Update to enable multiple records per file more consistent with SciDAC.

open, close, write records...

.gitignore

@@ -92,6 +92,7 @@ build*/*
 #####################
 *.xcodeproj/*
 build.sh
+.vscode
 
 # Eigen source #
 ################
@@ -106,6 +107,10 @@ lib/fftw/*
 m4/lt*
 m4/libtool.m4
 
+# github pages #
+################
+gh-pages/
+
 # Buck files #
 ##############
 .buck*
@@ -117,3 +122,4 @@ make-bin-BUCK.sh
 #####################
 lib/qcd/spin/gamma-gen/*.h
 lib/qcd/spin/gamma-gen/*.cc
+

.travis.yml

@@ -7,9 +7,11 @@ cache:
 matrix:
   include:
     - os:        osx
-      osx_image: xcode7.2
+      osx_image: xcode8.3
       compiler: clang
     - compiler: gcc
+      dist: trusty
+      sudo: required
       addons:
         apt:
           sources:
@@ -24,6 +26,8 @@ matrix:
             - binutils-dev
       env: VERSION=-4.9
     - compiler: gcc
+      dist: trusty
+      sudo: required
       addons:
         apt:
           sources:
@@ -38,6 +42,7 @@ matrix:
             - binutils-dev
       env: VERSION=-5
     - compiler: clang
+      dist: trusty
       addons:
         apt:
           sources:
@@ -52,6 +57,7 @@ matrix:
             - binutils-dev
       env: CLANG_LINK=http://llvm.org/releases/3.8.0/clang+llvm-3.8.0-x86_64-linux-gnu-ubuntu-14.04.tar.xz
     - compiler: clang
+      dist: trusty
       addons:
         apt:
           sources:
@@ -73,13 +79,15 @@ before_install:
     - 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; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install openmpi; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]] && [[ "$CC" == "gcc" ]]; then brew install gcc5; fi
     
 install:
     - export CC=$CC$VERSION
     - export CXX=$CXX$VERSION
     - echo $PATH
+    - which autoconf
+    - autoconf  --version
+    - which automake
+    - automake  --version
     - which $CC
     - $CC  --version
     - which $CXX
@@ -92,15 +100,15 @@ script:
     - cd build
     - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=none
     - make -j4 
-    - ./benchmarks/Benchmark_dwf --threads 1
+    - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
     - echo make clean
     - ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none
     - make -j4
-    - ./benchmarks/Benchmark_dwf --threads 1
+    - ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
+    - make check
     - echo make clean
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then export CXXFLAGS='-DMPI_UINT32_T=MPI_UNSIGNED -DMPI_UINT64_T=MPI_UNSIGNED_LONG'; fi
-    - ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto
-    - make -j4
-    - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
-    - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then mpirun -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto ; fi
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then make -j4; fi
+    - if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi
+
 

Makefile.am

@@ -3,10 +3,15 @@ SUBDIRS = lib benchmarks tests extras
 
 include $(top_srcdir)/doxygen.inc
 
-tests: all
-	$(MAKE) -C tests tests
+bin_SCRIPTS=grid-config
 
-.PHONY: tests doxygen-run doxygen-doc $(DX_PS_GOAL) $(DX_PDF_GOAL)
+
+.PHONY: bench check tests doxygen-run doxygen-doc $(DX_PS_GOAL) $(DX_PDF_GOAL)
+
+tests-local: all
+bench-local: all
+check-local: all
 
 AM_CXXFLAGS += -I$(top_builddir)/include
+
 ACLOCAL_AMFLAGS = -I m4

README.md

@@ -22,6 +22,26 @@ Last update Nov 2016.
 
 _Please do not send pull requests to the `master` branch which is reserved for releases._
 
+### Compilers
+
+Intel ICPC v16.0.3 and later
+
+Clang v3.5 and later (need 3.8 and later for OpenMP)
+
+GCC   v4.9.x (recommended)
+
+GCC   v6.3 and later
+
+### Important: 
+
+Some versions of GCC appear to have a bug under high optimisation (-O2, -O3).
+
+The safety of these compiler versions cannot be guaranteed at this time. Follow Issue 100 for details and updates.
+
+GCC   v5.x
+
+GCC   v6.1, v6.2
+
 ### Bug report
 
 _To help us tracking and solving more efficiently issues with Grid, please report problems using the issue system of GitHub rather than sending emails to Grid developers._
@@ -32,7 +52,7 @@ When you file an issue, please go though the following checklist:
 2. 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.
 3. Give the exact `configure` command used.
 4. Attach `config.log`.
-5. Attach `config.summary`.
+5. Attach `grid.config.summary`.
 6. Attach the output of `make V=1`.
 7. Describe the issue and any previous attempt to solve it. If relevant, show how to reproduce the issue using a minimal working example.
 
@@ -95,10 +115,10 @@ install Grid. Other options are detailed in the next section, you can also use `
 `CXX`, `CXXFLAGS`, `LDFLAGS`, ... environment variables can be modified to
 customise the build.
 
-Finally, you can build and install Grid:
+Finally, you can build, check, and install Grid:
 
 ``` bash
-make; make install
+make; make check; make install
 ```
 
 To minimise the build time, only the tests at the root of the `tests` directory are built by default. If you want to build tests in the sub-directory `<subdir>` you can execute:
@@ -121,7 +141,7 @@ If you want to build all the tests at once just use `make tests`.
 - `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 32 bytes).
 - `--enable-precision={single|double}`: set the default precision (default: `double`).
 - `--enable-precision=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
-- `--enable-rng={ranlux48|mt19937}`: choose the RNG (default: `ranlux48 `).
+- `--enable-rng={sitmo|ranlux48|mt19937}`: choose the RNG (default: `sitmo `).
 - `--disable-timers`: disable system dependent high-resolution timers.
 - `--enable-chroma`: enable Chroma regression tests.
 - `--enable-doxygen-doc`: enable the Doxygen documentation generation (build with `make doxygen-doc`)
@@ -159,7 +179,6 @@ Alternatively, some CPU codenames can be directly used:
 
 | `<code>`    | Description                            |
 | ----------- | -------------------------------------- |
-| `KNC`       | [Intel Xeon Phi codename Knights Corner](http://ark.intel.com/products/codename/57721/Knights-Corner) |
 | `KNL`       | [Intel Xeon Phi codename Knights Landing](http://ark.intel.com/products/codename/48999/Knights-Landing) |
 | `BGQ`       | Blue Gene/Q                            |
 

TODO

@@ -1,6 +1,27 @@
 TODO:
 ---------------
 
+Peter's work list:
+1)- Precision conversion and sort out localConvert      <-- 
+2)- Remove DenseVector, DenseMatrix; Use Eigen instead. <-- 
+
+-- Profile CG, BlockCG, etc... Flop count/rate -- PARTIAL, time but no flop/s yet
+-- Physical propagator interface
+-- Conserved currents
+-- GaugeFix into central location
+-- Multigrid Wilson and DWF, compare to other Multigrid implementations
+-- HDCR resume
+
+Recent DONE 
+-- Binary I/O speed up & x-strips                      <-- DONE
+-- Cut down the exterior overhead                      <-- DONE
+-- Interior legs from SHM comms                        <-- DONE
+-- Half-precision comms                                <-- DONE
+-- Merge high precision reduction into develop        
+-- multiRHS DWF; benchmark on Cori/BNL for comms elimination
+   -- slice* linalg routines for multiRHS, BlockCG    
+
+-----
 * Forces; the UdSdU  term in gauge force term is half of what I think it should
   be. This is a consequence of taking ONLY the first term in:
 
@@ -21,16 +42,8 @@ TODO:
   This means we must double the force in the Test_xxx_force routines, and is the origin of the factor of two.
   This 2x is applied by hand in the fermion routines and in the Test_rect_force routine.
 
-
-Policies:
-
-* Link smearing/boundary conds; Policy class based implementation ; framework more in place
-
 * Support different boundary conditions (finite temp, chem. potential ... )
 
-* Support different fermion representations? 
-  - contained entirely within the integrator presently
-
 - Sign of force term.
 
 - Reversibility test.
@@ -41,11 +54,6 @@ Policies:
 
 - Audit oIndex usage for cb behaviour
 
-- Rectangle gauge actions.
-  Iwasaki,
-  Symanzik,
-  ... etc...
-
 - Prepare multigrid for HMC. - Alternate setup schemes.
 
 - Support for ILDG --- ugly, not done
@@ -55,9 +63,11 @@ Policies:
 - FFTnD ?
 
 - Gparity; hand opt use template specialisation elegance to enable the optimised paths ?
+
 - Gparity force term; Gparity (R)HMC.
-- Random number state save restore
+
 - Mobius implementation clean up to rmove #if 0 stale code sequences
+
 - CG -- profile carefully, kernel fusion, whole CG performance measurements.
 
 ================================================================
@@ -90,6 +100,7 @@ Insert/Extract
 Not sure of status of this -- reverify. Things are working nicely now though.
 
 * Make the Tensor types and Complex etc... play more nicely.
+
   - TensorRemove is a hack, come up with a long term rationalised approach to Complex vs. Scalar<Scalar<Scalar<Complex > > >
     QDP forces use of "toDouble" to get back to non tensor scalar. This role is presently taken TensorRemove, but I
     want to introduce a syntax that does not require this.
@@ -112,6 +123,8 @@ Not sure of status of this -- reverify. Things are working nicely now though.
 RECENT
 ---------------
 
+  - Support different fermion representations? -- DONE
+  - contained entirely within the integrator presently
   - Clean up HMC                                                             -- DONE
   - LorentzScalar<GaugeField> gets Gauge link type (cleaner).                -- DONE
   - Simplified the integrators a bit.                                        -- DONE
@@ -123,6 +136,26 @@ RECENT
   - Parallel io improvements                                  -- DONE
   - Plaquette and link trace checks into nersc reader from the Grid_nersc_io.cc test. -- DONE
 
+
+DONE:
+- MultiArray -- MultiRHS done
+- ConjugateGradientMultiShift -- DONE
+- MCR                         -- DONE
+- Remez -- Mike or Boost?     -- DONE
+- Proto (ET)                  -- DONE
+- uBlas                       -- DONE ; Eigen
+- Potentially Useful Boost libraries -- DONE ; Eigen
+- Aligned allocator; memory pool -- DONE
+- Multiprecision              -- DONE
+- Serialization               -- DONE
+- Regex -- Not needed
+- Tokenize -- Why?
+
+- Random number state save restore -- DONE
+- Rectangle gauge actions. -- DONE
+  Iwasaki,
+  Symanzik,
+  ... etc...
 Done: Cayley, Partial , ContFrac force terms.
 
 DONE
@@ -207,6 +240,7 @@ Done
 FUNCTIONALITY: it pleases me to keep track of things I have done (keeps me arguably sane)
 ======================================================================================================
 
+* Link smearing/boundary conds; Policy class based implementation ; framework more in place -- DONE
 * Command line args for geometry, simd, etc. layout. Is it necessary to have -- DONE
   user pass these? Is this a QCD specific?
 

VERSION

@@ -1,6 +1,5 @@
-Version : 0.6.0
+Version : 0.7.0
 
-- AVX512, AVX2, AVX, SSE good
-- Clang 3.5 and above, ICPC v16 and above, GCC 4.9 and above
-- MPI and MPI3
-- HiRep, Smearing, Generic gauge group
+- Clang 3.5 and above, ICPC v16 and above, GCC 6.3 and above recommended
+- MPI and MPI3 comms optimisations for KNL and OPA finished
+- Half precision comms

benchmarks/Benchmark_comms.cc

@@ -31,6 +31,32 @@ using namespace std;
 using namespace Grid;
 using namespace Grid::QCD;
 
+struct time_statistics{
+  double mean;
+  double err;
+  double min;
+  double max;
+
+  void statistics(std::vector<double> v){
+      double sum = std::accumulate(v.begin(), v.end(), 0.0);
+      mean = sum / v.size();
+
+      std::vector<double> diff(v.size());
+      std::transform(v.begin(), v.end(), diff.begin(), [=](double x) { return x - mean; });
+      double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
+      err = std::sqrt(sq_sum / (v.size()*(v.size() - 1)));
+
+      auto result = std::minmax_element(v.begin(), v.end());
+      min = *result.first;
+      max = *result.second;
+}
+};
+
+void header(){
+  std::cout <<GridLogMessage << " L  "<<"\t"<<" Ls  "<<"\t"
+            <<std::setw(11)<<"bytes"<<"MB/s uni (err/min/max)"<<"\t\t"<<"MB/s bidi (err/min/max)"<<std::endl;
+};
+
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
@@ -40,17 +66,21 @@ int main (int argc, char ** argv)
   int threads = GridThread::GetThreads();
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
 
-  int Nloop=10;
+  int Nloop=100;
   int nmu=0;
+  int maxlat=24;
   for(int mu=0;mu<Nd;mu++) if (mpi_layout[mu]>1) nmu++;
 
+  std::cout << GridLogMessage << "Number of iterations to average: "<< Nloop << std::endl;
+  std::vector<double> t_time(Nloop);
+  time_statistics timestat;
+
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking concurrent halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
-  int maxlat=16;
-  for(int lat=4;lat<=maxlat;lat+=2){
-    for(int Ls=1;Ls<=16;Ls*=2){
+  header();
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=32;Ls*=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],
       				    lat*mpi_layout[1],
@@ -58,6 +88,9 @@ int main (int argc, char ** argv)
       				    lat*mpi_layout[3]});
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      RealD Nrank = Grid._Nprocessors;
+      RealD Nnode = Grid.NodeCount();
+      RealD ppn = Nrank/Nnode;
 
       std::vector<std::vector<HalfSpinColourVectorD> > xbuf(8,std::vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
       std::vector<std::vector<HalfSpinColourVectorD> > rbuf(8,std::vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
@@ -65,8 +98,8 @@ int main (int argc, char ** argv)
       int ncomm;
       int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
 
-      double start=usecond();
       for(int i=0;i<Nloop;i++){
+      double start=usecond();
 
 	std::vector<CartesianCommunicator::CommsRequest_t> requests;
 
@@ -102,18 +135,24 @@ int main (int argc, char ** argv)
 	}
 	Grid.SendToRecvFromComplete(requests);
 	Grid.Barrier();
-
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
       }
-      double stop=usecond();
 
-      double dbytes    = bytes;
-      double xbytes    = Nloop*dbytes*2.0*ncomm;
+      timestat.statistics(t_time);
+
+      double dbytes    = bytes*ppn;
+      double xbytes    = dbytes*2.0*ncomm;
       double rbytes    = xbytes;
       double bidibytes = xbytes+rbytes;
 
-      double time = stop-start; // microseconds
+      std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
+               <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
+               <<std::right<< xbytes/timestat.mean<<"  "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
+               <<xbytes/timestat.max <<" "<< xbytes/timestat.min  
+               << "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< "  " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
+               << bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
 
-      std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
     }
   }    
 
@@ -121,15 +160,17 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
+  header();
 
-
-  for(int lat=4;lat<=maxlat;lat+=2){
-    for(int Ls=1;Ls<=16;Ls*=2){
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=32;Ls*=2){
 
       std::vector<int> latt_size  ({lat,lat,lat,lat});
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      RealD Nrank = Grid._Nprocessors;
+      RealD Nnode = Grid.NodeCount();
+      RealD ppn = Nrank/Nnode;
 
       std::vector<std::vector<HalfSpinColourVectorD> > xbuf(8,std::vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
       std::vector<std::vector<HalfSpinColourVectorD> > rbuf(8,std::vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
@@ -138,8 +179,8 @@ int main (int argc, char ** argv)
       int ncomm;
       int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
 
-      double start=usecond();
       for(int i=0;i<Nloop;i++){
+      double start=usecond();
     
 	ncomm=0;
 	for(int mu=0;mu<4;mu++){
@@ -178,30 +219,37 @@ int main (int argc, char ** argv)
 	  }
 	}
 	Grid.Barrier();
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
+
       }
 
-      double stop=usecond();
+      timestat.statistics(t_time);
       
-      double dbytes    = bytes;
-      double xbytes    = Nloop*dbytes*2.0*ncomm;
+      double dbytes    = bytes*ppn;
+      double xbytes    = dbytes*2.0*ncomm;
       double rbytes    = xbytes;
       double bidibytes = xbytes+rbytes;
 
-      double time = stop-start;
+    std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
+               <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
+               <<std::right<< xbytes/timestat.mean<<"  "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
+               <<xbytes/timestat.max <<" "<< xbytes/timestat.min  
+               << "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< "  " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
+               << bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
+
       
-      std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
     }
   }  
 
 
-  Nloop=100;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking concurrent STENCIL halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
+  header();
 
-  for(int lat=4;lat<=maxlat;lat+=2){
-    for(int Ls=1;Ls<=16;Ls*=2){
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=32;Ls*=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],
       				    lat*mpi_layout[1],
@@ -209,6 +257,9 @@ int main (int argc, char ** argv)
       				    lat*mpi_layout[3]});
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      RealD Nrank = Grid._Nprocessors;
+      RealD Nnode = Grid.NodeCount();
+      RealD ppn = Nrank/Nnode;
 
       std::vector<HalfSpinColourVectorD *> xbuf(8);
       std::vector<HalfSpinColourVectorD *> rbuf(8);
@@ -216,73 +267,86 @@ int main (int argc, char ** argv)
       for(int d=0;d<8;d++){
 	xbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
 	rbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
       }
 
       int ncomm;
       int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
 
-      double start=usecond();
+      double dbytes;
       for(int i=0;i<Nloop;i++){
+	double start=usecond();
+
+	dbytes=0;
+	ncomm=0;
 
 	std::vector<CartesianCommunicator::CommsRequest_t> requests;
 
-	ncomm=0;
 	for(int mu=0;mu<4;mu++){
 	
+
 	  if (mpi_layout[mu]>1 ) {
 	  
 	    ncomm++;
 	    int comm_proc=1;
 	    int xmit_to_rank;
 	    int recv_from_rank;
-	    
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	    Grid.StencilSendToRecvFromBegin(requests,
-					    (void *)&xbuf[mu][0],
-					    xmit_to_rank,
-					    (void *)&rbuf[mu][0],
-					    recv_from_rank,
-					    bytes);
+	    dbytes+=
+	      Grid.StencilSendToRecvFromBegin(requests,
+					      (void *)&xbuf[mu][0],
+					      xmit_to_rank,
+					      (void *)&rbuf[mu][0],
+					      recv_from_rank,
+					      bytes);
 	
 	    comm_proc = mpi_layout[mu]-1;
 	  
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	    Grid.StencilSendToRecvFromBegin(requests,
-					    (void *)&xbuf[mu+4][0],
-					    xmit_to_rank,
-					    (void *)&rbuf[mu+4][0],
-					    recv_from_rank,
-					    bytes);
+	    dbytes+=
+	      Grid.StencilSendToRecvFromBegin(requests,
+					      (void *)&xbuf[mu+4][0],
+					      xmit_to_rank,
+					      (void *)&rbuf[mu+4][0],
+					      recv_from_rank,
+					      bytes);
 	  
 	  }
 	}
 	Grid.StencilSendToRecvFromComplete(requests);
 	Grid.Barrier();
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
 	
       }
-      double stop=usecond();
 
-      double dbytes    = bytes;
-      double xbytes    = Nloop*dbytes*2.0*ncomm;
-      double rbytes    = xbytes;
-      double bidibytes = xbytes+rbytes;
+      timestat.statistics(t_time);
+
+      dbytes=dbytes*ppn;
+      double xbytes    = dbytes*0.5;
+      double rbytes    = dbytes*0.5;
+      double bidibytes = dbytes;
+
+      std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
+               <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
+               <<std::right<< xbytes/timestat.mean<<"  "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
+               <<xbytes/timestat.max <<" "<< xbytes/timestat.min  
+               << "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< "  " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
+               << bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
 
-      double time = stop-start; // microseconds
 
-      std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
     }
   }    
 
 
-
-  Nloop=100;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking sequential STENCIL halo exchange in "<<nmu<<" dimensions"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<" Ls  "<<"\t\t"<<"bytes"<<"\t\t"<<"MB/s uni"<<"\t\t"<<"MB/s bidi"<<std::endl;
+  header();
 
-  for(int lat=4;lat<=maxlat;lat+=2){
-    for(int Ls=1;Ls<=16;Ls*=2){
+  for(int lat=4;lat<=maxlat;lat+=4){
+    for(int Ls=8;Ls<=32;Ls*=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],
       				    lat*mpi_layout[1],
@@ -290,6 +354,9 @@ int main (int argc, char ** argv)
       				    lat*mpi_layout[3]});
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
+      RealD Nrank = Grid._Nprocessors;
+      RealD Nnode = Grid.NodeCount();
+      RealD ppn = Nrank/Nnode;
 
       std::vector<HalfSpinColourVectorD *> xbuf(8);
       std::vector<HalfSpinColourVectorD *> rbuf(8);
@@ -297,16 +364,18 @@ int main (int argc, char ** argv)
       for(int d=0;d<8;d++){
 	xbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
 	rbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
+	bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
       }
 
       int ncomm;
       int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
-
-      double start=usecond();
+      double dbytes;
       for(int i=0;i<Nloop;i++){
+	double start=usecond();
 
 	std::vector<CartesianCommunicator::CommsRequest_t> requests;
-
+	dbytes=0;
 	ncomm=0;
 	for(int mu=0;mu<4;mu++){
 	
@@ -318,42 +387,52 @@ int main (int argc, char ** argv)
 	    int recv_from_rank;
 	    
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	    Grid.StencilSendToRecvFromBegin(requests,
-					    (void *)&xbuf[mu][0],
-					    xmit_to_rank,
-					    (void *)&rbuf[mu][0],
-					    recv_from_rank,
-					    bytes);
-	    //	    Grid.StencilSendToRecvFromComplete(requests);
-	    //	    requests.resize(0);
+	    dbytes+=
+	      Grid.StencilSendToRecvFromBegin(requests,
+					      (void *)&xbuf[mu][0],
+					      xmit_to_rank,
+					      (void *)&rbuf[mu][0],
+					      recv_from_rank,
+					      bytes);
+	    Grid.StencilSendToRecvFromComplete(requests);
+	    requests.resize(0);
 
 	    comm_proc = mpi_layout[mu]-1;
 	  
 	    Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
-	    Grid.StencilSendToRecvFromBegin(requests,
-					    (void *)&xbuf[mu+4][0],
-					    xmit_to_rank,
-					    (void *)&rbuf[mu+4][0],
-					    recv_from_rank,
-					    bytes);
+	    dbytes+=
+	      Grid.StencilSendToRecvFromBegin(requests,
+					      (void *)&xbuf[mu+4][0],
+					      xmit_to_rank,
+					      (void *)&rbuf[mu+4][0],
+					      recv_from_rank,
+					      bytes);
 	    Grid.StencilSendToRecvFromComplete(requests);
 	    requests.resize(0);
 	  
 	  }
 	}
 	Grid.Barrier();
+	double stop=usecond();
+	t_time[i] = stop-start; // microseconds
 	
       }
-      double stop=usecond();
 
-      double dbytes    = bytes;
-      double xbytes    = Nloop*dbytes*2.0*ncomm;
-      double rbytes    = xbytes;
-      double bidibytes = xbytes+rbytes;
+      timestat.statistics(t_time);
 
-      double time = stop-start; // microseconds
+      dbytes=dbytes*ppn;
+      double xbytes    = dbytes*0.5;
+      double rbytes    = dbytes*0.5;
+      double bidibytes = dbytes;
+
+
+      std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
+               <<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
+               <<std::right<< xbytes/timestat.mean<<"  "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
+               <<xbytes/timestat.max <<" "<< xbytes/timestat.min  
+               << "\t\t"<<std::setw(7)<< bidibytes/timestat.mean<< "  " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
+               << bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
  
-      std::cout<<GridLogMessage << lat<<"\t\t"<<Ls<<"\t\t"<<bytes<<"\t\t"<<xbytes/time<<"\t\t"<<bidibytes/time<<std::endl;
     }
   }    
 

benchmarks/Benchmark_dwf.cc

@@ -1,28 +1,22 @@
-    /*************************************************************************************
-
+ /*************************************************************************************
     Grid physics library, www.github.com/paboyle/Grid 
-
     Source file: ./benchmarks/Benchmark_dwf.cc
-
     Copyright (C) 2015
 
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: paboyle <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 */
@@ -48,16 +42,16 @@ typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
 typedef WilsonFermion5D<DomainWallVec5dImplF> WilsonFermion5DF;
 typedef WilsonFermion5D<DomainWallVec5dImplD> WilsonFermion5DD;
 
-
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
 
+
   int threads = GridThread::GetThreads();
   std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
 
   std::vector<int> latt4 = GridDefaultLatt();
-  const int Ls=8;
+  const int Ls=16;
   GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
   GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
   GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
@@ -72,34 +66,65 @@ int main (int argc, char ** argv)
   std::vector<int> seeds4({1,2,3,4});
   std::vector<int> seeds5({5,6,7,8});
   
+  std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
   GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+  std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
   GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
 
   LatticeFermion src   (FGrid); random(RNG5,src);
+#if 0
+  src = zero;
+  {
+    std::vector<int> origin({0,0,0,latt4[2]-1,0});
+    SpinColourVectorF tmp;
+    tmp=zero;
+    tmp()(0)(0)=Complex(-2.0,0.0);
+    std::cout << " source site 0 " << tmp<<std::endl;
+    pokeSite(tmp,src,origin);
+  }
+#else
+  RealD N2 = 1.0/::sqrt(norm2(src));
+  src = src*N2;
+#endif
+
+
   LatticeFermion result(FGrid); result=zero;
   LatticeFermion    ref(FGrid);    ref=zero;
   LatticeFermion    tmp(FGrid);
   LatticeFermion    err(FGrid);
 
+  std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
   LatticeGaugeField Umu(UGrid); 
-  random(RNG4,Umu);
-
-  LatticeGaugeField Umu5d(FGrid); 
+  SU3::HotConfiguration(RNG4,Umu); 
+  std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
+#if 0
+  Umu=1.0;
+  for(int mu=0;mu<Nd;mu++){
+    LatticeColourMatrix ttmp(UGrid);
+    ttmp = PeekIndex<LorentzIndex>(Umu,mu);
+    //    if (mu !=2 ) ttmp = 0;
+    //    ttmp = ttmp* pow(10.0,mu);
+    PokeIndex<LorentzIndex>(Umu,ttmp,mu);
+  }
+  std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
+#endif
 
+  ////////////////////////////////////
+  // Naive wilson implementation
+  ////////////////////////////////////
   // replicate across fifth dimension
+  LatticeGaugeField Umu5d(FGrid); 
+  std::vector<LatticeColourMatrix> U(4,FGrid);
   for(int ss=0;ss<Umu._grid->oSites();ss++){
     for(int s=0;s<Ls;s++){
       Umu5d._odata[Ls*ss+s] = Umu._odata[ss];
     }
   }
-
-  ////////////////////////////////////
-  // Naive wilson implementation
-  ////////////////////////////////////
-  std::vector<LatticeColourMatrix> U(4,FGrid);
   for(int mu=0;mu<Nd;mu++){
     U[mu] = PeekIndex<LorentzIndex>(Umu5d,mu);
   }
+  std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
 
   if (1)
   {
@@ -120,8 +145,7 @@ int main (int argc, char ** argv)
   RealD M5  =1.8;
 
   RealD NP = UGrid->_Nprocessors;
-
-  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  RealD NN = UGrid->NodeCount();
 
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
   std::cout << GridLogMessage<< "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" <<std::endl;
@@ -131,15 +155,22 @@ int main (int argc, char ** argv)
   std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplex::Nsimd()<<std::endl;
   if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
   if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
+#ifdef GRID_OMP
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+#endif
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
-  int ncall =100;
+  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  int ncall =1000;
   if (1) {
     FGrid->Barrier();
     Dw.ZeroCounters();
+    Dw.Dhop(src,result,0);
+    std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
     double t0=usecond();
     for(int i=0;i<ncall;i++){
       __SSC_START;
@@ -153,16 +184,55 @@ int main (int argc, char ** argv)
     double flops=1344*volume*ncall;
 
     std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
-    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
+    //    std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
+    //    std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
     err = ref-result; 
     std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+    /*
+    if(( norm2(err)>1.0e-4) ) { 
+      std::cout << "RESULT\n " << result<<std::endl;
+      std::cout << "REF   \n " << ref   <<std::endl;
+      std::cout << "ERR   \n " << err   <<std::endl;
+      FGrid->Barrier();
+      exit(-1);
+    }
+    */
     assert (norm2(err)< 1.0e-4 );
     Dw.Report();
   }
 
+  DomainWallFermionRL DwH(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  if (1) {
+    FGrid->Barrier();
+    DwH.ZeroCounters();
+    DwH.Dhop(src,result,0);
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      __SSC_START;
+      DwH.Dhop(src,result,0);
+      __SSC_STOP;
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+    
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=1344*volume*ncall;
+
+    std::cout<<GridLogMessage << "Called half prec comms Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
+    err = ref-result; 
+    std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+    assert (norm2(err)< 1.0e-3 );
+    DwH.Report();
+  }
+
   if (1)
   {
 
@@ -171,6 +241,10 @@ int main (int argc, char ** argv)
     std::cout << GridLogMessage<< "* Vectorising fifth dimension by "<<vComplex::Nsimd()<<std::endl;
     if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
     if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
+#ifdef GRID_OMP
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+#endif
     if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
     if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
     if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
@@ -183,20 +257,12 @@ int main (int argc, char ** argv)
 
     WilsonFermion5DR sDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,M5);
 
-    for(int x=0;x<latt4[0];x++){
-    for(int y=0;y<latt4[1];y++){
-    for(int z=0;z<latt4[2];z++){
-    for(int t=0;t<latt4[3];t++){
-    for(int s=0;s<Ls;s++){
-      std::vector<int> site({s,x,y,z,t});
-      SpinColourVector tmp;
-      peekSite(tmp,src,site);
-      pokeSite(tmp,ssrc,site);
-    }}}}}
+    localConvert(src,ssrc);
     std::cout<<GridLogMessage<< "src norms "<< norm2(src)<<" " <<norm2(ssrc)<<std::endl;
     FGrid->Barrier();
-    double t0=usecond();
+    sDw.Dhop(ssrc,sresult,0);
     sDw.ZeroCounters();
+    double t0=usecond();
     for(int i=0;i<ncall;i++){
       __SSC_START;
       sDw.Dhop(ssrc,sresult,0);
@@ -210,46 +276,52 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << "Called Dw s_inner "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
     std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per node =  "<< flops/(t1-t0)/NN<<std::endl;
+    //    std::cout<<GridLogMessage<< "res norms "<< norm2(result)<<" " <<norm2(sresult)<<std::endl;
     sDw.Report();
-  
-    if(0){
-      for(int i=0;i< PerformanceCounter::NumTypes(); i++ ){
-	sDw.Dhop(ssrc,sresult,0);
-	PerformanceCounter Counter(i);
-	Counter.Start();
-	sDw.Dhop(ssrc,sresult,0);
-	Counter.Stop();
-	Counter.Report();
-      }
-    }
-
-    std::cout<<GridLogMessage<< "res norms "<< norm2(result)<<" " <<norm2(sresult)<<std::endl;
-
     RealD sum=0;
-    for(int x=0;x<latt4[0];x++){
-    for(int y=0;y<latt4[1];y++){
-    for(int z=0;z<latt4[2];z++){
-    for(int t=0;t<latt4[3];t++){
-    for(int s=0;s<Ls;s++){
-      std::vector<int> site({s,x,y,z,t});
-      SpinColourVector normal, simd;
-      peekSite(normal,result,site);
-      peekSite(simd,sresult,site);
-      sum=sum+norm2(normal-simd);
-      if (norm2(normal-simd) > 1.0e-6 ) {
-	std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" "<<norm2(normal-simd)<<std::endl;
-	std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" normal "<<normal<<std::endl;
-	std::cout << "site "<<x<<","<<y<<","<<z<<","<<t<<","<<s<<" simd   "<<simd<<std::endl;
-      }
-    }}}}}
-    std::cout<<GridLogMessage<<" difference between normal and simd is "<<sum<<std::endl;
-    assert (sum< 1.0e-4 );
 
+    err=zero;
+    localConvert(sresult,err);
+    err = err - ref;
+    sum = norm2(err);
+    std::cout<<GridLogMessage<<" difference between normal ref and simd is "<<sum<<std::endl;
+    if(sum > 1.0e-4 ){
+      std::cout<< "sD REF\n " <<ref << std::endl;
+      std::cout<< "sD ERR   \n " <<err  <<std::endl;
+    }
+    //    assert(sum < 1.0e-4);
 
-    if (1) {
+    err=zero;
+    localConvert(sresult,err);
+    err = err - result;
+    sum = norm2(err);
+    std::cout<<GridLogMessage<<" difference between normal result and simd is "<<sum<<std::endl;
+    if(sum > 1.0e-4 ){
+      std::cout<< "sD REF\n " <<result << std::endl;
+      std::cout<< "sD ERR   \n " << err  <<std::endl;
+    }
+    assert(sum < 1.0e-4);
+    
+    if(1){
+      std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
+      std::cout << GridLogMessage<< "* Benchmarking WilsonFermion5D<DomainWallVec5dImplR>::DhopEO "<<std::endl;
+      std::cout << GridLogMessage<< "* Vectorising fifth dimension by "<<vComplex::Nsimd()<<std::endl;
+      if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
+      if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
+#ifdef GRID_OMP
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+#endif
+      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) 
+	std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
+      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) 
+	std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
+      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) 
+	std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
+      std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
 
       LatticeFermion sr_eo(sFGrid);
-
       LatticeFermion ssrc_e (sFrbGrid);
       LatticeFermion ssrc_o (sFrbGrid);
       LatticeFermion sr_e   (sFrbGrid);
@@ -257,39 +329,30 @@ int main (int argc, char ** argv)
 
       pickCheckerboard(Even,ssrc_e,ssrc);
       pickCheckerboard(Odd,ssrc_o,ssrc);
-
-      setCheckerboard(sr_eo,ssrc_o);
-      setCheckerboard(sr_eo,ssrc_e);
+      //      setCheckerboard(sr_eo,ssrc_o);
+      //      setCheckerboard(sr_eo,ssrc_e);
 
       sr_e = zero;
       sr_o = zero;
 
-      std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
-      std::cout << GridLogMessage<< "* Benchmarking WilsonFermion5D<DomainWallVec5dImplR>::DhopEO "<<std::endl;
-      std::cout << GridLogMessage<< "* Vectorising fifth dimension by "<<vComplex::Nsimd()<<std::endl;
-      if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
-      if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
-      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
-      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
-      if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
-      std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
-
       FGrid->Barrier();
+      sDw.DhopEO(ssrc_o, sr_e, DaggerNo);
       sDw.ZeroCounters();
-      sDw.stat.init("DhopEO");
+      //      sDw.stat.init("DhopEO");
       double t0=usecond();
       for (int i = 0; i < ncall; i++) {
         sDw.DhopEO(ssrc_o, sr_e, DaggerNo);
       }
       double t1=usecond();
       FGrid->Barrier();
-      sDw.stat.print();
+      //      sDw.stat.print();
 
       double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
       double flops=(1344.0*volume*ncall)/2;
 
       std::cout<<GridLogMessage << "sDeo mflop/s =   "<< flops/(t1-t0)<<std::endl;
       std::cout<<GridLogMessage << "sDeo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
+      std::cout<<GridLogMessage << "sDeo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
       sDw.Report();
 
       sDw.DhopEO(ssrc_o,sr_e,DaggerNo);
@@ -298,22 +361,26 @@ int main (int argc, char ** argv)
 
       pickCheckerboard(Even,ssrc_e,sresult);
       pickCheckerboard(Odd ,ssrc_o,sresult);
+
       ssrc_e = ssrc_e - sr_e;
       RealD error = norm2(ssrc_e);
-
       std::cout<<GridLogMessage << "sE norm diff   "<< norm2(ssrc_e)<< "  vec nrm"<<norm2(sr_e) <<std::endl;
-      ssrc_o = ssrc_o - sr_o;
 
+      ssrc_o = ssrc_o - sr_o;
       error+= norm2(ssrc_o);
       std::cout<<GridLogMessage << "sO norm diff   "<< norm2(ssrc_o)<< "  vec nrm"<<norm2(sr_o) <<std::endl;
-      if(error>1.0e-4) { 
+
+      if(( error>1.0e-4) ) { 
 	setCheckerboard(ssrc,ssrc_o);
 	setCheckerboard(ssrc,ssrc_e);
-	std::cout<< ssrc << std::endl;
+	std::cout<< "DIFF\n " <<ssrc << std::endl;
+	setCheckerboard(ssrc,sr_o);
+	setCheckerboard(ssrc,sr_e);
+	std::cout<< "CBRESULT\n " <<ssrc << std::endl;
+	std::cout<< "RESULT\n " <<sresult<< std::endl;
       }
+      assert(error<1.0e-4);
     }
-
-
   }
 
   if (1)
@@ -324,25 +391,30 @@ int main (int argc, char ** argv)
       //    ref =  src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
       tmp = U[mu]*Cshift(src,mu+1,1);
       for(int i=0;i<ref._odata.size();i++){
-  ref._odata[i]+= tmp._odata[i] + Gamma(Gmu[mu])*tmp._odata[i]; ;
+	ref._odata[i]+= tmp._odata[i] + Gamma(Gmu[mu])*tmp._odata[i]; ;
       }
 
       tmp =adj(U[mu])*src;
       tmp =Cshift(tmp,mu+1,-1);
       for(int i=0;i<ref._odata.size();i++){
-  ref._odata[i]+= tmp._odata[i] - Gamma(Gmu[mu])*tmp._odata[i]; ;
+	ref._odata[i]+= tmp._odata[i] - Gamma(Gmu[mu])*tmp._odata[i]; ;
       }
     }
     ref = -0.5*ref;
   }
+  //  dump=1;
   Dw.Dhop(src,result,1);
   std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
   std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
-  std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
-  std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
+  std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
+  std::cout<<GridLogMessage << "norm dag ref    "<< norm2(ref)<<std::endl;
   err = ref-result; 
-  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-  assert(norm2(err)<1.0e-4);
+  std::cout<<GridLogMessage << "norm dag diff   "<< norm2(err)<<std::endl;
+  if((norm2(err)>1.0e-4)){
+	std::cout<< "DAG RESULT\n "  <<ref     << std::endl;
+	std::cout<< "DAG sRESULT\n " <<result  << std::endl;
+	std::cout<< "DAG ERR   \n "  << err    <<std::endl;
+  }
   LatticeFermion src_e (FrbGrid);
   LatticeFermion src_o (FrbGrid);
   LatticeFermion r_e   (FrbGrid);
@@ -350,18 +422,24 @@ int main (int argc, char ** argv)
   LatticeFermion r_eo  (FGrid);
 
 
-  std::cout<<GridLogMessage << "Calling Deo and Doe and assert Deo+Doe == Dunprec"<<std::endl;
+  std::cout<<GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec"<<std::endl;
   pickCheckerboard(Even,src_e,src);
   pickCheckerboard(Odd,src_o,src);
 
   std::cout<<GridLogMessage << "src_e"<<norm2(src_e)<<std::endl;
   std::cout<<GridLogMessage << "src_o"<<norm2(src_o)<<std::endl;
 
+
+  // S-direction is INNERMOST and takes no part in the parity.
   std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
   std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionR::DhopEO                "<<std::endl;
   std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplex::Nsimd()<<std::endl;
   if ( sizeof(Real)==4 )   std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
   if ( sizeof(Real)==8 )   std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
+#ifdef GRID_OMP
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
+  if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
+#endif
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric   ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3       WilsonKernels" <<std::endl;
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
@@ -369,6 +447,7 @@ int main (int argc, char ** argv)
   {
     Dw.ZeroCounters();
     FGrid->Barrier();
+    Dw.DhopEO(src_o,r_e,DaggerNo);
     double t0=usecond();
     for(int i=0;i<ncall;i++){
       Dw.DhopEO(src_o,r_e,DaggerNo);
@@ -381,6 +460,7 @@ int main (int argc, char ** argv)
 
     std::cout<<GridLogMessage << "Deo mflop/s =   "<< flops/(t1-t0)<<std::endl;
     std::cout<<GridLogMessage << "Deo mflop/s per rank   "<< flops/(t1-t0)/NP<<std::endl;
+    std::cout<<GridLogMessage << "Deo mflop/s per node   "<< flops/(t1-t0)/NN<<std::endl;
     Dw.Report();
   }
   Dw.DhopEO(src_o,r_e,DaggerNo);
@@ -396,14 +476,20 @@ int main (int argc, char ** argv)
 
   err = r_eo-result; 
   std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
-  assert(norm2(err)<1.0e-4);
+  if((norm2(err)>1.0e-4)){
+	std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
+	std::cout<< "Deo REF\n " <<result  << std::endl;
+	std::cout<< "Deo ERR   \n " << err <<std::endl;
+  }
 
   pickCheckerboard(Even,src_e,err);
   pickCheckerboard(Odd,src_o,err);
   std::cout<<GridLogMessage << "norm diff even  "<< norm2(src_e)<<std::endl;
   std::cout<<GridLogMessage << "norm diff odd   "<< norm2(src_o)<<std::endl;
-  assert(norm2(src_e)<1.0e-4);
-  assert(norm2(src_o)<1.0e-4);
+
+  //assert(norm2(src_e)<1.0e-4);
+  //assert(norm2(src_o)<1.0e-4);
 
   Grid_finalize();
 }
+

benchmarks/Benchmark_memory_asynch.cc

@@ -66,7 +66,8 @@ int main (int argc, char ** argv)
 
     Vec tsum; tsum = zero;
 
-    GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+    GridParallelRNG          pRNG(&Grid);      
+    pRNG.SeedFixedIntegers(std::vector<int>({56,17,89,101}));
 
     std::vector<double> stop(threads);
     Vector<Vec> sum(threads);
@@ -77,8 +78,7 @@ int main (int argc, char ** argv)
     }
 
     double start=usecond();
-PARALLEL_FOR_LOOP
-    for(int t=0;t<threads;t++){
+    parallel_for(int t=0;t<threads;t++){
 
       sum[t] = x[t]._odata[0];
       for(int i=0;i<Nloop;i++){

benchmarks/Benchmark_memory_bandwidth.cc

@@ -55,8 +55,8 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
-  uint64_t lmax=44;
-#define NLOOP (1*lmax*lmax*lmax*lmax/vol)
+  uint64_t lmax=64;
+#define NLOOP (100*lmax*lmax*lmax*lmax/vol)
   for(int lat=4;lat<=lmax;lat+=4){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
@@ -65,7 +65,7 @@ int main (int argc, char ** argv)
 
       uint64_t Nloop=NLOOP;
 
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
 
       LatticeVec z(&Grid); //random(pRNG,z);
       LatticeVec x(&Grid); //random(pRNG,x);
@@ -100,7 +100,7 @@ int main (int argc, char ** argv)
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
 
       LatticeVec z(&Grid); //random(pRNG,z);
       LatticeVec x(&Grid); //random(pRNG,x);
@@ -138,7 +138,7 @@ int main (int argc, char ** argv)
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
 
       LatticeVec z(&Grid); //random(pRNG,z);
       LatticeVec x(&Grid); //random(pRNG,x);
@@ -173,7 +173,7 @@ int main (int argc, char ** argv)
       uint64_t Nloop=NLOOP;
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
       LatticeVec z(&Grid); //random(pRNG,z);
       LatticeVec x(&Grid); //random(pRNG,x);
       LatticeVec y(&Grid); //random(pRNG,y);

benchmarks/Benchmark_staggered.cc

@@ -0,0 +1,134 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./benchmarks/Benchmark_staggered.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: paboyle <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 */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(latt_size,simd_layout,mpi_layout);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REALF"<< sizeof(RealF)<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REALD"<< sizeof(RealD)<<std::endl;
+  std::cout<<GridLogMessage << "Grid floating point word size is REAL"<< sizeof(Real)<<std::endl;
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);
+  pRNG.SeedFixedIntegers(seeds);
+  //  pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
+
+  typedef typename ImprovedStaggeredFermionR::FermionField FermionField; 
+  typename ImprovedStaggeredFermionR::ImplParams params; 
+
+  FermionField src   (&Grid); random(pRNG,src);
+  FermionField result(&Grid); result=zero;
+  FermionField    ref(&Grid);    ref=zero;
+  FermionField    tmp(&Grid);    tmp=zero;
+  FermionField    err(&Grid);    tmp=zero;
+  LatticeGaugeField Umu(&Grid); random(pRNG,Umu);
+  std::vector<LatticeColourMatrix> U(4,&Grid);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }  
+
+  // Only one non-zero (y)
+#if 0
+  Umu=zero;
+  Complex cone(1.0,0.0);
+  for(int nn=0;nn<Nd;nn++){
+    random(pRNG,U[nn]);
+    if(1) {
+      if (nn!=2) { U[nn]=zero; std::cout<<GridLogMessage << "zeroing gauge field in dir "<<nn<<std::endl; }
+      //      else       { U[nn]= cone;std::cout<<GridLogMessage << "unit gauge field in dir "<<nn<<std::endl; }
+      else       { std::cout<<GridLogMessage << "random gauge field in dir "<<nn<<std::endl; }
+    }
+    PokeIndex<LorentzIndex>(Umu,U[nn],nn);
+  }
+#endif
+
+  for(int mu=0;mu<Nd;mu++){
+    U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
+  }
+  ref = zero;
+  /*  
+  { // Naive wilson implementation
+    ref = zero;
+    for(int mu=0;mu<Nd;mu++){
+      //    ref =  src + Gamma(Gamma::GammaX)* src ; // 1-gamma_x
+      tmp = U[mu]*Cshift(src,mu,1);
+      for(int i=0;i<ref._odata.size();i++){
+	ref._odata[i]+= tmp._odata[i] - Gamma(Gmu[mu])*tmp._odata[i]; ;
+      }
+
+      tmp =adj(U[mu])*src;
+      tmp =Cshift(tmp,mu,-1);
+      for(int i=0;i<ref._odata.size();i++){
+	ref._odata[i]+= tmp._odata[i] + Gamma(Gmu[mu])*tmp._odata[i]; ;
+      }
+    }
+  }
+  ref = -0.5*ref;
+  */
+
+  RealD mass=0.1;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0,params);
+  
+  std::cout<<GridLogMessage << "Calling Ds"<<std::endl;
+  int ncall=1000;
+  double t0=usecond();
+  for(int i=0;i<ncall;i++){
+    Ds.Dhop(src,result,0);
+  }
+  double t1=usecond();
+  double flops=(16*(3*(6+8+8)) + 15*3*2)*volume*ncall; // == 66*16 +  == 1146
+  
+  std::cout<<GridLogMessage << "Called Ds"<<std::endl;
+  std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
+  std::cout<<GridLogMessage << "norm ref    "<< norm2(ref)<<std::endl;
+  std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+  err = ref-result; 
+  std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+  Grid_finalize();
+}

benchmarks/Benchmark_su3.cc

@@ -35,8 +35,9 @@ using namespace Grid::QCD;
 int main (int argc, char ** argv)
 {
   Grid_init(&argc,&argv);
+#define LMAX (64)
 
-  int Nloop=1000;
+  int Nloop=20;
 
   std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
   std::vector<int> mpi_layout  = GridDefaultMpi();
@@ -50,12 +51,12 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=32;lat+=2){
+  for(int lat=2;lat<=LMAX;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
 
       LatticeColourMatrix z(&Grid);// random(pRNG,z);
       LatticeColourMatrix x(&Grid);// random(pRNG,x);
@@ -82,13 +83,13 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=32;lat+=2){
+  for(int lat=2;lat<=LMAX;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
 
       LatticeColourMatrix z(&Grid); //random(pRNG,z);
       LatticeColourMatrix x(&Grid); //random(pRNG,x);
@@ -113,13 +114,13 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=32;lat+=2){
+  for(int lat=2;lat<=LMAX;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
 
       LatticeColourMatrix z(&Grid); //random(pRNG,z);
       LatticeColourMatrix x(&Grid); //random(pRNG,x);
@@ -144,13 +145,13 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=32;lat+=2){
+  for(int lat=2;lat<=LMAX;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
 
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
-      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedRandomDevice();
+      //      GridParallelRNG          pRNG(&Grid);      pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
 
       LatticeColourMatrix z(&Grid); //random(pRNG,z);
       LatticeColourMatrix x(&Grid); //random(pRNG,x);

benchmarks/Benchmark_wilson.cc

@@ -69,7 +69,7 @@ int main (int argc, char ** argv)
   std::vector<int> seeds({1,2,3,4});
   GridParallelRNG          pRNG(&Grid);
   pRNG.SeedFixedIntegers(seeds);
-  //  pRNG.SeedRandomDevice();
+  //  pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9});
 
   LatticeFermion src   (&Grid); random(pRNG,src);
   LatticeFermion result(&Grid); result=zero;

benchmarks/Makefile.am

@@ -1,11 +1,7 @@
 include Make.inc
 
-simple: simple_su3_test.o simple_su3_expr.o simple_simd_test.o
-
-EXTRA_LIBRARIES = libsimple_su3_test.a libsimple_su3_expr.a libsimple_simd_test.a
-
-libsimple_su3_test_a_SOURCES = simple_su3_test.cc
-
-libsimple_su3_expr_a_SOURCES = simple_su3_expr.cc
-
-libsimple_simd_test_a_SOURCES = simple_simd_test.cc
+bench-local: all
+	./Benchmark_su3
+	./Benchmark_memory_bandwidth
+	./Benchmark_wilson
+	./Benchmark_dwf --dslash-unroll

bootstrap.sh

@@ -1,6 +1,6 @@
-#!/usr/bin/env bash
+]#!/usr/bin/env bash
 
-EIGEN_URL='http://bitbucket.org/eigen/eigen/get/3.2.9.tar.bz2'
+EIGEN_URL='http://bitbucket.org/eigen/eigen/get/3.3.3.tar.bz2'
 
 echo "-- deploying Eigen source..."
 wget ${EIGEN_URL} --no-check-certificate

configure.ac

@@ -1,16 +1,19 @@
 AC_PREREQ([2.63])
-AC_INIT([Grid], [0.6.0], [https://github.com/paboyle/Grid], [Grid])
+AC_INIT([Grid], [0.7.0], [https://github.com/paboyle/Grid], [Grid])
 AC_CANONICAL_BUILD
 AC_CANONICAL_HOST
 AC_CANONICAL_TARGET
-AM_INIT_AUTOMAKE(subdir-objects)
+AM_INIT_AUTOMAKE([subdir-objects 1.13])
+AM_EXTRA_RECURSIVE_TARGETS([tests bench])
 AC_CONFIG_MACRO_DIR([m4])
 AC_CONFIG_SRCDIR([lib/Grid.h])
 AC_CONFIG_HEADERS([lib/Config.h],[sed -i 's|PACKAGE_|GRID_|' lib/Config.h])
 m4_ifdef([AM_SILENT_RULES], [AM_SILENT_RULES([yes])])
 
+################ Get git info
+#AC_REVISION([m4_esyscmd_s([./scripts/configure.commit])])
+
 ############### Checks for programs
-CXXFLAGS="-O3 $CXXFLAGS"
 AC_PROG_CXX
 AC_PROG_RANLIB
 
@@ -24,6 +27,9 @@ AX_GXX_VERSION
 AC_DEFINE_UNQUOTED([GXX_VERSION],["$GXX_VERSION"],
       [version of g++ that will compile the code])
 
+CXXFLAGS="-g $CXXFLAGS"
+
+
 ############### Checks for typedefs, structures, and compiler characteristics
 AC_TYPE_SIZE_T
 AC_TYPE_UINT32_T
@@ -67,6 +73,13 @@ AC_ARG_WITH([fftw],
             [AM_CXXFLAGS="-I$with_fftw/include $AM_CXXFLAGS"]
             [AM_LDFLAGS="-L$with_fftw/lib $AM_LDFLAGS"])
 
+############### LIME
+AC_ARG_WITH([lime],
+            [AS_HELP_STRING([--with-lime=prefix],
+            [try this for a non-standard install prefix of the LIME library])],
+            [AM_CXXFLAGS="-I$with_lime/include $AM_CXXFLAGS"]
+            [AM_LDFLAGS="-L$with_lime/lib $AM_LDFLAGS"])
+
 ############### lapack
 AC_ARG_ENABLE([lapack],
     [AC_HELP_STRING([--enable-lapack=yes|no|prefix], [enable LAPACK])],
@@ -83,6 +96,18 @@ case ${ac_LAPACK} in
         AC_DEFINE([USE_LAPACK],[1],[use LAPACK]);;
 esac
 
+############### FP16 conversions
+AC_ARG_ENABLE([sfw-fp16],
+    [AC_HELP_STRING([--enable-sfw-fp16=yes|no], [enable software fp16 comms])],
+    [ac_SFW_FP16=${enable_sfw_fp16}], [ac_SFW_FP16=yes])
+case ${ac_SFW_FP16} in
+    yes)
+      AC_DEFINE([SFW_FP16],[1],[software conversion to fp16]);;
+    no);;
+    *)
+      AC_MSG_ERROR(["SFW FP16 option not supported ${ac_SFW_FP16}"]);;
+esac
+
 ############### MKL
 AC_ARG_ENABLE([mkl],
     [AC_HELP_STRING([--enable-mkl=yes|no|prefix], [enable Intel MKL for LAPACK & FFTW])],
@@ -152,6 +177,18 @@ AC_SEARCH_LIBS([fftw_execute], [fftw3],
                [AC_DEFINE([HAVE_FFTW], [1], [Define to 1 if you have the `FFTW' library])]
                [have_fftw=true])
 
+AC_SEARCH_LIBS([limeCreateReader], [lime],
+               [AC_DEFINE([HAVE_LIME], [1], [Define to 1 if you have the `LIME' library])]
+               [have_lime=true],
+	       [AC_MSG_WARN(C-LIME library was not found in your system.
+In order to use ILGG file format please install or provide the correct path to your installation
+Info at: http://usqcd.jlab.org/usqcd-docs/c-lime/)])
+
+AC_SEARCH_LIBS([crc32], [z],
+               [AC_DEFINE([HAVE_ZLIB], [1], [Define to 1 if you have the `LIBZ' library])]
+               [have_zlib=true],
+	       [AC_MSG_ERROR(zlib library was not found in your system.)])
+
 AC_SEARCH_LIBS([H5Fopen], [hdf5_cpp],
                [AC_DEFINE([HAVE_HDF5], [1], [Define to 1 if you have the `HDF5' library])]
                [have_hdf5=true]
@@ -176,19 +213,26 @@ case ${ax_cv_cxx_compiler_vendor} in
     case ${ac_SIMD} in
       SSE4)
         AC_DEFINE([SSE4],[1],[SSE4 intrinsics])
-        SIMD_FLAGS='-msse4.2';;
+	case ${ac_SFW_FP16} in
+	  yes)
+	  SIMD_FLAGS='-msse4.2';;
+	  no)
+	  SIMD_FLAGS='-msse4.2 -mf16c';;
+	  *)
+          AC_MSG_ERROR(["SFW_FP16 must be either yes or no value ${ac_SFW_FP16} "]);;
+	esac;;
       AVX)
         AC_DEFINE([AVX1],[1],[AVX intrinsics])
-        SIMD_FLAGS='-mavx';;
+        SIMD_FLAGS='-mavx -mf16c';;
       AVXFMA4)
         AC_DEFINE([AVXFMA4],[1],[AVX intrinsics with FMA4])
-        SIMD_FLAGS='-mavx -mfma4';;
+        SIMD_FLAGS='-mavx -mfma4 -mf16c';;
       AVXFMA)
         AC_DEFINE([AVXFMA],[1],[AVX intrinsics with FMA3])
-        SIMD_FLAGS='-mavx -mfma';;
+        SIMD_FLAGS='-mavx -mfma -mf16c';;
       AVX2)
         AC_DEFINE([AVX2],[1],[AVX2 intrinsics])
-        SIMD_FLAGS='-mavx2 -mfma';;
+        SIMD_FLAGS='-mavx2 -mfma -mf16c';;
       AVX512)
         AC_DEFINE([AVX512],[1],[AVX512 intrinsics])
         SIMD_FLAGS='-mavx512f -mavx512pf -mavx512er -mavx512cd';;
@@ -321,7 +365,7 @@ AM_CONDITIONAL(BUILD_COMMS_NONE,  [ test "${comms_type}X" == "noneX" ])
 ############### RNG selection
 AC_ARG_ENABLE([rng],[AC_HELP_STRING([--enable-rng=ranlux48|mt19937|sitmo],\
 	            [Select Random Number Generator to be used])],\
-	            [ac_RNG=${enable_rng}],[ac_RNG=ranlux48])
+	            [ac_RNG=${enable_rng}],[ac_RNG=sitmo])
 
 case ${ac_RNG} in
      ranlux48)
@@ -384,32 +428,31 @@ DX_INIT_DOXYGEN([$PACKAGE_NAME], [doxygen.cfg])
 
 ############### Ouput
 cwd=`pwd -P`; cd ${srcdir}; abs_srcdir=`pwd -P`; cd ${cwd}
+GRID_CXXFLAGS="$AM_CXXFLAGS $CXXFLAGS"
+GRID_LDFLAGS="$AM_LDFLAGS $LDFLAGS"
+GRID_LIBS=$LIBS
+GRID_SHORT_SHA=`git rev-parse --short HEAD`
+GRID_SHA=`git rev-parse HEAD`
+GRID_BRANCH=`git rev-parse --abbrev-ref HEAD`
 AM_CXXFLAGS="-I${abs_srcdir}/include $AM_CXXFLAGS"
 AM_CFLAGS="-I${abs_srcdir}/include $AM_CFLAGS"
 AM_LDFLAGS="-L${cwd}/lib $AM_LDFLAGS"
 AC_SUBST([AM_CFLAGS])
 AC_SUBST([AM_CXXFLAGS])
 AC_SUBST([AM_LDFLAGS])
-AC_CONFIG_FILES(Makefile)
-AC_CONFIG_FILES(lib/Makefile)
-AC_CONFIG_FILES(tests/Makefile)
-AC_CONFIG_FILES(tests/IO/Makefile)
-AC_CONFIG_FILES(tests/core/Makefile)
-AC_CONFIG_FILES(tests/debug/Makefile)
-AC_CONFIG_FILES(tests/forces/Makefile)
-AC_CONFIG_FILES(tests/hadrons/Makefile)
-AC_CONFIG_FILES(tests/hmc/Makefile)
-AC_CONFIG_FILES(tests/solver/Makefile)
-AC_CONFIG_FILES(tests/qdpxx/Makefile)
-AC_CONFIG_FILES(benchmarks/Makefile)
-AC_CONFIG_FILES(extras/Makefile)
-AC_CONFIG_FILES(extras/Hadrons/Makefile)
-AC_OUTPUT
+AC_SUBST([GRID_CXXFLAGS])
+AC_SUBST([GRID_LDFLAGS])
+AC_SUBST([GRID_LIBS])
+AC_SUBST([GRID_SHA])
+AC_SUBST([GRID_BRANCH])
+
+git_commit=`cd $srcdir && ./scripts/configure.commit`
 
 echo "~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Summary of configuration for $PACKAGE v$VERSION
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
+----- GIT VERSION -------------------------------------
+$git_commit
 ----- PLATFORM ----------------------------------------
 architecture (build)        : $build_cpu
 os (build)                  : $build_os
@@ -422,10 +465,12 @@ SIMD                        : ${ac_SIMD}${SIMD_GEN_WIDTH_MSG}
 Threading                   : ${ac_openmp}
 Communications type         : ${comms_type}
 Default precision           : ${ac_PRECISION}
+Software FP16 conversion    : ${ac_SFW_FP16}
 RNG choice                  : ${ac_RNG}
 GMP                         : `if test "x$have_gmp" = xtrue; then echo yes; else echo no; fi`
 LAPACK                      : ${ac_LAPACK}
 FFTW                        : `if test "x$have_fftw" = xtrue; then echo yes; else echo no; fi`
+LIME (ILDG support)         : `if test "x$have_lime" = xtrue; then echo yes; else echo no; fi`
 HDF5                        : `if test "x$have_hdf5" = xtrue; then echo yes; else echo no; fi`
 build DOXYGEN documentation : `if test "$DX_FLAG_doc" = '1'; then echo yes; else echo no; fi`
 ----- BUILD FLAGS -------------------------------------
@@ -435,7 +480,32 @@ LDFLAGS:
 `echo ${AM_LDFLAGS} ${LDFLAGS} | tr ' ' '\n' | sed 's/^-/    -/g'`
 LIBS:
 `echo ${LIBS} | tr ' ' '\n' | sed 's/^-/    -/g'`
--------------------------------------------------------" > config.summary
+-------------------------------------------------------" > grid.configure.summary
+
+GRID_SUMMARY="`cat grid.configure.summary`"
+AM_SUBST_NOTMAKE([GRID_SUMMARY])
+AC_SUBST([GRID_SUMMARY])
+
+AC_CONFIG_FILES([grid-config], [chmod +x grid-config])
+AC_CONFIG_FILES(Makefile)
+AC_CONFIG_FILES(lib/Makefile)
+AC_CONFIG_FILES(tests/Makefile)
+AC_CONFIG_FILES(tests/IO/Makefile)
+AC_CONFIG_FILES(tests/core/Makefile)
+AC_CONFIG_FILES(tests/debug/Makefile)
+AC_CONFIG_FILES(tests/forces/Makefile)
+AC_CONFIG_FILES(tests/hadrons/Makefile)
+AC_CONFIG_FILES(tests/hmc/Makefile)
+AC_CONFIG_FILES(tests/solver/Makefile)
+AC_CONFIG_FILES(tests/smearing/Makefile)
+AC_CONFIG_FILES(tests/qdpxx/Makefile)
+AC_CONFIG_FILES(tests/testu01/Makefile)
+AC_CONFIG_FILES(benchmarks/Makefile)
+AC_CONFIG_FILES(extras/Makefile)
+AC_CONFIG_FILES(extras/Hadrons/Makefile)
+AC_OUTPUT
+
 echo ""
-cat config.summary
+cat grid.configure.summary
 echo ""
+

extras/Hadrons/Application.cc

@@ -162,7 +162,8 @@ void Application::saveParameterFile(const std::string parameterFileName)
 sizeString((size)*locVol_) << " (" << sizeString(size)  << "/site)"
 
 #define DEFINE_MEMPEAK \
-auto memPeak = [this](const std::vector<unsigned int> &program)\
+GeneticScheduler<unsigned int>::ObjFunc memPeak = \
+[this](const std::vector<unsigned int> &program)\
 {\
     unsigned int memPeak;\
     bool         msg;\

extras/Hadrons/Global.hpp

@@ -145,6 +145,15 @@ std::string typeName(void)
     return typeName(typeIdPt<T>());
 }
 
+// default writers/readers
+#ifdef HAVE_HDF5
+typedef Hdf5Reader CorrReader;
+typedef Hdf5Writer CorrWriter;
+#else
+typedef XmlReader CorrReader;
+typedef XmlWriter CorrWriter;
+#endif
+
 END_HADRONS_NAMESPACE
 
 #endif // Hadrons_Global_hpp_

extras/Hadrons/Modules.hpp

@@ -29,12 +29,20 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #include <Grid/Hadrons/Modules/MAction/DWF.hpp>
 #include <Grid/Hadrons/Modules/MAction/Wilson.hpp>
 #include <Grid/Hadrons/Modules/MContraction/Baryon.hpp>
+#include <Grid/Hadrons/Modules/MContraction/DiscLoop.hpp>
+#include <Grid/Hadrons/Modules/MContraction/Gamma3pt.hpp>
 #include <Grid/Hadrons/Modules/MContraction/Meson.hpp>
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp>
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp>
+#include <Grid/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp>
 #include <Grid/Hadrons/Modules/MGauge/Load.hpp>
 #include <Grid/Hadrons/Modules/MGauge/Random.hpp>
 #include <Grid/Hadrons/Modules/MGauge/Unit.hpp>
+#include <Grid/Hadrons/Modules/MLoop/NoiseLoop.hpp>
 #include <Grid/Hadrons/Modules/MSolver/RBPrecCG.hpp>
 #include <Grid/Hadrons/Modules/MSource/Point.hpp>
 #include <Grid/Hadrons/Modules/MSource/SeqGamma.hpp>
+#include <Grid/Hadrons/Modules/MSource/Wall.hpp>
 #include <Grid/Hadrons/Modules/MSource/Z2.hpp>
 #include <Grid/Hadrons/Modules/Quark.hpp>

extras/Hadrons/Modules/MContraction/Baryon.hpp

@@ -112,7 +112,7 @@ void TBaryon<FImpl1, FImpl2, FImpl3>::execute(void)
                  << " quarks '" << par().q1 << "', '" << par().q2 << "', and '"
                  << par().q3 << "'" << std::endl;
     
-    XmlWriter             writer(par().output);
+    CorrWriter             writer(par().output);
     PropagatorField1      &q1 = *env().template getObject<PropagatorField1>(par().q1);
     PropagatorField2      &q2 = *env().template getObject<PropagatorField2>(par().q2);
     PropagatorField3      &q3 = *env().template getObject<PropagatorField3>(par().q2);

extras/Hadrons/Modules/MContraction/DiscLoop.hpp

@@ -0,0 +1,144 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/DiscLoop.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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 Hadrons_DiscLoop_hpp_
+#define Hadrons_DiscLoop_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                                DiscLoop                                    *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+class DiscLoopPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(DiscLoopPar,
+                                    std::string,    q_loop,
+                                    Gamma::Algebra, gamma,
+                                    std::string,    output);
+};
+
+template <typename FImpl>
+class TDiscLoop: public Module<DiscLoopPar>
+{
+    TYPE_ALIASES(FImpl,);
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        Gamma::Algebra, gamma,
+                                        std::vector<Complex>, corr);
+    };
+public:
+    // constructor
+    TDiscLoop(const std::string name);
+    // destructor
+    virtual ~TDiscLoop(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(DiscLoop, TDiscLoop<FIMPL>, MContraction);
+
+/******************************************************************************
+ *                       TDiscLoop implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TDiscLoop<FImpl>::TDiscLoop(const std::string name)
+: Module<DiscLoopPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TDiscLoop<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().q_loop};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TDiscLoop<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TDiscLoop<FImpl>::setup(void)
+{
+    
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TDiscLoop<FImpl>::execute(void)
+{
+    LOG(Message) << "Computing disconnected loop contraction '" << getName() 
+                 << "' using '" << par().q_loop << "' with " << par().gamma 
+                 << " insertion." << std::endl;
+
+    CorrWriter            writer(par().output);
+    PropagatorField       &q_loop = *env().template getObject<PropagatorField>(par().q_loop);
+    LatticeComplex        c(env().getGrid());
+    Gamma                 gamma(par().gamma);
+    std::vector<TComplex> buf;
+    Result                result;
+
+    c = trace(gamma*q_loop);
+    sliceSum(c, buf, Tp);
+
+    result.gamma = par().gamma;
+    result.corr.resize(buf.size());
+    for (unsigned int t = 0; t < buf.size(); ++t)
+    {
+        result.corr[t] = TensorRemove(buf[t]);
+    }
+
+    write(writer, "disc", result);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_DiscLoop_hpp_

extras/Hadrons/Modules/MContraction/Gamma3pt.hpp

@@ -0,0 +1,170 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/Gamma3pt.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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 Hadrons_Gamma3pt_hpp_
+#define Hadrons_Gamma3pt_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ * 3pt contraction with gamma matrix insertion.
+ *
+ * Schematic:
+ *
+ *             q2           q3
+ *        /----<------*------<----¬
+ *       /          gamma          \
+ *      /                           \
+ *   i *                            * f
+ *      \                          /
+ *       \                        /
+ *        \----------->----------/
+ *                   q1
+ *
+ *      trace(g5*q1*adj(q2)*g5*gamma*q3)
+ */
+
+/******************************************************************************
+ *                               Gamma3pt                                     *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+class Gamma3ptPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(Gamma3ptPar,
+                                    std::string,    q1,
+                                    std::string,    q2,
+                                    std::string,    q3,
+                                    Gamma::Algebra, gamma,
+                                    std::string,    output);
+};
+
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+class TGamma3pt: public Module<Gamma3ptPar>
+{
+    TYPE_ALIASES(FImpl1, 1);
+    TYPE_ALIASES(FImpl2, 2);
+    TYPE_ALIASES(FImpl3, 3);
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        Gamma::Algebra, gamma,
+                                        std::vector<Complex>, corr);
+    };
+public:
+    // constructor
+    TGamma3pt(const std::string name);
+    // destructor
+    virtual ~TGamma3pt(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(Gamma3pt, ARG(TGamma3pt<FIMPL, FIMPL, FIMPL>), MContraction);
+
+/******************************************************************************
+ *                       TGamma3pt implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+TGamma3pt<FImpl1, FImpl2, FImpl3>::TGamma3pt(const std::string name)
+: Module<Gamma3ptPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+std::vector<std::string> TGamma3pt<FImpl1, FImpl2, FImpl3>::getInput(void)
+{
+    std::vector<std::string> in = {par().q1, par().q2, par().q3};
+    
+    return in;
+}
+
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+std::vector<std::string> TGamma3pt<FImpl1, FImpl2, FImpl3>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+void TGamma3pt<FImpl1, FImpl2, FImpl3>::setup(void)
+{
+    
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+void TGamma3pt<FImpl1, FImpl2, FImpl3>::execute(void)
+{
+    LOG(Message) << "Computing 3pt contractions '" << getName() << "' using"
+                 << " quarks '" << par().q1 << "', '" << par().q2 << "' and '"
+                 << par().q3 << "', with " << par().gamma << " insertion." 
+                 << std::endl;
+
+    CorrWriter            writer(par().output);
+    PropagatorField1      &q1 = *env().template getObject<PropagatorField1>(par().q1);
+    PropagatorField2      &q2 = *env().template getObject<PropagatorField2>(par().q2);
+    PropagatorField3      &q3 = *env().template getObject<PropagatorField3>(par().q3);
+    LatticeComplex        c(env().getGrid());
+    Gamma                 g5(Gamma::Algebra::Gamma5);
+    Gamma                 gamma(par().gamma);
+    std::vector<TComplex> buf;
+    Result                result;
+
+    c = trace(g5*q1*adj(q2)*(g5*gamma)*q3);
+    sliceSum(c, buf, Tp);
+
+    result.gamma = par().gamma;
+    result.corr.resize(buf.size());
+    for (unsigned int t = 0; t < buf.size(); ++t)
+    {
+        result.corr[t] = TensorRemove(buf[t]);
+    }
+
+    write(writer, "gamma3pt", result);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Gamma3pt_hpp_

extras/Hadrons/Modules/MContraction/Meson.hpp

@@ -6,8 +6,10 @@ Source file: extras/Hadrons/Modules/MContraction/Meson.hpp
 
 Copyright (C) 2015
 Copyright (C) 2016
+Copyright (C) 2017
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+        Andrew Lawson    <andrew.lawson1991@gmail.com>
 
 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
@@ -36,20 +38,39 @@ See the full license in the file "LICENSE" in the top level distribution directo
 
 BEGIN_HADRONS_NAMESPACE
 
+/*
+ 
+ Meson contractions
+ -----------------------------
+ 
+ * options:
+ - q1: input propagator 1 (string)
+ - q2: input propagator 2 (string)
+ - gammas: gamma products to insert at sink & source, pairs of gamma matrices 
+           (space-separated strings) in angled brackets (i.e. <g_sink g_src>),
+           in a sequence (e.g. "<Gamma5 Gamma5><Gamma5 GammaT>").
+
+           Special values: "all" - perform all possible contractions.
+ - mom: momentum insertion, space-separated float sequence (e.g ".1 .2 1. 0."),
+        given as multiples of (2*pi) / L.
+*/
+
 /******************************************************************************
  *                                TMeson                                       *
  ******************************************************************************/
 BEGIN_MODULE_NAMESPACE(MContraction)
 
+typedef std::pair<Gamma::Algebra, Gamma::Algebra> GammaPair;
+
 class MesonPar: Serializable
 {
 public:
     GRID_SERIALIZABLE_CLASS_MEMBERS(MesonPar,
-                                    std::string,    q1,
-                                    std::string,    q2,
-                                    std::string,    output,
-                                    Gamma::Algebra, gammaSource,
-                                    Gamma::Algebra, gammaSink);
+                                    std::string, q1,
+                                    std::string, q2,
+                                    std::string, gammas,
+                                    std::string, mom,
+                                    std::string, output);
 };
 
 template <typename FImpl1, typename FImpl2>
@@ -61,7 +82,10 @@ public:
     class Result: Serializable
     {
     public:
-        GRID_SERIALIZABLE_CLASS_MEMBERS(Result, std::vector<Complex>, corr);
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        Gamma::Algebra, gamma_snk,
+                                        Gamma::Algebra, gamma_src,
+                                        std::vector<Complex>, corr);
     };
 public:
     // constructor
@@ -71,6 +95,7 @@ public:
     // dependencies/products
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+    virtual void parseGammaString(std::vector<GammaPair> &gammaList);
     // execution
     virtual void execute(void);
 };
@@ -103,6 +128,32 @@ std::vector<std::string> TMeson<FImpl1, FImpl2>::getOutput(void)
     return output;
 }
 
+template <typename FImpl1, typename FImpl2>
+void TMeson<FImpl1, FImpl2>::parseGammaString(std::vector<GammaPair> &gammaList)
+{
+    // Determine gamma matrices to insert at source/sink.
+    if (par().gammas.compare("all") == 0)
+    {
+        // Do all contractions.
+        unsigned int n_gam = Ns * Ns;
+        gammaList.resize(n_gam*n_gam);
+        for (unsigned int i = 1; i < Gamma::nGamma; i += 2)
+        {
+            for (unsigned int j = 1; j < Gamma::nGamma; j += 2)
+            {
+                gammaList.push_back(std::make_pair((Gamma::Algebra)i, 
+                                                   (Gamma::Algebra)j));
+            }
+        }
+    }
+    else
+    {
+        // Parse individual contractions from input string.
+        gammaList = strToVec<GammaPair>(par().gammas);
+    }
+}
+
+
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl1, typename FImpl2>
 void TMeson<FImpl1, FImpl2>::execute(void)
@@ -111,21 +162,44 @@ void TMeson<FImpl1, FImpl2>::execute(void)
                  << " quarks '" << par().q1 << "' and '" << par().q2 << "'"
                  << std::endl;
     
-    XmlWriter             writer(par().output);
-    PropagatorField1      &q1 = *env().template getObject<PropagatorField1>(par().q1);
-    PropagatorField2      &q2 = *env().template getObject<PropagatorField2>(par().q2);
-    LatticeComplex        c(env().getGrid());
-    Gamma                 gSrc(par().gammaSource), gSnk(par().gammaSink);
-    Gamma                 g5(Gamma::Algebra::Gamma5);
-    std::vector<TComplex> buf;
-    Result                result;
+    CorrWriter              writer(par().output);
+    PropagatorField1       &q1 = *env().template getObject<PropagatorField1>(par().q1);
+    PropagatorField2       &q2 = *env().template getObject<PropagatorField2>(par().q2);
+    LatticeComplex         c(env().getGrid());
+    Gamma                  g5(Gamma::Algebra::Gamma5);
+    std::vector<GammaPair> gammaList;
+    std::vector<TComplex>  buf;
+    std::vector<Result>    result;
+    std::vector<Real>      p;
 
-    c = trace(gSnk*q1*adj(gSrc)*g5*adj(q2)*g5);
-    sliceSum(c, buf, Tp);
-    result.corr.resize(buf.size());
-    for (unsigned int t = 0; t < buf.size(); ++t)
+    p  = strToVec<Real>(par().mom);
+    LatticeComplex         ph(env().getGrid()), coor(env().getGrid());
+    Complex                i(0.0,1.0);
+    ph = zero;
+    for(unsigned int mu = 0; mu < env().getNd(); mu++)
     {
-        result.corr[t] = TensorRemove(buf[t]);
+        LatticeCoordinate(coor, mu);
+        ph = ph + p[mu]*coor*((1./(env().getGrid()->_fdimensions[mu])));
+    }
+    ph = exp((Real)(2*M_PI)*i*ph);
+    
+    parseGammaString(gammaList);
+
+    result.resize(gammaList.size());
+    for (unsigned int i = 0; i < result.size(); ++i)
+    {
+        Gamma gSnk(gammaList[i].first);
+        Gamma gSrc(gammaList[i].second);
+        c = trace((g5*gSnk)*q1*(adj(gSrc)*g5)*adj(q2))*ph;
+        sliceSum(c, buf, Tp);
+
+        result[i].gamma_snk = gammaList[i].first;
+        result[i].gamma_src = gammaList[i].second;
+        result[i].corr.resize(buf.size());
+        for (unsigned int t = 0; t < buf.size(); ++t)
+        {
+            result[i].corr[t] = TensorRemove(buf[t]);
+        }
     }
     write(writer, "meson", result);
 }

extras/Hadrons/Modules/MContraction/WeakHamiltonian.hpp

@@ -0,0 +1,114 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonian.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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 Hadrons_WeakHamiltonian_hpp_
+#define Hadrons_WeakHamiltonian_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         WeakHamiltonian                                    *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+/*******************************************************************************
+ * Utilities for contractions involving the Weak Hamiltonian.
+ ******************************************************************************/
+//// Sum and store correlator.
+#define MAKE_DIAG(exp, buf, res, n)\
+sliceSum(exp, buf, Tp);\
+res.name = (n);\
+res.corr.resize(buf.size());\
+for (unsigned int t = 0; t < buf.size(); ++t)\
+{\
+    res.corr[t] = TensorRemove(buf[t]);\
+}
+
+//// Contraction of mu index: use 'mu' variable in exp.
+#define SUM_MU(buf,exp)\
+buf = zero;\
+for (unsigned int mu = 0; mu < ndim; ++mu)\
+{\
+    buf += exp;\
+}
+
+enum 
+{
+  i_V = 0,
+  i_A = 1,
+  n_i = 2
+};
+
+class WeakHamiltonianPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(WeakHamiltonianPar,
+                                    std::string, q1,
+                                    std::string, q2,
+                                    std::string, q3,
+                                    std::string, q4,
+                                    std::string, output);
+};
+
+#define MAKE_WEAK_MODULE(modname)\
+class T##modname: public Module<WeakHamiltonianPar>\
+{\
+public:\
+    TYPE_ALIASES(FIMPL,)\
+    class Result: Serializable\
+    {\
+    public:\
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,\
+                                        std::string, name,\
+                                        std::vector<Complex>, corr);\
+    };\
+public:\
+    /* constructor */ \
+    T##modname(const std::string name);\
+    /* destructor */ \
+    virtual ~T##modname(void) = default;\
+    /* dependency relation */ \
+    virtual std::vector<std::string> getInput(void);\
+    virtual std::vector<std::string> getOutput(void);\
+    /* setup */ \
+    virtual void setup(void);\
+    /* execution */ \
+    virtual void execute(void);\
+    std::vector<std::string> VA_label = {"V", "A"};\
+};\
+MODULE_REGISTER_NS(modname, T##modname, MContraction);
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_WeakHamiltonian_hpp_

extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.cc

@@ -0,0 +1,137 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.cc
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MContraction;
+
+/*
+ * Weak Hamiltonian current-current contractions, Eye-type.
+ * 
+ * These contractions are generated by the Q1 and Q2 operators in the physical
+ * basis (see e.g. Fig 3 of arXiv:1507.03094).
+ * 
+ * Schematics:        q4                 |                  
+ *                  /-<-¬                |                             
+ *                 /     \               |             q2           q3
+ *                 \     /               |        /----<------*------<----¬                        
+ *            q2    \   /    q3          |       /          /-*-¬          \
+ *       /-----<-----* *-----<----¬      |      /          /     \          \
+ *    i *            H_W           * f   |   i *           \     /  q4      * f
+ *       \                        /      |      \           \->-/          /   
+ *        \                      /       |       \                        /       
+ *         \---------->---------/        |        \----------->----------/        
+ *                   q1                  |                   q1                  
+ *                                       |
+ *                Saucer (S)             |                  Eye (E)
+ * 
+ * S: trace(q3*g5*q1*adj(q2)*g5*gL[mu][p_1]*q4*gL[mu][p_2])
+ * E: trace(q3*g5*q1*adj(q2)*g5*gL[mu][p_1])*trace(q4*gL[mu][p_2])
+ */
+
+/******************************************************************************
+ *                  TWeakHamiltonianEye implementation                        *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TWeakHamiltonianEye::TWeakHamiltonianEye(const std::string name)
+: Module<WeakHamiltonianPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TWeakHamiltonianEye::getInput(void)
+{
+    std::vector<std::string> in = {par().q1, par().q2, par().q3, par().q4};
+    
+    return in;
+}
+
+std::vector<std::string> TWeakHamiltonianEye::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TWeakHamiltonianEye::setup(void)
+{
+
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TWeakHamiltonianEye::execute(void)
+{
+    LOG(Message) << "Computing Weak Hamiltonian (Eye type) contractions '" 
+                 << getName() << "' using quarks '" << par().q1 << "', '" 
+                 << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
+                 << "'." << std::endl;
+
+    CorrWriter             writer(par().output);
+    PropagatorField &q1 = *env().template getObject<PropagatorField>(par().q1);
+    PropagatorField &q2 = *env().template getObject<PropagatorField>(par().q2);
+    PropagatorField &q3 = *env().template getObject<PropagatorField>(par().q3);
+    PropagatorField &q4 = *env().template getObject<PropagatorField>(par().q4);
+    Gamma g5            = Gamma(Gamma::Algebra::Gamma5);
+    LatticeComplex        expbuf(env().getGrid());
+    std::vector<TComplex> corrbuf;
+    std::vector<Result>   result(n_eye_diag);
+    unsigned int ndim   = env().getNd();
+
+    PropagatorField              tmp1(env().getGrid());
+    LatticeComplex               tmp2(env().getGrid());
+    std::vector<PropagatorField> S_body(ndim, tmp1);
+    std::vector<PropagatorField> S_loop(ndim, tmp1);
+    std::vector<LatticeComplex>  E_body(ndim, tmp2);
+    std::vector<LatticeComplex>  E_loop(ndim, tmp2);
+
+    // Setup for S-type contractions.
+    for (int mu = 0; mu < ndim; ++mu)
+    {
+        S_body[mu] = MAKE_SE_BODY(q1, q2, q3, GammaL(Gamma::gmu[mu]));
+        S_loop[mu] = MAKE_SE_LOOP(q4, GammaL(Gamma::gmu[mu]));
+    }
+
+    // Perform S-type contractions.    
+    SUM_MU(expbuf, trace(S_body[mu]*S_loop[mu]))
+    MAKE_DIAG(expbuf, corrbuf, result[S_diag], "HW_S")
+
+    // Recycle sub-expressions for E-type contractions.
+    for (unsigned int mu = 0; mu < ndim; ++mu)
+    {
+        E_body[mu] = trace(S_body[mu]);
+        E_loop[mu] = trace(S_loop[mu]);
+    }
+
+    // Perform E-type contractions.
+    SUM_MU(expbuf, E_body[mu]*E_loop[mu])
+    MAKE_DIAG(expbuf, corrbuf, result[E_diag], "HW_E")
+
+    write(writer, "HW_Eye", result);
+}

extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp

@@ -0,0 +1,58 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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 Hadrons_WeakHamiltonianEye_hpp_
+#define Hadrons_WeakHamiltonianEye_hpp_
+
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         WeakHamiltonianEye                                 *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+enum
+{
+    S_diag = 0,
+    E_diag = 1,
+    n_eye_diag = 2
+};
+
+// Saucer and Eye subdiagram contractions.
+#define MAKE_SE_BODY(Q_1, Q_2, Q_3, gamma) (Q_3*g5*Q_1*adj(Q_2)*g5*gamma)
+#define MAKE_SE_LOOP(Q_loop, gamma) (Q_loop*gamma)
+
+MAKE_WEAK_MODULE(WeakHamiltonianEye)
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_WeakHamiltonianEye_hpp_

extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.cc

@@ -0,0 +1,139 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.cc
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MContraction;
+
+/*
+ * Weak Hamiltonian current-current contractions, Non-Eye-type.
+ * 
+ * These contractions are generated by the Q1 and Q2 operators in the physical
+ * basis (see e.g. Fig 3 of arXiv:1507.03094).
+ * 
+ * Schematic:     
+ *            q2             q3          |           q2              q3
+ *          /--<--¬       /--<--¬        |        /--<--¬         /--<--¬       
+ *         /       \     /       \       |       /       \       /       \      
+ *        /         \   /         \      |      /         \     /         \     
+ *       /           \ /           \     |     /           \   /           \    
+ *    i *             * H_W         *  f |  i *             * * H_W         * f 
+ *      \             *             |    |     \           /   \           /
+ *       \           / \           /     |      \         /     \         /    
+ *        \         /   \         /      |       \       /       \       /  
+ *         \       /     \       /       |        \-->--/         \-->--/      
+ *          \-->--/       \-->--/        |          q1               q4 
+ *            q1             q4          |
+ *                Connected (C)          |                 Wing (W)
+ *
+ * C: trace(q1*adj(q2)*g5*gL[mu]*q3*adj(q4)*g5*gL[mu])
+ * W: trace(q1*adj(q2)*g5*gL[mu])*trace(q3*adj(q4)*g5*gL[mu])
+ * 
+ */
+
+/******************************************************************************
+ *                  TWeakHamiltonianNonEye implementation                     *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TWeakHamiltonianNonEye::TWeakHamiltonianNonEye(const std::string name)
+: Module<WeakHamiltonianPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TWeakHamiltonianNonEye::getInput(void)
+{
+    std::vector<std::string> in = {par().q1, par().q2, par().q3, par().q4};
+    
+    return in;
+}
+
+std::vector<std::string> TWeakHamiltonianNonEye::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TWeakHamiltonianNonEye::setup(void)
+{
+
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TWeakHamiltonianNonEye::execute(void)
+{
+    LOG(Message) << "Computing Weak Hamiltonian (Non-Eye type) contractions '" 
+                 << getName() << "' using quarks '" << par().q1 << "', '" 
+                 << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
+                 << "'." << std::endl;
+    
+    CorrWriter             writer(par().output);
+    PropagatorField &q1 = *env().template getObject<PropagatorField>(par().q1);
+    PropagatorField &q2 = *env().template getObject<PropagatorField>(par().q2);
+    PropagatorField &q3 = *env().template getObject<PropagatorField>(par().q3);
+    PropagatorField &q4 = *env().template getObject<PropagatorField>(par().q4);
+    Gamma g5            = Gamma(Gamma::Algebra::Gamma5);
+    LatticeComplex        expbuf(env().getGrid());
+    std::vector<TComplex> corrbuf;
+    std::vector<Result>   result(n_noneye_diag); 
+    unsigned int ndim   = env().getNd();
+
+    PropagatorField              tmp1(env().getGrid());
+    LatticeComplex               tmp2(env().getGrid());
+    std::vector<PropagatorField> C_i_side_loop(ndim, tmp1);
+    std::vector<PropagatorField> C_f_side_loop(ndim, tmp1);
+    std::vector<LatticeComplex>  W_i_side_loop(ndim, tmp2);
+    std::vector<LatticeComplex>  W_f_side_loop(ndim, tmp2);
+
+    // Setup for C-type contractions.
+    for (int mu = 0; mu < ndim; ++mu)
+    {
+        C_i_side_loop[mu] = MAKE_CW_SUBDIAG(q1, q2, GammaL(Gamma::gmu[mu]));
+        C_f_side_loop[mu] = MAKE_CW_SUBDIAG(q3, q4, GammaL(Gamma::gmu[mu]));
+    }
+
+    // Perform C-type contractions.    
+    SUM_MU(expbuf, trace(C_i_side_loop[mu]*C_f_side_loop[mu]))
+    MAKE_DIAG(expbuf, corrbuf, result[C_diag], "HW_C")
+
+    // Recycle sub-expressions for W-type contractions.
+    for (unsigned int mu = 0; mu < ndim; ++mu)
+    {
+        W_i_side_loop[mu] = trace(C_i_side_loop[mu]);
+        W_f_side_loop[mu] = trace(C_f_side_loop[mu]);
+    }
+
+    // Perform W-type contractions.
+    SUM_MU(expbuf, W_i_side_loop[mu]*W_f_side_loop[mu])
+    MAKE_DIAG(expbuf, corrbuf, result[W_diag], "HW_W")
+
+    write(writer, "HW_NonEye", result);
+}

extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp

@@ -0,0 +1,57 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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 Hadrons_WeakHamiltonianNonEye_hpp_
+#define Hadrons_WeakHamiltonianNonEye_hpp_
+
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         WeakHamiltonianNonEye                              *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+enum
+{
+    W_diag = 0,
+    C_diag = 1,
+    n_noneye_diag = 2
+};
+
+// Wing and Connected subdiagram contractions
+#define MAKE_CW_SUBDIAG(Q_1, Q_2, gamma) (Q_1*adj(Q_2)*g5*gamma)
+
+MAKE_WEAK_MODULE(WeakHamiltonianNonEye)
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_WeakHamiltonianNonEye_hpp_

extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.cc

@@ -0,0 +1,135 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.cc
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MContraction;
+
+/*
+ * Weak Hamiltonian + current contractions, disconnected topology for neutral 
+ * mesons.
+ * 
+ * These contractions are generated by operators Q_1,...,10 of the dS=1 Weak
+ * Hamiltonian in the physical basis and an additional current J (see e.g. 
+ * Fig 11 of arXiv:1507.03094).
+ * 
+ * Schematic:
+ *                        
+ *           q2          q4             q3
+ *       /--<--¬     /---<--¬       /---<--¬
+ *     /         \ /         \     /        \
+ *  i *           * H_W      |  J *          * f
+ *     \         / \         /     \        /
+ *      \--->---/   \-------/       \------/
+ *          q1 
+ * 
+ * options
+ * - q1: input propagator 1 (string)
+ * - q2: input propagator 2 (string)
+ * - q3: input propagator 3 (string), assumed to be sequential propagator 
+ * - q4: input propagator 4 (string), assumed to be a loop
+ * 
+ * type 1: trace(q1*adj(q2)*g5*gL[mu])*trace(loop*gL[mu])*trace(q3*g5)
+ * type 2: trace(q1*adj(q2)*g5*gL[mu]*loop*gL[mu])*trace(q3*g5)
+ */
+
+/*******************************************************************************
+ *                  TWeakNeutral4ptDisc implementation                         *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+TWeakNeutral4ptDisc::TWeakNeutral4ptDisc(const std::string name)
+: Module<WeakHamiltonianPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+std::vector<std::string> TWeakNeutral4ptDisc::getInput(void)
+{
+    std::vector<std::string> in = {par().q1, par().q2, par().q3, par().q4};
+    
+    return in;
+}
+
+std::vector<std::string> TWeakNeutral4ptDisc::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+void TWeakNeutral4ptDisc::setup(void)
+{
+
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+void TWeakNeutral4ptDisc::execute(void)
+{
+    LOG(Message) << "Computing Weak Hamiltonian neutral disconnected contractions '" 
+                 << getName() << "' using quarks '" << par().q1 << "', '" 
+                 << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
+                 << "'." << std::endl;
+
+    CorrWriter             writer(par().output);
+    PropagatorField &q1 = *env().template getObject<PropagatorField>(par().q1);
+    PropagatorField &q2 = *env().template getObject<PropagatorField>(par().q2);
+    PropagatorField &q3 = *env().template getObject<PropagatorField>(par().q3);
+    PropagatorField &q4 = *env().template getObject<PropagatorField>(par().q4);
+    Gamma g5            = Gamma(Gamma::Algebra::Gamma5);
+    LatticeComplex        expbuf(env().getGrid());
+    std::vector<TComplex> corrbuf;
+    std::vector<Result>   result(n_neut_disc_diag);
+    unsigned int ndim   = env().getNd();
+
+    PropagatorField              tmp(env().getGrid());
+    std::vector<PropagatorField> meson(ndim, tmp);
+    std::vector<PropagatorField> loop(ndim, tmp);
+    LatticeComplex               curr(env().getGrid());
+
+    // Setup for type 1 contractions.
+    for (int mu = 0; mu < ndim; ++mu)
+    {
+        meson[mu] = MAKE_DISC_MESON(q1, q2, GammaL(Gamma::gmu[mu]));
+        loop[mu] = MAKE_DISC_LOOP(q4, GammaL(Gamma::gmu[mu]));
+    }
+    curr = MAKE_DISC_CURR(q3, GammaL(Gamma::Algebra::Gamma5));
+
+    // Perform type 1 contractions.    
+    SUM_MU(expbuf, trace(meson[mu]*loop[mu]))
+    expbuf *= curr;
+    MAKE_DIAG(expbuf, corrbuf, result[neut_disc_1_diag], "HW_disc0_1")
+
+    // Perform type 2 contractions.
+    SUM_MU(expbuf, trace(meson[mu])*trace(loop[mu]))
+    expbuf *= curr;
+    MAKE_DIAG(expbuf, corrbuf, result[neut_disc_2_diag], "HW_disc0_2")
+
+    write(writer, "HW_disc0", result);
+}

extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp

@@ -0,0 +1,59 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+
+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 Hadrons_WeakNeutral4ptDisc_hpp_
+#define Hadrons_WeakNeutral4ptDisc_hpp_
+
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         WeakNeutral4ptDisc                                 *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MContraction)
+
+enum
+{
+    neut_disc_1_diag = 0,
+    neut_disc_2_diag = 1,
+    n_neut_disc_diag = 2
+};
+
+// Neutral 4pt disconnected subdiagram contractions.
+#define MAKE_DISC_MESON(Q_1, Q_2, gamma) (Q_1*adj(Q_2)*g5*gamma)
+#define MAKE_DISC_LOOP(Q_LOOP, gamma) (Q_LOOP*gamma)
+#define MAKE_DISC_CURR(Q_c, gamma) (trace(Q_c*gamma))
+
+MAKE_WEAK_MODULE(WeakNeutral4ptDisc)
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_WeakNeutral4ptDisc_hpp_

extras/Hadrons/Modules/MGauge/Load.cc

@@ -65,7 +65,7 @@ void TLoad::setup(void)
 // execution ///////////////////////////////////////////////////////////////////
 void TLoad::execute(void)
 {
-    NerscField  header;
+    FieldMetaData  header;
     std::string fileName = par().file + "."
                            + std::to_string(env().getTrajectory());
     
@@ -74,5 +74,5 @@ void TLoad::execute(void)
     LatticeGaugeField &U = *env().createLattice<LatticeGaugeField>(getName());
     NerscIO::readConfiguration(U, header, fileName);
     LOG(Message) << "NERSC header:" << std::endl;
-    dump_nersc_header(header, LOG(Message));
+    dump_meta_data(header, LOG(Message));
 }

extras/Hadrons/Modules/MLoop/NoiseLoop.hpp

@@ -0,0 +1,132 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MLoop/NoiseLoop.hpp
+
+Copyright (C) 2016
+
+Author: Andrew Lawson <andrew.lawson1991@gmail.com>
+
+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 Hadrons_NoiseLoop_hpp_
+#define Hadrons_NoiseLoop_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ 
+ Noise loop propagator
+ -----------------------------
+ * loop_x = q_x * adj(eta_x)
+ 
+ * options:
+ - q = Result of inversion on noise source.
+ - eta = noise source.
+
+ */
+
+
+/******************************************************************************
+ *                         NoiseLoop                                          *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MLoop)
+
+class NoiseLoopPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(NoiseLoopPar,
+                                    std::string, q,
+                                    std::string, eta);
+};
+
+template <typename FImpl>
+class TNoiseLoop: public Module<NoiseLoopPar>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TNoiseLoop(const std::string name);
+    // destructor
+    virtual ~TNoiseLoop(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(NoiseLoop, TNoiseLoop<FIMPL>, MLoop);
+
+/******************************************************************************
+ *                 TNoiseLoop implementation                                  *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TNoiseLoop<FImpl>::TNoiseLoop(const std::string name)
+: Module<NoiseLoopPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TNoiseLoop<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().q, par().eta};
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TNoiseLoop<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TNoiseLoop<FImpl>::setup(void)
+{
+    env().template registerLattice<PropagatorField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TNoiseLoop<FImpl>::execute(void)
+{
+    PropagatorField &loop = *env().template createLattice<PropagatorField>(getName());
+    PropagatorField &q    = *env().template getObject<PropagatorField>(par().q);
+    PropagatorField &eta  = *env().template getObject<PropagatorField>(par().eta);
+    loop = q*adj(eta);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_NoiseLoop_hpp_

extras/Hadrons/Modules/MSource/SeqGamma.hpp

@@ -6,6 +6,7 @@ Source file: extras/Hadrons/Modules/MSource/SeqGamma.hpp
 
 Copyright (C) 2015
 Copyright (C) 2016
+Copyright (C) 2017
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -149,9 +150,9 @@ void TSeqGamma<FImpl>::execute(void)
     for(unsigned int mu = 0; mu < env().getNd(); mu++)
     {
         LatticeCoordinate(coor, mu);
-        ph = ph + p[mu]*coor;
+        ph = ph + p[mu]*coor*((1./(env().getGrid()->_fdimensions[mu])));
     }
-    ph = exp(i*ph);
+    ph = exp((Real)(2*M_PI)*i*ph);
     LatticeCoordinate(t, Tp);
     src = where((t >= par().tA) and (t <= par().tB), ph*(g*q), 0.*q);
 }

extras/Hadrons/Modules/MSource/Wall.hpp

@@ -0,0 +1,147 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MSource/Wall.hpp
+
+Copyright (C) 2017
+
+Author: Andrew Lawson <andrew.lawson1991@gmail.com>
+
+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 Hadrons_WallSource_hpp_
+#define Hadrons_WallSource_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/*
+ 
+ Wall source
+ -----------------------------
+ * src_x = delta(x_3 - tW) * exp(i x.mom)
+ 
+ * options:
+ - tW: source timeslice (integer)
+ - mom: momentum insertion, space-separated float sequence (e.g ".1 .2 1. 0.")
+ 
+ */
+
+/******************************************************************************
+ *                         Wall                                               *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MSource)
+
+class WallPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(WallPar,
+                                    unsigned int, tW,
+                                    std::string, mom);
+};
+
+template <typename FImpl>
+class TWall: public Module<WallPar>
+{
+public:
+    TYPE_ALIASES(FImpl,);
+public:
+    // constructor
+    TWall(const std::string name);
+    // destructor
+    virtual ~TWall(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(Wall, TWall<FIMPL>, MSource);
+
+/******************************************************************************
+ *                 TWall implementation                                       *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TWall<FImpl>::TWall(const std::string name)
+: Module<WallPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TWall<FImpl>::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TWall<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TWall<FImpl>::setup(void)
+{
+    env().template registerLattice<PropagatorField>(getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TWall<FImpl>::execute(void)
+{    
+    LOG(Message) << "Generating wall source at t = " << par().tW 
+                 << " with momentum " << par().mom << std::endl;
+    
+    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    Lattice<iScalar<vInteger>> t(env().getGrid());
+    LatticeComplex             ph(env().getGrid()), coor(env().getGrid());
+    std::vector<Real>          p;
+    Complex                    i(0.0,1.0);
+    
+    p  = strToVec<Real>(par().mom);
+    ph = zero;
+    for(unsigned int mu = 0; mu < Nd; mu++)
+    {
+        LatticeCoordinate(coor, mu);
+        ph = ph + p[mu]*coor*((1./(env().getGrid()->_fdimensions[mu])));
+    }
+    ph = exp((Real)(2*M_PI)*i*ph);
+    LatticeCoordinate(t, Tp);
+    src = 1.;
+    src = where((t == par().tW), src*ph, 0.*src);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_WallSource_hpp_

extras/Hadrons/Modules/Quark.hpp

@@ -173,7 +173,7 @@ void TQuark<FImpl>::execute(void)
                 *env().template getObject<PropagatorField>(getName());
             
             axpby_ssp_pminus(sol, 0., sol, 1., sol, 0, 0);
-            axpby_ssp_pplus(sol, 0., sol, 1., sol, 0, Ls_-1);
+            axpby_ssp_pplus(sol, 1., sol, 1., sol, 0, Ls_-1);
             ExtractSlice(tmp, sol, 0, 0);
             FermToProp(p4d, tmp, s, c);
         }

extras/Hadrons/modules.inc

@@ -1,4 +1,7 @@
 modules_cc =\
+  Modules/MContraction/WeakHamiltonianEye.cc \
+  Modules/MContraction/WeakHamiltonianNonEye.cc \
+  Modules/MContraction/WeakNeutral4ptDisc.cc \
   Modules/MGauge/Load.cc \
   Modules/MGauge/Random.cc \
   Modules/MGauge/Unit.cc
@@ -7,13 +10,21 @@ modules_hpp =\
   Modules/MAction/DWF.hpp \
   Modules/MAction/Wilson.hpp \
   Modules/MContraction/Baryon.hpp \
+  Modules/MContraction/DiscLoop.hpp \
+  Modules/MContraction/Gamma3pt.hpp \
   Modules/MContraction/Meson.hpp \
+  Modules/MContraction/WeakHamiltonian.hpp \
+  Modules/MContraction/WeakHamiltonianEye.hpp \
+  Modules/MContraction/WeakHamiltonianNonEye.hpp \
+  Modules/MContraction/WeakNeutral4ptDisc.hpp \
   Modules/MGauge/Load.hpp \
   Modules/MGauge/Random.hpp \
   Modules/MGauge/Unit.hpp \
+  Modules/MLoop/NoiseLoop.hpp \
   Modules/MSolver/RBPrecCG.hpp \
   Modules/MSource/Point.hpp \
   Modules/MSource/SeqGamma.hpp \
+  Modules/MSource/Wall.hpp \
   Modules/MSource/Z2.hpp \
   Modules/Quark.hpp
 

gcc-bug-report/README

@@ -21,3 +21,16 @@ problem. The test case works with icpc and with clang++, but fails consistently
 current variants.
 
 Peter
+
+
+************
+
+Second GCC bug reported, see Issue 100.
+
+https://wandbox.org/permlink/tzssJza6R9XnqANw
+https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80652
+
+Getting Travis fails under gcc-5 for Test_simd, now that I added more comprehensive testing to the
+CI test suite. The limitations of Travis runtime limits & weak cores are being shown.
+
+Travis uses 5.4.1 for g++-5.

grid-config.in

@@ -0,0 +1,86 @@
+#! /bin/sh
+
+prefix=@prefix@
+exec_prefix=@exec_prefix@
+includedir=@includedir@
+
+usage()
+{
+  cat <<EOF
+Usage: grid-config [OPTION]
+
+Known values for OPTION are:
+
+  --prefix     show Grid installation prefix
+  --cxxflags   print pre-processor and compiler flags
+  --ldflags    print library linking flags
+  --libs       print library linking information
+  --summary    print full build summary
+  --help       display this help and exit
+  --version    output version information
+  --git        print git revision
+
+EOF
+  
+  exit $1
+}
+
+if test $# -eq 0; then
+  usage 1
+fi
+
+cflags=false
+libs=false
+
+while test $# -gt 0; do
+  case "$1" in
+    -*=*) optarg=`echo "$1" | sed 's/[-_a-zA-Z0-9]*=//'` ;;
+    *) optarg= ;;
+  esac
+  
+  case "$1" in
+    --prefix)
+      echo $prefix
+    ;;
+    
+    --version)
+      echo @VERSION@
+      exit 0
+    ;;
+    
+    --git)
+      echo "@GRID_BRANCH@ @GRID_SHA@"
+      exit 0
+    ;;
+    
+    --help)
+      usage 0
+    ;;
+    
+    --cxxflags)
+      echo @GRID_CXXFLAGS@
+    ;;
+    
+    --ldflags)
+      echo @GRID_LDFLAGS@
+    ;;
+    
+    --libs)
+      echo @GRID_LIBS@
+    ;;
+    
+    --summary)
+      echo ""
+      echo "@GRID_SUMMARY@"
+      echo ""
+    ;;
+    
+    *)
+      usage
+      exit 1
+    ;;
+  esac
+  shift
+done
+
+exit 0

lib/DisableWarnings.h

@@ -0,0 +1,37 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/DisableWarnings.h
+
+Copyright (C) 2016
+
+Author: Guido Cossu <guido.cossu@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 DISABLE_WARNINGS_H
+#define DISABLE_WARNINGS_H
+
+ //disables and intel compiler specific warning (in json.hpp)
+#pragma warning disable 488  
+
+
+#endif

lib/Grid.h

@@ -38,52 +38,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_H
 #define GRID_H
 
-///////////////////
-// Std C++ dependencies
-///////////////////
-#include <cassert>
-#include <complex>
-#include <vector>
-#include <iostream>
-#include <iomanip>
-#include <random>
-#include <functional>
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdio.h>
-#include <signal.h>
-#include <ctime>
-#include <sys/time.h>
-#include <chrono>
-
-///////////////////
-// Grid headers
-///////////////////
-#include "Config.h"
-#include <Grid/Timer.h>
-#include <Grid/PerfCount.h>
-#include <Grid/Log.h>
-#include <Grid/AlignedAllocator.h>
-#include <Grid/Simd.h>
-#include <Grid/serialisation/Serialisation.h>
-#include <Grid/Threads.h>
-#include <Grid/Lexicographic.h>
-#include <Grid/Init.h>
-#include <Grid/Communicator.h> 
-#include <Grid/Cartesian.h>    
-#include <Grid/Tensors.h>      
-#include <Grid/Lattice.h>      
-#include <Grid/Cshift.h>       
-#include <Grid/Stencil.h>      
-#include <Grid/Algorithms.h>   
-#include <Grid/parallelIO/BinaryIO.h>
-#include <Grid/FFT.h>
-
-#include <Grid/qcd/QCD.h>
-#include <Grid/parallelIO/NerscIO.h>
-#include <Grid/qcd/hmc/NerscCheckpointer.h>
-#include <Grid/qcd/hmc/HmcRunner.h>
-
-
+#include <Grid/GridCore.h>
+#include <Grid/GridQCDcore.h>
+#include <Grid/qcd/action/Action.h>
+#include <Grid/qcd/smearing/Smearing.h>
+#include <Grid/parallelIO/MetaData.h>
+#include <Grid/qcd/hmc/HMC_aggregate.h>
 
 #endif

lib/GridCore.h

@@ -2,11 +2,13 @@
 
     Grid physics library, www.github.com/paboyle/Grid 
 
-    Source file: ./lib/algorithms/iterative/MatrixUtils.h
+    Source file: ./lib/Grid.h
 
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+Author: azusayamaguchi <ayamaguc@YAMAKAZE.local>
+Author: paboyle <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
@@ -25,51 +27,34 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_MATRIX_UTILS_H
-#define GRID_MATRIX_UTILS_H
+//
+//  Grid.h
+//  simd
+//
+//  Created by Peter Boyle on 09/05/2014.
+//  Copyright (c) 2014 University of Edinburgh. All rights reserved.
+//
 
-namespace Grid {
+#ifndef GRID_BASE_H
+#define GRID_BASE_H
 
-  namespace MatrixUtils { 
+#include <Grid/GridStd.h>
 
-    template<class T> inline void Size(Matrix<T>& A,int &N,int &M){
-      N=A.size(); assert(N>0);
-      M=A[0].size();
-      for(int i=0;i<N;i++){
-	assert(A[i].size()==M);
-      }
-    }
+#include <Grid/perfmon/Timer.h>
+#include <Grid/perfmon/PerfCount.h>
+#include <Grid/log/Log.h>
+#include <Grid/allocator/AlignedAllocator.h>
+#include <Grid/simd/Simd.h>
+#include <Grid/serialisation/Serialisation.h>
+#include <Grid/threads/Threads.h>
+#include <Grid/util/Util.h>
+#include <Grid/communicator/Communicator.h> 
+#include <Grid/cartesian/Cartesian.h>    
+#include <Grid/tensors/Tensors.h>      
+#include <Grid/lattice/Lattice.h>      
+#include <Grid/cshift/Cshift.h>       
+#include <Grid/stencil/Stencil.h>      
+#include <Grid/parallelIO/BinaryIO.h>
+#include <Grid/algorithms/Algorithms.h>   
 
-    template<class T> inline void SizeSquare(Matrix<T>& A,int &N)
-    {
-      int M;
-      Size(A,N,M);
-      assert(N==M);
-    }
-
-    template<class T> inline void Fill(Matrix<T>& A,T & val)
-    { 
-      int N,M;
-      Size(A,N,M);
-      for(int i=0;i<N;i++){
-      for(int j=0;j<M;j++){
-	A[i][j]=val;
-      }}
-    }
-    template<class T> inline void Diagonal(Matrix<T>& A,T & val)
-    { 
-      int N;
-      SizeSquare(A,N);
-      for(int i=0;i<N;i++){
-	A[i][i]=val;
-      }
-    }
-    template<class T> inline void Identity(Matrix<T>& A)
-    {
-      Fill(A,0.0);
-      Diagonal(A,1.0);
-    }
-
-  };
-}
 #endif

lib/GridQCDcore.h

@@ -2,12 +2,12 @@
 
     Grid physics library, www.github.com/paboyle/Grid 
 
-    Source file: ./lib/qcd/hmc/HMC.cc
+    Source file: ./lib/Grid.h
 
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
+Author: azusayamaguchi <ayamaguc@YAMAKAZE.local>
 Author: paboyle <paboyle@ph.ed.ac.uk>
 
     This program is free software; you can redistribute it and/or modify
@@ -27,10 +27,16 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid/Grid.h>
+#ifndef GRID_QCD_CORE_H
+#define GRID_QCD_CORE_H
 
-namespace Grid{
-  namespace QCD{
+/////////////////////////
+// Core Grid QCD headers
+/////////////////////////
+#include <Grid/GridCore.h>
+#include <Grid/qcd/QCD.h>
+#include <Grid/qcd/spin/Spin.h>
+#include <Grid/qcd/utils/Utils.h>
+#include <Grid/qcd/representations/Representations.h>
 
-  }
-}
+#endif

lib/GridStd.h

@@ -0,0 +1,29 @@
+#ifndef GRID_STD_H
+#define GRID_STD_H
+
+///////////////////
+// Std C++ dependencies
+///////////////////
+#include <cassert>
+#include <complex>
+#include <vector>
+#include <string>
+#include <iostream>
+#include <iomanip>
+#include <random>
+#include <functional>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <signal.h>
+#include <ctime>
+#include <sys/time.h>
+#include <chrono>
+#include <zlib.h>
+
+///////////////////
+// Grid config
+///////////////////
+#include "Config.h"
+
+#endif /* GRID_STD_H */

lib/Grid_Eigen_Dense.h

@@ -0,0 +1,9 @@
+#pragma once
+#if defined __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+#endif
+#include <Grid/Eigen/Dense>
+#if defined __GNUC__
+#pragma GCC diagnostic pop
+#endif

lib/Old/Tensor_peek.h

@@ -1,154 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/Old/Tensor_peek.h
-
-    Copyright (C) 2015
-
-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 */
-#ifndef GRID_MATH_PEEK_H
-#define GRID_MATH_PEEK_H
-namespace Grid {
-
-//////////////////////////////////////////////////////////////////////////////
-// Peek on a specific index; returns a scalar in that index, tensor inherits rest
-//////////////////////////////////////////////////////////////////////////////
-// If we hit the right index, return scalar with no further recursion
-
-//template<int Level> inline ComplexF peekIndex(const ComplexF arg) { return arg;}
-//template<int Level> inline ComplexD peekIndex(const ComplexD arg) { return arg;}
-//template<int Level> inline RealF peekIndex(const RealF arg) { return arg;}
-//template<int Level> inline RealD peekIndex(const RealD arg) { return arg;}
-#if 0
-// Scalar peek, no indices
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  auto peekIndex(const iScalar<vtype> &arg) ->  iScalar<vtype> 
-{
-  return arg;
-}
-// Vector peek, one index
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  auto peekIndex(const iVector<vtype,N> &arg,int i) -> iScalar<vtype> // Index matches
-{
-  iScalar<vtype> ret;                              // return scalar
-  ret._internal = arg._internal[i];
-  return ret;
-}
-// Matrix peek, two indices
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  auto peekIndex(const iMatrix<vtype,N> &arg,int i,int j) ->  iScalar<vtype>
-{
-  iScalar<vtype> ret;                              // return scalar
-  ret._internal = arg._internal[i][j];
-  return ret;
-}
-
-/////////////
-// No match peek for scalar,vector,matrix must forward on either 0,1,2 args. Must have 9 routines with notvalue
-/////////////
-// scalar
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iScalar<vtype> &arg) -> iScalar<decltype(peekIndex<Level>(arg._internal))>
-{
-  iScalar<decltype(peekIndex<Level>(arg._internal))> ret;
-  ret._internal= peekIndex<Level>(arg._internal);
-  return ret;
-}
-template<int Level,class vtype, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iScalar<vtype> &arg,int i) ->  iScalar<decltype(peekIndex<Level>(arg._internal,i))> 
-{
-  iScalar<decltype(peekIndex<Level>(arg._internal,i))> ret;
-  ret._internal=peekIndex<Level>(arg._internal,i);
-  return ret;
-}
-template<int Level,class vtype, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iScalar<vtype> &arg,int i,int j) ->  iScalar<decltype(peekIndex<Level>(arg._internal,i,j))>
-{
-  iScalar<decltype(peekIndex<Level>(arg._internal,i,j))> ret;
-  ret._internal=peekIndex<Level>(arg._internal,i,j);
-  return ret;
-}
-// vector
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-auto peekIndex(const iVector<vtype,N> &arg) ->   iVector<decltype(peekIndex<Level>(arg._internal[0])),N>
-{
-  iVector<decltype(peekIndex<Level>(arg._internal[0])),N> ret;
-  for(int ii=0;ii<N;ii++){
-    ret._internal[ii]=peekIndex<Level>(arg._internal[ii]);
-  }
-  return ret;
-}
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iVector<vtype,N> &arg,int i) ->  iVector<decltype(peekIndex<Level>(arg._internal[0],i)),N>
-{
-  iVector<decltype(peekIndex<Level>(arg._internal[0],i)),N> ret;
-  for(int ii=0;ii<N;ii++){
-    ret._internal[ii]=peekIndex<Level>(arg._internal[ii],i);
-  }
-  return ret;
-}
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iVector<vtype,N> &arg,int i,int j) ->  iVector<decltype(peekIndex<Level>(arg._internal[0],i,j)),N> 
-{
-  iVector<decltype(peekIndex<Level>(arg._internal[0],i,j)),N> ret;
-  for(int ii=0;ii<N;ii++){
-    ret._internal[ii]=peekIndex<Level>(arg._internal[ii],i,j);
-  }
-  return ret;
-}
-
-// matrix
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-auto peekIndex(const iMatrix<vtype,N> &arg) ->   iMatrix<decltype(peekIndex<Level>(arg._internal[0][0])),N> 
-{
-  iMatrix<decltype(peekIndex<Level>(arg._internal[0][0])),N> ret;
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    ret._internal[ii][jj]=peekIndex<Level>(arg._internal[ii][jj]);// Could avoid this because peeking a scalar is dumb
-  }}
-  return ret;
-}
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iMatrix<vtype,N> &arg,int i) ->   iMatrix<decltype(peekIndex<Level>(arg._internal[0][0],i)),N>
-{
-  iMatrix<decltype(peekIndex<Level>(arg._internal[0][0],i)),N> ret;
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    ret._internal[ii][jj]=peekIndex<Level>(arg._internal[ii][jj],i);
-  }}
-  return ret;
-}
-template<int Level,class vtype,int N, typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  auto peekIndex(const iMatrix<vtype,N> &arg,int i,int j) ->   iMatrix<decltype(peekIndex<Level>(arg._internal[0][0],i,j)),N>
-{
-  iMatrix<decltype(peekIndex<Level>(arg._internal[0][0],i,j)),N> ret;
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    ret._internal[ii][jj]=peekIndex<Level>(arg._internal[ii][jj],i,j);
-  }}
-  return ret;
-}
-#endif
-
-
-}
-#endif

lib/Old/Tensor_poke.h

@@ -1,127 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/Old/Tensor_poke.h
-
-    Copyright (C) 2015
-
-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 */
-#ifndef GRID_MATH_POKE_H
-#define GRID_MATH_POKE_H
-namespace Grid {
-
-//////////////////////////////////////////////////////////////////////////////
-// Poke a specific index; 
-//////////////////////////////////////////////////////////////////////////////
-#if 0
-// Scalar poke
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  void pokeIndex(iScalar<vtype> &ret, const iScalar<vtype> &arg)
-{
-  ret._internal = arg._internal;
-}
-// Vector poke, one index
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  void pokeIndex(iVector<vtype,N> &ret, const iScalar<vtype> &arg,int i)
-{
-  ret._internal[i] = arg._internal;
-}
-//Matrix poke, two indices
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel == Level >::type * =nullptr> inline 
-  void pokeIndex(iMatrix<vtype,N> &ret, const iScalar<vtype> &arg,int i,int j)
-{
-  ret._internal[i][j] = arg._internal;
-}
-
-/////////////
-// No match poke for scalar,vector,matrix must forward on either 0,1,2 args. Must have 9 routines with notvalue
-/////////////
-// scalar
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-void pokeIndex(iScalar<vtype> &ret, const iScalar<decltype(peekIndex<Level>(ret._internal))>  &arg)
-{
-  pokeIndex<Level>(ret._internal,arg._internal);
-}
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iScalar<vtype> &ret, const iScalar<decltype(peekIndex<Level>(ret._internal,0))> &arg, int i)
-		 
-{
-  pokeIndex<Level>(ret._internal,arg._internal,i);
-}
-template<int Level,class vtype,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iScalar<vtype> &ret, const iScalar<decltype(peekIndex<Level>(ret._internal,0,0))> &arg,int i,int j)
-{
-  pokeIndex<Level>(ret._internal,arg._internal,i,j);
-}
-
-// Vector
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iVector<vtype,N> &ret, iVector<decltype(peekIndex<Level>(ret._internal)),N>  &arg)
-{
-  for(int ii=0;ii<N;ii++){
-    pokeIndex<Level>(ret._internal[ii],arg._internal[ii]);
-  }
-}
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iVector<vtype,N> &ret, const iVector<decltype(peekIndex<Level>(ret._internal,0)),N> &arg,int i)
-{
-  for(int ii=0;ii<N;ii++){
-    pokeIndex<Level>(ret._internal[ii],arg._internal[ii],i);
-  }
-}
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iVector<vtype,N> &ret, const iVector<decltype(peekIndex<Level>(ret._internal,0,0)),N> &arg,int i,int j)
-{
-  for(int ii=0;ii<N;ii++){
-    pokeIndex<Level>(ret._internal[ii],arg._internal[ii],i,j);
-  }
-}
-
-// Matrix
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iMatrix<vtype,N> &ret, const iMatrix<decltype(peekIndex<Level>(ret._internal)),N> &arg)		 
-{
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    pokeIndex<Level>(ret._internal[ii][jj],arg._internal[ii][jj]);
-  }}
-}
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iMatrix<vtype,N> &ret, const iMatrix<decltype(peekIndex<Level>(ret._internal,0)),N> &arg,int i)
-{
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    pokeIndex<Level>(ret._internal[ii][jj],arg._internal[ii][jj],i);
-  }}
-}
-template<int Level,class vtype,int N,typename std::enable_if< iScalar<vtype>::TensorLevel != Level >::type * =nullptr> inline 
-  void pokeIndex(iMatrix<vtype,N> &ret, const iMatrix<decltype(peekIndex<Level>(ret._internal,0,0)),N> &arg, int i,int j)
-{
-  for(int ii=0;ii<N;ii++){
-  for(int jj=0;jj<N;jj++){
-    pokeIndex<Level>(ret._internal[ii][jj],arg._internal[ii][jj],i,j);
-  }}
-}
-#endif
-
-}
-#endif

lib/algorithms/Algorithms.h

@@ -39,19 +39,17 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/approx/MultiShiftFunction.h>
 
 #include <Grid/algorithms/iterative/ConjugateGradient.h>
-#include <Grid/algorithms/iterative/ConjugateGradientShifted.h>
 #include <Grid/algorithms/iterative/ConjugateResidual.h>
 #include <Grid/algorithms/iterative/NormalEquations.h>
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
-
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
 
 // Lanczos support
-#include <Grid/algorithms/iterative/MatrixUtils.h>
+//#include <Grid/algorithms/iterative/MatrixUtils.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
-
 #include <Grid/algorithms/CoarsenedMatrix.h>
+#include <Grid/algorithms/FFT.h>
 
 // Eigen/lanczos
 // EigCg

lib/algorithms/CoarsenedMatrix.h

@@ -267,8 +267,7 @@ namespace Grid {
       SimpleCompressor<siteVector> compressor;
       Stencil.HaloExchange(in,compressor);
 
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<Grid()->oSites();ss++){
+      parallel_for(int ss=0;ss<Grid()->oSites();ss++){
         siteVector res = zero;
 	siteVector nbr;
 	int ptype;
@@ -380,8 +379,7 @@ PARALLEL_FOR_LOOP
 	  Subspace.ProjectToSubspace(oProj,oblock);
 	  //	  blockProject(iProj,iblock,Subspace.subspace);
 	  //	  blockProject(oProj,oblock,Subspace.subspace);
-PARALLEL_FOR_LOOP
-	  for(int ss=0;ss<Grid()->oSites();ss++){
+	  parallel_for(int ss=0;ss<Grid()->oSites();ss++){
 	    for(int j=0;j<nbasis;j++){
 	      if( disp!= 0 ) {
 		A[p]._odata[ss](j,i) = oProj._odata[ss](j);
@@ -427,7 +425,7 @@ PARALLEL_FOR_LOOP
 	A[p]=zero;
       }
 
-      GridParallelRNG  RNG(Grid()); RNG.SeedRandomDevice();
+      GridParallelRNG  RNG(Grid()); RNG.SeedFixedIntegers(std::vector<int>({55,72,19,17,34}));
       Lattice<iScalar<CComplex> > val(Grid()); random(RNG,val);
 
       Complex one(1.0);

lib/algorithms/LinearOperator.h

@@ -235,7 +235,7 @@ namespace Grid {
 	Field tmp(in._grid);
 
 	_Mat.MeooeDag(in,tmp);
-	_Mat.MooeeInvDag(tmp,out);
+        _Mat.MooeeInvDag(tmp,out);
 	_Mat.MeooeDag(out,tmp);
 
 	_Mat.MooeeDag(in,out);

lib/algorithms/approx/Chebyshev.h

@@ -197,8 +197,9 @@ namespace Grid {
     void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 
       GridBase *grid=in._grid;
-//std::cout << "Chevyshef(): in._grid="<<in._grid<<std::endl;
-//<<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
+
+      // std::cout << "Chevyshef(): in._grid="<<in._grid<<std::endl;
+      //std::cout <<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
 
       int vol=grid->gSites();
 

lib/algorithms/approx/MultiShiftFunction.cc

@@ -25,7 +25,7 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid/Grid.h>
+#include <Grid/GridCore.h>
 
 namespace Grid {
 double MultiShiftFunction::approx(double x)

lib/algorithms/approx/Remez.h

@@ -16,7 +16,7 @@
 #define INCLUDED_ALG_REMEZ_H
 
 #include <stddef.h>
-#include <Config.h>
+#include <Grid/GridStd.h>
 
 #ifdef HAVE_LIBGMP
 #include "bigfloat.h"

lib/algorithms/iterative/BlockConjugateGradient.h

@@ -0,0 +1,366 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/BlockConjugateGradient.h
+
+Copyright (C) 2017
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+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 */
+#ifndef GRID_BLOCK_CONJUGATE_GRADIENT_H
+#define GRID_BLOCK_CONJUGATE_GRADIENT_H
+
+
+namespace Grid {
+
+//////////////////////////////////////////////////////////////////////////
+// Block conjugate gradient. Dimension zero should be the block direction
+//////////////////////////////////////////////////////////////////////////
+template <class Field>
+class BlockConjugateGradient : public OperatorFunction<Field> {
+ public:
+
+  typedef typename Field::scalar_type scomplex;
+
+  const int blockDim = 0;
+
+  int Nblock;
+  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+                           // Defaults true.
+  RealD Tolerance;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  
+  BlockConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
+    : Tolerance(tol),
+    MaxIterations(maxit),
+    ErrorOnNoConverge(err_on_no_conv){};
+
+void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
+{
+  int Orthog = 0; // First dimension is block dim
+  Nblock = Src._grid->_fdimensions[Orthog];
+
+  std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
+
+  Psi.checkerboard = Src.checkerboard;
+  conformable(Psi, Src);
+
+  Field P(Src);
+  Field AP(Src);
+  Field R(Src);
+  
+  Eigen::MatrixXcd m_pAp    = Eigen::MatrixXcd::Identity(Nblock,Nblock);
+  Eigen::MatrixXcd m_pAp_inv= Eigen::MatrixXcd::Identity(Nblock,Nblock);
+  Eigen::MatrixXcd m_rr     = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  Eigen::MatrixXcd m_rr_inv = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+  Eigen::MatrixXcd m_alpha      = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  Eigen::MatrixXcd m_beta   = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+
+  // Initial residual computation & set up
+  std::vector<RealD> residuals(Nblock);
+  std::vector<RealD> ssq(Nblock);
+
+  sliceNorm(ssq,Src,Orthog);
+  RealD sssum=0;
+  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
+
+  sliceNorm(residuals,Src,Orthog);
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  sliceNorm(residuals,Psi,Orthog);
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  // Initial search dir is guess
+  Linop.HermOp(Psi, AP);
+  
+
+  /************************************************************************
+   * Block conjugate gradient (Stephen Pickles, thesis 1995, pp 71, O Leary 1980)
+   ************************************************************************
+   * O'Leary : R = B - A X
+   * O'Leary : P = M R ; preconditioner M = 1
+   * O'Leary : alpha = PAP^{-1} RMR
+   * O'Leary : beta  = RMR^{-1}_old RMR_new
+   * O'Leary : X=X+Palpha
+   * O'Leary : R_new=R_old-AP alpha
+   * O'Leary : P=MR_new+P beta
+   */
+
+  R = Src - AP;  
+  P = R;
+  sliceInnerProductMatrix(m_rr,R,R,Orthog);
+
+  GridStopWatch sliceInnerTimer;
+  GridStopWatch sliceMaddTimer;
+  GridStopWatch MatrixTimer;
+  GridStopWatch SolverTimer;
+  SolverTimer.Start();
+
+  int k;
+  for (k = 1; k <= MaxIterations; k++) {
+
+    RealD rrsum=0;
+    for(int b=0;b<Nblock;b++) rrsum+=real(m_rr(b,b));
+
+    std::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
+	      <<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
+
+    MatrixTimer.Start();
+    Linop.HermOp(P, AP);
+    MatrixTimer.Stop();
+
+    // Alpha
+    sliceInnerTimer.Start();
+    sliceInnerProductMatrix(m_pAp,P,AP,Orthog);
+    sliceInnerTimer.Stop();
+    m_pAp_inv = m_pAp.inverse();
+    m_alpha   = m_pAp_inv * m_rr ;
+
+    // Psi, R update
+    sliceMaddTimer.Start();
+    sliceMaddMatrix(Psi,m_alpha, P,Psi,Orthog);     // add alpha *  P to psi
+    sliceMaddMatrix(R  ,m_alpha,AP,  R,Orthog,-1.0);// sub alpha * AP to resid
+    sliceMaddTimer.Stop();
+
+    // Beta
+    m_rr_inv = m_rr.inverse();
+    sliceInnerTimer.Start();
+    sliceInnerProductMatrix(m_rr,R,R,Orthog);
+    sliceInnerTimer.Stop();
+    m_beta = m_rr_inv *m_rr;
+
+    // Search update
+    sliceMaddTimer.Start();
+    sliceMaddMatrix(AP,m_beta,P,R,Orthog);
+    sliceMaddTimer.Stop();
+    P= AP;
+
+    /*********************
+     * convergence monitor
+     *********************
+     */
+    RealD max_resid=0;
+    for(int b=0;b<Nblock;b++){
+      RealD rr = real(m_rr(b,b))/ssq[b];
+      if ( rr > max_resid ) max_resid = rr;
+    }
+    
+    if ( max_resid < Tolerance*Tolerance ) { 
+
+      SolverTimer.Stop();
+
+      std::cout << GridLogMessage<<"BlockCG converged in "<<k<<" iterations"<<std::endl;
+      for(int b=0;b<Nblock;b++){
+	std::cout << GridLogMessage<< "\t\tblock "<<b<<" resid "<< std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
+      }
+      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
+
+      Linop.HermOp(Psi, AP);
+      AP = AP-Src;
+      std::cout << GridLogMessage <<"\tTrue residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
+
+      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
+      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
+      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
+	    
+      IterationsToComplete = k;
+      return;
+    }
+
+  }
+  std::cout << GridLogMessage << "BlockConjugateGradient did NOT converge" << std::endl;
+
+  if (ErrorOnNoConverge) assert(0);
+  IterationsToComplete = k;
+}
+};
+
+
+//////////////////////////////////////////////////////////////////////////
+// multiRHS conjugate gradient. Dimension zero should be the block direction
+//////////////////////////////////////////////////////////////////////////
+template <class Field>
+class MultiRHSConjugateGradient : public OperatorFunction<Field> {
+ public:
+
+  typedef typename Field::scalar_type scomplex;
+
+  const int blockDim = 0;
+
+  int Nblock;
+  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+                           // Defaults true.
+  RealD Tolerance;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  
+   MultiRHSConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
+    : Tolerance(tol),
+    MaxIterations(maxit),
+    ErrorOnNoConverge(err_on_no_conv){};
+
+void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi) 
+{
+  int Orthog = 0; // First dimension is block dim
+  Nblock = Src._grid->_fdimensions[Orthog];
+
+  std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
+
+  Psi.checkerboard = Src.checkerboard;
+  conformable(Psi, Src);
+
+  Field P(Src);
+  Field AP(Src);
+  Field R(Src);
+  
+  std::vector<ComplexD> v_pAp(Nblock);
+  std::vector<RealD> v_rr (Nblock);
+  std::vector<RealD> v_rr_inv(Nblock);
+  std::vector<RealD> v_alpha(Nblock);
+  std::vector<RealD> v_beta(Nblock);
+
+  // Initial residual computation & set up
+  std::vector<RealD> residuals(Nblock);
+  std::vector<RealD> ssq(Nblock);
+
+  sliceNorm(ssq,Src,Orthog);
+  RealD sssum=0;
+  for(int b=0;b<Nblock;b++) sssum+=ssq[b];
+
+  sliceNorm(residuals,Src,Orthog);
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  sliceNorm(residuals,Psi,Orthog);
+  for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
+
+  // Initial search dir is guess
+  Linop.HermOp(Psi, AP);
+
+  R = Src - AP;  
+  P = R;
+  sliceNorm(v_rr,R,Orthog);
+
+  GridStopWatch sliceInnerTimer;
+  GridStopWatch sliceMaddTimer;
+  GridStopWatch sliceNormTimer;
+  GridStopWatch MatrixTimer;
+  GridStopWatch SolverTimer;
+
+  SolverTimer.Start();
+  int k;
+  for (k = 1; k <= MaxIterations; k++) {
+
+    RealD rrsum=0;
+    for(int b=0;b<Nblock;b++) rrsum+=real(v_rr[b]);
+
+    std::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
+	      <<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
+
+    MatrixTimer.Start();
+    Linop.HermOp(P, AP);
+    MatrixTimer.Stop();
+
+    // Alpha
+    //    sliceInnerProductVectorTest(v_pAp_test,P,AP,Orthog);
+    sliceInnerTimer.Start();
+    sliceInnerProductVector(v_pAp,P,AP,Orthog);
+    sliceInnerTimer.Stop();
+    for(int b=0;b<Nblock;b++){
+      //      std::cout << " "<< v_pAp[b]<<" "<< v_pAp_test[b]<<std::endl;
+      v_alpha[b] = v_rr[b]/real(v_pAp[b]);
+    }
+
+    // Psi, R update
+    sliceMaddTimer.Start();
+    sliceMaddVector(Psi,v_alpha, P,Psi,Orthog);     // add alpha *  P to psi
+    sliceMaddVector(R  ,v_alpha,AP,  R,Orthog,-1.0);// sub alpha * AP to resid
+    sliceMaddTimer.Stop();
+
+    // Beta
+    for(int b=0;b<Nblock;b++){
+      v_rr_inv[b] = 1.0/v_rr[b];
+    }
+    sliceNormTimer.Start();
+    sliceNorm(v_rr,R,Orthog);
+    sliceNormTimer.Stop();
+    for(int b=0;b<Nblock;b++){
+      v_beta[b] = v_rr_inv[b] *v_rr[b];
+    }
+
+    // Search update
+    sliceMaddTimer.Start();
+    sliceMaddVector(P,v_beta,P,R,Orthog);
+    sliceMaddTimer.Stop();
+
+    /*********************
+     * convergence monitor
+     *********************
+     */
+    RealD max_resid=0;
+    for(int b=0;b<Nblock;b++){
+      RealD rr = v_rr[b]/ssq[b];
+      if ( rr > max_resid ) max_resid = rr;
+    }
+    
+    if ( max_resid < Tolerance*Tolerance ) { 
+
+      SolverTimer.Stop();
+
+      std::cout << GridLogMessage<<"MultiRHS solver converged in " <<k<<" iterations"<<std::endl;
+      for(int b=0;b<Nblock;b++){
+	std::cout << GridLogMessage<< "\t\tBlock "<<b<<" resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
+      }
+      std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
+
+      Linop.HermOp(Psi, AP);
+      AP = AP-Src;
+      std::cout <<GridLogMessage << "\tTrue residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
+
+      std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
+      std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed()     <<std::endl;
+      std::cout << GridLogMessage << "\tInnerProd  " << sliceInnerTimer.Elapsed() <<std::endl;
+      std::cout << GridLogMessage << "\tNorm       " << sliceNormTimer.Elapsed() <<std::endl;
+      std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed()  <<std::endl;
+
+
+      IterationsToComplete = k;
+      return;
+    }
+
+  }
+  std::cout << GridLogMessage << "MultiRHSConjugateGradient did NOT converge" << std::endl;
+
+  if (ErrorOnNoConverge) assert(0);
+  IterationsToComplete = k;
+}
+};
+
+
+
+}
+#endif

lib/algorithms/iterative/ConjugateGradient.h

@@ -45,6 +45,8 @@ class ConjugateGradient : public OperatorFunction<Field> {
                            // Defaults true.
   RealD Tolerance;
   Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  
   ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
       : Tolerance(tol),
         MaxIterations(maxit),
@@ -76,18 +78,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
     cp = a;
     ssq = norm2(src);
 
-    std::cout << GridLogIterative << std::setprecision(4)
-              << "ConjugateGradient: guess " << guess << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4)
-              << "ConjugateGradient:   src " << ssq << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4)
-              << "ConjugateGradient:    mp " << d << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4)
-              << "ConjugateGradient:   mmp " << b << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4)
-              << "ConjugateGradient:  cp,r " << cp << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4)
-              << "ConjugateGradient:     p " << a << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: guess " << guess << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:   src " << ssq << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:    mp " << d << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:   mmp " << b << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:  cp,r " << cp << std::endl;
+    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:     p " << a << std::endl;
 
     RealD rsq = Tolerance * Tolerance * ssq;
 
@@ -97,8 +93,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
     }
 
     std::cout << GridLogIterative << std::setprecision(4)
-              << "ConjugateGradient: k=0 residual " << cp << " target " << rsq
-              << std::endl;
+              << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
 
     GridStopWatch LinalgTimer;
     GridStopWatch MatrixTimer;
@@ -128,8 +123,11 @@ class ConjugateGradient : public OperatorFunction<Field> {
       p = p * b + r;
 
       LinalgTimer.Stop();
+
       std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
                 << " residual " << cp << " target " << rsq << std::endl;
+      std::cout << GridLogDebug << "a = "<< a << " b_pred = "<< b_pred << "  b = "<< b << std::endl;
+      std::cout << GridLogDebug << "qq = "<< qq << " d = "<< d << "  c = "<< c << std::endl;
 
       // Stopping condition
       if (cp <= rsq) {
@@ -137,31 +135,33 @@ class ConjugateGradient : public OperatorFunction<Field> {
         Linop.HermOpAndNorm(psi, mmp, d, qq);
         p = mmp - src;
 
-        RealD mmpnorm = sqrt(norm2(mmp));
-        RealD psinorm = sqrt(norm2(psi));
         RealD srcnorm = sqrt(norm2(src));
         RealD resnorm = sqrt(norm2(p));
         RealD true_residual = resnorm / srcnorm;
 
-        std::cout << GridLogMessage
-                  << "ConjugateGradient: Converged on iteration " << k << std::endl;
-        std::cout << GridLogMessage << "Computed residual " << sqrt(cp / ssq)
-                  << " true residual " << true_residual << " target "
-                  << Tolerance << std::endl;
-        std::cout << GridLogMessage << "Time elapsed: Iterations "
-                  << SolverTimer.Elapsed() << " Matrix  "
-                  << MatrixTimer.Elapsed() << " Linalg "
-                  << LinalgTimer.Elapsed();
-        std::cout << std::endl;
+        std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k << std::endl;
+        std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+	std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
+	std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "Time breakdown "<<std::endl;
+	std::cout << GridLogMessage << "\tElapsed    " << SolverTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tMatrix     " << MatrixTimer.Elapsed() <<std::endl;
+	std::cout << GridLogMessage << "\tLinalg     " << LinalgTimer.Elapsed() <<std::endl;
 
         if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
 
+	IterationsToComplete = k;	
+
         return;
       }
     }
     std::cout << GridLogMessage << "ConjugateGradient did NOT converge"
               << std::endl;
+
     if (ErrorOnNoConverge) assert(0);
+    IterationsToComplete = k;
+
   }
 };
 }

lib/algorithms/iterative/ConjugateGradientMixedPrec.h

@@ -35,6 +35,7 @@ namespace Grid {
   class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
   public:                                                
     RealD   Tolerance;
+    RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
     Integer MaxInnerIterations;
     Integer MaxOuterIterations;
     GridBase* SinglePrecGrid; //Grid for single-precision fields
@@ -42,12 +43,16 @@ namespace Grid {
     LinearOperatorBase<FieldF> &Linop_f;
     LinearOperatorBase<FieldD> &Linop_d;
 
+    Integer TotalInnerIterations; //Number of inner CG iterations
+    Integer TotalOuterIterations; //Number of restarts
+    Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
+
     //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
     LinearFunction<FieldF> *guesser;
     
     MixedPrecisionConjugateGradient(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d) :
       Linop_f(_Linop_f), Linop_d(_Linop_d),
-      Tolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
+      Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
       OuterLoopNormMult(100.), guesser(NULL){ };
 
     void useGuesser(LinearFunction<FieldF> &g){
@@ -55,9 +60,8 @@ namespace Grid {
     }
   
     void operator() (const FieldD &src_d_in, FieldD &sol_d){
-	(*this)(src_d_in,sol_d,NULL);
-    }
-    void operator() (const FieldD &src_d_in, FieldD &sol_d, RealD *shift){
+      TotalInnerIterations = 0;
+	
       GridStopWatch TotalTimer;
       TotalTimer.Start();
     
@@ -77,7 +81,7 @@ namespace Grid {
       FieldD src_d(DoublePrecGrid);
       src_d = src_d_in; //source for next inner iteration, computed from residual during operation
     
-      RealD inner_tol = Tolerance;
+      RealD inner_tol = InnerTolerance;
     
       FieldF src_f(SinglePrecGrid);
       src_f.checkerboard = cb;
@@ -85,17 +89,18 @@ namespace Grid {
       FieldF sol_f(SinglePrecGrid);
       sol_f.checkerboard = cb;
     
-      ConjugateGradientShifted<FieldF> CG_f(inner_tol, MaxInnerIterations);
+      ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
       CG_f.ErrorOnNoConverge = false;
 
       GridStopWatch InnerCGtimer;
 
       GridStopWatch PrecChangeTimer;
     
-      for(Integer outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+      Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
+      
+      for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
 	//Compute double precision rsd and also new RHS vector.
 	Linop_d.HermOp(sol_d, tmp_d);
-	if(shift) axpy(tmp_d,*shift,sol_d,tmp_d);
 	RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
       
 	std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
@@ -119,8 +124,9 @@ namespace Grid {
 	//Inner CG
 	CG_f.Tolerance = inner_tol;
 	InnerCGtimer.Start();
-	CG_f(Linop_f, src_f, sol_f,shift);
+	CG_f(Linop_f, src_f, sol_f);
 	InnerCGtimer.Stop();
+	TotalInnerIterations += CG_f.IterationsToComplete;
       
 	//Convert sol back to double and add to double prec solution
 	PrecChangeTimer.Start();
@@ -133,11 +139,13 @@ namespace Grid {
       //Final trial CG
       std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting final patch-up double-precision solve"<<std::endl;
     
-      ConjugateGradientShifted<FieldD> CG_d(Tolerance, MaxInnerIterations);
-      CG_d(Linop_d, src_d_in, sol_d,shift);
+      ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
+      CG_d(Linop_d, src_d_in, sol_d);
+      TotalFinalStepIterations = CG_d.IterationsToComplete;
 
       TotalTimer.Stop();
-      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
     }
   };
 

lib/algorithms/iterative/ConjugateGradientMultiShift.h

@@ -45,7 +45,6 @@ public:
     Integer MaxIterations;
     int verbose;
     MultiShiftFunction shifts;
-    int iter;
 
     ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) : 
 	MaxIterations(maxit),
@@ -61,7 +60,6 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
   std::vector<Field> results(nshift,grid);
   (*this)(Linop,src,results,psi);
 }
-
 void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &results, Field &psi)
 {
   int nshift = shifts.order;
@@ -107,12 +105,11 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   RealD a,b,c,d;
   RealD cp,bp,qq; //prev
   
-  int cb=src.checkerboard;
   // Matrix mult fields
   Field r(grid);
-  Field p(grid); p.checkerboard = src.checkerboard;
+  Field p(grid);
   Field tmp(grid);
-  Field mmp(grid);mmp.checkerboard = src.checkerboard;
+  Field mmp(grid);
   
   // Check lightest mass
   for(int s=0;s<nshift;s++){
@@ -135,9 +132,6 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
   p=src;
   
   //MdagM+m[0]
-  std::cout << "p.checkerboard " << p.checkerboard
-  << "mmp.checkerboard " << mmp.checkerboard << std::endl;
-
   Linop.HermOpAndNorm(p,mmp,d,qq);
   axpy(mmp,mass[0],p,mmp);
   RealD rn = norm2(p);
@@ -275,7 +269,6 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
 	RealD cn = norm2(src);
 	std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(rn/cn)<<std::endl;
       }
-      iter = k;
       return;
     }
   }

lib/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h

@@ -1,404 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h
-
-    Copyright (C) 2015
-
-Author: Chulwoo Jung <chulwoo@quark.phy.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_MIXED_PREC_H
-#define GRID_CONJUGATE_GRADIENT_MULTI_MIXED_PREC_H
-
-namespace 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 MixedPrecisionConjugateGradientMultiShift : public LinearFunction<FieldD> {
-  public:                                                
-//    RealD   Tolerance;
-    Integer MaxInnerIterations;
-    Integer MaxOuterIterations;
-    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;
-    MultiShiftFunction shifts;
-    Integer iter;
-
-    //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
-//    LinearFunction<FieldF> *guesser;
-    
-    MixedPrecisionConjugateGradientMultiShift(GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d, 
-Integer maxinnerit,	MultiShiftFunction &_shifts ) :
-      Linop_f(_Linop_f), Linop_d(_Linop_d),
-      MaxInnerIterations(maxinnerit), SinglePrecGrid(_sp_grid),
-      OuterLoopNormMult(100.), shifts(_shifts) {};
-
-  
-    void operator() (const FieldD &src_d_in, FieldD &sol_d){
-	assert(0); // not yet implemented
-    }
-    void operator() (const FieldD &src_d_in, std::vector<FieldD> &sol_d){
-      GridStopWatch TotalTimer;
-      TotalTimer.Start();
-    
-      int cb = src_d_in.checkerboard;
-
-      int nshift = shifts.order;
-      assert(nshift == sol_d.size());
-      for(int i=0;i<nshift;i++) sol_d[i].checkerboard = cb;
-    
-      RealD src_norm = norm2(src_d_in);
-//      RealD stop = src_norm * Tolerance*Tolerance;
-
-      GridBase* DoublePrecGrid = src_d_in._grid;
-      FieldD tmp_d(DoublePrecGrid); tmp_d.checkerboard = cb;
-    
-      FieldD tmp2_d(DoublePrecGrid); tmp2_d.checkerboard = cb;
-    
-      FieldD src_d(DoublePrecGrid);
-      src_d = src_d_in; //source for next inner iteration, computed from residual during operation
-    
-//      RealD inner_tol = Tolerance;
-  	FieldD psi_d(DoublePrecGrid);psi_d.checkerboard = cb;
-    
-      FieldF src_f(SinglePrecGrid);
-      src_f.checkerboard = cb;
-    
-      std::vector<FieldF> sol_f(nshift,SinglePrecGrid);
-      for(int i=0;i<nshift;i++) sol_f[i].checkerboard = cb;
-    
-//      ConjugateGradientShifted<FieldF> CG_f(inner_tol, MaxInnerIterations);
-      ConjugateGradientMultiShift<FieldF> MSCG(MaxInnerIterations,shifts);
-//      CG_f.ErrorOnNoConverge = false;
-
-      GridStopWatch InnerCGtimer;
-
-      GridStopWatch PrecChangeTimer;
-    
-{
-//	std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
-
-//	if(norm < OuterLoopNormMult * stop){
-//	  std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
-//	  break;
-//	}
-//	while(norm * inner_tol * inner_tol < stop) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
-
-	PrecChangeTimer.Start();
-	precisionChange(src_f, src_d);
-	PrecChangeTimer.Stop();
-      
-//	zeroit(sol_f);
-
-
-	//Inner CG
-	InnerCGtimer.Start();
-  int if_relup = 0;
-#if 0
-        MSCG(Linop_f,src_f,sol_f);
-#else
-{
-  
-  GridBase *grid = SinglePrecGrid;
-  
-  ////////////////////////////////////////////////////////////////////////
-  // 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.);
-  std::vector<FieldF>   ps(nshift,grid);// Search directions
-
-  assert(sol_f.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  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
-  
-  int cb=src_f.checkerboard;
-  // Matrix mult fields
-  FieldF r(grid); r.checkerboard = src_f.checkerboard;
-  FieldF p(grid); p.checkerboard = src_f.checkerboard;
-  FieldF tmp(grid); tmp.checkerboard = src_f.checkerboard;
-  FieldF mmp(grid);mmp.checkerboard = src_f.checkerboard;
-  FieldF psi(grid);psi.checkerboard = src_f.checkerboard;
-    std::cout.precision(12);
-    std::cout<<GridLogMessage<<"norm2(psi_d)= "<<norm2(psi_d)<<std::endl;
-    std::cout<<GridLogMessage<<"norm2(psi)= "<<norm2(psi)<<std::endl;
-  
-  
-  // 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_f);
-  Real c_relup = cp;
-  for(int s=0;s<nshift;s++){
-    rsq[s] = cp * mresidual[s] * mresidual[s];
-    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientMultiShift: shift "<<s
-	     <<" target resid "<<rsq[s]<<std::endl;
-    ps[s] = src_f;
-  }
-  // r and p for primary
-  r=src_f;
-  p=src_f;
-  
-  //MdagM+m[0]
-  std::cout << "p.checkerboard " << p.checkerboard
-  << "mmp.checkerboard " << mmp.checkerboard << std::endl;
-
-  Linop_f.HermOpAndNorm(p,mmp,d,qq);
-  axpy(mmp,mass[0],p,mmp);
-  RealD rn = norm2(p);
-  d += rn*mass[0];
-  
-  // have verified that inner product of 
-  // p and mmp is equal to d after this since
-  // the d computation is tricky
-  //  qq = real(innerProduct(p,mmp));
-  //  std::cout<<GridLogMessage << "debug equal ?  qq "<<qq<<" d "<< d<<std::endl;
-  
-  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,b,mmp,r);
-  
- axpby(psi,0.,-bs[0],src_f,src_f);
-  for(int s=0;s<nshift;s++) {
-    axpby(sol_f[s],0.,-bs[s]*alpha[s],src_f,src_f);
-  }
-  
-  
-  // Iteration loop
-  int k;
- // inefficient zeroing, please replace!
-//  RealD sol_norm = axpy_norm(sol_d[0],-1.,sol_d[0],sol_d[0]);
-  zeroit(sol_d[0]);
-  std::cout<<GridLogMessage<<"norm(sol_d[0])= "<<norm2(sol_d[0])<<std::endl;
-  
-
-  int all_converged = 1;
-	RealD tmp1,tmp2;
-  for (k=1;k<=MaxOuterIterations;k++){
-    
-    a = c /cp;
-    axpy(p,a,p,r);
-    
-    // Note to self - direction ps is iterated seperately
-    // for each shift. Does not appear to have any scope
-    // for avoiding linear algebra in "single" case.
-    // 
-    // However SAME r is used. Could load "r" and update
-    // ALL ps[s]. 2/3 Bandwidth saving
-    // New Kernel: Load r, vector of coeffs, vector of pointers ps
-    for(int s=0;s<nshift;s++){
-      if ( ! converged[s] ) { 
-	if (s==0){
-	  axpy(ps[s],a,ps[s],r);
-	} else{
-	  RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
-	  axpby(ps[s],z[s][iz],as,r,ps[s]);
-	}
-      }
-    }
-    
-    cp=c;
-    
-    Linop_f.HermOpAndNorm(p,mmp,d,qq);
-    axpy(mmp,mass[0],p,mmp);
-    RealD rn = norm2(p);
-    d += rn*mass[0];
-    
-    bp=b;
-    b=-cp/d;
-    
-    c=axpy_norm(r,b,mmp,r);
-
-
-    // Toggle the recurrence history
-    bs[0] = b;
-    iz = 1-iz;
-    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
-      }
-    }
-    
-    axpy(psi,-bs[0],ps[0],psi);
-    for(int s=0;s<nshift;s++){
-      int ss = s;
-      // Scope for optimisation here in case of "single".
-      // Could load sol_f[0] and pull all ps[s] in.
-      //      if ( single ) ss=primary;
-      // Bandwith saving in single case is Ls * 3 -> 2+Ls, so ~ 3x saving
-      // Pipelined CG gain:
-      //
-      // New Kernel: Load r, vector of coeffs, vector of pointers ps
-      // New Kernel: Load sol_f[0], vector of coeffs, vector of pointers ps
-      // If can predict the coefficient bs then we can fuse these and avoid write reread cyce
-      //  on ps[s].
-      // Before:  3 x npole  + 3 x npole
-      // After :  2 x npole (ps[s])        => 3x speed up of multishift CG.
-      
-      if( (!converged[s]) ) { 
-	axpy(sol_f[ss],-bs[s]*alpha[s],ps[s],sol_f[ss]);
-      }
-    }
-
-
-    if (k%MaxInnerIterations==0){
-//    if (c < 1e-4*c_relup){
-       RealD c_f=c;
-       precisionChange(tmp_d,psi);
-       RealD sol_norm =axpy_norm (psi_d,1.,tmp_d,psi_d);
-       tmp1 = norm2(psi);
-       zeroit(psi);
-       tmp2 = norm2(psi);
-       std::cout<<GridLogMessage<<"k= "<<k<<" norm2(sol)= "<<sol_norm<<" "<<tmp1<<" "<<tmp2<<std::endl;
-//       precisionChange(sol_d[0],sol_f[0]);
-       Linop_d.HermOpAndNorm(psi_d,tmp_d,tmp1,tmp2);
-       axpy(tmp2_d,mass[0],psi_d,tmp_d);
-       axpy(tmp_d,-1.,tmp2_d,src_d);
-       precisionChange(r,tmp_d);
-	c_relup = norm2(r);
-       std::cout<<GridLogMessage<<"k= "<<k<<" norm2(r)= "<<c<<" "<<c_relup<<" "<<c_f<<std::endl;
-	if_relup=1;
-    }
-    
-    // Convergence checks
-  all_converged=1;
-    for(int s=0;s<nshift;s++){
-      
-      if ( (!converged[s]) ){
-	
-	RealD css  = c * z[s][iz]* z[s][iz];
-	
-	if(css<rsq[s]){
-	  if ( ! converged[s] )
-	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShift k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
-	      converged[s]=1;
-	} else {
-		if (k%MaxInnerIterations==0)
-	    std::cout<<GridLogMessage<<"ConjugateGradientMultiShift k="<<k<<" Shift "<<s<<" has not converged "<<css<<"<"<<rsq[s]<<std::endl;
-	  all_converged=0;
-	}
-
-      }
-    }
-    
-#if 0
-    if ( all_converged ){
-      std::cout<<GridLogMessage<< "CGMultiShift: All shifts have converged iteration "<<k<<std::endl;
-#else
-    if ( converged[0] ){
-      std::cout<<GridLogMessage<< "CGMultiShift: Shift 0 have converged iteration, terminating  "<<k<<std::endl;
-#endif
-      
-#if 1
-      for(int s=1; s < nshift; s++) { 
-	Linop_f.HermOpAndNorm(sol_f[s],mmp,d,qq);
-	axpy(tmp,mass[s],sol_f[s],mmp);
-	axpy(r,-alpha[s],src_f,tmp);
-	RealD rn = norm2(r);
-	RealD cn = norm2(src_f);
-	std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(rn/cn)<<std::endl;
-      }
-#endif
-     iter = k;
-      break;
-    }
-  }
-  // ugly hack
-  if ( !all_converged )
-  std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
-//  assert(0);
-}
-	
-#endif
-	InnerCGtimer.Stop();
-      
-	//Convert sol back to double and add to double prec solution
-	PrecChangeTimer.Start();
-	sol_d[0]=psi_d;
-	for(int i=1;i<nshift;i++)precisionChange(sol_d[i], sol_f[i]);
-      std::cout<<GridLogMessage<< "CGMultiShift: Checking solutions"<<std::endl;
-      // Check answers 
-      for(int s=0; s < nshift; s++) { 
-	RealD tmp1,tmp2;
-       Linop_d.HermOpAndNorm(sol_d[s],tmp_d,tmp1,tmp2);
-       axpy(tmp2_d,shifts.poles[s],sol_d[s],tmp_d);
-       axpy(tmp_d,-1.,src_d,tmp2_d);
-	std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(norm2(tmp_d)/norm2(src_d))<<std::endl;
-      }
-	PrecChangeTimer.Stop();
-      
-}
-    
-      //Final trial CG
- //     std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting final patch-up double-precision solve"<<std::endl;
-    
-      TotalTimer.Stop();
-      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
-    }
-  };
-
-}
-
-#endif

lib/algorithms/iterative/ConjugateGradientShifted.h

@@ -1,168 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/ConjugateGradient.h
-
-    Copyright (C) 2015
-
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <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 */
-#ifndef GRID_CONJUGATE_GRADIENT_SHIFTED_H
-#define GRID_CONJUGATE_GRADIENT_SHIFTED_H
-
-namespace Grid {
-
-    /////////////////////////////////////////////////////////////
-    // Base classes for iterative processes based on operators
-    // single input vec, single output vec.
-    /////////////////////////////////////////////////////////////
-
-  template<class Field> 
-    class ConjugateGradientShifted : public OperatorFunction<Field> {
-public:                                                
-    bool ErrorOnNoConverge; //throw an assert when the CG fails to converge. Defaults true.
-    RealD   Tolerance;
-    Integer MaxIterations;
-    ConjugateGradientShifted(RealD tol,Integer maxit, bool err_on_no_conv = true) : Tolerance(tol), MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv) { 
-    };
-
-    void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi ){
-	(*this)(Linop,src,psi,NULL);
-    }
-
-    void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi, RealD *shift){
-
-      psi.checkerboard = src.checkerboard;
-      conformable(psi,src);
-
-      RealD cp,c,a,d,b,ssq,qq,b_pred;
-      
-      Field   p(src);
-      Field mmp(src);
-      Field   r(src);
-      
-      //Initial residual computation & set up
-      RealD guess = norm2(psi);
-      assert(std::isnan(guess)==0);
-
-      Linop.HermOpAndNorm(psi,mmp,d,b);
-	if(shift) axpy(mmp,*shift,psi,mmp);
-	RealD rn = norm2(psi);
-	if(shift) d += rn*(*shift);
-	RealD d2 = real(innerProduct(psi,mmp));
-	b= norm2(mmp);
-      RealD src_norm=norm2(src);
-      r= src-mmp;
-      p= r;
-      
-      a  =norm2(p);
-      cp =a;
-      ssq=norm2(src);
-
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient: guess "<<guess<<std::endl;
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:   src "<<ssq  <<std::endl;
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:    mp "<<d    <<std::endl;
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:   mmp "<<b    <<std::endl;
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:  cp,r "<<cp   <<std::endl;
-      std::cout<<GridLogIterative <<std::setprecision(4)<< "ConjugateGradient:     p "<<a    <<std::endl;
-
-      RealD rsq =  Tolerance* Tolerance*ssq;
-      
-      //Check if guess is really REALLY good :)
-      if ( cp <= rsq ) {
-	return;
-      }
-      
-      std::cout<<GridLogIterative << std::setprecision(4)<< "ConjugateGradient: k=0 residual "<<cp<<" target "<<rsq<<std::endl;
-
-      GridStopWatch LinalgTimer;
-      GridStopWatch MatrixTimer;
-      GridStopWatch SolverTimer;
-
-      SolverTimer.Start();
-      int k;
-      for (k=1;k<=MaxIterations;k++){
-	
-	c=cp;
-
-	MatrixTimer.Start();
-	Linop.HermOpAndNorm(p,mmp,d,qq);
-	MatrixTimer.Stop();
-	LinalgTimer.Start();
-	if(shift) axpy(mmp,*shift,p,mmp);
-	RealD rn = norm2(p);
-	if(shift) d += rn*(*shift);
-	RealD d2 = real(innerProduct(p,mmp));
-	qq = norm2(mmp);
-      if (k%10==1) std::cout<< std::setprecision(4)<< "d:  "<<d<<" d2= "<<d2<<std::endl;
-
-	//	RealD    qqck = norm2(mmp);
-	//	ComplexD dck  = innerProduct(p,mmp);
-      
-	a      = c/d;
-	b_pred = a*(a*qq-d)/c;
-
-	cp = axpy_norm(r,-a,mmp,r);
-	b = cp/c;
-      if (k%10==1) std::cout<< std::setprecision(4)<<"k= "<<k<<" src:  "<<src_norm<<" r= "<<cp<<std::endl;
-	
-	// Fuse these loops ; should be really easy
-	psi= a*p+psi;
-	p  = p*b+r;
-	  
-	LinalgTimer.Stop();
-	std::cout<<GridLogIterative<<"ConjugateGradient: Iteration " <<k<<" residual "<<cp<< " target "<< rsq<<std::endl;
-	
-	// Stopping condition
-	if ( cp <= rsq ) { 
-	  
-	  SolverTimer.Stop();
-	  Linop.HermOpAndNorm(psi,mmp,d,qq);
-	  if(shift) mmp = mmp + (*shift) * psi;
-	  p=mmp-src;
-	  
-	  RealD mmpnorm = sqrt(norm2(mmp));
-	  RealD psinorm = sqrt(norm2(psi));
-	  RealD srcnorm = sqrt(norm2(src));
-	  RealD resnorm = sqrt(norm2(p));
-	  RealD true_residual = resnorm/srcnorm;
-
-	  std::cout<<GridLogMessage<<"ConjugateGradient: Converged on iteration " <<k
-		   <<" computed residual "<<sqrt(cp/ssq)
-		   <<" true residual "    <<true_residual
-		   <<" target "<<Tolerance<<std::endl;
-	  std::cout<<GridLogMessage<<"Time elapsed: Total "<< SolverTimer.Elapsed() << " Matrix  "<<MatrixTimer.Elapsed() << " Linalg "<<LinalgTimer.Elapsed();
-	  std::cout<<std::endl;
-	  
-	if(ErrorOnNoConverge)
-	  assert(true_residual/Tolerance < 1000.0);
-
-	  return;
-	}
-      }
-      std::cout<<GridLogMessage<<"ConjugateGradient did NOT converge"<<std::endl;
-//      assert(0);
-    }
-  };
-}
-#endif

lib/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -30,20 +30,17 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define GRID_IRL_H
 
 #include <string.h> //memset
+
 #ifdef USE_LAPACK
-#ifdef USE_MKL
-#include<mkl_lapack.h>
-#else
 void LAPACK_dstegr(char *jobz, char *range, int *n, double *d, double *e,
                    double *vl, double *vu, int *il, int *iu, double *abstol,
                    int *m, double *w, double *z, int *ldz, int *isuppz,
                    double *work, int *lwork, int *iwork, int *liwork,
                    int *info);
-//#include <lapacke/lapacke.h>
 #endif
-#endif
-#include "DenseMatrix.h"
-#include "EigenSort.h"
+
+#include <Grid/algorithms/densematrix/DenseMatrix.h>
+#include <Grid/algorithms/iterative/EigenSort.h>
 
 namespace Grid {
 
@@ -67,13 +64,12 @@ public:
     int Np;      // Np -- Number of spare vecs in kryloc space
     int Nm;      // Nm -- total number of vectors
 
-
-    RealD OrthoTime;
-
     RealD eresid;
 
     SortEigen<Field> _sort;
 
+//    GridCartesian &_fgrid;
+
     LinearOperatorBase<Field> &_Linop;
 
     OperatorFunction<Field>   &_poly;
@@ -130,23 +126,23 @@ public:
 
       GridBase *grid = evec[0]._grid;
       Field w(grid);
-      std::cout<<GridLogMessage << "RitzMatrix "<<std::endl;
+      std::cout << "RitzMatrix "<<std::endl;
       for(int i=0;i<k;i++){
 	_poly(_Linop,evec[i],w);
-	std::cout<<GridLogMessage << "["<<i<<"] ";
+	std::cout << "["<<i<<"] ";
 	for(int j=0;j<k;j++){
 	  ComplexD in = innerProduct(evec[j],w);
 	  if ( fabs((double)i-j)>1 ) { 
 	    if (abs(in) >1.0e-9 )  { 
-	      std::cout<<GridLogMessage<<"oops"<<std::endl;
+	      std::cout<<"oops"<<std::endl;
 	      abort();
 	    } else 
-	      std::cout<<GridLogMessage << " 0 ";
+	      std::cout << " 0 ";
 	  } else { 
-	    std::cout<<GridLogMessage << " "<<in<<" ";
+	    std::cout << " "<<in<<" ";
 	  }
 	}
-	std::cout<<GridLogMessage << std::endl;
+	std::cout << std::endl;
       }
     }
 
@@ -180,10 +176,10 @@ public:
       RealD beta = normalise(w); // 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
                                  // 7. vk+1 := wk/βk+1
 
-	std::cout<<GridLogMessage << "alpha = " << zalph << " beta "<<beta<<std::endl;
+//	std::cout << "alpha = " << zalph << " beta "<<beta<<std::endl;
       const RealD tiny = 1.0e-20;
       if ( beta < tiny ) { 
-	std::cout<<GridLogMessage << " beta is tiny "<<beta<<std::endl;
+	std::cout << " beta is tiny "<<beta<<std::endl;
      }
       lmd[k] = alph;
       lme[k]  = beta;
@@ -259,7 +255,6 @@ public:
     }
 
 #ifdef USE_LAPACK
-#define LAPACK_INT long long
     void diagonalize_lapack(DenseVector<RealD>& lmd,
 		     DenseVector<RealD>& lme, 
 		     int N1,
@@ -269,7 +264,7 @@ public:
   const int size = Nm;
 //  tevals.resize(size);
 //  tevecs.resize(size);
-  LAPACK_INT NN = N1;
+  int NN = N1;
   double evals_tmp[NN];
   double evec_tmp[NN][NN];
   memset(evec_tmp[0],0,sizeof(double)*NN*NN);
@@ -283,19 +278,19 @@ public:
         if (i==j) evals_tmp[i] = lmd[i];
         if (j==(i-1)) EE[j] = lme[j];
       }
-  LAPACK_INT evals_found;
-  LAPACK_INT lwork = ( (18*NN) > (1+4*NN+NN*NN)? (18*NN):(1+4*NN+NN*NN)) ;
-  LAPACK_INT liwork =  3+NN*10 ;
-  LAPACK_INT iwork[liwork];
+  int evals_found;
+  int lwork = ( (18*NN) > (1+4*NN+NN*NN)? (18*NN):(1+4*NN+NN*NN)) ;
+  int liwork =  3+NN*10 ;
+  int iwork[liwork];
   double work[lwork];
-  LAPACK_INT isuppz[2*NN];
+  int isuppz[2*NN];
   char jobz = 'V'; // calculate evals & evecs
   char range = 'I'; // calculate all evals
   //    char range = 'A'; // calculate all evals
   char uplo = 'U'; // refer to upper half of original matrix
   char compz = 'I'; // Compute eigenvectors of tridiagonal matrix
   int ifail[NN];
-  long long info;
+  int info;
 //  int total = QMP_get_number_of_nodes();
 //  int node = QMP_get_node_number();
 //  GridBase *grid = evec[0]._grid;
@@ -303,18 +298,14 @@ public:
   int node = grid->_processor;
   int interval = (NN/total)+1;
   double vl = 0.0, vu = 0.0;
-  LAPACK_INT il = interval*node+1 , iu = interval*(node+1);
+  int il = interval*node+1 , iu = interval*(node+1);
   if (iu > NN)  iu=NN;
   double tol = 0.0;
     if (1) {
       memset(evals_tmp,0,sizeof(double)*NN);
       if ( il <= NN){
         printf("total=%d node=%d il=%d iu=%d\n",total,node,il,iu);
-#ifdef USE_MKL
-        dstegr(&jobz, &range, &NN,
-#else
         LAPACK_dstegr(&jobz, &range, &NN,
-#endif
             (double*)DD, (double*)EE,
             &vl, &vu, &il, &iu, // these four are ignored if second parameteris 'A'
             &tol, // tolerance
@@ -346,7 +337,6 @@ public:
       lmd [NN-1-i]=evals_tmp[i];
   }
 }
-#undef LAPACK_INT 
 #endif
 
 
@@ -377,14 +367,12 @@ public:
 //	diagonalize_lapack(lmd2,lme2,Nm2,Nm,Qt,grid);
 #endif
 
-      int Niter = 10000*N1;
+      int Niter = 100*N1;
       int kmin = 1;
       int kmax = N2;
       // (this should be more sophisticated)
 
-      for(int iter=0; ; ++iter){
-      if ( (iter+1)%(100*N1)==0) 
-      std::cout<<GridLogMessage << "[QL method] Not converged - iteration "<<iter+1<<"\n";
+      for(int iter=0; iter<Niter; ++iter){
 
 	// determination of 2x2 leading submatrix
 	RealD dsub = lmd[kmax-1]-lmd[kmax-2];
@@ -413,11 +401,11 @@ public:
         _sort.push(lmd3,N2);
         _sort.push(lmd2,N2);
          for(int k=0; k<N2; ++k){
-	    if (fabs(lmd2[k] - lmd3[k]) >SMALL)  std::cout<<GridLogMessage <<"lmd(qr) lmd(lapack) "<< k << ": " << lmd2[k] <<" "<< lmd3[k] <<std::endl;
-//	    if (fabs(lme2[k] - lme[k]) >SMALL)  std::cout<<GridLogMessage <<"lme(qr)-lme(lapack) "<< k << ": " << lme2[k] - lme[k] <<std::endl;
+	    if (fabs(lmd2[k] - lmd3[k]) >SMALL)  std::cout <<"lmd(qr) lmd(lapack) "<< k << ": " << lmd2[k] <<" "<< lmd3[k] <<std::endl;
+//	    if (fabs(lme2[k] - lme[k]) >SMALL)  std::cout <<"lme(qr)-lme(lapack) "<< k << ": " << lme2[k] - lme[k] <<std::endl;
 	  }
          for(int k=0; k<N1*N1; ++k){
-//	    if (fabs(Qt2[k] - Qt[k]) >SMALL)  std::cout<<GridLogMessage <<"Qt(qr)-Qt(lapack) "<< k << ": " << Qt2[k] - Qt[k] <<std::endl;
+//	    if (fabs(Qt2[k] - Qt[k]) >SMALL)  std::cout <<"Qt(qr)-Qt(lapack) "<< k << ": " << Qt2[k] - Qt[k] <<std::endl;
 	}
     }
 #endif
@@ -432,7 +420,7 @@ public:
 	  }
 	}
       }
-      std::cout<<GridLogMessage << "[QL method] Error - Too many iteration: "<<Niter<<"\n";
+      std::cout << "[QL method] Error - Too many iteration: "<<Niter<<"\n";
       abort();
     }
 
@@ -449,7 +437,6 @@ public:
 		       DenseVector<Field>& evec,
 		       int k)
     {
-      double t0=-usecond()/1e6;
       typedef typename Field::scalar_type MyComplex;
       MyComplex ip;
 
@@ -468,8 +455,6 @@ public:
 	w = w - ip * evec[j];
       }
       normalise(w);
-      t0+=usecond()/1e6;
-      OrthoTime +=t0;
     }
 
     void setUnit_Qt(int Nm, DenseVector<RealD> &Qt) {
@@ -503,10 +488,10 @@ until convergence
 	GridBase *grid = evec[0]._grid;
 	assert(grid == src._grid);
 
-	std::cout<<GridLogMessage << " -- Nk = " << Nk << " Np = "<< Np << std::endl;
-	std::cout<<GridLogMessage << " -- Nm = " << Nm << std::endl;
-	std::cout<<GridLogMessage << " -- size of eval   = " << eval.size() << std::endl;
-	std::cout<<GridLogMessage << " -- size of evec  = " << evec.size() << std::endl;
+	std::cout << " -- Nk = " << Nk << " Np = "<< Np << std::endl;
+	std::cout << " -- Nm = " << Nm << std::endl;
+	std::cout << " -- size of eval   = " << eval.size() << std::endl;
+	std::cout << " -- size of evec  = " << evec.size() << std::endl;
 	
 	assert(Nm == evec.size() && Nm == eval.size());
 	
@@ -517,7 +502,6 @@ until convergence
 	DenseVector<int>   Iconv(Nm);
 
 	DenseVector<Field>  B(Nm,grid); // waste of space replicating
-//	DenseVector<Field>  Btemp(Nm,grid); // waste of space replicating
 	
 	Field f(grid);
 	Field v(grid);
@@ -533,48 +517,35 @@ until convergence
 	// (uniform vector) Why not src??
 	//	evec[0] = 1.0;
 	evec[0] = src;
-	std:: cout<<GridLogMessage <<"norm2(src)= " << norm2(src)<<std::endl;
+	std:: cout <<"norm2(src)= " << norm2(src)<<std::endl;
 // << src._grid  << std::endl;
 	normalise(evec[0]);
-	std:: cout<<GridLogMessage <<"norm2(evec[0])= " << norm2(evec[0]) <<std::endl;
+	std:: cout <<"norm2(evec[0])= " << norm2(evec[0]) <<std::endl;
 // << evec[0]._grid << std::endl;
 	
 	// Initial Nk steps
-	OrthoTime=0.;
-	double t0=usecond()/1e6;
 	for(int k=0; k<Nk; ++k) step(eval,lme,evec,f,Nm,k);
-	double t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::Initial steps: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
-	std::cout<<GridLogMessage <<"IRL::Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
-//	std:: cout<<GridLogMessage <<"norm2(evec[1])= " << norm2(evec[1]) << std::endl;
-//	std:: cout<<GridLogMessage <<"norm2(evec[2])= " << norm2(evec[2]) << std::endl;
+//	std:: cout <<"norm2(evec[1])= " << norm2(evec[1]) << std::endl;
+//	std:: cout <<"norm2(evec[2])= " << norm2(evec[2]) << std::endl;
 	RitzMatrix(evec,Nk);
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::RitzMatrix: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
 	for(int k=0; k<Nk; ++k){
-//	std:: cout<<GridLogMessage <<"eval " << k << " " <<eval[k] << std::endl;
-//	std:: cout<<GridLogMessage <<"lme " << k << " " << lme[k] << std::endl;
+//	std:: cout <<"eval " << k << " " <<eval[k] << std::endl;
+//	std:: cout <<"lme " << k << " " << lme[k] << std::endl;
 	}
 
 	// Restarting loop begins
 	for(int iter = 0; iter<Niter; ++iter){
 
-	  std::cout<<GridLogMessage<<"\n Restart iteration = "<< iter << std::endl;
+	  std::cout<<"\n Restart iteration = "<< iter << std::endl;
 
 	  // 
 	  // Rudy does a sort first which looks very different. Getting fed up with sorting out the algo defs.
 	  // We loop over 
 	  //
-	OrthoTime=0.;
 	  for(int k=Nk; k<Nm; ++k) step(eval,lme,evec,f,Nm,k);
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL:: "<<Np <<" steps: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
-	std::cout<<GridLogMessage <<"IRL::Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
 	  f *= lme[Nm-1];
 
 	  RitzMatrix(evec,k2);
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL:: RitzMatrix: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
 	  
 	  // getting eigenvalues
 	  for(int k=0; k<Nm; ++k){
@@ -583,27 +554,18 @@ until convergence
 	  }
 	  setUnit_Qt(Nm,Qt);
 	  diagonalize(eval2,lme2,Nm,Nm,Qt,grid);
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL:: diagonalize: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
 
 	  // sorting
 	  _sort.push(eval2,Nm);
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL:: eval sorting: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
 	  
 	  // Implicitly shifted QR transformations
 	  setUnit_Qt(Nm,Qt);
-	  for(int ip=0; ip<k2; ++ip){
-	std::cout<<GridLogMessage << "eval "<< ip << " "<< eval2[ip] << std::endl;
-	}
 	  for(int ip=k2; ip<Nm; ++ip){ 
-	std::cout<<GridLogMessage << "qr_decomp "<< ip << " "<< eval2[ip] << std::endl;
+	std::cout << "qr_decomp "<< ip << " "<< eval2[ip] << std::endl;
 	    qr_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
 		
 	}
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::qr_decomp: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
-if (0) {  
+    
 	  for(int i=0; i<(Nk+1); ++i) B[i] = 0.0;
 	  
 	  for(int j=k1-1; j<k2+1; ++j){
@@ -612,38 +574,14 @@ if (0) {
 	      B[j] += Qt[k+Nm*j] * evec[k];
 	    }
 	  }
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::QR Rotate: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
-}
-
-if (1) {
-	for(int i=0; i<(Nk+1); ++i) {
-		B[i] = 0.0;
-	  	B[i].checkerboard = evec[0].checkerboard;
-	}
-
-	int j_block = 24; int k_block=24;
-PARALLEL_FOR_LOOP
-	for(int ss=0;ss < grid->oSites();ss++){
-	for(int jj=k1-1; jj<k2+1; jj += j_block)
-	for(int kk=0; kk<Nm; kk += k_block)
-	for(int j=jj; (j<(k2+1)) && j<(jj+j_block); ++j){
-	for(int k=kk; (k<Nm) && k<(kk+k_block) ; ++k){
-	    B[j]._odata[ss] +=Qt[k+Nm*j] * evec[k]._odata[ss]; 
-	}
-	}
-	}
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::QR rotation: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
-}
-	for(int j=k1-1; j<k2+1; ++j) evec[j] = B[j];
+	  for(int j=k1-1; j<k2+1; ++j) evec[j] = B[j];
 
 	  // Compressed vector f and beta(k2)
 	  f *= Qt[Nm-1+Nm*(k2-1)];
 	  f += lme[k2-1] * evec[k2];
 	  beta_k = norm2(f);
 	  beta_k = sqrt(beta_k);
-	  std::cout<<GridLogMessage<<" beta(k) = "<<beta_k<<std::endl;
+	  std::cout<<" beta(k) = "<<beta_k<<std::endl;
 
 	  RealD betar = 1.0/beta_k;
 	  evec[k2] = betar * f;
@@ -656,10 +594,7 @@ PARALLEL_FOR_LOOP
 	  }
 	  setUnit_Qt(Nm,Qt);
 	  diagonalize(eval2,lme2,Nk,Nm,Qt,grid);
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::diagonalize: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
 	  
-if (0) {
 	  for(int k = 0; k<Nk; ++k) B[k]=0.0;
 	  
 	  for(int j = 0; j<Nk; ++j){
@@ -667,34 +602,12 @@ if (0) {
 	    B[j].checkerboard = evec[k].checkerboard;
 	      B[j] += Qt[k+j*Nm] * evec[k];
 	    }
-	    std::cout<<GridLogMessage << "norm(B["<<j<<"])="<<norm2(B[j])<<std::endl;
+//	    std::cout << "norm(B["<<j<<"])="<<norm2(B[j])<<std::endl;
 	  }
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::Convergence rotation: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
-}
-if (1) {
-	for(int i=0; i<(Nk+1); ++i) {
-		B[i] = 0.0;
-	  	B[i].checkerboard = evec[0].checkerboard;
-	}
-
-	int j_block = 24; int k_block=24;
-PARALLEL_FOR_LOOP
-	for(int ss=0;ss < grid->oSites();ss++){
-	for(int jj=0; jj<Nk; jj += j_block)
-	for(int kk=0; kk<Nk; kk += k_block)
-	for(int j=jj; (j<Nk) && j<(jj+j_block); ++j){
-	for(int k=kk; (k<Nk) && k<(kk+k_block) ; ++k){
-	    B[j]._odata[ss] +=Qt[k+Nm*j] * evec[k]._odata[ss]; 
-	}
-	}
-	}
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::convergence rotation : "<<t1-t0<< "seconds"<<std::endl; t0=t1;
-}
+//	_sort.push(eval2,B,Nk);
 
 	  Nconv = 0;
-	  //	  std::cout<<GridLogMessage << std::setiosflags(std::ios_base::scientific);
+	  //	  std::cout << std::setiosflags(std::ios_base::scientific);
 	  for(int i=0; i<Nk; ++i){
 
 //	    _poly(_Linop,B[i],v);
@@ -702,16 +615,14 @@ PARALLEL_FOR_LOOP
 	    
 	    RealD vnum = real(innerProduct(B[i],v)); // HermOp.
 	    RealD vden = norm2(B[i]);
-	    RealD vv0 = norm2(v);
 	    eval2[i] = vnum/vden;
 	    v -= eval2[i]*B[i];
 	    RealD vv = norm2(v);
 	    
 	    std::cout.precision(13);
-	    std::cout<<GridLogMessage << "[" << std::setw(3)<< std::setiosflags(std::ios_base::right) <<i<<"] ";
-	    std::cout<<"eval = "<<std::setw(25)<< std::setiosflags(std::ios_base::left)<< eval2[i];
-	    std::cout<<"|H B[i] - eval[i]B[i]|^2 "<< std::setw(25)<< std::setiosflags(std::ios_base::right)<< vv;
-	    std::cout<<" "<< vnum/(sqrt(vden)*sqrt(vv0)) << std::endl;
+	    std::cout << "[" << std::setw(3)<< std::setiosflags(std::ios_base::right) <<i<<"] ";
+	    std::cout << "eval = "<<std::setw(25)<< std::setiosflags(std::ios_base::left)<< eval2[i];
+	    std::cout <<" |H B[i] - eval[i]B[i]|^2 "<< std::setw(25)<< std::setiosflags(std::ios_base::right)<< vv<< std::endl;
 	    
 	// change the criteria as evals are supposed to be sorted, all evals smaller(larger) than Nstop should have converged
 	    if((vv<eresid*eresid) && (i == Nconv) ){
@@ -720,19 +631,17 @@ PARALLEL_FOR_LOOP
 	    }
 
 	  }  // i-loop end
-	  //	  std::cout<<GridLogMessage << std::resetiosflags(std::ios_base::scientific);
-	t1=usecond()/1e6;
-	std::cout<<GridLogMessage <<"IRL::convergence testing: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
+	  //	  std::cout << std::resetiosflags(std::ios_base::scientific);
 
 
-	  std::cout<<GridLogMessage<<" #modes converged: "<<Nconv<<std::endl;
+	  std::cout<<" #modes converged: "<<Nconv<<std::endl;
 
 	  if( Nconv>=Nstop ){
 	    goto converged;
 	  }
 	} // end of iter loop
 	
-	std::cout<<GridLogMessage<<"\n NOT converged.\n";
+	std::cout<<"\n NOT converged.\n";
 	abort();
 	
       converged:
@@ -745,10 +654,10 @@ PARALLEL_FOR_LOOP
        }
       _sort.push(eval,evec,Nconv);
 
-      std::cout<<GridLogMessage << "\n Converged\n Summary :\n";
-      std::cout<<GridLogMessage << " -- Iterations  = "<< Nconv  << "\n";
-      std::cout<<GridLogMessage << " -- beta(k)     = "<< beta_k << "\n";
-      std::cout<<GridLogMessage << " -- Nconv       = "<< Nconv  << "\n";
+      std::cout << "\n Converged\n Summary :\n";
+      std::cout << " -- Iterations  = "<< Nconv  << "\n";
+      std::cout << " -- beta(k)     = "<< beta_k << "\n";
+      std::cout << " -- Nconv       = "<< Nconv  << "\n";
      }
 
     /////////////////////////////////////////////////
@@ -771,25 +680,25 @@ PARALLEL_FOR_LOOP
 	}
       }
 
-      std::cout<<GridLogMessage<<"Lanczos_Factor start/end " <<start <<"/"<<end<<std::endl;
+      std::cout<<"Lanczos_Factor start/end " <<start <<"/"<<end<<std::endl;
 
       // Starting from scratch, bq[0] contains a random vector and |bq[0]| = 1
       int first;
       if(start == 0){
 
-	std::cout<<GridLogMessage << "start == 0\n"; //TESTING
+	std::cout << "start == 0\n"; //TESTING
 
 	_poly(_Linop,bq[0],bf);
 
 	alpha = real(innerProduct(bq[0],bf));//alpha =  bq[0]^dag A bq[0]
 
-	std::cout<<GridLogMessage << "alpha = " << alpha << std::endl;
+	std::cout << "alpha = " << alpha << std::endl;
 	
 	bf = bf - alpha * bq[0];  //bf =  A bq[0] - alpha bq[0]
 
 	H[0][0]=alpha;
 
-	std::cout<<GridLogMessage << "Set H(0,0) to " << H[0][0] << std::endl;
+	std::cout << "Set H(0,0) to " << H[0][0] << std::endl;
 
 	first = 1;
 
@@ -809,19 +718,19 @@ PARALLEL_FOR_LOOP
 
 	beta = 0;sqbt = 0;
 
-	std::cout<<GridLogMessage << "cont is true so setting beta to zero\n";
+	std::cout << "cont is true so setting beta to zero\n";
 
       }	else {
 
 	beta = norm2(bf);
 	sqbt = sqrt(beta);
 
-	std::cout<<GridLogMessage << "beta = " << beta << std::endl;
+	std::cout << "beta = " << beta << std::endl;
       }
 
       for(int j=first;j<end;j++){
 
-	std::cout<<GridLogMessage << "Factor j " << j <<std::endl;
+	std::cout << "Factor j " << j <<std::endl;
 
 	if(cont){ // switches to factoring; understand start!=0 and initial bf value is right.
 	  bq[j] = bf; cont = false;
@@ -844,7 +753,7 @@ PARALLEL_FOR_LOOP
 
 	beta = fnorm;
 	sqbt = sqrt(beta);
-	std::cout<<GridLogMessage << "alpha = " << alpha << " fnorm = " << fnorm << '\n';
+	std::cout << "alpha = " << alpha << " fnorm = " << fnorm << '\n';
 
 	///Iterative refinement of orthogonality V = [ bq[0]  bq[1]  ...  bq[M] ]
 	int re = 0;
@@ -879,8 +788,8 @@ PARALLEL_FOR_LOOP
 	  bck = sqrt( nmbex );
 	  re++;
 	}
-	std::cout<<GridLogMessage << "Iteratively refined orthogonality, changes alpha\n";
-	if(re > 1) std::cout<<GridLogMessage << "orthagonality refined " << re << " times" <<std::endl;
+	std::cout << "Iteratively refined orthogonality, changes alpha\n";
+	if(re > 1) std::cout << "orthagonality refined " << re << " times" <<std::endl;
 	H[j][j]=alpha;
       }
 
@@ -895,13 +804,11 @@ PARALLEL_FOR_LOOP
 
     void ImplicitRestart(int TM, DenseVector<RealD> &evals,  DenseVector<DenseVector<RealD> > &evecs, DenseVector<Field> &bq, Field &bf, int cont)
     {
-      std::cout<<GridLogMessage << "ImplicitRestart begin. Eigensort starting\n";
+      std::cout << "ImplicitRestart begin. Eigensort starting\n";
 
       DenseMatrix<RealD> H; Resize(H,Nm,Nm);
 
-#ifndef USE_LAPACK
       EigenSort(evals, evecs);
-#endif
 
       ///Assign shifts
       int K=Nk;
@@ -924,15 +831,15 @@ PARALLEL_FOR_LOOP
       /// Shifted H defines a new K step Arnoldi factorization
       RealD  beta = H[ff][ff-1]; 
       RealD  sig  = Q[TM - 1][ff - 1];
-      std::cout<<GridLogMessage << "beta = " << beta << " sig = " << real(sig) <<std::endl;
+      std::cout << "beta = " << beta << " sig = " << real(sig) <<std::endl;
 
-      std::cout<<GridLogMessage << "TM = " << TM << " ";
-      std::cout<<GridLogMessage << norm2(bq[0]) << " -- before" <<std::endl;
+      std::cout << "TM = " << TM << " ";
+      std::cout << norm2(bq[0]) << " -- before" <<std::endl;
 
       /// q -> q Q
       times_real(bq, Q, TM);
 
-      std::cout<<GridLogMessage << norm2(bq[0]) << " -- after " << ff <<std::endl;
+      std::cout << norm2(bq[0]) << " -- after " << ff <<std::endl;
       bf =  beta* bq[ff] + sig* bf;
 
       /// Do the rest of the factorization
@@ -956,7 +863,7 @@ PARALLEL_FOR_LOOP
       int ff = Lanczos_Factor(0, M, cont, bq,bf,H); // 0--M to begin with
 
       if(ff < M) {
-	std::cout<<GridLogMessage << "Krylov: aborting ff "<<ff <<" "<<M<<std::endl;
+	std::cout << "Krylov: aborting ff "<<ff <<" "<<M<<std::endl;
 	abort(); // Why would this happen?
       }
 
@@ -965,7 +872,7 @@ PARALLEL_FOR_LOOP
 
       for(int it = 0; it < Niter && (converged < Nk); ++it) {
 
-	std::cout<<GridLogMessage << "Krylov: Iteration --> " << it << std::endl;
+	std::cout << "Krylov: Iteration --> " << it << std::endl;
 	int lock_num = lock ? converged : 0;
 	DenseVector<RealD> tevals(M - lock_num );
 	DenseMatrix<RealD> tevecs; Resize(tevecs,M - lock_num,M - lock_num);
@@ -981,7 +888,7 @@ PARALLEL_FOR_LOOP
       Wilkinson<RealD>(H, evals, evecs, small); 
       //      Check();
 
-      std::cout<<GridLogMessage << "Done  "<<std::endl;
+      std::cout << "Done  "<<std::endl;
 
     }
 
@@ -1046,7 +953,7 @@ PARALLEL_FOR_LOOP
 		  DenseVector<RealD> &tevals, DenseVector<DenseVector<RealD> > &tevecs, 
 		  int lock, int converged)
     {
-      std::cout<<GridLogMessage << "Converged " << converged << " so far." << std::endl;
+      std::cout << "Converged " << converged << " so far." << std::endl;
       int lock_num = lock ? converged : 0;
       int M = Nm;
 
@@ -1061,9 +968,7 @@ PARALLEL_FOR_LOOP
       RealD small=1.0e-16;
       Wilkinson<RealD>(AH, tevals, tevecs, small);
 
-#ifndef USE_LAPACK
       EigenSort(tevals, tevecs);
-#endif
 
       RealD resid_nrm=  norm2(bf);
 
@@ -1074,7 +979,7 @@ PARALLEL_FOR_LOOP
 	RealD diff = 0;
 	diff = abs( tevecs[i][Nm - 1 - lock_num] ) * resid_nrm;
 
-	std::cout<<GridLogMessage << "residual estimate " << SS-1-i << " " << diff << " of (" << tevals[i] << ")" << std::endl;
+	std::cout << "residual estimate " << SS-1-i << " " << diff << " of (" << tevals[i] << ")" << std::endl;
 
 	if(diff < converged) {
 
@@ -1090,13 +995,13 @@ PARALLEL_FOR_LOOP
 	    lock_num++;
 	  }
 	  converged++;
-	  std::cout<<GridLogMessage << " converged on eval " << converged << " of " << Nk << std::endl;
+	  std::cout << " converged on eval " << converged << " of " << Nk << std::endl;
 	} else {
 	  break;
 	}
       }
 #endif
-      std::cout<<GridLogMessage << "Got " << converged << " so far " <<std::endl;	
+      std::cout << "Got " << converged << " so far " <<std::endl;	
     }
 
     ///Check
@@ -1105,9 +1010,7 @@ PARALLEL_FOR_LOOP
 
       DenseVector<RealD> goodval(this->get);
 
-#ifndef USE_LAPACK
       EigenSort(evals,evecs);
-#endif
 
       int NM = Nm;
 
@@ -1179,16 +1082,14 @@ say con = 2
 **/
 
 template<class T>
-static void Lock(DenseMatrix<T> &H, 	///Hess mtx	
-		 DenseMatrix<T> &Q, 	///Lock Transform
-		 T val, 		///value to be locked
-		 int con, 	///number already locked
+static void Lock(DenseMatrix<T> &H, 	// Hess mtx	
+		 DenseMatrix<T> &Q, 	// Lock Transform
+		 T val, 		// value to be locked
+		 int con, 	// number already locked
 		 RealD small,
 		 int dfg,
 		 bool herm)
 {	
-
-
   //ForceTridiagonal(H);
 
   int M = H.dim;
@@ -1221,7 +1122,6 @@ static void Lock(DenseMatrix<T> &H, 	///Hess mtx
   AH = Hermitian(QQ)*AH;
   AH = AH*QQ;
 
-
   for(int i=con;i<M;i++){
     for(int j=con;j<M;j++){
       Q[i][j]=QQ[i-con][j-con];

lib/algorithms/iterative/Matrix.h

@@ -1,453 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/algorithms/iterative/Matrix.h
-
-    Copyright (C) 2015
-
-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 */
-#ifndef MATRIX_H
-#define MATRIX_H
-
-#include <cstdlib>
-#include <string>
-#include <cmath>
-#include <vector>
-#include <iostream>
-#include <iomanip>
-#include <complex>
-#include <typeinfo>
-#include <Grid.h>
-
-
-/** Sign function **/
-template <class T> T sign(T p){return ( p/abs(p) );}
-
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-///////////////////// Hijack STL containers for our wicked means /////////////////////////////////////////
-/////////////////////////////////////////////////////////////////////////////////////////////////////////
-template<class T> using Vector = Vector<T>;
-template<class T> using Matrix = Vector<Vector<T> >;
-
-template<class T> void Resize(Vector<T > & vec, int N) { vec.resize(N); }
-
-template<class T> void Resize(Matrix<T > & mat, int N, int M) { 
-  mat.resize(N);
-  for(int i=0;i<N;i++){
-    mat[i].resize(M);
-  }
-}
-template<class T> void Size(Vector<T> & vec, int &N) 
-{ 
-  N= vec.size();
-}
-template<class T> void Size(Matrix<T> & mat, int &N,int &M) 
-{ 
-  N= mat.size();
-  M= mat[0].size();
-}
-template<class T> void SizeSquare(Matrix<T> & mat, int &N) 
-{ 
-  int M; Size(mat,N,M);
-  assert(N==M);
-}
-template<class T> void SizeSame(Matrix<T> & mat1,Matrix<T> &mat2, int &N1,int &M1) 
-{ 
-  int N2,M2;
-  Size(mat1,N1,M1);
-  Size(mat2,N2,M2);
-  assert(N1==N2);
-  assert(M1==M2);
-}
-
-//*****************************************
-//*	(Complex) Vector operations	*
-//*****************************************
-
-/**Conj of a Vector **/
-template <class T> Vector<T> conj(Vector<T> p){
-	Vector<T> q(p.size());
-	for(int i=0;i<p.size();i++){q[i] = conj(p[i]);}
-	return q;
-}
-
-/** Norm of a Vector**/
-template <class T> T norm(Vector<T> p){
-	T sum = 0;
-	for(int i=0;i<p.size();i++){sum = sum + p[i]*conj(p[i]);}
-	return abs(sqrt(sum));
-}
-
-/** Norm squared of a Vector **/
-template <class T> T norm2(Vector<T> p){
-	T sum = 0;
-	for(int i=0;i<p.size();i++){sum = sum + p[i]*conj(p[i]);}
-	return abs((sum));
-}
-
-/** Sum elements of a Vector **/
-template <class T> T trace(Vector<T> p){
-	T sum = 0;
-	for(int i=0;i<p.size();i++){sum = sum + p[i];}
-	return sum;
-}
-
-/** Fill a Vector with constant c **/
-template <class T> void Fill(Vector<T> &p, T c){
-	for(int i=0;i<p.size();i++){p[i] = c;}
-}
-/** Normalize a Vector **/
-template <class T> void normalize(Vector<T> &p){
-	T m = norm(p);
-	if( abs(m) > 0.0) for(int i=0;i<p.size();i++){p[i] /= m;}
-}
-/** Vector by scalar **/
-template <class T, class U> Vector<T> times(Vector<T> p, U s){
-	for(int i=0;i<p.size();i++){p[i] *= s;}
-	return p;
-}
-template <class T, class U> Vector<T> times(U s, Vector<T> p){
-	for(int i=0;i<p.size();i++){p[i] *= s;}
-	return p;
-}
-/** inner product of a and b = conj(a) . b **/
-template <class T> T inner(Vector<T> a, Vector<T> b){
-	T m = 0.;
-	for(int i=0;i<a.size();i++){m = m + conj(a[i])*b[i];}
-	return m;
-}
-/** sum of a and b = a + b **/
-template <class T> Vector<T> add(Vector<T> a, Vector<T> b){
-	Vector<T> m(a.size());
-	for(int i=0;i<a.size();i++){m[i] = a[i] + b[i];}
-	return m;
-}
-/** sum of a and b = a - b **/
-template <class T> Vector<T> sub(Vector<T> a, Vector<T> b){
-	Vector<T> m(a.size());
-	for(int i=0;i<a.size();i++){m[i] = a[i] - b[i];}
-	return m;
-}
-
-/** 
- *********************************
- *	Matrices	         *
- *********************************
- **/
-
-template<class T> void Fill(Matrix<T> & mat, T&val) { 
-  int N,M;
-  Size(mat,N,M);
-  for(int i=0;i<N;i++){
-  for(int j=0;j<M;j++){
-    mat[i][j] = val;
-  }}
-}
-
-/** Transpose of a matrix **/
-Matrix<T> Transpose(Matrix<T> & mat){
-  int N,M;
-  Size(mat,N,M);
-  Matrix C; Resize(C,M,N);
-  for(int i=0;i<M;i++){
-  for(int j=0;j<N;j++){
-    C[i][j] = mat[j][i];
-  }} 
-  return C;
-}
-/** Set Matrix to unit matrix **/
-template<class T> void Unity(Matrix<T> &mat){
-  int N;  SizeSquare(mat,N);
-  for(int i=0;i<N;i++){
-    for(int j=0;j<N;j++){
-      if ( i==j ) A[i][j] = 1;
-      else        A[i][j] = 0;
-    } 
-  } 
-}
-/** Add C * I to matrix **/
-template<class T>
-void PlusUnit(Matrix<T> & A,T c){
-  int dim;  SizeSquare(A,dim);
-  for(int i=0;i<dim;i++){A[i][i] = A[i][i] + c;} 
-}
-
-/** return the Hermitian conjugate of matrix **/
-Matrix<T> HermitianConj(Matrix<T> &mat){
-
-  int dim; SizeSquare(mat,dim);
-
-  Matrix<T> C; Resize(C,dim,dim);
-
-  for(int i=0;i<dim;i++){
-    for(int j=0;j<dim;j++){
-      C[i][j] = conj(mat[j][i]);
-    } 
-  } 
-  return C;
-}
-
-/** return diagonal entries as a Vector **/
-Vector<T> diag(Matrix<T> &A)
-{
-  int dim; SizeSquare(A,dim);
-  Vector<T> d; Resize(d,dim);
-
-  for(int i=0;i<dim;i++){
-    d[i] = A[i][i];
-  }
-  return d;
-}
-
-/** Left multiply by a Vector **/
-Vector<T> operator *(Vector<T> &B,Matrix<T> &A)
-{
-  int K,M,N; 
-  Size(B,K);
-  Size(A,M,N);
-  assert(K==M);
-  
-  Vector<T> C; Resize(C,N);
-
-  for(int j=0;j<N;j++){
-    T sum = 0.0;
-    for(int i=0;i<M;i++){
-      sum += B[i] * A[i][j];
-    }
-    C[j] =  sum;
-  }
-  return C; 
-}
-
-/** return 1/diagonal entries as a Vector **/
-Vector<T> inv_diag(Matrix<T> & A){
-  int dim; SizeSquare(A,dim);
-  Vector<T> d; Resize(d,dim);
-  for(int i=0;i<dim;i++){
-    d[i] = 1.0/A[i][i];
-  }
-  return d;
-}
-/** Matrix Addition **/
-inline Matrix<T> operator + (Matrix<T> &A,Matrix<T> &B)
-{
-  int N,M  ; SizeSame(A,B,N,M);
-  Matrix C; Resize(C,N,M);
-  for(int i=0;i<N;i++){
-    for(int j=0;j<M;j++){
-      C[i][j] = A[i][j] +  B[i][j];
-    } 
-  } 
-  return C;
-} 
-/** Matrix Subtraction **/
-inline Matrix<T> operator- (Matrix<T> & A,Matrix<T> &B){
-  int N,M  ; SizeSame(A,B,N,M);
-  Matrix C; Resize(C,N,M);
-  for(int i=0;i<N;i++){
-  for(int j=0;j<M;j++){
-    C[i][j] = A[i][j] -  B[i][j];
-  }}
-  return C;
-} 
-
-/** Matrix scalar multiplication **/
-inline Matrix<T> operator* (Matrix<T> & A,T c){
-  int N,M; Size(A,N,M);
-  Matrix C; Resize(C,N,M);
-  for(int i=0;i<N;i++){
-  for(int j=0;j<M;j++){
-    C[i][j] = A[i][j]*c;
-  }} 
-  return C;
-} 
-/** Matrix Matrix multiplication **/
-inline Matrix<T> operator* (Matrix<T> &A,Matrix<T> &B){
-  int K,L,N,M;
-  Size(A,K,L);
-  Size(B,N,M); assert(L==N);
-  Matrix C; Resize(C,K,M);
-
-  for(int i=0;i<K;i++){
-    for(int j=0;j<M;j++){
-      T sum = 0.0;
-      for(int k=0;k<N;k++) sum += A[i][k]*B[k][j];
-      C[i][j] =sum;
-    }
-  }
-  return C; 
-} 
-/** Matrix Vector multiplication **/
-inline Vector<T> operator* (Matrix<T> &A,Vector<T> &B){
-  int M,N,K;
-  Size(A,N,M);
-  Size(B,K); assert(K==M);
-  Vector<T> C; Resize(C,N);
-  for(int i=0;i<N;i++){
-    T sum = 0.0;
-    for(int j=0;j<M;j++) sum += A[i][j]*B[j];
-    C[i] =  sum;
-  }
-  return C; 
-} 
-
-/** Some version of Matrix norm **/
-/*
-inline T Norm(){ // this is not a usual L2 norm
-    T norm = 0;
-    for(int i=0;i<dim;i++){
-      for(int j=0;j<dim;j++){
-	norm += abs(A[i][j]);
-    }}
-    return norm;
-  }
-*/
-
-/** Some version of Matrix norm **/
-template<class T> T LargestDiag(Matrix<T> &A)
-{
-  int dim ; SizeSquare(A,dim); 
-
-  T ld = abs(A[0][0]);
-  for(int i=1;i<dim;i++){
-    T cf = abs(A[i][i]);
-    if(abs(cf) > abs(ld) ){ld = cf;}
-  }
-  return ld;
-}
-
-/** Look for entries on the leading subdiagonal that are smaller than 'small' **/
-template <class T,class U> int Chop_subdiag(Matrix<T> &A,T norm, int offset, U small)
-{
-  int dim; SizeSquare(A,dim);
-  for(int l = dim - 1 - offset; l >= 1; l--) {             		
-    if((U)abs(A[l][l - 1]) < (U)small) {
-      A[l][l-1]=(U)0.0;
-      return l;
-    }
-  }
-  return 0;
-}
-
-/** Look for entries on the leading subdiagonal that are smaller than 'small' **/
-template <class T,class U> int Chop_symm_subdiag(Matrix<T> & A,T norm, int offset, U small) 
-{
-  int dim; SizeSquare(A,dim);
-  for(int l = dim - 1 - offset; l >= 1; l--) {
-    if((U)abs(A[l][l - 1]) < (U)small) {
-      A[l][l - 1] = (U)0.0;
-      A[l - 1][l] = (U)0.0;
-      return l;
-    }
-  }
-  return 0;
-}
-/**Assign a submatrix to a larger one**/
-template<class T>
-void AssignSubMtx(Matrix<T> & A,int row_st, int row_end, int col_st, int col_end, Matrix<T> &S)
-{
-  for(int i = row_st; i<row_end; i++){
-    for(int j = col_st; j<col_end; j++){
-      A[i][j] = S[i - row_st][j - col_st];
-    }
-  }
-}
-
-/**Get a square submatrix**/
-template <class T>
-Matrix<T> GetSubMtx(Matrix<T> &A,int row_st, int row_end, int col_st, int col_end)
-{
-  Matrix<T> H; Resize(row_end - row_st,col_end-col_st);
-
-  for(int i = row_st; i<row_end; i++){
-  for(int j = col_st; j<col_end; j++){
-    H[i-row_st][j-col_st]=A[i][j];
-  }}
-  return H;
-}
-  
- /**Assign a submatrix to a larger one NB remember Vector Vectors are transposes of the matricies they represent**/
-template<class T>
-void AssignSubMtx(Matrix<T> & A,int row_st, int row_end, int col_st, int col_end, Matrix<T> &S)
-{
-  for(int i = row_st; i<row_end; i++){
-  for(int j = col_st; j<col_end; j++){
-    A[i][j] = S[i - row_st][j - col_st];
-  }}
-}
-  
-/** compute b_i A_ij b_j **/ // surprised no Conj
-template<class T> T proj(Matrix<T> A, Vector<T> B){
-  int dim; SizeSquare(A,dim);
-  int dimB; Size(B,dimB);
-  assert(dimB==dim);
-  T C = 0;
-  for(int i=0;i<dim;i++){
-    T sum = 0.0;
-    for(int j=0;j<dim;j++){
-      sum += A[i][j]*B[j];
-    }
-    C +=  B[i]*sum; // No conj?
-  }
-  return C; 
-}
-
-
-/*
- *************************************************************
- *
- * Matrix Vector products
- *
- *************************************************************
- */
-// Instead make a linop and call my CG;
-
-/// q -> q Q
-template <class T,class Fermion> void times(Vector<Fermion> &q, Matrix<T> &Q)
-{
-  int M; SizeSquare(Q,M);
-  int N; Size(q,N); 
-  assert(M==N);
-
-  times(q,Q,N);
-}
-
-/// q -> q Q
-template <class T> void times(multi1d<LatticeFermion> &q, Matrix<T> &Q, int N)
-{
-  GridBase *grid = q[0]._grid;
-  int M; SizeSquare(Q,M);
-  int K; Size(q,K); 
-  assert(N<M);
-  assert(N<K);
-  Vector<Fermion> S(N,grid );
-  for(int j=0;j<N;j++){
-    S[j] = zero;
-    for(int k=0;k<N;k++){
-      S[j] = S[j] +  q[k]* Q[k][j]; 
-    }
-  }
-  for(int j=0;j<q.size();j++){
-    q[j] = S[j];
-  }
-}
-#endif

lib/algorithms/iterative/TODO

@@ -1,15 +0,0 @@
-- ConjugateGradientMultiShift
-- MCR
-
-- Potentially Useful Boost libraries
-
-- MultiArray
-- Aligned allocator; memory pool
-- Remez -- Mike or Boost?
-- Multiprecision
-- quaternians
-- Tokenize
-- Serialization
-- Regex
-- Proto (ET)
-- uBlas

lib/algorithms/iterative/bisec.c

@@ -1,122 +0,0 @@
-#include <math.h>
-#include <stdlib.h>
-#include <vector>
-
-struct Bisection {
-
-static void get_eig2(int row_num,std::vector<RealD> &ALPHA,std::vector<RealD> &BETA, std::vector<RealD> & eig)
-{
-  int i,j;
-  std::vector<RealD> evec1(row_num+3);
-  std::vector<RealD> evec2(row_num+3);
-  RealD eps2;
-  ALPHA[1]=0.;
-  BETHA[1]=0.;
-  for(i=0;i<row_num-1;i++) {
-    ALPHA[i+1] = A[i*(row_num+1)].real();
-    BETHA[i+2] = A[i*(row_num+1)+1].real();
-  }
-  ALPHA[row_num] = A[(row_num-1)*(row_num+1)].real();
-  bisec(ALPHA,BETHA,row_num,1,row_num,1e-10,1e-10,evec1,eps2);
-  bisec(ALPHA,BETHA,row_num,1,row_num,1e-16,1e-16,evec2,eps2);
-
-  // Do we really need to sort here?
-  int begin=1;
-  int end = row_num;
-  int swapped=1;
-  while(swapped) {
-    swapped=0;
-    for(i=begin;i<end;i++){
-      if(mag(evec2[i])>mag(evec2[i+1]))	{
-	swap(evec2+i,evec2+i+1);
-	swapped=1;
-      }
-    }
-    end--;
-    for(i=end-1;i>=begin;i--){
-      if(mag(evec2[i])>mag(evec2[i+1]))	{
-	swap(evec2+i,evec2+i+1);
-	swapped=1;
-      }
-    }
-    begin++;
-  }
-
-  for(i=0;i<row_num;i++){
-    for(j=0;j<row_num;j++) {
-      if(i==j) H[i*row_num+j]=evec2[i+1];
-      else H[i*row_num+j]=0.;
-    }
-  }
-}
-
-static void bisec(std::vector<RealD> &c,   
-		  std::vector<RealD> &b,
-		  int n,
-		  int m1,
-		  int m2,
-		  RealD eps1,
-		  RealD relfeh,
-		  std::vector<RealD> &x,
-		  RealD &eps2)
-{
-  std::vector<RealD> wu(n+2);
-
-  RealD h,q,x1,xu,x0,xmin,xmax; 
-  int i,a,k;
-
-  b[1]=0.0;
-  xmin=c[n]-fabs(b[n]);
-  xmax=c[n]+fabs(b[n]);
-  for(i=1;i<n;i++){
-    h=fabs(b[i])+fabs(b[i+1]);
-    if(c[i]+h>xmax) xmax= c[i]+h;
-    if(c[i]-h<xmin) xmin= c[i]-h;
-  }
-  xmax *=2.;
-
-  eps2=relfeh*((xmin+xmax)>0.0 ? xmax : -xmin);
-  if(eps1<=0.0) eps1=eps2;
-  eps2=0.5*eps1+7.0*(eps2);
-  x0=xmax;
-  for(i=m1;i<=m2;i++){
-    x[i]=xmax;
-    wu[i]=xmin;
-  }
-
-  for(k=m2;k>=m1;k--){
-    xu=xmin;
-    i=k;
-    do{
-      if(xu<wu[i]){
-	xu=wu[i];
-	i=m1-1;
-      }
-      i--;
-    }while(i>=m1);
-    if(x0>x[k]) x0=x[k];
-    while((x0-xu)>2*relfeh*(fabs(xu)+fabs(x0))+eps1){
-      x1=(xu+x0)/2;
-
-      a=0;
-      q=1.0;
-      for(i=1;i<=n;i++){
-	q=c[i]-x1-((q!=0.0)? b[i]*b[i]/q:fabs(b[i])/relfeh);
-	if(q<0) a++;
-      }
-      //			printf("x1=%e a=%d\n",x1,a);
-      if(a<k){
-	if(a<m1){
-	  xu=x1;
-	  wu[m1]=x1;
-	}else {
-	  xu=x1;
-	  wu[a+1]=x1;
-	  if(x[a]>x1) x[a]=x1;
-	}
-      }else x0=x1;
-    }
-    x[k]=(x0+xu)/2;
-  }
-}
-}

lib/algorithms/iterative/get_eig.c

@@ -1 +0,0 @@
-

lib/allocator/AlignedAllocator.cc

@@ -1,7 +1,7 @@
 
 
 
-#include <Grid/Grid.h>
+#include <Grid/GridCore.h>
 
 namespace Grid {
 
@@ -13,9 +13,10 @@ void *PointerCache::Insert(void *ptr,size_t bytes) {
 
   if (bytes < 4096 ) return NULL;
 
-#ifdef _OPENMP
+#ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
+
   void * ret = NULL;
   int v = -1;
 

lib/cartesian/Cartesian_base.h

@@ -6,8 +6,9 @@
 
     Copyright (C) 2015
 
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Guido Cossu <guido.cossu@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
@@ -49,10 +50,9 @@ public:
 
     GridBase(const std::vector<int> & processor_grid) : CartesianCommunicator(processor_grid) {};
 
-
     // Physics Grid information.
     std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
-    std::vector<int> _fdimensions;// Global dimensions of array prior to cb removal
+    std::vector<int> _fdimensions;// (full) Global dimensions of array prior to cb removal
     std::vector<int> _gdimensions;// Global dimensions of array after cb removal
     std::vector<int> _ldimensions;// local dimensions of array with processor images removed
     std::vector<int> _rdimensions;// Reduced local dimensions with simd lane images and processor images removed 
@@ -62,13 +62,12 @@ public:
     int _isites;
     int _fsites;                  // _isites*_osites = product(dimensions).
     int _gsites;
-    std::vector<int> _slice_block;   // subslice information
+    std::vector<int> _slice_block;// subslice information
     std::vector<int> _slice_stride;
     std::vector<int> _slice_nblock;
 
-    // Might need these at some point
-    //    std::vector<int> _lstart;     // local start of array in gcoors. _processor_coor[d]*_ldimensions[d]
-    //    std::vector<int> _lend;       // local end of array in gcoors    _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
+    std::vector<int> _lstart;     // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
+    std::vector<int> _lend  ;     // local end of array in gcoors   _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
 
 public:
 
@@ -99,7 +98,7 @@ public:
     virtual int oIndex(std::vector<int> &coor)
     {
         int idx=0;
-	// Works with either global or local coordinates
+        // Works with either global or local coordinates
         for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
         return idx;
     }
@@ -121,6 +120,11 @@ public:
       Lexicographic::CoorFromIndex(coor,Oindex,_rdimensions);
     }
 
+    inline void InOutCoorToLocalCoor (std::vector<int> &ocoor, std::vector<int> &icoor, std::vector<int> &lcoor) {
+      lcoor.resize(_ndimension);
+      for (int d = 0; d < _ndimension; d++)
+        lcoor[d] = ocoor[d] + _rdimensions[d] * icoor[d];
+    }
 
     //////////////////////////////////////////////////////////
     // SIMD lane addressing
@@ -129,6 +133,7 @@ public:
     {
       Lexicographic::CoorFromIndex(coor,lane,_simd_layout);
     }
+
     inline int PermuteDim(int dimension){
       return _simd_layout[dimension]>1;
     }
@@ -146,15 +151,15 @@ public:
       // Distance should be either 0,1,2..
       //
       if ( _simd_layout[dimension] > 2 ) { 
-	for(int d=0;d<_ndimension;d++){
-	  if ( d != dimension ) assert ( (_simd_layout[d]==1)  );
-	}
-	permute_type = RotateBit; // How to specify distance; this is not just direction.
-	return permute_type;
+        for(int d=0;d<_ndimension;d++){
+          if ( d != dimension ) assert ( (_simd_layout[d]==1)  );
+        }
+        permute_type = RotateBit; // How to specify distance; this is not just direction.
+        return permute_type;
       }
 
       for(int d=_ndimension-1;d>dimension;d--){
-	if (_simd_layout[d]>1 ) permute_type++;
+        if (_simd_layout[d]>1 ) permute_type++;
       }
       return permute_type;
     }
@@ -169,26 +174,50 @@ public:
     inline int gSites(void) const { return _isites*_osites*_Nprocessors; }; 
     inline int Nd    (void) const { return _ndimension;};
 
+    inline const std::vector<int> LocalStarts(void)             { return _lstart;    };
     inline const std::vector<int> &FullDimensions(void)         { return _fdimensions;};
     inline const std::vector<int> &GlobalDimensions(void)       { return _gdimensions;};
     inline const std::vector<int> &LocalDimensions(void)        { return _ldimensions;};
     inline const std::vector<int> &VirtualLocalDimensions(void) { return _ldimensions;};
 
+    ////////////////////////////////////////////////////////////////
+    // Utility to print the full decomposition details 
+    ////////////////////////////////////////////////////////////////
+
+    void show_decomposition(){
+      std::cout << GridLogMessage << "Full Dimensions    : " << _fdimensions << std::endl;
+      std::cout << GridLogMessage << "Global Dimensions  : " << _gdimensions << std::endl;
+      std::cout << GridLogMessage << "Local Dimensions   : " << _ldimensions << std::endl;
+      std::cout << GridLogMessage << "Reduced Dimensions : " << _rdimensions << std::endl;
+      std::cout << GridLogMessage << "Outer strides      : " << _ostride << std::endl;
+      std::cout << GridLogMessage << "Inner strides      : " << _istride << std::endl;
+      std::cout << GridLogMessage << "iSites             : " << _isites << std::endl;
+      std::cout << GridLogMessage << "oSites             : " << _osites << std::endl;
+      std::cout << GridLogMessage << "lSites             : " << lSites() << std::endl;        
+      std::cout << GridLogMessage << "gSites             : " << gSites() << std::endl;
+      std::cout << GridLogMessage << "Nd                 : " << _ndimension << std::endl;             
+    } 
+
     ////////////////////////////////////////////////////////////////
     // Global addressing
     ////////////////////////////////////////////////////////////////
     void GlobalIndexToGlobalCoor(int gidx,std::vector<int> &gcoor){
+      assert(gidx< gSites());
       Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
     }
     void LocalIndexToLocalCoor(int lidx,std::vector<int> &lcoor){
+      assert(lidx<lSites());
       Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
     }
+
+
+
     void GlobalCoorToGlobalIndex(const std::vector<int> & gcoor,int & gidx){
       gidx=0;
       int mult=1;
       for(int mu=0;mu<_ndimension;mu++) {
-	gidx+=mult*gcoor[mu];
-	mult*=_gdimensions[mu];
+        gidx+=mult*gcoor[mu];
+        mult*=_gdimensions[mu];
       }
     }
     void GlobalCoorToProcessorCoorLocalCoor(std::vector<int> &pcoor,std::vector<int> &lcoor,const std::vector<int> &gcoor)
@@ -196,9 +225,9 @@ public:
       pcoor.resize(_ndimension);
       lcoor.resize(_ndimension);
       for(int mu=0;mu<_ndimension;mu++){
-	int _fld  = _fdimensions[mu]/_processors[mu];
-	pcoor[mu] = gcoor[mu]/_fld;
-	lcoor[mu] = gcoor[mu]%_fld;
+        int _fld  = _fdimensions[mu]/_processors[mu];
+        pcoor[mu] = gcoor[mu]/_fld;
+        lcoor[mu] = gcoor[mu]%_fld;
       }
     }
     void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const std::vector<int> &gcoor)
@@ -207,16 +236,16 @@ public:
       std::vector<int> lcoor;
       GlobalCoorToProcessorCoorLocalCoor(pcoor,lcoor,gcoor);
       rank = RankFromProcessorCoor(pcoor);
-
+      /*
       std::vector<int> cblcoor(lcoor);
       for(int d=0;d<cblcoor.size();d++){
-	if( this->CheckerBoarded(d) ) {
-	  cblcoor[d] = lcoor[d]/2;
-	}
+        if( this->CheckerBoarded(d) ) {
+          cblcoor[d] = lcoor[d]/2;
+        }
       }
-
-      i_idx= iIndex(cblcoor);// this does not imply divide by 2 on checker dim
-      o_idx= oIndex(lcoor);  // this implies divide by 2 on checkerdim
+      */
+      i_idx= iIndex(lcoor);
+      o_idx= oIndex(lcoor);
     }
 
     void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , std::vector<int> &gcoor)
@@ -238,7 +267,7 @@ public:
     {
       RankIndexToGlobalCoor(rank,o_idx,i_idx ,fcoor);
       if(CheckerBoarded(0)){
-	fcoor[0] = fcoor[0]*2+cb;
+        fcoor[0] = fcoor[0]*2+cb;
       }
     }
     void ProcessorCoorLocalCoorToGlobalCoor(std::vector<int> &Pcoor,std::vector<int> &Lcoor,std::vector<int> &gcoor)

lib/cartesian/Cartesian_full.h

@@ -76,6 +76,8 @@ public:
         _ldimensions.resize(_ndimension);
         _rdimensions.resize(_ndimension);
         _simd_layout.resize(_ndimension);
+	_lstart.resize(_ndimension);
+	_lend.resize(_ndimension);
             
         _ostride.resize(_ndimension);
         _istride.resize(_ndimension);
@@ -94,8 +96,10 @@ public:
 	  // Use a reduced simd grid
 	  _ldimensions[d]= _gdimensions[d]/_processors[d];  //local dimensions
 	  _rdimensions[d]= _ldimensions[d]/_simd_layout[d]; //overdecomposition
-	  _osites *= _rdimensions[d];
-	  _isites *= _simd_layout[d];
+	  _lstart[d]     = _processor_coor[d]*_ldimensions[d];
+	  _lend[d]       = _processor_coor[d]*_ldimensions[d]+_ldimensions[d]-1;
+	  _osites  *= _rdimensions[d];
+	  _isites  *= _simd_layout[d];
                 
 	  // Addressing support
 	  if ( d==0 ) {

lib/cartesian/Cartesian_red_black.h

@@ -151,6 +151,8 @@ public:
       _ldimensions.resize(_ndimension);
       _rdimensions.resize(_ndimension);
       _simd_layout.resize(_ndimension);
+      _lstart.resize(_ndimension);
+      _lend.resize(_ndimension);
       
       _ostride.resize(_ndimension);
       _istride.resize(_ndimension);
@@ -169,6 +171,8 @@ public:
 	  _gdimensions[d] = _gdimensions[d]/2; // Remove a checkerboard
 	}
 	_ldimensions[d] = _gdimensions[d]/_processors[d];
+	_lstart[d]     = _processor_coor[d]*_ldimensions[d];
+	_lend[d]       = _processor_coor[d]*_ldimensions[d]+_ldimensions[d]-1;
 
 	// Use a reduced simd grid
 	_simd_layout[d] = simd_layout[d];

lib/communicator/Communicator_base.cc

@@ -25,7 +25,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid/Grid.h>
+#include <Grid/GridCore.h>
+
 namespace Grid {
 
 ///////////////////////////////////////////////////////////////
@@ -33,6 +34,7 @@ namespace Grid {
 ///////////////////////////////////////////////////////////////
 void *              CartesianCommunicator::ShmCommBuf;
 uint64_t            CartesianCommunicator::MAX_MPI_SHM_BYTES   = 128*1024*1024; 
+CartesianCommunicator::CommunicatorPolicy_t  CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
 
 /////////////////////////////////
 // Alloc, free shmem region
@@ -58,6 +60,7 @@ void CartesianCommunicator::ShmBufferFreeAll(void) {
 /////////////////////////////////
 // Grid information queries
 /////////////////////////////////
+int                      CartesianCommunicator::Dimensions(void)         { return _ndimension; };
 int                      CartesianCommunicator::IsBoss(void)            { return _processor==0; };
 int                      CartesianCommunicator::BossRank(void)          { return 0; };
 int                      CartesianCommunicator::ThisRank(void)          { return _processor; };
@@ -88,7 +91,10 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
 
 #if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPI3L)
 
-void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+int                      CartesianCommunicator::NodeCount(void)    { return ProcessorCount();};
+int                      CartesianCommunicator::RankCount(void)    { return ProcessorCount();};
+
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						       void *xmit,
 						       int xmit_to_rank,
 						       void *recv,
@@ -96,6 +102,7 @@ void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_
 						       int bytes)
 {
   SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+  return 2.0*bytes;
 }
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
 {

lib/communicator/Communicator_base.h

@@ -116,6 +116,12 @@ class CartesianCommunicator {
   // Implemented in Communicator_base.C
   /////////////////////////////////
   static void * ShmCommBuf;
+
+  // Isend/Irecv/Wait, or Sendrecv blocking
+  enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
+  static CommunicatorPolicy_t CommunicatorPolicy;
+  static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
+
   size_t heap_top;
   size_t heap_bytes;
 
@@ -142,12 +148,15 @@ class CartesianCommunicator {
   int  RankFromProcessorCoor(std::vector<int> &coor);
   void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
   
+  int                      Dimensions(void)        ;
   int                      IsBoss(void)            ;
   int                      BossRank(void)          ;
   int                      ThisRank(void)          ;
   const std::vector<int> & ThisProcessorCoor(void) ;
   const std::vector<int> & ProcessorGrid(void)     ;
   int                      ProcessorCount(void)    ;
+  int                      NodeCount(void)    ;
+  int                      RankCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
@@ -168,6 +177,8 @@ class CartesianCommunicator {
   void GlobalSumVector(ComplexF *c,int N);
   void GlobalSum(ComplexD &c);
   void GlobalSumVector(ComplexD *c,int N);
+  void GlobalXOR(uint32_t &);
+  void GlobalXOR(uint64_t &);
   
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
@@ -200,7 +211,7 @@ class CartesianCommunicator {
   
   void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
 
-  void StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+  double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 				  void *xmit,
 				  int xmit_to_rank,
 				  void *recv,

lib/communicator/Communicator_mpi.cc

@@ -25,7 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid/Grid.h>
+#include <Grid/GridCore.h>
+#include <Grid/GridQCDcore.h>
+#include <Grid/qcd/action/ActionCore.h>
 #include <mpi.h>
 
 namespace Grid {
@@ -39,9 +41,13 @@ MPI_Comm CartesianCommunicator::communicator_world;
 // Should error check all MPI calls.
 void CartesianCommunicator::Init(int *argc, char ***argv) {
   int flag;
+  int provided;
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
-    MPI_Init(argc,argv);
+    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+    if ( provided != MPI_THREAD_MULTIPLE ) {
+      QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
+    }
   }
   MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
   ShmInitGeneric();
@@ -77,6 +83,14 @@ void CartesianCommunicator::GlobalSum(uint64_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
   assert(ierr==0);
 }
+void CartesianCommunicator::GlobalXOR(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint64_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
+  assert(ierr==0);
+}
 void CartesianCommunicator::GlobalSum(float &f){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
@@ -152,24 +166,34 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
 						int from,
 						int bytes)
 {
-  MPI_Request xrq;
-  MPI_Request rrq;
-  int rank = _processor;
+  int myrank = _processor;
   int ierr;
-  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-  ierr|=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    MPI_Request xrq;
+    MPI_Request rrq;
 
-  assert(ierr==0);
+    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
     
-  list.push_back(xrq);
-  list.push_back(rrq);
+    assert(ierr==0);
+    list.push_back(xrq);
+    list.push_back(rrq);
+  } else { 
+    // Give the CPU to MPI immediately; can use threads to overlap optionally
+    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
+		      recv,bytes,MPI_CHAR,from, from,
+		      communicator,MPI_STATUS_IGNORE);
+    assert(ierr==0);
+  }
 }
 void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
-  int nreq=list.size();
-  std::vector<MPI_Status> status(nreq);
-  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-  assert(ierr==0);
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    int nreq=list.size();
+    std::vector<MPI_Status> status(nreq);
+    int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
+    assert(ierr==0);
+  }
 }
 
 void CartesianCommunicator::Barrier(void)

lib/communicator/Communicator_mpi3.cc

@@ -1,4 +1,4 @@
-    /*************************************************************************************
+/*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
 
@@ -25,9 +25,23 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid/Grid.h>
+#include <Grid/GridCore.h>
+
 #include <mpi.h>
 
+#include <semaphore.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <limits.h>
+#include <sys/types.h>
+#include <sys/ipc.h>
+#include <sys/shm.h>
+#include <sys/mman.h>
+//#include <zlib.h>
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
 namespace Grid {
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -50,6 +64,11 @@ std::vector<int> CartesianCommunicator::GroupRanks;
 std::vector<int> CartesianCommunicator::MyGroup;
 std::vector<void *> CartesianCommunicator::ShmCommBufs;
 
+int CartesianCommunicator::NodeCount(void)    { return GroupSize;};
+int CartesianCommunicator::RankCount(void)    { return WorldSize;};
+
+
+#undef FORCE_COMMS
 void *CartesianCommunicator::ShmBufferSelf(void)
 {
   return ShmCommBufs[ShmRank];
@@ -57,6 +76,9 @@ void *CartesianCommunicator::ShmBufferSelf(void)
 void *CartesianCommunicator::ShmBuffer(int rank)
 {
   int gpeer = GroupRanks[rank];
+#ifdef FORCE_COMMS
+  return NULL;
+#endif
   if (gpeer == MPI_UNDEFINED){
     return NULL;
   } else { 
@@ -65,7 +87,13 @@ void *CartesianCommunicator::ShmBuffer(int rank)
 }
 void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p)
 {
+  static int count =0;
   int gpeer = GroupRanks[rank];
+  assert(gpeer!=ShmRank); // never send to self
+  assert(rank!=WorldRank);// never send to self
+#ifdef FORCE_COMMS
+  return NULL;
+#endif
   if (gpeer == MPI_UNDEFINED){
     return NULL;
   } else { 
@@ -76,16 +104,27 @@ void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p)
 }
 
 void CartesianCommunicator::Init(int *argc, char ***argv) {
+
   int flag;
+  int provided;
+  //  mtrace();
+
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
-    MPI_Init(argc,argv);
+    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
+    assert (provided == MPI_THREAD_MULTIPLE);
   }
 
+  Grid_quiesce_nodes();
+
   MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
   MPI_Comm_rank(communicator_world,&WorldRank);
   MPI_Comm_size(communicator_world,&WorldSize);
 
+  if ( WorldRank == 0 ) {
+    std::cout << GridLogMessage<< "Initialising MPI "<< WorldRank <<"/"<<WorldSize <<std::endl;
+  }
+
   /////////////////////////////////////////////////////////////////////
   // Split into groups that can share memory
   /////////////////////////////////////////////////////////////////////
@@ -131,7 +170,6 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
   ///////////////////////////////////////////////////////////////////
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,communicator_world);
   assert(ierr==0);
-  
   ///////////////////////////////////////////////////////////////////
   // find the group leaders world rank
   ///////////////////////////////////////////////////////////////////
@@ -141,7 +179,6 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
       leaders_group[group++] = l;
     }
   }
-  
   ///////////////////////////////////////////////////////////////////
   // Identify the rank of the group in which I (and my leader) live
   ///////////////////////////////////////////////////////////////////
@@ -152,39 +189,114 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
     }
   }
   assert(GroupRank!=-1);
-  
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // allocate the shared window for our group
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(ShmComm);
 
   ShmCommBuf = 0;
-  ierr = MPI_Win_allocate_shared(MAX_MPI_SHM_BYTES,1,MPI_INFO_NULL,ShmComm,&ShmCommBuf,&ShmWindow);
-  assert(ierr==0);
-  // KNL hack -- force to numa-domain 1 in flat
-#if 0
-  //#include <numaif.h>
-  for(uint64_t page=0;page<MAX_MPI_SHM_BYTES;page+=4096){
-    void *pages = (void *) ( page + ShmCommBuf );
-    int status;
-    int flags=MPOL_MF_MOVE_ALL;
-    int nodes=1; // numa domain == MCDRAM
-    unsigned long count=1;
-    ierr= move_pages(0,count, &pages,&nodes,&status,flags);
-    if (ierr && (page==0)) perror("numa relocate command failed");
-  }
-#endif
-  MPI_Win_lock_all (MPI_MODE_NOCHECK, ShmWindow);
-  
-  /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Plan: allocate a fixed SHM region. Scratch that is just used via some scheme during stencil comms, with no allocate free.
-  /////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
   ShmCommBufs.resize(ShmSize);
-  for(int r=0;r<ShmSize;r++){
-    MPI_Aint sz;
-    int dsp_unit;
-    MPI_Win_shared_query (ShmWindow, r, &sz, &dsp_unit, &ShmCommBufs[r]);
+
+#if 1
+  char shm_name [NAME_MAX];
+  if ( ShmRank == 0 ) {
+    for(int r=0;r<ShmSize;r++){
+
+      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
+
+      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",GroupRank,r);
+
+      shm_unlink(shm_name);
+      int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
+      if ( fd < 0 ) {	perror("failed shm_open");	assert(0);      }
+      ftruncate(fd, size);
+
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      assert(((uint64_t)ptr&0x3F)==0);
+      ShmCommBufs[r] =ptr;
+      
+    }
   }
 
+  MPI_Barrier(ShmComm);
+
+  if ( ShmRank != 0 ) { 
+    for(int r=0;r<ShmSize;r++){
+      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES ;
+    
+      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",GroupRank,r);
+
+      int fd=shm_open(shm_name,O_RDWR,0666);
+      if ( fd<0 ) {	perror("failed shm_open");	assert(0);      }
+
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      assert(((uint64_t)ptr&0x3F)==0);
+      ShmCommBufs[r] =ptr;
+    }
+  }
+
+#else
+  std::vector<int> shmids(ShmSize);
+
+  if ( ShmRank == 0 ) {
+    for(int r=0;r<ShmSize;r++){
+      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
+      key_t key   = 0x4545 + r;
+      if ((shmids[r]= shmget(key,size, SHM_HUGETLB | IPC_CREAT | SHM_R | SHM_W)) < 0) {
+	int errsv = errno;
+	printf("Errno %d\n",errsv);
+	perror("shmget");
+	exit(1);
+      }
+      printf("shmid: 0x%x\n", shmids[r]);
+    }
+  }
+  MPI_Barrier(ShmComm);
+  MPI_Bcast(&shmids[0],ShmSize*sizeof(int),MPI_BYTE,0,ShmComm);
+  MPI_Barrier(ShmComm);
+
+  for(int r=0;r<ShmSize;r++){
+    ShmCommBufs[r] = (uint64_t *)shmat(shmids[r], NULL,0);
+    if (ShmCommBufs[r] == (uint64_t *)-1) {
+      perror("Shared memory attach failure");
+      shmctl(shmids[r], IPC_RMID, NULL);
+      exit(2);
+    }
+    printf("shmaddr: %p\n", ShmCommBufs[r]);
+  }
+  MPI_Barrier(ShmComm);
+  // Mark for clean up
+  for(int r=0;r<ShmSize;r++){
+    shmctl(shmids[r], IPC_RMID,(struct shmid_ds *)NULL);
+  }
+  MPI_Barrier(ShmComm);
+
+#endif
+  ShmCommBuf         = ShmCommBufs[ShmRank];
+
+  MPI_Barrier(ShmComm);
+  if ( ShmRank == 0 ) {
+    for(int r=0;r<ShmSize;r++){
+      uint64_t * check = (uint64_t *) ShmCommBufs[r];
+      check[0] = GroupRank;
+      check[1] = r;
+      check[2] = 0x5A5A5A;
+    }
+  }
+
+  MPI_Barrier(ShmComm);
+  for(int r=0;r<ShmSize;r++){
+    uint64_t * check = (uint64_t *) ShmCommBufs[r];
+    
+    assert(check[0]==GroupRank);
+    assert(check[1]==r);
+    assert(check[2]==0x5A5A5A);
+
+  }
+  MPI_Barrier(ShmComm);
+
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
   // Verbose for now
   //////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -192,7 +304,7 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
     std::cout<<GridLogMessage<< "Grid MPI-3 configuration: detected ";
     std::cout<< WorldSize << " Ranks " ;
     std::cout<< GroupSize << " Nodes " ;
-    std::cout<<  ShmSize  << " with ranks-per-node "<<std::endl;
+    std::cout<< " with "<< ShmSize  << " ranks-per-node "<<std::endl;
     
     std::cout<<GridLogMessage     <<"Grid MPI-3 configuration: allocated shared memory region of size ";
     std::cout<<std::hex << MAX_MPI_SHM_BYTES <<" ShmCommBuf address = "<<ShmCommBuf << std::dec<<std::endl;
@@ -207,7 +319,6 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
       if(g!=ShmSize-1) std::cout<<",";
       else std::cout<<"}"<<std::endl;
     }
-
   }
   
   for(int g=0;g<GroupSize;g++){
@@ -216,7 +327,7 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
       if ( (ShmRank == 0) && (GroupRank==g) ) {
 	std::cout<<MyGroup[r];
 	if(r<ShmSize-1) std::cout<<",";
-	else std::cout<<"}"<<std::endl;
+	else std::cout<<"}"<<std::endl<<std::flush;
       }
       MPI_Barrier(communicator_world);
     }
@@ -225,14 +336,12 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
   assert(ShmSetup==0);  ShmSetup=1;
 }
 
-
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Want to implement some magic ... Group sub-cubes into those on same node
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
-void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
+void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &dest,int &source)
 {
-  std::vector<int> coor = _processor_coor;
-
+  std::vector<int> coor = _processor_coor; // my coord
   assert(std::abs(shift) <_processors[dim]);
 
   coor[dim] = (_processor_coor[dim] + shift + _processors[dim])%_processors[dim];
@@ -242,26 +351,30 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
   coor[dim] = (_processor_coor[dim] - shift + _processors[dim])%_processors[dim];
   Lexicographic::IndexFromCoor(coor,dest,_processors);
   dest = LexicographicToWorldRank[dest];
-}
+
+}// rank is world rank.
+
 int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
 {
   int rank;
   Lexicographic::IndexFromCoor(coor,rank,_processors);
   rank = LexicographicToWorldRank[rank];
   return rank;
-}
+}// rank is world rank
+
 void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
 {
-  Lexicographic::CoorFromIndex(coor,rank,_processors);
-  rank = LexicographicToWorldRank[rank];
+  int lr=-1;
+  for(int r=0;r<WorldSize;r++){// map world Rank to lexico and then to coor
+    if( LexicographicToWorldRank[r]==rank) lr = r;
+  }
+  assert(lr!=-1);
+  Lexicographic::CoorFromIndex(coor,lr,_processors);
 }
-
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 { 
   int ierr;
-
   communicator=communicator_world;
-
   _ndimension = processors.size();
 
   ////////////////////////////////////////////////////////////////
@@ -280,19 +393,17 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
   // Identify subblock of ranks on node spreading across dims
   // in a maximally symmetrical way
   ////////////////////////////////////////////////////////////////
-  int dim = 0;
-  
   std::vector<int> WorldDims = processors;
 
-  ShmDims.resize(_ndimension,1);
+  ShmDims.resize  (_ndimension,1);
   GroupDims.resize(_ndimension);
-    
-  ShmCoor.resize(_ndimension);
+  ShmCoor.resize  (_ndimension);
   GroupCoor.resize(_ndimension);
   WorldCoor.resize(_ndimension);
 
+  int dim = 0;
   for(int l2=0;l2<log2size;l2++){
-    while ( WorldDims[dim] / ShmDims[dim] <= 1 ) dim=(dim+1)%_ndimension;
+    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%_ndimension;
     ShmDims[dim]*=2;
     dim=(dim+1)%_ndimension;
   }
@@ -304,6 +415,29 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
     GroupDims[d] = WorldDims[d]/ShmDims[d];
   }
 
+  ////////////////////////////////////////////////////////////////
+  // Verbose
+  ////////////////////////////////////////////////////////////////
+#if 0
+  std::cout<< GridLogMessage << "MPI-3 usage "<<std::endl;
+  std::cout<< GridLogMessage << "SHM   ";
+  for(int d=0;d<_ndimension;d++){
+    std::cout<< ShmDims[d] <<" ";
+  }
+  std::cout<< std::endl;
+
+  std::cout<< GridLogMessage << "Group ";
+  for(int d=0;d<_ndimension;d++){
+    std::cout<< GroupDims[d] <<" ";
+  }
+  std::cout<< std::endl;
+
+  std::cout<< GridLogMessage<<"World ";
+  for(int d=0;d<_ndimension;d++){
+    std::cout<< WorldDims[d] <<" ";
+  }
+  std::cout<< std::endl;
+#endif
   ////////////////////////////////////////////////////////////////
   // Check processor counts match
   ////////////////////////////////////////////////////////////////
@@ -317,29 +451,57 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
       
   ////////////////////////////////////////////////////////////////
   // Establish mapping between lexico physics coord and WorldRank
-  // 
   ////////////////////////////////////////////////////////////////
-  LexicographicToWorldRank.resize(WorldSize,0);
   Lexicographic::CoorFromIndex(GroupCoor,GroupRank,GroupDims);
   Lexicographic::CoorFromIndex(ShmCoor,ShmRank,ShmDims);
   for(int d=0;d<_ndimension;d++){
     WorldCoor[d] = GroupCoor[d]*ShmDims[d]+ShmCoor[d];
   }
   _processor_coor = WorldCoor;
-
-  int lexico;
-  Lexicographic::IndexFromCoor(WorldCoor,lexico,WorldDims);
-  LexicographicToWorldRank[lexico]=WorldRank;
-  _processor = lexico;
+  _processor      = WorldRank;
 
   ///////////////////////////////////////////////////////////////////
   // global sum Lexico to World mapping
   ///////////////////////////////////////////////////////////////////
+  int lexico;
+  LexicographicToWorldRank.resize(WorldSize,0);
+  Lexicographic::IndexFromCoor(WorldCoor,lexico,WorldDims);
+  LexicographicToWorldRank[lexico] = WorldRank;
   ierr=MPI_Allreduce(MPI_IN_PLACE,&LexicographicToWorldRank[0],WorldSize,MPI_INT,MPI_SUM,communicator);
   assert(ierr==0);
 
-};
+  for(int i=0;i<WorldSize;i++){
 
+    int wr = LexicographicToWorldRank[i];
+    //    int wr = i;
+
+    std::vector<int> coor(_ndimension);
+    ProcessorCoorFromRank(wr,coor); // from world rank
+    int ck = RankFromProcessorCoor(coor);
+    assert(ck==wr);
+
+    if ( wr == WorldRank ) { 
+      for(int j=0;j<coor.size();j++) {
+	assert(coor[j] == _processor_coor[j]);
+      }
+    }
+    /*
+    std::cout << GridLogMessage<< " Lexicographic "<<i;
+    std::cout << " MPI rank      "<<wr;
+    std::cout << " Coor          ";
+    for(int j=0;j<coor.size();j++) std::cout << coor[j];
+    std::cout<< std::endl;
+    */
+    /////////////////////////////////////////////////////
+    // Check everyone agrees on everyone elses coords
+    /////////////////////////////////////////////////////
+    std::vector<int> mcoor = coor;
+    this->Broadcast(0,(void *)&mcoor[0],mcoor.size()*sizeof(int));
+    for(int d = 0 ; d< _ndimension; d++) {
+      assert(coor[d] == mcoor[d]);
+    }
+  }
+};
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
   assert(ierr==0);
@@ -348,6 +510,14 @@ void CartesianCommunicator::GlobalSum(uint64_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
   assert(ierr==0);
 }
+void CartesianCommunicator::GlobalXOR(uint32_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
+  assert(ierr==0);
+}
+void CartesianCommunicator::GlobalXOR(uint64_t &u){
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
+  assert(ierr==0);
+}
 void CartesianCommunicator::GlobalSum(float &f){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
   assert(ierr==0);
@@ -367,8 +537,6 @@ void CartesianCommunicator::GlobalSumVector(double *d,int N)
   int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
   assert(ierr==0);
 }
-
-
 // Basic Halo comms primitive
 void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int dest,
@@ -377,10 +545,14 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
 					   int bytes)
 {
   std::vector<CommsRequest_t> reqs(0);
+  //    unsigned long  xcrc = crc32(0L, Z_NULL, 0);
+  //    unsigned long  rcrc = crc32(0L, Z_NULL, 0);
+  //    xcrc = crc32(xcrc,(unsigned char *)xmit,bytes);
   SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
   SendToRecvFromComplete(reqs);
+  //    rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
+  //    printf("proc %d SendToRecvFrom %d bytes %lx %lx\n",_processor,bytes,xcrc,rcrc);
 }
-
 void CartesianCommunicator::SendRecvPacket(void *xmit,
 					   void *recv,
 					   int sender,
@@ -397,7 +569,6 @@ void CartesianCommunicator::SendRecvPacket(void *xmit,
     MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
   }
 }
-
 // Basic Halo comms primitive
 void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						void *xmit,
@@ -406,95 +577,29 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
 						int from,
 						int bytes)
 {
-#if 0
-  this->StencilBarrier();
-
-  MPI_Request xrq;
-  MPI_Request rrq;
-  
-  static int sequence;
-
+  int myrank = _processor;
   int ierr;
-  int tag;
-  int check;
 
-  assert(dest != _processor);
-  assert(from != _processor);
+  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
+    MPI_Request xrq;
+    MPI_Request rrq;
 
-  int gdest = GroupRanks[dest];
-  int gfrom = GroupRanks[from];
-  int gme   = GroupRanks[_processor];
+    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
     
-  sequence++;
-  
-  char *from_ptr = (char *)ShmCommBufs[ShmRank];
-
-  int small = (bytes<MAX_MPI_SHM_BYTES);
-
-  typedef uint64_t T;
-  int words = bytes/sizeof(T);
-
-  assert(((size_t)bytes &(sizeof(T)-1))==0);
-  assert(gme == ShmRank);
-
-  if ( small && (gdest !=MPI_UNDEFINED) ) {
-
-    char *to_ptr   = (char *)ShmCommBufs[gdest];
-
-    assert(gme != gdest);
-
-    T *ip = (T *)xmit;
-    T *op = (T *)to_ptr;
-PARALLEL_FOR_LOOP 
-    for(int w=0;w<words;w++) {
-      op[w]=ip[w];
-    }
-
-    bcopy(&_processor,&to_ptr[bytes],sizeof(_processor));
-    bcopy(&  sequence,&to_ptr[bytes+4],sizeof(sequence));
-  } else { 
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
     assert(ierr==0);
     list.push_back(xrq);
-  }
-
-  this->StencilBarrier();
-  
-  if (small && (gfrom !=MPI_UNDEFINED) ) {
-    T *ip = (T *)from_ptr;
-    T *op = (T *)recv;
-PARALLEL_FOR_LOOP 
-    for(int w=0;w<words;w++) {
-      op[w]=ip[w];
-    }
-    bcopy(&from_ptr[bytes]  ,&tag  ,sizeof(tag));
-    bcopy(&from_ptr[bytes+4],&check,sizeof(check));
-    assert(check==sequence);
-    assert(tag==from);
-  } else { 
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-    assert(ierr==0);
     list.push_back(rrq);
+  } else { 
+    // Give the CPU to MPI immediately; can use threads to overlap optionally
+    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
+		      recv,bytes,MPI_CHAR,from, from,
+		      communicator,MPI_STATUS_IGNORE);
+    assert(ierr==0);
   }
-
-  this->StencilBarrier();
-
-#else
-  MPI_Request xrq;
-  MPI_Request rrq;
-  int rank = _processor;
-  int ierr;
-  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-  ierr|=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-  
-  assert(ierr==0);
-
-  list.push_back(xrq);
-  list.push_back(rrq);
-#endif
 }
 
-void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
 						       void *xmit,
 						       int dest,
 						       void *recv,
@@ -505,57 +610,63 @@ void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_
   MPI_Request rrq;
 
   int ierr;
-
-  assert(dest != _processor);
-  assert(from != _processor);
-  
   int gdest = GroupRanks[dest];
   int gfrom = GroupRanks[from];
   int gme   = GroupRanks[_processor];
 
-  assert(gme == ShmRank);
+  assert(dest != _processor);
+  assert(from != _processor);
+  assert(gme  == ShmRank);
+  double off_node_bytes=0.0;
+
+#ifdef FORCE_COMMS
+  gdest = MPI_UNDEFINED;
+  gfrom = MPI_UNDEFINED;
+#endif
+  if ( gfrom ==MPI_UNDEFINED) {
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
+    assert(ierr==0);
+    list.push_back(rrq);
+    off_node_bytes+=bytes;
+  }
 
   if ( gdest == MPI_UNDEFINED ) {
     ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
     assert(ierr==0);
     list.push_back(xrq);
+    off_node_bytes+=bytes;
   }
 
-  if ( gfrom ==MPI_UNDEFINED) {
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-    assert(ierr==0);
-    list.push_back(rrq);
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) { 
+    this->StencilSendToRecvFromComplete(list);
   }
 
+  return off_node_bytes;
 }
-
-
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list)
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
 {
-  SendToRecvFromComplete(list);
+  SendToRecvFromComplete(waitall);
 }
-
 void CartesianCommunicator::StencilBarrier(void)
 {
-  MPI_Win_sync (ShmWindow);   
   MPI_Barrier  (ShmComm);
-  MPI_Win_sync (ShmWindow);   
 }
-
 void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
   int nreq=list.size();
+
+  if (nreq==0) return;
+
   std::vector<MPI_Status> status(nreq);
   int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
   assert(ierr==0);
+  list.resize(0);
 }
-
 void CartesianCommunicator::Barrier(void)
 {
   int ierr = MPI_Barrier(communicator);
   assert(ierr==0);
 }
-
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 {
   int ierr=MPI_Bcast(data,
@@ -565,7 +676,11 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
 		     communicator);
   assert(ierr==0);
 }
-
+int CartesianCommunicator::RankWorld(void){ 
+  int r; 
+  MPI_Comm_rank(communicator_world,&r);
+  return r;
+}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 {
   int ierr= MPI_Bcast(data,

lib/communicator/Communicator_mpi3_leader.cc

@@ -27,6 +27,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     /*  END LEGAL */
 #include "Grid.h"
 #include <mpi.h>
+//#include <numaif.h>
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 /// Workarounds:
@@ -42,19 +43,27 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <fcntl.h>
 #include <unistd.h>
 #include <limits.h>
-
 typedef sem_t *Grid_semaphore;
 
+
+#error  /*THis is deprecated*/
+
+#if 0 
 #define SEM_INIT(S)      S = sem_open(sem_name,0,0600,0); assert ( S != SEM_FAILED );
 #define SEM_INIT_EXCL(S) sem_unlink(sem_name); S = sem_open(sem_name,O_CREAT|O_EXCL,0600,0); assert ( S != SEM_FAILED );
 #define SEM_POST(S) assert ( sem_post(S) == 0 ); 
 #define SEM_WAIT(S) assert ( sem_wait(S) == 0 );
-
+#else
+#define SEM_INIT(S)      ;
+#define SEM_INIT_EXCL(S) ;
+#define SEM_POST(S) ;
+#define SEM_WAIT(S) ;
+#endif
 #include <sys/mman.h>
 
 namespace Grid {
 
-enum { COMMAND_ISEND, COMMAND_IRECV, COMMAND_WAITALL };
+enum { COMMAND_ISEND, COMMAND_IRECV, COMMAND_WAITALL, COMMAND_SENDRECV };
 
 struct Descriptor {
   uint64_t buf;
@@ -62,6 +71,12 @@ struct Descriptor {
   int rank;
   int tag;
   int command;
+  uint64_t xbuf;
+  uint64_t rbuf;
+  int xtag;
+  int rtag;
+  int src;
+  int dest;
   MPI_Request request;
 };
 
@@ -94,18 +109,14 @@ public:
 
   void SemInit(void) {
     sprintf(sem_name,"/Grid_mpi3_sem_head_%d",universe_rank);
-    //    printf("SEM_NAME: %s \n",sem_name);
     SEM_INIT(sem_head);
     sprintf(sem_name,"/Grid_mpi3_sem_tail_%d",universe_rank);
-    //    printf("SEM_NAME: %s \n",sem_name);
     SEM_INIT(sem_tail);
   }  
   void SemInitExcl(void) {
     sprintf(sem_name,"/Grid_mpi3_sem_head_%d",universe_rank);
-    //    printf("SEM_INIT_EXCL: %s \n",sem_name);
     SEM_INIT_EXCL(sem_head);
     sprintf(sem_name,"/Grid_mpi3_sem_tail_%d",universe_rank);
-    //    printf("SEM_INIT_EXCL: %s \n",sem_name);
     SEM_INIT_EXCL(sem_tail);
   }  
   void WakeUpDMA(void) { 
@@ -125,6 +136,13 @@ public:
     while(1){
       WaitForCommand();
       //      std::cout << "Getting command "<<std::endl;
+#if 0
+      _mm_monitor((void *)&state->head,0,0);
+      int s=state->start;
+      if ( s != state->head ) {
+	_mm_mwait(0,0);
+      }
+#endif
       Event();
     }
   }
@@ -132,6 +150,7 @@ public:
   int Event (void) ;
 
   uint64_t QueueCommand(int command,void *buf, int bytes, int hashtag, MPI_Comm comm,int u_rank) ;
+  void QueueSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) ;
 
   void WaitAll() {
     //    std::cout << "Queueing WAIT command  "<<std::endl;
@@ -141,7 +160,7 @@ public:
     //    std::cout << "Waiting from semaphore "<<std::endl;
     WaitForComplete();
     //    std::cout << "Checking FIFO is empty "<<std::endl;
-    assert ( state->tail == state->head );
+    while ( state->tail != state->head );
   }
 };
 
@@ -196,6 +215,12 @@ public:
     //    std::cout << "Waking up DMA "<< slave<<std::endl;
   };
 
+  static void QueueSendRecv(int slave,void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) 
+  {
+    Slaves[slave].QueueSendRecv(xbuf,rbuf,bytes,xtag,rtag,comm,dest,src);
+    Slaves[slave].WakeUpDMA();
+  }
+
   static void QueueRecv(int slave, void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
     //    std::cout<< " Queueing recv "<< bytes<< " slave "<< slave << " from comm "<<rank  <<std::endl;
     Slaves[slave].QueueCommand(COMMAND_IRECV,buf,bytes,tag,comm,rank);
@@ -226,6 +251,28 @@ public:
     return;
   };
 
+  static void QueueRoundRobinSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) {
+    uint8_t * cxbuf = (uint8_t *) xbuf;
+    uint8_t * crbuf = (uint8_t *) rbuf;
+    static int rrp=0;
+    int procs = VerticalSize-1;
+    int myoff=0;
+    int mywork=bytes;
+    QueueSendRecv(rrp+1,&cxbuf[myoff],&crbuf[myoff],mywork,xtag,rtag,comm,dest,src);
+    rrp = rrp+1;
+    if ( rrp == (VerticalSize-1) ) rrp = 0;
+  }
+
+  static void QueueMultiplexedSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) {
+    uint8_t * cxbuf = (uint8_t *) xbuf;
+    uint8_t * crbuf = (uint8_t *) rbuf;
+    int mywork, myoff, procs;
+    procs = VerticalSize-1;
+    for(int s=0;s<procs;s++) {
+      GetWork(bytes,s,mywork,myoff,procs);
+      QueueSendRecv(s+1,&cxbuf[myoff],&crbuf[myoff],mywork,xtag,rtag,comm,dest,src);
+    }
+  };
   static void QueueMultiplexedSend(void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
     uint8_t * cbuf = (uint8_t *) buf;
     int mywork, myoff, procs;
@@ -275,6 +322,7 @@ std::vector<void *>            MPIoffloadEngine::VerticalShmBufs;
 std::vector<std::vector<int> > MPIoffloadEngine::UniverseRanks;
 std::vector<int>               MPIoffloadEngine::UserCommunicatorToWorldRanks; 
 
+int CartesianCommunicator::NodeCount(void)    { return HorizontalSize;};
 int MPIoffloadEngine::ShmSetup = 0;
 
 void MPIoffloadEngine::CommunicatorInit (MPI_Comm &communicator_world,
@@ -370,12 +418,22 @@ void MPIoffloadEngine::CommunicatorInit (MPI_Comm &communicator_world,
       ftruncate(fd, size);
 
       VerticalShmBufs[r] = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-
       if ( VerticalShmBufs[r] == MAP_FAILED ) { 
 	perror("failed mmap");
 	assert(0);
       }
 
+      /*
+      for(uint64_t page=0;page<size;page+=4096){
+	void *pages = (void *) ( page + (uint64_t)VerticalShmBufs[r] );
+	int status;
+	int flags=MPOL_MF_MOVE_ALL;
+	int nodes=1; // numa domain == MCDRAM
+	unsigned long count=1;
+	ierr= move_pages(0,count, &pages,&nodes,&status,flags);
+	if (ierr && (page==0)) perror("numa relocate command failed");
+      }
+      */
       uint64_t * check = (uint64_t *) VerticalShmBufs[r];
       check[0] = WorldRank;
       check[1] = r;
@@ -404,7 +462,7 @@ void MPIoffloadEngine::CommunicatorInit (MPI_Comm &communicator_world,
     uint64_t * check = (uint64_t *) VerticalShmBufs[r];
     assert(check[0]== WorldRank);
     assert(check[1]== r);
-    std::cerr<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
+    //    std::cerr<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
   }
   }
 #endif
@@ -542,6 +600,8 @@ int Slave::Event (void) {
   static int head_last;
   static int start_last;
   int ierr;
+  MPI_Status stat;
+  static int i=0;
 
   ////////////////////////////////////////////////////
   // Try to advance the start pointers
@@ -550,11 +610,6 @@ int Slave::Event (void) {
   if ( s != state->head ) {
     switch ( state->Descrs[s].command ) {
     case COMMAND_ISEND:
-      /*
-            std::cout<< " Send "<<s << " ptr "<< state<<" "<< state->Descrs[s].buf<< "["<<state->Descrs[s].bytes<<"]"
-      	       << " to " << state->Descrs[s].rank<< " tag" << state->Descrs[s].tag
-       << " Comm " << MPIoffloadEngine::communicator_universe<< " me " <<universe_rank<< std::endl;
-      */
       ierr = MPI_Isend((void *)(state->Descrs[s].buf+base), 
 		       state->Descrs[s].bytes, 
 		       MPI_CHAR,
@@ -568,11 +623,6 @@ int Slave::Event (void) {
       break;
 
     case COMMAND_IRECV:
-      /*
-      std::cout<< " Recv "<<s << " ptr "<< state<<" "<< state->Descrs[s].buf<< "["<<state->Descrs[s].bytes<<"]"
-	       << " from " << state->Descrs[s].rank<< " tag" << state->Descrs[s].tag
-	       << " Comm " << MPIoffloadEngine::communicator_universe<< " me "<< universe_rank<< std::endl;
-      */
       ierr=MPI_Irecv((void *)(state->Descrs[s].buf+base), 
 		     state->Descrs[s].bytes, 
 		     MPI_CHAR,
@@ -588,10 +638,32 @@ int Slave::Event (void) {
       return 1;
       break;
 
+    case COMMAND_SENDRECV:
+
+      //      fprintf(stderr,"Sendrecv ->%d %d : <-%d %d \n",state->Descrs[s].dest, state->Descrs[s].xtag+i*10,state->Descrs[s].src, state->Descrs[s].rtag+i*10);
+
+      ierr=MPI_Sendrecv((void *)(state->Descrs[s].xbuf+base), state->Descrs[s].bytes, MPI_CHAR, state->Descrs[s].dest, state->Descrs[s].xtag+i*10,
+			(void *)(state->Descrs[s].rbuf+base), state->Descrs[s].bytes, MPI_CHAR, state->Descrs[s].src , state->Descrs[s].rtag+i*10,
+			MPIoffloadEngine::communicator_universe,MPI_STATUS_IGNORE);
+
+      assert(ierr==0);
+
+      //      fprintf(stderr,"Sendrecv done %d %d\n",ierr,i);
+      //      MPI_Barrier(MPIoffloadEngine::HorizontalComm);
+      //      fprintf(stderr,"Barrier\n");
+      i++;
+
+      state->start = PERI_PLUS(s);
+
+      return 1;
+      break;
+
     case COMMAND_WAITALL:
 
       for(int t=state->tail;t!=s; t=PERI_PLUS(t) ){
-	MPI_Wait((MPI_Request *)&state->Descrs[t].request,MPI_STATUS_IGNORE);
+	if ( state->Descrs[t].command != COMMAND_SENDRECV ) {
+	  MPI_Wait((MPI_Request *)&state->Descrs[t].request,MPI_STATUS_IGNORE);
+	}
       };
       s=PERI_PLUS(s);
       state->start = s;
@@ -613,6 +685,45 @@ int Slave::Event (void) {
   // External interaction with the queue
   //////////////////////////////////////////////////////////////////////////////
   
+void Slave::QueueSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) 
+{
+  int head =state->head;
+  int next = PERI_PLUS(head);
+  
+  // Set up descriptor
+  int worldrank;
+  int hashtag;
+  MPI_Comm    communicator;
+  MPI_Request request;
+  uint64_t relative;
+  
+  relative = (uint64_t)xbuf - base;
+  state->Descrs[head].xbuf    = relative;
+  
+  relative= (uint64_t)rbuf - base;
+  state->Descrs[head].rbuf    = relative;
+  
+  state->Descrs[head].bytes  = bytes;
+  
+  MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,xtag,comm,dest);
+  state->Descrs[head].dest   = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
+  state->Descrs[head].xtag    = hashtag;
+  
+  MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,rtag,comm,src);
+  state->Descrs[head].src    = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
+  state->Descrs[head].rtag    = hashtag;
+  
+  state->Descrs[head].command= COMMAND_SENDRECV;
+  
+  // Block until FIFO has space
+  while( state->tail==next );
+  
+  // Msync on weak order architectures
+  
+  // Advance pointer
+  state->head = next;
+  
+};
 uint64_t Slave::QueueCommand(int command,void *buf, int bytes, int tag, MPI_Comm comm,int commrank) 
 {
   /////////////////////////////////////////
@@ -812,19 +923,22 @@ void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_
   assert( (recv_i >= shm) && (recv_i+bytes <= shm+MAX_MPI_SHM_BYTES) );
   assert(from!=_processor);
   assert(dest!=_processor);
-  MPIoffloadEngine::QueueMultiplexedSend(xmit,bytes,_processor,communicator,dest);
-  MPIoffloadEngine::QueueMultiplexedRecv(recv,bytes,from,communicator,from);
-}
 
+  MPIoffloadEngine::QueueMultiplexedSendRecv(xmit,recv,bytes,_processor,from,communicator,dest,from);
+
+  //MPIoffloadEngine::QueueRoundRobinSendRecv(xmit,recv,bytes,_processor,from,communicator,dest,from);
+
+  //MPIoffloadEngine::QueueMultiplexedSend(xmit,bytes,_processor,communicator,dest);
+  //MPIoffloadEngine::QueueMultiplexedRecv(recv,bytes,from,communicator,from);
+}
 
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
   MPIoffloadEngine::WaitAll();
+  //this->Barrier();
 }
 
-void CartesianCommunicator::StencilBarrier(void)
-{
-}
+void CartesianCommunicator::StencilBarrier(void) { }
 
 void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {

lib/communicator/Communicator_none.cc

@@ -25,7 +25,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#include <Grid/Grid.h>
+#include <Grid/GridCore.h>
+
 namespace Grid {
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -58,6 +59,8 @@ void CartesianCommunicator::GlobalSum(double &){}
 void CartesianCommunicator::GlobalSum(uint32_t &){}
 void CartesianCommunicator::GlobalSum(uint64_t &){}
 void CartesianCommunicator::GlobalSumVector(double *,int N){}
+void CartesianCommunicator::GlobalXOR(uint32_t &){}
+void CartesianCommunicator::GlobalXOR(uint64_t &){}
 
 void CartesianCommunicator::SendRecvPacket(void *xmit,
 					   void *recv,
@@ -87,6 +90,7 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
 {
   assert(0);
 }
+
 void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
   assert(0);
@@ -97,7 +101,7 @@ void CartesianCommunicator::Barrier(void){}
 void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
 void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
 int  CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor) {  return 0;}
-void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor){ coor = _processor_coor ;}
+void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor){  coor = _processor_coor; }
 void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
 {
   source =0;

lib/communicator/Communicator_shmem.cc

@@ -27,6 +27,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     /*  END LEGAL */
 #include <Grid/Grid.h>
 #include <mpp/shmem.h>
+#include <array>
 
 namespace Grid {
 
@@ -51,7 +52,7 @@ typedef struct HandShake_t {
 } HandShake;
 
 std::array<long,_SHMEM_REDUCE_SYNC_SIZE> make_psync_init(void) {
-  array<long,_SHMEM_REDUCE_SYNC_SIZE> ret;
+  std::array<long,_SHMEM_REDUCE_SYNC_SIZE> ret;
   ret.fill(SHMEM_SYNC_VALUE);
   return ret;
 }
@@ -109,7 +110,7 @@ void CartesianCommunicator::GlobalSum(uint32_t &u){
 
   source = u;
   dest   = 0;
-  shmem_longlong_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync);
+  shmem_longlong_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
   shmem_barrier_all(); // necessary?
   u = dest;
 }
@@ -125,7 +126,7 @@ void CartesianCommunicator::GlobalSum(uint64_t &u){
 
   source = u;
   dest   = 0;
-  shmem_longlong_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync);
+  shmem_longlong_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
   shmem_barrier_all(); // necessary?
   u = dest;
 }
@@ -137,7 +138,8 @@ void CartesianCommunicator::GlobalSum(float &f){
 
   source = f;
   dest   =0.0;
-  shmem_float_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync);
+  shmem_float_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
+  shmem_barrier_all();
   f = dest;
 }
 void CartesianCommunicator::GlobalSumVector(float *f,int N)
@@ -148,14 +150,16 @@ void CartesianCommunicator::GlobalSumVector(float *f,int N)
   static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
 
   if ( shmem_addr_accessible(f,_processor)  ){
-    shmem_float_sum_to_all(f,f,N,0,0,_Nprocessors,llwrk,psync);
+    shmem_float_sum_to_all(f,f,N,0,0,_Nprocessors,llwrk,psync.data());
+    shmem_barrier_all();
     return;
   }
 
   for(int i=0;i<N;i++){
     dest   =0.0;
     source = f[i];
-    shmem_float_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync);
+    shmem_float_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
+    shmem_barrier_all();
     f[i] = dest;
   }
 }
@@ -168,7 +172,8 @@ void CartesianCommunicator::GlobalSum(double &d)
 
   source = d;
   dest   = 0;
-  shmem_double_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync);
+  shmem_double_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
+  shmem_barrier_all();
   d = dest;
 }
 void CartesianCommunicator::GlobalSumVector(double *d,int N)
@@ -180,14 +185,16 @@ void CartesianCommunicator::GlobalSumVector(double *d,int N)
 
 
   if ( shmem_addr_accessible(d,_processor)  ){
-    shmem_double_sum_to_all(d,d,N,0,0,_Nprocessors,llwrk,psync);
+    shmem_double_sum_to_all(d,d,N,0,0,_Nprocessors,llwrk,psync.data());
+    shmem_barrier_all();
     return;
   }
 
   for(int i=0;i<N;i++){
     source = d[i];
     dest   =0.0;
-    shmem_double_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync);
+    shmem_double_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
+    shmem_barrier_all();
     d[i] = dest;
   }
 }
@@ -282,11 +289,13 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
   SHMEM_VET(recv);
   //  shmem_putmem_nb(recv,xmit,bytes,dest,NULL);
   shmem_putmem(recv,xmit,bytes,dest);
+
+  if ( CommunicatorPolicy == CommunicatorPolicySequential ) shmem_barrier_all(); 
 }
 void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
 {
   //  shmem_quiet();      // I'm done
-  shmem_barrier_all();// He's done too
+  if( CommunicatorPolicy == CommunicatorPolicyConcurrent ) shmem_barrier_all();// He's done too
 }
 void CartesianCommunicator::Barrier(void)
 {
@@ -301,13 +310,13 @@ void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
   int words = bytes/4;
 
   if ( shmem_addr_accessible(data,_processor)  ){
-    shmem_broadcast32(data,data,words,root,0,0,shmem_n_pes(),psync);
+    shmem_broadcast32(data,data,words,root,0,0,shmem_n_pes(),psync.data());
     return;
   }
 
   for(int w=0;w<words;w++){
     word = array[w];
-    shmem_broadcast32((void *)&word,(void *)&word,1,root,0,0,shmem_n_pes(),psync);
+    shmem_broadcast32((void *)&word,(void *)&word,1,root,0,0,shmem_n_pes(),psync.data());
     if ( shmem_my_pe() != root ) {
       array[w] = word;
     }
@@ -325,7 +334,7 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
 
   for(int w=0;w<words;w++){
     word = array[w];
-    shmem_broadcast32((void *)&word,(void *)&word,1,root,0,0,shmem_n_pes(),psync);
+    shmem_broadcast32((void *)&word,(void *)&word,1,root,0,0,shmem_n_pes(),psync.data());
     if ( shmem_my_pe() != root ) {
       array[w]= word;
     }
@@ -333,5 +342,9 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
   }
 }
   
+int CartesianCommunicator::RankWorld(void){ 
+  return shmem_my_pe();
+}
+
 }
 

lib/cshift/Cshift_common.h

@@ -1,5 +1,4 @@
-
-    /*************************************************************************************
+/*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
 
@@ -31,21 +30,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 namespace Grid {
 
-template<class vobj>
-class SimpleCompressor {
-public:
-  void Point(int) {};
-
-  vobj operator() (const vobj &arg) {
-    return arg;
-  }
-};
-
 ///////////////////////////////////////////////////////////////////
-// Gather for when there is no need to SIMD split with compression
+// Gather for when there is no need to SIMD split 
 ///////////////////////////////////////////////////////////////////
-template<class vobj,class cobj,class compressor> void 
-Gather_plane_simple (const Lattice<vobj> &rhs,commVector<cobj> &buffer,int dimension,int plane,int cbmask,compressor &compress, int off=0)
+template<class vobj> void 
+Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
 {
   int rd = rhs._grid->_rdimensions[dimension];
 
@@ -53,19 +42,17 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<cobj> &buffer,int dimen
     cbmask = 0x3;
   }
   
-  int so  = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
-  
+  int so=plane*rhs._grid->_ostride[dimension]; // base offset for start of plane 
   int e1=rhs._grid->_slice_nblock[dimension];
   int e2=rhs._grid->_slice_block[dimension];
 
   int stride=rhs._grid->_slice_stride[dimension];
   if ( cbmask == 0x3 ) { 
-PARALLEL_NESTED_LOOP2
-    for(int n=0;n<e1;n++){
+    parallel_for_nest2(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o  = n*stride;
 	int bo = n*e2;
-	buffer[off+bo+b]=compress(rhs._odata[so+o+b]);
+	buffer[off+bo+b]=rhs._odata[so+o+b];
       }
     }
   } else { 
@@ -74,25 +61,23 @@ PARALLEL_NESTED_LOOP2
      for(int n=0;n<e1;n++){
        for(int b=0;b<e2;b++){
 	 int o  = n*stride;
-	 int ocb=1<<rhs._grid->CheckerBoardFromOindexTable(o+b);
+	 int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
 	 if ( ocb &cbmask ) {
 	   table.push_back(std::pair<int,int> (bo++,o+b));
 	 }
        }
      }
-PARALLEL_FOR_LOOP     
-     for(int i=0;i<table.size();i++){
-       buffer[off+table[i].first]=compress(rhs._odata[so+table[i].second]);
+     parallel_for(int i=0;i<table.size();i++){
+       buffer[off+table[i].first]=rhs._odata[so+table[i].second];
      }
   }
 }
 
-
 ///////////////////////////////////////////////////////////////////
-// Gather for when there *is* need to SIMD split with compression
+// Gather for when there *is* need to SIMD split 
 ///////////////////////////////////////////////////////////////////
-template<class cobj,class vobj,class compressor> void 
-Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename cobj::scalar_object *> pointers,int dimension,int plane,int cbmask,compressor &compress)
+template<class vobj> void 
+Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
 {
   int rd = rhs._grid->_rdimensions[dimension];
 
@@ -105,57 +90,40 @@ Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename cobj::scalar_
   int e1=rhs._grid->_slice_nblock[dimension];
   int e2=rhs._grid->_slice_block[dimension];
   int n1=rhs._grid->_slice_stride[dimension];
-  int n2=rhs._grid->_slice_block[dimension];
+
   if ( cbmask ==0x3){
-PARALLEL_NESTED_LOOP2
-    for(int n=0;n<e1;n++){
+    parallel_for_nest2(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 
 	int o      =   n*n1;
-	int offset = b+n*n2;
-	cobj temp =compress(rhs._odata[so+o+b]);
+	int offset = b+n*e2;
 	
-	extract<cobj>(temp,pointers,offset);
+	vobj temp =rhs._odata[so+o+b];
+	extract<vobj>(temp,pointers,offset);
 
       }
     }
   } else { 
 
-    assert(0); //Fixme think this is buggy
-
-    for(int n=0;n<e1;n++){
+    // Case of SIMD split AND checker dim cannot currently be hit, except in 
+    // Test_cshift_red_black code.
+    std::cout << " Dense packed buffer WARNING " <<std::endl;
+    parallel_for_nest2(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
-	int o=n*rhs._grid->_slice_stride[dimension];
+
+	int o=n*n1;
 	int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
-	int offset = b+n*rhs._grid->_slice_block[dimension];
+	int offset = b+n*e2;
 
 	if ( ocb & cbmask ) {
-	  cobj temp =compress(rhs._odata[so+o+b]);
-	  extract<cobj>(temp,pointers,offset);
+	  vobj temp =rhs._odata[so+o+b];
+	  extract<vobj>(temp,pointers,offset);
 	}
       }
     }
   }
 }
 
-//////////////////////////////////////////////////////
-// Gather for when there is no need to SIMD split
-//////////////////////////////////////////////////////
-template<class vobj> void Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
-{
-  SimpleCompressor<vobj> dontcompress;
-  Gather_plane_simple (rhs,buffer,dimension,plane,cbmask,dontcompress);
-}
-
-//////////////////////////////////////////////////////
-// Gather for when there *is* need to SIMD split
-//////////////////////////////////////////////////////
-template<class vobj> void Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
-{
-  SimpleCompressor<vobj> dontcompress;
-  Gather_plane_extract<vobj,vobj,decltype(dontcompress)>(rhs,pointers,dimension,plane,cbmask,dontcompress);
-}
-
 //////////////////////////////////////////////////////
 // Scatter for when there is no need to SIMD split
 //////////////////////////////////////////////////////
@@ -171,10 +139,10 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
     
   int e1=rhs._grid->_slice_nblock[dimension];
   int e2=rhs._grid->_slice_block[dimension];
+  int stride=rhs._grid->_slice_stride[dimension];
   
   if ( cbmask ==0x3 ) {
-PARALLEL_NESTED_LOOP2
-    for(int n=0;n<e1;n++){
+    parallel_for_nest2(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o   =n*rhs._grid->_slice_stride[dimension];
 	int bo  =n*rhs._grid->_slice_block[dimension];
@@ -182,24 +150,28 @@ PARALLEL_NESTED_LOOP2
       }
     }
   } else { 
+    std::vector<std::pair<int,int> > table;
     int bo=0;
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o   =n*rhs._grid->_slice_stride[dimension];
-	int bo  =n*rhs._grid->_slice_block[dimension];
 	int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
 	if ( ocb & cbmask ) {
-	  rhs._odata[so+o+b]=buffer[bo++];
+	  table.push_back(std::pair<int,int> (so+o+b,bo++));
 	}
       }
     }
+    parallel_for(int i=0;i<table.size();i++){
+       //       std::cout << "Rcv"<< table[i].first << " " << table[i].second << " " <<buffer[table[i].second]<<std::endl;
+       rhs._odata[table[i].first]=buffer[table[i].second];
+     }
   }
 }
 
 //////////////////////////////////////////////////////
 // Scatter for when there *is* need to SIMD split
 //////////////////////////////////////////////////////
- template<class vobj,class cobj> void Scatter_plane_merge(Lattice<vobj> &rhs,std::vector<cobj *> pointers,int dimension,int plane,int cbmask)
+template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
 {
   int rd = rhs._grid->_rdimensions[dimension];
 
@@ -213,8 +185,7 @@ PARALLEL_NESTED_LOOP2
   int e2=rhs._grid->_slice_block[dimension];
 
   if(cbmask ==0x3 ) {
-PARALLEL_NESTED_LOOP2
-    for(int n=0;n<e1;n++){
+    parallel_for_nest2(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o      = n*rhs._grid->_slice_stride[dimension];
 	int offset = b+n*rhs._grid->_slice_block[dimension];
@@ -222,7 +193,11 @@ PARALLEL_NESTED_LOOP2
       }
     }
   } else { 
-    assert(0); // think this is buggy FIXME
+
+    // Case of SIMD split AND checker dim cannot currently be hit, except in 
+    // 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<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
 	int o      = n*rhs._grid->_slice_stride[dimension];
@@ -254,8 +229,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   int e2=rhs._grid->_slice_block[dimension];
   int stride = rhs._grid->_slice_stride[dimension];
   if(cbmask == 0x3 ){
-PARALLEL_NESTED_LOOP2
-    for(int n=0;n<e1;n++){
+    parallel_for_nest2(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
  
         int o =n*stride+b;
@@ -264,8 +238,7 @@ PARALLEL_NESTED_LOOP2
       }
     }
   } else { 
-PARALLEL_NESTED_LOOP2
-    for(int n=0;n<e1;n++){
+    parallel_for_nest2(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
  
         int o =n*stride+b;
@@ -295,8 +268,8 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   int e1=rhs._grid->_slice_nblock[dimension];
   int e2=rhs._grid->_slice_block [dimension];
   int stride = rhs._grid->_slice_stride[dimension];
-PARALLEL_NESTED_LOOP2
-  for(int n=0;n<e1;n++){
+
+  parallel_for_nest2(int n=0;n<e1;n++){
   for(int b=0;b<e2;b++){
 
       int o  =n*stride;
@@ -338,8 +311,8 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
   // Map to always positive shift modulo global full dimension.
   shift = (shift+fd)%fd;
 
-  ret.checkerboard = grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
   // the permute type
+  ret.checkerboard = grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
   int permute_dim =grid->PermuteDim(dimension);
   int permute_type=grid->PermuteType(dimension);
   int permute_type_dist;
@@ -348,7 +321,6 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
 
     int o   = 0;
     int bo  = x * grid->_ostride[dimension];
-    
     int cb= (cbmask==0x2)? Odd : Even;
 
     int sshift = grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
@@ -361,9 +333,23 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
     // wrap is whether sshift > rd.
     //  num is sshift mod rd.
     // 
+    //  shift 7
+    //
+    //  XoXo YcYc 
+    //  oXoX cYcY
+    //  XoXo YcYc
+    //  oXoX cYcY
+    //
+    //  sshift -- 
+    //
+    //  XX YY ; 3
+    //  XX YY ; 0
+    //  XX YY ; 3
+    //  XX YY ; 0
+    //
     int permute_slice=0;
     if(permute_dim){
-      int wrap = sshift/rd;
+      int wrap = sshift/rd; wrap=wrap % ly;
       int  num = sshift%rd;
 
       if ( x< rd-num ) permute_slice=wrap;
@@ -375,7 +361,6 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
       } else {
 	permute_type_dist = permute_type;
       }
-      
     }
 
     if ( permute_slice ) Copy_plane_permute(ret,rhs,dimension,x,sx,cbmask,permute_type_dist);

lib/cshift/Cshift_mpi.h

@@ -74,7 +74,6 @@ template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &r
   sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
 
   //  std::cout << "Cshift_comms dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
-
   if ( sshift[0] == sshift[1] ) {
     //    std::cout << "Single pass Cshift_comms" <<std::endl;
     Cshift_comms(ret,rhs,dimension,shift,0x3);
@@ -154,10 +153,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
 			   (void *)&recv_buf[0],
 			   recv_from_rank,
 			   bytes);
+      grid->Barrier();
 
-      //      for(int i=0;i<words;i++){
-      //	std::cout << "SendRecv ["<<i<<"] snd "<<send_buf[i]<<" rcv " << recv_buf[i] << "  0x" << cbmask<<std::endl;
-      //      }
       Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
     }
   }
@@ -243,7 +240,7 @@ template<class vobj> void  Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
 			     (void *)&recv_buf_extract[i][0],
 			     recv_from_rank,
 			     bytes);
-
+	grid->Barrier();
 	rpointers[i] = &recv_buf_extract[i][0];
       } else { 
 	rpointers[i] = &send_buf_extract[nbr_lane][0];

lib/lattice/Lattice_arith.h

@@ -39,8 +39,7 @@ namespace Grid {
     ret.checkerboard = lhs.checkerboard;
     conformable(ret,rhs);
     conformable(lhs,rhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<lhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       mult(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
@@ -56,8 +55,7 @@ PARALLEL_FOR_LOOP
     ret.checkerboard = lhs.checkerboard;
     conformable(ret,rhs);
     conformable(lhs,rhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<lhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       mac(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
@@ -73,8 +71,7 @@ PARALLEL_FOR_LOOP
     ret.checkerboard = lhs.checkerboard;
     conformable(ret,rhs);
     conformable(lhs,rhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<lhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       sub(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
@@ -89,8 +86,7 @@ PARALLEL_FOR_LOOP
     ret.checkerboard = lhs.checkerboard;
     conformable(ret,rhs);
     conformable(lhs,rhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<lhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       add(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
@@ -108,8 +104,7 @@ PARALLEL_FOR_LOOP
     void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
     ret.checkerboard = lhs.checkerboard;
     conformable(lhs,ret);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<lhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
       obj1 tmp;
       mult(&tmp,&lhs._odata[ss],&rhs);
       vstream(ret._odata[ss],tmp);
@@ -120,8 +115,7 @@ PARALLEL_FOR_LOOP
     void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
     ret.checkerboard = lhs.checkerboard;
     conformable(ret,lhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<lhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
       obj1 tmp;
       mac(&tmp,&lhs._odata[ss],&rhs);
       vstream(ret._odata[ss],tmp);
@@ -132,8 +126,7 @@ PARALLEL_FOR_LOOP
     void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
     ret.checkerboard = lhs.checkerboard;
     conformable(ret,lhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<lhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       sub(&tmp,&lhs._odata[ss],&rhs);
@@ -147,8 +140,7 @@ PARALLEL_FOR_LOOP
     void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
     ret.checkerboard = lhs.checkerboard;
     conformable(lhs,ret);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<lhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       add(&tmp,&lhs._odata[ss],&rhs);
@@ -166,8 +158,7 @@ PARALLEL_FOR_LOOP
     void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
     ret.checkerboard = rhs.checkerboard;
     conformable(ret,rhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<rhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       mult(&tmp,&lhs,&rhs._odata[ss]);
@@ -182,8 +173,7 @@ PARALLEL_FOR_LOOP
     void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
     ret.checkerboard = rhs.checkerboard;
     conformable(ret,rhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<rhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       mac(&tmp,&lhs,&rhs._odata[ss]);
@@ -198,8 +188,7 @@ PARALLEL_FOR_LOOP
     void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
     ret.checkerboard = rhs.checkerboard;
     conformable(ret,rhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<rhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       sub(&tmp,&lhs,&rhs._odata[ss]);
@@ -213,8 +202,7 @@ PARALLEL_FOR_LOOP
     void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
     ret.checkerboard = rhs.checkerboard;
     conformable(ret,rhs);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<rhs._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       obj1 tmp;
       add(&tmp,&lhs,&rhs._odata[ss]);
@@ -230,8 +218,7 @@ PARALLEL_FOR_LOOP
     ret.checkerboard = x.checkerboard;
     conformable(ret,x);
     conformable(x,y);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<x._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<x._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       vobj tmp = a*x._odata[ss]+y._odata[ss];
       vstream(ret._odata[ss],tmp);
@@ -245,8 +232,7 @@ PARALLEL_FOR_LOOP
     ret.checkerboard = x.checkerboard;
     conformable(ret,x);
     conformable(x,y);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<x._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<x._grid->oSites();ss++){
 #ifdef STREAMING_STORES
       vobj tmp = a*x._odata[ss]+b*y._odata[ss];
       vstream(ret._odata[ss],tmp);

lib/lattice/Lattice_base.h

@@ -121,8 +121,7 @@ public:
     assert( (cb==Odd) || (cb==Even));
     checkerboard=cb;
 
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<_grid->oSites();ss++){
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
 #ifdef STREAMING_STORES
       vobj tmp = eval(ss,expr);
       vstream(_odata[ss] ,tmp);
@@ -144,8 +143,7 @@ PARALLEL_FOR_LOOP
     assert( (cb==Odd) || (cb==Even));
     checkerboard=cb;
 
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<_grid->oSites();ss++){
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
 #ifdef STREAMING_STORES
       vobj tmp = eval(ss,expr);
       vstream(_odata[ss] ,tmp);
@@ -167,8 +165,7 @@ PARALLEL_FOR_LOOP
     assert( (cb==Odd) || (cb==Even));
     checkerboard=cb;
 
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<_grid->oSites();ss++){
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
 #ifdef STREAMING_STORES
       //vobj tmp = eval(ss,expr);
       vstream(_odata[ss] ,eval(ss,expr));
@@ -191,8 +188,7 @@ PARALLEL_FOR_LOOP
     checkerboard=cb;
 
     _odata.resize(_grid->oSites());
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<_grid->oSites();ss++){
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
 #ifdef STREAMING_STORES
       vobj tmp = eval(ss,expr);
       vstream(_odata[ss] ,tmp);
@@ -213,8 +209,7 @@ PARALLEL_FOR_LOOP
     checkerboard=cb;
 
     _odata.resize(_grid->oSites());
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<_grid->oSites();ss++){
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
 #ifdef STREAMING_STORES
       vobj tmp = eval(ss,expr);
       vstream(_odata[ss] ,tmp);
@@ -235,72 +230,78 @@ PARALLEL_FOR_LOOP
     checkerboard=cb;
 
     _odata.resize(_grid->oSites());
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<_grid->oSites();ss++){
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
       vstream(_odata[ss] ,eval(ss,expr));
     }
   };
 
-    //////////////////////////////////////////////////////////////////
-    // Constructor requires "grid" passed.
-    // what about a default grid?
-    //////////////////////////////////////////////////////////////////
-    Lattice(GridBase *grid) : _odata(grid->oSites()) {
-        _grid = grid;
+  //////////////////////////////////////////////////////////////////
+  // Constructor requires "grid" passed.
+  // what about a default grid?
+  //////////////////////////////////////////////////////////////////
+  Lattice(GridBase *grid) : _odata(grid->oSites()) {
+    _grid = grid;
     //        _odata.reserve(_grid->oSites());
     //        _odata.resize(_grid->oSites());
     //      std::cout << "Constructing lattice object with Grid pointer "<<_grid<<std::endl;
-        assert((((uint64_t)&_odata[0])&0xF) ==0);
-        checkerboard=0;
+    assert((((uint64_t)&_odata[0])&0xF) ==0);
+    checkerboard=0;
+  }
+  
+  Lattice(const Lattice& r){ // copy constructor
+    _grid = r._grid;
+    checkerboard = r.checkerboard;
+    _odata.resize(_grid->oSites());// essential
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+      _odata[ss]=r._odata[ss];
     }  	
+  }
   
-    Lattice(const Lattice& r){ // copy constructor
-    	_grid = r._grid;
-    	checkerboard = r.checkerboard;
-    	_odata.resize(_grid->oSites());// essential
-  		PARALLEL_FOR_LOOP
-        for(int ss=0;ss<_grid->oSites();ss++){
-            _odata[ss]=r._odata[ss];
-        }  	
+  
+  
+  virtual ~Lattice(void) = default;
+    
+  void reset(GridBase* grid) {
+    if (_grid != grid) {
+      _grid = grid;
+      _odata.resize(grid->oSites());
+      checkerboard = 0;
     }
+  }
   
 
-
-    virtual ~Lattice(void) = default;
-    
-    template<class sobj> strong_inline Lattice<vobj> & operator = (const sobj & r){
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<_grid->oSites();ss++){
-            this->_odata[ss]=r;
-        }
-        return *this;
+  template<class sobj> strong_inline Lattice<vobj> & operator = (const sobj & r){
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+      this->_odata[ss]=r;
     }
-    template<class robj> strong_inline Lattice<vobj> & operator = (const Lattice<robj> & r){
-      this->checkerboard = r.checkerboard;
-      conformable(*this,r);
+    return *this;
+  }
   
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<_grid->oSites();ss++){
-            this->_odata[ss]=r._odata[ss];
-        }
-        return *this;
-    }
+  template<class robj> strong_inline Lattice<vobj> & operator = (const Lattice<robj> & r){
+    this->checkerboard = r.checkerboard;
+    conformable(*this,r);
     
-    // *=,+=,-= operators inherit behvour from correspond */+/- operation
-    template<class T> strong_inline Lattice<vobj> &operator *=(const T &r) {
-        *this = (*this)*r;
-        return *this;
+    parallel_for(int ss=0;ss<_grid->oSites();ss++){
+      this->_odata[ss]=r._odata[ss];
     }
+    return *this;
+  }
   
-    template<class T> strong_inline Lattice<vobj> &operator -=(const T &r) {
-        *this = (*this)-r;
-        return *this;
-    }
-    template<class T> strong_inline Lattice<vobj> &operator +=(const T &r) {
-        *this = (*this)+r;
-        return *this;
-    }
- }; // class Lattice
+  // *=,+=,-= operators inherit behvour from correspond */+/- operation
+  template<class T> strong_inline Lattice<vobj> &operator *=(const T &r) {
+    *this = (*this)*r;
+    return *this;
+  }
+  
+  template<class T> strong_inline Lattice<vobj> &operator -=(const T &r) {
+    *this = (*this)-r;
+    return *this;
+  }
+  template<class T> strong_inline Lattice<vobj> &operator +=(const T &r) {
+    *this = (*this)+r;
+    return *this;
+  }
+}; // class Lattice
   
   template<class vobj> std::ostream& operator<< (std::ostream& stream, const Lattice<vobj> &o){
     std::vector<int> gcoor;

lib/lattice/Lattice_comparison.h

@@ -45,90 +45,87 @@ namespace Grid {
   //////////////////////////////////////////////////////////////////////////
   template<class vfunctor,class lobj,class robj>  
     inline Lattice<vInteger> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
-    {
-      Lattice<vInteger> ret(rhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	  ret._odata[ss]=op(lhs._odata[ss],rhs._odata[ss]);
-        }
-        return ret;
+  {
+    Lattice<vInteger> ret(rhs._grid);
+    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+      ret._odata[ss]=op(lhs._odata[ss],rhs._odata[ss]);
     }
+    return ret;
+  }
   //////////////////////////////////////////////////////////////////////////
   // compare lattice to scalar
   //////////////////////////////////////////////////////////////////////////
-    template<class vfunctor,class lobj,class robj> 
+  template<class vfunctor,class lobj,class robj> 
     inline Lattice<vInteger> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
-    {
-      Lattice<vInteger> ret(lhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<lhs._grid->oSites(); ss++){
-	  ret._odata[ss]=op(lhs._odata[ss],rhs);
-        }
-        return ret;
+  {
+    Lattice<vInteger> ret(lhs._grid);
+    parallel_for(int ss=0;ss<lhs._grid->oSites(); ss++){
+      ret._odata[ss]=op(lhs._odata[ss],rhs);
     }
+    return ret;
+  }
   //////////////////////////////////////////////////////////////////////////
   // compare scalar to lattice
   //////////////////////////////////////////////////////////////////////////
-    template<class vfunctor,class lobj,class robj> 
+  template<class vfunctor,class lobj,class robj> 
     inline Lattice<vInteger> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
-    {
-      Lattice<vInteger> ret(rhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	  ret._odata[ss]=op(lhs._odata[ss],rhs);
-        }
-        return ret;
+  {
+    Lattice<vInteger> ret(rhs._grid);
+    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+      ret._odata[ss]=op(lhs._odata[ss],rhs);
     }
+    return ret;
+  }
   
   //////////////////////////////////////////////////////////////////////////
   // Map to functors
   //////////////////////////////////////////////////////////////////////////
-    // Less than
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-     return LLComparison(vlt<lobj,robj>(),lhs,rhs);
-   }
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
-     return LSComparison(vlt<lobj,robj>(),lhs,rhs);
-   }
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
-     return SLComparison(vlt<lobj,robj>(),lhs,rhs);
-   }
+  // Less than
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+    return LLComparison(vlt<lobj,robj>(),lhs,rhs);
+  }
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
+    return LSComparison(vlt<lobj,robj>(),lhs,rhs);
+  }
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
+    return SLComparison(vlt<lobj,robj>(),lhs,rhs);
+  }
   
-   // Less than equal
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-     return LLComparison(vle<lobj,robj>(),lhs,rhs);
-   }
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
-     return LSComparison(vle<lobj,robj>(),lhs,rhs);
-   }
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
-     return SLComparison(vle<lobj,robj>(),lhs,rhs);
-   }
+  // Less than equal
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+    return LLComparison(vle<lobj,robj>(),lhs,rhs);
+  }
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
+    return LSComparison(vle<lobj,robj>(),lhs,rhs);
+  }
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
+    return SLComparison(vle<lobj,robj>(),lhs,rhs);
+  }
   
-   // Greater than 
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
-     return LLComparison(vgt<lobj,robj>(),lhs,rhs);
-   }
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
-     return LSComparison(vgt<lobj,robj>(),lhs,rhs);
-   }
-   template<class lobj,class robj>
-   inline Lattice<vInteger> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
+  // Greater than 
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+    return LLComparison(vgt<lobj,robj>(),lhs,rhs);
+  }
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
+    return LSComparison(vgt<lobj,robj>(),lhs,rhs);
+  }
+  template<class lobj,class robj>
+    inline Lattice<vInteger> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
      return SLComparison(vgt<lobj,robj>(),lhs,rhs);
-   }
+  }
   
   
-   // Greater than equal
+  // Greater than equal
    template<class lobj,class robj>
-   inline Lattice<vInteger> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+     inline Lattice<vInteger> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
      return LLComparison(vge<lobj,robj>(),lhs,rhs);
    }
    template<class lobj,class robj>
@@ -136,38 +133,37 @@ PARALLEL_FOR_LOOP
      return LSComparison(vge<lobj,robj>(),lhs,rhs);
    }
    template<class lobj,class robj>
-   inline Lattice<vInteger> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
+     inline Lattice<vInteger> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
      return SLComparison(vge<lobj,robj>(),lhs,rhs);
    }
    
    // equal
    template<class lobj,class robj>
-   inline Lattice<vInteger> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+     inline Lattice<vInteger> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
      return LLComparison(veq<lobj,robj>(),lhs,rhs);
    }
    template<class lobj,class robj>
-   inline Lattice<vInteger> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
+     inline Lattice<vInteger> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
      return LSComparison(veq<lobj,robj>(),lhs,rhs);
    }
    template<class lobj,class robj>
-   inline Lattice<vInteger> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
+     inline Lattice<vInteger> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
      return SLComparison(veq<lobj,robj>(),lhs,rhs);
    }
    
    
    // not equal
    template<class lobj,class robj>
-   inline Lattice<vInteger> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
+     inline Lattice<vInteger> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
      return LLComparison(vne<lobj,robj>(),lhs,rhs);
    }
    template<class lobj,class robj>
-   inline Lattice<vInteger> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
+     inline Lattice<vInteger> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
      return LSComparison(vne<lobj,robj>(),lhs,rhs);
    }
    template<class lobj,class robj>
-   inline Lattice<vInteger> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
+     inline Lattice<vInteger> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
      return SLComparison(vne<lobj,robj>(),lhs,rhs);
    }
-
 }
 #endif

lib/lattice/Lattice_local.h

@@ -34,47 +34,42 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 namespace Grid {
 
-    /////////////////////////////////////////////////////
-    // Non site, reduced locally reduced routines
-    /////////////////////////////////////////////////////
+  /////////////////////////////////////////////////////
+  // Non site, reduced locally reduced routines
+  /////////////////////////////////////////////////////
   
-    // localNorm2,
-    template<class vobj>
+  // localNorm2,
+  template<class vobj>
     inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
     {
       Lattice<typename vobj::tensor_reduced> ret(rhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<rhs._grid->oSites(); ss++){
-	  ret._odata[ss]=innerProduct(rhs._odata[ss],rhs._odata[ss]);
-        }
-        return ret;
+      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+	ret._odata[ss]=innerProduct(rhs._odata[ss],rhs._odata[ss]);
+      }
+      return ret;
     }
   
-    // localInnerProduct
-    template<class vobj>
+  // localInnerProduct
+  template<class vobj>
     inline auto localInnerProduct (const Lattice<vobj> &lhs,const Lattice<vobj> &rhs) -> Lattice<typename vobj::tensor_reduced>
     {
       Lattice<typename vobj::tensor_reduced> ret(rhs._grid);
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<rhs._grid->oSites(); ss++){
+      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
 	ret._odata[ss]=innerProduct(lhs._odata[ss],rhs._odata[ss]);
       }
       return ret;
     }
   
-    // outerProduct Scalar x Scalar -> Scalar
-    //              Vector x Vector -> Matrix
-    template<class ll,class rr>
+  // outerProduct Scalar x Scalar -> Scalar
+  //              Vector x Vector -> Matrix
+  template<class ll,class rr>
     inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Lattice<decltype(outerProduct(lhs._odata[0],rhs._odata[0]))>
-    {
-        Lattice<decltype(outerProduct(lhs._odata[0],rhs._odata[0]))> ret(rhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<rhs._grid->oSites(); ss++){
-            ret._odata[ss]=outerProduct(lhs._odata[ss],rhs._odata[ss]);
-        }
-        return ret;
-     }
-
+  {
+    Lattice<decltype(outerProduct(lhs._odata[0],rhs._odata[0]))> ret(rhs._grid);
+    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
+      ret._odata[ss]=outerProduct(lhs._odata[ss],rhs._odata[ss]);
+    }
+    return ret;
+  }
 }
-
 #endif

lib/lattice/Lattice_overload.h

@@ -37,8 +37,7 @@ namespace Grid {
   inline Lattice<vobj> operator -(const Lattice<vobj> &r)
   {
     Lattice<vobj> ret(r._grid);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<r._grid->oSites();ss++){
+    parallel_for(int ss=0;ss<r._grid->oSites();ss++){
       vstream(ret._odata[ss], -r._odata[ss]);
     }
     return ret;
@@ -74,8 +73,7 @@ PARALLEL_FOR_LOOP
   inline auto operator * (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs*rhs._odata[0])>
   {
     Lattice<decltype(lhs*rhs._odata[0])> ret(rhs._grid);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<rhs._grid->oSites(); ss++){
+    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
       decltype(lhs*rhs._odata[0]) tmp=lhs*rhs._odata[ss]; 
       vstream(ret._odata[ss],tmp);
 	   //      ret._odata[ss]=lhs*rhs._odata[ss];
@@ -86,8 +84,7 @@ PARALLEL_FOR_LOOP
     inline auto operator + (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs+rhs._odata[0])>
     {
       Lattice<decltype(lhs+rhs._odata[0])> ret(rhs._grid);
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<rhs._grid->oSites(); ss++){
+      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
 	decltype(lhs+rhs._odata[0]) tmp =lhs-rhs._odata[ss];  
 	vstream(ret._odata[ss],tmp);
 	//	ret._odata[ss]=lhs+rhs._odata[ss];
@@ -98,11 +95,9 @@ PARALLEL_FOR_LOOP
     inline auto operator - (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs-rhs._odata[0])>
   {
     Lattice<decltype(lhs-rhs._odata[0])> ret(rhs._grid);
-PARALLEL_FOR_LOOP
-    for(int ss=0;ss<rhs._grid->oSites(); ss++){
+    parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
       decltype(lhs-rhs._odata[0]) tmp=lhs-rhs._odata[ss];  
       vstream(ret._odata[ss],tmp);
-      //      ret._odata[ss]=lhs-rhs._odata[ss];
     }
     return ret;
   }
@@ -110,8 +105,7 @@ PARALLEL_FOR_LOOP
       inline auto operator * (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]*rhs)>
     {
       Lattice<decltype(lhs._odata[0]*rhs)> ret(lhs._grid);
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<lhs._grid->oSites(); ss++){
+      parallel_for(int ss=0;ss<lhs._grid->oSites(); ss++){
 	decltype(lhs._odata[0]*rhs) tmp =lhs._odata[ss]*rhs;
 	vstream(ret._odata[ss],tmp);
 	//            ret._odata[ss]=lhs._odata[ss]*rhs;
@@ -122,8 +116,7 @@ PARALLEL_FOR_LOOP
       inline auto operator + (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]+rhs)>
     {
         Lattice<decltype(lhs._odata[0]+rhs)> ret(lhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<rhs._grid->oSites(); ss++){
+	parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
 	  decltype(lhs._odata[0]+rhs) tmp=lhs._odata[ss]+rhs; 
 	  vstream(ret._odata[ss],tmp);
 	  //	  ret._odata[ss]=lhs._odata[ss]+rhs;
@@ -134,15 +127,12 @@ PARALLEL_FOR_LOOP
       inline auto operator - (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]-rhs)>
     {
       Lattice<decltype(lhs._odata[0]-rhs)> ret(lhs._grid);
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<rhs._grid->oSites(); ss++){
+      parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
 	  decltype(lhs._odata[0]-rhs) tmp=lhs._odata[ss]-rhs;
 	  vstream(ret._odata[ss],tmp);
 	  //	ret._odata[ss]=lhs._odata[ss]-rhs;
       }
       return ret;
     }
-
-
 }
 #endif

lib/lattice/Lattice_peekpoke.h

@@ -44,22 +44,20 @@ namespace Grid {
     {
       Lattice<decltype(peekIndex<Index>(lhs._odata[0],i))> ret(lhs._grid);
       ret.checkerboard=lhs.checkerboard;
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<lhs._grid->oSites();ss++){
-	  ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i);
-        }
-        return ret;
+      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+	ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i);
+      }
+      return ret;
     };
     template<int Index,class vobj>
-       auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekIndex<Index>(lhs._odata[0],i,j))>
+      auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekIndex<Index>(lhs._odata[0],i,j))>
     {
       Lattice<decltype(peekIndex<Index>(lhs._odata[0],i,j))> ret(lhs._grid);
       ret.checkerboard=lhs.checkerboard;
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<lhs._grid->oSites();ss++){
-	  ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i,j);
-        }
-        return ret;
+      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+	ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i,j);
+      }
+      return ret;
     };
 
     ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -68,18 +66,16 @@ PARALLEL_FOR_LOOP
     template<int Index,class vobj> 
     void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(lhs._odata[0],0))> & rhs,int i)
     {
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<lhs._grid->oSites();ss++){
-	  pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i);
-	}      
+      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+	pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i);
+      }      
     }
     template<int Index,class vobj>
       void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(lhs._odata[0],0,0))> & rhs,int i,int j)
     {
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<lhs._grid->oSites();ss++){
-	  pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i,j);
-	}      
+      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+	pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i,j);
+      }      
     }
 
     //////////////////////////////////////////////////////
@@ -131,9 +127,6 @@ PARALLEL_FOR_LOOP
 
       assert( l.checkerboard == l._grid->CheckerBoard(site));
 
-      // FIXME
-      //      assert( sizeof(sobj)*Nsimd == sizeof(vobj));
-
       int rank,odx,idx;
       grid->GlobalCoorToRankIndex(rank,odx,idx,site);
 

lib/lattice/Lattice_reality.h

@@ -40,8 +40,7 @@ namespace Grid {
 
     template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
         Lattice<vobj> ret(lhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<lhs._grid->oSites();ss++){
+	parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
             ret._odata[ss] = adj(lhs._odata[ss]);
         }
         return ret;
@@ -49,13 +48,10 @@ PARALLEL_FOR_LOOP
 
     template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
         Lattice<vobj> ret(lhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<lhs._grid->oSites();ss++){
-            ret._odata[ss] = conjugate(lhs._odata[ss]);
+	parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
+	  ret._odata[ss] = conjugate(lhs._odata[ss]);
         }
         return ret;
     };
-
-
 }
 #endif

lib/lattice/Lattice_reduction.h

@@ -1,159 +1,154 @@
-    /*************************************************************************************
-
+ /*************************************************************************************
     Grid physics library, www.github.com/paboyle/Grid 
-
     Source file: ./lib/lattice/Lattice_reduction.h
-
     Copyright (C) 2015
-
 Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: paboyle <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 */
 #ifndef GRID_LATTICE_REDUCTION_H
 #define GRID_LATTICE_REDUCTION_H
 
+#include <Grid/Grid_Eigen_Dense.h>
+
 namespace Grid {
 #ifdef GRID_WARN_SUBOPTIMAL
 #warning "Optimisation alert all these reduction loops are NOT threaded "
 #endif     
 
-    ////////////////////////////////////////////////////////////////////////////////////////////////////
-    // Deterministic Reduction operations
-    ////////////////////////////////////////////////////////////////////////////////////////////////////
-  template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
-    ComplexD nrm = innerProduct(arg,arg);
-    return std::real(nrm); 
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Deterministic Reduction operations
+  ////////////////////////////////////////////////////////////////////////////////////////////////////
+template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
+  ComplexD nrm = innerProduct(arg,arg);
+  return std::real(nrm); 
+}
+
+// Double inner product
+template<class vobj>
+inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) 
+{
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_typeD vector_type;
+  scalar_type  nrm;
+  
+  GridBase *grid = left._grid;
+  
+  std::vector<vector_type,alignedAllocator<vector_type> > sumarray(grid->SumArraySize());
+  
+  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
+    int nwork, mywork, myoff;
+    GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
+    
+    decltype(innerProductD(left._odata[0],right._odata[0])) vnrm=zero; // private to thread; sub summation
+    for(int ss=myoff;ss<mywork+myoff; ss++){
+      vnrm = vnrm + innerProductD(left._odata[ss],right._odata[ss]);
+    }
+    sumarray[thr]=TensorRemove(vnrm) ;
   }
   
-    template<class vobj>
-    inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) 
-    {
-      typedef typename vobj::scalar_type scalar_type;
-      typedef typename vobj::vector_type vector_type;
-      scalar_type  nrm;
+  vector_type vvnrm; vvnrm=zero;  // sum across threads
+  for(int i=0;i<grid->SumArraySize();i++){
+    vvnrm = vvnrm+sumarray[i];
+  } 
+  nrm = Reduce(vvnrm);// sum across simd
+  right._grid->GlobalSum(nrm);
+  return nrm;
+}
  
-      GridBase *grid = left._grid;
+template<class Op,class T1>
+inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
+  ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second))))::scalar_object
+{
+  return sum(closure(expr));
+}
 
-      std::vector<vector_type,alignedAllocator<vector_type> > sumarray(grid->SumArraySize());
-      for(int i=0;i<grid->SumArraySize();i++){
-	sumarray[i]=zero;
-      }
-
-PARALLEL_FOR_LOOP
-      for(int thr=0;thr<grid->SumArraySize();thr++){
-	int nwork, mywork, myoff;
-	GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
-	
-	decltype(innerProduct(left._odata[0],right._odata[0])) vnrm=zero; // private to thread; sub summation
-        for(int ss=myoff;ss<mywork+myoff; ss++){
-	  vnrm = vnrm + innerProduct(left._odata[ss],right._odata[ss]);
-	}
-	sumarray[thr]=TensorRemove(vnrm) ;
-      }
-    
-      vector_type vvnrm; vvnrm=zero;  // sum across threads
-      for(int i=0;i<grid->SumArraySize();i++){
-	vvnrm = vvnrm+sumarray[i];
-      } 
-      nrm = Reduce(vvnrm);// sum across simd
-      right._grid->GlobalSum(nrm);
-      return nrm;
-    }
-
-    template<class Op,class T1>
-      inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
-      ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second))))::scalar_object
-    {
-      return sum(closure(expr));
-    }
-
-    template<class Op,class T1,class T2>
-      inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
+template<class Op,class T1,class T2>
+inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
       ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),eval(0,std::get<1>(expr.second))))::scalar_object
-    {
-      return sum(closure(expr));
-    }
-
-
-    template<class Op,class T1,class T2,class T3>
-      inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
-      ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),
-				 eval(0,std::get<1>(expr.second)),
-				 eval(0,std::get<2>(expr.second))
-				 ))::scalar_object
-    {
-      return sum(closure(expr));
-    }
-
-    template<class vobj>
-    inline typename vobj::scalar_object sum(const Lattice<vobj> &arg){
-
-      GridBase *grid=arg._grid;
-      int Nsimd = grid->Nsimd();
-
-      std::vector<vobj,alignedAllocator<vobj> > sumarray(grid->SumArraySize());
-      for(int i=0;i<grid->SumArraySize();i++){
-	sumarray[i]=zero;
-      }
-
-PARALLEL_FOR_LOOP
-      for(int thr=0;thr<grid->SumArraySize();thr++){
-	int nwork, mywork, myoff;
-	GridThread::GetWork(grid->oSites(),thr,mywork,myoff);
-
-	vobj vvsum=zero;
-        for(int ss=myoff;ss<mywork+myoff; ss++){
-	  vvsum = vvsum + arg._odata[ss];
-	}
-	sumarray[thr]=vvsum;
-      }
-
-      vobj vsum=zero;  // sum across threads
-      for(int i=0;i<grid->SumArraySize();i++){
-	vsum = vsum+sumarray[i];
-      } 
-
-      typedef typename vobj::scalar_object sobj;
-      sobj ssum=zero;
-
-      std::vector<sobj>               buf(Nsimd);
-      extract(vsum,buf);
-
-      for(int i=0;i<Nsimd;i++) ssum = ssum + buf[i];
-      arg._grid->GlobalSum(ssum);
-
-      return ssum;
+{
+  return sum(closure(expr));
+}
+
+
+template<class Op,class T1,class T2,class T3>
+inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
+  ->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),
+				      eval(0,std::get<1>(expr.second)),
+				      eval(0,std::get<2>(expr.second))
+				      ))::scalar_object
+{
+  return sum(closure(expr));
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+{
+  GridBase *grid=arg._grid;
+  int Nsimd = grid->Nsimd();
+  
+  std::vector<vobj,alignedAllocator<vobj> > sumarray(grid->SumArraySize());
+  for(int i=0;i<grid->SumArraySize();i++){
+    sumarray[i]=zero;
+  }
+  
+  parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
+    int nwork, mywork, myoff;
+    GridThread::GetWork(grid->oSites(),thr,mywork,myoff);
+    
+    vobj vvsum=zero;
+    for(int ss=myoff;ss<mywork+myoff; ss++){
+      vvsum = vvsum + arg._odata[ss];
     }
+    sumarray[thr]=vvsum;
+  }
+  
+  vobj vsum=zero;  // sum across threads
+  for(int i=0;i<grid->SumArraySize();i++){
+    vsum = vsum+sumarray[i];
+  } 
+  
+  typedef typename vobj::scalar_object sobj;
+  sobj ssum=zero;
+  
+  std::vector<sobj>               buf(Nsimd);
+  extract(vsum,buf);
+  
+  for(int i=0;i<Nsimd;i++) ssum = ssum + buf[i];
+  arg._grid->GlobalSum(ssum);
+  
+  return ssum;
+}
 
 
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
 template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
 {
+  ///////////////////////////////////////////////////////
+  // FIXME precision promoted summation
+  // may be important for correlation functions
+  // But easily avoided by using double precision fields
+  ///////////////////////////////////////////////////////
   typedef typename vobj::scalar_object sobj;
   GridBase  *grid = Data._grid;
   assert(grid!=NULL);
 
-  // FIXME
-  // std::cout<<GridLogMessage<<"WARNING ! SliceSum is unthreaded "<<grid->SumArraySize()<<" threads "<<std::endl;
-
   const int    Nd = grid->_ndimension;
   const int Nsimd = grid->Nsimd();
 
@@ -165,22 +160,30 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
   int rd=grid->_rdimensions[orthogdim];
 
   std::vector<vobj,alignedAllocator<vobj> > lvSum(rd); // will locally sum vectors first
-  std::vector<sobj> lsSum(ld,zero); // sum across these down to scalars
-  std::vector<sobj> extracted(Nsimd);     // splitting the SIMD
+  std::vector<sobj> lsSum(ld,zero);                    // sum across these down to scalars
+  std::vector<sobj> extracted(Nsimd);                  // splitting the SIMD
 
-  result.resize(fd); // And then global sum to return the same vector to every node for IO to file
+  result.resize(fd); // And then global sum to return the same vector to every node 
   for(int r=0;r<rd;r++){
     lvSum[r]=zero;
   }
 
-  std::vector<int>  coor(Nd);  
+  int e1=    grid->_slice_nblock[orthogdim];
+  int e2=    grid->_slice_block [orthogdim];
+  int stride=grid->_slice_stride[orthogdim];
 
   // sum over reduced dimension planes, breaking out orthog dir
+  // Parallel over orthog direction
+  parallel_for(int r=0;r<rd;r++){
 
-  for(int ss=0;ss<grid->oSites();ss++){
-    Lexicographic::CoorFromIndex(coor,ss,grid->_rdimensions);
-    int r = coor[orthogdim];
-    lvSum[r]=lvSum[r]+Data._odata[ss];
+    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._odata[ss];
+      }
+    }
   }
 
   // Sum across simd lanes in the plane, breaking out orthog dir.
@@ -216,10 +219,305 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
 
     result[t]=gsum;
   }
+}
 
+template<class vobj>
+static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim) 
+{
+  typedef typename vobj::vector_type   vector_type;
+  typedef typename vobj::scalar_type   scalar_type;
+  GridBase  *grid = lhs._grid;
+  assert(grid!=NULL);
+  conformable(grid,rhs._grid);
+
+  const int    Nd = grid->_ndimension;
+  const int Nsimd = grid->Nsimd();
+
+  assert(orthogdim >= 0);
+  assert(orthogdim < Nd);
+
+  int fd=grid->_fdimensions[orthogdim];
+  int ld=grid->_ldimensions[orthogdim];
+  int rd=grid->_rdimensions[orthogdim];
+
+  std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd); // will locally sum vectors first
+  std::vector<scalar_type > lsSum(ld,scalar_type(0.0));                    // sum across these down to scalars
+  std::vector<iScalar<scalar_type> > extracted(Nsimd);                  // splitting the SIMD
+
+  result.resize(fd); // And then global sum to return the same vector to every node for IO to file
+  for(int r=0;r<rd;r++){
+    lvSum[r]=zero;
+  }
+
+  int e1=    grid->_slice_nblock[orthogdim];
+  int e2=    grid->_slice_block [orthogdim];
+  int stride=grid->_slice_stride[orthogdim];
+
+  parallel_for(int r=0;r<rd;r++){
+
+    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;
+	vector_type vv = TensorRemove(innerProduct(lhs._odata[ss],rhs._odata[ss]));
+	lvSum[r]=lvSum[r]+vv;
+      }
+    }
+  }
+
+  // Sum across simd lanes in the plane, breaking out orthog dir.
+  std::vector<int> icoor(Nd);
+  for(int rt=0;rt<rd;rt++){
+
+    iScalar<vector_type> temp; 
+    temp._internal = lvSum[rt];
+    extract(temp,extracted);
+
+    for(int idx=0;idx<Nsimd;idx++){
+
+      grid->iCoorFromIindex(icoor,idx);
+
+      int ldx =rt+icoor[orthogdim]*rd;
+
+      lsSum[ldx]=lsSum[ldx]+extracted[idx]._internal;
+
+    }
+  }
+  
+  // sum over nodes.
+  scalar_type gsum;
+  for(int t=0;t<fd;t++){
+    int pt = t/ld; // processor plane
+    int lt = t%ld;
+    if ( pt == grid->_processor_coor[orthogdim] ) {
+      gsum=lsSum[lt];
+    } else {
+      gsum=scalar_type(0.0);
+    }
+
+    grid->GlobalSum(gsum);
+
+    result[t]=gsum;
+  }
+}
+template<class vobj>
+static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Orthog) 
+{
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+  
+  int Nblock = rhs._grid->GlobalDimensions()[Orthog];
+  std::vector<ComplexD> ip(Nblock);
+  sn.resize(Nblock);
+  
+  sliceInnerProductVector(ip,rhs,rhs,Orthog);
+  for(int ss=0;ss<Nblock;ss++){
+    sn[ss] = real(ip[ss]);
+  }
+};
+
+
+template<class vobj>
+static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
+			    int orthogdim,RealD scale=1.0) 
+{    
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+  typedef typename vobj::tensor_reduced tensor_reduced;
+  
+  scalar_type zscale(scale);
+
+  GridBase *grid  = X._grid;
+
+  int Nsimd  =grid->Nsimd();
+  int Nblock =grid->GlobalDimensions()[orthogdim];
+
+  int fd     =grid->_fdimensions[orthogdim];
+  int ld     =grid->_ldimensions[orthogdim];
+  int rd     =grid->_rdimensions[orthogdim];
+
+  int e1     =grid->_slice_nblock[orthogdim];
+  int e2     =grid->_slice_block [orthogdim];
+  int stride =grid->_slice_stride[orthogdim];
+
+  std::vector<int> icoor;
+
+  for(int r=0;r<rd;r++){
+
+    int so=r*grid->_ostride[orthogdim]; // base offset for start of plane 
+
+    vector_type    av;
+
+    for(int l=0;l<Nsimd;l++){
+      grid->iCoorFromIindex(icoor,l);
+      int ldx =r+icoor[orthogdim]*rd;
+      scalar_type *as =(scalar_type *)&av;
+      as[l] = scalar_type(a[ldx])*zscale;
+    }
+
+    tensor_reduced at; at=av;
+
+    parallel_for_nest2(int n=0;n<e1;n++){
+      for(int b=0;b<e2;b++){
+	int ss= so+n*stride+b;
+	R._odata[ss] = at*X._odata[ss]+Y._odata[ss];
+      }
+    }
+  }
+};
+
+
+/*
+template<class vobj>
+static void sliceMaddVectorSlow (Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
+			     int Orthog,RealD scale=1.0) 
+{    
+  // FIXME: Implementation is slow
+  // Best base the linear combination by constructing a 
+  // set of vectors of size grid->_rdimensions[Orthog].
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+  
+  int Nblock = X._grid->GlobalDimensions()[Orthog];
+  
+  GridBase *FullGrid  = X._grid;
+  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  
+  Lattice<vobj> Xslice(SliceGrid);
+  Lattice<vobj> Rslice(SliceGrid);
+  // If we based this on Cshift it would work for spread out
+  // but it would be even slower
+  for(int i=0;i<Nblock;i++){
+    ExtractSlice(Rslice,Y,i,Orthog);
+    ExtractSlice(Xslice,X,i,Orthog);
+    Rslice = Rslice + Xslice*(scale*a[i]);
+    InsertSlice(Rslice,R,i,Orthog);
+  }
+};
+template<class vobj>
+static void sliceInnerProductVectorSlow( std::vector<ComplexD> & vec, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
+  {
+    // FIXME: Implementation is slow
+    // Look at localInnerProduct implementation,
+    // and do inside a site loop with block strided iterators
+    typedef typename vobj::scalar_object sobj;
+    typedef typename vobj::scalar_type scalar_type;
+    typedef typename vobj::vector_type vector_type;
+    typedef typename vobj::tensor_reduced scalar;
+    typedef typename scalar::scalar_object  scomplex;
+  
+    int Nblock = lhs._grid->GlobalDimensions()[Orthog];
+    vec.resize(Nblock);
+    std::vector<scomplex> sip(Nblock);
+    Lattice<scalar> IP(lhs._grid); 
+    IP=localInnerProduct(lhs,rhs);
+    sliceSum(IP,sip,Orthog);
+  
+    for(int ss=0;ss<Nblock;ss++){
+      vec[ss] = TensorRemove(sip[ss]);
+    }
+  }
+*/
+
+//////////////////////////////////////////////////////////////////////////////////////////
+// FIXME: Implementation is slow
+// If we based this on Cshift it would work for spread out
+// but it would be even slower
+//
+// Repeated extract slice is inefficient
+//
+// Best base the linear combination by constructing a 
+// set of vectors of size grid->_rdimensions[Orthog].
+//////////////////////////////////////////////////////////////////////////////////////////
+
+inline GridBase         *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
+{
+  int NN    = BlockSolverGrid->_ndimension;
+  int nsimd = BlockSolverGrid->Nsimd();
+  
+  std::vector<int> latt_phys(0);
+  std::vector<int> simd_phys(0);
+  std::vector<int>  mpi_phys(0);
+  
+  for(int d=0;d<NN;d++){
+    if( d!=Orthog ) { 
+      latt_phys.push_back(BlockSolverGrid->_fdimensions[d]);
+      simd_phys.push_back(BlockSolverGrid->_simd_layout[d]);
+      mpi_phys.push_back(BlockSolverGrid->_processors[d]);
+    }
+  }
+  return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys); 
 }
 
 
+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) 
+{    
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+
+  int Nblock = X._grid->GlobalDimensions()[Orthog];
+  
+  GridBase *FullGrid  = X._grid;
+  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  
+  Lattice<vobj> Xslice(SliceGrid);
+  Lattice<vobj> Rslice(SliceGrid);
+  
+  for(int i=0;i<Nblock;i++){
+    ExtractSlice(Rslice,Y,i,Orthog);
+    for(int j=0;j<Nblock;j++){
+      ExtractSlice(Xslice,X,j,Orthog);
+      Rslice = Rslice + Xslice*(scale*aa(j,i));
+    }
+    InsertSlice(Rslice,R,i,Orthog);
+  }
+};
+
+template<class vobj>
+static void sliceInnerProductMatrix(  Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog) 
+{
+  // FIXME: Implementation is slow
+  // Not sure of best solution.. think about it
+  typedef typename vobj::scalar_object sobj;
+  typedef typename vobj::scalar_type scalar_type;
+  typedef typename vobj::vector_type vector_type;
+  
+  GridBase *FullGrid  = lhs._grid;
+  GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
+  
+  int Nblock = FullGrid->GlobalDimensions()[Orthog];
+  
+  Lattice<vobj> Lslice(SliceGrid);
+  Lattice<vobj> Rslice(SliceGrid);
+  
+  mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
+  
+  for(int i=0;i<Nblock;i++){
+    ExtractSlice(Lslice,lhs,i,Orthog);
+    for(int j=0;j<Nblock;j++){
+      ExtractSlice(Rslice,rhs,j,Orthog);
+      mat(i,j) = innerProduct(Lslice,Rslice);
+    }
+  }
+#undef FORCE_DIAG
+#ifdef FORCE_DIAG
+  for(int i=0;i<Nblock;i++){
+    for(int j=0;j<Nblock;j++){
+      if ( i != j ) mat(i,j)=0.0;
+    }
+  }
+#endif
+  return;
 }
+
+} /*END NAMESPACE GRID*/
 #endif
 
+
+

lib/lattice/Lattice_rng.h

@@ -6,8 +6,8 @@
 
     Copyright (C) 2015
 
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: Guido Cossu <guido.cossu@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
@@ -30,11 +30,19 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define GRID_LATTICE_RNG_H
 
 #include <random>
+
+#ifdef RNG_SITMO
 #include <Grid/sitmo_rng/sitmo_prng_engine.hpp>
+#endif 
+
+#if defined(RNG_SITMO)
+#define RNG_FAST_DISCARD
+#else 
+#undef  RNG_FAST_DISCARD
+#endif
 
 namespace Grid {
 
-
   //////////////////////////////////////////////////////////////
   // Allow the RNG state to be less dense than the fine grid
   //////////////////////////////////////////////////////////////
@@ -64,115 +72,188 @@ namespace Grid {
 
       multiplicity = multiplicity *fine->_rdimensions[fd] / coarse->_rdimensions[d]; 
     }
-
     return multiplicity;
   }
 
+  
+// merge of April 11 2017
+//<<<<<<< HEAD
+
+
+  // this function is necessary for the LS vectorised field
+  inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
+  {
+    int rngdims = coarse->_ndimension;
+    
+    // trivially extended in higher dims, with locality guaranteeing RNG state is local to node
+    int lowerdims   = fine->_ndimension - coarse->_ndimension;  assert(lowerdims >= 0);
+    // assumes that the higher dimensions are not using more processors
+    // all further divisions are local
+    for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
+    for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
+    
+
+    // then divide the number of local sites
+    // check that the total number of sims agree, meanse the iSites are the same
+    assert(fine->Nsimd() == coarse->Nsimd());
+
+    // check that the two grids divide cleanly
+    assert( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
+
+    return fine->lSites() / coarse->lSites();
+  }
+
+  /*
   // Wrap seed_seq to give common interface with random_device
   class fixedSeed {
   public:
-
     typedef std::seed_seq::result_type result_type;
-
     std::seed_seq src;
     
     fixedSeed(const std::vector<int> &seeds) : src(seeds.begin(),seeds.end()) {};
 
     result_type operator () (void){
-
       std::vector<result_type> list(1);
-
       src.generate(list.begin(),list.end());
-
       return list[0];
-
     }
 
   };
 
+=======
+>>>>>>> develop
+  */
+  
   // real scalars are one component
-  template<class scalar,class distribution,class generator> void fillScalar(scalar &s,distribution &dist,generator & gen)
+  template<class scalar,class distribution,class generator> 
+  void fillScalar(scalar &s,distribution &dist,generator & gen)
   {
     s=dist(gen);
   }
-  template<class distribution,class generator> void fillScalar(ComplexF &s,distribution &dist, generator &gen)
+  template<class distribution,class generator> 
+  void fillScalar(ComplexF &s,distribution &dist, generator &gen)
   {
     s=ComplexF(dist(gen),dist(gen));
   }
-  template<class distribution,class generator> void fillScalar(ComplexD &s,distribution &dist,generator &gen)
+  template<class distribution,class generator> 
+  void fillScalar(ComplexD &s,distribution &dist,generator &gen)
   {
     s=ComplexD(dist(gen),dist(gen));
   }
   
   class GridRNGbase {
-
   public:
-
-    int _seeded;
     // One generator per site.
     // Uniform and Gaussian distributions from these generators.
 #ifdef RNG_RANLUX
-    typedef uint64_t      RngStateType;
     typedef std::ranlux48 RngEngine;
+    typedef uint64_t      RngStateType;
     static const int RngStateCount = 15;
-#elif RNG_MT19937 
+#endif 
+#ifdef RNG_MT19937 
     typedef std::mt19937 RngEngine;
     typedef uint32_t     RngStateType;
     static const int     RngStateCount = std::mt19937::state_size;
-#elif RNG_SITMO
+#endif
+#ifdef RNG_SITMO
     typedef sitmo::prng_engine 	RngEngine;
     typedef uint64_t    	RngStateType;
-    static const int    	RngStateCount = 4;
+    static const int    	RngStateCount = 13;
 #endif
-    std::vector<RngEngine>                             _generators;
-    std::vector<std::uniform_real_distribution<RealD>> _uniform;
-    std::vector<std::normal_distribution<RealD>>       _gaussian;
-    std::vector<std::discrete_distribution<int32_t>>     _bernoulli;
 
-    void GetState(std::vector<RngStateType> & saved,int gen) {
+    std::vector<RngEngine>                             _generators;
+    std::vector<std::uniform_real_distribution<RealD> > _uniform;
+    std::vector<std::normal_distribution<RealD> >       _gaussian;
+    std::vector<std::discrete_distribution<int32_t> >   _bernoulli;
+    std::vector<std::uniform_int_distribution<uint32_t> > _uid;
+
+    ///////////////////////
+    // support for parallel init
+    ///////////////////////
+#ifdef RNG_FAST_DISCARD
+    static void Skip(RngEngine &eng)
+    {
+      /////////////////////////////////////////////////////////////////////////////////////
+      // Skip by 2^40 elements between successive lattice sites
+      // This goes by 10^12.
+      // Consider quenched updating; likely never exceeding rate of 1000 sweeps
+      // per second on any machine. This gives us of order 10^9 seconds, or 100 years
+      // skip ahead.
+      // For HMC unlikely to go at faster than a solve per second, and 
+      // tens of seconds per trajectory so this is clean in all reasonable cases,
+      // and margin of safety is orders of magnitude.
+      // We could hack Sitmo to skip in the higher order words of state if necessary
+      /////////////////////////////////////////////////////////////////////////////////////
+      uint64_t skip = 0x1; skip = skip<<40;
+      eng.discard(skip);
+    } 
+#endif
+    static RngEngine Reseed(RngEngine &eng)
+    {
+      std::vector<uint32_t> newseed;
+      std::uniform_int_distribution<uint32_t> uid;
+      return Reseed(eng,newseed,uid);
+    }
+    static RngEngine Reseed(RngEngine &eng,std::vector<uint32_t> & newseed,
+			    std::uniform_int_distribution<uint32_t> &uid)
+    {
+      const int reseeds=4;
+      
+      newseed.resize(reseeds);
+      for(int i=0;i<reseeds;i++){
+	newseed[i] = uid(eng);
+      }
+      std::seed_seq sseq(newseed.begin(),newseed.end());
+      return RngEngine(sseq);
+    }    
+
+    void GetState(std::vector<RngStateType> & saved,RngEngine &eng) {
       saved.resize(RngStateCount);
       std::stringstream ss;
-      ss<<_generators[gen];
+      ss<<eng;
       ss.seekg(0,ss.beg);
       for(int i=0;i<RngStateCount;i++){
-	ss>>saved[i];
+        ss>>saved[i];
       }
     }
-    void SetState(std::vector<RngStateType> & saved,int gen){
+    void GetState(std::vector<RngStateType> & saved,int gen) {
+      GetState(saved,_generators[gen]);
+    }
+    void SetState(std::vector<RngStateType> & saved,RngEngine &eng){
       assert(saved.size()==RngStateCount);
       std::stringstream ss;
       for(int i=0;i<RngStateCount;i++){
-	ss<< saved[i]<<" ";
+        ss<< saved[i]<<" ";
       }
       ss.seekg(0,ss.beg);
-      ss>>_generators[gen];
+      ss>>eng;
+    }
+    void SetState(std::vector<RngStateType> & saved,int gen){
+      SetState(saved,_generators[gen]);
+    }
+    void SetEngine(RngEngine &Eng, int gen){
+      _generators[gen]=Eng;
+    }
+    void GetEngine(RngEngine &Eng, int gen){
+      Eng=_generators[gen];
+    }
+    template<class source> void Seed(source &src, int gen)
+    {
+      _generators[gen] = RngEngine(src);
     }    
   };
 
   class GridSerialRNG : public GridRNGbase {
   public:
 
-    // FIXME ... do we require lockstep draws of randoms 
-    // from all nodes keeping seeds consistent.
-    // place a barrier/broadcast in the fill routine
-    template<class source> void Seed(source &src)
-    {
-      typename source::result_type init = src();
-      CartesianCommunicator::BroadcastWorld(0,(void *)&init,sizeof(init));
-      _generators[0] = RngEngine(init);
-      _seeded=1;
-    }    
-
     GridSerialRNG() : GridRNGbase() {
       _generators.resize(1);
       _uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
       _gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
       _bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
-      _seeded=0;
+      _uid.resize(1,std::uniform_int_distribution<uint32_t>() );
     }
 
-
-
     template <class sobj,class distribution> inline void fill(sobj &l,std::vector<distribution> &dist){
 
       typedef typename sobj::scalar_type scalar_type;
@@ -183,7 +264,7 @@ namespace Grid {
 
       dist[0].reset();
       for(int idx=0;idx<words;idx++){
-	fillScalar(buf[idx],dist[0],_generators[0]);
+  fillScalar(buf[idx],dist[0],_generators[0]);
       }
 
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
@@ -215,7 +296,7 @@ namespace Grid {
       RealF *pointer=(RealF *)&l;
       dist[0].reset();
       for(int i=0;i<2*vComplexF::Nsimd();i++){
-	fillScalar(pointer[i],dist[0],_generators[0]);
+  fillScalar(pointer[i],dist[0],_generators[0]);
       }
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
     }
@@ -223,7 +304,7 @@ namespace Grid {
       RealD *pointer=(RealD *)&l;
       dist[0].reset();
       for(int i=0;i<2*vComplexD::Nsimd();i++){
-	fillScalar(pointer[i],dist[0],_generators[0]);
+  fillScalar(pointer[i],dist[0],_generators[0]);
       }
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
     }
@@ -231,7 +312,7 @@ namespace Grid {
       RealF *pointer=(RealF *)&l;
       dist[0].reset();
       for(int i=0;i<vRealF::Nsimd();i++){
-	fillScalar(pointer[i],dist[0],_generators[0]);
+  fillScalar(pointer[i],dist[0],_generators[0]);
       }
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
     }
@@ -244,155 +325,184 @@ namespace Grid {
       CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
     }
     
-
-    void SeedRandomDevice(void){
-      std::random_device rd;
-      Seed(rd);
-    }
     void SeedFixedIntegers(const std::vector<int> &seeds){
-      fixedSeed src(seeds);
-      Seed(src);
+      CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
+      std::seed_seq src(seeds.begin(),seeds.end());
+      Seed(src,0);
     }
-
   };
 
   class GridParallelRNG : public GridRNGbase {
+
+    double _time_counter;
+
   public:
-
     GridBase *_grid;
-    int _vol;
+    unsigned int _vol;
 
-    int generator_idx(int os,int is){
+    int generator_idx(int os,int is) {
       return is*_grid->oSites()+os;
     }
 
     GridParallelRNG(GridBase *grid) : GridRNGbase() {
-      _grid=grid;
-      _vol =_grid->iSites()*_grid->oSites();
+      _grid = grid;
+      _vol  =_grid->iSites()*_grid->oSites();
 
       _generators.resize(_vol);
       _uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
       _gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
       _bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
-      _seeded=0;
+      _uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
     }
 
-
-    // This loop could be made faster to avoid the Ahmdahl by
-    // i)  seed generators on each timeslice, for x=y=z=0;
-    // ii) seed generators on each z for x=y=0
-    // iii)seed generators on each y,z for x=0
-    // iv) seed generators on each y,z,x 
-    // made possible by physical indexing.
-    template<class source> void Seed(source &src)
-    {
-      std::vector<int> gcoor;
-
-      int gsites = _grid->_gsites;
-
-      typename source::result_type init = src();
-      RngEngine pseeder(init);
-      std::uniform_int_distribution<uint64_t> ui;
-
-      for(int gidx=0;gidx<gsites;gidx++){
-
-	int rank,o_idx,i_idx;
-	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
-	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
-        
-	int l_idx=generator_idx(o_idx,i_idx);
-
-	const int num_rand_seed=16;
-	std::vector<int> site_seeds(num_rand_seed);
-	for(int i=0;i<site_seeds.size();i++){
-	  site_seeds[i]= ui(pseeder);
-	}
-
-	_grid->Broadcast(0,(void *)&site_seeds[0],sizeof(int)*site_seeds.size());
-
-	if( rank == _grid->ThisRank() ){
-	  fixedSeed ssrc(site_seeds);
-	  typename source::result_type sinit = ssrc();
-	  _generators[l_idx] = RngEngine(sinit);
-	}
-      }
-      _seeded=1;
-    }    
-
-    //FIXME implement generic IO and create state save/restore
-    //void SaveState(const std::string<char> &file);
-    //void LoadState(const std::string<char> &file);
-
     template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist){
 
       typedef typename vobj::scalar_object scalar_object;
       typedef typename vobj::scalar_type scalar_type;
       typedef typename vobj::vector_type vector_type;
 
-      int multiplicity = RNGfillable(_grid,l._grid);
+      double inner_time_counter = usecond();
 
-      int     Nsimd =_grid->Nsimd();
-      int     osites=_grid->oSites();
-      int words=sizeof(scalar_object)/sizeof(scalar_type);
+      int multiplicity = RNGfillable_general(_grid, l._grid); // l has finer or same grid
+      int Nsimd  = _grid->Nsimd();  // guaranteed to be the same for l._grid too
+      int osites = _grid->oSites();  // guaranteed to be <= l._grid->oSites() by a factor multiplicity
+      int words  = sizeof(scalar_object) / sizeof(scalar_type);
 
+      parallel_for(int ss=0;ss<osites;ss++){
+        std::vector<scalar_object> buf(Nsimd);
+        for (int m = 0; m < multiplicity; m++) {  // Draw from same generator multiplicity times
 
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<osites;ss++){
+          int sm = multiplicity * ss + m;  // Maps the generator site to the fine site
 
-	std::vector<scalar_object> buf(Nsimd);
-	for(int m=0;m<multiplicity;m++) {// Draw from same generator multiplicity times
+          for (int si = 0; si < Nsimd; si++) {
             
-	  int sm=multiplicity*ss+m;      // Maps the generator site to the fine site
-
-	  for(int si=0;si<Nsimd;si++){
-	    int gdx = generator_idx(ss,si); // index of generator state
-	    scalar_type *pointer = (scalar_type *)&buf[si];
-	    dist[gdx].reset();
-	    for(int idx=0;idx<words;idx++){
-	      fillScalar(pointer[idx],dist[gdx],_generators[gdx]);
-	    }
-	  }
-
-	  // merge into SIMD lanes
-	  merge(l._odata[sm],buf);
-	}
+            int gdx = generator_idx(ss, si);  // index of generator state
+            scalar_type *pointer = (scalar_type *)&buf[si];
+            dist[gdx].reset();
+            for (int idx = 0; idx < words; idx++) 
+              fillScalar(pointer[idx], dist[gdx], _generators[gdx]);
+          }
+          // merge into SIMD lanes, FIXME suboptimal implementation
+          merge(l._odata[sm], buf);
+        }
       }
+
+      _time_counter += usecond()- inner_time_counter;
     };
 
-    void SeedRandomDevice(void){
-      std::random_device rd;
-      Seed(rd);
-    }
     void SeedFixedIntegers(const std::vector<int> &seeds){
-      fixedSeed src(seeds);
-      Seed(src);
+
+      // Everyone generates the same seed_seq based on input seeds
+      CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
+
+      std::seed_seq source(seeds.begin(),seeds.end());
+
+      RngEngine master_engine(source);
+
+#ifdef RNG_FAST_DISCARD
+      ////////////////////////////////////////////////
+      // Skip ahead through a single stream.
+      // Applicable to SITMO and other has based/crypto RNGs
+      // Should be applicable to Mersenne Twister, but the C++11
+      // MT implementation does not implement fast discard even though
+      // in principle this is possible
+      ////////////////////////////////////////////////
+      std::vector<int> gcoor;
+      int rank,o_idx,i_idx;
+
+      // Everybody loops over global volume.
+      for(int gidx=0;gidx<_grid->_gsites;gidx++){
+
+	Skip(master_engine); // Skip to next RNG sequence
+
+	// Where is it?
+	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
+	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
+
+	// If this is one of mine we take it
+	if( rank == _grid->ThisRank() ){
+	  int l_idx=generator_idx(o_idx,i_idx);
+	  _generators[l_idx] = master_engine;
+	}
+
+      }
+#else 
+      ////////////////////////////////////////////////////////////////
+      // Machine and thread decomposition dependent seeding is efficient
+      // and maximally parallel; but NOT reproducible from machine to machine. 
+      // Not ideal, but fastest way to reseed all nodes.
+      ////////////////////////////////////////////////////////////////
+      {
+	// Obtain one Reseed per processor
+	int Nproc = _grid->ProcessorCount();
+	std::vector<RngEngine> seeders(Nproc);
+	int me= _grid->ThisRank();
+	for(int p=0;p<Nproc;p++){
+	  seeders[p] = Reseed(master_engine);
+	}
+	master_engine = seeders[me];
+      }
+
+      {
+	// Obtain one reseeded generator per thread
+	int Nthread = GridThread::GetThreads();
+	std::vector<RngEngine> seeders(Nthread);
+	for(int t=0;t<Nthread;t++){
+	  seeders[t] = Reseed(master_engine);
+	}
+
+	parallel_for(int t=0;t<Nthread;t++) {
+	  // set up one per local site in threaded fashion
+	  std::vector<uint32_t> newseeds;
+	  std::uniform_int_distribution<uint32_t> uid;	
+	  for(int l=0;l<_grid->lSites();l++) {
+	    if ( (l%Nthread)==t ) {
+	      _generators[l] = Reseed(seeders[t],newseeds,uid);
+	    }
+	  }
+	}
+      }
+#endif
+    }
+
+    void Report(){
+      std::cout << GridLogMessage << "Time spent in the fill() routine by GridParallelRNG: "<< _time_counter/1e3 << " ms" << std::endl;
+    }
+
+
+    ////////////////////////////////////////////////////////////////////////
+    // Support for rigorous test of RNG's
+    // Return uniform random uint32_t from requested site generator
+    ////////////////////////////////////////////////////////////////////////
+    uint32_t GlobalU01(int gsite){
+
+      uint32_t the_number;
+      // who
+      std::vector<int> gcoor;
+      int rank,o_idx,i_idx;
+      _grid->GlobalIndexToGlobalCoor(gsite,gcoor);
+      _grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
+
+      // draw
+      int l_idx=generator_idx(o_idx,i_idx);
+      if( rank == _grid->ThisRank() ){
+	the_number = _uid[l_idx](_generators[l_idx]);
+      }
+      
+      // share & return
+      _grid->Broadcast(rank,(void *)&the_number,sizeof(the_number));
+      return the_number;
     }
 
   };
 
-  template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l){
-    rng.fill(l,rng._uniform);
-  }
+  template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l)   { rng.fill(l,rng._uniform);  }
+  template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._gaussian); }
+  template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
 
-  template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l){
-    rng.fill(l,rng._gaussian);
-  }
-  
-  template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){
-    rng.fill(l,rng._bernoulli);
-  }
-
-  template <class sobj> inline void random(GridSerialRNG &rng,sobj &l){
-    rng.fill(l,rng._uniform);
-  }
-  
-  template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l){
-    rng.fill(l,rng._gaussian);
-  }
-  
-  template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){
-    rng.fill(l,rng._bernoulli);
-  }
+  template <class sobj> inline void random(GridSerialRNG &rng,sobj &l)   { rng.fill(l,rng._uniform  ); }
+  template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
+  template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
 
 }
 #endif

lib/lattice/Lattice_trace.h

@@ -42,8 +42,7 @@ namespace Grid {
       -> Lattice<decltype(trace(lhs._odata[0]))>
     {
       Lattice<decltype(trace(lhs._odata[0]))> ret(lhs._grid);
-PARALLEL_FOR_LOOP
-        for(int ss=0;ss<lhs._grid->oSites();ss++){
+      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
             ret._odata[ss] = trace(lhs._odata[ss]);
         }
         return ret;
@@ -56,8 +55,7 @@ PARALLEL_FOR_LOOP
     inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(lhs._odata[0]))>
     {
       Lattice<decltype(traceIndex<Index>(lhs._odata[0]))> ret(lhs._grid);
-PARALLEL_FOR_LOOP
-      for(int ss=0;ss<lhs._grid->oSites();ss++){
+      parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
 	ret._odata[ss] = traceIndex<Index>(lhs._odata[ss]);
       }
       return ret;

lib/lattice/Lattice_transfer.h

@@ -1,4 +1,4 @@
-    /*************************************************************************************
+/*************************************************************************************
 
     Grid physics library, www.github.com/paboyle/Grid 
 
@@ -51,7 +51,7 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
   template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full){
     half.checkerboard = cb;
     int ssh=0;
-    //PARALLEL_FOR_LOOP
+    //parallel_for
     for(int ss=0;ss<full._grid->oSites();ss++){
       std::vector<int> coor;
       int cbos;
@@ -68,7 +68,7 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
   template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
     int cb = half.checkerboard;
     int ssh=0;
-    //PARALLEL_FOR_LOOP
+    //parallel_for
     for(int ss=0;ss<full._grid->oSites();ss++){
       std::vector<int> coor;
       int cbos;
@@ -153,8 +153,7 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
     assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
   }
 
-PARALLEL_FOR_LOOP
-  for(int sf=0;sf<fine->oSites();sf++){
+  parallel_for(int sf=0;sf<fine->oSites();sf++){
     
     int sc;
     std::vector<int> coor_c(_ndimension);
@@ -186,8 +185,7 @@ template<class vobj,class CComplex>
 
   fine_inner = localInnerProduct(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
-PARALLEL_FOR_LOOP
-  for(int ss=0;ss<coarse->oSites();ss++){
+  parallel_for(int ss=0;ss<coarse->oSites();ss++){
     CoarseInner._odata[ss] = coarse_inner._odata[ss];
   }
 }
@@ -333,9 +331,6 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
   typedef typename vobj::scalar_object sobj;
   typedef typename vvobj::scalar_object ssobj;
 
-  sobj s;
-  ssobj ss;
-
   GridBase *ig = in._grid;
   GridBase *og = out._grid;
 
@@ -347,10 +342,13 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
   for(int d=0;d<no;d++){
     assert(ig->_processors[d]  == og->_processors[d]);
     assert(ig->_ldimensions[d] == og->_ldimensions[d]);
+    assert(ig->lSites() == og->lSites());
   }
 
-  //PARALLEL_FOR_LOOP
-  for(int idx=0;idx<ig->lSites();idx++){
+  parallel_for(int idx=0;idx<ig->lSites();idx++){
+    sobj s;
+    ssobj ss;
+
     std::vector<int> lcoor(ni);
     ig->LocalIndexToLocalCoor(idx,lcoor);
     peekLocalSite(s,in,lcoor);
@@ -361,10 +359,9 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
 
 
 template<class vobj>
-void InsertSlice(Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int orthog)
+void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int orthog)
 {
   typedef typename vobj::scalar_object sobj;
-  sobj s;
 
   GridBase *lg = lowDim._grid;
   GridBase *hg = higherDim._grid;
@@ -386,17 +383,16 @@ void InsertSlice(Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int
   }
 
   // the above should guarantee that the operations are local
-  // Guido: check the threading here
-  //PARALLEL_FOR_LOOP
-  for(int idx=0;idx<lg->lSites();idx++){
+  parallel_for(int idx=0;idx<lg->lSites();idx++){
+    sobj s;
     std::vector<int> lcoor(nl);
     std::vector<int> hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
-    dl=0;
+    int ddl=0;
     hcoor[orthog] = slice;
     for(int d=0;d<nh;d++){
       if ( d!=orthog ) { 
-	hcoor[d]=lcoor[dl++];
+	hcoor[d]=lcoor[ddl++];
       }
     }
     peekLocalSite(s,lowDim,lcoor);
@@ -405,10 +401,9 @@ void InsertSlice(Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int
 }
 
 template<class vobj>
-void ExtractSlice(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice, int orthog)
+void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slice, int orthog)
 {
   typedef typename vobj::scalar_object sobj;
-  sobj s;
 
   GridBase *lg = lowDim._grid;
   GridBase *hg = higherDim._grid;
@@ -429,16 +424,16 @@ void ExtractSlice(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice, in
     }
   }
   // the above should guarantee that the operations are local
-  //PARALLEL_FOR_LOOP
-  for(int idx=0;idx<lg->lSites();idx++){
+  parallel_for(int idx=0;idx<lg->lSites();idx++){
+    sobj s;
     std::vector<int> lcoor(nl);
     std::vector<int> hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
-    dl=0;
+    int ddl=0;
     hcoor[orthog] = slice;
     for(int d=0;d<nh;d++){
       if ( d!=orthog ) { 
-	hcoor[d]=lcoor[dl++];
+	hcoor[d]=lcoor[ddl++];
       }
     }
     peekLocalSite(s,higherDim,hcoor);
@@ -449,10 +444,9 @@ void ExtractSlice(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice, in
 
 
 template<class vobj>
-void InsertSliceLocal(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)
 {
   typedef typename vobj::scalar_object sobj;
-  sobj s;
 
   GridBase *lg = lowDim._grid;
   GridBase *hg = higherDim._grid;
@@ -469,8 +463,8 @@ void InsertSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice
   }
 
   // the above should guarantee that the operations are local
-  //PARALLEL_FOR_LOOP
-  for(int idx=0;idx<lg->lSites();idx++){
+  parallel_for(int idx=0;idx<lg->lSites();idx++){
+    sobj s;
     std::vector<int> lcoor(nl);
     std::vector<int> hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
@@ -488,7 +482,6 @@ template<class vobj>
 void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
   typedef typename vobj::scalar_object sobj;
-  sobj s;
 
   GridBase *lg = lowDim._grid;
   GridBase *hg = higherDim._grid;
@@ -505,8 +498,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slic
   }
 
   // the above should guarantee that the operations are local
-  //PARALLEL_FOR_LOOP
-  for(int idx=0;idx<lg->lSites();idx++){
+  parallel_for(int idx=0;idx<lg->lSites();idx++){
+    sobj s;
     std::vector<int> lcoor(nl);
     std::vector<int> hcoor(nh);
     lg->LocalIndexToLocalCoor(idx,lcoor);
@@ -558,7 +551,10 @@ void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
 
 //Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
 template<typename vobj, typename sobj>
-typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in){
+typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type 
+unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
+{
+
   typedef typename vobj::vector_type vtype;
   
   GridBase* in_grid = in._grid;
@@ -574,8 +570,7 @@ typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>
     in_grid->iCoorFromIindex(in_icoor[lane], lane);
   }
   
-PARALLEL_FOR_LOOP
-  for(int in_oidx = 0; in_oidx < in_grid->oSites(); in_oidx++){ //loop over outer index
+  parallel_for(int in_oidx = 0; in_oidx < in_grid->oSites(); in_oidx++){ //loop over outer index
     //Assemble vector of pointers to output elements
     std::vector<sobj*> out_ptrs(in_nsimd);
 
@@ -598,6 +593,54 @@ PARALLEL_FOR_LOOP
     extract1(in_vobj, out_ptrs, 0);
   }
 }
+//Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
+template<typename vobj, typename sobj>
+typename std::enable_if<isSIMDvectorized<vobj>::value 
+                    && !isSIMDvectorized<sobj>::value, void>::type 
+vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
+{
+
+  typedef typename vobj::vector_type vtype;
+  
+  GridBase* grid = out._grid;
+  assert(in.size()==grid->lSites());
+  
+  int ndim     = grid->Nd();
+  int nsimd    = vtype::Nsimd();
+
+  std::vector<std::vector<int> > icoor(nsimd);
+      
+  for(int lane=0; lane < nsimd; lane++){
+    icoor[lane].resize(ndim);
+    grid->iCoorFromIindex(icoor[lane],lane);
+  }
+  
+  parallel_for(uint64_t oidx = 0; oidx < grid->oSites(); oidx++){ //loop over outer index
+    //Assemble vector of pointers to output elements
+    std::vector<sobj*> ptrs(nsimd);
+
+    std::vector<int> ocoor(ndim);
+    grid->oCoorFromOindex(ocoor, oidx);
+
+    std::vector<int> lcoor(grid->Nd());
+      
+    for(int lane=0; lane < nsimd; lane++){
+
+      for(int mu=0;mu<ndim;mu++){
+	lcoor[mu] = ocoor[mu] + grid->_rdimensions[mu]*icoor[lane][mu];
+      }
+
+      int lex;
+      Lexicographic::IndexFromCoor(lcoor, lex, grid->_ldimensions);
+      ptrs[lane] = &in[lex];
+    }
+    
+    //pack from those ptrs
+    vobj vecobj;
+    merge1(vecobj, ptrs, 0);
+    out._odata[oidx] = vecobj; 
+  }
+}
 
 //Convert a Lattice from one precision to another
 template<class VobjOut, class VobjIn>
@@ -623,8 +666,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
   std::vector<SobjOut> in_slex_conv(in_grid->lSites());
   unvectorizeToLexOrdArray(in_slex_conv, in);
     
-  PARALLEL_FOR_LOOP
-  for(int out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
+  parallel_for(uint64_t out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
     std::vector<int> out_ocoor(ndim);
     out_grid->oCoorFromOindex(out_ocoor, out_oidx);
 
@@ -643,9 +685,5 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
   }
 }
  
-
-  
-
- 
 }
 #endif