portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-09-20 09:15:38 +01:00
Code Releases Activity

Attempts at some optimisation

Browse Source
This commit is contained in:
paboyle 2016-02-18 22:33:58 +00:00
parent a3fbabf404
commit db5e8050a8
4 changed files with 4034 additions and 3350 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7328
configure vendored
View File

File diff suppressed because it is too large Load Diff

36
lib/Config.h.in
View File

@ -30,6 +30,15 @@
/* GRID_DEFAULT_PRECISION is SINGLE */
#undef GRID_DEFAULT_PRECISION_SINGLE
/* Support Altivec instructions */
#undef HAVE_ALTIVEC
/* Support AVX (Advanced Vector Extensions) instructions */
#undef HAVE_AVX
/* Support AVX2 (Advanced Vector Extensions 2) instructions */
#undef HAVE_AVX2
/* Define to 1 if you have the declaration of `be64toh', and to 0 if you
don't. */
#undef HAVE_DECL_BE64TOH
@ -44,6 +53,9 @@
/* Define to 1 if you have the <execinfo.h> header file. */
#undef HAVE_EXECINFO_H
/* Support FMA3 (Fused Multiply-Add) instructions */
#undef HAVE_FMA
/* Define to 1 if you have the `gettimeofday' function. */
#undef HAVE_GETTIMEOFDAY
@ -62,9 +74,30 @@
/* Define to 1 if you have the <memory.h> header file. */
#undef HAVE_MEMORY_H
/* Support mmx instructions */
#undef HAVE_MMX
/* Define to 1 if you have the <mm_malloc.h> header file. */
#undef HAVE_MM_MALLOC_H
/* Support SSE (Streaming SIMD Extensions) instructions */
#undef HAVE_SSE
/* Support SSE2 (Streaming SIMD Extensions 2) instructions */
#undef HAVE_SSE2
/* Support SSE3 (Streaming SIMD Extensions 3) instructions */
#undef HAVE_SSE3
/* Support SSSE4.1 (Streaming SIMD Extensions 4.1) instructions */
#undef HAVE_SSE4_1
/* Support SSSE4.2 (Streaming SIMD Extensions 4.2) instructions */
#undef HAVE_SSE4_2
/* Support SSSE3 (Supplemental Streaming SIMD Extensions 3) instructions */
#undef HAVE_SSSE3
/* Define to 1 if you have the <stdint.h> header file. */
#undef HAVE_STDINT_H
@ -107,6 +140,9 @@
/* Define to the one symbol short name of this package. */
#undef PACKAGE_TARNAME
/* Define to the home page for this package. */
#undef PACKAGE_URL
/* Define to the version of this package. */
#undef PACKAGE_VERSION

2
lib/Make.inc
View File

@ -1,4 +1,4 @@
HFILES=./algorithms/approx/bigfloat.h ./algorithms/approx/bigfloat_double.h ./algorithms/approx/Chebyshev.h ./algorithms/approx/MultiShiftFunction.h ./algorithms/approx/Remez.h ./algorithms/approx/Zolotarev.h ./algorithms/CoarsenedMatrix.h ./algorithms/iterative/AdefGeneric.h ./algorithms/iterative/ConjugateGradient.h ./algorithms/iterative/ConjugateGradientMultiShift.h ./algorithms/iterative/ConjugateResidual.h ./algorithms/iterative/DenseMatrix.h ./algorithms/iterative/EigenSort.h ./algorithms/iterative/Francis.h ./algorithms/iterative/Householder.h ./algorithms/iterative/ImplicitlyRestartedLanczos.h ./algorithms/iterative/Matrix.h ./algorithms/iterative/MatrixUtils.h ./algorithms/iterative/NormalEquations.h ./algorithms/iterative/PrecConjugateResidual.h ./algorithms/iterative/PrecGeneralisedConjugateResidual.h ./algorithms/iterative/SchurRedBlack.h ./algorithms/LinearOperator.h ./algorithms/Preconditioner.h ./algorithms/SparseMatrix.h ./Algorithms.h ./AlignedAllocator.h ./cartesian/Cartesian_base.h ./cartesian/Cartesian_full.h ./cartesian/Cartesian_red_black.h ./Cartesian.h ./communicator/Communicator_base.h ./Communicator.h ./cshift/Cshift_common.h ./cshift/Cshift_mpi.h ./cshift/Cshift_none.h ./Cshift.h ./Grid.h ./Init.h ./lattice/Lattice_arith.h ./lattice/Lattice_base.h ./lattice/Lattice_comparison.h ./lattice/Lattice_comparison_utils.h ./lattice/Lattice_conformable.h ./lattice/Lattice_coordinate.h ./lattice/Lattice_ET.h ./lattice/Lattice_local.h ./lattice/Lattice_overload.h ./lattice/Lattice_peekpoke.h ./lattice/Lattice_reality.h ./lattice/Lattice_reduction.h ./lattice/Lattice_rng.h ./lattice/Lattice_trace.h ./lattice/Lattice_transfer.h ./lattice/Lattice_transpose.h ./lattice/Lattice_unary.h ./lattice/Lattice_where.h ./Lattice.h ./Log.h ./Old/Tensor_peek.h ./Old/Tensor_poke.h ./parallelIO/BinaryIO.h ./parallelIO/NerscIO.h ./PerfCount.h ./pugixml/pugixml.h ./qcd/action/ActionBase.h ./qcd/action/ActionParams.h ./qcd/action/Actions.h ./qcd/action/fermion/CayleyFermion5D.h ./qcd/action/fermion/ContinuedFractionFermion5D.h ./qcd/action/fermion/DomainWallFermion.h ./qcd/action/fermion/FermionOperator.h ./qcd/action/fermion/FermionOperatorImpl.h ./qcd/action/fermion/g5HermitianLinop.h ./qcd/action/fermion/MobiusFermion.h ./qcd/action/fermion/MobiusZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h ./qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonContfracTanhFermion.h ./qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h ./qcd/action/fermion/PartialFractionFermion5D.h ./qcd/action/fermion/ScaledShamirFermion.h ./qcd/action/fermion/ShamirZolotarevFermion.h ./qcd/action/fermion/WilsonCompressor.h ./qcd/action/fermion/WilsonFermion.h ./qcd/action/fermion/WilsonFermion5D.h ./qcd/action/fermion/WilsonKernels.h ./qcd/action/fermion/WilsonTMFermion.h ./qcd/action/gauge/GaugeImpl.h ./qcd/action/gauge/PlaqPlusRectangleAction.h ./qcd/action/gauge/WilsonGaugeAction.h ./qcd/action/pseudofermion/EvenOddSchurDifferentiable.h ./qcd/action/pseudofermion/OneFlavourEvenOddRational.h ./qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h ./qcd/action/pseudofermion/OneFlavourRational.h ./qcd/action/pseudofermion/OneFlavourRationalRatio.h ./qcd/action/pseudofermion/TwoFlavour.h ./qcd/action/pseudofermion/TwoFlavourEvenOdd.h ./qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h ./qcd/action/pseudofermion/TwoFlavourRatio.h ./qcd/hmc/HMC.h ./qcd/hmc/HmcRunner.h ./qcd/hmc/integrators/Integrator.h ./qcd/hmc/integrators/Integrator_algorithm.h ./qcd/hmc/NerscCheckpointer.h ./qcd/QCD.h ./qcd/spin/Dirac.h ./qcd/spin/TwoSpinor.h ./qcd/utils/CovariantCshift.h ./qcd/utils/LinalgUtils.h ./qcd/utils/SpaceTimeGrid.h ./qcd/utils/SUn.h ./qcd/utils/WilsonLoops.h ./serialisation/BaseIO.h ./serialisation/BinaryIO.h ./serialisation/MacroMagic.h ./serialisation/Serialisation.h ./serialisation/TextIO.h ./serialisation/XmlIO.h ./simd/Avx512Asm.h ./simd/Grid_avx.h ./simd/Grid_avx512.h ./simd/Grid_empty.h ./simd/Grid_imci.h ./simd/Grid_neon.h ./simd/Grid_qpx.h ./simd/Grid_sse4.h ./simd/Grid_vector_types.h ./simd/Grid_vector_unops.h ./Simd.h ./stencil/Lebesgue.h ./Stencil.h ./tensors/Tensor_arith.h ./tensors/Tensor_arith_add.h ./tensors/Tensor_arith_mac.h ./tensors/Tensor_arith_mul.h ./tensors/Tensor_arith_scalar.h ./tensors/Tensor_arith_sub.h ./tensors/Tensor_class.h ./tensors/Tensor_determinant.h ./tensors/Tensor_exp.h ./tensors/Tensor_extract_merge.h ./tensors/Tensor_index.h ./tensors/Tensor_inner.h ./tensors/Tensor_logical.h ./tensors/Tensor_outer.h ./tensors/Tensor_reality.h ./tensors/Tensor_Ta.h ./tensors/Tensor_trace.h ./tensors/Tensor_traits.h ./tensors/Tensor_transpose.h ./tensors/Tensor_unary.h ./Tensors.h ./Threads.h ./Timer.h
HFILES=./algorithms/approx/bigfloat.h ./algorithms/approx/bigfloat_double.h ./algorithms/approx/Chebyshev.h ./algorithms/approx/MultiShiftFunction.h ./algorithms/approx/Remez.h ./algorithms/approx/Zolotarev.h ./algorithms/CoarsenedMatrix.h ./algorithms/iterative/AdefGeneric.h ./algorithms/iterative/ConjugateGradient.h ./algorithms/iterative/ConjugateGradientMultiShift.h ./algorithms/iterative/ConjugateResidual.h ./algorithms/iterative/DenseMatrix.h ./algorithms/iterative/EigenSort.h ./algorithms/iterative/Francis.h ./algorithms/iterative/Householder.h ./algorithms/iterative/ImplicitlyRestartedLanczos.h ./algorithms/iterative/Matrix.h ./algorithms/iterative/MatrixUtils.h ./algorithms/iterative/NormalEquations.h ./algorithms/iterative/PrecConjugateResidual.h ./algorithms/iterative/PrecGeneralisedConjugateResidual.h ./algorithms/iterative/SchurRedBlack.h ./algorithms/LinearOperator.h ./algorithms/Preconditioner.h ./algorithms/SparseMatrix.h ./Algorithms.h ./AlignedAllocator.h ./cartesian/Cartesian_base.h ./cartesian/Cartesian_full.h ./cartesian/Cartesian_red_black.h ./Cartesian.h ./communicator/Communicator_base.h ./Communicator.h ./cshift/Cshift_common.h ./cshift/Cshift_mpi.h ./cshift/Cshift_none.h ./Cshift.h ./Grid.h ./Init.h ./lattice/Lattice_arith.h ./lattice/Lattice_base.h ./lattice/Lattice_comparison.h ./lattice/Lattice_comparison_utils.h ./lattice/Lattice_conformable.h ./lattice/Lattice_coordinate.h ./lattice/Lattice_ET.h ./lattice/Lattice_local.h ./lattice/Lattice_overload.h ./lattice/Lattice_peekpoke.h ./lattice/Lattice_reality.h ./lattice/Lattice_reduction.h ./lattice/Lattice_rng.h ./lattice/Lattice_trace.h ./lattice/Lattice_transfer.h ./lattice/Lattice_transpose.h ./lattice/Lattice_unary.h ./lattice/Lattice_where.h ./Lattice.h ./Lexicographic.h ./Log.h ./Old/Tensor_peek.h ./Old/Tensor_poke.h ./parallelIO/BinaryIO.h ./parallelIO/NerscIO.h ./PerfCount.h ./pugixml/pugixml.h ./qcd/action/ActionBase.h ./qcd/action/ActionParams.h ./qcd/action/Actions.h ./qcd/action/fermion/CayleyFermion5D.h ./qcd/action/fermion/ContinuedFractionFermion5D.h ./qcd/action/fermion/DomainWallFermion.h ./qcd/action/fermion/FermionOperator.h ./qcd/action/fermion/FermionOperatorImpl.h ./qcd/action/fermion/g5HermitianLinop.h ./qcd/action/fermion/MobiusFermion.h ./qcd/action/fermion/MobiusZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h ./qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonContfracTanhFermion.h ./qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h ./qcd/action/fermion/PartialFractionFermion5D.h ./qcd/action/fermion/ScaledShamirFermion.h ./qcd/action/fermion/ShamirZolotarevFermion.h ./qcd/action/fermion/WilsonCompressor.h ./qcd/action/fermion/WilsonFermion.h ./qcd/action/fermion/WilsonFermion5D.h ./qcd/action/fermion/WilsonKernels.h ./qcd/action/fermion/WilsonTMFermion.h ./qcd/action/gauge/GaugeImpl.h ./qcd/action/gauge/PlaqPlusRectangleAction.h ./qcd/action/gauge/WilsonGaugeAction.h ./qcd/action/pseudofermion/EvenOddSchurDifferentiable.h ./qcd/action/pseudofermion/OneFlavourEvenOddRational.h ./qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h ./qcd/action/pseudofermion/OneFlavourRational.h ./qcd/action/pseudofermion/OneFlavourRationalRatio.h ./qcd/action/pseudofermion/TwoFlavour.h ./qcd/action/pseudofermion/TwoFlavourEvenOdd.h ./qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h ./qcd/action/pseudofermion/TwoFlavourRatio.h ./qcd/hmc/HMC.h ./qcd/hmc/HmcRunner.h ./qcd/hmc/integrators/Integrator.h ./qcd/hmc/integrators/Integrator_algorithm.h ./qcd/hmc/NerscCheckpointer.h ./qcd/QCD.h ./qcd/spin/Dirac.h ./qcd/spin/TwoSpinor.h ./qcd/utils/CovariantCshift.h ./qcd/utils/LinalgUtils.h ./qcd/utils/SpaceTimeGrid.h ./qcd/utils/SUn.h ./qcd/utils/WilsonLoops.h ./serialisation/BaseIO.h ./serialisation/BinaryIO.h ./serialisation/MacroMagic.h ./serialisation/Serialisation.h ./serialisation/TextIO.h ./serialisation/XmlIO.h ./simd/Avx512Asm.h ./simd/Grid_avx.h ./simd/Grid_avx512.h ./simd/Grid_empty.h ./simd/Grid_imci.h ./simd/Grid_neon.h ./simd/Grid_qpx.h ./simd/Grid_sse4.h ./simd/Grid_vector_types.h ./simd/Grid_vector_unops.h ./Simd.h ./stencil/Lebesgue.h ./Stencil.h ./tensors/Tensor_arith.h ./tensors/Tensor_arith_add.h ./tensors/Tensor_arith_mac.h ./tensors/Tensor_arith_mul.h ./tensors/Tensor_arith_scalar.h ./tensors/Tensor_arith_sub.h ./tensors/Tensor_class.h ./tensors/Tensor_determinant.h ./tensors/Tensor_exp.h ./tensors/Tensor_extract_merge.h ./tensors/Tensor_index.h ./tensors/Tensor_inner.h ./tensors/Tensor_logical.h ./tensors/Tensor_outer.h ./tensors/Tensor_reality.h ./tensors/Tensor_Ta.h ./tensors/Tensor_trace.h ./tensors/Tensor_traits.h ./tensors/Tensor_transpose.h ./tensors/Tensor_unary.h ./Tensors.h ./Threads.h ./Timer.h
CCFILES=./algorithms/approx/MultiShiftFunction.cc ./algorithms/approx/Remez.cc ./algorithms/approx/Zolotarev.cc ./Init.cc ./Log.cc ./PerfCount.cc ./pugixml/pugixml.cc ./qcd/action/fermion/CayleyFermion5D.cc ./qcd/action/fermion/ContinuedFractionFermion5D.cc ./qcd/action/fermion/PartialFractionFermion5D.cc ./qcd/action/fermion/WilsonFermion.cc ./qcd/action/fermion/WilsonFermion5D.cc ./qcd/action/fermion/WilsonKernels.cc ./qcd/action/fermion/WilsonKernelsAsm.cc ./qcd/action/fermion/WilsonKernelsHand.cc ./qcd/action/fermion/WilsonTMFermion.cc ./qcd/hmc/HMC.cc ./qcd/spin/Dirac.cc ./qcd/utils/SpaceTimeGrid.cc ./serialisation/BinaryIO.cc ./serialisation/TextIO.cc ./serialisation/XmlIO.cc ./stencil/Lebesgue.cc ./stencil/Stencil_common.cc

18
lib/cshift/Cshift_common.h
View File

@ -56,13 +56,13 @@ Gather_plane_simple (const Lattice<vobj> &rhs,std::vector<cobj,alignedAllocator<
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++){
for(int b=0;b<e2;b++){
int o = n*rhs._grid->_slice_stride[dimension];
int bo = n*rhs._grid->_slice_block[dimension];
int o = n*stride;
int bo = n*e2;
buffer[off+bo+b]=compress(rhs._odata[so+o+b]);
}
}
@ -70,7 +70,7 @@ PARALLEL_NESTED_LOOP2
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 o = n*stride;
int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
if ( ocb &cbmask ) {
buffer[off+bo++]=compress(rhs._odata[so+o+b]);
@ -106,7 +106,6 @@ PARALLEL_NESTED_LOOP2
int o = n*n1;
int offset = b+n*n2;
cobj temp =compress(rhs._odata[so+o+b]);
extract<cobj>(temp,pointers,offset);
@ -244,13 +243,13 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
int e1=rhs._grid->_slice_nblock[dimension]; // clearly loop invariant for icpc
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++){
for(int b=0;b<e2;b++){
int o =n*rhs._grid->_slice_stride[dimension]+b;
int o =n*stride+b;
//lhs._odata[lo+o]=rhs._odata[ro+o];
vstream(lhs._odata[lo+o],rhs._odata[ro+o]);
}
@ -260,7 +259,7 @@ PARALLEL_NESTED_LOOP2
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*rhs._grid->_slice_stride[dimension]+b;
int o =n*stride+b;
int ocb=1<<lhs._grid->CheckerBoardFromOindex(o);
if ( ocb&cbmask ) {
//lhs._odata[lo+o]=rhs._odata[ro+o];
@ -286,11 +285,12 @@ 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++){
for(int b=0;b<e2;b++){
int o =n*rhs._grid->_slice_stride[dimension];
int o =n*stride;
int ocb=1<<lhs._grid->CheckerBoardFromOindex(o+b);
if ( ocb&cbmask ) {
permute(lhs._odata[lo+o+b],rhs._odata[ro+o+b],permute_type);