  | ||||||
| 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format, | 	  (Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format, | ||||||
| 	   nersc_csum,scidac_csuma,scidac_csumb); | 	   nersc_csum,scidac_csuma,scidac_csumb); | ||||||
|       } |       } | ||||||
|       if ( ieee64 || ieee64big ) { |       if ( ieee64 || ieee64big ) { | ||||||
| 	BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3D>  | 	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D>  | ||||||
| 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format, | 	  (Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format, | ||||||
| 	   nersc_csum,scidac_csuma,scidac_csumb); | 	   nersc_csum,scidac_csuma,scidac_csumb); | ||||||
|       } |       } | ||||||
|     } else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) { |     } else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) { | ||||||
|       if ( ieee32 || ieee32big ) { |       if ( ieee32 || ieee32big ) { | ||||||
| 	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF> | 	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF> | ||||||
| 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format, | 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format, | ||||||
| 	   nersc_csum,scidac_csuma,scidac_csumb); | 	   nersc_csum,scidac_csuma,scidac_csumb); | ||||||
|       } |       } | ||||||
|       if ( ieee64 || ieee64big ) { |       if ( ieee64 || ieee64big ) { | ||||||
| 	BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixD> | 	BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD> | ||||||
| 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format, | 	  (Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format, | ||||||
| 	   nersc_csum,scidac_csuma,scidac_csumb); | 	   nersc_csum,scidac_csuma,scidac_csumb); | ||||||
|       } |       } | ||||||
| @@ -186,7 +177,7 @@ public: | |||||||
|       assert(0); |       assert(0); | ||||||
|     } |     } | ||||||
|  |  | ||||||
|     GaugeStats Stats; Stats(Umu,clone); |     GaugeStatistics(Umu,clone); | ||||||
|  |  | ||||||
|     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec |     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec | ||||||
| 	     <<" header   "<<std::hex<<header.checksum<<std::dec <<std::endl; | 	     <<" header   "<<std::hex<<header.checksum<<std::dec <<std::endl; | ||||||
| @@ -205,40 +196,31 @@ public: | |||||||
|       std::cerr << " nersc_csum  " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl; |       std::cerr << " nersc_csum  " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl; | ||||||
|       exit(0); |       exit(0); | ||||||
|     } |     } | ||||||
|     if(exitOnReadPlaquetteMismatch()) assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 ); |     assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 ); | ||||||
|     assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 ); |     assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 ); | ||||||
|     assert(nersc_csum == header.checksum ); |     assert(nersc_csum == header.checksum ); | ||||||
|        |        | ||||||
|     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl; |     std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl; | ||||||
|   } |   } | ||||||
|  |  | ||||||
|   // Preferred interface |   template<class vsimd> | ||||||
|   template<class GaugeStats=PeriodicGaugeStatistics> |   static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu, | ||||||
|   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu, |  | ||||||
| 					std::string file,  |  | ||||||
| 					std::string ens_label = std::string("DWF"), |  | ||||||
| 					std::string ens_id = std::string("UKQCD"), |  | ||||||
| 					unsigned int sequence_number = 1) |  | ||||||
|   { |  | ||||||
|     writeConfiguration(Umu,file,0,1,ens_label,ens_id,sequence_number); |  | ||||||
|   } |  | ||||||
|   template<class GaugeStats=PeriodicGaugeStatistics> |  | ||||||
|   static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu, |  | ||||||
| 					std::string file,  | 					std::string file,  | ||||||
| 					int two_row, | 					int two_row, | ||||||
| 					int bits32, | 					int bits32) | ||||||
| 					std::string ens_label = std::string("DWF"), |  | ||||||
| 					std::string ens_id = std::string("UKQCD"), |  | ||||||
| 					unsigned int sequence_number = 1) |  | ||||||
|   { |   { | ||||||
|     typedef vLorentzColourMatrixD vobj; |     typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField; | ||||||
|  |  | ||||||
|  |     typedef iLorentzColourMatrix<vsimd> vobj; | ||||||
|     typedef typename vobj::scalar_object sobj; |     typedef typename vobj::scalar_object sobj; | ||||||
|  |  | ||||||
|     FieldMetaData header; |     FieldMetaData header; | ||||||
|     header.sequence_number = sequence_number; |     /////////////////////////////////////////// | ||||||
|     header.ensemble_id     = ens_id; |     // Following should become arguments | ||||||
|     header.ensemble_label  = ens_label; |     /////////////////////////////////////////// | ||||||
|     header.hdr_version     = "1.0" ; |     header.sequence_number = 1; | ||||||
|  |     header.ensemble_id     = "UKQCD"; | ||||||
|  |     header.ensemble_label  = "DWF"; | ||||||
|  |  | ||||||
|     typedef LorentzColourMatrixD fobj3D; |     typedef LorentzColourMatrixD fobj3D; | ||||||
|     typedef LorentzColour2x3D    fobj2D; |     typedef LorentzColour2x3D    fobj2D; | ||||||
| @@ -247,39 +229,28 @@ public: | |||||||
|  |  | ||||||
|     GridMetaData(grid,header); |     GridMetaData(grid,header); | ||||||
|     assert(header.nd==4); |     assert(header.nd==4); | ||||||
|     GaugeStats Stats; Stats(Umu,header); |     GaugeStatistics(Umu,header); | ||||||
|     MachineCharacteristics(header); |     MachineCharacteristics(header); | ||||||
|  |  | ||||||
|     uint64_t offset; | 	uint64_t offset; | ||||||
|  |  | ||||||
|     // Sod it -- always write NcxNc double |     // Sod it -- always write 3x3 double | ||||||
|     header.floating_point  = std::string("IEEE64BIG"); |     header.floating_point = std::string("IEEE64BIG"); | ||||||
|     const std::string stNC = std::to_string( Nc ) ; |     header.data_type      = std::string("4D_SU3_GAUGE_3x3"); | ||||||
|     if( two_row ) { |     GaugeSimpleUnmunger<fobj3D,sobj> munge; | ||||||
|       header.data_type = std::string("4D_SU" + stNC + "_GAUGE" ); | 	if ( grid->IsBoss() ) {  | ||||||
|     } else { | 	  truncate(file); | ||||||
|       header.data_type = std::string("4D_SU" + stNC + "_GAUGE_" + stNC + "x" + stNC ); |     offset = writeHeader(header,file); | ||||||
|     } | 	} | ||||||
|     if ( grid->IsBoss() ) {  | 	grid->Broadcast(0,(void *)&offset,sizeof(offset)); | ||||||
|       truncate(file); |  | ||||||
|       offset = writeHeader(header,file); |  | ||||||
|     } |  | ||||||
|     grid->Broadcast(0,(void *)&offset,sizeof(offset)); |  | ||||||
|  |  | ||||||
|     uint32_t nersc_csum,scidac_csuma,scidac_csumb; |     uint32_t nersc_csum,scidac_csuma,scidac_csumb; | ||||||
|     if( two_row ) { |     BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point, | ||||||
|       Gauge3x2unmunger<fobj2D,sobj> munge; | 					      nersc_csum,scidac_csuma,scidac_csumb); | ||||||
|       BinaryIO::writeLatticeObject<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point, |  | ||||||
| 						nersc_csum,scidac_csuma,scidac_csumb); |  | ||||||
|     } else { |  | ||||||
|       GaugeSimpleUnmunger<fobj3D,sobj> munge; |  | ||||||
|       BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point, |  | ||||||
| 						nersc_csum,scidac_csuma,scidac_csumb); |  | ||||||
|     } |  | ||||||
|     header.checksum = nersc_csum; |     header.checksum = nersc_csum; | ||||||
|     if ( grid->IsBoss() ) {  | 	if ( grid->IsBoss() ) {  | ||||||
|       writeHeader(header,file); |     writeHeader(header,file); | ||||||
|     } | 	} | ||||||
|  |  | ||||||
|     std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum " |     std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum " | ||||||
| 	     <<std::hex<<header.checksum | 	     <<std::hex<<header.checksum | ||||||
| @@ -307,7 +278,8 @@ public: | |||||||
|     header.plaquette=0.0; |     header.plaquette=0.0; | ||||||
|     MachineCharacteristics(header); |     MachineCharacteristics(header); | ||||||
|  |  | ||||||
|     uint64_t offset; | 	uint64_t offset; | ||||||
|  |    | ||||||
| #ifdef RNG_RANLUX | #ifdef RNG_RANLUX | ||||||
|     header.floating_point = std::string("UINT64"); |     header.floating_point = std::string("UINT64"); | ||||||
|     header.data_type      = std::string("RANLUX48"); |     header.data_type      = std::string("RANLUX48"); | ||||||
| @@ -347,7 +319,7 @@ public: | |||||||
|  |  | ||||||
|     GridBase *grid = parallel.Grid(); |     GridBase *grid = parallel.Grid(); | ||||||
|  |  | ||||||
|     uint64_t offset = readHeader(file,grid,header); | 	uint64_t offset = readHeader(file,grid,header); | ||||||
|  |  | ||||||
|     FieldMetaData clone(header); |     FieldMetaData clone(header); | ||||||
|  |  | ||||||
|   | |||||||
| @@ -154,7 +154,7 @@ public: | |||||||
|     grid->Barrier(); timer.Stop(); |     grid->Barrier(); timer.Stop(); | ||||||
|     std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: redistribute overhead " << timer.Elapsed() << std::endl; |     std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: redistribute overhead " << timer.Elapsed() << std::endl; | ||||||
|  |  | ||||||
|     PeriodicGaugeStatistics Stats; Stats(Umu, clone); |     GaugeStatistics(Umu, clone); | ||||||
|  |  | ||||||
|     RealD plaq_diff = fabs(clone.plaquette - header.plaquette); |     RealD plaq_diff = fabs(clone.plaquette - header.plaquette); | ||||||
|  |  | ||||||
|   | |||||||
| @@ -208,7 +208,7 @@ public: | |||||||
|  |  | ||||||
|     FieldMetaData clone(header); |     FieldMetaData clone(header); | ||||||
|  |  | ||||||
|     PeriodicGaugeStatistics Stats; Stats(Umu, clone); |     GaugeStatistics(Umu, clone); | ||||||
|  |  | ||||||
|     RealD plaq_diff = fabs(clone.plaquette - header.plaquette); |     RealD plaq_diff = fabs(clone.plaquette - header.plaquette); | ||||||
|  |  | ||||||
|   | |||||||
| @@ -27,12 +27,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk> | |||||||
| /*  END LEGAL */ | /*  END LEGAL */ | ||||||
|  |  | ||||||
| #include <Grid/GridCore.h> | #include <Grid/GridCore.h> | ||||||
|  |  | ||||||
| #include <Grid/perfmon/Timer.h> |  | ||||||
| #include <Grid/perfmon/PerfCount.h> | #include <Grid/perfmon/PerfCount.h> | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| GridTimePoint theProgramStart = GridClock::now(); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16)) | #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16)) | ||||||
| #define RawConfig(A,B) (A<<8|B) | #define RawConfig(A,B) (A<<8|B) | ||||||
|   | |||||||
| @@ -30,12 +30,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk> | |||||||
| #ifndef GRID_PERFCOUNT_H | #ifndef GRID_PERFCOUNT_H | ||||||
| #define GRID_PERFCOUNT_H | #define GRID_PERFCOUNT_H | ||||||
|  |  | ||||||
|  |  | ||||||
| #ifndef __SSC_START |  | ||||||
| #define __SSC_START |  | ||||||
| #define __SSC_STOP |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
| #include <sys/time.h> | #include <sys/time.h> | ||||||
| #include <ctime> | #include <ctime> | ||||||
| #include <chrono> | #include <chrono> | ||||||
| @@ -78,9 +72,17 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid, | |||||||
| inline uint64_t cyclecount(void){  | inline uint64_t cyclecount(void){  | ||||||
|   return 0; |   return 0; | ||||||
| } | } | ||||||
|  | #define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx") | ||||||
|  | #define __SSC_STOP  __SSC_MARK(0x110) | ||||||
|  | #define __SSC_START __SSC_MARK(0x111) | ||||||
|  |  | ||||||
|  |  | ||||||
| #else | #else | ||||||
|  |  | ||||||
|  | #define __SSC_MARK(mark)  | ||||||
|  | #define __SSC_STOP   | ||||||
|  | #define __SSC_START  | ||||||
|  |  | ||||||
| /* | /* | ||||||
|  * cycle counters arch dependent |  * cycle counters arch dependent | ||||||
|  */ |  */ | ||||||
|   | |||||||
| @@ -35,8 +35,17 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk> | |||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid) | NAMESPACE_BEGIN(Grid) | ||||||
|  |  | ||||||
| //typedef  std::chrono::system_clock          GridClock; | // Dress the output; use std::chrono | ||||||
| typedef  std::chrono::high_resolution_clock   GridClock; | // C++11 time facilities better? | ||||||
|  | inline double usecond(void) { | ||||||
|  |   struct timeval tv; | ||||||
|  | #ifdef TIMERS_ON | ||||||
|  |   gettimeofday(&tv,NULL); | ||||||
|  | #endif | ||||||
|  |   return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec; | ||||||
|  | } | ||||||
|  |  | ||||||
|  | typedef  std::chrono::system_clock          GridClock; | ||||||
| typedef  std::chrono::time_point<GridClock> GridTimePoint; | typedef  std::chrono::time_point<GridClock> GridTimePoint; | ||||||
|  |  | ||||||
| typedef  std::chrono::seconds               GridSecs; | typedef  std::chrono::seconds               GridSecs; | ||||||
| @@ -44,15 +53,6 @@ typedef  std::chrono::milliseconds          GridMillisecs; | |||||||
| typedef  std::chrono::microseconds          GridUsecs; | typedef  std::chrono::microseconds          GridUsecs; | ||||||
| typedef  std::chrono::microseconds          GridTime; | typedef  std::chrono::microseconds          GridTime; | ||||||
|  |  | ||||||
| extern GridTimePoint theProgramStart; |  | ||||||
| // Dress the output; use std::chrono |  | ||||||
| // C++11 time facilities better? |  | ||||||
| inline double usecond(void) { |  | ||||||
|   auto usecs = std::chrono::duration_cast<GridUsecs>(GridClock::now()-theProgramStart);  |  | ||||||
|   return 1.0*usecs.count(); |  | ||||||
| } |  | ||||||
|  |  | ||||||
|  |  | ||||||
| inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time) | inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time) | ||||||
| { | { | ||||||
|   stream << time.count()<<" s"; |   stream << time.count()<<" s"; | ||||||
|   | |||||||
| @@ -1,70 +0,0 @@ | |||||||
| #pragma once |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| #ifdef GRID_TRACING_NVTX |  | ||||||
| #include <nvToolsExt.h> |  | ||||||
| class GridTracer { |  | ||||||
| public: |  | ||||||
|   GridTracer(const char* name) { |  | ||||||
|     nvtxRangePushA(name); |  | ||||||
|   } |  | ||||||
|   ~GridTracer() { |  | ||||||
|     nvtxRangePop(); |  | ||||||
|   } |  | ||||||
| }; |  | ||||||
| inline void tracePush(const char *name) { nvtxRangePushA(name); } |  | ||||||
| inline void tracePop(const char *name) { nvtxRangePop(); } |  | ||||||
| inline int  traceStart(const char *name) {  } |  | ||||||
| inline void traceStop(int ID) {  } |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
| #ifdef GRID_TRACING_ROCTX |  | ||||||
| #include <roctracer/roctx.h> |  | ||||||
| class GridTracer { |  | ||||||
|  public: |  | ||||||
|   GridTracer(const char* name) { |  | ||||||
|     roctxRangePushA(name); |  | ||||||
|     std::cout << "roctxRangePush "<<name<<std::endl; |  | ||||||
|   } |  | ||||||
|   ~GridTracer() { |  | ||||||
|     roctxRangePop(); |  | ||||||
|     std::cout << "roctxRangePop "<<std::endl; |  | ||||||
|   } |  | ||||||
| }; |  | ||||||
| inline void tracePush(const char *name) { roctxRangePushA(name); } |  | ||||||
| inline void tracePop(const char *name) { roctxRangePop(); } |  | ||||||
| inline int  traceStart(const char *name) { return roctxRangeStart(name); } |  | ||||||
| inline void traceStop(int ID) { roctxRangeStop(ID); } |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
| #ifdef GRID_TRACING_TIMER |  | ||||||
| class GridTracer { |  | ||||||
|  public: |  | ||||||
|   const char *name; |  | ||||||
|   double elapsed; |  | ||||||
|   GridTracer(const char* _name) { |  | ||||||
|     name = _name; |  | ||||||
|     elapsed=-usecond(); |  | ||||||
|   } |  | ||||||
|   ~GridTracer() { |  | ||||||
|     elapsed+=usecond(); |  | ||||||
|     std::cout << GridLogTracing << name << " took " <<elapsed<< " us" <<std::endl; |  | ||||||
|   } |  | ||||||
| }; |  | ||||||
| inline void tracePush(const char *name) {  } |  | ||||||
| inline void tracePop(const char *name) {  } |  | ||||||
| inline int  traceStart(const char *name) { return 0; } |  | ||||||
| inline void traceStop(int ID) {  } |  | ||||||
| #endif |  | ||||||
|  |  | ||||||
| #ifdef GRID_TRACING_NONE |  | ||||||
| #define GRID_TRACE(name)  |  | ||||||
| inline void tracePush(const char *name) {  } |  | ||||||
| inline void tracePop(const char *name) {  } |  | ||||||
| inline int  traceStart(const char *name) { return 0;  } |  | ||||||
| inline void traceStop(int ID) {  } |  | ||||||
| #else |  | ||||||
| #define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name); |  | ||||||
| #endif |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
| @@ -16,12 +16,8 @@ | |||||||
|  |  | ||||||
| #ifdef __NVCC__ | #ifdef __NVCC__ | ||||||
| #pragma push | #pragma push | ||||||
| #ifdef __NVCC_DIAG_PRAGMA_SUPPORT__ |  | ||||||
| #pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning" |  | ||||||
| #else |  | ||||||
| #pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning" | #pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning" | ||||||
| #endif | #endif | ||||||
| #endif |  | ||||||
|  |  | ||||||
| #include "pugixml.h" | #include "pugixml.h" | ||||||
|  |  | ||||||
|   | |||||||
							
								
								
									
										146
									
								
								Grid/qcd/QCD.h
									
									
									
									
									
								
							
							
						
						
									
										146
									
								
								Grid/qcd/QCD.h
									
									
									
									
									
								
							| @@ -63,7 +63,6 @@ static constexpr int Ngp=2; // gparity index range | |||||||
| #define ColourIndex  (2) | #define ColourIndex  (2) | ||||||
| #define SpinIndex    (1) | #define SpinIndex    (1) | ||||||
| #define LorentzIndex (0) | #define LorentzIndex (0) | ||||||
| #define GparityFlavourIndex (0) |  | ||||||
|  |  | ||||||
| // Also should make these a named enum type | // Also should make these a named enum type | ||||||
| static constexpr int DaggerNo=0; | static constexpr int DaggerNo=0; | ||||||
| @@ -81,15 +80,6 @@ template<typename T> struct isSpinor { | |||||||
| template <typename T> using IfSpinor    = Invoke<std::enable_if< isSpinor<T>::value,int> > ; | template <typename T> using IfSpinor    = Invoke<std::enable_if< isSpinor<T>::value,int> > ; | ||||||
| template <typename T> using IfNotSpinor = Invoke<std::enable_if<!isSpinor<T>::value,int> > ; | template <typename T> using IfNotSpinor = Invoke<std::enable_if<!isSpinor<T>::value,int> > ; | ||||||
|  |  | ||||||
| const int CoarseIndex = 4; |  | ||||||
| template<typename T> struct isCoarsened { |  | ||||||
|    static constexpr bool value = (CoarseIndex<=T::TensorLevel); |  | ||||||
| }; |  | ||||||
| template <typename T> using IfCoarsened    = Invoke<std::enable_if< isCoarsened<T>::value,int> > ; |  | ||||||
| template <typename T> using IfNotCoarsened = Invoke<std::enable_if<!isCoarsened<T>::value,int> > ; |  | ||||||
|  |  | ||||||
| const int GparityFlavourTensorIndex = 3; //TensorLevel counts from the bottom! |  | ||||||
|  |  | ||||||
| // ChrisK very keen to add extra space for Gparity doubling. | // ChrisK very keen to add extra space for Gparity doubling. | ||||||
| // | // | ||||||
| // Also add domain wall index, in a way where Wilson operator  | // Also add domain wall index, in a way where Wilson operator  | ||||||
| @@ -104,7 +94,6 @@ template<typename vtype> using iSpinMatrix                = iScalar<iMatrix<iSca | |||||||
| template<typename vtype> using iColourMatrix              = iScalar<iScalar<iMatrix<vtype, Nc> > > ; | template<typename vtype> using iColourMatrix              = iScalar<iScalar<iMatrix<vtype, Nc> > > ; | ||||||
| template<typename vtype> using iSpinColourMatrix          = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >; | template<typename vtype> using iSpinColourMatrix          = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >; | ||||||
| template<typename vtype> using iLorentzColourMatrix       = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ; | template<typename vtype> using iLorentzColourMatrix       = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ; | ||||||
| template<typename vtype> using iLorentzComplex            = iVector<iScalar<iScalar<vtype> >, Nd > ; |  | ||||||
| template<typename vtype> using iDoubleStoredColourMatrix  = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ; | template<typename vtype> using iDoubleStoredColourMatrix  = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ; | ||||||
| template<typename vtype> using iSpinVector                = iScalar<iVector<iScalar<vtype>, Ns> >; | template<typename vtype> using iSpinVector                = iScalar<iVector<iScalar<vtype>, Ns> >; | ||||||
| template<typename vtype> using iColourVector              = iScalar<iScalar<iVector<vtype, Nc> > >; | template<typename vtype> using iColourVector              = iScalar<iScalar<iVector<vtype, Nc> > >; | ||||||
| @@ -114,10 +103,8 @@ template<typename vtype> using iHalfSpinColourVector      = iScalar<iVector<iVec | |||||||
|     template<typename vtype> using iSpinColourSpinColourMatrix  = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >; |     template<typename vtype> using iSpinColourSpinColourMatrix  = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >; | ||||||
|  |  | ||||||
|  |  | ||||||
| template<typename vtype> using iGparityFlavourVector                = iVector<iScalar<iScalar<vtype> >, Ngp>; |  | ||||||
| template<typename vtype> using iGparitySpinColourVector       = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >; | template<typename vtype> using iGparitySpinColourVector       = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >; | ||||||
| template<typename vtype> using iGparityHalfSpinColourVector   = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >; | template<typename vtype> using iGparityHalfSpinColourVector   = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >; | ||||||
| template<typename vtype> using iGparityFlavourMatrix = iMatrix<iScalar<iScalar<vtype> >, Ngp>; |  | ||||||
|  |  | ||||||
| // Spin matrix | // Spin matrix | ||||||
| typedef iSpinMatrix<Complex  >          SpinMatrix; | typedef iSpinMatrix<Complex  >          SpinMatrix; | ||||||
| @@ -127,7 +114,6 @@ typedef iSpinMatrix<ComplexD >          SpinMatrixD; | |||||||
| typedef iSpinMatrix<vComplex >          vSpinMatrix; | typedef iSpinMatrix<vComplex >          vSpinMatrix; | ||||||
| typedef iSpinMatrix<vComplexF>          vSpinMatrixF; | typedef iSpinMatrix<vComplexF>          vSpinMatrixF; | ||||||
| typedef iSpinMatrix<vComplexD>          vSpinMatrixD; | typedef iSpinMatrix<vComplexD>          vSpinMatrixD; | ||||||
| typedef iSpinMatrix<vComplexD2>         vSpinMatrixD2; |  | ||||||
|  |  | ||||||
| // Colour Matrix | // Colour Matrix | ||||||
| typedef iColourMatrix<Complex  >        ColourMatrix; | typedef iColourMatrix<Complex  >        ColourMatrix; | ||||||
| @@ -137,7 +123,6 @@ typedef iColourMatrix<ComplexD >        ColourMatrixD; | |||||||
| typedef iColourMatrix<vComplex >        vColourMatrix; | typedef iColourMatrix<vComplex >        vColourMatrix; | ||||||
| typedef iColourMatrix<vComplexF>        vColourMatrixF; | typedef iColourMatrix<vComplexF>        vColourMatrixF; | ||||||
| typedef iColourMatrix<vComplexD>        vColourMatrixD; | typedef iColourMatrix<vComplexD>        vColourMatrixD; | ||||||
| typedef iColourMatrix<vComplexD2>       vColourMatrixD2; |  | ||||||
|  |  | ||||||
| // SpinColour matrix | // SpinColour matrix | ||||||
| typedef iSpinColourMatrix<Complex  >    SpinColourMatrix; | typedef iSpinColourMatrix<Complex  >    SpinColourMatrix; | ||||||
| @@ -147,7 +132,6 @@ typedef iSpinColourMatrix<ComplexD >    SpinColourMatrixD; | |||||||
| typedef iSpinColourMatrix<vComplex >    vSpinColourMatrix; | typedef iSpinColourMatrix<vComplex >    vSpinColourMatrix; | ||||||
| typedef iSpinColourMatrix<vComplexF>    vSpinColourMatrixF; | typedef iSpinColourMatrix<vComplexF>    vSpinColourMatrixF; | ||||||
| typedef iSpinColourMatrix<vComplexD>    vSpinColourMatrixD; | typedef iSpinColourMatrix<vComplexD>    vSpinColourMatrixD; | ||||||
| typedef iSpinColourMatrix<vComplexD2>   vSpinColourMatrixD2; |  | ||||||
|  |  | ||||||
| // SpinColourSpinColour matrix | // SpinColourSpinColour matrix | ||||||
| typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix; | typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix; | ||||||
| @@ -157,7 +141,6 @@ typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD; | |||||||
| typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix; | typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix; | ||||||
| typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF; | typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF; | ||||||
| typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD; | typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD; | ||||||
| typedef iSpinColourSpinColourMatrix<vComplexD2>   vSpinColourSpinColourMatrixD2; |  | ||||||
|  |  | ||||||
| // SpinColourSpinColour matrix | // SpinColourSpinColour matrix | ||||||
| typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix; | typedef iSpinColourSpinColourMatrix<Complex  >    SpinColourSpinColourMatrix; | ||||||
| @@ -167,47 +150,24 @@ typedef iSpinColourSpinColourMatrix<ComplexD >    SpinColourSpinColourMatrixD; | |||||||
| typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix; | typedef iSpinColourSpinColourMatrix<vComplex >    vSpinColourSpinColourMatrix; | ||||||
| typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF; | typedef iSpinColourSpinColourMatrix<vComplexF>    vSpinColourSpinColourMatrixF; | ||||||
| typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD; | typedef iSpinColourSpinColourMatrix<vComplexD>    vSpinColourSpinColourMatrixD; | ||||||
| typedef iSpinColourSpinColourMatrix<vComplexD2>   vSpinColourSpinColourMatrixD2; |  | ||||||
|  |  | ||||||
| // LorentzColour | // LorentzColour | ||||||
| typedef iLorentzColourMatrix<Complex  > LorentzColourMatrix; | typedef iLorentzColourMatrix<Complex  > LorentzColourMatrix; | ||||||
| typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF; | typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF; | ||||||
| typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD; | typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD; | ||||||
|  |  | ||||||
| typedef iLorentzColourMatrix<vComplex >  vLorentzColourMatrix; | typedef iLorentzColourMatrix<vComplex > vLorentzColourMatrix; | ||||||
| typedef iLorentzColourMatrix<vComplexF>  vLorentzColourMatrixF; | typedef iLorentzColourMatrix<vComplexF> vLorentzColourMatrixF; | ||||||
| typedef iLorentzColourMatrix<vComplexD>  vLorentzColourMatrixD; | typedef iLorentzColourMatrix<vComplexD> vLorentzColourMatrixD; | ||||||
| typedef iLorentzColourMatrix<vComplexD2> vLorentzColourMatrixD2; |  | ||||||
|  |  | ||||||
| // LorentzComplex |  | ||||||
| typedef iLorentzComplex<Complex  > LorentzComplex; |  | ||||||
| typedef iLorentzComplex<ComplexF > LorentzComplexF; |  | ||||||
| typedef iLorentzComplex<ComplexD > LorentzComplexD; |  | ||||||
|  |  | ||||||
| typedef iLorentzComplex<vComplex > vLorentzComplex; |  | ||||||
| typedef iLorentzComplex<vComplexF> vLorentzComplexF; |  | ||||||
| typedef iLorentzComplex<vComplexD> vLorentzComplexD; |  | ||||||
|  |  | ||||||
| // DoubleStored gauge field | // DoubleStored gauge field | ||||||
| typedef iDoubleStoredColourMatrix<Complex  > DoubleStoredColourMatrix; | typedef iDoubleStoredColourMatrix<Complex  > DoubleStoredColourMatrix; | ||||||
| typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF; | typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF; | ||||||
| typedef iDoubleStoredColourMatrix<ComplexD > DoubleStoredColourMatrixD; | typedef iDoubleStoredColourMatrix<ComplexD > DoubleStoredColourMatrixD; | ||||||
|  |  | ||||||
| typedef iDoubleStoredColourMatrix<vComplex >  vDoubleStoredColourMatrix; | typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix; | ||||||
| typedef iDoubleStoredColourMatrix<vComplexF>  vDoubleStoredColourMatrixF; | typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF; | ||||||
| typedef iDoubleStoredColourMatrix<vComplexD>  vDoubleStoredColourMatrixD; | typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD; | ||||||
| typedef iDoubleStoredColourMatrix<vComplexD2> vDoubleStoredColourMatrixD2; |  | ||||||
|  |  | ||||||
| //G-parity flavour matrix |  | ||||||
| typedef iGparityFlavourMatrix<Complex> GparityFlavourMatrix; |  | ||||||
| typedef iGparityFlavourMatrix<ComplexF> GparityFlavourMatrixF; |  | ||||||
| typedef iGparityFlavourMatrix<ComplexD> GparityFlavourMatrixD; |  | ||||||
|  |  | ||||||
| typedef iGparityFlavourMatrix<vComplex>   vGparityFlavourMatrix; |  | ||||||
| typedef iGparityFlavourMatrix<vComplexF>  vGparityFlavourMatrixF; |  | ||||||
| typedef iGparityFlavourMatrix<vComplexD>  vGparityFlavourMatrixD; |  | ||||||
| typedef iGparityFlavourMatrix<vComplexD2> vGparityFlavourMatrixD2; |  | ||||||
|  |  | ||||||
|  |  | ||||||
| // Spin vector | // Spin vector | ||||||
| typedef iSpinVector<Complex >           SpinVector; | typedef iSpinVector<Complex >           SpinVector; | ||||||
| @@ -217,7 +177,6 @@ typedef iSpinVector<ComplexD>           SpinVectorD; | |||||||
| typedef iSpinVector<vComplex >           vSpinVector; | typedef iSpinVector<vComplex >           vSpinVector; | ||||||
| typedef iSpinVector<vComplexF>           vSpinVectorF; | typedef iSpinVector<vComplexF>           vSpinVectorF; | ||||||
| typedef iSpinVector<vComplexD>           vSpinVectorD; | typedef iSpinVector<vComplexD>           vSpinVectorD; | ||||||
| typedef iSpinVector<vComplexD2>          vSpinVectorD2; |  | ||||||
|  |  | ||||||
| // Colour vector | // Colour vector | ||||||
| typedef iColourVector<Complex >         ColourVector; | typedef iColourVector<Complex >         ColourVector; | ||||||
| @@ -227,7 +186,6 @@ typedef iColourVector<ComplexD>         ColourVectorD; | |||||||
| typedef iColourVector<vComplex >         vColourVector; | typedef iColourVector<vComplex >         vColourVector; | ||||||
| typedef iColourVector<vComplexF>         vColourVectorF; | typedef iColourVector<vComplexF>         vColourVectorF; | ||||||
| typedef iColourVector<vComplexD>         vColourVectorD; | typedef iColourVector<vComplexD>         vColourVectorD; | ||||||
| typedef iColourVector<vComplexD2>        vColourVectorD2; |  | ||||||
|  |  | ||||||
| // SpinColourVector | // SpinColourVector | ||||||
| typedef iSpinColourVector<Complex >     SpinColourVector; | typedef iSpinColourVector<Complex >     SpinColourVector; | ||||||
| @@ -237,7 +195,6 @@ typedef iSpinColourVector<ComplexD>     SpinColourVectorD; | |||||||
| typedef iSpinColourVector<vComplex >     vSpinColourVector; | typedef iSpinColourVector<vComplex >     vSpinColourVector; | ||||||
| typedef iSpinColourVector<vComplexF>     vSpinColourVectorF; | typedef iSpinColourVector<vComplexF>     vSpinColourVectorF; | ||||||
| typedef iSpinColourVector<vComplexD>     vSpinColourVectorD; | typedef iSpinColourVector<vComplexD>     vSpinColourVectorD; | ||||||
| typedef iSpinColourVector<vComplexD2>    vSpinColourVectorD2; |  | ||||||
|  |  | ||||||
| // HalfSpin vector | // HalfSpin vector | ||||||
| typedef iHalfSpinVector<Complex >       HalfSpinVector; | typedef iHalfSpinVector<Complex >       HalfSpinVector; | ||||||
| @@ -247,27 +204,15 @@ typedef iHalfSpinVector<ComplexD>       HalfSpinVectorD; | |||||||
| typedef iHalfSpinVector<vComplex >       vHalfSpinVector; | typedef iHalfSpinVector<vComplex >       vHalfSpinVector; | ||||||
| typedef iHalfSpinVector<vComplexF>       vHalfSpinVectorF; | typedef iHalfSpinVector<vComplexF>       vHalfSpinVectorF; | ||||||
| typedef iHalfSpinVector<vComplexD>       vHalfSpinVectorD; | typedef iHalfSpinVector<vComplexD>       vHalfSpinVectorD; | ||||||
| typedef iHalfSpinVector<vComplexD2>      vHalfSpinVectorD2; |  | ||||||
|  |  | ||||||
| // HalfSpinColour vector | // HalfSpinColour vector | ||||||
| typedef iHalfSpinColourVector<Complex > HalfSpinColourVector; | typedef iHalfSpinColourVector<Complex > HalfSpinColourVector; | ||||||
| typedef iHalfSpinColourVector<ComplexF> HalfSpinColourVectorF; | typedef iHalfSpinColourVector<ComplexF> HalfSpinColourVectorF; | ||||||
| typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD; | typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD; | ||||||
|      |      | ||||||
| typedef iHalfSpinColourVector<vComplex >  vHalfSpinColourVector; | typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector; | ||||||
| typedef iHalfSpinColourVector<vComplexF>  vHalfSpinColourVectorF; | typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF; | ||||||
| typedef iHalfSpinColourVector<vComplexD>  vHalfSpinColourVectorD; | typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD; | ||||||
| typedef iHalfSpinColourVector<vComplexD2> vHalfSpinColourVectorD2; |  | ||||||
|  |  | ||||||
| //G-parity flavour vector |  | ||||||
| typedef iGparityFlavourVector<Complex >         GparityFlavourVector; |  | ||||||
| typedef iGparityFlavourVector<ComplexF>         GparityFlavourVectorF; |  | ||||||
| typedef iGparityFlavourVector<ComplexD>         GparityFlavourVectorD; |  | ||||||
|  |  | ||||||
| typedef iGparityFlavourVector<vComplex >         vGparityFlavourVector; |  | ||||||
| typedef iGparityFlavourVector<vComplexF>         vGparityFlavourVectorF; |  | ||||||
| typedef iGparityFlavourVector<vComplexD>         vGparityFlavourVectorD; |  | ||||||
| typedef iGparityFlavourVector<vComplexD2>        vGparityFlavourVectorD2; |  | ||||||
|      |      | ||||||
| // singlets | // singlets | ||||||
| typedef iSinglet<Complex >         TComplex;     // FIXME This is painful. Tensor singlet complex type. | typedef iSinglet<Complex >         TComplex;     // FIXME This is painful. Tensor singlet complex type. | ||||||
| @@ -277,7 +222,6 @@ typedef iSinglet<ComplexD>         TComplexD;    // FIXME This is painful. Tenso | |||||||
| typedef iSinglet<vComplex >        vTComplex ;   // what if we don't know the tensor structure | typedef iSinglet<vComplex >        vTComplex ;   // what if we don't know the tensor structure | ||||||
| typedef iSinglet<vComplexF>        vTComplexF;   // what if we don't know the tensor structure | typedef iSinglet<vComplexF>        vTComplexF;   // what if we don't know the tensor structure | ||||||
| typedef iSinglet<vComplexD>        vTComplexD;   // what if we don't know the tensor structure | typedef iSinglet<vComplexD>        vTComplexD;   // what if we don't know the tensor structure | ||||||
| typedef iSinglet<vComplexD2>       vTComplexD2;   // what if we don't know the tensor structure |  | ||||||
|  |  | ||||||
| typedef iSinglet<Real >            TReal;        // Shouldn't need these; can I make it work without? | typedef iSinglet<Real >            TReal;        // Shouldn't need these; can I make it work without? | ||||||
| typedef iSinglet<RealF>            TRealF;       // Shouldn't need these; can I make it work without? | typedef iSinglet<RealF>            TRealF;       // Shouldn't need these; can I make it work without? | ||||||
| @@ -295,62 +239,47 @@ typedef iSinglet<Integer >         TInteger; | |||||||
| typedef Lattice<vColourMatrix>          LatticeColourMatrix; | typedef Lattice<vColourMatrix>          LatticeColourMatrix; | ||||||
| typedef Lattice<vColourMatrixF>         LatticeColourMatrixF; | typedef Lattice<vColourMatrixF>         LatticeColourMatrixF; | ||||||
| typedef Lattice<vColourMatrixD>         LatticeColourMatrixD; | typedef Lattice<vColourMatrixD>         LatticeColourMatrixD; | ||||||
| typedef Lattice<vColourMatrixD2>        LatticeColourMatrixD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vSpinMatrix>            LatticeSpinMatrix; | typedef Lattice<vSpinMatrix>            LatticeSpinMatrix; | ||||||
| typedef Lattice<vSpinMatrixF>           LatticeSpinMatrixF; | typedef Lattice<vSpinMatrixF>           LatticeSpinMatrixF; | ||||||
| typedef Lattice<vSpinMatrixD>           LatticeSpinMatrixD; | typedef Lattice<vSpinMatrixD>           LatticeSpinMatrixD; | ||||||
| typedef Lattice<vSpinMatrixD2>          LatticeSpinMatrixD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vSpinColourMatrix>      LatticeSpinColourMatrix; | typedef Lattice<vSpinColourMatrix>      LatticeSpinColourMatrix; | ||||||
| typedef Lattice<vSpinColourMatrixF>     LatticeSpinColourMatrixF; | typedef Lattice<vSpinColourMatrixF>     LatticeSpinColourMatrixF; | ||||||
| typedef Lattice<vSpinColourMatrixD>     LatticeSpinColourMatrixD; | typedef Lattice<vSpinColourMatrixD>     LatticeSpinColourMatrixD; | ||||||
| typedef Lattice<vSpinColourMatrixD2>    LatticeSpinColourMatrixD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vSpinColourSpinColourMatrix>      LatticeSpinColourSpinColourMatrix; | typedef Lattice<vSpinColourSpinColourMatrix>      LatticeSpinColourSpinColourMatrix; | ||||||
| typedef Lattice<vSpinColourSpinColourMatrixF>     LatticeSpinColourSpinColourMatrixF; | typedef Lattice<vSpinColourSpinColourMatrixF>     LatticeSpinColourSpinColourMatrixF; | ||||||
| typedef Lattice<vSpinColourSpinColourMatrixD>     LatticeSpinColourSpinColourMatrixD; | typedef Lattice<vSpinColourSpinColourMatrixD>     LatticeSpinColourSpinColourMatrixD; | ||||||
| typedef Lattice<vSpinColourSpinColourMatrixD2>    LatticeSpinColourSpinColourMatrixD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vLorentzColourMatrix>   LatticeLorentzColourMatrix; | typedef Lattice<vLorentzColourMatrix>  LatticeLorentzColourMatrix; | ||||||
| typedef Lattice<vLorentzColourMatrixF>  LatticeLorentzColourMatrixF; | typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF; | ||||||
| typedef Lattice<vLorentzColourMatrixD>  LatticeLorentzColourMatrixD; | typedef Lattice<vLorentzColourMatrixD> LatticeLorentzColourMatrixD; | ||||||
| typedef Lattice<vLorentzColourMatrixD2> LatticeLorentzColourMatrixD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vLorentzComplex>  LatticeLorentzComplex; |  | ||||||
| typedef Lattice<vLorentzComplexF> LatticeLorentzComplexF; |  | ||||||
| typedef Lattice<vLorentzComplexD> LatticeLorentzComplexD; |  | ||||||
|  |  | ||||||
| // DoubleStored gauge field | // DoubleStored gauge field | ||||||
| typedef Lattice<vDoubleStoredColourMatrix>   LatticeDoubleStoredColourMatrix; | typedef Lattice<vDoubleStoredColourMatrix>  LatticeDoubleStoredColourMatrix; | ||||||
| typedef Lattice<vDoubleStoredColourMatrixF>  LatticeDoubleStoredColourMatrixF; | typedef Lattice<vDoubleStoredColourMatrixF> LatticeDoubleStoredColourMatrixF; | ||||||
| typedef Lattice<vDoubleStoredColourMatrixD>  LatticeDoubleStoredColourMatrixD; | typedef Lattice<vDoubleStoredColourMatrixD> LatticeDoubleStoredColourMatrixD; | ||||||
| typedef Lattice<vDoubleStoredColourMatrixD2> LatticeDoubleStoredColourMatrixD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vSpinVector>            LatticeSpinVector; | typedef Lattice<vSpinVector>            LatticeSpinVector; | ||||||
| typedef Lattice<vSpinVectorF>           LatticeSpinVectorF; | typedef Lattice<vSpinVectorF>           LatticeSpinVectorF; | ||||||
| typedef Lattice<vSpinVectorD>           LatticeSpinVectorD; | typedef Lattice<vSpinVectorD>           LatticeSpinVectorD; | ||||||
| typedef Lattice<vSpinVectorD2>          LatticeSpinVectorD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vColourVector>          LatticeColourVector; | typedef Lattice<vColourVector>          LatticeColourVector; | ||||||
| typedef Lattice<vColourVectorF>         LatticeColourVectorF; | typedef Lattice<vColourVectorF>         LatticeColourVectorF; | ||||||
| typedef Lattice<vColourVectorD>         LatticeColourVectorD; | typedef Lattice<vColourVectorD>         LatticeColourVectorD; | ||||||
| typedef Lattice<vColourVectorD2>        LatticeColourVectorD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vSpinColourVector>      LatticeSpinColourVector; | typedef Lattice<vSpinColourVector>      LatticeSpinColourVector; | ||||||
| typedef Lattice<vSpinColourVectorF>     LatticeSpinColourVectorF; | typedef Lattice<vSpinColourVectorF>     LatticeSpinColourVectorF; | ||||||
| typedef Lattice<vSpinColourVectorD>     LatticeSpinColourVectorD; | typedef Lattice<vSpinColourVectorD>     LatticeSpinColourVectorD; | ||||||
| typedef Lattice<vSpinColourVectorD2>    LatticeSpinColourVectorD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vHalfSpinVector>        LatticeHalfSpinVector; | typedef Lattice<vHalfSpinVector>        LatticeHalfSpinVector; | ||||||
| typedef Lattice<vHalfSpinVectorF>       LatticeHalfSpinVectorF; | typedef Lattice<vHalfSpinVectorF>       LatticeHalfSpinVectorF; | ||||||
| typedef Lattice<vHalfSpinVectorD>       LatticeHalfSpinVectorD; | typedef Lattice<vHalfSpinVectorD>       LatticeHalfSpinVectorD; | ||||||
| typedef Lattice<vHalfSpinVectorD2>      LatticeHalfSpinVectorD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vHalfSpinColourVector>   LatticeHalfSpinColourVector; | typedef Lattice<vHalfSpinColourVector>  LatticeHalfSpinColourVector; | ||||||
| typedef Lattice<vHalfSpinColourVectorF>  LatticeHalfSpinColourVectorF; | typedef Lattice<vHalfSpinColourVectorF> LatticeHalfSpinColourVectorF; | ||||||
| typedef Lattice<vHalfSpinColourVectorD>  LatticeHalfSpinColourVectorD; | typedef Lattice<vHalfSpinColourVectorD> LatticeHalfSpinColourVectorD; | ||||||
| typedef Lattice<vHalfSpinColourVectorD2> LatticeHalfSpinColourVectorD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vTReal>            LatticeReal; | typedef Lattice<vTReal>            LatticeReal; | ||||||
| typedef Lattice<vTRealF>           LatticeRealF; | typedef Lattice<vTRealF>           LatticeRealF; | ||||||
| @@ -359,7 +288,6 @@ typedef Lattice<vTRealD>           LatticeRealD; | |||||||
| typedef Lattice<vTComplex>         LatticeComplex; | typedef Lattice<vTComplex>         LatticeComplex; | ||||||
| typedef Lattice<vTComplexF>        LatticeComplexF; | typedef Lattice<vTComplexF>        LatticeComplexF; | ||||||
| typedef Lattice<vTComplexD>        LatticeComplexD; | typedef Lattice<vTComplexD>        LatticeComplexD; | ||||||
| typedef Lattice<vTComplexD2>       LatticeComplexD2; |  | ||||||
|  |  | ||||||
| typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where" | typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where" | ||||||
|  |  | ||||||
| @@ -367,42 +295,37 @@ typedef Lattice<vTInteger>         LatticeInteger; // Predicates for "where" | |||||||
| /////////////////////////////////////////// | /////////////////////////////////////////// | ||||||
| // Physical names for things | // Physical names for things | ||||||
| /////////////////////////////////////////// | /////////////////////////////////////////// | ||||||
| typedef LatticeHalfSpinColourVector   LatticeHalfFermion; | typedef LatticeHalfSpinColourVector  LatticeHalfFermion; | ||||||
| typedef LatticeHalfSpinColourVectorF  LatticeHalfFermionF; | typedef LatticeHalfSpinColourVectorF LatticeHalfFermionF; | ||||||
| typedef LatticeHalfSpinColourVectorD  LatticeHalfFermionD; | typedef LatticeHalfSpinColourVectorF LatticeHalfFermionD; | ||||||
| typedef LatticeHalfSpinColourVectorD2 LatticeHalfFermionD2; |  | ||||||
|  |  | ||||||
| typedef LatticeSpinColourVector      LatticeFermion; | typedef LatticeSpinColourVector      LatticeFermion; | ||||||
| typedef LatticeSpinColourVectorF     LatticeFermionF; | typedef LatticeSpinColourVectorF     LatticeFermionF; | ||||||
| typedef LatticeSpinColourVectorD     LatticeFermionD; | typedef LatticeSpinColourVectorD     LatticeFermionD; | ||||||
| typedef LatticeSpinColourVectorD2    LatticeFermionD2; |  | ||||||
|  |  | ||||||
| typedef LatticeSpinColourMatrix                LatticePropagator; | typedef LatticeSpinColourMatrix                LatticePropagator; | ||||||
| typedef LatticeSpinColourMatrixF               LatticePropagatorF; | typedef LatticeSpinColourMatrixF               LatticePropagatorF; | ||||||
| typedef LatticeSpinColourMatrixD               LatticePropagatorD; | typedef LatticeSpinColourMatrixD               LatticePropagatorD; | ||||||
| typedef LatticeSpinColourMatrixD2              LatticePropagatorD2; |  | ||||||
|  |  | ||||||
| typedef LatticeLorentzColourMatrix             LatticeGaugeField; | typedef LatticeLorentzColourMatrix             LatticeGaugeField; | ||||||
| typedef LatticeLorentzColourMatrixF            LatticeGaugeFieldF; | typedef LatticeLorentzColourMatrixF            LatticeGaugeFieldF; | ||||||
| typedef LatticeLorentzColourMatrixD            LatticeGaugeFieldD; | typedef LatticeLorentzColourMatrixD            LatticeGaugeFieldD; | ||||||
| typedef LatticeLorentzColourMatrixD2           LatticeGaugeFieldD2; |  | ||||||
|  |  | ||||||
| typedef LatticeDoubleStoredColourMatrix        LatticeDoubledGaugeField; | typedef LatticeDoubleStoredColourMatrix        LatticeDoubledGaugeField; | ||||||
| typedef LatticeDoubleStoredColourMatrixF       LatticeDoubledGaugeFieldF; | typedef LatticeDoubleStoredColourMatrixF       LatticeDoubledGaugeFieldF; | ||||||
| typedef LatticeDoubleStoredColourMatrixD       LatticeDoubledGaugeFieldD; | typedef LatticeDoubleStoredColourMatrixD       LatticeDoubledGaugeFieldD; | ||||||
| typedef LatticeDoubleStoredColourMatrixD2      LatticeDoubledGaugeFieldD2; |  | ||||||
|  |  | ||||||
| template<class GF> using LorentzScalar = Lattice<iScalar<typename GF::vector_object::element> >; | template<class GF> using LorentzScalar = Lattice<iScalar<typename GF::vector_object::element> >; | ||||||
|  |  | ||||||
|  | // Uhgg... typing this hurt  ;) | ||||||
|  | // (my keyboard got burning hot when I typed this, must be the anti-Fermion) | ||||||
| typedef Lattice<vColourVector>          LatticeStaggeredFermion;     | typedef Lattice<vColourVector>          LatticeStaggeredFermion;     | ||||||
| typedef Lattice<vColourVectorF>         LatticeStaggeredFermionF;     | typedef Lattice<vColourVectorF>         LatticeStaggeredFermionF;     | ||||||
| typedef Lattice<vColourVectorD>         LatticeStaggeredFermionD;     | typedef Lattice<vColourVectorD>         LatticeStaggeredFermionD;     | ||||||
| typedef Lattice<vColourVectorD2>        LatticeStaggeredFermionD2;     |  | ||||||
|  |  | ||||||
| typedef Lattice<vColourMatrix>          LatticeStaggeredPropagator;  | typedef Lattice<vColourMatrix>          LatticeStaggeredPropagator;  | ||||||
| typedef Lattice<vColourMatrixF>         LatticeStaggeredPropagatorF;  | typedef Lattice<vColourMatrixF>         LatticeStaggeredPropagatorF;  | ||||||
| typedef Lattice<vColourMatrixD>         LatticeStaggeredPropagatorD;  | typedef Lattice<vColourMatrixD>         LatticeStaggeredPropagatorD;  | ||||||
| typedef Lattice<vColourMatrixD2>        LatticeStaggeredPropagatorD2;  |  | ||||||
|  |  | ||||||
| ////////////////////////////////////////////////////////////////////////////// | ////////////////////////////////////////////////////////////////////////////// | ||||||
| // Peek and Poke named after physics attributes | // Peek and Poke named after physics attributes | ||||||
| @@ -521,20 +444,9 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz | |||||||
| // Fermion <-> propagator assignements | // Fermion <-> propagator assignements | ||||||
| ////////////////////////////////////////////// | ////////////////////////////////////////////// | ||||||
| //template <class Prop, class Ferm> | //template <class Prop, class Ferm> | ||||||
| #define FAST_FERM_TO_PROP |  | ||||||
| template <class Fimpl> | template <class Fimpl> | ||||||
| void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c) | void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c) | ||||||
| { | { | ||||||
| #ifdef FAST_FERM_TO_PROP |  | ||||||
|   autoView(p_v,p,CpuWrite); |  | ||||||
|   autoView(f_v,f,CpuRead); |  | ||||||
|   thread_for(idx,p_v.oSites(),{ |  | ||||||
|       for(int ss = 0; ss < Ns; ++ss) { |  | ||||||
|       for(int cc = 0; cc < Fimpl::Dimension; ++cc) { |  | ||||||
| 	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.) |  | ||||||
|       }} |  | ||||||
|     }); |  | ||||||
| #else |  | ||||||
|   for(int j = 0; j < Ns; ++j) |   for(int j = 0; j < Ns; ++j) | ||||||
|     { |     { | ||||||
|       auto pjs = peekSpin(p, j, s); |       auto pjs = peekSpin(p, j, s); | ||||||
| @@ -546,23 +458,12 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio | |||||||
| 	} | 	} | ||||||
|       pokeSpin(p, pjs, j, s); |       pokeSpin(p, pjs, j, s); | ||||||
|     } |     } | ||||||
| #endif |  | ||||||
| } | } | ||||||
|      |      | ||||||
| //template <class Prop, class Ferm> | //template <class Prop, class Ferm> | ||||||
| template <class Fimpl> | template <class Fimpl> | ||||||
| void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c) | void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c) | ||||||
| { | { | ||||||
| #ifdef FAST_FERM_TO_PROP |  | ||||||
|   autoView(p_v,p,CpuRead); |  | ||||||
|   autoView(f_v,f,CpuWrite); |  | ||||||
|   thread_for(idx,p_v.oSites(),{ |  | ||||||
|       for(int ss = 0; ss < Ns; ++ss) { |  | ||||||
|       for(int cc = 0; cc < Fimpl::Dimension; ++cc) { |  | ||||||
| 	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index |  | ||||||
|       }} |  | ||||||
|     }); |  | ||||||
| #else |  | ||||||
|   for(int j = 0; j < Ns; ++j) |   for(int j = 0; j < Ns; ++j) | ||||||
|     { |     { | ||||||
|       auto pjs = peekSpin(p, j, s); |       auto pjs = peekSpin(p, j, s); | ||||||
| @@ -574,7 +475,6 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato | |||||||
| 	} | 	} | ||||||
|       pokeSpin(f, fj, j); |       pokeSpin(f, fj, j); | ||||||
|     } |     } | ||||||
| #endif |  | ||||||
| } | } | ||||||
|      |      | ||||||
| ////////////////////////////////////////////// | ////////////////////////////////////////////// | ||||||
|   | |||||||
| @@ -34,117 +34,21 @@ directory | |||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| /////////////////////////////////// |  | ||||||
| // Smart configuration base class |  | ||||||
| /////////////////////////////////// |  | ||||||
| template< class Field > |  | ||||||
| class ConfigurationBase |  | ||||||
| { |  | ||||||
| public: |  | ||||||
|   ConfigurationBase() {} |  | ||||||
|   virtual ~ConfigurationBase() {} |  | ||||||
|   virtual void set_Field(Field& U) =0; |  | ||||||
|   virtual void smeared_force(Field&) = 0; |  | ||||||
|   virtual Field& get_SmearedU() =0; |  | ||||||
|   virtual Field &get_U(bool smeared = false) = 0; |  | ||||||
| }; |  | ||||||
|  |  | ||||||
| template <class GaugeField > | template <class GaugeField > | ||||||
| class Action  | class Action  | ||||||
| { | { | ||||||
|  |  | ||||||
| public: | public: | ||||||
|   bool is_smeared = false; |   bool is_smeared = false; | ||||||
|   RealD deriv_norm_sum; |  | ||||||
|   RealD deriv_max_sum; |  | ||||||
|   RealD Fdt_norm_sum; |  | ||||||
|   RealD Fdt_max_sum; |  | ||||||
|   int   deriv_num; |  | ||||||
|   RealD deriv_us; |  | ||||||
|   RealD S_us; |  | ||||||
|   RealD refresh_us; |  | ||||||
|   void  reset_timer(void)        { |  | ||||||
|     deriv_us = S_us = refresh_us = 0.0; |  | ||||||
|     deriv_norm_sum = deriv_max_sum=0.0; |  | ||||||
|     Fdt_max_sum =  Fdt_norm_sum = 0.0; |  | ||||||
|     deriv_num=0; |  | ||||||
|   } |  | ||||||
|   void  deriv_log(RealD nrm, RealD max,RealD Fdt_nrm,RealD Fdt_max) { |  | ||||||
|     if ( max > deriv_max_sum ) { |  | ||||||
|       deriv_max_sum=max; |  | ||||||
|     } |  | ||||||
|     deriv_norm_sum+=nrm; |  | ||||||
|     if ( Fdt_max > Fdt_max_sum ) { |  | ||||||
|       Fdt_max_sum=Fdt_max; |  | ||||||
|     } |  | ||||||
|     Fdt_norm_sum+=Fdt_nrm; deriv_num++; |  | ||||||
|   } |  | ||||||
|   RealD deriv_max_average(void)       { return deriv_max_sum; }; |  | ||||||
|   RealD deriv_norm_average(void)      { return deriv_norm_sum/deriv_num; }; |  | ||||||
|   RealD Fdt_max_average(void)         { return Fdt_max_sum; }; |  | ||||||
|   RealD Fdt_norm_average(void)        { return Fdt_norm_sum/deriv_num; }; |  | ||||||
|   RealD deriv_timer(void)        { return deriv_us; }; |  | ||||||
|   RealD S_timer(void)            { return S_us; }; |  | ||||||
|   RealD refresh_timer(void)      { return refresh_us; }; |  | ||||||
|   void deriv_timer_start(void)   { deriv_us-=usecond(); } |  | ||||||
|   void deriv_timer_stop(void)    { deriv_us+=usecond(); } |  | ||||||
|   void refresh_timer_start(void) { refresh_us-=usecond(); } |  | ||||||
|   void refresh_timer_stop(void)  { refresh_us+=usecond(); } |  | ||||||
|   void S_timer_start(void)       { S_us-=usecond(); } |  | ||||||
|   void S_timer_stop(void)        { S_us+=usecond(); } |  | ||||||
|   ///////////////////////////// |  | ||||||
|   // Heatbath? |   // Heatbath? | ||||||
|   ///////////////////////////// |   virtual void refresh(const GaugeField& U, GridParallelRNG& pRNG) = 0; // refresh pseudofermions | ||||||
|   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions |  | ||||||
|   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action |   virtual RealD S(const GaugeField& U) = 0;                             // evaluate the action | ||||||
|   virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ;  // if the refresh computes the action, can cache it. Alternately refreshAndAction() ? |  | ||||||
|   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative |   virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0;        // evaluate the action derivative | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////////////////////// |  | ||||||
|   // virtual smeared interface through configuration container |  | ||||||
|   ///////////////////////////////////////////////////////////// |  | ||||||
|   virtual void refresh(ConfigurationBase<GaugeField> & U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) |  | ||||||
|   { |  | ||||||
|     refresh(U.get_U(is_smeared),sRNG,pRNG); |  | ||||||
|   } |  | ||||||
|   virtual RealD S(ConfigurationBase<GaugeField>& U) |  | ||||||
|   { |  | ||||||
|     return S(U.get_U(is_smeared)); |  | ||||||
|   } |  | ||||||
|   virtual RealD Sinitial(ConfigurationBase<GaugeField>& U)  |  | ||||||
|   { |  | ||||||
|     return Sinitial(U.get_U(is_smeared)); |  | ||||||
|   } |  | ||||||
|   virtual void deriv(ConfigurationBase<GaugeField>& U, GaugeField& dSdU) |  | ||||||
|   { |  | ||||||
|     deriv(U.get_U(is_smeared),dSdU);  |  | ||||||
|     if ( is_smeared ) { |  | ||||||
|       U.smeared_force(dSdU); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|   /////////////////////////////// |  | ||||||
|   // Logging |  | ||||||
|   /////////////////////////////// |  | ||||||
|   virtual std::string action_name()    = 0;                             // return the action name |   virtual std::string action_name()    = 0;                             // return the action name | ||||||
|   virtual std::string LogParameters()  = 0;                             // prints action parameters |   virtual std::string LogParameters()  = 0;                             // prints action parameters | ||||||
|   virtual ~Action(){} |   virtual ~Action(){} | ||||||
| }; | }; | ||||||
|  |  | ||||||
| template <class GaugeField > |  | ||||||
| class EmptyAction : public Action <GaugeField> |  | ||||||
| { |  | ||||||
|   virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) { assert(0);}; // refresh pseudofermions |  | ||||||
|   virtual RealD S(const GaugeField& U) { return 0.0;};                             // evaluate the action |  | ||||||
|   virtual void deriv(const GaugeField& U, GaugeField& dSdU) { assert(0); };        // evaluate the action derivative |  | ||||||
|  |  | ||||||
|   /////////////////////////////// |  | ||||||
|   // Logging |  | ||||||
|   /////////////////////////////// |  | ||||||
|   virtual std::string action_name()    { return std::string("Level Force Log"); }; |  | ||||||
|   virtual std::string LogParameters()  { return std::string("No parameters");}; |  | ||||||
| }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|  |  | ||||||
| #endif // ACTION_BASE_H | #endif // ACTION_BASE_H | ||||||
|   | |||||||
| @@ -30,8 +30,6 @@ directory | |||||||
| #ifndef QCD_ACTION_CORE | #ifndef QCD_ACTION_CORE | ||||||
| #define QCD_ACTION_CORE | #define QCD_ACTION_CORE | ||||||
|  |  | ||||||
| #include <Grid/qcd/action/gauge/GaugeImplementations.h> |  | ||||||
|  |  | ||||||
| #include <Grid/qcd/action/ActionBase.h> | #include <Grid/qcd/action/ActionBase.h> | ||||||
| NAMESPACE_CHECK(ActionBase); | NAMESPACE_CHECK(ActionBase); | ||||||
| #include <Grid/qcd/action/ActionSet.h> | #include <Grid/qcd/action/ActionSet.h> | ||||||
| @@ -39,10 +37,6 @@ NAMESPACE_CHECK(ActionSet); | |||||||
| #include <Grid/qcd/action/ActionParams.h> | #include <Grid/qcd/action/ActionParams.h> | ||||||
| NAMESPACE_CHECK(ActionParams); | NAMESPACE_CHECK(ActionParams); | ||||||
|  |  | ||||||
| #include <Grid/qcd/action/filters/MomentumFilter.h> |  | ||||||
| #include <Grid/qcd/action/filters/DirichletFilter.h> |  | ||||||
| #include <Grid/qcd/action/filters/DDHMCFilter.h> |  | ||||||
|  |  | ||||||
| //////////////////////////////////////////// | //////////////////////////////////////////// | ||||||
| // Gauge Actions | // Gauge Actions | ||||||
| //////////////////////////////////////////// | //////////////////////////////////////////// | ||||||
|   | |||||||
| @@ -34,45 +34,27 @@ directory | |||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
|  | // These can move into a params header and be given MacroMagic serialisation | ||||||
| struct GparityWilsonImplParams { | struct GparityWilsonImplParams { | ||||||
|   Coordinate twists; |   Coordinate twists; | ||||||
|                      //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs |   GparityWilsonImplParams() : twists(Nd, 0) {}; | ||||||
|   Coordinate dirichlet; // Blocksize of dirichlet BCs |  | ||||||
|   int  partialDirichlet; |  | ||||||
|   GparityWilsonImplParams() : twists(Nd, 0) { |  | ||||||
|     dirichlet.resize(0); |  | ||||||
|     partialDirichlet=0; |  | ||||||
|   }; |  | ||||||
| }; | }; | ||||||
|    |    | ||||||
| struct WilsonImplParams { | struct WilsonImplParams { | ||||||
|   bool overlapCommsCompute; |   bool overlapCommsCompute; | ||||||
|   Coordinate dirichlet; // Blocksize of dirichlet BCs |  | ||||||
|   int  partialDirichlet; |  | ||||||
|   AcceleratorVector<Real,Nd> twist_n_2pi_L; |   AcceleratorVector<Real,Nd> twist_n_2pi_L; | ||||||
|   AcceleratorVector<Complex,Nd> boundary_phases; |   AcceleratorVector<Complex,Nd> boundary_phases; | ||||||
|   WilsonImplParams()  { |   WilsonImplParams()  { | ||||||
|     dirichlet.resize(0); |  | ||||||
|     partialDirichlet=0; |  | ||||||
|     boundary_phases.resize(Nd, 1.0); |     boundary_phases.resize(Nd, 1.0); | ||||||
|       twist_n_2pi_L.resize(Nd, 0.0); |       twist_n_2pi_L.resize(Nd, 0.0); | ||||||
|   }; |   }; | ||||||
|   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) { |   WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) { | ||||||
|     twist_n_2pi_L.resize(Nd, 0.0); |     twist_n_2pi_L.resize(Nd, 0.0); | ||||||
|     partialDirichlet=0; |  | ||||||
|     dirichlet.resize(0); |  | ||||||
|   } |   } | ||||||
| }; | }; | ||||||
|  |  | ||||||
| struct StaggeredImplParams { | struct StaggeredImplParams { | ||||||
|   Coordinate dirichlet; // Blocksize of dirichlet BCs |   StaggeredImplParams()  {}; | ||||||
|   int  partialDirichlet; |  | ||||||
|   StaggeredImplParams() |  | ||||||
|   { |  | ||||||
|     partialDirichlet=0; |  | ||||||
|     dirichlet.resize(0); |  | ||||||
|   }; |  | ||||||
| }; | }; | ||||||
|    |    | ||||||
|   struct OneFlavourRationalParams : Serializable { |   struct OneFlavourRationalParams : Serializable { | ||||||
| @@ -81,11 +63,9 @@ struct StaggeredImplParams { | |||||||
| 				    RealD, hi,  | 				    RealD, hi,  | ||||||
| 				    int,   MaxIter,  | 				    int,   MaxIter,  | ||||||
| 				    RealD, tolerance,  | 				    RealD, tolerance,  | ||||||
| 				    RealD, mdtolerance,  |  | ||||||
| 				    int,   degree,  | 				    int,   degree,  | ||||||
| 				    int,   precision, | 				    int,   precision, | ||||||
| 				    int,   BoundsCheckFreq, | 				    int,   BoundsCheckFreq); | ||||||
| 				    RealD, BoundsCheckTol); |  | ||||||
|      |      | ||||||
|   // MaxIter and tolerance, vectors?? |   // MaxIter and tolerance, vectors?? | ||||||
|      |      | ||||||
| @@ -96,62 +76,16 @@ struct StaggeredImplParams { | |||||||
| 				RealD tol      = 1.0e-8,  | 				RealD tol      = 1.0e-8,  | ||||||
|                            	int _degree    = 10, |                            	int _degree    = 10, | ||||||
| 				int _precision = 64, | 				int _precision = 64, | ||||||
| 				int _BoundsCheckFreq=20, | 				int _BoundsCheckFreq=20) | ||||||
| 				RealD mdtol    = 1.0e-6, |  | ||||||
| 				double _BoundsCheckTol=1e-6) |  | ||||||
|       : lo(_lo), |       : lo(_lo), | ||||||
| 	hi(_hi), | 	hi(_hi), | ||||||
| 	MaxIter(_maxit), | 	MaxIter(_maxit), | ||||||
| 	tolerance(tol), | 	tolerance(tol), | ||||||
|         mdtolerance(mdtol), |  | ||||||
| 	degree(_degree), | 	degree(_degree), | ||||||
|         precision(_precision), |         precision(_precision), | ||||||
|         BoundsCheckFreq(_BoundsCheckFreq), |         BoundsCheckFreq(_BoundsCheckFreq){}; | ||||||
|         BoundsCheckTol(_BoundsCheckTol){}; |  | ||||||
|   }; |   }; | ||||||
|    |    | ||||||
|   /*Action parameters for the generalized rational action |  | ||||||
|     The approximation is for (M^dag M)^{1/inv_pow} |  | ||||||
|     where inv_pow is the denominator of the fractional power. |  | ||||||
|     Default inv_pow=2 for square root, making this equivalent to  |  | ||||||
|     the OneFlavourRational action |  | ||||||
|   */ |  | ||||||
|     struct RationalActionParams : Serializable { |  | ||||||
|     GRID_SERIALIZABLE_CLASS_MEMBERS(RationalActionParams,  |  | ||||||
| 				    int, inv_pow,  |  | ||||||
| 				    RealD, lo, //low eigenvalue bound of rational approx |  | ||||||
| 				    RealD, hi, //high eigenvalue bound of rational approx |  | ||||||
| 				    int,   MaxIter,  //maximum iterations in msCG |  | ||||||
| 				    RealD, action_tolerance,  //msCG tolerance in action evaluation |  | ||||||
| 				    int,   action_degree, //rational approx tolerance in action evaluation |  | ||||||
| 				    RealD, md_tolerance,  //msCG tolerance in MD integration |  | ||||||
| 				    int,   md_degree, //rational approx tolerance in MD integration |  | ||||||
| 				    int,   precision, //precision of floating point arithmetic |  | ||||||
| 				    int,   BoundsCheckFreq); //frequency the approximation is tested (with Metropolis degree/tolerance); 0 disables the check |  | ||||||
|   // constructor  |  | ||||||
|   RationalActionParams(int _inv_pow = 2, |  | ||||||
| 		       RealD _lo      = 0.0,  |  | ||||||
| 		       RealD _hi      = 1.0,  |  | ||||||
| 		       int _maxit     = 1000, |  | ||||||
| 		       RealD _action_tolerance      = 1.0e-8,  |  | ||||||
| 		       int _action_degree    = 10, |  | ||||||
| 		       RealD _md_tolerance      = 1.0e-8,  |  | ||||||
| 		       int _md_degree    = 10, |  | ||||||
| 		       int _precision = 64, |  | ||||||
| 		       int _BoundsCheckFreq=20) |  | ||||||
|     : inv_pow(_inv_pow),  |  | ||||||
|       lo(_lo), |  | ||||||
|       hi(_hi), |  | ||||||
|       MaxIter(_maxit), |  | ||||||
|       action_tolerance(_action_tolerance), |  | ||||||
|       action_degree(_action_degree), |  | ||||||
|       md_tolerance(_md_tolerance), |  | ||||||
|       md_degree(_md_degree), |  | ||||||
|       precision(_precision), |  | ||||||
|       BoundsCheckFreq(_BoundsCheckFreq){}; |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|  |  | ||||||
| #endif | #endif | ||||||
|   | |||||||
| @@ -68,17 +68,9 @@ public: | |||||||
|   /////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////// | ||||||
|   // Support for MADWF tricks |   // Support for MADWF tricks | ||||||
|   /////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////// | ||||||
|   RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; }; |   RealD Mass(void) { return mass; }; | ||||||
|   RealD MassPlus(void) { return mass_plus; }; |  | ||||||
|   RealD MassMinus(void) { return mass_minus; }; |  | ||||||
|  |  | ||||||
|   void  SetMass(RealD _mass) {  |   void  SetMass(RealD _mass) {  | ||||||
|     mass_plus=mass_minus=_mass;  |     mass=_mass;  | ||||||
|     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs |  | ||||||
|   } ; |  | ||||||
|   void  SetMass(RealD _mass_plus, RealD _mass_minus) {  |  | ||||||
|     mass_plus=_mass_plus; |  | ||||||
|     mass_minus=_mass_minus; |  | ||||||
|     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs |     SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c);  // Reset coeffs | ||||||
|   } ; |   } ; | ||||||
|   void  P(const FermionField &psi, FermionField &chi); |   void  P(const FermionField &psi, FermionField &chi); | ||||||
| @@ -116,7 +108,7 @@ public: | |||||||
|   void   MeooeDag5D    (const FermionField &in, FermionField &out); |   void   MeooeDag5D    (const FermionField &in, FermionField &out); | ||||||
|  |  | ||||||
|   //    protected: |   //    protected: | ||||||
|   RealD mass_plus, mass_minus; |   RealD mass; | ||||||
|  |  | ||||||
|   // Save arguments to SetCoefficientsInternal |   // Save arguments to SetCoefficientsInternal | ||||||
|   Vector<Coeff_t> _gamma; |   Vector<Coeff_t> _gamma; | ||||||
| @@ -183,6 +175,16 @@ public: | |||||||
| 		  GridRedBlackCartesian &FourDimRedBlackGrid, | 		  GridRedBlackCartesian &FourDimRedBlackGrid, | ||||||
| 		  RealD _mass,RealD _M5,const ImplParams &p= ImplParams()); | 		  RealD _mass,RealD _M5,const ImplParams &p= ImplParams()); | ||||||
|  |  | ||||||
|  |   void CayleyReport(void); | ||||||
|  |   void CayleyZeroCounters(void); | ||||||
|  |  | ||||||
|  |   double M5Dflops; | ||||||
|  |   double M5Dcalls; | ||||||
|  |   double M5Dtime; | ||||||
|  |  | ||||||
|  |   double MooeeInvFlops; | ||||||
|  |   double MooeeInvCalls; | ||||||
|  |   double MooeeInvTime; | ||||||
|  |  | ||||||
| protected: | protected: | ||||||
|   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c); |   virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c); | ||||||
|   | |||||||
| @@ -1,334 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid |  | ||||||
|  |  | ||||||
|     Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2017 - 2022 |  | ||||||
|  |  | ||||||
|     Author: paboyle <paboyle@ph.ed.ac.uk> |  | ||||||
|     Author: Daniel Richtmann <daniel.richtmann@gmail.com> |  | ||||||
|     Author: Mattia Bruno <mattia.bruno@cern.ch> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
|     *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
|  |  | ||||||
| #pragma once |  | ||||||
|  |  | ||||||
| #include <Grid/Grid.h> |  | ||||||
| #include <Grid/qcd/spin/Dirac.h> |  | ||||||
| #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h> |  | ||||||
|  |  | ||||||
| //////////////////////////////////////////// |  | ||||||
| // Standard Clover |  | ||||||
| //   (4+m0) + csw * clover_term |  | ||||||
| // Exp Clover |  | ||||||
| //   (4+m0) * exp(csw/(4+m0) clover_term) |  | ||||||
| //   = (4+m0) + csw * clover_term + ... |  | ||||||
| //////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
|  |  | ||||||
| ////////////////////////////////// |  | ||||||
| // Generic Standard Clover |  | ||||||
| ////////////////////////////////// |  | ||||||
|  |  | ||||||
| template<class Impl> |  | ||||||
| class CloverHelpers: public WilsonCloverHelpers<Impl> { |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   INHERIT_IMPL_TYPES(Impl); |  | ||||||
|   INHERIT_CLOVER_TYPES(Impl); |  | ||||||
|  |  | ||||||
|   typedef WilsonCloverHelpers<Impl> Helpers; |  | ||||||
|  |  | ||||||
|   static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) { |  | ||||||
|     GridBase *grid = CloverTerm.Grid(); |  | ||||||
|     CloverTerm += diag_mass; |  | ||||||
|  |  | ||||||
|     int lvol = grid->lSites(); |  | ||||||
|     int DimRep = Impl::Dimension; |  | ||||||
|     { |  | ||||||
|       autoView(CTv,CloverTerm,CpuRead); |  | ||||||
|       autoView(CTIv,CloverTermInv,CpuWrite); |  | ||||||
|       thread_for(site, lvol, { |  | ||||||
|         Coordinate lcoor; |  | ||||||
|         grid->LocalIndexToLocalCoor(site, lcoor); |  | ||||||
|         Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep); |  | ||||||
|         Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep); |  | ||||||
|         typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero(); |  | ||||||
|         peekLocalSite(Qx, CTv, lcoor); |  | ||||||
|  |  | ||||||
|         for (int j = 0; j < Ns; j++) |  | ||||||
|           for (int k = 0; k < Ns; k++) |  | ||||||
|             for (int a = 0; a < DimRep; a++) |  | ||||||
|               for (int b = 0; b < DimRep; b++){ |  | ||||||
|                 auto zz =  Qx()(j, k)(a, b); |  | ||||||
|                 EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz); |  | ||||||
|               } |  | ||||||
|  |  | ||||||
|         EigenInvCloverOp = EigenCloverOp.inverse(); |  | ||||||
|         for (int j = 0; j < Ns; j++) |  | ||||||
|           for (int k = 0; k < Ns; k++) |  | ||||||
|             for (int a = 0; a < DimRep; a++) |  | ||||||
|               for (int b = 0; b < DimRep; b++) |  | ||||||
|                 Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep); |  | ||||||
|                pokeLocalSite(Qxinv, CTIv, lcoor); |  | ||||||
|       }); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) { |  | ||||||
|     return Helpers::Cmunu(U, lambda, mu, nu); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
| }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
| ////////////////////////////////// |  | ||||||
| // Generic Exp Clover |  | ||||||
| ////////////////////////////////// |  | ||||||
|  |  | ||||||
| template<class Impl> |  | ||||||
| class ExpCloverHelpers: public WilsonCloverHelpers<Impl> { |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   INHERIT_IMPL_TYPES(Impl); |  | ||||||
|   INHERIT_CLOVER_TYPES(Impl); |  | ||||||
|  |  | ||||||
|   template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>; |  | ||||||
|   typedef WilsonCloverHelpers<Impl> Helpers; |  | ||||||
|  |  | ||||||
|   // Can this be avoided? |  | ||||||
|   static void IdentityTimesC(const CloverField& in, RealD c) { |  | ||||||
|     int DimRep = Impl::Dimension; |  | ||||||
|  |  | ||||||
|     autoView(in_v, in, AcceleratorWrite); |  | ||||||
|  |  | ||||||
|     accelerator_for(ss, in.Grid()->oSites(), 1, { |  | ||||||
|       for (int sa=0; sa<Ns; sa++) |  | ||||||
|         for (int ca=0; ca<DimRep; ca++) |  | ||||||
|           in_v[ss]()(sa,sa)(ca,ca) = c; |  | ||||||
|     }); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static int getNMAX(RealD prec, RealD R) { |  | ||||||
|     /* compute stop condition for exponential */ |  | ||||||
|     int NMAX=1; |  | ||||||
|     RealD cond=R*R/2.; |  | ||||||
|  |  | ||||||
|     while (cond*std::exp(R)>prec) { |  | ||||||
|       NMAX++; |  | ||||||
|       cond*=R/(double)(NMAX+1); |  | ||||||
|     } |  | ||||||
|     return NMAX; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static int getNMAX(Lattice<iImplClover<vComplexD2>> &t, RealD R) {return getNMAX(1e-12,R);} |  | ||||||
|   static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);} |  | ||||||
|   static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);} |  | ||||||
|  |  | ||||||
|   static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) { |  | ||||||
|     GridBase* grid = Clover.Grid(); |  | ||||||
|     CloverField ExpClover(grid); |  | ||||||
|  |  | ||||||
|     int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass); |  | ||||||
|  |  | ||||||
|     Clover *= (1.0/diag_mass); |  | ||||||
|  |  | ||||||
|     // Taylor expansion, slow but generic |  | ||||||
|     // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...)) |  | ||||||
|     // qN = cN |  | ||||||
|     // qn = cn + qn+1 X |  | ||||||
|     std::vector<RealD> cn(NMAX+1); |  | ||||||
|     cn[0] = 1.0; |  | ||||||
|     for (int i=1; i<=NMAX; i++) |  | ||||||
|       cn[i] = cn[i-1] / RealD(i); |  | ||||||
|  |  | ||||||
|     ExpClover = Zero(); |  | ||||||
|     IdentityTimesC(ExpClover, cn[NMAX]); |  | ||||||
|     for (int i=NMAX-1; i>=0; i--) |  | ||||||
|       ExpClover = ExpClover * Clover + cn[i]; |  | ||||||
|  |  | ||||||
|     // prepare inverse |  | ||||||
|     CloverInv = (-1.0)*Clover; |  | ||||||
|  |  | ||||||
|     Clover = ExpClover * diag_mass; |  | ||||||
|  |  | ||||||
|     ExpClover = Zero(); |  | ||||||
|     IdentityTimesC(ExpClover, cn[NMAX]); |  | ||||||
|     for (int i=NMAX-1; i>=0; i--) |  | ||||||
|       ExpClover = ExpClover * CloverInv + cn[i]; |  | ||||||
|  |  | ||||||
|     CloverInv = ExpClover * (1.0/diag_mass); |  | ||||||
|  |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) { |  | ||||||
|     assert(0); |  | ||||||
|     return lambda; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
| }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
| ////////////////////////////////// |  | ||||||
| // Compact Standard Clover |  | ||||||
| ////////////////////////////////// |  | ||||||
|  |  | ||||||
|  |  | ||||||
| template<class Impl> |  | ||||||
| class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>, |  | ||||||
|                             public WilsonCloverHelpers<Impl> { |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   INHERIT_IMPL_TYPES(Impl); |  | ||||||
|   INHERIT_CLOVER_TYPES(Impl); |  | ||||||
|   INHERIT_COMPACT_CLOVER_TYPES(Impl); |  | ||||||
|  |  | ||||||
|   typedef WilsonCloverHelpers<Impl> Helpers; |  | ||||||
|   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers; |  | ||||||
|  |  | ||||||
|   static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) { |  | ||||||
|     Clover += diag_mass; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static void InvertClover(CloverField& InvClover, |  | ||||||
|                             const CloverDiagonalField& diagonal, |  | ||||||
|                             const CloverTriangleField& triangle, |  | ||||||
|                             CloverDiagonalField&       diagonalInv, |  | ||||||
|                             CloverTriangleField&       triangleInv, |  | ||||||
|                             bool fixedBoundaries) { |  | ||||||
|  |  | ||||||
|     CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   // TODO: implement Cmunu for better performances with compact layout, but don't do it |  | ||||||
|   // here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers |  | ||||||
|   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) { |  | ||||||
|     return Helpers::Cmunu(U, lambda, mu, nu); |  | ||||||
|   } |  | ||||||
| }; |  | ||||||
|  |  | ||||||
| ////////////////////////////////// |  | ||||||
| // Compact Exp Clover |  | ||||||
| ////////////////////////////////// |  | ||||||
|  |  | ||||||
| template<class Impl> |  | ||||||
| class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> { |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   INHERIT_IMPL_TYPES(Impl); |  | ||||||
|   INHERIT_CLOVER_TYPES(Impl); |  | ||||||
|   INHERIT_COMPACT_CLOVER_TYPES(Impl); |  | ||||||
|  |  | ||||||
|   template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>; |  | ||||||
|   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers; |  | ||||||
|  |  | ||||||
|   // Can this be avoided? |  | ||||||
|   static void IdentityTimesC(const CloverField& in, RealD c) { |  | ||||||
|     int DimRep = Impl::Dimension; |  | ||||||
|  |  | ||||||
|     autoView(in_v, in, AcceleratorWrite); |  | ||||||
|  |  | ||||||
|     accelerator_for(ss, in.Grid()->oSites(), 1, { |  | ||||||
|       for (int sa=0; sa<Ns; sa++) |  | ||||||
|         for (int ca=0; ca<DimRep; ca++) |  | ||||||
|           in_v[ss]()(sa,sa)(ca,ca) = c; |  | ||||||
|     }); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static int getNMAX(RealD prec, RealD R) { |  | ||||||
|     /* compute stop condition for exponential */ |  | ||||||
|     int NMAX=1; |  | ||||||
|     RealD cond=R*R/2.; |  | ||||||
|  |  | ||||||
|     while (cond*std::exp(R)>prec) { |  | ||||||
|       NMAX++; |  | ||||||
|       cond*=R/(double)(NMAX+1); |  | ||||||
|     } |  | ||||||
|     return NMAX; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);} |  | ||||||
|   static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);} |  | ||||||
|  |  | ||||||
|   static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) { |  | ||||||
|  |  | ||||||
|     GridBase* grid = Clover.Grid(); |  | ||||||
|     CloverField ExpClover(grid); |  | ||||||
|  |  | ||||||
|     int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass); |  | ||||||
|  |  | ||||||
|     Clover *= (1.0/diag_mass); |  | ||||||
|  |  | ||||||
|     // Taylor expansion, slow but generic |  | ||||||
|     // Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...)) |  | ||||||
|     // qN = cN |  | ||||||
|     // qn = cn + qn+1 X |  | ||||||
|     std::vector<RealD> cn(NMAX+1); |  | ||||||
|     cn[0] = 1.0; |  | ||||||
|     for (int i=1; i<=NMAX; i++) |  | ||||||
|       cn[i] = cn[i-1] / RealD(i); |  | ||||||
|  |  | ||||||
|     ExpClover = Zero(); |  | ||||||
|     IdentityTimesC(ExpClover, cn[NMAX]); |  | ||||||
|     for (int i=NMAX-1; i>=0; i--) |  | ||||||
|       ExpClover = ExpClover * Clover + cn[i]; |  | ||||||
|  |  | ||||||
|     // prepare inverse |  | ||||||
|     CloverInv = (-1.0)*Clover; |  | ||||||
|  |  | ||||||
|     Clover = ExpClover * diag_mass; |  | ||||||
|  |  | ||||||
|     ExpClover = Zero(); |  | ||||||
|     IdentityTimesC(ExpClover, cn[NMAX]); |  | ||||||
|     for (int i=NMAX-1; i>=0; i--) |  | ||||||
|       ExpClover = ExpClover * CloverInv + cn[i]; |  | ||||||
|  |  | ||||||
|     CloverInv = ExpClover * (1.0/diag_mass); |  | ||||||
|  |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static void InvertClover(CloverField& InvClover, |  | ||||||
|                             const CloverDiagonalField& diagonal, |  | ||||||
|                             const CloverTriangleField& triangle, |  | ||||||
|                             CloverDiagonalField&       diagonalInv, |  | ||||||
|                             CloverTriangleField&       triangleInv, |  | ||||||
|                             bool fixedBoundaries) { |  | ||||||
|  |  | ||||||
|     if (fixedBoundaries) |  | ||||||
|     { |  | ||||||
|       CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv); |  | ||||||
|     } |  | ||||||
|     else |  | ||||||
|     { |  | ||||||
|       CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) { |  | ||||||
|     assert(0); |  | ||||||
|     return lambda; |  | ||||||
|   } |  | ||||||
|  |  | ||||||
| }; |  | ||||||
|  |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
| @@ -1,241 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
|     Grid physics library, www.github.com/paboyle/Grid |  | ||||||
|  |  | ||||||
|     Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion.h |  | ||||||
|  |  | ||||||
|     Copyright (C) 2020 - 2022 |  | ||||||
|  |  | ||||||
|     Author: Daniel Richtmann <daniel.richtmann@gmail.com> |  | ||||||
|     Author: Nils Meyer <nils.meyer@ur.de> |  | ||||||
|  |  | ||||||
|     This program is free software; you can redistribute it and/or modify |  | ||||||
|     it under the terms of the GNU General Public License as published by |  | ||||||
|     the Free Software Foundation; either version 2 of the License, or |  | ||||||
|     (at your option) any later version. |  | ||||||
|  |  | ||||||
|     This program is distributed in the hope that it will be useful, |  | ||||||
|     but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
|     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
|     GNU General Public License for more details. |  | ||||||
|  |  | ||||||
|     You should have received a copy of the GNU General Public License along |  | ||||||
|     with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
|     51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
|     See the full license in the file "LICENSE" in the top level distribution directory |  | ||||||
|     *************************************************************************************/ |  | ||||||
| /*  END LEGAL */ |  | ||||||
|  |  | ||||||
| #pragma once |  | ||||||
|  |  | ||||||
| #include <Grid/qcd/action/fermion/WilsonCloverTypes.h> |  | ||||||
| #include <Grid/qcd/action/fermion/WilsonCloverHelpers.h> |  | ||||||
| #include <Grid/qcd/action/fermion/CloverHelpers.h> |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| // see Grid/qcd/action/fermion/WilsonCloverFermion.h for description |  | ||||||
| // |  | ||||||
| // Modifications done here: |  | ||||||
| // |  | ||||||
| // Original: clover term = 12x12 matrix per site |  | ||||||
| // |  | ||||||
| // But: Only two diagonal 6x6 hermitian blocks are non-zero (also true for original, verified by running) |  | ||||||
| // Sufficient to store/transfer only the real parts of the diagonal and one triangular part |  | ||||||
| // 2 * (6 + 15 * 2) = 72 real or 36 complex words to be stored/transfered |  | ||||||
| // |  | ||||||
| // Here: Above but diagonal as complex numbers, i.e., need to store/transfer |  | ||||||
| // 2 * (6 * 2 + 15 * 2) = 84 real or 42 complex words |  | ||||||
| // |  | ||||||
| // Words per site and improvement compared to original (combined with the input and output spinors): |  | ||||||
| // |  | ||||||
| // - Original: 2*12 + 12*12 = 168 words -> 1.00 x less |  | ||||||
| // - Minimal:  2*12 + 36    =  60 words -> 2.80 x less |  | ||||||
| // - Here:     2*12 + 42    =  66 words -> 2.55 x less |  | ||||||
| // |  | ||||||
| // These improvements directly translate to wall-clock time |  | ||||||
| // |  | ||||||
| // Data layout: |  | ||||||
| // |  | ||||||
| // - diagonal and triangle part as separate lattice fields, |  | ||||||
| //   this was faster than as 1 combined field on all tested machines |  | ||||||
| // - diagonal: as expected |  | ||||||
| // - triangle: store upper right triangle in row major order |  | ||||||
| // - graphical: |  | ||||||
| //        0  1  2  3  4 |  | ||||||
| //           5  6  7  8 |  | ||||||
| //              9 10 11 = upper right triangle indices |  | ||||||
| //                12 13 |  | ||||||
| //                   14 |  | ||||||
| //     0 |  | ||||||
| //        1 |  | ||||||
| //           2 |  | ||||||
| //              3       = diagonal indices |  | ||||||
| //                 4 |  | ||||||
| //                    5 |  | ||||||
| //     0 |  | ||||||
| //     1  5 |  | ||||||
| //     2  6  9          = lower left triangle indices |  | ||||||
| //     3  7 10 12 |  | ||||||
| //     4  8 11 13 14 |  | ||||||
| // |  | ||||||
| // Impact on total memory consumption: |  | ||||||
| // - Original: (2 * 1 + 8 * 1/2) 12x12 matrices = 6 12x12 matrices = 864 complex words per site |  | ||||||
| // - Here:     (2 * 1 + 4 * 1/2) diagonal parts = 4 diagonal parts =  24 complex words per site |  | ||||||
| //           + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts =  60 complex words per site |  | ||||||
| //                                                                 =  84 complex words per site |  | ||||||
|  |  | ||||||
| template<class Impl, class CloverHelpers> |  | ||||||
| class CompactWilsonCloverFermion : public WilsonFermion<Impl>, |  | ||||||
|                                    public WilsonCloverHelpers<Impl>, |  | ||||||
|                                    public CompactWilsonCloverHelpers<Impl> { |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|   // Sizes |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   INHERIT_COMPACT_CLOVER_SIZES(Impl); |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|   // Type definitions |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   INHERIT_IMPL_TYPES(Impl); |  | ||||||
|   INHERIT_CLOVER_TYPES(Impl); |  | ||||||
|   INHERIT_COMPACT_CLOVER_TYPES(Impl); |  | ||||||
|  |  | ||||||
|   typedef WilsonFermion<Impl>              WilsonBase; |  | ||||||
|   typedef WilsonCloverHelpers<Impl>        Helpers; |  | ||||||
|   typedef CompactWilsonCloverHelpers<Impl> CompactHelpers; |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|   // Constructors |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   CompactWilsonCloverFermion(GaugeField& _Umu, |  | ||||||
| 			    GridCartesian& Fgrid, |  | ||||||
| 			    GridRedBlackCartesian& Hgrid, |  | ||||||
| 			    const RealD _mass, |  | ||||||
| 			    const RealD _csw_r = 0.0, |  | ||||||
| 			    const RealD _csw_t = 0.0, |  | ||||||
| 			    const RealD _cF = 1.0, |  | ||||||
| 			    const WilsonAnisotropyCoefficients& clover_anisotropy = WilsonAnisotropyCoefficients(), |  | ||||||
| 			    const ImplParams& impl_p = ImplParams()); |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|   // Member functions (implementing interface) |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   virtual void Instantiatable() {}; |  | ||||||
|   int          ConstEE()     override { return 0; }; |  | ||||||
|   int          isTrivialEE() override { return 0; }; |  | ||||||
|  |  | ||||||
|   void Dhop(const FermionField& in, FermionField& out, int dag) override; |  | ||||||
|  |  | ||||||
|   void DhopOE(const FermionField& in, FermionField& out, int dag) override; |  | ||||||
|  |  | ||||||
|   void DhopEO(const FermionField& in, FermionField& out, int dag) override; |  | ||||||
|  |  | ||||||
|   void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override; |  | ||||||
|  |  | ||||||
|   void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */; |  | ||||||
|  |  | ||||||
|   void M(const FermionField& in, FermionField& out) override; |  | ||||||
|  |  | ||||||
|   void Mdag(const FermionField& in, FermionField& out) override; |  | ||||||
|  |  | ||||||
|   void Meooe(const FermionField& in, FermionField& out) override; |  | ||||||
|  |  | ||||||
|   void MeooeDag(const FermionField& in, FermionField& out) override; |  | ||||||
|  |  | ||||||
|   void Mooee(const FermionField& in, FermionField& out) override; |  | ||||||
|  |  | ||||||
|   void MooeeDag(const FermionField& in, FermionField& out) override; |  | ||||||
|  |  | ||||||
|   void MooeeInv(const FermionField& in, FermionField& out) override; |  | ||||||
|  |  | ||||||
|   void MooeeInvDag(const FermionField& in, FermionField& out) override; |  | ||||||
|  |  | ||||||
|   void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override; |  | ||||||
|  |  | ||||||
|   void MdirAll(const FermionField& in, std::vector<FermionField>& out) override; |  | ||||||
|  |  | ||||||
|   void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override; |  | ||||||
|  |  | ||||||
|   void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override; |  | ||||||
|  |  | ||||||
|   void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override; |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|   // Member functions (internals) |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|  |  | ||||||
|   void MooeeInternal(const FermionField&        in, |  | ||||||
|                      FermionField&              out, |  | ||||||
|                      const CloverDiagonalField& diagonal, |  | ||||||
|                      const CloverTriangleField& triangle); |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|   // Helpers |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|  |  | ||||||
|   void ImportGauge(const GaugeField& _Umu) override; |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|   // Helpers |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| private: |  | ||||||
|  |  | ||||||
|   template<class Field> |  | ||||||
|   const MaskField* getCorrectMaskField(const Field &in) const { |  | ||||||
|     if(in.Grid()->_isCheckerBoarded) { |  | ||||||
|       if(in.Checkerboard() == Odd) { |  | ||||||
|         return &this->BoundaryMaskOdd; |  | ||||||
|       } else { |  | ||||||
|         return &this->BoundaryMaskEven; |  | ||||||
|       } |  | ||||||
|     } else { |  | ||||||
|       return &this->BoundaryMask; |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   template<class Field> |  | ||||||
|   void ApplyBoundaryMask(Field& f) { |  | ||||||
|     const MaskField* m = getCorrectMaskField(f); assert(m != nullptr); |  | ||||||
|     assert(m != nullptr); |  | ||||||
|     CompactHelpers::ApplyBoundaryMask(f, *m); |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|   // Member Data |  | ||||||
|   ///////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| public: |  | ||||||
|  |  | ||||||
|   RealD csw_r; |  | ||||||
|   RealD csw_t; |  | ||||||
|   RealD cF; |  | ||||||
|  |  | ||||||
|   bool fixedBoundaries; |  | ||||||
|  |  | ||||||
|   CloverDiagonalField Diagonal,    DiagonalEven,    DiagonalOdd; |  | ||||||
|   CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd; |  | ||||||
|  |  | ||||||
|   CloverTriangleField Triangle,    TriangleEven,    TriangleOdd; |  | ||||||
|   CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd; |  | ||||||
|  |  | ||||||
|   FermionField Tmp; |  | ||||||
|  |  | ||||||
|   MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd; |  | ||||||
| }; |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
| @@ -1,291 +0,0 @@ | |||||||
| /************************************************************************************* |  | ||||||
|  |  | ||||||
| Grid physics library, www.github.com/paboyle/Grid |  | ||||||
|  |  | ||||||
| Source file: ./lib/qcd/action/fermion/DWFSlow.h |  | ||||||
|  |  | ||||||
| Copyright (C) 2022 |  | ||||||
|  |  | ||||||
| Author: Peter Boyle <pboyle@bnl.gov> |  | ||||||
|  |  | ||||||
| This program is free software; you can redistribute it and/or modify |  | ||||||
| it under the terms of the GNU General Public License as published by |  | ||||||
| the Free Software Foundation; either version 2 of the License, or |  | ||||||
| (at your option) any later version. |  | ||||||
|  |  | ||||||
| This program is distributed in the hope that it will be useful, |  | ||||||
| but WITHOUT ANY WARRANTY; without even the implied warranty of |  | ||||||
| MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the |  | ||||||
| GNU General Public License for more details. |  | ||||||
|  |  | ||||||
| You should have received a copy of the GNU General Public License along |  | ||||||
| with this program; if not, write to the Free Software Foundation, Inc., |  | ||||||
| 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |  | ||||||
|  |  | ||||||
| See the full license in the file "LICENSE" in the top level distribution |  | ||||||
| directory |  | ||||||
| *************************************************************************************/ |  | ||||||
| 			   /*  END LEGAL */ |  | ||||||
| #pragma once |  | ||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); |  | ||||||
|  |  | ||||||
| template <class Impl> |  | ||||||
| class DWFSlowFermion : public FermionOperator<Impl> |  | ||||||
| { |  | ||||||
| public: |  | ||||||
|   INHERIT_IMPL_TYPES(Impl); |  | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   // Implement the abstract base |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   GridBase *GaugeGrid(void) { return _grid4; } |  | ||||||
|   GridBase *GaugeRedBlackGrid(void) { return _cbgrid4; } |  | ||||||
|   GridBase *FermionGrid(void) { return _grid; } |  | ||||||
|   GridBase *FermionRedBlackGrid(void) { return _cbgrid; } |  | ||||||
|  |  | ||||||
|   FermionField _tmp; |  | ||||||
|   FermionField &tmp(void) { return _tmp; } |  | ||||||
|  |  | ||||||
|   ////////////////////////////////////////////////////////////////// |  | ||||||
|   // override multiply; cut number routines if pass dagger argument |  | ||||||
|   // and also make interface more uniformly consistent |  | ||||||
|   ////////////////////////////////////////////////////////////////// |  | ||||||
|   virtual void  M(const FermionField &in, FermionField &out) |  | ||||||
|   { |  | ||||||
|     FermionField tmp(_grid); |  | ||||||
|     out = (5.0 - M5) * in; |  | ||||||
|     Dhop(in,tmp,DaggerNo); |  | ||||||
|     out = out + tmp; |  | ||||||
|   } |  | ||||||
|   virtual void  Mdag(const FermionField &in, FermionField &out) |  | ||||||
|   { |  | ||||||
|     FermionField tmp(_grid); |  | ||||||
|     out = (5.0 - M5) * in; |  | ||||||
|     Dhop(in,tmp,DaggerYes); |  | ||||||
|     out = out + tmp; |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   ///////////////////////////////////////////////////////// |  | ||||||
|   // half checkerboard operations 5D redblack so just site identiy |  | ||||||
|   ///////////////////////////////////////////////////////// |  | ||||||
|   void Meooe(const FermionField &in, FermionField &out) |  | ||||||
|   { |  | ||||||
|     if ( in.Checkerboard() == Odd ) { |  | ||||||
|       this->DhopEO(in,out,DaggerNo); |  | ||||||
|     } else { |  | ||||||
|       this->DhopOE(in,out,DaggerNo); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|   void MeooeDag(const FermionField &in, FermionField &out) |  | ||||||
|   { |  | ||||||
|     if ( in.Checkerboard() == Odd ) { |  | ||||||
|       this->DhopEO(in,out,DaggerYes); |  | ||||||
|     } else { |  | ||||||
|       this->DhopOE(in,out,DaggerYes); |  | ||||||
|     } |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   // allow override for twisted mass and clover |  | ||||||
|   virtual void Mooee(const FermionField &in, FermionField &out) |  | ||||||
|   { |  | ||||||
|     out = (5.0 - M5) * in; |  | ||||||
|   } |  | ||||||
|   virtual void MooeeDag(const FermionField &in, FermionField &out) |  | ||||||
|   { |  | ||||||
|     out = (5.0 - M5) * in; |  | ||||||
|   } |  | ||||||
|   virtual void MooeeInv(const FermionField &in, FermionField &out) |  | ||||||
|   { |  | ||||||
|     out = (1.0/(5.0 - M5)) * in; |  | ||||||
|   }; |  | ||||||
|   virtual void MooeeInvDag(const FermionField &in, FermionField &out) |  | ||||||
|   { |  | ||||||
|     out = (1.0/(5.0 - M5)) * in; |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   virtual void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _mass,std::vector<double> twist) {} ; |  | ||||||
|  |  | ||||||
|   //////////////////////// |  | ||||||
|   // Derivative interface |  | ||||||
|   //////////////////////// |  | ||||||
|   // Interface calls an internal routine |  | ||||||
|   void DhopDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)  { assert(0);}; |  | ||||||
|   void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);}; |  | ||||||
|   void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);}; |  | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   // non-hermitian hopping term; half cb or both |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   void Dhop(const FermionField &in, FermionField &out, int dag) |  | ||||||
|   { |  | ||||||
|     FermionField tmp(in.Grid()); |  | ||||||
|     Dhop5(in,out,MassField,MassField,dag ); |  | ||||||
|     for(int mu=0;mu<4;mu++){ |  | ||||||
|       DhopDirU(in,Umu[mu],Umu[mu],tmp,mu,dag );    out = out + tmp; |  | ||||||
|     } |  | ||||||
|   }; |  | ||||||
|   void DhopOE(const FermionField &in, FermionField &out, int dag) |  | ||||||
|   { |  | ||||||
|     FermionField tmp(in.Grid()); |  | ||||||
|     assert(in.Checkerboard()==Even); |  | ||||||
|     Dhop5(in,out,MassFieldOdd,MassFieldEven,dag); |  | ||||||
|     for(int mu=0;mu<4;mu++){ |  | ||||||
|       DhopDirU(in,UmuOdd[mu],UmuEven[mu],tmp,mu,dag );    out = out + tmp; |  | ||||||
|     } |  | ||||||
|   }; |  | ||||||
|   void DhopEO(const FermionField &in, FermionField &out, int dag) |  | ||||||
|   { |  | ||||||
|     FermionField tmp(in.Grid()); |  | ||||||
|     assert(in.Checkerboard()==Odd); |  | ||||||
|     Dhop5(in,out, MassFieldEven,MassFieldOdd ,dag );   |  | ||||||
|     for(int mu=0;mu<4;mu++){ |  | ||||||
|       DhopDirU(in,UmuEven[mu],UmuOdd[mu],tmp,mu,dag );    out = out + tmp; |  | ||||||
|     } |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   // Multigrid assistance; force term uses too |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   void Mdir(const FermionField &in, FermionField &out, int dir, int disp){ assert(0);}; |  | ||||||
|   void MdirAll(const FermionField &in, std::vector<FermionField> &out)   { assert(0);}; |  | ||||||
|   void DhopDir(const FermionField &in, FermionField &out, int dir, int disp) { assert(0);}; |  | ||||||
|   void DhopDirAll(const FermionField &in, std::vector<FermionField> &out)    { assert(0);}; |  | ||||||
|   void DhopDirCalc(const FermionField &in, FermionField &out, int dirdisp,int gamma, int dag) { assert(0);}; |  | ||||||
|  |  | ||||||
|   void DhopDirU(const FermionField &in, const GaugeLinkField &U5e, const GaugeLinkField &U5o, FermionField &out, int mu, int dag) |  | ||||||
|   { |  | ||||||
|     RealD     sgn= 1.0; |  | ||||||
|     if (dag ) sgn=-1.0; |  | ||||||
|  |  | ||||||
|     Gamma::Algebra Gmu [] = { |  | ||||||
| 			 Gamma::Algebra::GammaX, |  | ||||||
| 			 Gamma::Algebra::GammaY, |  | ||||||
| 			 Gamma::Algebra::GammaZ, |  | ||||||
| 			 Gamma::Algebra::GammaT |  | ||||||
|     }; |  | ||||||
|  |  | ||||||
|     //    mass is  1,1,1,1,-m has to multiply the round the world term |  | ||||||
|     FermionField tmp (in.Grid()); |  | ||||||
|     tmp = U5e * Cshift(in,mu+1,1); |  | ||||||
|     out = tmp - Gamma(Gmu[mu])*tmp*sgn; |  | ||||||
|      |  | ||||||
|     tmp = Cshift(adj(U5o)*in,mu+1,-1); |  | ||||||
|     out = out + tmp + Gamma(Gmu[mu])*tmp*sgn; |  | ||||||
|  |  | ||||||
|     out = -0.5*out; |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   void Dhop5(const FermionField &in, FermionField &out, ComplexField &massE, ComplexField &massO, int dag) |  | ||||||
|   { |  | ||||||
|     // Mass term.... must multiple the round world with mass = 1,1,1,1, -m |  | ||||||
|     RealD     sgn= 1.0; |  | ||||||
|     if (dag ) sgn=-1.0; |  | ||||||
|  |  | ||||||
|     Gamma G5(Gamma::Algebra::Gamma5); |  | ||||||
|  |  | ||||||
|     FermionField tmp (in.Grid()); |  | ||||||
|     tmp = massE*Cshift(in,0,1); |  | ||||||
|     out = tmp - G5*tmp*sgn; |  | ||||||
|      |  | ||||||
|     tmp = Cshift(massO*in,0,-1); |  | ||||||
|     out = out + tmp + G5*tmp*sgn; |  | ||||||
|     out = -0.5*out; |  | ||||||
|   }; |  | ||||||
|  |  | ||||||
|   // Constructor |  | ||||||
|   DWFSlowFermion(GaugeField &_Umu, GridCartesian &Fgrid, |  | ||||||
| 		 GridRedBlackCartesian &Hgrid, RealD _mass, RealD _M5) |  | ||||||
|     : |  | ||||||
|     _grid(&Fgrid), |  | ||||||
|     _cbgrid(&Hgrid), |  | ||||||
|     _grid4(_Umu.Grid()), |  | ||||||
|     Umu(Nd,&Fgrid), |  | ||||||
|     UmuEven(Nd,&Hgrid), |  | ||||||
|     UmuOdd(Nd,&Hgrid), |  | ||||||
|     MassField(&Fgrid), |  | ||||||
|     MassFieldEven(&Hgrid), |  | ||||||
|     MassFieldOdd(&Hgrid), |  | ||||||
|     M5(_M5), |  | ||||||
|     mass(_mass), |  | ||||||
|     _tmp(&Hgrid) |  | ||||||
|     { |  | ||||||
|       Ls=Fgrid._fdimensions[0]; |  | ||||||
|       ImportGauge(_Umu); |  | ||||||
|  |  | ||||||
|       typedef typename FermionField::scalar_type scalar; |  | ||||||
|  |  | ||||||
|       Lattice<iScalar<vInteger> > coor(&Fgrid); |  | ||||||
|       LatticeCoordinate(coor, 0); // Scoor |  | ||||||
|       ComplexField one(&Fgrid); |  | ||||||
|       MassField =scalar(-mass); |  | ||||||
|       one       =scalar(1.0); |  | ||||||
|       MassField =where(coor==Integer(Ls-1),MassField,one); |  | ||||||
|       for(int mu=0;mu<Nd;mu++){ |  | ||||||
| 	pickCheckerboard(Even,UmuEven[mu],Umu[mu]); |  | ||||||
| 	pickCheckerboard(Odd ,UmuOdd[mu],Umu[mu]); |  | ||||||
|       } |  | ||||||
|       pickCheckerboard(Even,MassFieldEven,MassField); |  | ||||||
|       pickCheckerboard(Odd ,MassFieldOdd,MassField); |  | ||||||
|        |  | ||||||
|     } |  | ||||||
|    |  | ||||||
|   // DoubleStore impl dependent |  | ||||||
|   void ImportGauge(const GaugeField &_Umu4) |  | ||||||
|   { |  | ||||||
|     GaugeLinkField U4(_grid4); |  | ||||||
|     for(int mu=0;mu<Nd;mu++){ |  | ||||||
|       U4 = PeekIndex<LorentzIndex>(_Umu4, mu); |  | ||||||
|       for(int s=0;s<this->Ls;s++){ |  | ||||||
| 	InsertSlice(U4,Umu[mu],s,0); |  | ||||||
|       } |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   // Data members require to support the functionality |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|  |  | ||||||
| public: |  | ||||||
|   virtual RealD Mass(void) { return mass; } |  | ||||||
|   virtual int   isTrivialEE(void) { return 1; }; |  | ||||||
|   RealD mass; |  | ||||||
|   RealD M5; |  | ||||||
|   int Ls; |  | ||||||
|  |  | ||||||
|   GridBase *_grid4; |  | ||||||
|   GridBase *_grid; |  | ||||||
|   GridBase *_cbgrid4; |  | ||||||
|   GridBase *_cbgrid; |  | ||||||
|  |  | ||||||
|   // Copy of the gauge field , with even and odd subsets |  | ||||||
|   std::vector<GaugeLinkField> Umu; |  | ||||||
|   std::vector<GaugeLinkField> UmuEven; |  | ||||||
|   std::vector<GaugeLinkField> UmuOdd; |  | ||||||
|   ComplexField MassField; |  | ||||||
|   ComplexField MassFieldEven; |  | ||||||
|   ComplexField MassFieldOdd; |  | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   // Conserved current utilities |  | ||||||
|   /////////////////////////////////////////////////////////////// |  | ||||||
|   void ContractConservedCurrent(PropagatorField &q_in_1, |  | ||||||
|                                 PropagatorField &q_in_2, |  | ||||||
|                                 PropagatorField &q_out, |  | ||||||
|                                 PropagatorField &phys_src, |  | ||||||
|                                 Current curr_type, |  | ||||||
|                                 unsigned int mu){} |  | ||||||
|   void SeqConservedCurrent(PropagatorField &q_in, |  | ||||||
|                            PropagatorField &q_out, |  | ||||||
|                            PropagatorField &phys_src, |  | ||||||
|                            Current curr_type, |  | ||||||
|                            unsigned int mu, |  | ||||||
|                            unsigned int tmin, |  | ||||||
| 			   unsigned int tmax, |  | ||||||
| 			   ComplexField &lattice_cmplx){} |  | ||||||
| }; |  | ||||||
|  |  | ||||||
| typedef DWFSlowFermion<WilsonImplF> DWFSlowFermionF; |  | ||||||
| typedef DWFSlowFermion<WilsonImplD> DWFSlowFermionD; |  | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); |  | ||||||
| @@ -47,14 +47,12 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk> | |||||||
| //////////////////////////////////////////// | //////////////////////////////////////////// | ||||||
| // Fermion operators / actions | // Fermion operators / actions | ||||||
| //////////////////////////////////////////// | //////////////////////////////////////////// | ||||||
| #include <Grid/qcd/action/fermion/DWFSlow.h>       // Slow DWF |  | ||||||
|  |  | ||||||
| #include <Grid/qcd/action/fermion/WilsonFermion.h>       // 4d wilson like | #include <Grid/qcd/action/fermion/WilsonFermion.h>       // 4d wilson like | ||||||
| NAMESPACE_CHECK(Wilson); | NAMESPACE_CHECK(Wilson); | ||||||
| #include <Grid/qcd/action/fermion/WilsonTMFermion.h>       // 4d wilson like | #include <Grid/qcd/action/fermion/WilsonTMFermion.h>       // 4d wilson like | ||||||
| NAMESPACE_CHECK(WilsonTM); | NAMESPACE_CHECK(WilsonTM); | ||||||
| #include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions | #include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions | ||||||
| #include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h> // 4d compact wilson clover fermions |  | ||||||
| NAMESPACE_CHECK(WilsonClover); | NAMESPACE_CHECK(WilsonClover); | ||||||
| #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types | #include <Grid/qcd/action/fermion/WilsonFermion5D.h>     // 5d base used by all 5d overlap types | ||||||
| NAMESPACE_CHECK(Wilson5D); | NAMESPACE_CHECK(Wilson5D); | ||||||
| @@ -113,164 +111,192 @@ NAMESPACE_CHECK(DWFutils); | |||||||
| // Cayley 5d | // Cayley 5d | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| typedef WilsonFermion<WilsonImplD2> WilsonFermionD2; | typedef WilsonFermion<WilsonImplR> WilsonFermionR; | ||||||
| typedef WilsonFermion<WilsonImplF> WilsonFermionF; | typedef WilsonFermion<WilsonImplF> WilsonFermionF; | ||||||
| typedef WilsonFermion<WilsonImplD> WilsonFermionD; | typedef WilsonFermion<WilsonImplD> WilsonFermionD; | ||||||
|  |  | ||||||
|  | typedef WilsonFermion<WilsonImplRL> WilsonFermionRL; | ||||||
|  | typedef WilsonFermion<WilsonImplFH> WilsonFermionFH; | ||||||
|  | typedef WilsonFermion<WilsonImplDF> WilsonFermionDF; | ||||||
|  |  | ||||||
|  | typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR; | ||||||
| typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF; | typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF; | ||||||
| typedef WilsonFermion<WilsonAdjImplD> WilsonAdjFermionD; | typedef WilsonFermion<WilsonAdjImplD> WilsonAdjFermionD; | ||||||
|  |  | ||||||
|  | typedef WilsonFermion<WilsonTwoIndexSymmetricImplR> WilsonTwoIndexSymmetricFermionR; | ||||||
| typedef WilsonFermion<WilsonTwoIndexSymmetricImplF> WilsonTwoIndexSymmetricFermionF; | typedef WilsonFermion<WilsonTwoIndexSymmetricImplF> WilsonTwoIndexSymmetricFermionF; | ||||||
| typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermionD; | typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermionD; | ||||||
|  |  | ||||||
|  | typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonTwoIndexAntiSymmetricFermionR; | ||||||
| typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonTwoIndexAntiSymmetricFermionF; | typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonTwoIndexAntiSymmetricFermionF; | ||||||
| typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonTwoIndexAntiSymmetricFermionD; | typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonTwoIndexAntiSymmetricFermionD; | ||||||
|  |  | ||||||
| // Sp(2n) |  | ||||||
| typedef WilsonFermion<SpWilsonImplF> SpWilsonFermionF; |  | ||||||
| typedef WilsonFermion<SpWilsonImplD> SpWilsonFermionD; |  | ||||||
|  |  | ||||||
| typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplF> SpWilsonTwoIndexAntiSymmetricFermionF; |  | ||||||
| typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplD> SpWilsonTwoIndexAntiSymmetricFermionD; |  | ||||||
|  |  | ||||||
| typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplF> SpWilsonTwoIndexSymmetricFermionF; |  | ||||||
| typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplD> SpWilsonTwoIndexSymmetricFermionD; |  | ||||||
|  |  | ||||||
| // Twisted mass fermion | // Twisted mass fermion | ||||||
| typedef WilsonTMFermion<WilsonImplD2> WilsonTMFermionD2; | typedef WilsonTMFermion<WilsonImplR> WilsonTMFermionR; | ||||||
| typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF; | typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF; | ||||||
| typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD; | typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD; | ||||||
|  |  | ||||||
| // Clover fermions | // Clover fermions | ||||||
| template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>; | typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR; | ||||||
| template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>; | typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF; | ||||||
|  | typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD; | ||||||
|  |  | ||||||
| typedef WilsonClover<WilsonImplD2> WilsonCloverFermionD2; | typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR; | ||||||
| typedef WilsonClover<WilsonImplF> WilsonCloverFermionF; | typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF; | ||||||
| typedef WilsonClover<WilsonImplD> WilsonCloverFermionD; | typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD; | ||||||
|  |  | ||||||
| typedef WilsonExpClover<WilsonImplD2> WilsonExpCloverFermionD2; | typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR; | ||||||
| typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF; | typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF; | ||||||
| typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD; | typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD; | ||||||
|  |  | ||||||
| typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF; | typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR; | ||||||
| typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD; | typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF; | ||||||
|  | typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD; | ||||||
| typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF; |  | ||||||
| typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD; |  | ||||||
|  |  | ||||||
| typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF; |  | ||||||
| typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD; |  | ||||||
|  |  | ||||||
| // Compact Clover fermions |  | ||||||
| template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>; |  | ||||||
| template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>; |  | ||||||
|  |  | ||||||
| typedef CompactWilsonClover<WilsonImplD2> CompactWilsonCloverFermionD2; |  | ||||||
| typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF; |  | ||||||
| typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD; |  | ||||||
|  |  | ||||||
| typedef CompactWilsonExpClover<WilsonImplD2> CompactWilsonExpCloverFermionD2; |  | ||||||
| typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF; |  | ||||||
| typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD; |  | ||||||
|  |  | ||||||
| typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF; |  | ||||||
| typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD; |  | ||||||
|  |  | ||||||
| typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF; |  | ||||||
| typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD; |  | ||||||
|  |  | ||||||
| typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF; |  | ||||||
| typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD; |  | ||||||
|  |  | ||||||
| // Domain Wall fermions | // Domain Wall fermions | ||||||
|  | typedef DomainWallFermion<WilsonImplR> DomainWallFermionR; | ||||||
| typedef DomainWallFermion<WilsonImplF> DomainWallFermionF; | typedef DomainWallFermion<WilsonImplF> DomainWallFermionF; | ||||||
| typedef DomainWallFermion<WilsonImplD> DomainWallFermionD; | typedef DomainWallFermion<WilsonImplD> DomainWallFermionD; | ||||||
| typedef DomainWallFermion<WilsonImplD2> DomainWallFermionD2; |  | ||||||
|  |  | ||||||
| typedef DomainWallEOFAFermion<WilsonImplD2> DomainWallEOFAFermionD2; | typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL; | ||||||
|  | typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH; | ||||||
|  | typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF; | ||||||
|  |  | ||||||
|  | typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR; | ||||||
| typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF; | typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF; | ||||||
| typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD; | typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD; | ||||||
|  |  | ||||||
| typedef MobiusFermion<WilsonImplD2> MobiusFermionD2; | typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL; | ||||||
|  | typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH; | ||||||
|  | typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF; | ||||||
|  |  | ||||||
|  | typedef MobiusFermion<WilsonImplR> MobiusFermionR; | ||||||
| typedef MobiusFermion<WilsonImplF> MobiusFermionF; | typedef MobiusFermion<WilsonImplF> MobiusFermionF; | ||||||
| typedef MobiusFermion<WilsonImplD> MobiusFermionD; | typedef MobiusFermion<WilsonImplD> MobiusFermionD; | ||||||
|  |  | ||||||
| typedef MobiusEOFAFermion<WilsonImplD2> MobiusEOFAFermionD2; | typedef MobiusFermion<WilsonImplRL> MobiusFermionRL; | ||||||
|  | typedef MobiusFermion<WilsonImplFH> MobiusFermionFH; | ||||||
|  | typedef MobiusFermion<WilsonImplDF> MobiusFermionDF; | ||||||
|  |  | ||||||
|  | typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR; | ||||||
| typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF; | typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF; | ||||||
| typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD; | typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD; | ||||||
|  |  | ||||||
| typedef ZMobiusFermion<ZWilsonImplD2> ZMobiusFermionD2; | typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL; | ||||||
|  | typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH; | ||||||
|  | typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF; | ||||||
|  |  | ||||||
|  | typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR; | ||||||
| typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF; | typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF; | ||||||
| typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD; | typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD; | ||||||
|  |  | ||||||
| typedef ScaledShamirFermion<WilsonImplD2> ScaledShamirFermionD2; | typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL; | ||||||
|  | typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH; | ||||||
|  | typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF; | ||||||
|  |  | ||||||
|  | // Ls vectorised | ||||||
|  | typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR; | ||||||
| typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF; | typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF; | ||||||
| typedef ScaledShamirFermion<WilsonImplD> ScaledShamirFermionD; | typedef ScaledShamirFermion<WilsonImplD> ScaledShamirFermionD; | ||||||
|  |  | ||||||
| typedef MobiusZolotarevFermion<WilsonImplD2> MobiusZolotarevFermionD2; | typedef MobiusZolotarevFermion<WilsonImplR> MobiusZolotarevFermionR; | ||||||
| typedef MobiusZolotarevFermion<WilsonImplF> MobiusZolotarevFermionF; | typedef MobiusZolotarevFermion<WilsonImplF> MobiusZolotarevFermionF; | ||||||
| typedef MobiusZolotarevFermion<WilsonImplD> MobiusZolotarevFermionD; | typedef MobiusZolotarevFermion<WilsonImplD> MobiusZolotarevFermionD; | ||||||
| typedef ShamirZolotarevFermion<WilsonImplD2> ShamirZolotarevFermionD2; | typedef ShamirZolotarevFermion<WilsonImplR> ShamirZolotarevFermionR; | ||||||
| typedef ShamirZolotarevFermion<WilsonImplF> ShamirZolotarevFermionF; | typedef ShamirZolotarevFermion<WilsonImplF> ShamirZolotarevFermionF; | ||||||
| typedef ShamirZolotarevFermion<WilsonImplD> ShamirZolotarevFermionD; | typedef ShamirZolotarevFermion<WilsonImplD> ShamirZolotarevFermionD; | ||||||
|  |  | ||||||
| typedef OverlapWilsonCayleyTanhFermion<WilsonImplD2> OverlapWilsonCayleyTanhFermionD2; | typedef OverlapWilsonCayleyTanhFermion<WilsonImplR> OverlapWilsonCayleyTanhFermionR; | ||||||
| typedef OverlapWilsonCayleyTanhFermion<WilsonImplF> OverlapWilsonCayleyTanhFermionF; | typedef OverlapWilsonCayleyTanhFermion<WilsonImplF> OverlapWilsonCayleyTanhFermionF; | ||||||
| typedef OverlapWilsonCayleyTanhFermion<WilsonImplD> OverlapWilsonCayleyTanhFermionD; | typedef OverlapWilsonCayleyTanhFermion<WilsonImplD> OverlapWilsonCayleyTanhFermionD; | ||||||
| typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD2> OverlapWilsonCayleyZolotarevFermionD2; | typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplR> OverlapWilsonCayleyZolotarevFermionR; | ||||||
| typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplF> OverlapWilsonCayleyZolotarevFermionF; | typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplF> OverlapWilsonCayleyZolotarevFermionF; | ||||||
| typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD> OverlapWilsonCayleyZolotarevFermionD; | typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD> OverlapWilsonCayleyZolotarevFermionD; | ||||||
|  |  | ||||||
| // Continued fraction | // Continued fraction | ||||||
| typedef OverlapWilsonContFracTanhFermion<WilsonImplD2> OverlapWilsonContFracTanhFermionD2; | typedef OverlapWilsonContFracTanhFermion<WilsonImplR> OverlapWilsonContFracTanhFermionR; | ||||||
| typedef OverlapWilsonContFracTanhFermion<WilsonImplF> OverlapWilsonContFracTanhFermionF; | typedef OverlapWilsonContFracTanhFermion<WilsonImplF> OverlapWilsonContFracTanhFermionF; | ||||||
| typedef OverlapWilsonContFracTanhFermion<WilsonImplD> OverlapWilsonContFracTanhFermionD; | typedef OverlapWilsonContFracTanhFermion<WilsonImplD> OverlapWilsonContFracTanhFermionD; | ||||||
| typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD2> OverlapWilsonContFracZolotarevFermionD2; | typedef OverlapWilsonContFracZolotarevFermion<WilsonImplR> OverlapWilsonContFracZolotarevFermionR; | ||||||
| typedef OverlapWilsonContFracZolotarevFermion<WilsonImplF> OverlapWilsonContFracZolotarevFermionF; | typedef OverlapWilsonContFracZolotarevFermion<WilsonImplF> OverlapWilsonContFracZolotarevFermionF; | ||||||
| typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD> OverlapWilsonContFracZolotarevFermionD; | typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD> OverlapWilsonContFracZolotarevFermionD; | ||||||
|  |  | ||||||
| // Partial fraction | // Partial fraction | ||||||
| typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD2> OverlapWilsonPartialFractionTanhFermionD2; | typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplR> OverlapWilsonPartialFractionTanhFermionR; | ||||||
| typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplF> OverlapWilsonPartialFractionTanhFermionF; | typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplF> OverlapWilsonPartialFractionTanhFermionF; | ||||||
| typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD> OverlapWilsonPartialFractionTanhFermionD; | typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD> OverlapWilsonPartialFractionTanhFermionD; | ||||||
|  |  | ||||||
| typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD2> OverlapWilsonPartialFractionZolotarevFermionD2; | typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplR> OverlapWilsonPartialFractionZolotarevFermionR; | ||||||
| typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplF> OverlapWilsonPartialFractionZolotarevFermionF; | typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplF> OverlapWilsonPartialFractionZolotarevFermionF; | ||||||
| typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD> OverlapWilsonPartialFractionZolotarevFermionD; | typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD> OverlapWilsonPartialFractionZolotarevFermionD; | ||||||
|  |  | ||||||
| // Gparity cases; partial list until tested | // Gparity cases; partial list until tested | ||||||
|  | typedef WilsonFermion<GparityWilsonImplR>     GparityWilsonFermionR; | ||||||
| typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF; | typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF; | ||||||
| typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD; | typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD; | ||||||
|  |  | ||||||
|  | typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL; | ||||||
|  | typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH; | ||||||
|  | typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF; | ||||||
|  |  | ||||||
|  | typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR; | ||||||
| typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF; | typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF; | ||||||
| typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD; | typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD; | ||||||
|  |  | ||||||
| typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionD2; | typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL; | ||||||
|  | typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH; | ||||||
|  | typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF; | ||||||
|  |  | ||||||
|  | typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR; | ||||||
| typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF; | typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF; | ||||||
| typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD; | typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD; | ||||||
|  |  | ||||||
| typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionD2; | typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL; | ||||||
|  | typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH; | ||||||
|  | typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF; | ||||||
|  |  | ||||||
|  | typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR; | ||||||
| typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF; | typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF; | ||||||
| typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD; | typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD; | ||||||
|  |  | ||||||
| typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionD2; | typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL; | ||||||
|  | typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH; | ||||||
|  | typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF; | ||||||
|  |  | ||||||
|  | typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR; | ||||||
| typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF; | typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF; | ||||||
| typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD; | typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD; | ||||||
|  |  | ||||||
| typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionD2; | typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL; | ||||||
|  | typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH; | ||||||
|  | typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF; | ||||||
|  |  | ||||||
|  | typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR; | ||||||
| typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF; | typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF; | ||||||
| typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD; | typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD; | ||||||
|  |  | ||||||
|  | typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL; | ||||||
|  | typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH; | ||||||
|  | typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF; | ||||||
|  |  | ||||||
|  | typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR; | ||||||
| typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF; | typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF; | ||||||
| typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD; | typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD; | ||||||
|  |  | ||||||
|  | typedef NaiveStaggeredFermion<StaggeredImplR> NaiveStaggeredFermionR; | ||||||
| typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF; | typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF; | ||||||
| typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD; | typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD; | ||||||
|  |  | ||||||
|  | typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR; | ||||||
| typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF; | typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF; | ||||||
| typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD; | typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD; | ||||||
|  |  | ||||||
|  | #ifndef GRID_CUDA | ||||||
|  | typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplR> ImprovedStaggeredFermionVec5dR; | ||||||
|  | typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplF> ImprovedStaggeredFermionVec5dF; | ||||||
|  | typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermionVec5dD; | ||||||
|  | #endif | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|  |  | ||||||
| //////////////////// | //////////////////// | ||||||
|   | |||||||
| @@ -49,8 +49,6 @@ public: | |||||||
|  |  | ||||||
|   virtual FermionField &tmp(void) = 0; |   virtual FermionField &tmp(void) = 0; | ||||||
|  |  | ||||||
|   virtual void DirichletBlock(const Coordinate & _Block) { assert(0); }; |  | ||||||
|    |  | ||||||
|   GridBase * Grid(void)   { return FermionGrid(); };   // this is all the linalg routines need to know |   GridBase * Grid(void)   { return FermionGrid(); };   // this is all the linalg routines need to know | ||||||
|   GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); }; |   GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); }; | ||||||
|  |  | ||||||
|   | |||||||
| @@ -153,8 +153,8 @@ public: | |||||||
|   typedef typename Impl::StencilImpl             StencilImpl;		\ |   typedef typename Impl::StencilImpl             StencilImpl;		\ | ||||||
|   typedef typename Impl::ImplParams               ImplParams;	        \ |   typedef typename Impl::ImplParams               ImplParams;	        \ | ||||||
|   typedef typename Impl::StencilImpl::View_type  StencilView;		\ |   typedef typename Impl::StencilImpl::View_type  StencilView;		\ | ||||||
|   typedef const typename ViewMap<FermionField>::Type      FermionFieldView;	\ |   typedef typename ViewMap<FermionField>::Type      FermionFieldView;	\ | ||||||
|   typedef const typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView; |   typedef typename ViewMap<DoubledGaugeField>::Type DoubledGaugeFieldView; | ||||||
|  |  | ||||||
| #define INHERIT_IMPL_TYPES(Base)		\ | #define INHERIT_IMPL_TYPES(Base)		\ | ||||||
|   INHERIT_GIMPL_TYPES(Base)			\ |   INHERIT_GIMPL_TYPES(Base)			\ | ||||||
| @@ -183,8 +183,7 @@ NAMESPACE_CHECK(ImplStaggered); | |||||||
| ///////////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////////// | ||||||
| // Single flavour one component spinors with colour index. 5d vec | // Single flavour one component spinors with colour index. 5d vec | ||||||
| ///////////////////////////////////////////////////////////////////////////// | ///////////////////////////////////////////////////////////////////////////// | ||||||
| // Deprecate Vec5d | #include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h>  | ||||||
| //#include <Grid/qcd/action/fermion/StaggeredVec5dImpl.h>  | NAMESPACE_CHECK(ImplStaggered5dVec);   | ||||||
| //NAMESPACE_CHECK(ImplStaggered5dVec);   |  | ||||||
|  |  | ||||||
|  |  | ||||||
|   | |||||||
| @@ -30,18 +30,6 @@ directory | |||||||
|  |  | ||||||
| NAMESPACE_BEGIN(Grid); | NAMESPACE_BEGIN(Grid); | ||||||
|  |  | ||||||
| /* |  | ||||||
|   Policy implementation for G-parity boundary conditions |  | ||||||
|  |  | ||||||
|   Rather than treating the gauge field as a flavored field, the Grid implementation of G-parity treats the gauge field as a regular |  | ||||||
|   field with complex conjugate boundary conditions. In order to ensure the second flavor interacts with the conjugate links and the first |  | ||||||
|   with the regular links we overload the functionality of doubleStore, whose purpose is to store the gauge field and the barrel-shifted gauge field |  | ||||||
|   to avoid communicating links when applying the Dirac operator, such that the double-stored field contains also a flavor index which maps to |  | ||||||
|   either the link or the conjugate link. This flavored field is then used by multLink to apply the correct link to a spinor. |  | ||||||
|  |  | ||||||
|   Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction.  |  | ||||||
|   mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs |  | ||||||
|  */ |  | ||||||
| template <class S, class Representation = FundamentalRepresentation, class Options=CoeffReal> | template <class S, class Representation = FundamentalRepresentation, class Options=CoeffReal> | ||||||
| class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > { | class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > { | ||||||
| public: | public: | ||||||
| @@ -109,30 +97,42 @@ public: | |||||||
|     Coordinate icoor; |     Coordinate icoor; | ||||||
|  |  | ||||||
| #ifdef GRID_SIMT | #ifdef GRID_SIMT | ||||||
|  |     _Spinor tmp; | ||||||
|  |  | ||||||
|     const int Nsimd =SiteDoubledGaugeField::Nsimd(); |     const int Nsimd =SiteDoubledGaugeField::Nsimd(); | ||||||
|     int s = acceleratorSIMTlane(Nsimd); |     int s = acceleratorSIMTlane(Nsimd); | ||||||
|     St.iCoorFromIindex(icoor,s); |     St.iCoorFromIindex(icoor,s); | ||||||
|  |  | ||||||
|     int mmu = mu % Nd; |     int mmu = mu % Nd; | ||||||
|  |     if ( SE->_around_the_world && St.parameters.twists[mmu] ) { | ||||||
|  |        | ||||||
|  |       int permute_lane = (sl==1)  | ||||||
|  |     	|| ((distance== 1)&&(icoor[direction]==1)) | ||||||
|  | 	|| ((distance==-1)&&(icoor[direction]==0)); | ||||||
|  |  | ||||||
|     auto UU0=coalescedRead(U(0)(mu)); |       if ( permute_lane ) {  | ||||||
|     auto UU1=coalescedRead(U(1)(mu)); | 	tmp(0) = chi(1); | ||||||
|      | 	tmp(1) = chi(0); | ||||||
|     //Decide whether we do a G-parity flavor twist |       } else { | ||||||
|     //Note: this assumes (but does not check) that sl==1 || sl==2 i.e. max 2 SIMD lanes in G-parity dir | 	tmp(0) = chi(0); | ||||||
|     //It also assumes (but does not check) that abs(distance) == 1 | 	tmp(1) = chi(1); | ||||||
|     int permute_lane = (sl==1)  |       } | ||||||
|     || ((distance== 1)&&(icoor[direction]==1)) |  | ||||||
|     || ((distance==-1)&&(icoor[direction]==0)); |  | ||||||
|  |  | ||||||
|     permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu] && mmu < Nd-1; //only if we are going around the world in a spatial direction |       auto UU0=coalescedRead(U(0)(mu)); | ||||||
|  |       auto UU1=coalescedRead(U(1)(mu)); | ||||||
|  |  | ||||||
|     //Apply the links |       mult(&phi(0),&UU0,&tmp(0)); | ||||||
|     int f_upper = permute_lane ? 1 : 0; |       mult(&phi(1),&UU1,&tmp(1)); | ||||||
|     int f_lower = !f_upper; |  | ||||||
|  |  | ||||||
|     mult(&phi(0),&UU0,&chi(f_upper)); |     } else { | ||||||
|     mult(&phi(1),&UU1,&chi(f_lower)); |  | ||||||
|  |       auto UU0=coalescedRead(U(0)(mu)); | ||||||
|  |       auto UU1=coalescedRead(U(1)(mu)); | ||||||
|  |  | ||||||
|  |       mult(&phi(0),&UU0,&chi(0)); | ||||||
|  |       mult(&phi(1),&UU1,&chi(1)); | ||||||
|  |  | ||||||
|  |     } | ||||||
|  |  | ||||||
| #else | #else | ||||||
|     typedef _Spinor vobj; |     typedef _Spinor vobj; | ||||||
| @@ -151,10 +151,10 @@ public: | |||||||
|     assert((distance == 1) || (distance == -1));  // nearest neighbour stencil hard code |     assert((distance == 1) || (distance == -1));  // nearest neighbour stencil hard code | ||||||
|     assert((sl == 1) || (sl == 2)); |     assert((sl == 1) || (sl == 2)); | ||||||
|  |  | ||||||
|     //If this site is an global boundary site, perform the G-parity flavor twist |     if ( SE->_around_the_world && St.parameters.twists[mmu] ) { | ||||||
|     if ( mmu < Nd-1 && SE->_around_the_world && St.parameters.twists[mmu] ) { |  | ||||||
|       if ( sl == 2 ) { |       if ( sl == 2 ) { | ||||||
| 	//Only do the twist for lanes on the edge of the physical node |         | ||||||
| 	ExtractBuffer<sobj> vals(Nsimd); | 	ExtractBuffer<sobj> vals(Nsimd); | ||||||
|  |  | ||||||
| 	extract(chi,vals); | 	extract(chi,vals); | ||||||
| @@ -209,19 +209,6 @@ public: | |||||||
|     reg = memory; |     reg = memory; | ||||||
|   } |   } | ||||||
|  |  | ||||||
|  |  | ||||||
|   //Poke 'poke_f0' onto flavor 0 and 'poke_f1' onto flavor 1 in direction mu of the doubled gauge field Uds |  | ||||||
|   inline void pokeGparityDoubledGaugeField(DoubledGaugeField &Uds, const GaugeLinkField &poke_f0, const GaugeLinkField &poke_f1, const int mu){ |  | ||||||
|     autoView(poke_f0_v, poke_f0, CpuRead); |  | ||||||
|     autoView(poke_f1_v, poke_f1, CpuRead); |  | ||||||
|     autoView(Uds_v, Uds, CpuWrite); |  | ||||||
|     thread_foreach(ss,poke_f0_v,{ |  | ||||||
| 	Uds_v[ss](0)(mu) = poke_f0_v[ss](); |  | ||||||
| 	Uds_v[ss](1)(mu) = poke_f1_v[ss](); |  | ||||||
|       }); |  | ||||||
|   } |  | ||||||
|      |  | ||||||
|  |  | ||||||
|   inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu) |   inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu) | ||||||
|   { |   { | ||||||
|     conformable(Uds.Grid(),GaugeGrid); |     conformable(Uds.Grid(),GaugeGrid); | ||||||
| @@ -232,19 +219,14 @@ public: | |||||||
|     GaugeLinkField Uconj(GaugeGrid); |     GaugeLinkField Uconj(GaugeGrid); | ||||||
|     |     | ||||||
|     Lattice<iScalar<vInteger> > coor(GaugeGrid); |     Lattice<iScalar<vInteger> > coor(GaugeGrid); | ||||||
|  |          | ||||||
|     //Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction.  |     for(int mu=0;mu<Nd;mu++){ | ||||||
|     //mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs         |            | ||||||
|     for(int mu=0;mu<Nd-1;mu++){ |       LatticeCoordinate(coor,mu); | ||||||
|  |  | ||||||
|       if( Params.twists[mu] ){ |  | ||||||
| 	LatticeCoordinate(coor,mu); |  | ||||||
|       } |  | ||||||
|            |            | ||||||
|       U     = PeekIndex<LorentzIndex>(Umu,mu); |       U     = PeekIndex<LorentzIndex>(Umu,mu); | ||||||
|       Uconj = conjugate(U); |       Uconj = conjugate(U); | ||||||
|       |       | ||||||
|       // Implement the isospin rotation sign on the boundary between f=1 and f=0 |  | ||||||
|       // This phase could come from a simple bc 1,1,-1,1 .. |       // This phase could come from a simple bc 1,1,-1,1 .. | ||||||
|       int neglink = GaugeGrid->GlobalDimensions()[mu]-1; |       int neglink = GaugeGrid->GlobalDimensions()[mu]-1; | ||||||
|       if ( Params.twists[mu] ) {  |       if ( Params.twists[mu] ) {  | ||||||
| @@ -259,7 +241,7 @@ public: | |||||||
| 	thread_foreach(ss,U_v,{ | 	thread_foreach(ss,U_v,{ | ||||||
| 	    Uds_v[ss](0)(mu) = U_v[ss](); | 	    Uds_v[ss](0)(mu) = U_v[ss](); | ||||||
| 	    Uds_v[ss](1)(mu) = Uconj_v[ss](); | 	    Uds_v[ss](1)(mu) = Uconj_v[ss](); | ||||||
| 	}); | 	  }); | ||||||
|       } |       } | ||||||
|            |            | ||||||
|       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary |       U     = adj(Cshift(U    ,mu,-1));      // correct except for spanning the boundary | ||||||
| @@ -290,38 +272,6 @@ public: | |||||||
|         }); |         }); | ||||||
|       } |       } | ||||||
|     } |     } | ||||||
|  |  | ||||||
|     { //periodic / antiperiodic temporal BCs |  | ||||||
|       int mu = Nd-1; |  | ||||||
|       int L   = GaugeGrid->GlobalDimensions()[mu]; |  | ||||||
|       int Lmu = L - 1; |  | ||||||
|  |  | ||||||
|       LatticeCoordinate(coor, mu); |  | ||||||
|  |  | ||||||
|       U = PeekIndex<LorentzIndex>(Umu, mu); //Get t-directed links |  | ||||||
|        |  | ||||||
|       GaugeLinkField *Upoke = &U; |  | ||||||
|  |  | ||||||
|       if(Params.twists[mu]){ //antiperiodic |  | ||||||
| 	Utmp =  where(coor == Lmu, -U, U); |  | ||||||
| 	Upoke = &Utmp; |  | ||||||
|       } |  | ||||||
|      |  | ||||||
|       Uconj = conjugate(*Upoke); //second flavor interacts with conjugate links       |  | ||||||
|       pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu); |  | ||||||
|  |  | ||||||
|       //Get the barrel-shifted field |  | ||||||
|       Utmp = adj(Cshift(U, mu, -1)); //is a forward shift! |  | ||||||
|       Upoke = &Utmp; |  | ||||||
|  |  | ||||||
|       if(Params.twists[mu]){ |  | ||||||
| 	U = where(coor == 0, -Utmp, Utmp);  //boundary phase |  | ||||||
| 	Upoke = &U; |  | ||||||
|       } |  | ||||||
|        |  | ||||||
|       Uconj = conjugate(*Upoke); |  | ||||||
|       pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu + 4); |  | ||||||
|     } |  | ||||||
|   } |   } | ||||||
|        |        | ||||||
|   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) { |   inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) { | ||||||
| @@ -360,48 +310,28 @@ public: | |||||||
|   inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){ |   inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){ | ||||||
|     assert(0); |     assert(0); | ||||||
|   } |   } | ||||||
|   |  | ||||||
|   inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField Ã, int mu) { |  | ||||||
|     int Ls=Btilde.Grid()->_fdimensions[0]; |  | ||||||
|      |  | ||||||
|     { |  | ||||||
|       GridBase *GaugeGrid = mat.Grid(); |  | ||||||
|       Lattice<iScalar<vInteger> > coor(GaugeGrid); |  | ||||||
|  |  | ||||||
|       if( Params.twists[mu] ){ |  | ||||||
| 	LatticeCoordinate(coor,mu); |  | ||||||
|       } |  | ||||||
|  |  | ||||||
|       autoView( mat_v , mat, AcceleratorWrite); |  | ||||||
|       autoView( Btilde_v , Btilde, AcceleratorRead); |  | ||||||
|       autoView( Atilde_v , Atilde, AcceleratorRead); |  | ||||||
|       accelerator_for(sss,mat.Grid()->oSites(), FermionField::vector_type::Nsimd(),{	   |  | ||||||
|   	  int sU=sss; |  | ||||||
|   	  typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType; |  | ||||||
|   	  ColorMatrixType sum; |  | ||||||
|   	  zeroit(sum); |  | ||||||
|   	  for(int s=0;s<Ls;s++){ |  | ||||||
|   	    int sF = s+Ls*sU; |  | ||||||
|   	    for(int spn=0;spn<Ns;spn++){ //sum over spin |  | ||||||
| 	      //Flavor 0 |  | ||||||
|   	      auto bb = coalescedRead(Btilde_v[sF](0)(spn) ); //color vector |  | ||||||
|   	      auto aa = coalescedRead(Atilde_v[sF](0)(spn) ); |  | ||||||
|   	      sum = sum + outerProduct(bb,aa); |  | ||||||
|  |  | ||||||
|   	      //Flavor 1 |  | ||||||
|   	      bb = coalescedRead(Btilde_v[sF](1)(spn) ); |  | ||||||
|   	      aa = coalescedRead(Atilde_v[sF](1)(spn) ); |  | ||||||
|   	      sum = sum + conjugate(outerProduct(bb,aa)); |  | ||||||
|   	    } |  | ||||||
|   	  }	     |  | ||||||
|   	  coalescedWrite(mat_v[sU](mu)(), sum); |  | ||||||
|   	}); |  | ||||||
|     } |  | ||||||
|   } |  | ||||||
|  |  | ||||||
|  |  | ||||||
|    |    | ||||||
|  |   inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField Ã, int mu) { | ||||||
|  |  | ||||||
|  |     int Ls = Btilde.Grid()->_fdimensions[0]; | ||||||
|  |          | ||||||
|  |     GaugeLinkField tmp(mat.Grid()); | ||||||
|  |     tmp = Zero(); | ||||||
|  |     { | ||||||
|  |       autoView( tmp_v , tmp, CpuWrite); | ||||||
|  |       autoView( Atilde_v , Atilde, CpuRead); | ||||||
|  |       autoView( Btilde_v , Btilde, CpuRead); | ||||||
|  |       thread_for(ss,tmp.Grid()->oSites(),{ | ||||||
|  | 	  for (int s = 0; s < Ls; s++) { | ||||||
|  | 	    int sF = s + Ls * ss; | ||||||
|  | 	    auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF])); | ||||||
|  | 	    tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1)); | ||||||
|  | 	  } | ||||||
|  | 	}); | ||||||
|  |     } | ||||||
|  |     PokeIndex<LorentzIndex>(mat, tmp, mu); | ||||||
|  |     return; | ||||||
|  |   } | ||||||
|    |    | ||||||
| }; | }; | ||||||
|  |  | ||||||
| @@ -409,8 +339,8 @@ typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffReal> Gparit | |||||||
| typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffReal> GparityWilsonImplF;  // Float | typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffReal> GparityWilsonImplF;  // Float | ||||||
| typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffReal> GparityWilsonImplD;  // Double | typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffReal> GparityWilsonImplD;  // Double | ||||||
|   |   | ||||||
| //typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec | typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec | ||||||
| //typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float | typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float | ||||||
| //typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double | typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double | ||||||
|  |  | ||||||
| NAMESPACE_END(Grid); | NAMESPACE_END(Grid); | ||||||
|   | |||||||
| @@ -47,6 +47,18 @@ public: | |||||||
|   FermionField _tmp; |   FermionField _tmp; | ||||||
|   FermionField &tmp(void) { return _tmp; } |   FermionField &tmp(void) { return _tmp; } | ||||||
|  |  | ||||||
|  |   //////////////////////////////////////// | ||||||
|  |   // Performance monitoring | ||||||
|  |   //////////////////////////////////////// | ||||||
|  |   void Report(void); | ||||||
|  |   void ZeroCounters(void); | ||||||
|  |   double DhopTotalTime; | ||||||
|  |   double DhopCalls; | ||||||
|  |   double DhopCommTime; | ||||||
|  |   double DhopComputeTime; | ||||||
|  |   double DhopComputeTime2; | ||||||
|  |   double DhopFaceTime; | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////// | ||||||
|   // Implement the abstract base |   // Implement the abstract base | ||||||
|   /////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////// | ||||||
|   | |||||||
| @@ -52,6 +52,18 @@ public: | |||||||
|   FermionField _tmp; |   FermionField _tmp; | ||||||
|   FermionField &tmp(void) { return _tmp; } |   FermionField &tmp(void) { return _tmp; } | ||||||
|  |  | ||||||
|  |   //////////////////////////////////////// | ||||||
|  |   // Performance monitoring | ||||||
|  |   //////////////////////////////////////// | ||||||
|  |   void Report(void); | ||||||
|  |   void ZeroCounters(void); | ||||||
|  |   double DhopTotalTime; | ||||||
|  |   double DhopCalls; | ||||||
|  |   double DhopCommTime; | ||||||
|  |   double DhopComputeTime; | ||||||
|  |   double DhopComputeTime2; | ||||||
|  |   double DhopFaceTime; | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////// | ||||||
|   // Implement the abstract base |   // Implement the abstract base | ||||||
|   /////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////// | ||||||
|   | |||||||
| @@ -85,7 +85,7 @@ class MADWF | |||||||
|       maxiter     =_maxiter; |       maxiter     =_maxiter; | ||||||
|     }; |     }; | ||||||
|     |     | ||||||
|   void operator() (const FermionFieldo &src,FermionFieldo &sol5) |   void operator() (const FermionFieldo &src4,FermionFieldo &sol5) | ||||||
|   { |   { | ||||||
|     std::cout << GridLogMessage<< " ************************************************" << std::endl; |     std::cout << GridLogMessage<< " ************************************************" << std::endl; | ||||||
|     std::cout << GridLogMessage<< "  MADWF-like algorithm                           " << std::endl; |     std::cout << GridLogMessage<< "  MADWF-like algorithm                           " << std::endl; | ||||||
| @@ -114,16 +114,8 @@ class MADWF | |||||||
|     /////////////////////////////////////// |     /////////////////////////////////////// | ||||||
|     //Import source, include Dminus factors |     //Import source, include Dminus factors | ||||||
|     /////////////////////////////////////// |     /////////////////////////////////////// | ||||||
|     GridBase *src_grid = src.Grid(); |     Mato.ImportPhysicalFermionSource(src4,b);  | ||||||
|  |     std::cout << GridLogMessage << " src4 " <<norm2(src4)<<std::endl; | ||||||
|     assert( (src_grid == Mato.GaugeGrid()) || (src_grid == Mato.FermionGrid())); |  | ||||||
|  |  | ||||||
|     if ( src_grid == Mato.GaugeGrid() ) { |  | ||||||
|       Mato.ImportPhysicalFermionSource(src,b); |  | ||||||
|     } else { |  | ||||||
|       b=src; |  | ||||||
|     } |  | ||||||
|     std::cout << GridLogMessage << " src " <<norm2(src)<<std::endl; |  | ||||||
|     std::cout << GridLogMessage << " b    " <<norm2(b)<<std::endl; |     std::cout << GridLogMessage << " b    " <<norm2(b)<<std::endl; | ||||||
|  |  | ||||||
|     defect = b; |     defect = b; | ||||||
|   | |||||||
| @@ -47,6 +47,18 @@ public: | |||||||
|   FermionField _tmp; |   FermionField _tmp; | ||||||
|   FermionField &tmp(void) { return _tmp; } |   FermionField &tmp(void) { return _tmp; } | ||||||
|  |  | ||||||
|  |   //////////////////////////////////////// | ||||||
|  |   // Performance monitoring | ||||||
|  |   //////////////////////////////////////// | ||||||
|  |   void Report(void); | ||||||
|  |   void ZeroCounters(void); | ||||||
|  |   double DhopTotalTime; | ||||||
|  |   double DhopCalls; | ||||||
|  |   double DhopCommTime; | ||||||
|  |   double DhopComputeTime; | ||||||
|  |   double DhopComputeTime2; | ||||||
|  |   double DhopFaceTime; | ||||||
|  |  | ||||||
|   /////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////// | ||||||
|   // Implement the abstract base |   // Implement the abstract base | ||||||
|   /////////////////////////////////////////////////////////////// |   /////////////////////////////////////////////////////////////// | ||||||
|   | |||||||
Some files were not shown because too many files have changed in this diff Show More
		Reference in New Issue
	
	Block a user