portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-09 23:45:36 +00:00
Code Releases Activity

Merge branch 'feature/omp-offload' of github.com:BNL-HPC/Grid into feature/omp-offload

Browse Source
This commit is contained in:
Meifeng Lin 2023-12-14 15:33:52 -05:00
commit cc5ab624a2
501 changed files with 51328 additions and 6622 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
.github/ISSUE_TEMPLATE/bug-report.yml vendored Normal file
View File

@ -0,0 +1,54 @@
name: Bug report
description: Report a bug.
title: "<insert title>"
labels: [bug]
body:
- type: markdown
attributes:
value: >
Thank you for taking the time to file a bug report.
Please check that the code is pointing to the HEAD of develop
or any commit in master which is tagged with a version number.
- type: textarea
attributes:
label: "Describe the issue:"
description: >
Describe the issue and any previous attempt to solve it.
validations:
required: true
- type: textarea
attributes:
label: "Code example:"
description: >
If relevant, show how to reproduce the issue using a minimal working
example.
placeholder: |
<< your code here >>
render: shell
validations:
required: false
- type: textarea
attributes:
label: "Target platform:"
description: >
Give a description of the target platform (CPU, network, compiler).
Please give the full CPU part description, using for example
`cat /proc/cpuinfo | grep 'model name' | uniq` (Linux)
or `sysctl machdep.cpu.brand_string` (macOS) and the full output
the `--version` option of your compiler.
validations:
required: true
- type: textarea
attributes:
label: "Configure options:"
description: >
Please give the exact configure command used and attach
`config.log`, `grid.config.summary` and the output of `make V=1`.
render: shell
validations:
required: true

2
Grid/DisableWarnings.h
View File

@ -45,7 +45,7 @@ directory
//disables nvcc specific warning in json.hpp
#pragma clang diagnostic ignored "-Wdeprecated-register"
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
//disables nvcc specific warning in json.hpp
#pragma nv_diag_suppress unsigned_compare_with_zero
#pragma nv_diag_suppress cast_to_qualified_type

3
Grid/GridCore.h
View File

@ -44,9 +44,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridStd.h>
#include <Grid/threads/Pragmas.h>
#include <Grid/perfmon/Timer.h>
#include <Grid/perfmon/PerfCount.h>
//#include <Grid/perfmon/PerfCount.h>
#include <Grid/util/Util.h>
#include <Grid/log/Log.h>
#include <Grid/perfmon/Tracing.h>
#include <Grid/allocator/Allocator.h>
#include <Grid/simd/Simd.h>
#include <Grid/threads/ThreadReduction.h>

1
Grid/GridQCDcore.h
View File

@ -36,6 +36,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
#include <Grid/qcd/QCD.h>
#include <Grid/qcd/spin/Spin.h>
#include <Grid/qcd/gparity/Gparity.h>
#include <Grid/qcd/utils/Utils.h>
#include <Grid/qcd/representations/Representations.h>
NAMESPACE_CHECK(GridQCDCore);

2
Grid/Grid_Eigen_Dense.h
View File

@ -14,7 +14,7 @@
/* NVCC save and restore compile environment*/
#ifdef __NVCC__
#pragma push
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
#pragma nv_diag_suppress code_is_unreachable
#else
#pragma diag_suppress code_is_unreachable

4
Grid/Makefile.am
View File

@ -66,6 +66,10 @@ if BUILD_FERMION_REPS
extra_sources+=$(ADJ_FERMION_FILES)
extra_sources+=$(TWOIND_FERMION_FILES)
endif
if BUILD_SP
extra_sources+=$(SP_FERMION_FILES)
extra_sources+=$(SP_TWOIND_FERMION_FILES)
endif
lib_LIBRARIES = libGrid.a

2
Grid/algorithms/Algorithms.h
View File

@ -54,6 +54,8 @@ NAMESPACE_CHECK(BiCGSTAB);
#include <Grid/algorithms/iterative/SchurRedBlack.h>
#include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
#include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
#include <Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h>
#include <Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h>
#include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
#include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>

22
Grid/algorithms/CoarsenedMatrix.h
View File

@ -324,9 +324,9 @@ public:
GridBase* _cbgrid;
int hermitian;
CartesianStencil<siteVector,siteVector,int> Stencil;
CartesianStencil<siteVector,siteVector,int> StencilEven;
CartesianStencil<siteVector,siteVector,int> StencilOdd;
CartesianStencil<siteVector,siteVector,DefaultImplParams> Stencil;
CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilEven;
CartesianStencil<siteVector,siteVector,DefaultImplParams> StencilOdd;
std::vector<CoarseMatrix> A;
std::vector<CoarseMatrix> Aeven;
@ -631,7 +631,7 @@ public:
assert(Aself != nullptr);
}
void DselfInternal(CartesianStencil<siteVector,siteVector,int> &st, CoarseMatrix &a,
void DselfInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, CoarseMatrix &a,
const CoarseVector &in, CoarseVector &out, int dag) {
int point = geom.npoint-1;
autoView( out_v, out, AcceleratorWrite);
@ -694,7 +694,7 @@ public:
}
}
void DhopInternal(CartesianStencil<siteVector,siteVector,int> &st, std::vector<CoarseMatrix> &a,
void DhopInternal(CartesianStencil<siteVector,siteVector,DefaultImplParams> &st, std::vector<CoarseMatrix> &a,
const CoarseVector &in, CoarseVector &out, int dag) {
SimpleCompressor<siteVector> compressor;
@ -784,9 +784,9 @@ public:
_cbgrid(new GridRedBlackCartesian(&CoarseGrid)),
geom(CoarseGrid._ndimension),
hermitian(hermitian_),
Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements,0),
StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
StencilEven(_cbgrid,geom.npoint,Even,geom.directions,geom.displacements),
StencilOdd(_cbgrid,geom.npoint,Odd,geom.directions,geom.displacements),
A(geom.npoint,&CoarseGrid),
Aeven(geom.npoint,_cbgrid),
Aodd(geom.npoint,_cbgrid),
@ -804,9 +804,9 @@ public:
_cbgrid(&CoarseRBGrid),
geom(CoarseGrid._ndimension),
hermitian(hermitian_),
Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements,0),
Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
StencilEven(&CoarseRBGrid,geom.npoint,Even,geom.directions,geom.displacements),
StencilOdd(&CoarseRBGrid,geom.npoint,Odd,geom.directions,geom.displacements),
A(geom.npoint,&CoarseGrid),
Aeven(geom.npoint,&CoarseRBGrid),
Aodd(geom.npoint,&CoarseRBGrid),

2
Grid/algorithms/LinearOperator.h
View File

@ -526,6 +526,7 @@ public:
(*this)(Linop,in[k],out[k]);
}
};
virtual ~OperatorFunction(){};
};
template<class Field> class LinearFunction {
@ -541,6 +542,7 @@ public:
(*this)(in[i], out[i]);
}
}
virtual ~LinearFunction(){};
};
template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {

16
Grid/algorithms/approx/Chebyshev.h
View File

@ -258,26 +258,12 @@ public:
for(int n=2;n<order;n++){
Linop.HermOp(*Tn,y);
#if 0
auto y_v = y.View();
auto Tn_v = Tn->View();
auto Tnp_v = Tnp->View();
auto Tnm_v = Tnm->View();
constexpr int Nsimd = vector_type::Nsimd();
accelerator_for(ss, in.Grid()->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
});
if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#else
axpby(y,xscale,mscale,y,(*Tn));
axpby(*Tnp,2.0,-1.0,y,(*Tnm));
if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#endif
// Cycle pointers to avoid copies
Field *swizzle = Tnm;
Tnm =Tn;

35
Grid/algorithms/iterative/ConjugateGradient.h
View File

@ -58,6 +58,7 @@ public:
void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
GRID_TRACE("ConjugateGradient");
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
@ -117,9 +118,13 @@ public:
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
RealD usecs = -usecond();
SolverTimer.Start();
int k;
for (k = 1; k <= MaxIterations; k++) {
GridStopWatch IterationTimer;
IterationTimer.Start();
c = cp;
MatrixTimer.Start();
@ -152,31 +157,41 @@ public:
LinearCombTimer.Stop();
LinalgTimer.Stop();
std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
IterationTimer.Stop();
if ( (k % 500) == 0 ) {
std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k
<< " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
} else {
std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
<< " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl;
}
// Stopping condition
if (cp <= rsq) {
usecs +=usecond();
SolverTimer.Stop();
Linop.HermOpAndNorm(psi, mmp, d, qq);
p = mmp - src;
GridBase *grid = src.Grid();
RealD DwfFlops = (1452. )*grid->gSites()*4*k
+ (8+4+8+4+4)*12*grid->gSites()*k; // CG linear algebra
RealD srcnorm = std::sqrt(norm2(src));
RealD resnorm = std::sqrt(norm2(p));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k
<< "\tComputed residual " << std::sqrt(cp / ssq)
<< "\tTrue residual " << true_residual
<< "\tTarget " << Tolerance << std::endl;
std::cout << GridLogIterative << "Time breakdown "<<std::endl;
std::cout << GridLogIterative << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogIterative << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogIterative << "\tLinalg " << LinalgTimer.Elapsed() <<std::endl;
std::cout << GridLogIterative << "\tInner " << InnerTimer.Elapsed() <<std::endl;
std::cout << GridLogIterative << "\tAxpyNorm " << AxpyNormTimer.Elapsed() <<std::endl;
std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<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;
std::cout << GridLogMessage << "\tInner " << InnerTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tAxpyNorm " << AxpyNormTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);

14
Grid/algorithms/iterative/ConjugateGradientMixedPrec.h
View File

@ -49,6 +49,7 @@ NAMESPACE_BEGIN(Grid);
Integer TotalInnerIterations; //Number of inner CG iterations
Integer TotalOuterIterations; //Number of restarts
Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
RealD TrueResidual;
//Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
LinearFunction<FieldF> *guesser;
@ -68,6 +69,7 @@ NAMESPACE_BEGIN(Grid);
}
void operator() (const FieldD &src_d_in, FieldD &sol_d){
std::cout << GridLogMessage << "MixedPrecisionConjugateGradient: Starting mixed precision CG with outer tolerance " << Tolerance << " and inner tolerance " << InnerTolerance << std::endl;
TotalInnerIterations = 0;
GridStopWatch TotalTimer;
@ -97,6 +99,7 @@ NAMESPACE_BEGIN(Grid);
FieldF sol_f(SinglePrecGrid);
sol_f.Checkerboard() = cb;
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting initial inner CG with tolerance " << inner_tol << std::endl;
ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
CG_f.ErrorOnNoConverge = false;
@ -105,7 +108,10 @@ NAMESPACE_BEGIN(Grid);
GridStopWatch PrecChangeTimer;
Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
precisionChangeWorkspace pc_wk_sp_to_dp(DoublePrecGrid, SinglePrecGrid);
precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, DoublePrecGrid);
for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
//Compute double precision rsd and also new RHS vector.
Linop_d.HermOp(sol_d, tmp_d);
@ -120,7 +126,7 @@ NAMESPACE_BEGIN(Grid);
while(norm * inner_tol * inner_tol < stop) inner_tol *= 2; // inner_tol = sqrt(stop/norm) ??
PrecChangeTimer.Start();
precisionChange(src_f, src_d);
precisionChange(src_f, src_d, pc_wk_dp_to_sp);
PrecChangeTimer.Stop();
sol_f = Zero();
@ -130,6 +136,7 @@ NAMESPACE_BEGIN(Grid);
(*guesser)(src_f, sol_f);
//Inner CG
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " << outer_iter << " starting inner CG with tolerance " << inner_tol << std::endl;
CG_f.Tolerance = inner_tol;
InnerCGtimer.Start();
CG_f(Linop_f, src_f, sol_f);
@ -138,7 +145,7 @@ NAMESPACE_BEGIN(Grid);
//Convert sol back to double and add to double prec solution
PrecChangeTimer.Start();
precisionChange(tmp_d, sol_f);
precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp);
PrecChangeTimer.Stop();
axpy(sol_d, 1.0, tmp_d, sol_d);
@ -150,6 +157,7 @@ NAMESPACE_BEGIN(Grid);
ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
CG_d(Linop_d, src_d_in, sol_d);
TotalFinalStepIterations = CG_d.IterationsToComplete;
TrueResidual = CG_d.TrueResidual;
TotalTimer.Stop();
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;

213
Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h Normal file
View File

@ -0,0 +1,213 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h
Copyright (C) 2015
Author: Raoul Hodgson <raoul.hodgson@ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
NAMESPACE_BEGIN(Grid);
//Mixed precision restarted defect correction CG
template<class FieldD,class FieldF,
typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> {
public:
using LinearFunction<FieldD>::operator();
RealD Tolerance;
RealD InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
Integer MaxInnerIterations;
Integer MaxOuterIterations;
Integer MaxPatchupIterations;
GridBase* SinglePrecGrid; //Grid for single-precision fields
RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
LinearOperatorBase<FieldF> &Linop_f;
LinearOperatorBase<FieldD> &Linop_d;
//Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
LinearFunction<FieldF> *guesser;
bool updateResidual;
MixedPrecisionConjugateGradientBatched(RealD tol,
Integer maxinnerit,
Integer maxouterit,
Integer maxpatchit,
GridBase* _sp_grid,
LinearOperatorBase<FieldF> &_Linop_f,
LinearOperatorBase<FieldD> &_Linop_d,
bool _updateResidual=true) :
Linop_f(_Linop_f), Linop_d(_Linop_d),
Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { };
void useGuesser(LinearFunction<FieldF> &g){
guesser = &g;
}
void operator() (const FieldD &src_d_in, FieldD &sol_d){
std::vector<FieldD> srcs_d_in{src_d_in};
std::vector<FieldD> sols_d{sol_d};
(*this)(srcs_d_in,sols_d);
sol_d = sols_d[0];
}
void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
assert(src_d_in.size() == sol_d.size());
int NBatch = src_d_in.size();
std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;
Integer TotalOuterIterations = 0; //Number of restarts
std::vector<Integer> TotalInnerIterations(NBatch,0); //Number of inner CG iterations
std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step
GridStopWatch TotalTimer;
TotalTimer.Start();
GridStopWatch InnerCGtimer;
GridStopWatch PrecChangeTimer;
int cb = src_d_in[0].Checkerboard();
std::vector<RealD> src_norm;
std::vector<RealD> norm;
std::vector<RealD> stop;
GridBase* DoublePrecGrid = src_d_in[0].Grid();
FieldD tmp_d(DoublePrecGrid);
tmp_d.Checkerboard() = cb;
FieldD tmp2_d(DoublePrecGrid);
tmp2_d.Checkerboard() = cb;
std::vector<FieldD> src_d;
std::vector<FieldF> src_f;
std::vector<FieldF> sol_f;
for (int i=0; i<NBatch; i++) {
sol_d[i].Checkerboard() = cb;
src_norm.push_back(norm2(src_d_in[i]));
norm.push_back(0.);
stop.push_back(src_norm[i] * Tolerance*Tolerance);
src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation
src_f.push_back(SinglePrecGrid);
src_f[i].Checkerboard() = cb;
sol_f.push_back(SinglePrecGrid);
sol_f[i].Checkerboard() = cb;
}
RealD inner_tol = InnerTolerance;
ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
CG_f.ErrorOnNoConverge = false;
Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
std::cout << GridLogMessage << std::endl;
std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
bool allConverged = true;
for (int i=0; i<NBatch; i++) {
//Compute double precision rsd and also new RHS vector.
Linop_d.HermOp(sol_d[i], tmp_d);
norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
PrecChangeTimer.Start();
precisionChange(src_f[i], src_d[i]);
PrecChangeTimer.Stop();
sol_f[i] = Zero();
if(norm[i] > OuterLoopNormMult * stop[i]) {
allConverged = false;
}
}
if (allConverged) break;
if (updateResidual) {
RealD normMax = *std::max_element(std::begin(norm), std::end(norm));
RealD stopMax = *std::max_element(std::begin(stop), std::end(stop));
while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2; // inner_tol = sqrt(stop/norm) ??
CG_f.Tolerance = inner_tol;
}
//Optionally improve inner solver guess (eg using known eigenvectors)
if(guesser != NULL) {
(*guesser)(src_f, sol_f);
}
for (int i=0; i<NBatch; i++) {
//Inner CG
InnerCGtimer.Start();
CG_f(Linop_f, src_f[i], sol_f[i]);
InnerCGtimer.Stop();
TotalInnerIterations[i] += CG_f.IterationsToComplete;
//Convert sol back to double and add to double prec solution
PrecChangeTimer.Start();
precisionChange(tmp_d, sol_f[i]);
PrecChangeTimer.Stop();
axpy(sol_d[i], 1.0, tmp_d, sol_d[i]);
}
}
//Final trial CG
std::cout << GridLogMessage << std::endl;
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
for (int i=0; i<NBatch; i++) {
ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
CG_d(Linop_d, src_d_in[i], sol_d[i]);
TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
}
TotalTimer.Stop();
std::cout << GridLogMessage << std::endl;
for (int i=0; i<NBatch; i++) {
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
}
std::cout << GridLogMessage << std::endl;
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
}
};
NAMESPACE_END(Grid);
#endif

12
Grid/algorithms/iterative/ConjugateGradientMultiShift.h
View File

@ -44,7 +44,7 @@ public:
using OperatorFunction<Field>::operator();
RealD Tolerance;
// RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
std::vector<int> IterationsToCompleteShift; // Iterations for this shift
@ -52,7 +52,7 @@ public:
MultiShiftFunction shifts;
std::vector<RealD> TrueResidualShift;
ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) :
ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) :
MaxIterations(maxit),
shifts(_shifts)
{
@ -84,6 +84,7 @@ public:
void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
{
GRID_TRACE("ConjugateGradientMultiShift");
GridBase *grid = src.Grid();
@ -182,6 +183,9 @@ public:
for(int s=0;s<nshift;s++) {
axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
}
std::cout << GridLogIterative << "ConjugateGradientMultiShift: initial rn (|src|^2) =" << rn << " qq (|MdagM src|^2) =" << qq << " d ( dot(src, [MdagM + m_0]src) ) =" << d << " c=" << c << std::endl;
///////////////////////////////////////
// Timers
@ -321,8 +325,8 @@ public:
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tAXPY " << AXPYTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMarix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tAXPY " << AXPYTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tShift " << ShiftTimer.Elapsed() <<std::endl;
IterationsToComplete = k;

373
Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h Normal file
View File

@ -0,0 +1,373 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Christopher Kelly <ckelly@bnl.gov>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#pragma once
NAMESPACE_BEGIN(Grid);
//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision.
//The residual is stored in single precision, but the search directions and solution are stored in double precision.
//Every update_freq iterations the residual is corrected in double precision.
//For safety the a final regular CG is applied to clean up if necessary
//PB Pure single, then double fixup
template<class FieldD, class FieldF,
typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class ConjugateGradientMultiShiftMixedPrecCleanup : public OperatorMultiFunction<FieldD>,
public OperatorFunction<FieldD>
{
public:
using OperatorFunction<FieldD>::operator();
RealD Tolerance;
Integer MaxIterationsMshift;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
std::vector<int> IterationsToCompleteShift; // Iterations for this shift
int verbose;
MultiShiftFunction shifts;
std::vector<RealD> TrueResidualShift;
int ReliableUpdateFreq; //number of iterations between reliable updates
GridBase* SinglePrecGrid; //Grid for single-precision fields
LinearOperatorBase<FieldF> &Linop_f; //single precision
ConjugateGradientMultiShiftMixedPrecCleanup(Integer maxit, const MultiShiftFunction &_shifts,
GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
int _ReliableUpdateFreq) :
MaxIterationsMshift(maxit), shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
MaxIterations(20000)
{
verbose=1;
IterationsToCompleteShift.resize(_shifts.order);
TrueResidualShift.resize(_shifts.order);
}
void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
{
GridBase *grid = src.Grid();
int nshift = shifts.order;
std::vector<FieldD> results(nshift,grid);
(*this)(Linop,src,results,psi);
}
void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
{
int nshift = shifts.order;
(*this)(Linop,src,results);
psi = shifts.norm*src;
for(int i=0;i<nshift;i++){
psi = psi + shifts.residues[i]*results[i];
}
return;
}
void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
{
GRID_TRACE("ConjugateGradientMultiShiftMixedPrecCleanup");
GridBase *DoublePrecGrid = src_d.Grid();
////////////////////////////////////////////////////////////////////////
// Convenience references to the info stored in "MultiShiftFunction"
////////////////////////////////////////////////////////////////////////
int nshift = shifts.order;
std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
std::vector<RealD> &mresidual(shifts.tolerances);
std::vector<RealD> alpha(nshift,1.0);
//Double precision search directions
FieldD p_d(DoublePrecGrid);
std::vector<FieldF> ps_f (nshift, SinglePrecGrid);// Search directions (single precision)
std::vector<FieldF> psi_f(nshift, SinglePrecGrid);// solutions (single precision)
FieldD tmp_d(DoublePrecGrid);
FieldD r_d(DoublePrecGrid);
FieldF r_f(SinglePrecGrid);
FieldD mmp_d(DoublePrecGrid);
assert(psi_d.size()==nshift);
assert(mass.size()==nshift);
assert(mresidual.size()==nshift);
// dynamic sized arrays on stack; 2d is a pain with vector
RealD bs[nshift];
RealD rsq[nshift];
RealD rsqf[nshift];
RealD z[nshift][2];
int converged[nshift];
const int primary =0;
//Primary shift fields CG iteration
RealD a,b,c,d;
RealD cp,bp,qq; //prev
// Matrix mult fields
FieldF p_f(SinglePrecGrid);
FieldF mmp_f(SinglePrecGrid);
// Check lightest mass
for(int s=0;s<nshift;s++){
assert( mass[s]>= mass[primary] );
converged[s]=0;
}
// Wire guess to zero
// Residuals "r" are src
// First search direction "p" is also src
cp = norm2(src_d);
// Handle trivial case of zero src.
if( cp == 0. ){
for(int s=0;s<nshift;s++){
psi_d[s] = Zero();
psi_f[s] = Zero();
IterationsToCompleteShift[s] = 1;
TrueResidualShift[s] = 0.;
}
return;
}
for(int s=0;s<nshift;s++){
rsq[s] = cp * mresidual[s] * mresidual[s];
rsqf[s] =rsq[s];
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
// ps_d[s] = src_d;
precisionChange(ps_f[s],src_d);
}
// r and p for primary
p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
r_d = p_d;
//MdagM+m[0]
precisionChange(p_f,p_d);
Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p d=real(dot(p, mmp)), qq=norm2(mmp)
precisionChange(tmp_d,mmp_f);
Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p d=real(dot(p, mmp)), qq=norm2(mmp)
tmp_d = tmp_d - mmp_d;
std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
// assert(norm2(tmp_d)< 1.0e-4);
axpy(mmp_d,mass[0],p_d,mmp_d);
RealD rn = norm2(p_d);
d += rn*mass[0];
b = -cp /d;
// Set up the various shift variables
int iz=0;
z[0][1-iz] = 1.0;
z[0][iz] = 1.0;
bs[0] = b;
for(int s=1;s<nshift;s++){
z[s][1-iz] = 1.0;
z[s][iz] = 1.0/( 1.0 - b*(mass[s]-mass[0]));
bs[s] = b*z[s][iz];
}
// r += b[0] A.p[0]
// c= norm(r)
c=axpy_norm(r_d,b,mmp_d,r_d);
for(int s=0;s<nshift;s++) {
axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
precisionChange(psi_f[s],psi_d[s]);
}
///////////////////////////////////////
// Timers
///////////////////////////////////////
GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
SolverTimer.Start();
// Iteration loop
int k;
for (k=1;k<=MaxIterationsMshift;k++){
a = c /cp;
AXPYTimer.Start();
axpy(p_d,a,p_d,r_d);
AXPYTimer.Stop();
PrecChangeTimer.Start();
precisionChange(r_f, r_d);
PrecChangeTimer.Stop();
AXPYTimer.Start();
for(int s=0;s<nshift;s++){
if ( ! converged[s] ) {
if (s==0){
axpy(ps_f[s],a,ps_f[s],r_f);
} else{
RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
axpby(ps_f[s],z[s][iz],as,r_f,ps_f[s]);
}
}
}
AXPYTimer.Stop();
cp=c;
PrecChangeTimer.Start();
precisionChange(p_f, p_d); //get back single prec search direction for linop
PrecChangeTimer.Stop();
MatrixTimer.Start();
Linop_f.HermOp(p_f,mmp_f);
MatrixTimer.Stop();
PrecChangeTimer.Start();
precisionChange(mmp_d, mmp_f); // From Float to Double
PrecChangeTimer.Stop();
d=real(innerProduct(p_d,mmp_d));
axpy(mmp_d,mass[0],p_d,mmp_d);
RealD rn = norm2(p_d);
d += rn*mass[0];
bp=b;
b=-cp/d;
// Toggle the recurrence history
bs[0] = b;
iz = 1-iz;
ShiftTimer.Start();
for(int s=1;s<nshift;s++){
if((!converged[s])){
RealD z0 = z[s][1-iz];
RealD z1 = z[s][iz];
z[s][iz] = z0*z1*bp
/ (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b));
bs[s] = b*z[s][iz]/z0; // NB sign rel to Mike
}
}
ShiftTimer.Stop();
//Update single precision solutions
AXPYTimer.Start();
for(int s=0;s<nshift;s++){
int ss = s;
if( (!converged[s]) ) {
axpy(psi_f[ss],-bs[s]*alpha[s],ps_f[s],psi_f[ss]);
}
}
c = axpy_norm(r_d,b,mmp_d,r_d);
AXPYTimer.Stop();
// Convergence checks
int all_converged = 1;
for(int s=0;s<nshift;s++){
if ( (!converged[s]) ){
IterationsToCompleteShift[s] = k;
RealD css = c * z[s][iz]* z[s][iz];
if(css<rsqf[s]){
if ( ! converged[s] )
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
converged[s]=1;
} else {
all_converged=0;
}
}
}
if ( all_converged || k == MaxIterationsMshift-1){
SolverTimer.Stop();
for(int s=0;s<nshift;s++){
precisionChange(psi_d[s],psi_f[s]);
}
if ( all_converged ){
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: All shifts have converged iteration "<<k<<std::endl;
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Checking solutions"<<std::endl;
} else {
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrecCleanup: Not all shifts have converged iteration "<<k<<std::endl;
}
// Check answers
for(int s=0; s < nshift; s++) {
Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
axpy(tmp_d,mass[s],psi_d[s],mmp_d);
axpy(r_d,-alpha[s],src_d,tmp_d);
RealD rn = norm2(r_d);
RealD cn = norm2(src_d);
TrueResidualShift[s] = std::sqrt(rn/cn);
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
//If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
if(rn >= rsq[s]){
CleanupTimer.Start();
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrecCleanup: performing cleanup step for shift " << s << std::endl;
//Setup linear operators for final cleanup
ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d);
cg(src_d, psi_d[s]);
TrueResidualShift[s] = cg.TrueResidual;
CleanupTimer.Stop();
}
}
std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrecCleanup: Time Breakdown for body"<<std::endl;
std::cout << GridLogMessage << "\tSolver " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\t\tAXPY " << AXPYTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\t\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\t\tShift " << ShiftTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
IterationsToComplete = k;
return;
}
}
std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
assert(0);
}
};
NAMESPACE_END(Grid);

416
Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h Normal file
View File

@ -0,0 +1,416 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Christopher Kelly <ckelly@bnl.gov>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
#define GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
NAMESPACE_BEGIN(Grid);
//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision.
//The residual is stored in single precision, but the search directions and solution are stored in double precision.
//Every update_freq iterations the residual is corrected in double precision.
//For safety the a final regular CG is applied to clean up if necessary
//Linop to add shift to input linop, used in cleanup CG
namespace ConjugateGradientMultiShiftMixedPrecSupport{
template<typename Field>
class ShiftedLinop: public LinearOperatorBase<Field>{
public:
LinearOperatorBase<Field> &linop_base;
RealD shift;
ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
void OpDiag (const Field &in, Field &out){ assert(0); }
void OpDir (const Field &in, Field &out,int dir,int disp){ assert(0); }
void OpDirAll (const Field &in, std::vector<Field> &out){ assert(0); }
void Op (const Field &in, Field &out){ assert(0); }
void AdjOp (const Field &in, Field &out){ assert(0); }
void HermOp(const Field &in, Field &out){
linop_base.HermOp(in, out);
axpy(out, shift, in, out);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
HermOp(in,out);
ComplexD dot = innerProduct(in,out);
n1=real(dot);
n2=norm2(out);
}
};
};
template<class FieldD, class FieldF,
typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class ConjugateGradientMultiShiftMixedPrec : public OperatorMultiFunction<FieldD>,
public OperatorFunction<FieldD>
{
public:
using OperatorFunction<FieldD>::operator();
RealD Tolerance;
Integer MaxIterationsMshift;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
std::vector<int> IterationsToCompleteShift; // Iterations for this shift
int verbose;
MultiShiftFunction shifts;
std::vector<RealD> TrueResidualShift;
int ReliableUpdateFreq; //number of iterations between reliable updates
GridBase* SinglePrecGrid; //Grid for single-precision fields
LinearOperatorBase<FieldF> &Linop_f; //single precision
ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
int _ReliableUpdateFreq) :
MaxIterationsMshift(maxit), shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
MaxIterations(20000)
{
verbose=1;
IterationsToCompleteShift.resize(_shifts.order);
TrueResidualShift.resize(_shifts.order);
}
void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
{
GridBase *grid = src.Grid();
int nshift = shifts.order;
std::vector<FieldD> results(nshift,grid);
(*this)(Linop,src,results,psi);
}
void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &psi)
{
int nshift = shifts.order;
(*this)(Linop,src,results);
psi = shifts.norm*src;
for(int i=0;i<nshift;i++){
psi = psi + shifts.residues[i]*results[i];
}
return;
}
void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
{
GRID_TRACE("ConjugateGradientMultiShiftMixedPrec");
GridBase *DoublePrecGrid = src_d.Grid();
precisionChangeWorkspace pc_wk_s_to_d(DoublePrecGrid,SinglePrecGrid);
precisionChangeWorkspace pc_wk_d_to_s(SinglePrecGrid,DoublePrecGrid);
////////////////////////////////////////////////////////////////////////
// Convenience references to the info stored in "MultiShiftFunction"
////////////////////////////////////////////////////////////////////////
int nshift = shifts.order;
std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
std::vector<RealD> &mresidual(shifts.tolerances);
std::vector<RealD> alpha(nshift,1.0);
//Double precision search directions
FieldD p_d(DoublePrecGrid);
std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
FieldD tmp_d(DoublePrecGrid);
FieldD r_d(DoublePrecGrid);
FieldD mmp_d(DoublePrecGrid);
assert(psi_d.size()==nshift);
assert(mass.size()==nshift);
assert(mresidual.size()==nshift);
// dynamic sized arrays on stack; 2d is a pain with vector
RealD bs[nshift];
RealD rsq[nshift];
RealD rsqf[nshift];
RealD z[nshift][2];
int converged[nshift];
const int primary =0;
//Primary shift fields CG iteration
RealD a,b,c,d;
RealD cp,bp,qq; //prev
// Matrix mult fields
FieldF p_f(SinglePrecGrid);
FieldF mmp_f(SinglePrecGrid);
// Check lightest mass
for(int s=0;s<nshift;s++){
assert( mass[s]>= mass[primary] );
converged[s]=0;
}
// Wire guess to zero
// Residuals "r" are src
// First search direction "p" is also src
cp = norm2(src_d);
// Handle trivial case of zero src.
if( cp == 0. ){
for(int s=0;s<nshift;s++){
psi_d[s] = Zero();
IterationsToCompleteShift[s] = 1;
TrueResidualShift[s] = 0.;
}
return;
}
for(int s=0;s<nshift;s++){
rsq[s] = cp * mresidual[s] * mresidual[s];
rsqf[s] =rsq[s];
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
ps_d[s] = src_d;
}
// r and p for primary
p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
r_d = p_d;
//MdagM+m[0]
precisionChange(p_f, p_d, pc_wk_d_to_s);
Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p d=real(dot(p, mmp)), qq=norm2(mmp)
precisionChange(tmp_d, mmp_f, pc_wk_s_to_d);
Linop_d.HermOpAndNorm(p_d,mmp_d,d,qq); // mmp = MdagM p d=real(dot(p, mmp)), qq=norm2(mmp)
tmp_d = tmp_d - mmp_d;
std::cout << " Testing operators match "<<norm2(mmp_d)<<" f "<<norm2(mmp_f)<<" diff "<< norm2(tmp_d)<<std::endl;
assert(norm2(tmp_d)< 1.0);
axpy(mmp_d,mass[0],p_d,mmp_d);
RealD rn = norm2(p_d);
d += rn*mass[0];
b = -cp /d;
// Set up the various shift variables
int iz=0;
z[0][1-iz] = 1.0;
z[0][iz] = 1.0;
bs[0] = b;
for(int s=1;s<nshift;s++){
z[s][1-iz] = 1.0;
z[s][iz] = 1.0/( 1.0 - b*(mass[s]-mass[0]));
bs[s] = b*z[s][iz];
}
// r += b[0] A.p[0]
// c= norm(r)
c=axpy_norm(r_d,b,mmp_d,r_d);
for(int s=0;s<nshift;s++) {
axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
}
///////////////////////////////////////
// Timers
///////////////////////////////////////
GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
SolverTimer.Start();
// Iteration loop
int k;
for (k=1;k<=MaxIterationsMshift;k++){
a = c /cp;
AXPYTimer.Start();
axpy(p_d,a,p_d,r_d);
for(int s=0;s<nshift;s++){
if ( ! converged[s] ) {
if (s==0){
axpy(ps_d[s],a,ps_d[s],r_d);
} else{
RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
}
}
}
AXPYTimer.Stop();
PrecChangeTimer.Start();
precisionChange(p_f, p_d, pc_wk_d_to_s); //get back single prec search direction for linop
PrecChangeTimer.Stop();
cp=c;
MatrixTimer.Start();
Linop_f.HermOp(p_f,mmp_f);
MatrixTimer.Stop();
PrecChangeTimer.Start();
precisionChange(mmp_d, mmp_f, pc_wk_s_to_d); // From Float to Double
PrecChangeTimer.Stop();
AXPYTimer.Start();
d=real(innerProduct(p_d,mmp_d));
axpy(mmp_d,mass[0],p_d,mmp_d);
AXPYTimer.Stop();
RealD rn = norm2(p_d);
d += rn*mass[0];
bp=b;
b=-cp/d;
// Toggle the recurrence history
bs[0] = b;
iz = 1-iz;
ShiftTimer.Start();
for(int s=1;s<nshift;s++){
if((!converged[s])){
RealD z0 = z[s][1-iz];
RealD z1 = z[s][iz];
z[s][iz] = z0*z1*bp
/ (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b));
bs[s] = b*z[s][iz]/z0; // NB sign rel to Mike
}
}
ShiftTimer.Stop();
//Update double precision solutions
AXPYTimer.Start();
for(int s=0;s<nshift;s++){
int ss = s;
if( (!converged[s]) ) {
axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
}
}
//Perform reliable update if necessary; otherwise update residual from single-prec mmp
c = axpy_norm(r_d,b,mmp_d,r_d);
AXPYTimer.Stop();
if(k % ReliableUpdateFreq == 0){
RealD c_old = c;
//Replace r with true residual
MatrixTimer.Start();
Linop_d.HermOp(psi_d[0],mmp_d);
MatrixTimer.Stop();
AXPYTimer.Start();
axpy(mmp_d,mass[0],psi_d[0],mmp_d);
c = axpy_norm(r_d, -1.0, mmp_d, src_d);
AXPYTimer.Stop();
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_old <<" with |r|^2 = "<<c<<std::endl;
}
// Convergence checks
int all_converged = 1;
for(int s=0;s<nshift;s++){
if ( (!converged[s]) ){
IterationsToCompleteShift[s] = k;
RealD css = c * z[s][iz]* z[s][iz];
if(css<rsqf[s]){
if ( ! converged[s] )
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
converged[s]=1;
} else {
all_converged=0;
}
}
}
if ( all_converged || k == MaxIterationsMshift-1){
SolverTimer.Stop();
if ( all_converged ){
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
} else {
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Not all shifts have converged iteration "<<k<<std::endl;
}
// Check answers
for(int s=0; s < nshift; s++) {
Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
axpy(tmp_d,mass[s],psi_d[s],mmp_d);
axpy(r_d,-alpha[s],src_d,tmp_d);
RealD rn = norm2(r_d);
RealD cn = norm2(src_d);
TrueResidualShift[s] = std::sqrt(rn/cn);
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
//If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
if(rn >= rsq[s]){
CleanupTimer.Start();
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
//Setup linear operators for final cleanup
ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d);
cg(src_d, psi_d[s]);
TrueResidualShift[s] = cg.TrueResidual;
CleanupTimer.Stop();
}
}
std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
std::cout << GridLogMessage << "\tSolver " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\t\tAXPY " << AXPYTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\t\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\t\tShift " << ShiftTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
IterationsToComplete = k;
return;
}
}
std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
assert(0);
}
};
NAMESPACE_END(Grid);
#endif

33
Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h
View File

@ -48,7 +48,7 @@ public:
LinearOperatorBase<FieldF> &Linop_f;
LinearOperatorBase<FieldD> &Linop_d;
GridBase* SinglePrecGrid;
RealD Delta; //reliable update parameter
RealD Delta; //reliable update parameter. A reliable update is performed when the residual drops by a factor of Delta relative to its value at the last update
//Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
LinearOperatorBase<FieldF> *Linop_fallback;
@ -65,7 +65,9 @@ public:
ErrorOnNoConverge(err_on_no_conv),
DoFinalCleanup(true),
Linop_fallback(NULL)
{};
{
assert(Delta > 0. && Delta < 1. && "Expect 0 < Delta < 1");
};
void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
Linop_fallback = &_Linop_fallback;
@ -73,6 +75,7 @@ public:
}
void operator()(const FieldD &src, FieldD &psi) {
GRID_TRACE("ConjugateGradientReliableUpdate");
LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
bool using_fallback = false;
@ -115,9 +118,12 @@ public:
}
//Single prec initialization
precisionChangeWorkspace pc_wk_sp_to_dp(src.Grid(), SinglePrecGrid);
precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, src.Grid());
FieldF r_f(SinglePrecGrid);
r_f.Checkerboard() = r.Checkerboard();
precisionChange(r_f, r);
precisionChange(r_f, r, pc_wk_dp_to_sp);
FieldF psi_f(r_f);
psi_f = Zero();
@ -133,7 +139,8 @@ public:
GridStopWatch LinalgTimer;
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
GridStopWatch PrecChangeTimer;
SolverTimer.Start();
int k = 0;
int l = 0;
@ -172,7 +179,9 @@ public:
// Stopping condition
if (cp <= rsq) {
//Although not written in the paper, I assume that I have to add on the final solution
precisionChange(mmp, psi_f);
PrecChangeTimer.Start();
precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
PrecChangeTimer.Stop();
psi = psi + mmp;
@ -193,7 +202,10 @@ public:
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tLinalg " << LinalgTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tPrecChange " << PrecChangeTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tPrecChange avg time " << PrecChangeTimer.Elapsed()/(2*l+1) <<std::endl;
IterationsToComplete = k;
ReliableUpdatesPerformed = l;
@ -213,14 +225,21 @@ public:
else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
<< cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
precisionChange(mmp, psi_f);
PrecChangeTimer.Start();
precisionChange(mmp, psi_f, pc_wk_sp_to_dp);
PrecChangeTimer.Stop();
psi = psi + mmp;
MatrixTimer.Start();
Linop_d.HermOpAndNorm(psi, mmp, d, qq);
MatrixTimer.Stop();
r = src - mmp;
psi_f = Zero();
precisionChange(r_f, r);
PrecChangeTimer.Start();
precisionChange(r_f, r, pc_wk_dp_to_sp);
PrecChangeTimer.Stop();
cp = norm2(r);
MaxResidSinceLastRelUp = cp;

1412
Grid/algorithms/iterative/ImplicitlyRestartedBlockLanczos.h Normal file
View File

File diff suppressed because it is too large Load Diff

9
Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@ -419,14 +419,15 @@ until convergence
}
}
if ( Nconv < Nstop )
if ( Nconv < Nstop ) {
std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
std::cout << GridLogIRL << "returning Nstop vectors, the last "<< Nstop-Nconv << "of which might meet convergence criterion only approximately" <<std::endl;
}
eval=eval2;
//Keep only converged
eval.resize(Nconv);// Nstop?
evec.resize(Nconv,grid);// Nstop?
eval.resize(Nstop);// was Nconv
evec.resize(Nstop,grid);// was Nconv
basisSortInPlace(evec,eval,reverse);
}

65
Grid/algorithms/iterative/LocalCoherenceLanczos.h
View File

@ -44,6 +44,7 @@ public:
int, MinRes); // Must restart
};
//This class is the input parameter class for some testing programs
struct LocalCoherenceLanczosParams : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
@ -145,16 +146,24 @@ public:
LinearOperatorBase<FineField> &_Linop;
RealD _coarse_relax_tol;
std::vector<FineField> &_subspace;
int _largestEvalIdxForReport; //The convergence of the LCL is based on the evals of the coarse grid operator, not those of the underlying fine grid operator
//As a result we do not know what the eval range of the fine operator is until the very end, making tuning the Cheby bounds very difficult
//To work around this issue, every restart we separately reconstruct the fine operator eval for the lowest and highest evec and print these
//out alongside the evals of the coarse operator. To do so we need to know the index of the largest eval (i.e. Nstop-1)
//NOTE: If largestEvalIdxForReport=-1 (default) then this is not performed
ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField> &Poly,
OperatorFunction<FineField> &smoother,
LinearOperatorBase<FineField> &Linop,
std::vector<FineField> &subspace,
RealD coarse_relax_tol=5.0e3)
RealD coarse_relax_tol=5.0e3,
int largestEvalIdxForReport=-1)
: _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
_coarse_relax_tol(coarse_relax_tol)
_coarse_relax_tol(coarse_relax_tol), _largestEvalIdxForReport(largestEvalIdxForReport)
{ };
//evalMaxApprox: approximation of largest eval of the fine Chebyshev operator (suitably wrapped by block projection)
int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
{
CoarseField v(B);
@ -177,12 +186,26 @@ public:
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
<<std::endl;
if(_largestEvalIdxForReport != -1 && (j==0 || j==_largestEvalIdxForReport)){
std::cout<<GridLogIRL << "Estimating true eval of fine grid operator for eval idx " << j << std::endl;
RealD tmp_eval;
ReconstructEval(j,eresid,B,tmp_eval,1.0); //don't use evalMaxApprox of coarse operator! (cf below)
}
int conv=0;
if( (vv<eresid*eresid) ) conv = 1;
return conv;
}
int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
//This function is called at the end of the coarse grid Lanczos. It promotes the coarse eigenvector 'B' to the fine grid,
//applies a smoother to the result then computes the computes the *fine grid* eigenvalue (output as 'eval').
//evalMaxApprox should be the approximation of the largest eval of the fine Hermop. However when this function is called by IRL it actually passes the largest eval of the *Chebyshev* operator (as this is the max approx used for the TestConvergence above)
//As the largest eval of the Chebyshev is typically several orders of magnitude larger this makes the convergence test pass even when it should not.
//We therefore ignore evalMaxApprox here and use a value of 1.0 (note this value is already used by TestCoarse)
int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
{
evalMaxApprox = 1.0; //cf above
GridBase *FineGrid = _subspace[0].Grid();
int checkerboard = _subspace[0].Checkerboard();
FineField fB(FineGrid);fB.Checkerboard() =checkerboard;
@ -201,13 +224,13 @@ public:
eval = vnum/vden;
fv -= eval*fB;
RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
std::cout.precision(13);
std::cout<<GridLogIRL << "[" << std::setw(3)<<j<<"] "
<<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv << " target " << eresid*eresid
<<std::endl;
if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
if( (vv<eresid*eresid) ) return 1;
return 0;
}
@ -285,6 +308,10 @@ public:
evals_coarse.resize(0);
};
//The block inner product is the inner product on the fine grid locally summed over the blocks
//to give a Lattice<Scalar> on the coarse grid. This function orthnormalizes the fine-grid subspace
//vectors under the block inner product. This step must be performed after computing the fine grid
//eigenvectors and before computing the coarse grid eigenvectors.
void Orthogonalise(void ) {
CoarseScalar InnerProd(_CoarseGrid);
std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
@ -328,6 +355,8 @@ public:
}
}
//While this method serves to check the coarse eigenvectors, it also recomputes the eigenvalues from the smoothed reconstructed eigenvectors
//hence the smoother can be tuned after running the coarse Lanczos by using a different smoother here
void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax)
{
assert(evals_fine.size() == nbasis);
@ -376,25 +405,31 @@ public:
evals_fine.resize(nbasis);
subspace.resize(nbasis,_FineGrid);
}
//cheby_op: Parameters of the fine grid Chebyshev polynomial used for the Lanczos acceleration
//cheby_smooth: Parameters of a separate Chebyshev polynomial used after the Lanczos has completed to smooth out high frequency noise in the reconstructed fine grid eigenvectors prior to computing the eigenvalue
//relax: Reconstructed eigenvectors (post smoothing) are naturally not as precise as true eigenvectors. This factor acts as a multiplier on the stopping condition when determining whether the results satisfy the user provided stopping condition
void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
int Nstop, int Nk, int Nm,RealD resid,
RealD MaxIt, RealD betastp, int MinRes)
{
Chebyshev<FineField> Cheby(cheby_op);
ProjectedHermOp<Fobj,CComplex,nbasis> Op(_FineOp,subspace);
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace);
Chebyshev<FineField> Cheby(cheby_op); //Chebyshev of fine operator on fine grid
ProjectedHermOp<Fobj,CComplex,nbasis> Op(_FineOp,subspace); //Fine operator on coarse grid with intermediate fine grid conversion
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace); //Chebyshev of fine operator on coarse grid with intermediate fine grid conversion
//////////////////////////////////////////////////////////////////////////////////////////////////
// create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
//////////////////////////////////////////////////////////////////////////////////////////////////
Chebyshev<FineField> ChebySmooth(cheby_smooth);
ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
Chebyshev<FineField> ChebySmooth(cheby_smooth); //lower order Chebyshev of fine operator on fine grid used to smooth regenerated eigenvectors
ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1);
evals_coarse.resize(Nm);
evec_coarse.resize(Nm,_CoarseGrid);
CoarseField src(_CoarseGrid); src=1.0;
//Note the "tester" here is also responsible for generating the fine grid eigenvalues which are output into the "evals_coarse" array
ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
int Nconv=0;
IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
@ -405,6 +440,14 @@ public:
std::cout << i << " Coarse eval = " << evals_coarse[i] << std::endl;
}
}
//Get the fine eigenvector 'i' by reconstruction
void getFineEvecEval(FineField &evec, RealD &eval, const int i) const{
blockPromote(evec_coarse[i],evec,subspace);
eval = evals_coarse[i];
}
};
NAMESPACE_END(Grid);

2
Grid/algorithms/iterative/PowerMethod.h
View File

@ -29,6 +29,8 @@ template<class Field> class PowerMethod
RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
RealD vden = norm2(src_n);
RealD na = vnum/vden;
std::cout << GridLogIterative << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) {
evalMaxApprox = na;

56
Grid/allocator/MemoryManager.cc
View File

@ -4,11 +4,14 @@ NAMESPACE_BEGIN(Grid);
/*Allocation types, saying which pointer cache should be used*/
#define Cpu (0)
#define CpuSmall (1)
#define Acc (2)
#define AccSmall (3)
#define Shared (4)
#define SharedSmall (5)
#define CpuHuge (1)
#define CpuSmall (2)
#define Acc (3)
#define AccHuge (4)
#define AccSmall (5)
#define Shared (6)
#define SharedHuge (7)
#define SharedSmall (8)
#undef GRID_MM_VERBOSE
uint64_t total_shared;
uint64_t total_device;
@ -35,12 +38,15 @@ void MemoryManager::PrintBytes(void)
}
uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; }
uint64_t MemoryManager::HostCacheBytes() { return CacheBytes[Cpu] + CacheBytes[CpuHuge] + CacheBytes[CpuSmall]; }
//////////////////////////////////////////////////////////////////////
// Data tables for recently freed pooiniter caches
//////////////////////////////////////////////////////////////////////
MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
int MemoryManager::Victim[MemoryManager::NallocType];
int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 8, 16, 8, 16 };
int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 0, 8, 8, 0, 16, 8, 0, 16 };
uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
//////////////////////////////////////////////////////////////////////
// Actual allocation and deallocation utils
@ -170,6 +176,16 @@ void MemoryManager::Init(void)
}
}
str= getenv("GRID_ALLOC_NCACHE_HUGE");
if ( str ) {
Nc = atoi(str);
if ( (Nc>=0) && (Nc < NallocCacheMax)) {
Ncache[CpuHuge]=Nc;
Ncache[AccHuge]=Nc;
Ncache[SharedHuge]=Nc;
}
}
str= getenv("GRID_ALLOC_NCACHE_SMALL");
if ( str ) {
Nc = atoi(str);
@ -190,7 +206,9 @@ void MemoryManager::InitMessage(void) {
std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
#ifdef ALLOCATION_CACHE
std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent host allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<" HUGE "<<Ncache[CpuHuge]<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent device allocations: SMALL "<<Ncache[AccSmall]<<" LARGE "<<Ncache[Acc]<<" Huge "<<Ncache[AccHuge]<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent shared allocations: SMALL "<<Ncache[SharedSmall]<<" LARGE "<<Ncache[Shared]<<" Huge "<<Ncache[SharedHuge]<<std::endl;
#endif
#ifdef GRID_UVM
@ -204,6 +222,9 @@ void MemoryManager::InitMessage(void) {
#ifdef GRID_SYCL
std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_shared"<<std::endl;
#endif
#ifdef GRID_OMPTARGET
std::cout << GridLogMessage<< "MemoryManager::Init() Using OMPTARGET managed memory"<<std::endl;
#endif
#else
std::cout << GridLogMessage<< "MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory"<<std::endl;
#ifdef GRID_CUDA
@ -215,6 +236,9 @@ void MemoryManager::InitMessage(void) {
#ifdef GRID_SYCL
std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_device"<<std::endl;
#endif
#ifdef GRID_OMPTARGET
std::cout << GridLogMessage<< "MemoryManager::Init() Using OMPTARGET omp_alloc_device"<<std::endl;
#endif
#endif
}
@ -222,8 +246,11 @@ void MemoryManager::InitMessage(void) {
void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
{
#ifdef ALLOCATION_CACHE
bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
int cache = type + small;
int cache;
if (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
else cache = type;
return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);
#else
return ptr;
@ -232,11 +259,12 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes)
{
assert(ncache>0);
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif
if (ncache == 0) return ptr;
void * ret = NULL;
int v = -1;
@ -271,8 +299,11 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
void *MemoryManager::Lookup(size_t bytes,int type)
{
#ifdef ALLOCATION_CACHE
bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
int cache = type+small;
int cache;
if (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
else cache = type;
return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
#else
return NULL;
@ -281,7 +312,6 @@ void *MemoryManager::Lookup(size_t bytes,int type)
void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes)
{
assert(ncache>0);
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif

39
Grid/allocator/MemoryManager.h
View File

@ -35,6 +35,7 @@ NAMESPACE_BEGIN(Grid);
// Move control to configure.ac and Config.h?
#define GRID_ALLOC_SMALL_LIMIT (4096)
#define GRID_ALLOC_HUGE_LIMIT (2147483648)
#define STRINGIFY(x) #x
#define TOSTRING(x) STRINGIFY(x)
@ -70,6 +71,21 @@ enum ViewMode {
CpuWriteDiscard = 0x10 // same for now
};
struct MemoryStatus {
uint64_t DeviceBytes;
uint64_t DeviceLRUBytes;
uint64_t DeviceMaxBytes;
uint64_t HostToDeviceBytes;
uint64_t DeviceToHostBytes;
uint64_t HostToDeviceXfer;
uint64_t DeviceToHostXfer;
uint64_t DeviceEvictions;
uint64_t DeviceDestroy;
uint64_t DeviceAllocCacheBytes;
uint64_t HostAllocCacheBytes;
};
class MemoryManager {
private:
@ -83,7 +99,7 @@ private:
} AllocationCacheEntry;
static const int NallocCacheMax=128;
static const int NallocType=6;
static const int NallocType=9;
static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
static int Victim[NallocType];
static int Ncache[NallocType];
@ -121,7 +137,26 @@ private:
static uint64_t DeviceToHostXfer;
static uint64_t DeviceEvictions;
static uint64_t DeviceDestroy;
static uint64_t DeviceCacheBytes();
static uint64_t HostCacheBytes();
static MemoryStatus GetFootprint(void) {
MemoryStatus stat;
stat.DeviceBytes = DeviceBytes;
stat.DeviceLRUBytes = DeviceLRUBytes;
stat.DeviceMaxBytes = DeviceMaxBytes;
stat.HostToDeviceBytes = HostToDeviceBytes;
stat.DeviceToHostBytes = DeviceToHostBytes;
stat.HostToDeviceXfer = HostToDeviceXfer;
stat.DeviceToHostXfer = DeviceToHostXfer;
stat.DeviceEvictions = DeviceEvictions;
stat.DeviceDestroy = DeviceDestroy;
stat.DeviceAllocCacheBytes = DeviceCacheBytes();
stat.HostAllocCacheBytes = HostCacheBytes();
return stat;
};
private:
#ifndef GRID_UVM
//////////////////////////////////////////////////////////////////////

8
Grid/allocator/MemoryManagerCache.cc
View File

@ -144,8 +144,8 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
(uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock);
assert(AccCache.accLock==0); // Cannot evict so logic bomb
assert(AccCache.CpuPtr!=(uint64_t)NULL);
if (AccCache.accLock!=0) return;
if (AccCache.cpuLock!=0) return;
if(AccCache.state==AccDirty) {
Flush(AccCache);
}
@ -519,7 +519,6 @@ void MemoryManager::Audit(std::string s)
uint64_t LruBytes1=0;
uint64_t LruBytes2=0;
uint64_t LruCnt=0;
uint64_t LockedBytes=0;
std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
for(auto it=LRU.begin();it!=LRU.end();it++){
@ -532,6 +531,7 @@ void MemoryManager::Audit(std::string s)
assert(AccCache.LRU_entry==it);
}
std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
auto &AccCache = it->second;
@ -548,6 +548,7 @@ void MemoryManager::Audit(std::string s)
if ( AccCache.cpuLock || AccCache.accLock ) {
assert(AccCache.LRU_valid==0);
std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
<< "\t cpuLock " << AccCache.cpuLock
@ -566,6 +567,7 @@ void MemoryManager::Audit(std::string s)
std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
assert(LruCnt == LRU.size());
std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
}
void MemoryManager::PrintState(void* _CpuPtr)

21
Grid/communicator/Communicator_base.h
View File

@ -53,10 +53,11 @@ public:
// Communicator should know nothing of the physics grid, only processor grid.
////////////////////////////////////////////
int _Nprocessors; // How many in all
Coordinate _processors; // Which dimensions get relayed out over processors lanes.
int _processor; // linear processor rank
Coordinate _processor_coor; // linear processor coordinate
unsigned long _ndimension;
Coordinate _shm_processors; // Which dimensions get relayed out over processors lanes.
Coordinate _processors; // Which dimensions get relayed out over processors lanes.
Coordinate _processor_coor; // linear processor coordinate
static Grid_MPI_Comm communicator_world;
Grid_MPI_Comm communicator;
std::vector<Grid_MPI_Comm> communicator_halo;
@ -97,14 +98,16 @@ public:
int BossRank(void) ;
int ThisRank(void) ;
const Coordinate & ThisProcessorCoor(void) ;
const Coordinate & ShmGrid(void) { return _shm_processors; } ;
const Coordinate & ProcessorGrid(void) ;
int ProcessorCount(void) ;
int ProcessorCount(void) ;
////////////////////////////////////////////////////////////////////////////////
// very VERY rarely (Log, serial RNG) we need world without a grid
////////////////////////////////////////////////////////////////////////////////
static int RankWorld(void) ;
static void BroadcastWorld(int root,void* data, int bytes);
static void BarrierWorld(void);
////////////////////////////////////////////////////////////
// Reduction
@ -128,7 +131,7 @@ public:
template<class obj> void GlobalSum(obj &o){
typedef typename obj::scalar_type scalar_type;
int words = sizeof(obj)/sizeof(scalar_type);
scalar_type * ptr = (scalar_type *)& o;
scalar_type * ptr = (scalar_type *)& o; // Safe alias
GlobalSumVector(ptr,words);
}
@ -142,17 +145,17 @@ public:
int bytes);
double StencilSendToRecvFrom(void *xmit,
int xmit_to_rank,
int xmit_to_rank,int do_xmit,
void *recv,
int recv_from_rank,
int recv_from_rank,int do_recv,
int bytes,int dir);
double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
int xmit_to_rank,int do_xmit,
void *recv,
int recv_from_rank,
int bytes,int dir);
int recv_from_rank,int do_recv,
int xbytes,int rbytes,int dir);
void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);

69
Grid/communicator/Communicator_mpi3.cc
View File

@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
// Remap using the shared memory optimising routine
// The remap creates a comm which must be freed
////////////////////////////////////////////////////
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm);
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm,_shm_processors);
InitFromMPICommunicator(processors,optimal_comm);
SetCommunicator(optimal_comm);
///////////////////////////////////////////////////
@ -124,12 +124,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
Coordinate parent_processor_coor(_ndimension,0);
Coordinate parent_processors (_ndimension,1);
Coordinate shm_processors (_ndimension,1);
// Can make 5d grid from 4d etc...
int pad = _ndimension-parent_ndimension;
for(int d=0;d<parent_ndimension;d++){
parent_processor_coor[pad+d]=parent._processor_coor[d];
parent_processors [pad+d]=parent._processors[d];
shm_processors [pad+d]=parent._shm_processors[d];
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
@ -154,6 +155,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
ccoor[d] = parent_processor_coor[d] % processors[d];
scoor[d] = parent_processor_coor[d] / processors[d];
ssize[d] = parent_processors[d] / processors[d];
if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
}
// rank within subcomm ; srank is rank of subcomm within blocks of subcomms
@ -335,23 +337,23 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
}
// Basic Halo comms primitive
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int dest,
int dest, int dox,
void *recv,
int from,
int from, int dor,
int bytes,int dir)
{
std::vector<CommsRequest_t> list;
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
StencilSendToRecvFromComplete(list,dir);
return offbytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
int dest,int dox,
void *recv,
int from,
int bytes,int dir)
int from,int dor,
int xbytes,int rbytes,int dir)
{
int ncomm =communicator_halo.size();
int commdir=dir%ncomm;
@ -370,39 +372,34 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
double off_node_bytes=0.0;
int tag;
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+from*32;
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
assert(ierr==0);
list.push_back(rrq);
off_node_bytes+=bytes;
if ( dor ) {
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+from*32;
ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
assert(ierr==0);
list.push_back(rrq);
off_node_bytes+=rbytes;
}
}
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+_processor*32;
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
assert(ierr==0);
list.push_back(xrq);
off_node_bytes+=bytes;
} else {
// TODO : make a OMP loop on CPU, call threaded bcopy
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
assert(shm!=NULL);
// std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
if (dox) {
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+_processor*32;
ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
assert(ierr==0);
list.push_back(xrq);
off_node_bytes+=xbytes;
} else {
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
assert(shm!=NULL);
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
}
}
// if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
// this->StencilSendToRecvFromComplete(list,dir);
// }
return off_node_bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
{
// std::cout << "Copy Synchronised\n"<<std::endl;
acceleratorCopySynchronise();
int nreq=list.size();
if (nreq==0) return;
@ -438,6 +435,10 @@ int CartesianCommunicator::RankWorld(void){
MPI_Comm_rank(communicator_world,&r);
return r;
}
void CartesianCommunicator::BarrierWorld(void){
int ierr = MPI_Barrier(communicator_world);
assert(ierr==0);
}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
{
int ierr= MPI_Bcast(data,

15
Grid/communicator/Communicator_none.cc
View File

@ -45,12 +45,14 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
: CartesianCommunicator(processors)
{
_shm_processors = Coordinate(processors.size(),1);
srank=0;
SetCommunicator(communicator_world);
}
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
{
_shm_processors = Coordinate(processors.size(),1);
_processors = processors;
_ndimension = processors.size(); assert(_ndimension>=1);
_processor_coor.resize(_ndimension);
@ -102,6 +104,7 @@ int CartesianCommunicator::RankWorld(void){return 0;}
void CartesianCommunicator::Barrier(void){}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
void CartesianCommunicator::BarrierWorld(void) { }
int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) { return 0;}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){ coor = _processor_coor; }
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
@ -111,21 +114,21 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
}
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int xmit_to_rank,
int xmit_to_rank,int dox,
void *recv,
int recv_from_rank,
int recv_from_rank,int dor,
int bytes, int dir)
{
return 2.0*bytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
int xmit_to_rank,int dox,
void *recv,
int recv_from_rank,
int bytes, int dir)
int recv_from_rank,int dor,
int xbytes,int rbytes, int dir)
{
return 2.0*bytes;
return xbytes+rbytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{

53
Grid/communicator/SharedMemory.cc
View File

@ -91,6 +91,59 @@ void *SharedMemory::ShmBufferSelf(void)
//std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
return ShmCommBufs[ShmRank];
}
static inline int divides(int a,int b)
{
return ( b == ( (b/a)*a ) );
}
void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
{
////////////////////////////////////////////////////////////////
// Allow user to configure through environment variable
////////////////////////////////////////////////////////////////
char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
if ( str ) {
std::vector<int> IntShmDims;
GridCmdOptionIntVector(std::string(str),IntShmDims);
assert(IntShmDims.size() == WorldDims.size());
long ShmSize = 1;
for (int dim=0;dim<WorldDims.size();dim++) {
ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
assert(divides(ShmDims[dim],WorldDims[dim]));
}
assert(ShmSize == WorldShmSize);
return;
}
////////////////////////////////////////////////////////////////
// Powers of 2,3,5 only in prime decomposition for now
////////////////////////////////////////////////////////////////
int ndimension = WorldDims.size();
ShmDims=Coordinate(ndimension,1);
std::vector<int> primes({2,3,5});
int dim = 0;
int last_dim = ndimension - 1;
int AutoShmSize = 1;
while(AutoShmSize != WorldShmSize) {
int p;
for(p=0;p<primes.size();p++) {
int prime=primes[p];
if ( divides(prime,WorldDims[dim]/ShmDims[dim])
&& divides(prime,WorldShmSize/AutoShmSize) ) {
AutoShmSize*=prime;
ShmDims[dim]*=prime;
last_dim = dim;
break;
}
}
if (p == primes.size() && last_dim == dim) {
std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
exit(EXIT_FAILURE);
}
dim=(dim+1) %ndimension;
}
}
NAMESPACE_END(Grid);

7
Grid/communicator/SharedMemory.h
View File

@ -93,9 +93,10 @@ public:
// Create an optimal reordered communicator that makes MPI_Cart_create get it right
//////////////////////////////////////////////////////////////////////////////////////
static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
static void OptimalCommunicator (const Coordinate &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicatorHypercube (const Coordinate &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
// Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicator (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);
static void OptimalCommunicatorHypercube (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);
static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);
static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims);
///////////////////////////////////////////////////
// Provide shared memory facilities off comm world

250
Grid/communicator/SharedMemoryMPI.cc
View File

@ -27,6 +27,8 @@ Author: Christoph Lehner <christoph@lhnr.de>
*************************************************************************************/
/* END LEGAL */
#define Mheader "SharedMemoryMpi: "
#include <Grid/GridCore.h>
#include <pwd.h>
@ -36,12 +38,120 @@ Author: Christoph Lehner <christoph@lhnr.de>
#ifdef GRID_HIP
#include <hip/hip_runtime_api.h>
#endif
#ifdef GRID_SYCl
#ifdef GRID_SYCL
#define GRID_SYCL_LEVEL_ZERO_IPC
#include <syscall.h>
#define SHM_SOCKETS
#endif
#include <sys/socket.h>
#include <sys/un.h>
NAMESPACE_BEGIN(Grid);
#define header "SharedMemoryMpi: "
#ifdef SHM_SOCKETS
/*
* Barbaric extra intranode communication route in case we need sockets to pass FDs
* Forced by level_zero not being nicely designed
*/
static int sock;
static const char *sock_path_fmt = "/tmp/GridUnixSocket.%d";
static char sock_path[256];
class UnixSockets {
public:
static void Open(int rank)
{
int errnum;
sock = socket(AF_UNIX, SOCK_DGRAM, 0); assert(sock>0);
struct sockaddr_un sa_un = { 0 };
sa_un.sun_family = AF_UNIX;
snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,rank);
unlink(sa_un.sun_path);
if (bind(sock, (struct sockaddr *)&sa_un, sizeof(sa_un))) {
perror("bind failure");
exit(EXIT_FAILURE);
}
}
static int RecvFileDescriptor(void)
{
int n;
int fd;
char buf[1];
struct iovec iov;
struct msghdr msg;
struct cmsghdr *cmsg;
char cms[CMSG_SPACE(sizeof(int))];
iov.iov_base = buf;
iov.iov_len = 1;
memset(&msg, 0, sizeof msg);
msg.msg_name = 0;
msg.msg_namelen = 0;
msg.msg_iov = &iov;
msg.msg_iovlen = 1;
msg.msg_control = (caddr_t)cms;
msg.msg_controllen = sizeof cms;
if((n=recvmsg(sock, &msg, 0)) < 0) {
perror("recvmsg failed");
return -1;
}
if(n == 0){
perror("recvmsg returned 0");
return -1;
}
cmsg = CMSG_FIRSTHDR(&msg);
memmove(&fd, CMSG_DATA(cmsg), sizeof(int));
return fd;
}
static void SendFileDescriptor(int fildes,int xmit_to_rank)
{
struct msghdr msg;
struct iovec iov;
struct cmsghdr *cmsg = NULL;
char ctrl[CMSG_SPACE(sizeof(int))];
char data = ' ';
memset(&msg, 0, sizeof(struct msghdr));
memset(ctrl, 0, CMSG_SPACE(sizeof(int)));
iov.iov_base = &data;
iov.iov_len = sizeof(data);
sprintf(sock_path,sock_path_fmt,xmit_to_rank);
struct sockaddr_un sa_un = { 0 };
sa_un.sun_family = AF_UNIX;
snprintf(sa_un.sun_path, sizeof(sa_un.sun_path),sock_path_fmt,xmit_to_rank);
msg.msg_name = (void *)&sa_un;
msg.msg_namelen = sizeof(sa_un);
msg.msg_iov = &iov;
msg.msg_iovlen = 1;
msg.msg_controllen = CMSG_SPACE(sizeof(int));
msg.msg_control = ctrl;
cmsg = CMSG_FIRSTHDR(&msg);
cmsg->cmsg_level = SOL_SOCKET;
cmsg->cmsg_type = SCM_RIGHTS;
cmsg->cmsg_len = CMSG_LEN(sizeof(int));
*((int *) CMSG_DATA(cmsg)) = fildes;
sendmsg(sock, &msg, 0);
};
};
#endif
/*Construct from an MPI communicator*/
void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
{
@ -64,8 +174,8 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
MPI_Comm_size(WorldShmComm ,&WorldShmSize);
if ( WorldRank == 0) {
std::cout << header " World communicator of size " <<WorldSize << std::endl;
std::cout << header " Node communicator of size " <<WorldShmSize << std::endl;
std::cout << Mheader " World communicator of size " <<WorldSize << std::endl;
std::cout << Mheader " Node communicator of size " <<WorldShmSize << std::endl;
}
// WorldShmComm, WorldShmSize, WorldShmRank
@ -152,7 +262,7 @@ int Log2Size(int TwoToPower,int MAXLOG2)
}
return log2size;
}
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
{
//////////////////////////////////////////////////////////////////////////////
// Look and see if it looks like an HPE 8600 based on hostname conventions
@ -165,63 +275,11 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
gethostname(name,namelen);
int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
else OptimalCommunicatorSharedMemory(processors,optimal_comm);
if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
else OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
}
static inline int divides(int a,int b)
{
return ( b == ( (b/a)*a ) );
}
void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
{
////////////////////////////////////////////////////////////////
// Allow user to configure through environment variable
////////////////////////////////////////////////////////////////
char* str = getenv(("GRID_SHM_DIMS_" + std::to_string(ShmDims.size())).c_str());
if ( str ) {
std::vector<int> IntShmDims;
GridCmdOptionIntVector(std::string(str),IntShmDims);
assert(IntShmDims.size() == WorldDims.size());
long ShmSize = 1;
for (int dim=0;dim<WorldDims.size();dim++) {
ShmSize *= (ShmDims[dim] = IntShmDims[dim]);
assert(divides(ShmDims[dim],WorldDims[dim]));
}
assert(ShmSize == WorldShmSize);
return;
}
////////////////////////////////////////////////////////////////
// Powers of 2,3,5 only in prime decomposition for now
////////////////////////////////////////////////////////////////
int ndimension = WorldDims.size();
ShmDims=Coordinate(ndimension,1);
std::vector<int> primes({2,3,5});
int dim = 0;
int last_dim = ndimension - 1;
int AutoShmSize = 1;
while(AutoShmSize != WorldShmSize) {
int p;
for(p=0;p<primes.size();p++) {
int prime=primes[p];
if ( divides(prime,WorldDims[dim]/ShmDims[dim])
&& divides(prime,WorldShmSize/AutoShmSize) ) {
AutoShmSize*=prime;
ShmDims[dim]*=prime;
last_dim = dim;
break;
}
}
if (p == primes.size() && last_dim == dim) {
std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
exit(EXIT_FAILURE);
}
dim=(dim+1) %ndimension;
}
}
void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
{
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
@ -294,7 +352,8 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
Coordinate HyperCoor(ndimension);
GetShmDims(WorldDims,ShmDims);
SHM = ShmDims;
////////////////////////////////////////////////////////////////
// Establish torus of processes and nodes with sub-blockings
////////////////////////////////////////////////////////////////
@ -341,7 +400,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
assert(ierr==0);
}
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
{
////////////////////////////////////////////////////////////////
// Identify subblock of ranks on node spreading across dims
@ -353,6 +412,8 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
Coordinate ShmCoor(ndimension); Coordinate NodeCoor(ndimension); Coordinate WorldCoor(ndimension);
GetShmDims(WorldDims,ShmDims);
SHM=ShmDims;
////////////////////////////////////////////////////////////////
// Establish torus of processes and nodes with sub-blockings
////////////////////////////////////////////////////////////////
@ -391,7 +452,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
#ifdef GRID_MPI3_SHMGET
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
@ -476,7 +537,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
exit(EXIT_FAILURE);
}
std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes
std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes
<< "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
SharedMemoryZero(ShmCommBuf,bytes);
@ -519,16 +580,21 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
exit(EXIT_FAILURE);
}
if ( WorldRank == 0 ){
std::cout << WorldRank << header " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes
<< "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes
<< "bytes at "<< std::hex<< ShmCommBuf << " - "<<(bytes-1+(uint64_t)ShmCommBuf) <<std::dec<<" for comms buffers " <<std::endl;
}
SharedMemoryZero(ShmCommBuf,bytes);
std::cout<< "Setting up IPC"<<std::endl;
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Loop over ranks/gpu's on our node
///////////////////////////////////////////////////////////////////////////////////////////////////////////
#ifdef SHM_SOCKETS
UnixSockets::Open(WorldShmRank);
#endif
for(int r=0;r<WorldShmSize;r++){
MPI_Barrier(WorldShmComm);
#ifndef GRID_MPI3_SHM_NONE
//////////////////////////////////////////////////
// If it is me, pass around the IPC access key
@ -536,24 +602,32 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
void * thisBuf = ShmCommBuf;
if(!Stencil_force_mpi) {
#ifdef GRID_SYCL_LEVEL_ZERO_IPC
typedef struct { int fd; pid_t pid ; } clone_mem_t;
typedef struct { int fd; pid_t pid ; ze_ipc_mem_handle_t ze; } clone_mem_t;
auto zeDevice = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_device());
auto zeContext = cl::sycl::get_native<cl::sycl::backend::level_zero>(theGridAccelerator->get_context());
auto zeDevice = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_device());
auto zeContext = cl::sycl::get_native<cl::sycl::backend::ext_oneapi_level_zero>(theGridAccelerator->get_context());
ze_ipc_mem_handle_t ihandle;
clone_mem_t handle;
if ( r==WorldShmRank ) {
auto err = zeMemGetIpcHandle(zeContext,ShmCommBuf,&ihandle);
if ( err != ZE_RESULT_SUCCESS ) {
std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
std::cerr << "SharedMemoryMPI.cc zeMemGetIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
exit(EXIT_FAILURE);
} else {
std::cout << "SharedMemoryMPI.cc zeMemGetIpcHandle succeeded for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
}
memcpy((void *)&handle.fd,(void *)&ihandle,sizeof(int));
handle.pid = getpid();
memcpy((void *)&handle.ze,(void *)&ihandle,sizeof(ihandle));
#ifdef SHM_SOCKETS
for(int rr=0;rr<WorldShmSize;rr++){
if(rr!=r){
UnixSockets::SendFileDescriptor(handle.fd,rr);
}
}
#endif
}
#endif
#ifdef GRID_CUDA
@ -581,6 +655,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
// Share this IPC handle across the Shm Comm
//////////////////////////////////////////////////
{
MPI_Barrier(WorldShmComm);
int ierr=MPI_Bcast(&handle,
sizeof(handle),
MPI_BYTE,
@ -596,6 +671,10 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
#ifdef GRID_SYCL_LEVEL_ZERO_IPC
if ( r!=WorldShmRank ) {
thisBuf = nullptr;
int myfd;
#ifdef SHM_SOCKETS
myfd=UnixSockets::RecvFileDescriptor();
#else
std::cout<<"mapping seeking remote pid/fd "
<<handle.pid<<"/"
<<handle.fd<<std::endl;
@ -603,16 +682,22 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
int pidfd = syscall(SYS_pidfd_open,handle.pid,0);
std::cout<<"Using IpcHandle pidfd "<<pidfd<<"\n";
// int myfd = syscall(SYS_pidfd_getfd,pidfd,handle.fd,0);
int myfd = syscall(438,pidfd,handle.fd,0);
std::cout<<"Using IpcHandle myfd "<<myfd<<"\n";
myfd = syscall(438,pidfd,handle.fd,0);
int err_t = errno;
if (myfd < 0) {
fprintf(stderr,"pidfd_getfd returned %d errno was %d\n", myfd,err_t); fflush(stderr);
perror("pidfd_getfd failed ");
assert(0);
}
#endif
std::cout<<"Using IpcHandle mapped remote pid "<<handle.pid <<" FD "<<handle.fd <<" to myfd "<<myfd<<"\n";
memcpy((void *)&ihandle,(void *)&handle.ze,sizeof(ihandle));
memcpy((void *)&ihandle,(void *)&myfd,sizeof(int));
auto err = zeMemOpenIpcHandle(zeContext,zeDevice,ihandle,0,&thisBuf);
if ( err != ZE_RESULT_SUCCESS ) {
std::cout << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
std::cerr << "SharedMemoryMPI.cc "<<zeContext<<" "<<zeDevice<<std::endl;
std::cerr << "SharedMemoryMPI.cc zeMemOpenIpcHandle failed for rank "<<r<<" "<<std::hex<<err<<std::dec<<std::endl;
exit(EXIT_FAILURE);
} else {
std::cout << "SharedMemoryMPI.cc zeMemOpenIpcHandle succeeded for rank "<<r<<std::endl;
@ -647,6 +732,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
#else
WorldShmCommBufs[r] = ShmCommBuf;
#endif
MPI_Barrier(WorldShmComm);
}
_ShmAllocBytes=bytes;
@ -658,7 +744,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
#ifdef GRID_MPI3_SHMMMAP
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -695,7 +781,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
assert(((uint64_t)ptr&0x3F)==0);
close(fd);
WorldShmCommBufs[r] =ptr;
// std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
// std::cout << Mheader "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
}
_ShmAlloc=1;
_ShmAllocBytes = bytes;
@ -705,7 +791,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
#ifdef GRID_MPI3_SHM_NONE
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -752,7 +838,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
////////////////////////////////////////////////////////////////////////////////////////////
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
std::cout << Mheader "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
MPI_Barrier(WorldShmComm);

6
Grid/communicator/SharedMemoryNone.cc
View File

@ -48,9 +48,10 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
_ShmSetup=1;
}
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
{
optimal_comm = WorldComm;
SHM = Coordinate(processors.size(),1);
}
////////////////////////////////////////////////////////////////////////////////////////////
@ -67,7 +68,8 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
ShmCommBuf = acceleratorAllocDevice(bytes);
ShmCommBuf = acceleratorAllocShared(bytes);
//ShmCommBuf = acceleratorAllocDevice(bytes);
if (ShmCommBuf == (void *)NULL ) {
std::cerr << " SharedMemoryNone.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;

31
Grid/cshift/Cshift_common.h
View File

@ -29,8 +29,27 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
NAMESPACE_BEGIN(Grid);
extern Vector<std::pair<int,int> > Cshift_table;
extern std::vector<std::pair<int,int> > Cshift_table;
extern commVector<std::pair<int,int> > Cshift_table_device;
inline std::pair<int,int> *MapCshiftTable(void)
{
// GPU version
#ifdef ACCELERATOR_CSHIFT
uint64_t sz=Cshift_table.size();
if (Cshift_table_device.size()!=sz ) {
Cshift_table_device.resize(sz);
}
acceleratorCopyToDevice((void *)&Cshift_table[0],
(void *)&Cshift_table_device[0],
sizeof(Cshift_table[0])*sz);
return &Cshift_table_device[0];
#else
return &Cshift_table[0];
#endif
// CPU version use identify map
}
///////////////////////////////////////////////////////////////////
// Gather for when there is no need to SIMD split
///////////////////////////////////////////////////////////////////
@ -74,8 +93,8 @@ Gather_plane_simple (const Lattice<vobj> &rhs,cshiftVector<vobj> &buffer,int dim
}
{
auto buffer_p = & buffer[0];
auto table = &Cshift_table[0];
#ifdef ACCELERATOR_CSHIFT
auto table = MapCshiftTable();
#ifdef ACCELERATOR_CSHIFT
autoView(rhs_v , rhs, AcceleratorRead);
accelerator_for(i,ent,vobj::Nsimd(),{
coalescedWrite(buffer_p[table[i].first],coalescedRead(rhs_v[table[i].second]));
@ -225,7 +244,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,cshiftVector<
{
auto buffer_p = & buffer[0];
auto table = &Cshift_table[0];
auto table = MapCshiftTable();
#ifdef ACCELERATOR_CSHIFT
autoView( rhs_v, rhs, AcceleratorWrite);
accelerator_for(i,ent,vobj::Nsimd(),{
@ -340,7 +359,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
}
{
auto table = &Cshift_table[0];
auto table = MapCshiftTable();
#ifdef ACCELERATOR_CSHIFT
autoView(rhs_v , rhs, AcceleratorRead);
autoView(lhs_v , lhs, AcceleratorWrite);
@ -392,7 +411,7 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
}
{
auto table = &Cshift_table[0];
auto table = MapCshiftTable();
#ifdef ACCELERATOR_CSHIFT
autoView( rhs_v, rhs, AcceleratorRead);
autoView( lhs_v, lhs, AcceleratorWrite);

111
Grid/cshift/Cshift_mpi.h
View File

@ -52,7 +52,8 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
int comm_dim = rhs.Grid()->_processors[dimension] >1 ;
int splice_dim = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim);
RealD t1,t0;
t0=usecond();
if ( !comm_dim ) {
//std::cout << "CSHIFT: Cshift_local" <<std::endl;
Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
@ -63,6 +64,8 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
//std::cout << "CSHIFT: Cshift_comms" <<std::endl;
Cshift_comms(ret,rhs,dimension,shift);
}
t1=usecond();
// std::cout << GridLogPerformance << "Cshift took "<< (t1-t0)/1e3 << " ms"<<std::endl;
return ret;
}
@ -127,16 +130,20 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
int cb= (cbmask==0x2)? Odd : Even;
int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
RealD tcopy=0.0;
RealD tgather=0.0;
RealD tscatter=0.0;
RealD tcomms=0.0;
uint64_t xbytes=0;
for(int x=0;x<rd;x++){
int sx = (x+sshift)%rd;
int comm_proc = ((x+sshift)/rd)%pd;
if (comm_proc==0) {
tcopy-=usecond();
Copy_plane(ret,rhs,dimension,x,sx,cbmask);
tcopy+=usecond();
} else {
int words = buffer_size;
@ -144,26 +151,39 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
int bytes = words * sizeof(vobj);
tgather-=usecond();
Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
tgather+=usecond();
// int rank = grid->_processor;
int recv_from_rank;
int xmit_to_rank;
grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
grid->Barrier();
tcomms-=usecond();
// grid->Barrier();
grid->SendToRecvFrom((void *)&send_buf[0],
xmit_to_rank,
(void *)&recv_buf[0],
recv_from_rank,
bytes);
xbytes+=bytes;
// grid->Barrier();
tcomms+=usecond();
grid->Barrier();
tscatter-=usecond();
Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
tscatter+=usecond();
}
}
/*
std::cout << GridLogPerformance << " Cshift copy "<<tcopy/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift gather "<<tgather/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift comm "<<tcomms/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift BW "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
*/
}
template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@ -190,6 +210,12 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
assert(shift>=0);
assert(shift<fd);
RealD tcopy=0.0;
RealD tgather=0.0;
RealD tscatter=0.0;
RealD tcomms=0.0;
uint64_t xbytes=0;
int permute_type=grid->PermuteType(dimension);
///////////////////////////////////////////////
@ -227,7 +253,9 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
pointers[i] = &send_buf_extract[i][0];
}
int sx = (x+sshift)%rd;
tgather-=usecond();
Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
tgather+=usecond();
for(int i=0;i<Nsimd;i++){
@ -252,7 +280,8 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
if(nbr_proc){
grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
grid->Barrier();
tcomms-=usecond();
// grid->Barrier();
send_buf_extract_mpi = &send_buf_extract[nbr_lane][0];
recv_buf_extract_mpi = &recv_buf_extract[i][0];
@ -262,7 +291,9 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
recv_from_rank,
bytes);
grid->Barrier();
xbytes+=bytes;
// grid->Barrier();
tcomms+=usecond();
rpointers[i] = &recv_buf_extract[i][0];
} else {
@ -270,9 +301,17 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
}
}
tscatter-=usecond();
Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
tscatter+=usecond();
}
/*
std::cout << GridLogPerformance << " Cshift (s) copy "<<tcopy/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift (s) gather "<<tgather/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift (s) comm "<<tcomms/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift BW "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
*/
}
#else
template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@ -292,6 +331,11 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
assert(comm_dim==1);
assert(shift>=0);
assert(shift<fd);
RealD tcopy=0.0;
RealD tgather=0.0;
RealD tscatter=0.0;
RealD tcomms=0.0;
uint64_t xbytes=0;
int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
static cshiftVector<vobj> send_buf_v; send_buf_v.resize(buffer_size);
@ -315,7 +359,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
if (comm_proc==0) {
tcopy-=usecond();
Copy_plane(ret,rhs,dimension,x,sx,cbmask);
tcopy+=usecond();
} else {
@ -324,7 +370,9 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
int bytes = words * sizeof(vobj);
tgather-=usecond();
Gather_plane_simple (rhs,send_buf_v,dimension,sx,cbmask);
tgather+=usecond();
// int rank = grid->_processor;
int recv_from_rank;
@ -332,7 +380,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
grid->Barrier();
tcomms-=usecond();
// grid->Barrier();
acceleratorCopyDeviceToDevice((void *)&send_buf_v[0],(void *)&send_buf[0],bytes);
grid->SendToRecvFrom((void *)&send_buf[0],
@ -340,13 +389,24 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
(void *)&recv_buf[0],
recv_from_rank,
bytes);
xbytes+=bytes;
acceleratorCopyDeviceToDevice((void *)&recv_buf[0],(void *)&recv_buf_v[0],bytes);
grid->Barrier();
// grid->Barrier();
tcomms+=usecond();
tscatter-=usecond();
Scatter_plane_simple (ret,recv_buf_v,dimension,x,cbmask);
tscatter+=usecond();
}
}
/*
std::cout << GridLogPerformance << " Cshift copy "<<tcopy/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift gather "<<tgather/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift scatter "<<tscatter/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift comm "<<tcomms/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift BW "<<(2.0*xbytes)/tcomms<<" MB/s "<<2*xbytes<< " Bytes "<<std::endl;
*/
}
template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
@ -372,6 +432,11 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
assert(simd_layout==2);
assert(shift>=0);
assert(shift<fd);
RealD tcopy=0.0;
RealD tgather=0.0;
RealD tscatter=0.0;
RealD tcomms=0.0;
uint64_t xbytes=0;
int permute_type=grid->PermuteType(dimension);
@ -414,8 +479,10 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
for(int i=0;i<Nsimd;i++){
pointers[i] = &send_buf_extract[i][0];
}
tgather-=usecond();
int sx = (x+sshift)%rd;
Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
tgather+=usecond();
for(int i=0;i<Nsimd;i++){
@ -440,7 +507,8 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
if(nbr_proc){
grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
grid->Barrier();
tcomms-=usecond();
// grid->Barrier();
acceleratorCopyDeviceToDevice((void *)&send_buf_extract[nbr_lane][0],(void *)send_buf_extract_mpi,bytes);
grid->SendToRecvFrom((void *)send_buf_extract_mpi,
@ -449,17 +517,28 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
recv_from_rank,
bytes);
acceleratorCopyDeviceToDevice((void *)recv_buf_extract_mpi,(void *)&recv_buf_extract[i][0],bytes);
xbytes+=bytes;
grid->Barrier();
// grid->Barrier();
tcomms+=usecond();
rpointers[i] = &recv_buf_extract[i][0];
} else {
rpointers[i] = &send_buf_extract[nbr_lane][0];
}
}
tscatter-=usecond();
Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
}
tscatter+=usecond();
}
/*
std::cout << GridLogPerformance << " Cshift (s) copy "<<tcopy/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift (s) gather "<<tgather/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift (s) scatter "<<tscatter/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift (s) comm "<<tcomms/1e3<<" ms"<<std::endl;
std::cout << GridLogPerformance << " Cshift BW "<<(2.0*xbytes)/tcomms<<" MB/s"<<std::endl;
*/
}
#endif
NAMESPACE_END(Grid);

3
Grid/cshift/Cshift_table.cc
View File

@ -1,4 +1,5 @@
#include <Grid/GridCore.h>
NAMESPACE_BEGIN(Grid);
Vector<std::pair<int,int> > Cshift_table;
std::vector<std::pair<int,int> > Cshift_table;
commVector<std::pair<int,int> > Cshift_table_device;
NAMESPACE_END(Grid);

1
Grid/lattice/Lattice.h
View File

@ -47,3 +47,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/lattice/Lattice_transfer.h>
#include <Grid/lattice/Lattice_basis.h>
#include <Grid/lattice/Lattice_crc.h>
#include <Grid/lattice/PaddedCell.h>

6
Grid/lattice/Lattice_ET.h
View File

@ -63,7 +63,7 @@ accelerator_inline vobj predicatedWhere(const iobj &predicate,
typename std::remove_const<vobj>::type ret;
typedef typename vobj::scalar_object scalar_object;
typedef typename vobj::scalar_type scalar_type;
// typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
const int Nsimd = vobj::vector_type::Nsimd();
@ -345,7 +345,9 @@ GridUnopClass(UnaryNot, Not(a));
GridUnopClass(UnaryTrace, trace(a));
GridUnopClass(UnaryTranspose, transpose(a));
GridUnopClass(UnaryTa, Ta(a));
GridUnopClass(UnarySpTa, SpTa(a));
GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
GridUnopClass(UnaryProjectOnSpGroup, ProjectOnSpGroup(a));
GridUnopClass(UnaryTimesI, timesI(a));
GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
GridUnopClass(UnaryAbs, abs(a));
@ -456,7 +458,9 @@ GRID_DEF_UNOP(operator!, UnaryNot);
GRID_DEF_UNOP(trace, UnaryTrace);
GRID_DEF_UNOP(transpose, UnaryTranspose);
GRID_DEF_UNOP(Ta, UnaryTa);
GRID_DEF_UNOP(SpTa, UnarySpTa);
GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
GRID_DEF_UNOP(ProjectOnSpGroup, UnaryProjectOnSpGroup);
GRID_DEF_UNOP(timesI, UnaryTimesI);
GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
GRID_DEF_UNOP(abs, UnaryAbs); // abs overloaded in cmath C++98; DON'T do the

53
Grid/lattice/Lattice_arith.h
View File

@ -36,6 +36,7 @@ NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
GRID_TRACE("mult");
ret.Checkerboard() = lhs.Checkerboard();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
@ -53,6 +54,7 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
template<class obj1,class obj2,class obj3> inline
void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
GRID_TRACE("mac");
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
@ -70,6 +72,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
template<class obj1,class obj2,class obj3> inline
void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
GRID_TRACE("sub");
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
@ -86,6 +89,7 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
}
template<class obj1,class obj2,class obj3> inline
void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
GRID_TRACE("add");
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
@ -106,6 +110,7 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
GRID_TRACE("mult");
ret.Checkerboard() = lhs.Checkerboard();
conformable(lhs,ret);
autoView( ret_v , ret, AcceleratorWrite);
@ -119,6 +124,7 @@ void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
template<class obj1,class obj2,class obj3> inline
void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
GRID_TRACE("mac");
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,lhs);
autoView( ret_v , ret, AcceleratorWrite);
@ -133,6 +139,7 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
template<class obj1,class obj2,class obj3> inline
void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
GRID_TRACE("sub");
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,lhs);
autoView( ret_v , ret, AcceleratorWrite);
@ -146,6 +153,7 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
}
template<class obj1,class obj2,class obj3> inline
void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
GRID_TRACE("add");
ret.Checkerboard() = lhs.Checkerboard();
conformable(lhs,ret);
autoView( ret_v , ret, AcceleratorWrite);
@ -163,6 +171,7 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
GRID_TRACE("mult");
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
autoView( ret_v , ret, AcceleratorWrite);
@ -177,6 +186,7 @@ void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
template<class obj1,class obj2,class obj3> inline
void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
GRID_TRACE("mac");
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
autoView( ret_v , ret, AcceleratorWrite);
@ -191,6 +201,7 @@ void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
template<class obj1,class obj2,class obj3> inline
void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
GRID_TRACE("sub");
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
autoView( ret_v , ret, AcceleratorWrite);
@ -204,6 +215,7 @@ void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
}
template<class obj1,class obj2,class obj3> inline
void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
GRID_TRACE("add");
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
autoView( ret_v , ret, AcceleratorWrite);
@ -218,6 +230,7 @@ void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
template<class sobj,class vobj> inline
void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
GRID_TRACE("axpy");
ret.Checkerboard() = x.Checkerboard();
conformable(ret,x);
conformable(x,y);
@ -231,6 +244,7 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
}
template<class sobj,class vobj> inline
void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
GRID_TRACE("axpby");
ret.Checkerboard() = x.Checkerboard();
conformable(ret,x);
conformable(x,y);
@ -246,13 +260,52 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
template<class sobj,class vobj> inline
RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
{
GRID_TRACE("axpy_norm");
return axpy_norm_fast(ret,a,x,y);
}
template<class sobj,class vobj> inline
RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
{
GRID_TRACE("axpby_norm");
return axpby_norm_fast(ret,a,b,x,y);
}
/// Trace product
template<class obj> auto traceProduct(const Lattice<obj> &rhs_1,const Lattice<obj> &rhs_2)
-> Lattice<decltype(trace(obj()))>
{
typedef decltype(trace(obj())) robj;
Lattice<robj> ret_i(rhs_1.Grid());
autoView( rhs1 , rhs_1, AcceleratorRead);
autoView( rhs2 , rhs_2, AcceleratorRead);
autoView( ret , ret_i, AcceleratorWrite);
ret.Checkerboard() = rhs_1.Checkerboard();
accelerator_for(ss,rhs1.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2(ss)));
});
return ret_i;
}
template<class obj1,class obj2> auto traceProduct(const Lattice<obj1> &rhs_1,const obj2 &rhs2)
-> Lattice<decltype(trace(obj1()))>
{
typedef decltype(trace(obj1())) robj;
Lattice<robj> ret_i(rhs_1.Grid());
autoView( rhs1 , rhs_1, AcceleratorRead);
autoView( ret , ret_i, AcceleratorWrite);
ret.Checkerboard() = rhs_1.Checkerboard();
accelerator_for(ss,rhs1.size(),obj1::Nsimd(),{
coalescedWrite(ret[ss],traceProduct(rhs1(ss),rhs2));
});
return ret_i;
}
template<class obj1,class obj2> auto traceProduct(const obj2 &rhs_2,const Lattice<obj1> &rhs_1)
-> Lattice<decltype(trace(obj1()))>
{
return traceProduct(rhs_1,rhs_2);
}
NAMESPACE_END(Grid);
#endif

15
Grid/lattice/Lattice_base.h
View File

@ -117,6 +117,7 @@ public:
////////////////////////////////////////////////////////////////////////////////
template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
{
GRID_TRACE("ExpressionTemplateEval");
GridBase *egrid(nullptr);
GridFromExpression(egrid,expr);
assert(egrid!=nullptr);
@ -129,7 +130,7 @@ public:
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWrite);
auto me = View(AcceleratorWriteDiscard);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
auto tmp = eval(ss,exprCopy);
coalescedWrite(me[ss],tmp);
@ -140,6 +141,7 @@ public:
}
template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
{
GRID_TRACE("ExpressionTemplateEval");
GridBase *egrid(nullptr);
GridFromExpression(egrid,expr);
assert(egrid!=nullptr);
@ -152,7 +154,7 @@ public:
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWrite);
auto me = View(AcceleratorWriteDiscard);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
auto tmp = eval(ss,exprCopy);
coalescedWrite(me[ss],tmp);
@ -163,6 +165,7 @@ public:
}
template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
{
GRID_TRACE("ExpressionTemplateEval");
GridBase *egrid(nullptr);
GridFromExpression(egrid,expr);
assert(egrid!=nullptr);
@ -174,7 +177,7 @@ public:
this->checkerboard=cb;
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWrite);
auto me = View(AcceleratorWriteDiscard);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
auto tmp = eval(ss,exprCopy);
coalescedWrite(me[ss],tmp);
@ -245,7 +248,7 @@ public:
///////////////////////////////////////////
// user defined constructor
///////////////////////////////////////////
Lattice(GridBase *grid,ViewMode mode=AcceleratorWrite) {
Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) {
this->_grid = grid;
resize(this->_grid->oSites());
assert((((uint64_t)&this->_odata[0])&0xF) ==0);
@ -288,8 +291,8 @@ public:
typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
conformable(*this,r);
this->checkerboard = r.Checkerboard();
auto me = View(AcceleratorWrite);
auto him= r.View(AcceleratorRead);
auto me = View(AcceleratorWriteDiscard);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
coalescedWrite(me[ss],him(ss));
});
@ -303,8 +306,8 @@ public:
inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
this->checkerboard = r.Checkerboard();
conformable(*this,r);
auto me = View(AcceleratorWrite);
auto him= r.View(AcceleratorRead);
auto me = View(AcceleratorWriteDiscard);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
coalescedWrite(me[ss],him(ss));
});

3
Grid/lattice/Lattice_matrix_reduction.h
View File

@ -32,7 +32,6 @@ 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];
@ -82,7 +81,6 @@ template<class vobj>
static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@ -130,7 +128,6 @@ template<class vobj>
static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
GridBase *FullGrid = lhs.Grid();

20
Grid/lattice/Lattice_peekpoke.h
View File

@ -96,9 +96,6 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
GridBase *grid=l.Grid();
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
@ -125,14 +122,17 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
//////////////////////////////////////////////////////////
// Peek a scalar object from the SIMD array
//////////////////////////////////////////////////////////
template<class vobj>
typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){
typename vobj::scalar_object s;
peekSite(s,l,site);
return s;
}
template<class vobj,class sobj>
void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
GridBase *grid=l.Grid();
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
@ -173,11 +173,11 @@ inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
idx= grid->iIndex(site);
odx= grid->oIndex(site);
scalar_type * vp = (scalar_type *)&l[odx];
const vector_type *vp = (const vector_type *) &l[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
pt[w] = vp[idx+w*Nsimd];
pt[w] = getlane(vp[w],idx);
}
return;
@ -210,10 +210,10 @@ inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
idx= grid->iIndex(site);
odx= grid->oIndex(site);
scalar_type * vp = (scalar_type *)&l[odx];
vector_type * vp = (vector_type *)&l[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
vp[idx+w*Nsimd] = pt[w];
putlane(vp[w],pt[w],idx);
}
return;
};

106
Grid/lattice/Lattice_reduction.h
View File

@ -94,10 +94,7 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
for(int i=0;i<nthread;i++){
ssum = ssum+sumarray[i];
}
typedef typename vobj::scalar_object ssobj;
ssobj ret = ssum;
return ret;
return ssum;
}
/*
Threaded max, don't use for now
@ -156,33 +153,44 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
}
template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg)
{
Integer osites = arg.Grid()->oSites();
#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
typename vobj::scalar_object ssum;
autoView( arg_v, arg, AcceleratorRead);
ssum= sum_gpu(&arg_v[0],osites);
return sum_gpu(&arg_v[0],osites);
#else
autoView(arg_v, arg, CpuRead);
auto ssum= sum_cpu(&arg_v[0],osites);
return sum_cpu(&arg_v[0],osites);
#endif
}
template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
{
auto ssum = rankSum(arg);
arg.Grid()->GlobalSum(ssum);
return ssum;
}
template<class vobj>
inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg)
{
#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
autoView( arg_v, arg, AcceleratorRead);
Integer osites = arg.Grid()->oSites();
auto ssum= sum_gpu_large(&arg_v[0],osites);
return sum_gpu_large(&arg_v[0],osites);
#else
autoView(arg_v, arg, CpuRead);
Integer osites = arg.Grid()->oSites();
auto ssum= sum_cpu(&arg_v[0],osites);
return sum_cpu(&arg_v[0],osites);
#endif
}
template<class vobj>
inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
{
auto ssum = rankSumLarge(arg);
arg.Grid()->GlobalSum(ssum);
return ssum;
}
@ -225,7 +233,6 @@ template<class vobj> inline RealD maxLocalNorm2(const Lattice<vobj> &arg)
template<class vobj>
inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
{
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_typeD vector_type;
ComplexD nrm;
@ -235,6 +242,7 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
const uint64_t sites = grid->oSites();
// Might make all code paths go this way.
#if 0
typedef decltype(innerProductD(vobj(),vobj())) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
@ -244,14 +252,37 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
// This code could read coalesce
// GPU - SIMT lane compliance...
accelerator_for( ss, sites, 1,{
auto x_l = left_v[ss];
auto y_l = right_v[ss];
inner_tmp_v[ss]=innerProductD(x_l,y_l);
auto x_l = left_v[ss];
auto y_l = right_v[ss];
inner_tmp_v[ss]=innerProductD(x_l,y_l);
});
}
#else
typedef decltype(innerProduct(vobj(),vobj())) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
{
autoView( left_v , left, AcceleratorRead);
autoView( right_v,right, AcceleratorRead);
// GPU - SIMT lane compliance...
//accelerator_for( ss, sites, nsimd,{
// auto x_l = left_v(ss);
// auto y_l = right_v(ss);
// coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
//});
#pragma omp target map ( to:left_v, right_v ) map ( tofrom:inner_tmp_v )
#pragma omp teams distribute parallel for thread_limit(THREAD_LIMIT) //nowait
for ( uint64_t ss=0;ss<sites;ss++) {
auto x_l = left_v[ss];
auto y_l = right_v[ss];
coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
}
}
#endif
// This is in single precision and fails some tests
auto anrm = sum(inner_tmp_v,sites);
auto anrm = sumD(inner_tmp_v,sites);
nrm = anrm;
return nrm;
}
@ -284,8 +315,7 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
conformable(z,x);
conformable(x,y);
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_typeD vector_type;
// typedef typename vobj::vector_typeD vector_type;
RealD nrm;
GridBase *grid = x.Grid();
@ -297,17 +327,29 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
autoView( x_v, x, AcceleratorRead);
autoView( y_v, y, AcceleratorRead);
autoView( z_v, z, AcceleratorWrite);
#if 0
typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
accelerator_for( ss, sites, 1,{
auto tmp = a*x_v[ss]+b*y_v[ss];
inner_tmp_v[ss]=innerProductD(tmp,tmp);
z_v[ss]=tmp;
accelerator_for( ss, sites, nsimd,{
auto tmp = a*x_v(ss)+b*y_v(ss);
coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
coalescedWrite(z_v[ss],tmp);
});
nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
#else
typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
accelerator_for( ss, sites, nsimd,{
auto tmp = a*x_v(ss)+b*y_v(ss);
coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
coalescedWrite(z_v[ss],tmp);
});
nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
#endif
grid->GlobalSum(nrm);
return nrm;
}
@ -317,7 +359,6 @@ innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Latti
{
conformable(left,right);
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_typeD vector_type;
Vector<ComplexD> tmp(2);
@ -461,6 +502,14 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
int words = fd*sizeof(sobj)/sizeof(scalar_type);
grid->GlobalSumVector(ptr, words);
}
template<class vobj> inline
std::vector<typename vobj::scalar_object>
sliceSum(const Lattice<vobj> &Data,int orthogdim)
{
std::vector<typename vobj::scalar_object> result;
sliceSum(Data,result,orthogdim);
return result;
}
template<class vobj>
static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim)
@ -565,7 +614,8 @@ static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Ortho
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)
{
{
// perhaps easier to just promote A to a field and use regular madd
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
@ -596,8 +646,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
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;
av.putlane(scalar_type(a[ldx])*zscale,l);
}
tensor_reduced at; at=av;
@ -637,7 +686,6 @@ 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];
@ -691,7 +739,6 @@ template<class vobj>
static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X.Grid()->GlobalDimensions()[Orthog];
@ -745,7 +792,6 @@ template<class vobj>
static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
GridBase *FullGrid = lhs.Grid();

29
Grid/lattice/Lattice_reduction_gpu.h
View File

@ -30,9 +30,12 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
cudaGetDevice(&device);
#endif
#ifdef GRID_HIP
hipGetDevice(&device);
auto r=hipGetDevice(&device);
#endif
#ifdef GRID_OMPTARGET
device = omp_get_device_num();
#endif
Iterator warpSize = gpu_props[device].warpSize;
Iterator sharedMemPerBlock = gpu_props[device].sharedMemPerBlock;
Iterator maxThreadsPerBlock = gpu_props[device].maxThreadsPerBlock;
@ -211,13 +214,25 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
assert(ok);
Integer smemSize = numThreads * sizeof(sobj);
// Move out of UVM
// Turns out I had messed up the synchronise after move to compute stream
// as running this on the default stream fools the synchronise
#undef UVM_BLOCK_BUFFER
#ifndef UVM_BLOCK_BUFFER
commVector<sobj> buffer(numBlocks);
sobj *buffer_v = &buffer[0];
sobj result;
reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
accelerator_barrier();
acceleratorCopyFromDevice(buffer_v,&result,sizeof(result));
#else
Vector<sobj> buffer(numBlocks);
sobj *buffer_v = &buffer[0];
reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
sobj result;
reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
accelerator_barrier();
auto result = buffer_v[0];
result = *buffer_v;
#endif
return result;
}
@ -250,8 +265,6 @@ inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osi
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
{
typedef typename vobj::vector_type vector;
typedef typename vobj::scalar_typeD scalarD;
typedef typename vobj::scalar_objectD sobj;
sobj ret;

25
Grid/lattice/Lattice_rng.h
View File

@ -424,9 +424,33 @@ public:
// MT implementation does not implement fast discard even though
// in principle this is possible
////////////////////////////////////////////////
#if 1
thread_for( lidx, _grid->lSites(), {
int gidx;
int o_idx;
int i_idx;
int rank;
Coordinate pcoor;
Coordinate lcoor;
Coordinate gcoor;
_grid->LocalIndexToLocalCoor(lidx,lcoor);
pcoor=_grid->ThisProcessorCoor();
_grid->ProcessorCoorLocalCoorToGlobalCoor(pcoor,lcoor,gcoor);
_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
assert(rank == _grid->ThisRank() );
int l_idx=generator_idx(o_idx,i_idx);
_generators[l_idx] = master_engine;
Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
});
#else
// Everybody loops over global volume.
thread_for( gidx, _grid->_gsites, {
// Where is it?
int rank;
int o_idx;
@ -443,6 +467,7 @@ public:
Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
}
});
#endif
#else
////////////////////////////////////////////////////////////////
// Machine and thread decomposition dependent seeding is efficient

59
Grid/lattice/Lattice_trace.h
View File

@ -66,6 +66,65 @@ inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<
return ret;
};
template<int N, class Vec>
Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
{
GridBase *grid=Umu.Grid();
auto lvol = grid->lSites();
Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid);
typedef typename Vec::scalar_type scalar;
autoView(Umu_v,Umu,CpuRead);
autoView(ret_v,ret,CpuWrite);
thread_for(site,lvol,{
Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
Coordinate lcoor;
grid->LocalIndexToLocalCoor(site, lcoor);
iScalar<iScalar<iMatrix<scalar, N> > > Us;
peekLocalSite(Us, Umu_v, lcoor);
for(int i=0;i<N;i++){
for(int j=0;j<N;j++){
scalar tmp= Us()()(i,j);
ComplexD ztmp(real(tmp),imag(tmp));
EigenU(i,j)=ztmp;
}}
ComplexD detD = EigenU.determinant();
typename Vec::scalar_type det(detD.real(),detD.imag());
pokeLocalSite(det,ret_v,lcoor);
});
return ret;
}
template<int N>
Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
{
GridBase *grid=Umu.Grid();
auto lvol = grid->lSites();
Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid);
autoView(Umu_v,Umu,CpuRead);
autoView(ret_v,ret,CpuWrite);
thread_for(site,lvol,{
Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
Coordinate lcoor;
grid->LocalIndexToLocalCoor(site, lcoor);
iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
iScalar<iScalar<iMatrix<ComplexD, N> > > Ui;
peekLocalSite(Us, Umu_v, lcoor);
for(int i=0;i<N;i++){
for(int j=0;j<N;j++){
EigenU(i,j) = Us()()(i,j);
}}
Eigen::MatrixXcd EigenUinv = EigenU.inverse();
for(int i=0;i<N;i++){
for(int j=0;j<N;j++){
Ui()()(i,j) = EigenUinv(i,j);
}}
pokeLocalSite(Ui,ret_v,lcoor);
});
return ret;
}
NAMESPACE_END(Grid);
#endif

354
Grid/lattice/Lattice_transfer.h
View File

@ -194,11 +194,11 @@ accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
#endif
accelerator_inline void convertType(vComplexF & out, const vComplexD2 & in) {
out.v = Optimization::PrecisionChange::DtoS(in._internal[0].v,in._internal[1].v);
precisionChange(out,in);
}
accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
precisionChange(out,in);
}
template<typename T1,typename T2>
@ -288,7 +288,36 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed);
}
}
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
const std::vector<Lattice<vobj>> &fineData,
const VLattice &Basis)
{
int NBatch = fineData.size();
assert(coarseData.size() == NBatch);
GridBase * fine = fineData[0].Grid();
GridBase * coarse= coarseData[0].Grid();
Lattice<iScalar<CComplex>> ip(coarse);
std::vector<Lattice<vobj>> fineDataCopy = fineData;
autoView(ip_, ip, AcceleratorWrite);
for(int v=0;v<nbasis;v++) {
for (int k=0; k<NBatch; k++) {
autoView( coarseData_ , coarseData[k], AcceleratorWrite);
blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // ip = <basis|fine>
accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
convertType(coarseData_[sc](v),ip_[sc]);
});
// improve numerical stability of projection
// |fine> = |fine> - <basis|fine> |basis>
ip=-ip;
blockZAXPY(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]);
}
}
}
template<class vobj,class vobj2,class CComplex>
inline void blockZAXPY(Lattice<vobj> &fineZ,
@ -590,6 +619,26 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
}
#endif
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
std::vector<Lattice<vobj>> &fineData,
const VLattice &Basis)
{
int NBatch = coarseData.size();
assert(fineData.size() == NBatch);
GridBase * fine = fineData[0].Grid();
GridBase * coarse = coarseData[0].Grid();
for (int k=0; k<NBatch; k++)
fineData[k]=Zero();
for (int i=0;i<nbasis;i++) {
for (int k=0; k<NBatch; k++) {
Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i);
blockZAXPY(fineData[k],ip,Basis[i],fineData[k]);
}
}
}
// Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars.
// Simd layouts need not match since we use peek/poke Local
template<class vobj,class vvobj>
@ -648,8 +697,68 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
for(int d=0;d<nd;d++){
assert(Fg->_processors[d] == Tg->_processors[d]);
}
// the above should guarantee that the operations are local
#if 1
size_t nsite = 1;
for(int i=0;i<nd;i++) nsite *= RegionSize[i];
size_t tbytes = 4*nsite*sizeof(int);
int *table = (int*)malloc(tbytes);
thread_for(idx, nsite, {
Coordinate from_coor, to_coor;
size_t rem = idx;
for(int i=0;i<nd;i++){
size_t base_i = rem % RegionSize[i]; rem /= RegionSize[i];
from_coor[i] = base_i + FromLowerLeft[i];
to_coor[i] = base_i + ToLowerLeft[i];
}
int foidx = Fg->oIndex(from_coor);
int fiidx = Fg->iIndex(from_coor);
int toidx = Tg->oIndex(to_coor);
int tiidx = Tg->iIndex(to_coor);
int* tt = table + 4*idx;
tt[0] = foidx;
tt[1] = fiidx;
tt[2] = toidx;
tt[3] = tiidx;
});
int* table_d = (int*)acceleratorAllocDevice(tbytes);
acceleratorCopyToDevice(table,table_d,tbytes);
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
autoView(from_v,From,AcceleratorRead);
autoView(to_v,To,AcceleratorWrite);
accelerator_for(idx,nsite,1,{
static const int words=sizeof(vobj)/sizeof(vector_type);
int* tt = table_d + 4*idx;
int from_oidx = *tt++;
int from_lane = *tt++;
int to_oidx = *tt++;
int to_lane = *tt;
const vector_type* from = (const vector_type *)&from_v[from_oidx];
vector_type* to = (vector_type *)&to_v[to_oidx];
scalar_type stmp;
for(int w=0;w<words;w++){
stmp = getlane(from[w], from_lane);
putlane(to[w], stmp, to_lane);
}
});
acceleratorFreeDevice(table_d);
free(table);
#else
Coordinate ldf = Fg->_ldimensions;
Coordinate rdf = Fg->_rdimensions;
Coordinate isf = Fg->_istride;
@ -658,9 +767,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
Coordinate ist = Tg->_istride;
Coordinate ost = Tg->_ostride;
autoView( t_v , To, AcceleratorWrite);
autoView( f_v , From, AcceleratorRead);
accelerator_for(idx,Fg->lSites(),1,{
autoView( t_v , To, CpuWrite);
autoView( f_v , From, CpuRead);
thread_for(idx,Fg->lSites(),{
sobj s;
Coordinate Fcoor(nd);
Coordinate Tcoor(nd);
@ -673,17 +782,24 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
}
if (in_region) {
Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]);
Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]);
Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]);
Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]);
#if 0
Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); // inner index from
Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); // inner index to
Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); // outer index from
Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); // outer index to
scalar_type * fp = (scalar_type *)&f_v[odx_f];
scalar_type * tp = (scalar_type *)&t_v[odx_t];
for(int w=0;w<words;w++){
tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd]; // FIXME IF RRII layout, type pun no worke
tp[w].putlane(fp[w].getlane(idx_f),idx_t);
}
#else
peekLocalSite(s,f_v,Fcoor);
pokeLocalSite(s,t_v,Tcoor);
#endif
}
});
#endif
}
@ -776,6 +892,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
}
//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
template<class vobj>
void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
{
@ -792,11 +910,70 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
for(int d=0;d<nh;d++){
if ( d!=orthog ) {
assert(lg->_processors[d] == hg->_processors[d]);
assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
}
assert(lg->_processors[d] == hg->_processors[d]);
assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
}
}
#if 1
size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
size_t tbytes = 4*nsite*sizeof(int);
int *table = (int*)malloc(tbytes);
thread_for(idx,nsite,{
Coordinate lcoor(nl);
Coordinate hcoor(nh);
lcoor[orthog] = slice_lo;
hcoor[orthog] = slice_hi;
size_t rem = idx;
for(int mu=0;mu<nl;mu++){
if(mu != orthog){
int xmu = rem % lg->LocalDimensions()[mu]; rem /= lg->LocalDimensions()[mu];
lcoor[mu] = hcoor[mu] = xmu;
}
}
int loidx = lg->oIndex(lcoor);
int liidx = lg->iIndex(lcoor);
int hoidx = hg->oIndex(hcoor);
int hiidx = hg->iIndex(hcoor);
int* tt = table + 4*idx;
tt[0] = loidx;
tt[1] = liidx;
tt[2] = hoidx;
tt[3] = hiidx;
});
int* table_d = (int*)acceleratorAllocDevice(tbytes);
acceleratorCopyToDevice(table,table_d,tbytes);
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
autoView(lowDim_v,lowDim,AcceleratorRead);
autoView(higherDim_v,higherDim,AcceleratorWrite);
accelerator_for(idx,nsite,1,{
static const int words=sizeof(vobj)/sizeof(vector_type);
int* tt = table_d + 4*idx;
int from_oidx = *tt++;
int from_lane = *tt++;
int to_oidx = *tt++;
int to_lane = *tt;
const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
vector_type* to = (vector_type *)&higherDim_v[to_oidx];
scalar_type stmp;
for(int w=0;w<words;w++){
stmp = getlane(from[w], from_lane);
putlane(to[w], stmp, to_lane);
}
});
acceleratorFreeDevice(table_d);
free(table);
#else
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuRead);
autoView(higherDimv,higherDim,CpuWrite);
@ -812,6 +989,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
pokeLocalSite(s,higherDimv,hcoor);
}
});
#endif
}
@ -855,7 +1033,7 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
template<class vobj>
void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
{
typedef typename vobj::scalar_object sobj;
@ -1080,9 +1258,27 @@ vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
});
}
//Convert a Lattice from one precision to another
//Very fast precision change. Requires in/out objects to reside on same Grid (e.g. by using double2 for the double-precision field)
template<class VobjOut, class VobjIn>
void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
void precisionChangeFast(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
{
typedef typename VobjOut::vector_type Vout;
typedef typename VobjIn::vector_type Vin;
const int N = sizeof(VobjOut)/sizeof(Vout);
conformable(out.Grid(),in.Grid());
out.Checkerboard() = in.Checkerboard();
int nsimd = out.Grid()->Nsimd();
autoView( out_v , out, AcceleratorWrite);
autoView( in_v , in, AcceleratorRead);
accelerator_for(idx,out.Grid()->oSites(),1,{
Vout *vout = (Vout *)&out_v[idx];
Vin *vin = (Vin *)&in_v[idx];
precisionChange(vout,vin,N);
});
}
//Convert a Lattice from one precision to another (original, slow implementation)
template<class VobjOut, class VobjIn>
void precisionChangeOrig(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
{
assert(out.Grid()->Nd() == in.Grid()->Nd());
for(int d=0;d<out.Grid()->Nd();d++){
@ -1097,7 +1293,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
int ndim = out.Grid()->Nd();
int out_nsimd = out_grid->Nsimd();
int in_nsimd = in_grid->Nsimd();
std::vector<Coordinate > out_icoor(out_nsimd);
for(int lane=0; lane < out_nsimd; lane++){
@ -1128,6 +1324,128 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
});
}
//The workspace for a precision change operation allowing for the reuse of the mapping to save time on subsequent calls
class precisionChangeWorkspace{
std::pair<Integer,Integer>* fmap_device; //device pointer
//maintain grids for checking
GridBase* _out_grid;
GridBase* _in_grid;
public:
precisionChangeWorkspace(GridBase *out_grid, GridBase *in_grid): _out_grid(out_grid), _in_grid(in_grid){
//Build a map between the sites and lanes of the output field and the input field as we cannot use the Grids on the device
assert(out_grid->Nd() == in_grid->Nd());
for(int d=0;d<out_grid->Nd();d++){
assert(out_grid->FullDimensions()[d] == in_grid->FullDimensions()[d]);
}
int Nsimd_out = out_grid->Nsimd();
std::vector<Coordinate> out_icorrs(out_grid->Nsimd()); //reuse these
for(int lane=0; lane < out_grid->Nsimd(); lane++)
out_grid->iCoorFromIindex(out_icorrs[lane], lane);
std::vector<std::pair<Integer,Integer> > fmap_host(out_grid->lSites()); //lsites = osites*Nsimd
thread_for(out_oidx,out_grid->oSites(),{
Coordinate out_ocorr;
out_grid->oCoorFromOindex(out_ocorr, out_oidx);
Coordinate lcorr; //the local coordinate (common to both in and out as full coordinate)
for(int out_lane=0; out_lane < Nsimd_out; out_lane++){
out_grid->InOutCoorToLocalCoor(out_ocorr, out_icorrs[out_lane], lcorr);
//int in_oidx = in_grid->oIndex(lcorr), in_lane = in_grid->iIndex(lcorr);
//Note oIndex and OcorrFromOindex (and same for iIndex) are not inverse for checkerboarded lattice, the former coordinates being defined on the full lattice and the latter on the reduced lattice
//Until this is fixed we need to circumvent the problem locally. Here I will use the coordinates defined on the reduced lattice for simplicity
int in_oidx = 0, in_lane = 0;
for(int d=0;d<in_grid->_ndimension;d++){
in_oidx += in_grid->_ostride[d] * ( lcorr[d] % in_grid->_rdimensions[d] );
in_lane += in_grid->_istride[d] * ( lcorr[d] / in_grid->_rdimensions[d] );
}
fmap_host[out_lane + Nsimd_out*out_oidx] = std::pair<Integer,Integer>( in_oidx, in_lane );
}
});
//Copy the map to the device (if we had a way to tell if an accelerator is in use we could avoid this copy for CPU-only machines)
size_t fmap_bytes = out_grid->lSites() * sizeof(std::pair<Integer,Integer>);
fmap_device = (std::pair<Integer,Integer>*)acceleratorAllocDevice(fmap_bytes);
acceleratorCopyToDevice(fmap_host.data(), fmap_device, fmap_bytes);
}
//Prevent moving or copying
precisionChangeWorkspace(const precisionChangeWorkspace &r) = delete;
precisionChangeWorkspace(precisionChangeWorkspace &&r) = delete;
precisionChangeWorkspace &operator=(const precisionChangeWorkspace &r) = delete;
precisionChangeWorkspace &operator=(precisionChangeWorkspace &&r) = delete;
std::pair<Integer,Integer> const* getMap() const{ return fmap_device; }
void checkGrids(GridBase* out, GridBase* in) const{
conformable(out, _out_grid);
conformable(in, _in_grid);
}
~precisionChangeWorkspace(){
acceleratorFreeDevice(fmap_device);
}
};
//We would like to use precisionChangeFast when possible. However usage of this requires the Grids to be the same (runtime check)
//*and* the precisionChange(VobjOut::vector_type, VobjIn, int) function to be defined for the types; this requires an extra compile-time check which we do using some SFINAE trickery
template<class VobjOut, class VobjIn>
auto _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, int dummy)->decltype( precisionChange( ((typename VobjOut::vector_type*)0), ((typename VobjIn::vector_type*)0), 1), int()){
if(out.Grid() == in.Grid()){
precisionChangeFast(out,in);
return 1;
}else{
return 0;
}
}
template<class VobjOut, class VobjIn>
int _precisionChangeFastWrap(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, long dummy){ //note long here is intentional; it means the above is preferred if available
return 0;
}
//Convert a lattice of one precision to another. Much faster than original implementation but requires a pregenerated workspace
//which contains the mapping data.
template<class VobjOut, class VobjIn>
void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in, const precisionChangeWorkspace &workspace){
if(_precisionChangeFastWrap(out,in,0)) return;
static_assert( std::is_same<typename VobjOut::scalar_typeD, typename VobjIn::scalar_typeD>::value == 1, "precisionChange: tensor types must be the same" ); //if tensor types are same the DoublePrecision type must be the same
out.Checkerboard() = in.Checkerboard();
constexpr int Nsimd_out = VobjOut::Nsimd();
workspace.checkGrids(out.Grid(),in.Grid());
std::pair<Integer,Integer> const* fmap_device = workspace.getMap();
//Do the copy/precision change
autoView( out_v , out, AcceleratorWrite);
autoView( in_v , in, AcceleratorRead);
accelerator_for(out_oidx, out.Grid()->oSites(), 1,{
std::pair<Integer,Integer> const* fmap_osite = fmap_device + out_oidx*Nsimd_out;
for(int out_lane=0; out_lane < Nsimd_out; out_lane++){
int in_oidx = fmap_osite[out_lane].first;
int in_lane = fmap_osite[out_lane].second;
copyLane(out_v[out_oidx], out_lane, in_v[in_oidx], in_lane);
}
});
}
//Convert a Lattice from one precision to another. Much faster than original implementation but slower than precisionChangeFast
//or precisionChange called with pregenerated workspace, as it needs to internally generate the workspace on the host and copy to device
template<class VobjOut, class VobjIn>
void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
if(_precisionChangeFastWrap(out,in,0)) return;
precisionChangeWorkspace workspace(out.Grid(), in.Grid());
precisionChange(out, in, workspace);
}
////////////////////////////////////////////////////////////////////////////////
// Communicate between grids
////////////////////////////////////////////////////////////////////////////////

2
Grid/lattice/Lattice_view.h
View File

@ -79,7 +79,7 @@ public:
accelerator_inline uint64_t end(void) const { return this->_odata_size; };
accelerator_inline uint64_t size(void) const { return this->_odata_size; };
LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me){}
LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me){ }
LatticeView(const LatticeView<vobj> &refer_to_me) = default; // Trivially copyable
LatticeView(const LatticeAccelerator<vobj> &refer_to_me,ViewMode mode) : LatticeAccelerator<vobj> (refer_to_me)
{

174
Grid/lattice/PaddedCell.h Normal file
View File

@ -0,0 +1,174 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/lattice/PaddedCell.h
Copyright (C) 2019
Author: Peter Boyle pboyle@bnl.gov
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#pragma once
#include<Grid/cshift/Cshift.h>
NAMESPACE_BEGIN(Grid);
//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
template<typename vobj>
struct CshiftImplBase{
virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
virtual ~CshiftImplBase(){}
};
template<typename vobj>
struct CshiftImplDefault: public CshiftImplBase<vobj>{
Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
};
template<typename Gimpl>
struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
};
class PaddedCell {
public:
GridCartesian * unpadded_grid;
int dims;
int depth;
std::vector<GridCartesian *> grids;
~PaddedCell()
{
DeleteGrids();
}
PaddedCell(int _depth,GridCartesian *_grid)
{
unpadded_grid = _grid;
depth=_depth;
dims=_grid->Nd();
AllocateGrids();
Coordinate local =unpadded_grid->LocalDimensions();
for(int d=0;d<dims;d++){
assert(local[d]>=depth);
}
}
void DeleteGrids(void)
{
for(int d=0;d<grids.size();d++){
delete grids[d];
}
grids.resize(0);
};
void AllocateGrids(void)
{
Coordinate local =unpadded_grid->LocalDimensions();
Coordinate simd =unpadded_grid->_simd_layout;
Coordinate processors=unpadded_grid->_processors;
Coordinate plocal =unpadded_grid->LocalDimensions();
Coordinate global(dims);
// expand up one dim at a time
for(int d=0;d<dims;d++){
plocal[d] += 2*depth;
for(int d=0;d<dims;d++){
global[d] = plocal[d]*processors[d];
}
grids.push_back(new GridCartesian(global,simd,processors));
}
};
template<class vobj>
inline Lattice<vobj> Extract(const Lattice<vobj> &in) const
{
Lattice<vobj> out(unpadded_grid);
Coordinate local =unpadded_grid->LocalDimensions();
Coordinate fll(dims,depth); // depends on the MPI spread
Coordinate tll(dims,0); // depends on the MPI spread
localCopyRegion(in,out,fll,tll,local);
return out;
}
template<class vobj>
inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
{
GridBase *old_grid = in.Grid();
int dims = old_grid->Nd();
Lattice<vobj> tmp = in;
for(int d=0;d<dims;d++){
tmp = Expand(d,tmp,cshift); // rvalue && assignment
}
return tmp;
}
// expand up one dim at a time
template<class vobj>
inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
{
GridBase *old_grid = in.Grid();
GridCartesian *new_grid = grids[dim];//These are new grids
Lattice<vobj> padded(new_grid);
Lattice<vobj> shifted(old_grid);
Coordinate local =old_grid->LocalDimensions();
Coordinate plocal =new_grid->LocalDimensions();
if(dim==0) conformable(old_grid,unpadded_grid);
else conformable(old_grid,grids[dim-1]);
std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
double tins=0, tshift=0;
// Middle bit
double t = usecond();
for(int x=0;x<local[dim];x++){
InsertSliceLocal(in,padded,x,depth+x,dim);
}
tins += usecond() - t;
// High bit
t = usecond();
shifted = cshift.Cshift(in,dim,depth);
tshift += usecond() - t;
t=usecond();
for(int x=0;x<depth;x++){
InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
}
tins += usecond() - t;
// Low bit
t = usecond();
shifted = cshift.Cshift(in,dim,-depth);
tshift += usecond() - t;
t = usecond();
for(int x=0;x<depth;x++){
InsertSliceLocal(shifted,padded,x,x,dim);
}
tins += usecond() - t;
std::cout << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
return padded;
}
};
NAMESPACE_END(Grid);

6
Grid/parallelIO/NerscIO.h
View File

@ -42,9 +42,11 @@ using namespace Grid;
////////////////////////////////////////////////////////////////////////////////
class NerscIO : public BinaryIO {
public:
typedef Lattice<vLorentzColourMatrixD> GaugeField;
// Enable/disable exiting if the plaquette in the header does not match the value computed (default true)
static bool & exitOnReadPlaquetteMismatch(){ static bool v=true; return v; }
static inline void truncate(std::string file){
std::ofstream fout(file,std::ios::out);
}
@ -203,7 +205,7 @@ public:
std::cerr << " nersc_csum " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
exit(0);
}
assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
if(exitOnReadPlaquetteMismatch()) assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
assert(nersc_csum == header.checksum );

2
Grid/pugixml/pugixml.cc
View File

@ -16,7 +16,7 @@
#ifdef __NVCC__
#pragma push
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
#pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
#else
#pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"

116
Grid/qcd/QCD.h
View File

@ -63,6 +63,7 @@ static constexpr int Ngp=2; // gparity index range
#define ColourIndex (2)
#define SpinIndex (1)
#define LorentzIndex (0)
#define GparityFlavourIndex (0)
// Also should make these a named enum type
static constexpr int DaggerNo=0;
@ -87,6 +88,8 @@ template<typename T> struct isCoarsened {
template <typename T> using IfCoarsened = Invoke<std::enable_if< isCoarsened<T>::value,int> > ;
template <typename T> using IfNotCoarsened = Invoke<std::enable_if<!isCoarsened<T>::value,int> > ;
const int GparityFlavourTensorIndex = 3; //TensorLevel counts from the bottom!
// ChrisK very keen to add extra space for Gparity doubling.
//
// Also add domain wall index, in a way where Wilson operator
@ -101,6 +104,7 @@ template<typename vtype> using iSpinMatrix = iScalar<iMatrix<iSca
template<typename vtype> using iColourMatrix = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
template<typename vtype> using iSpinColourMatrix = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >;
template<typename vtype> using iLorentzColourMatrix = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ;
template<typename vtype> using iLorentzComplex = iVector<iScalar<iScalar<vtype> >, Nd > ;
template<typename vtype> using iDoubleStoredColourMatrix = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ;
template<typename vtype> using iSpinVector = iScalar<iVector<iScalar<vtype>, Ns> >;
template<typename vtype> using iColourVector = iScalar<iScalar<iVector<vtype, Nc> > >;
@ -110,8 +114,10 @@ template<typename vtype> using iHalfSpinColourVector = iScalar<iVector<iVec
template<typename vtype> using iSpinColourSpinColourMatrix = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >;
template<typename vtype> using iGparityFlavourVector = iVector<iScalar<iScalar<vtype> >, Ngp>;
template<typename vtype> using iGparitySpinColourVector = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >;
template<typename vtype> using iGparityHalfSpinColourVector = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >;
template<typename vtype> using iGparityFlavourMatrix = iMatrix<iScalar<iScalar<vtype> >, Ngp>;
// Spin matrix
typedef iSpinMatrix<Complex > SpinMatrix;
@ -121,6 +127,7 @@ typedef iSpinMatrix<ComplexD > SpinMatrixD;
typedef iSpinMatrix<vComplex > vSpinMatrix;
typedef iSpinMatrix<vComplexF> vSpinMatrixF;
typedef iSpinMatrix<vComplexD> vSpinMatrixD;
typedef iSpinMatrix<vComplexD2> vSpinMatrixD2;
// Colour Matrix
typedef iColourMatrix<Complex > ColourMatrix;
@ -130,6 +137,7 @@ typedef iColourMatrix<ComplexD > ColourMatrixD;
typedef iColourMatrix<vComplex > vColourMatrix;
typedef iColourMatrix<vComplexF> vColourMatrixF;
typedef iColourMatrix<vComplexD> vColourMatrixD;
typedef iColourMatrix<vComplexD2> vColourMatrixD2;
// SpinColour matrix
typedef iSpinColourMatrix<Complex > SpinColourMatrix;
@ -139,6 +147,7 @@ typedef iSpinColourMatrix<ComplexD > SpinColourMatrixD;
typedef iSpinColourMatrix<vComplex > vSpinColourMatrix;
typedef iSpinColourMatrix<vComplexF> vSpinColourMatrixF;
typedef iSpinColourMatrix<vComplexD> vSpinColourMatrixD;
typedef iSpinColourMatrix<vComplexD2> vSpinColourMatrixD2;
// SpinColourSpinColour matrix
typedef iSpinColourSpinColourMatrix<Complex > SpinColourSpinColourMatrix;
@ -148,6 +157,7 @@ typedef iSpinColourSpinColourMatrix<ComplexD > SpinColourSpinColourMatrixD;
typedef iSpinColourSpinColourMatrix<vComplex > vSpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<vComplexF> vSpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<vComplexD> vSpinColourSpinColourMatrixD;
typedef iSpinColourSpinColourMatrix<vComplexD2> vSpinColourSpinColourMatrixD2;
// SpinColourSpinColour matrix
typedef iSpinColourSpinColourMatrix<Complex > SpinColourSpinColourMatrix;
@ -157,24 +167,47 @@ typedef iSpinColourSpinColourMatrix<ComplexD > SpinColourSpinColourMatrixD;
typedef iSpinColourSpinColourMatrix<vComplex > vSpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<vComplexF> vSpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<vComplexD> vSpinColourSpinColourMatrixD;
typedef iSpinColourSpinColourMatrix<vComplexD2> vSpinColourSpinColourMatrixD2;
// LorentzColour
typedef iLorentzColourMatrix<Complex > LorentzColourMatrix;
typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF;
typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD;
typedef iLorentzColourMatrix<vComplex > vLorentzColourMatrix;
typedef iLorentzColourMatrix<vComplexF> vLorentzColourMatrixF;
typedef iLorentzColourMatrix<vComplexD> vLorentzColourMatrixD;
typedef iLorentzColourMatrix<vComplex > vLorentzColourMatrix;
typedef iLorentzColourMatrix<vComplexF> vLorentzColourMatrixF;
typedef iLorentzColourMatrix<vComplexD> vLorentzColourMatrixD;
typedef iLorentzColourMatrix<vComplexD2> vLorentzColourMatrixD2;
// LorentzComplex
typedef iLorentzComplex<Complex > LorentzComplex;
typedef iLorentzComplex<ComplexF > LorentzComplexF;
typedef iLorentzComplex<ComplexD > LorentzComplexD;
typedef iLorentzComplex<vComplex > vLorentzComplex;
typedef iLorentzComplex<vComplexF> vLorentzComplexF;
typedef iLorentzComplex<vComplexD> vLorentzComplexD;
// DoubleStored gauge field
typedef iDoubleStoredColourMatrix<Complex > DoubleStoredColourMatrix;
typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF;
typedef iDoubleStoredColourMatrix<ComplexD > DoubleStoredColourMatrixD;
typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
typedef iDoubleStoredColourMatrix<vComplexD2> vDoubleStoredColourMatrixD2;
//G-parity flavour matrix
typedef iGparityFlavourMatrix<Complex> GparityFlavourMatrix;
typedef iGparityFlavourMatrix<ComplexF> GparityFlavourMatrixF;
typedef iGparityFlavourMatrix<ComplexD> GparityFlavourMatrixD;
typedef iGparityFlavourMatrix<vComplex> vGparityFlavourMatrix;
typedef iGparityFlavourMatrix<vComplexF> vGparityFlavourMatrixF;
typedef iGparityFlavourMatrix<vComplexD> vGparityFlavourMatrixD;
typedef iGparityFlavourMatrix<vComplexD2> vGparityFlavourMatrixD2;
// Spin vector
typedef iSpinVector<Complex > SpinVector;
@ -184,6 +217,7 @@ typedef iSpinVector<ComplexD> SpinVectorD;
typedef iSpinVector<vComplex > vSpinVector;
typedef iSpinVector<vComplexF> vSpinVectorF;
typedef iSpinVector<vComplexD> vSpinVectorD;
typedef iSpinVector<vComplexD2> vSpinVectorD2;
// Colour vector
typedef iColourVector<Complex > ColourVector;
@ -193,6 +227,7 @@ typedef iColourVector<ComplexD> ColourVectorD;
typedef iColourVector<vComplex > vColourVector;
typedef iColourVector<vComplexF> vColourVectorF;
typedef iColourVector<vComplexD> vColourVectorD;
typedef iColourVector<vComplexD2> vColourVectorD2;
// SpinColourVector
typedef iSpinColourVector<Complex > SpinColourVector;
@ -202,6 +237,7 @@ typedef iSpinColourVector<ComplexD> SpinColourVectorD;
typedef iSpinColourVector<vComplex > vSpinColourVector;
typedef iSpinColourVector<vComplexF> vSpinColourVectorF;
typedef iSpinColourVector<vComplexD> vSpinColourVectorD;
typedef iSpinColourVector<vComplexD2> vSpinColourVectorD2;
// HalfSpin vector
typedef iHalfSpinVector<Complex > HalfSpinVector;
@ -211,15 +247,27 @@ typedef iHalfSpinVector<ComplexD> HalfSpinVectorD;
typedef iHalfSpinVector<vComplex > vHalfSpinVector;
typedef iHalfSpinVector<vComplexF> vHalfSpinVectorF;
typedef iHalfSpinVector<vComplexD> vHalfSpinVectorD;
typedef iHalfSpinVector<vComplexD2> vHalfSpinVectorD2;
// HalfSpinColour vector
typedef iHalfSpinColourVector<Complex > HalfSpinColourVector;
typedef iHalfSpinColourVector<ComplexF> HalfSpinColourVectorF;
typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
typedef iHalfSpinColourVector<vComplexD2> vHalfSpinColourVectorD2;
//G-parity flavour vector
typedef iGparityFlavourVector<Complex > GparityFlavourVector;
typedef iGparityFlavourVector<ComplexF> GparityFlavourVectorF;
typedef iGparityFlavourVector<ComplexD> GparityFlavourVectorD;
typedef iGparityFlavourVector<vComplex > vGparityFlavourVector;
typedef iGparityFlavourVector<vComplexF> vGparityFlavourVectorF;
typedef iGparityFlavourVector<vComplexD> vGparityFlavourVectorD;
typedef iGparityFlavourVector<vComplexD2> vGparityFlavourVectorD2;
// singlets
typedef iSinglet<Complex > TComplex; // FIXME This is painful. Tensor singlet complex type.
@ -229,6 +277,7 @@ typedef iSinglet<ComplexD> TComplexD; // FIXME This is painful. Tenso
typedef iSinglet<vComplex > vTComplex ; // what if we don't know the tensor structure
typedef iSinglet<vComplexF> vTComplexF; // what if we don't know the tensor structure
typedef iSinglet<vComplexD> vTComplexD; // what if we don't know the tensor structure
typedef iSinglet<vComplexD2> vTComplexD2; // what if we don't know the tensor structure
typedef iSinglet<Real > TReal; // Shouldn't need these; can I make it work without?
typedef iSinglet<RealF> TRealF; // Shouldn't need these; can I make it work without?
@ -246,47 +295,62 @@ typedef iSinglet<Integer > TInteger;
typedef Lattice<vColourMatrix> LatticeColourMatrix;
typedef Lattice<vColourMatrixF> LatticeColourMatrixF;
typedef Lattice<vColourMatrixD> LatticeColourMatrixD;
typedef Lattice<vColourMatrixD2> LatticeColourMatrixD2;
typedef Lattice<vSpinMatrix> LatticeSpinMatrix;
typedef Lattice<vSpinMatrixF> LatticeSpinMatrixF;
typedef Lattice<vSpinMatrixD> LatticeSpinMatrixD;
typedef Lattice<vSpinMatrixD2> LatticeSpinMatrixD2;
typedef Lattice<vSpinColourMatrix> LatticeSpinColourMatrix;
typedef Lattice<vSpinColourMatrixF> LatticeSpinColourMatrixF;
typedef Lattice<vSpinColourMatrixD> LatticeSpinColourMatrixD;
typedef Lattice<vSpinColourMatrixD2> LatticeSpinColourMatrixD2;
typedef Lattice<vSpinColourSpinColourMatrix> LatticeSpinColourSpinColourMatrix;
typedef Lattice<vSpinColourSpinColourMatrixF> LatticeSpinColourSpinColourMatrixF;
typedef Lattice<vSpinColourSpinColourMatrixD> LatticeSpinColourSpinColourMatrixD;
typedef Lattice<vSpinColourSpinColourMatrixD2> LatticeSpinColourSpinColourMatrixD2;
typedef Lattice<vLorentzColourMatrix> LatticeLorentzColourMatrix;
typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF;
typedef Lattice<vLorentzColourMatrixD> LatticeLorentzColourMatrixD;
typedef Lattice<vLorentzColourMatrix> LatticeLorentzColourMatrix;
typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF;
typedef Lattice<vLorentzColourMatrixD> LatticeLorentzColourMatrixD;
typedef Lattice<vLorentzColourMatrixD2> LatticeLorentzColourMatrixD2;
typedef Lattice<vLorentzComplex> LatticeLorentzComplex;
typedef Lattice<vLorentzComplexF> LatticeLorentzComplexF;
typedef Lattice<vLorentzComplexD> LatticeLorentzComplexD;
// DoubleStored gauge field
typedef Lattice<vDoubleStoredColourMatrix> LatticeDoubleStoredColourMatrix;
typedef Lattice<vDoubleStoredColourMatrixF> LatticeDoubleStoredColourMatrixF;
typedef Lattice<vDoubleStoredColourMatrixD> LatticeDoubleStoredColourMatrixD;
typedef Lattice<vDoubleStoredColourMatrix> LatticeDoubleStoredColourMatrix;
typedef Lattice<vDoubleStoredColourMatrixF> LatticeDoubleStoredColourMatrixF;
typedef Lattice<vDoubleStoredColourMatrixD> LatticeDoubleStoredColourMatrixD;
typedef Lattice<vDoubleStoredColourMatrixD2> LatticeDoubleStoredColourMatrixD2;
typedef Lattice<vSpinVector> LatticeSpinVector;
typedef Lattice<vSpinVectorF> LatticeSpinVectorF;
typedef Lattice<vSpinVectorD> LatticeSpinVectorD;
typedef Lattice<vSpinVectorD2> LatticeSpinVectorD2;
typedef Lattice<vColourVector> LatticeColourVector;
typedef Lattice<vColourVectorF> LatticeColourVectorF;
typedef Lattice<vColourVectorD> LatticeColourVectorD;
typedef Lattice<vColourVectorD2> LatticeColourVectorD2;
typedef Lattice<vSpinColourVector> LatticeSpinColourVector;
typedef Lattice<vSpinColourVectorF> LatticeSpinColourVectorF;
typedef Lattice<vSpinColourVectorD> LatticeSpinColourVectorD;
typedef Lattice<vSpinColourVectorD2> LatticeSpinColourVectorD2;
typedef Lattice<vHalfSpinVector> LatticeHalfSpinVector;
typedef Lattice<vHalfSpinVectorF> LatticeHalfSpinVectorF;
typedef Lattice<vHalfSpinVectorD> LatticeHalfSpinVectorD;
typedef Lattice<vHalfSpinVectorD2> LatticeHalfSpinVectorD2;
typedef Lattice<vHalfSpinColourVector> LatticeHalfSpinColourVector;
typedef Lattice<vHalfSpinColourVectorF> LatticeHalfSpinColourVectorF;
typedef Lattice<vHalfSpinColourVectorD> LatticeHalfSpinColourVectorD;
typedef Lattice<vHalfSpinColourVector> LatticeHalfSpinColourVector;
typedef Lattice<vHalfSpinColourVectorF> LatticeHalfSpinColourVectorF;
typedef Lattice<vHalfSpinColourVectorD> LatticeHalfSpinColourVectorD;
typedef Lattice<vHalfSpinColourVectorD2> LatticeHalfSpinColourVectorD2;
typedef Lattice<vTReal> LatticeReal;
typedef Lattice<vTRealF> LatticeRealF;
@ -295,6 +359,7 @@ typedef Lattice<vTRealD> LatticeRealD;
typedef Lattice<vTComplex> LatticeComplex;
typedef Lattice<vTComplexF> LatticeComplexF;
typedef Lattice<vTComplexD> LatticeComplexD;
typedef Lattice<vTComplexD2> LatticeComplexD2;
typedef Lattice<vTInteger> LatticeInteger; // Predicates for "where"
@ -302,37 +367,42 @@ typedef Lattice<vTInteger> LatticeInteger; // Predicates for "where"
///////////////////////////////////////////
// Physical names for things
///////////////////////////////////////////
typedef LatticeHalfSpinColourVector LatticeHalfFermion;
typedef LatticeHalfSpinColourVectorF LatticeHalfFermionF;
typedef LatticeHalfSpinColourVectorF LatticeHalfFermionD;
typedef LatticeHalfSpinColourVector LatticeHalfFermion;
typedef LatticeHalfSpinColourVectorF LatticeHalfFermionF;
typedef LatticeHalfSpinColourVectorD LatticeHalfFermionD;
typedef LatticeHalfSpinColourVectorD2 LatticeHalfFermionD2;
typedef LatticeSpinColourVector LatticeFermion;
typedef LatticeSpinColourVectorF LatticeFermionF;
typedef LatticeSpinColourVectorD LatticeFermionD;
typedef LatticeSpinColourVectorD2 LatticeFermionD2;
typedef LatticeSpinColourMatrix LatticePropagator;
typedef LatticeSpinColourMatrixF LatticePropagatorF;
typedef LatticeSpinColourMatrixD LatticePropagatorD;
typedef LatticeSpinColourMatrixD2 LatticePropagatorD2;
typedef LatticeLorentzColourMatrix LatticeGaugeField;
typedef LatticeLorentzColourMatrixF LatticeGaugeFieldF;
typedef LatticeLorentzColourMatrixD LatticeGaugeFieldD;
typedef LatticeLorentzColourMatrixD2 LatticeGaugeFieldD2;
typedef LatticeDoubleStoredColourMatrix LatticeDoubledGaugeField;
typedef LatticeDoubleStoredColourMatrixF LatticeDoubledGaugeFieldF;
typedef LatticeDoubleStoredColourMatrixD LatticeDoubledGaugeFieldD;
typedef LatticeDoubleStoredColourMatrixD2 LatticeDoubledGaugeFieldD2;
template<class GF> using LorentzScalar = Lattice<iScalar<typename GF::vector_object::element> >;
// Uhgg... typing this hurt ;)
// (my keyboard got burning hot when I typed this, must be the anti-Fermion)
typedef Lattice<vColourVector> LatticeStaggeredFermion;
typedef Lattice<vColourVectorF> LatticeStaggeredFermionF;
typedef Lattice<vColourVectorD> LatticeStaggeredFermionD;
typedef Lattice<vColourVectorD2> LatticeStaggeredFermionD2;
typedef Lattice<vColourMatrix> LatticeStaggeredPropagator;
typedef Lattice<vColourMatrixF> LatticeStaggeredPropagatorF;
typedef Lattice<vColourMatrixD> LatticeStaggeredPropagatorD;
typedef Lattice<vColourMatrixD2> LatticeStaggeredPropagatorD2;
//////////////////////////////////////////////////////////////////////////////
// Peek and Poke named after physics attributes

98
Grid/qcd/action/ActionBase.h
View File

@ -34,21 +34,117 @@ directory
NAMESPACE_BEGIN(Grid);
///////////////////////////////////
// Smart configuration base class
///////////////////////////////////
template< class Field >
class ConfigurationBase
{
public:
ConfigurationBase() {}
virtual ~ConfigurationBase() {}
virtual void set_Field(Field& U) =0;
virtual void smeared_force(Field&) = 0;
virtual Field& get_SmearedU() =0;
virtual Field &get_U(bool smeared = false) = 0;
};
template <class GaugeField >
class Action
{
public:
bool is_smeared = false;
RealD deriv_norm_sum;
RealD deriv_max_sum;
RealD Fdt_norm_sum;
RealD Fdt_max_sum;
int deriv_num;
RealD deriv_us;
RealD S_us;
RealD refresh_us;
void reset_timer(void) {
deriv_us = S_us = refresh_us = 0.0;
deriv_norm_sum = deriv_max_sum=0.0;
Fdt_max_sum = Fdt_norm_sum = 0.0;
deriv_num=0;
}
void deriv_log(RealD nrm, RealD max,RealD Fdt_nrm,RealD Fdt_max) {
if ( max > deriv_max_sum ) {
deriv_max_sum=max;
}
deriv_norm_sum+=nrm;
if ( Fdt_max > Fdt_max_sum ) {
Fdt_max_sum=Fdt_max;
}
Fdt_norm_sum+=Fdt_nrm; deriv_num++;
}
RealD deriv_max_average(void) { return deriv_max_sum; };
RealD deriv_norm_average(void) { return deriv_norm_sum/deriv_num; };
RealD Fdt_max_average(void) { return Fdt_max_sum; };
RealD Fdt_norm_average(void) { return Fdt_norm_sum/deriv_num; };
RealD deriv_timer(void) { return deriv_us; };
RealD S_timer(void) { return S_us; };
RealD refresh_timer(void) { return refresh_us; };
void deriv_timer_start(void) { deriv_us-=usecond(); }
void deriv_timer_stop(void) { deriv_us+=usecond(); }
void refresh_timer_start(void) { refresh_us-=usecond(); }
void refresh_timer_stop(void) { refresh_us+=usecond(); }
void S_timer_start(void) { S_us-=usecond(); }
void S_timer_stop(void) { S_us+=usecond(); }
/////////////////////////////
// Heatbath?
/////////////////////////////
virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
virtual RealD S(const GaugeField& U) = 0; // evaluate the action
virtual RealD Sinitial(const GaugeField& U) { return this->S(U); } ; // if the refresh computes the action, can cache it. Alternately refreshAndAction() ?
virtual void deriv(const GaugeField& U, GaugeField& dSdU) = 0; // evaluate the action derivative
/////////////////////////////////////////////////////////////
// virtual smeared interface through configuration container
/////////////////////////////////////////////////////////////
virtual void refresh(ConfigurationBase<GaugeField> & U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
{
refresh(U.get_U(is_smeared),sRNG,pRNG);
}
virtual RealD S(ConfigurationBase<GaugeField>& U)
{
return S(U.get_U(is_smeared));
}
virtual RealD Sinitial(ConfigurationBase<GaugeField>& U)
{
return Sinitial(U.get_U(is_smeared));
}
virtual void deriv(ConfigurationBase<GaugeField>& U, GaugeField& dSdU)
{
deriv(U.get_U(is_smeared),dSdU);
if ( is_smeared ) {
U.smeared_force(dSdU);
}
}
///////////////////////////////
// Logging
///////////////////////////////
virtual std::string action_name() = 0; // return the action name
virtual std::string LogParameters() = 0; // prints action parameters
virtual ~Action(){}
};
template <class GaugeField >
class EmptyAction : public Action <GaugeField>
{
virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) { assert(0);}; // refresh pseudofermions
virtual RealD S(const GaugeField& U) { return 0.0;}; // evaluate the action
virtual void deriv(const GaugeField& U, GaugeField& dSdU) { assert(0); }; // evaluate the action derivative
///////////////////////////////
// Logging
///////////////////////////////
virtual std::string action_name() { return std::string("Level Force Log"); };
virtual std::string LogParameters() { return std::string("No parameters");};
};
NAMESPACE_END(Grid);
#endif // ACTION_BASE_H

6
Grid/qcd/action/ActionCore.h
View File

@ -30,6 +30,8 @@ directory
#ifndef QCD_ACTION_CORE
#define QCD_ACTION_CORE
#include <Grid/qcd/action/gauge/GaugeImplementations.h>
#include <Grid/qcd/action/ActionBase.h>
NAMESPACE_CHECK(ActionBase);
#include <Grid/qcd/action/ActionSet.h>
@ -37,6 +39,10 @@ NAMESPACE_CHECK(ActionSet);
#include <Grid/qcd/action/ActionParams.h>
NAMESPACE_CHECK(ActionParams);
#include <Grid/qcd/action/filters/MomentumFilter.h>
#include <Grid/qcd/action/filters/DirichletFilter.h>
#include <Grid/qcd/action/filters/DDHMCFilter.h>
////////////////////////////////////////////
// Gauge Actions
////////////////////////////////////////////

78
Grid/qcd/action/ActionParams.h
View File

@ -34,27 +34,45 @@ directory
NAMESPACE_BEGIN(Grid);
// These can move into a params header and be given MacroMagic serialisation
struct GparityWilsonImplParams {
Coordinate twists;
GparityWilsonImplParams() : twists(Nd, 0) {};
//mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
Coordinate dirichlet; // Blocksize of dirichlet BCs
int partialDirichlet;
GparityWilsonImplParams() : twists(Nd, 0) {
dirichlet.resize(0);
partialDirichlet=0;
};
};
struct WilsonImplParams {
bool overlapCommsCompute;
Coordinate dirichlet; // Blocksize of dirichlet BCs
int partialDirichlet;
AcceleratorVector<Real,Nd> twist_n_2pi_L;
AcceleratorVector<Complex,Nd> boundary_phases;
WilsonImplParams() {
dirichlet.resize(0);
partialDirichlet=0;
boundary_phases.resize(Nd, 1.0);
twist_n_2pi_L.resize(Nd, 0.0);
};
WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
twist_n_2pi_L.resize(Nd, 0.0);
partialDirichlet=0;
dirichlet.resize(0);
}
};
struct StaggeredImplParams {
StaggeredImplParams() {};
Coordinate dirichlet; // Blocksize of dirichlet BCs
int partialDirichlet;
StaggeredImplParams()
{
partialDirichlet=0;
dirichlet.resize(0);
};
};
struct OneFlavourRationalParams : Serializable {
@ -63,9 +81,11 @@ struct StaggeredImplParams {
RealD, hi,
int, MaxIter,
RealD, tolerance,
RealD, mdtolerance,
int, degree,
int, precision,
int, BoundsCheckFreq);
int, BoundsCheckFreq,
RealD, BoundsCheckTol);
// MaxIter and tolerance, vectors??
@ -76,16 +96,62 @@ struct StaggeredImplParams {
RealD tol = 1.0e-8,
int _degree = 10,
int _precision = 64,
int _BoundsCheckFreq=20)
int _BoundsCheckFreq=20,
RealD mdtol = 1.0e-6,
double _BoundsCheckTol=1e-6)
: lo(_lo),
hi(_hi),
MaxIter(_maxit),
tolerance(tol),
mdtolerance(mdtol),
degree(_degree),
precision(_precision),
BoundsCheckFreq(_BoundsCheckFreq){};
BoundsCheckFreq(_BoundsCheckFreq),
BoundsCheckTol(_BoundsCheckTol){};
};
/*Action parameters for the generalized rational action
The approximation is for (M^dag M)^{1/inv_pow}
where inv_pow is the denominator of the fractional power.
Default inv_pow=2 for square root, making this equivalent to
the OneFlavourRational action
*/
struct RationalActionParams : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(RationalActionParams,
int, inv_pow,
RealD, lo, //low eigenvalue bound of rational approx
RealD, hi, //high eigenvalue bound of rational approx
int, MaxIter, //maximum iterations in msCG
RealD, action_tolerance, //msCG tolerance in action evaluation
int, action_degree, //rational approx tolerance in action evaluation
RealD, md_tolerance, //msCG tolerance in MD integration
int, md_degree, //rational approx tolerance in MD integration
int, precision, //precision of floating point arithmetic
int, BoundsCheckFreq); //frequency the approximation is tested (with Metropolis degree/tolerance); 0 disables the check
// constructor
RationalActionParams(int _inv_pow = 2,
RealD _lo = 0.0,
RealD _hi = 1.0,
int _maxit = 1000,
RealD _action_tolerance = 1.0e-8,
int _action_degree = 10,
RealD _md_tolerance = 1.0e-8,
int _md_degree = 10,
int _precision = 64,
int _BoundsCheckFreq=20)
: inv_pow(_inv_pow),
lo(_lo),
hi(_hi),
MaxIter(_maxit),
action_tolerance(_action_tolerance),
action_degree(_action_degree),
md_tolerance(_md_tolerance),
md_degree(_md_degree),
precision(_precision),
BoundsCheckFreq(_BoundsCheckFreq){};
};
NAMESPACE_END(Grid);
#endif

11
Grid/qcd/action/fermion/CayleyFermion5D.h
View File

@ -71,6 +71,7 @@ public:
RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
RealD MassPlus(void) { return mass_plus; };
RealD MassMinus(void) { return mass_minus; };
void SetMass(RealD _mass) {
mass_plus=mass_minus=_mass;
SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c); // Reset coeffs
@ -182,16 +183,6 @@ public:
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mass,RealD _M5,const ImplParams &p= ImplParams());
void CayleyReport(void);
void CayleyZeroCounters(void);
double M5Dflops;
double M5Dcalls;
double M5Dtime;
double MooeeInvFlops;
double MooeeInvCalls;
double MooeeInvTime;
protected:
virtual void SetCoefficientsZolotarev(RealD zolohi,Approx::zolotarev_data *zdata,RealD b,RealD c);

1
Grid/qcd/action/fermion/CloverHelpers.h
View File

@ -140,6 +140,7 @@ public:
return NMAX;
}
static int getNMAX(Lattice<iImplClover<vComplexD2>> &t, RealD R) {return getNMAX(1e-12,R);}
static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}

291
Grid/qcd/action/fermion/DWFSlow.h Normal file
View File

@ -0,0 +1,291 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/DWFSlow.h
Copyright (C) 2022
Author: Peter Boyle <pboyle@bnl.gov>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#pragma once
NAMESPACE_BEGIN(Grid);
template <class Impl>
class DWFSlowFermion : public FermionOperator<Impl>
{
public:
INHERIT_IMPL_TYPES(Impl);
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////
GridBase *GaugeGrid(void) { return _grid4; }
GridBase *GaugeRedBlackGrid(void) { return _cbgrid4; }
GridBase *FermionGrid(void) { return _grid; }
GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
//////////////////////////////////////////////////////////////////
// override multiply; cut number routines if pass dagger argument
// and also make interface more uniformly consistent
//////////////////////////////////////////////////////////////////
virtual void M(const FermionField &in, FermionField &out)
{
FermionField tmp(_grid);
out = (5.0 - M5) * in;
Dhop(in,tmp,DaggerNo);
out = out + tmp;
}
virtual void Mdag(const FermionField &in, FermionField &out)
{
FermionField tmp(_grid);
out = (5.0 - M5) * in;
Dhop(in,tmp,DaggerYes);
out = out + tmp;
};
/////////////////////////////////////////////////////////
// half checkerboard operations 5D redblack so just site identiy
/////////////////////////////////////////////////////////
void Meooe(const FermionField &in, FermionField &out)
{
if ( in.Checkerboard() == Odd ) {
this->DhopEO(in,out,DaggerNo);
} else {
this->DhopOE(in,out,DaggerNo);
}
}
void MeooeDag(const FermionField &in, FermionField &out)
{
if ( in.Checkerboard() == Odd ) {
this->DhopEO(in,out,DaggerYes);
} else {
this->DhopOE(in,out,DaggerYes);
}
};
// allow override for twisted mass and clover
virtual void Mooee(const FermionField &in, FermionField &out)
{
out = (5.0 - M5) * in;
}
virtual void MooeeDag(const FermionField &in, FermionField &out)
{
out = (5.0 - M5) * in;
}
virtual void MooeeInv(const FermionField &in, FermionField &out)
{
out = (1.0/(5.0 - M5)) * in;
};
virtual void MooeeInvDag(const FermionField &in, FermionField &out)
{
out = (1.0/(5.0 - M5)) * in;
};
virtual void MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _mass,std::vector<double> twist) {} ;
////////////////////////
// Derivative interface
////////////////////////
// Interface calls an internal routine
void DhopDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag) { assert(0);};
void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);};
void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag){ assert(0);};
///////////////////////////////////////////////////////////////
// non-hermitian hopping term; half cb or both
///////////////////////////////////////////////////////////////
void Dhop(const FermionField &in, FermionField &out, int dag)
{
FermionField tmp(in.Grid());
Dhop5(in,out,MassField,MassField,dag );
for(int mu=0;mu<4;mu++){
DhopDirU(in,Umu[mu],Umu[mu],tmp,mu,dag ); out = out + tmp;
}
};
void DhopOE(const FermionField &in, FermionField &out, int dag)
{
FermionField tmp(in.Grid());
assert(in.Checkerboard()==Even);
Dhop5(in,out,MassFieldOdd,MassFieldEven,dag);
for(int mu=0;mu<4;mu++){
DhopDirU(in,UmuOdd[mu],UmuEven[mu],tmp,mu,dag ); out = out + tmp;
}
};
void DhopEO(const FermionField &in, FermionField &out, int dag)
{
FermionField tmp(in.Grid());
assert(in.Checkerboard()==Odd);
Dhop5(in,out, MassFieldEven,MassFieldOdd ,dag );
for(int mu=0;mu<4;mu++){
DhopDirU(in,UmuEven[mu],UmuOdd[mu],tmp,mu,dag ); out = out + tmp;
}
};
///////////////////////////////////////////////////////////////
// Multigrid assistance; force term uses too
///////////////////////////////////////////////////////////////
void Mdir(const FermionField &in, FermionField &out, int dir, int disp){ assert(0);};
void MdirAll(const FermionField &in, std::vector<FermionField> &out) { assert(0);};
void DhopDir(const FermionField &in, FermionField &out, int dir, int disp) { assert(0);};
void DhopDirAll(const FermionField &in, std::vector<FermionField> &out) { assert(0);};
void DhopDirCalc(const FermionField &in, FermionField &out, int dirdisp,int gamma, int dag) { assert(0);};
void DhopDirU(const FermionField &in, const GaugeLinkField &U5e, const GaugeLinkField &U5o, FermionField &out, int mu, int dag)
{
RealD sgn= 1.0;
if (dag ) sgn=-1.0;
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT
};
// mass is 1,1,1,1,-m has to multiply the round the world term
FermionField tmp (in.Grid());
tmp = U5e * Cshift(in,mu+1,1);
out = tmp - Gamma(Gmu[mu])*tmp*sgn;
tmp = Cshift(adj(U5o)*in,mu+1,-1);
out = out + tmp + Gamma(Gmu[mu])*tmp*sgn;
out = -0.5*out;
};
void Dhop5(const FermionField &in, FermionField &out, ComplexField &massE, ComplexField &massO, int dag)
{
// Mass term.... must multiple the round world with mass = 1,1,1,1, -m
RealD sgn= 1.0;
if (dag ) sgn=-1.0;
Gamma G5(Gamma::Algebra::Gamma5);
FermionField tmp (in.Grid());
tmp = massE*Cshift(in,0,1);
out = tmp - G5*tmp*sgn;
tmp = Cshift(massO*in,0,-1);
out = out + tmp + G5*tmp*sgn;
out = -0.5*out;
};
// Constructor
DWFSlowFermion(GaugeField &_Umu, GridCartesian &Fgrid,
GridRedBlackCartesian &Hgrid, RealD _mass, RealD _M5)
:
_grid(&Fgrid),
_cbgrid(&Hgrid),
_grid4(_Umu.Grid()),
Umu(Nd,&Fgrid),
UmuEven(Nd,&Hgrid),
UmuOdd(Nd,&Hgrid),
MassField(&Fgrid),
MassFieldEven(&Hgrid),
MassFieldOdd(&Hgrid),
M5(_M5),
mass(_mass),
_tmp(&Hgrid)
{
Ls=Fgrid._fdimensions[0];
ImportGauge(_Umu);
typedef typename FermionField::scalar_type scalar;
Lattice<iScalar<vInteger> > coor(&Fgrid);
LatticeCoordinate(coor, 0); // Scoor
ComplexField one(&Fgrid);
MassField =scalar(-mass);
one =scalar(1.0);
MassField =where(coor==Integer(Ls-1),MassField,one);
for(int mu=0;mu<Nd;mu++){
pickCheckerboard(Even,UmuEven[mu],Umu[mu]);
pickCheckerboard(Odd ,UmuOdd[mu],Umu[mu]);
}
pickCheckerboard(Even,MassFieldEven,MassField);
pickCheckerboard(Odd ,MassFieldOdd,MassField);
}
// DoubleStore impl dependent
void ImportGauge(const GaugeField &_Umu4)
{
GaugeLinkField U4(_grid4);
for(int mu=0;mu<Nd;mu++){
U4 = PeekIndex<LorentzIndex>(_Umu4, mu);
for(int s=0;s<this->Ls;s++){
InsertSlice(U4,Umu[mu],s,0);
}
}
}
///////////////////////////////////////////////////////////////
// Data members require to support the functionality
///////////////////////////////////////////////////////////////
public:
virtual RealD Mass(void) { return mass; }
virtual int isTrivialEE(void) { return 1; };
RealD mass;
RealD M5;
int Ls;
GridBase *_grid4;
GridBase *_grid;
GridBase *_cbgrid4;
GridBase *_cbgrid;
// Copy of the gauge field , with even and odd subsets
std::vector<GaugeLinkField> Umu;
std::vector<GaugeLinkField> UmuEven;
std::vector<GaugeLinkField> UmuOdd;
ComplexField MassField;
ComplexField MassFieldEven;
ComplexField MassFieldOdd;
///////////////////////////////////////////////////////////////
// Conserved current utilities
///////////////////////////////////////////////////////////////
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu){}
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx){}
};
typedef DWFSlowFermion<WilsonImplF> DWFSlowFermionF;
typedef DWFSlowFermion<WilsonImplD> DWFSlowFermionD;
NAMESPACE_END(Grid);

122
Grid/qcd/action/fermion/Fermion.h
View File

@ -47,6 +47,7 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
////////////////////////////////////////////
// Fermion operators / actions
////////////////////////////////////////////
#include <Grid/qcd/action/fermion/DWFSlow.h> // Slow DWF
#include <Grid/qcd/action/fermion/WilsonFermion.h> // 4d wilson like
NAMESPACE_CHECK(Wilson);
@ -112,28 +113,31 @@ NAMESPACE_CHECK(DWFutils);
// Cayley 5d
NAMESPACE_BEGIN(Grid);
typedef WilsonFermion<WilsonImplR> WilsonFermionR;
typedef WilsonFermion<WilsonImplD2> WilsonFermionD2;
typedef WilsonFermion<WilsonImplF> WilsonFermionF;
typedef WilsonFermion<WilsonImplD> WilsonFermionD;
//typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
//typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
//typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
typedef WilsonFermion<WilsonAdjImplD> WilsonAdjFermionD;
typedef WilsonFermion<WilsonTwoIndexSymmetricImplR> WilsonTwoIndexSymmetricFermionR;
typedef WilsonFermion<WilsonTwoIndexSymmetricImplF> WilsonTwoIndexSymmetricFermionF;
typedef WilsonFermion<WilsonTwoIndexSymmetricImplD> WilsonTwoIndexSymmetricFermionD;
typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonTwoIndexAntiSymmetricFermionR;
typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonTwoIndexAntiSymmetricFermionF;
typedef WilsonFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonTwoIndexAntiSymmetricFermionD;
// Sp(2n)
typedef WilsonFermion<SpWilsonImplF> SpWilsonFermionF;
typedef WilsonFermion<SpWilsonImplD> SpWilsonFermionD;
typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplF> SpWilsonTwoIndexAntiSymmetricFermionF;
typedef WilsonFermion<SpWilsonTwoIndexAntiSymmetricImplD> SpWilsonTwoIndexAntiSymmetricFermionD;
typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplF> SpWilsonTwoIndexSymmetricFermionF;
typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplD> SpWilsonTwoIndexSymmetricFermionD;
// Twisted mass fermion
typedef WilsonTMFermion<WilsonImplR> WilsonTMFermionR;
typedef WilsonTMFermion<WilsonImplD2> WilsonTMFermionD2;
typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
@ -141,23 +145,20 @@ typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
typedef WilsonClover<WilsonImplR> WilsonCloverFermionR;
typedef WilsonClover<WilsonImplD2> WilsonCloverFermionD2;
typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
typedef WilsonExpClover<WilsonImplD2> WilsonExpCloverFermionD2;
typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
typedef WilsonClover<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
@ -165,161 +166,108 @@ typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiS
template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
typedef CompactWilsonClover<WilsonImplR> CompactWilsonCloverFermionR;
typedef CompactWilsonClover<WilsonImplD2> CompactWilsonCloverFermionD2;
typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
typedef CompactWilsonExpClover<WilsonImplR> CompactWilsonExpCloverFermionR;
typedef CompactWilsonExpClover<WilsonImplD2> CompactWilsonExpCloverFermionD2;
typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
typedef CompactWilsonClover<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
// Domain Wall fermions
typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
typedef DomainWallFermion<WilsonImplD2> DomainWallFermionD2;
//typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
//typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
//typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
typedef DomainWallEOFAFermion<WilsonImplR> DomainWallEOFAFermionR;
typedef DomainWallEOFAFermion<WilsonImplD2> DomainWallEOFAFermionD2;
typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
//typedef DomainWallEOFAFermion<WilsonImplRL> DomainWallEOFAFermionRL;
//typedef DomainWallEOFAFermion<WilsonImplFH> DomainWallEOFAFermionFH;
//typedef DomainWallEOFAFermion<WilsonImplDF> DomainWallEOFAFermionDF;
typedef MobiusFermion<WilsonImplR> MobiusFermionR;
typedef MobiusFermion<WilsonImplD2> MobiusFermionD2;
typedef MobiusFermion<WilsonImplF> MobiusFermionF;
typedef MobiusFermion<WilsonImplD> MobiusFermionD;
//typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
//typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
//typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
typedef MobiusEOFAFermion<WilsonImplR> MobiusEOFAFermionR;
typedef MobiusEOFAFermion<WilsonImplD2> MobiusEOFAFermionD2;
typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
//typedef MobiusEOFAFermion<WilsonImplRL> MobiusEOFAFermionRL;
//typedef MobiusEOFAFermion<WilsonImplFH> MobiusEOFAFermionFH;
//typedef MobiusEOFAFermion<WilsonImplDF> MobiusEOFAFermionDF;
typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
typedef ZMobiusFermion<ZWilsonImplD2> ZMobiusFermionD2;
typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
//typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
//typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
//typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
// Ls vectorised
typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
typedef ScaledShamirFermion<WilsonImplD2> ScaledShamirFermionD2;
typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF;
typedef ScaledShamirFermion<WilsonImplD> ScaledShamirFermionD;
typedef MobiusZolotarevFermion<WilsonImplR> MobiusZolotarevFermionR;
typedef MobiusZolotarevFermion<WilsonImplD2> MobiusZolotarevFermionD2;
typedef MobiusZolotarevFermion<WilsonImplF> MobiusZolotarevFermionF;
typedef MobiusZolotarevFermion<WilsonImplD> MobiusZolotarevFermionD;
typedef ShamirZolotarevFermion<WilsonImplR> ShamirZolotarevFermionR;
typedef ShamirZolotarevFermion<WilsonImplD2> ShamirZolotarevFermionD2;
typedef ShamirZolotarevFermion<WilsonImplF> ShamirZolotarevFermionF;
typedef ShamirZolotarevFermion<WilsonImplD> ShamirZolotarevFermionD;
typedef OverlapWilsonCayleyTanhFermion<WilsonImplR> OverlapWilsonCayleyTanhFermionR;
typedef OverlapWilsonCayleyTanhFermion<WilsonImplD2> OverlapWilsonCayleyTanhFermionD2;
typedef OverlapWilsonCayleyTanhFermion<WilsonImplF> OverlapWilsonCayleyTanhFermionF;
typedef OverlapWilsonCayleyTanhFermion<WilsonImplD> OverlapWilsonCayleyTanhFermionD;
typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplR> OverlapWilsonCayleyZolotarevFermionR;
typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD2> OverlapWilsonCayleyZolotarevFermionD2;
typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplF> OverlapWilsonCayleyZolotarevFermionF;
typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD> OverlapWilsonCayleyZolotarevFermionD;
// Continued fraction
typedef OverlapWilsonContFracTanhFermion<WilsonImplR> OverlapWilsonContFracTanhFermionR;
typedef OverlapWilsonContFracTanhFermion<WilsonImplD2> OverlapWilsonContFracTanhFermionD2;
typedef OverlapWilsonContFracTanhFermion<WilsonImplF> OverlapWilsonContFracTanhFermionF;
typedef OverlapWilsonContFracTanhFermion<WilsonImplD> OverlapWilsonContFracTanhFermionD;
typedef OverlapWilsonContFracZolotarevFermion<WilsonImplR> OverlapWilsonContFracZolotarevFermionR;
typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD2> OverlapWilsonContFracZolotarevFermionD2;
typedef OverlapWilsonContFracZolotarevFermion<WilsonImplF> OverlapWilsonContFracZolotarevFermionF;
typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD> OverlapWilsonContFracZolotarevFermionD;
// Partial fraction
typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplR> OverlapWilsonPartialFractionTanhFermionR;
typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD2> OverlapWilsonPartialFractionTanhFermionD2;
typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplF> OverlapWilsonPartialFractionTanhFermionF;
typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD> OverlapWilsonPartialFractionTanhFermionD;
typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplR> OverlapWilsonPartialFractionZolotarevFermionR;
typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD2> OverlapWilsonPartialFractionZolotarevFermionD2;
typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplF> OverlapWilsonPartialFractionZolotarevFermionF;
typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD> OverlapWilsonPartialFractionZolotarevFermionD;
// Gparity cases; partial list until tested
typedef WilsonFermion<GparityWilsonImplR> GparityWilsonFermionR;
typedef WilsonFermion<GparityWilsonImplF> GparityWilsonFermionF;
typedef WilsonFermion<GparityWilsonImplD> GparityWilsonFermionD;
//typedef WilsonFermion<GparityWilsonImplRL> GparityWilsonFermionRL;
//typedef WilsonFermion<GparityWilsonImplFH> GparityWilsonFermionFH;
//typedef WilsonFermion<GparityWilsonImplDF> GparityWilsonFermionDF;
typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
//typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
//typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
//typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionR;
typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionD2;
typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
//typedef DomainWallEOFAFermion<GparityWilsonImplRL> GparityDomainWallEOFAFermionRL;
//typedef DomainWallEOFAFermion<GparityWilsonImplFH> GparityDomainWallEOFAFermionFH;
//typedef DomainWallEOFAFermion<GparityWilsonImplDF> GparityDomainWallEOFAFermionDF;
typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionD2;
typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
//typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
//typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
//typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionD2;
typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
//typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
//typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
//typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionR;
typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionD2;
typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
//typedef MobiusEOFAFermion<GparityWilsonImplRL> GparityMobiusEOFAFermionRL;
//typedef MobiusEOFAFermion<GparityWilsonImplFH> GparityMobiusEOFAFermionFH;
//typedef MobiusEOFAFermion<GparityWilsonImplDF> GparityMobiusEOFAFermionDF;
typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD;
typedef NaiveStaggeredFermion<StaggeredImplR> NaiveStaggeredFermionR;
typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR;
typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF;
typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD;

2
Grid/qcd/action/fermion/FermionOperator.h
View File

@ -49,6 +49,8 @@ public:
virtual FermionField &tmp(void) = 0;
virtual void DirichletBlock(const Coordinate & _Block) { assert(0); };
GridBase * Grid(void) { return FermionGrid(); }; // this is all the linalg routines need to know
GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); };

136
Grid/qcd/action/fermion/GparityWilsonImpl.h
View File

@ -30,6 +30,18 @@ directory
NAMESPACE_BEGIN(Grid);
/*
Policy implementation for G-parity boundary conditions
Rather than treating the gauge field as a flavored field, the Grid implementation of G-parity treats the gauge field as a regular
field with complex conjugate boundary conditions. In order to ensure the second flavor interacts with the conjugate links and the first
with the regular links we overload the functionality of doubleStore, whose purpose is to store the gauge field and the barrel-shifted gauge field
to avoid communicating links when applying the Dirac operator, such that the double-stored field contains also a flavor index which maps to
either the link or the conjugate link. This flavored field is then used by multLink to apply the correct link to a spinor.
Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction.
mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
*/
template <class S, class Representation = FundamentalRepresentation, class Options=CoeffReal>
class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > {
public:
@ -113,7 +125,7 @@ public:
|| ((distance== 1)&&(icoor[direction]==1))
|| ((distance==-1)&&(icoor[direction]==0));
permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu] && mmu < Nd-1; //only if we are going around the world in a spatial direction
//Apply the links
int f_upper = permute_lane ? 1 : 0;
@ -139,10 +151,10 @@ public:
assert((distance == 1) || (distance == -1)); // nearest neighbour stencil hard code
assert((sl == 1) || (sl == 2));
if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
//If this site is an global boundary site, perform the G-parity flavor twist
if ( mmu < Nd-1 && SE->_around_the_world && St.parameters.twists[mmu] ) {
if ( sl == 2 ) {
//Only do the twist for lanes on the edge of the physical node
ExtractBuffer<sobj> vals(Nsimd);
extract(chi,vals);
@ -197,6 +209,19 @@ public:
reg = memory;
}
//Poke 'poke_f0' onto flavor 0 and 'poke_f1' onto flavor 1 in direction mu of the doubled gauge field Uds
inline void pokeGparityDoubledGaugeField(DoubledGaugeField &Uds, const GaugeLinkField &poke_f0, const GaugeLinkField &poke_f1, const int mu){
autoView(poke_f0_v, poke_f0, CpuRead);
autoView(poke_f1_v, poke_f1, CpuRead);
autoView(Uds_v, Uds, CpuWrite);
thread_foreach(ss,poke_f0_v,{
Uds_v[ss](0)(mu) = poke_f0_v[ss]();
Uds_v[ss](1)(mu) = poke_f1_v[ss]();
});
}
inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
{
conformable(Uds.Grid(),GaugeGrid);
@ -207,14 +232,19 @@ public:
GaugeLinkField Uconj(GaugeGrid);
Lattice<iScalar<vInteger> > coor(GaugeGrid);
for(int mu=0;mu<Nd;mu++){
LatticeCoordinate(coor,mu);
//Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction.
//mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
for(int mu=0;mu<Nd-1;mu++){
if( Params.twists[mu] ){
LatticeCoordinate(coor,mu);
}
U = PeekIndex<LorentzIndex>(Umu,mu);
Uconj = conjugate(U);
// Implement the isospin rotation sign on the boundary between f=1 and f=0
// This phase could come from a simple bc 1,1,-1,1 ..
int neglink = GaugeGrid->GlobalDimensions()[mu]-1;
if ( Params.twists[mu] ) {
@ -229,7 +259,7 @@ public:
thread_foreach(ss,U_v,{
Uds_v[ss](0)(mu) = U_v[ss]();
Uds_v[ss](1)(mu) = Uconj_v[ss]();
});
});
}
U = adj(Cshift(U ,mu,-1)); // correct except for spanning the boundary
@ -260,6 +290,38 @@ public:
});
}
}
{ //periodic / antiperiodic temporal BCs
int mu = Nd-1;
int L = GaugeGrid->GlobalDimensions()[mu];
int Lmu = L - 1;
LatticeCoordinate(coor, mu);
U = PeekIndex<LorentzIndex>(Umu, mu); //Get t-directed links
GaugeLinkField *Upoke = &U;
if(Params.twists[mu]){ //antiperiodic
Utmp = where(coor == Lmu, -U, U);
Upoke = &Utmp;
}
Uconj = conjugate(*Upoke); //second flavor interacts with conjugate links
pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu);
//Get the barrel-shifted field
Utmp = adj(Cshift(U, mu, -1)); //is a forward shift!
Upoke = &Utmp;
if(Params.twists[mu]){
U = where(coor == 0, -Utmp, Utmp); //boundary phase
Upoke = &U;
}
Uconj = conjugate(*Upoke);
pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu + 4);
}
}
inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) {
@ -298,28 +360,48 @@ public:
inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
assert(0);
}
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
int Ls = Btilde.Grid()->_fdimensions[0];
GaugeLinkField tmp(mat.Grid());
tmp = Zero();
int Ls=Btilde.Grid()->_fdimensions[0];
{
autoView( tmp_v , tmp, CpuWrite);
autoView( Atilde_v , Atilde, CpuRead);
autoView( Btilde_v , Btilde, CpuRead);
thread_for(ss,tmp.Grid()->oSites(),{
for (int s = 0; s < Ls; s++) {
int sF = s + Ls * ss;
auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
}
});
GridBase *GaugeGrid = mat.Grid();
Lattice<iScalar<vInteger> > coor(GaugeGrid);
if( Params.twists[mu] ){
LatticeCoordinate(coor,mu);
}
autoView( mat_v , mat, AcceleratorWrite);
autoView( Btilde_v , Btilde, AcceleratorRead);
autoView( Atilde_v , Atilde, AcceleratorRead);
accelerator_for(sss,mat.Grid()->oSites(), FermionField::vector_type::Nsimd(),{
int sU=sss;
typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
ColorMatrixType sum;
zeroit(sum);
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
for(int spn=0;spn<Ns;spn++){ //sum over spin
//Flavor 0
auto bb = coalescedRead(Btilde_v[sF](0)(spn) ); //color vector
auto aa = coalescedRead(Atilde_v[sF](0)(spn) );
sum = sum + outerProduct(bb,aa);
//Flavor 1
bb = coalescedRead(Btilde_v[sF](1)(spn) );
aa = coalescedRead(Atilde_v[sF](1)(spn) );
sum = sum + conjugate(outerProduct(bb,aa));
}
}
coalescedWrite(mat_v[sU](mu)(), sum);
});
}
PokeIndex<LorentzIndex>(mat, tmp, mu);
return;
}
};

12
Grid/qcd/action/fermion/ImprovedStaggeredFermion.h
View File

@ -47,18 +47,6 @@ public:
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
////////////////////////////////////////
// Performance monitoring
////////////////////////////////////////
void Report(void);
void ZeroCounters(void);
double DhopTotalTime;
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
double DhopComputeTime2;
double DhopFaceTime;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////

12
Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h
View File

@ -52,18 +52,6 @@ public:
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
////////////////////////////////////////
// Performance monitoring
////////////////////////////////////////
void Report(void);
void ZeroCounters(void);
double DhopTotalTime;
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
double DhopComputeTime2;
double DhopFaceTime;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////

12
Grid/qcd/action/fermion/NaiveStaggeredFermion.h
View File

@ -47,18 +47,6 @@ public:
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
////////////////////////////////////////
// Performance monitoring
////////////////////////////////////////
void Report(void);
void ZeroCounters(void);
double DhopTotalTime;
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
double DhopComputeTime2;
double DhopFaceTime;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////

423
Grid/qcd/action/fermion/WilsonCompressor.h
View File

@ -32,17 +32,218 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////////
// Wilson compressor will need FaceGather policies for:
// Periodic, Dirichlet, and partial Dirichlet for DWF
///////////////////////////////////////////////////////////////
const int dwf_compressor_depth=2;
#define DWF_COMPRESS
class FaceGatherPartialDWF
{
public:
#ifdef DWF_COMPRESS
static int PartialCompressionFactor(GridBase *grid) {return grid->_fdimensions[0]/(2*dwf_compressor_depth);};
#else
static int PartialCompressionFactor(GridBase *grid) { return 1;}
#endif
template<class vobj,class cobj,class compressor>
static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
const Lattice<vobj> &rhs,
cobj *buffer,
compressor &compress,
int off,int so,int partial)
{
//DWF only hack: If a direction that is OFF node we use Partial Dirichlet
// Shrinks local and remote comms buffers
GridBase *Grid = rhs.Grid();
int Ls = Grid->_rdimensions[0];
#ifdef DWF_COMPRESS
int depth=dwf_compressor_depth;
#else
int depth=Ls/2;
#endif
std::pair<int,int> *table_v = & table[0];
auto rhs_v = rhs.View(AcceleratorRead);
int vol=table.size()/Ls;
accelerator_forNB( idx,table.size(), vobj::Nsimd(), {
Integer i=idx/Ls;
Integer s=idx%Ls;
Integer sc=depth+s-(Ls-depth);
if(s<depth) compress.Compress(buffer[off+i+s*vol],rhs_v[so+table_v[idx].second]);
if(s>=Ls-depth) compress.Compress(buffer[off+i+sc*vol],rhs_v[so+table_v[idx].second]);
});
rhs_v.ViewClose();
}
template<class decompressor,class Decompression>
static void DecompressFace(decompressor decompress,Decompression &dd)
{
auto Ls = dd.dims[0];
#ifdef DWF_COMPRESS
int depth=dwf_compressor_depth;
#else
int depth=Ls/2;
#endif
// Just pass in the Grid
auto kp = dd.kernel_p;
auto mp = dd.mpi_p;
int size= dd.buffer_size;
int vol= size/Ls;
accelerator_forNB(o,size,1,{
int idx=o/Ls;
int s=o%Ls;
if ( s < depth ) {
int oo=s*vol+idx;
kp[o]=mp[oo];
} else if ( s >= Ls-depth ) {
int sc = depth + s - (Ls-depth);
int oo=sc*vol+idx;
kp[o]=mp[oo];
} else {
kp[o] = Zero();//fill rest with zero if partial dirichlet
}
});
}
////////////////////////////////////////////////////////////////////////////////////////////
// Need to gather *interior portions* for ALL s-slices in simd directions
// Do the gather as need to treat SIMD lanes differently, and insert zeroes on receive side
// Reorder the fifth dim to be s=Ls-1 , s=0, s=1,...,Ls-2.
////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj,class cobj,class compressor>
static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
compressor &compress,int type,int partial)
{
GridBase *Grid = rhs.Grid();
int Ls = Grid->_rdimensions[0];
#ifdef DWF_COMPRESS
int depth=dwf_compressor_depth;
#else
int depth = Ls/2;
#endif
// insertion of zeroes...
assert( (table.size()&0x1)==0);
int num=table.size()/2;
int so = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
auto rhs_v = rhs.View(AcceleratorRead);
auto p0=&pointers[0][0];
auto p1=&pointers[1][0];
auto tp=&table[0];
int nnum=num/Ls;
accelerator_forNB(j, num, vobj::Nsimd(), {
// Reorders both local and remote comms buffers
//
int s = j % Ls;
int sp1 = (s+depth)%Ls; // peri incremented s slice
int hxyz= j/Ls;
int xyz0= hxyz*2; // xyzt part of coor
int xyz1= hxyz*2+1;
int jj= hxyz + sp1*nnum ; // 0,1,2,3 -> Ls-1 slice , 0-slice, 1-slice ....
int kk0= xyz0*Ls + s ; // s=0 goes to s=1
int kk1= xyz1*Ls + s ; // s=Ls-1 -> s=0
compress.CompressExchange(p0[jj],p1[jj],
rhs_v[so+tp[kk0 ].second], // Same s, consecutive xyz sites
rhs_v[so+tp[kk1 ].second],
type);
});
rhs_v.ViewClose();
}
// Merge routine is for SIMD faces
template<class decompressor,class Merger>
static void MergeFace(decompressor decompress,Merger &mm)
{
auto Ls = mm.dims[0];
#ifdef DWF_COMPRESS
int depth=dwf_compressor_depth;
#else
int depth = Ls/2;
#endif
int num= mm.buffer_size/2; // relate vol and Ls to buffer size
auto mp = &mm.mpointer[0];
auto vp0= &mm.vpointers[0][0]; // First arg is exchange first
auto vp1= &mm.vpointers[1][0];
auto type= mm.type;
int nnum = num/Ls;
accelerator_forNB(o,num,Merger::Nsimd,{
int s=o%Ls;
int hxyz=o/Ls; // xyzt related component
int xyz0=hxyz*2;
int xyz1=hxyz*2+1;
int sp = (s+depth)%Ls;
int jj= hxyz + sp*nnum ; // 0,1,2,3 -> Ls-1 slice , 0-slice, 1-slice ....
int oo0= s+xyz0*Ls;
int oo1= s+xyz1*Ls;
// same ss0, ss1 pair goes to new layout
decompress.Exchange(mp[oo0],mp[oo1],vp0[jj],vp1[jj],type);
});
}
};
class FaceGatherDWFMixedBCs
{
public:
#ifdef DWF_COMPRESS
static int PartialCompressionFactor(GridBase *grid) {return grid->_fdimensions[0]/(2*dwf_compressor_depth);};
#else
static int PartialCompressionFactor(GridBase *grid) {return 1;}
#endif
template<class vobj,class cobj,class compressor>
static void Gather_plane_simple (commVector<std::pair<int,int> >& table,
const Lattice<vobj> &rhs,
cobj *buffer,
compressor &compress,
int off,int so,int partial)
{
// std::cout << " face gather simple DWF partial "<<partial <<std::endl;
if(partial) FaceGatherPartialDWF::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
else FaceGatherSimple::Gather_plane_simple(table,rhs,buffer,compress,off,so,partial);
}
template<class vobj,class cobj,class compressor>
static void Gather_plane_exchange(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
compressor &compress,int type,int partial)
{
// std::cout << " face gather exch DWF partial "<<partial <<std::endl;
if(partial) FaceGatherPartialDWF::Gather_plane_exchange(table,rhs,pointers,dimension, plane,cbmask,compress,type,partial);
else FaceGatherSimple::Gather_plane_exchange (table,rhs,pointers,dimension, plane,cbmask,compress,type,partial);
}
template<class decompressor,class Merger>
static void MergeFace(decompressor decompress,Merger &mm)
{
int partial = mm.partial;
// std::cout << " merge DWF partial "<<partial <<std::endl;
if ( partial ) FaceGatherPartialDWF::MergeFace(decompress,mm);
else FaceGatherSimple::MergeFace(decompress,mm);
}
template<class decompressor,class Decompression>
static void DecompressFace(decompressor decompress,Decompression &dd)
{
int partial = dd.partial;
// std::cout << " decompress DWF partial "<<partial <<std::endl;
if ( partial ) FaceGatherPartialDWF::DecompressFace(decompress,dd);
else FaceGatherSimple::DecompressFace(decompress,dd);
}
};
/////////////////////////////////////////////////////////////////////////////////////////////
// optimised versions supporting half precision too
// optimised versions supporting half precision too??? Deprecate
/////////////////////////////////////////////////////////////////////////////////////////////
template<class _HCspinor,class _Hspinor,class _Spinor, class projector,typename SFINAE = void >
class WilsonCompressorTemplate;
//Could make FaceGather a template param, but then behaviour is runtime not compile time
template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
typename std::enable_if<std::is_same<_HCspinor,_Hspinor>::value>::type >
class WilsonCompressorTemplate : public FaceGatherDWFMixedBCs
// : public FaceGatherSimple
{
public:
@ -79,172 +280,81 @@ public:
/*****************************************************/
/* Exchange includes precision change if mpi data is not same */
/*****************************************************/
accelerator_inline void Exchange(SiteHalfSpinor *mp,
const SiteHalfSpinor * __restrict__ vp0,
const SiteHalfSpinor * __restrict__ vp1,
Integer type,Integer o) const {
accelerator_inline void Exchange(SiteHalfSpinor &mp0,
SiteHalfSpinor &mp1,
const SiteHalfSpinor & vp0,
const SiteHalfSpinor & vp1,
Integer type) const {
#ifdef GRID_SIMT
exchangeSIMT(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
exchangeSIMT(mp0,mp1,vp0,vp1,type);
#else
SiteHalfSpinor tmp1;
SiteHalfSpinor tmp2;
exchange(tmp1,tmp2,vp0[o],vp1[o],type);
vstream(mp[2*o ],tmp1);
vstream(mp[2*o+1],tmp2);
exchange(tmp1,tmp2,vp0,vp1,type);
vstream(mp0,tmp1);
vstream(mp1,tmp2);
#endif
}
/*****************************************************/
/* Have a decompression step if mpi data is not same */
/*****************************************************/
accelerator_inline void Decompress(SiteHalfSpinor * __restrict__ out,
SiteHalfSpinor * __restrict__ in, Integer o) const {
assert(0);
accelerator_inline void Decompress(SiteHalfSpinor &out,
SiteHalfSpinor &in) const {
out = in;
}
/*****************************************************/
/* Compress Exchange */
/*****************************************************/
accelerator_inline void CompressExchange(SiteHalfSpinor * __restrict__ out0,
SiteHalfSpinor * __restrict__ out1,
const SiteSpinor * __restrict__ in,
Integer j,Integer k, Integer m,Integer type) const
accelerator_inline void CompressExchange(SiteHalfSpinor &out0,
SiteHalfSpinor &out1,
const SiteSpinor &in0,
const SiteSpinor &in1,
Integer type) const
{
#ifdef GRID_SIMT
typedef SiteSpinor vobj;
typedef SiteHalfSpinor hvobj;
typedef decltype(coalescedRead(*in)) sobj;
typedef decltype(coalescedRead(*out0)) hsobj;
typedef decltype(coalescedRead(in0)) sobj;
typedef decltype(coalescedRead(out0)) hsobj;
constexpr unsigned int Nsimd = vobj::Nsimd();
unsigned int mask = Nsimd >> (type + 1);
int lane = acceleratorSIMTlane(Nsimd);
int j0 = lane &(~mask); // inner coor zero
int j1 = lane |(mask) ; // inner coor one
const vobj *vp0 = &in[k]; // out0[j] = merge low bit of type from in[k] and in[m]
const vobj *vp1 = &in[m]; // out1[j] = merge hi bit of type from in[k] and in[m]
const vobj *vp = (lane&mask) ? vp1:vp0;// if my lane has high bit take vp1, low bit take vp0
auto sa = coalescedRead(*vp,j0); // lane to read for out 0, NB 50% read coalescing
auto sb = coalescedRead(*vp,j1); // lane to read for out 1
const vobj *vp0 = &in0;
const vobj *vp1 = &in1;
const vobj *vp = (lane&mask) ? vp1:vp0;
auto sa = coalescedRead(*vp,j0);
auto sb = coalescedRead(*vp,j1);
hsobj psa, psb;
projector::Proj(psa,sa,mu,dag); // spin project the result0
projector::Proj(psb,sb,mu,dag); // spin project the result1
coalescedWrite(out0[j],psa);
coalescedWrite(out1[j],psb);
projector::Proj(psa,sa,mu,dag);
projector::Proj(psb,sb,mu,dag);
coalescedWrite(out0,psa);
coalescedWrite(out1,psb);
#else
SiteHalfSpinor temp1, temp2;
SiteHalfSpinor temp3, temp4;
projector::Proj(temp1,in[k],mu,dag);
projector::Proj(temp2,in[m],mu,dag);
projector::Proj(temp1,in0,mu,dag);
projector::Proj(temp2,in1,mu,dag);
exchange(temp3,temp4,temp1,temp2,type);
vstream(out0[j],temp3);
vstream(out1[j],temp4);
vstream(out0,temp3);
vstream(out1,temp4);
#endif
}
/*****************************************************/
/* Pass the info to the stencil */
/*****************************************************/
accelerator_inline bool DecompressionStep(void) const { return false; }
accelerator_inline bool DecompressionStep(void) const {
return false;
}
};
#if 0
template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
class WilsonCompressorTemplate< _HCspinor, _Hspinor, _Spinor, projector,
typename std::enable_if<!std::is_same<_HCspinor,_Hspinor>::value>::type >
{
public:
int mu,dag;
void Point(int p) { mu=p; };
WilsonCompressorTemplate(int _dag=0){
dag = _dag;
}
typedef _Spinor SiteSpinor;
typedef _Hspinor SiteHalfSpinor;
typedef _HCspinor SiteHalfCommSpinor;
typedef typename SiteHalfCommSpinor::vector_type vComplexLow;
typedef typename SiteHalfSpinor::vector_type vComplexHigh;
constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
accelerator_inline int CommDatumSize(void) const {
return sizeof(SiteHalfCommSpinor);
}
/*****************************************************/
/* Compress includes precision change if mpi data is not same */
/*****************************************************/
accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
SiteHalfSpinor hsp;
SiteHalfCommSpinor *hbuf = (SiteHalfCommSpinor *)buf;
projector::Proj(hsp,in,mu,dag);
precisionChange((vComplexLow *)&hbuf[o],(vComplexHigh *)&hsp,Nw);
}
accelerator_inline void Compress(SiteHalfSpinor &buf,const SiteSpinor &in) const {
#ifdef GRID_SIMT
typedef decltype(coalescedRead(buf)) sobj;
sobj sp;
auto sin = coalescedRead(in);
projector::Proj(sp,sin,mu,dag);
coalescedWrite(buf,sp);
#else
projector::Proj(buf,in,mu,dag);
#endif
}
/*****************************************************/
/* Exchange includes precision change if mpi data is not same */
/*****************************************************/
accelerator_inline void Exchange(SiteHalfSpinor *mp,
SiteHalfSpinor *vp0,
SiteHalfSpinor *vp1,
Integer type,Integer o) const {
SiteHalfSpinor vt0,vt1;
SiteHalfCommSpinor *vpp0 = (SiteHalfCommSpinor *)vp0;
SiteHalfCommSpinor *vpp1 = (SiteHalfCommSpinor *)vp1;
precisionChange((vComplexHigh *)&vt0,(vComplexLow *)&vpp0[o],Nw);
precisionChange((vComplexHigh *)&vt1,(vComplexLow *)&vpp1[o],Nw);
exchange(mp[2*o],mp[2*o+1],vt0,vt1,type);
}
/*****************************************************/
/* Have a decompression step if mpi data is not same */
/*****************************************************/
accelerator_inline void Decompress(SiteHalfSpinor *out, SiteHalfSpinor *in, Integer o) const {
SiteHalfCommSpinor *hin=(SiteHalfCommSpinor *)in;
precisionChange((vComplexHigh *)&out[o],(vComplexLow *)&hin[o],Nw);
}
/*****************************************************/
/* Compress Exchange */
/*****************************************************/
accelerator_inline void CompressExchange(SiteHalfSpinor *out0,
SiteHalfSpinor *out1,
const SiteSpinor *in,
Integer j,Integer k, Integer m,Integer type) const {
SiteHalfSpinor temp1, temp2,temp3,temp4;
SiteHalfCommSpinor *hout0 = (SiteHalfCommSpinor *)out0;
SiteHalfCommSpinor *hout1 = (SiteHalfCommSpinor *)out1;
projector::Proj(temp1,in[k],mu,dag);
projector::Proj(temp2,in[m],mu,dag);
exchange(temp3,temp4,temp1,temp2,type);
precisionChange((vComplexLow *)&hout0[j],(vComplexHigh *)&temp3,Nw);
precisionChange((vComplexLow *)&hout1[j],(vComplexHigh *)&temp4,Nw);
}
/*****************************************************/
/* Pass the info to the stencil */
/*****************************************************/
accelerator_inline bool DecompressionStep(void) const { return true; }
};
#endif
#define DECLARE_PROJ(Projector,Compressor,spProj) \
class Projector { \
public: \
@ -294,11 +404,7 @@ public:
typedef typename Base::View_type View_type;
typedef typename Base::StencilVector StencilVector;
void ZeroCountersi(void) { }
void Reporti(int calls) { }
std::vector<int> surface_list;
// Vector<int> surface_list;
WilsonStencil(GridBase *grid,
int npoints,
int checkerboard,
@ -306,11 +412,11 @@ public:
const std::vector<int> &distances,Parameters p)
: CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p)
{
ZeroCountersi();
surface_list.resize(0);
// surface_list.resize(0);
this->same_node.resize(npoints);
};
/*
void BuildSurfaceList(int Ls,int vol4){
// find same node for SHM
@ -331,7 +437,8 @@ public:
}
}
}
*/
template < class compressor>
void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress)
{
@ -377,24 +484,26 @@ public:
int dag = compress.dag;
int face_idx=0;
#define vet_same_node(a,b) \
{ auto tmp = b; }
if ( dag ) {
assert(this->same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
assert(this->same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
assert(this->same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
assert(this->same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
assert(this->same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
assert(this->same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
assert(this->same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
assert(this->same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XpCompress,Xp,face_idx));
vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YpCompress,Yp,face_idx));
vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TpCompress,Tp,face_idx));
vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XmCompress,Xm,face_idx));
vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YmCompress,Ym,face_idx));
vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TmCompress,Tm,face_idx));
} else {
assert(this->same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
assert(this->same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
assert(this->same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
assert(this->same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
assert(this->same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
assert(this->same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
assert(this->same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
assert(this->same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XmCompress,Xp,face_idx));
vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YmCompress,Yp,face_idx));
vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TmCompress,Tp,face_idx));
vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XpCompress,Xm,face_idx));
vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YpCompress,Ym,face_idx));
vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TpCompress,Tm,face_idx));
}
this->face_table_computed=1;
assert(this->u_comm_offset==this->_unified_buffer_size);

14
Grid/qcd/action/fermion/WilsonFermion.h
View File

@ -74,20 +74,6 @@ public:
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
void Report(void);
void ZeroCounters(void);
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
double DhopComputeTime2;
double DhopFaceTime;
double DhopTotalTime;
double DerivCalls;
double DerivCommTime;
double DerivComputeTime;
double DerivDhopComputeTime;
//////////////////////////////////////////////////////////////////
// override multiply; cut number routines if pass dagger argument
// and also make interface more uniformly consistent

20
Grid/qcd/action/fermion/WilsonFermion5D.h
View File

@ -75,19 +75,8 @@ public:
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
void Report(void);
void ZeroCounters(void);
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
double DhopComputeTime2;
double DhopFaceTime;
double DhopTotalTime;
double DerivCalls;
double DerivCommTime;
double DerivComputeTime;
double DerivDhopComputeTime;
int Dirichlet;
Coordinate Block;
///////////////////////////////////////////////////////////////
// Implement the abstract base
@ -173,7 +162,10 @@ public:
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
double _M5,const ImplParams &p= ImplParams());
virtual void DirichletBlock(const Coordinate & block)
{
}
// Constructors
/*
WilsonFermion5D(int simd,

30
Grid/qcd/action/fermion/WilsonImpl.h
View File

@ -37,7 +37,7 @@ NAMESPACE_BEGIN(Grid);
template <class S, class Representation = FundamentalRepresentation,class Options = CoeffReal >
class WilsonImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation::Dimension > > {
public:
static const int Dimension = Representation::Dimension;
static const bool isFundamental = Representation::isFundamental;
static const bool LsVectorised=false;
@ -246,19 +246,13 @@ public:
typedef WilsonImpl<vComplex, FundamentalRepresentation, CoeffReal > WilsonImplR; // Real.. whichever prec
typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffReal > WilsonImplF; // Float
typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffReal > WilsonImplD; // Double
//typedef WilsonImpl<vComplex, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL; // Real.. whichever prec
//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH; // Float
//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF; // Double
typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffReal > WilsonImplD2; // Double
typedef WilsonImpl<vComplex, FundamentalRepresentation, CoeffComplex > ZWilsonImplR; // Real.. whichever prec
typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplex > ZWilsonImplF; // Float
typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplex > ZWilsonImplD; // Double
typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffComplex > ZWilsonImplD2; // Double
//typedef WilsonImpl<vComplex, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
//typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
//typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
typedef WilsonImpl<vComplex, AdjointRepresentation, CoeffReal > WilsonAdjImplR; // Real.. whichever prec
typedef WilsonImpl<vComplexF, AdjointRepresentation, CoeffReal > WilsonAdjImplF; // Float
typedef WilsonImpl<vComplexD, AdjointRepresentation, CoeffReal > WilsonAdjImplD; // Double
@ -271,6 +265,22 @@ typedef WilsonImpl<vComplex, TwoIndexAntiSymmetricRepresentation, CoeffReal > W
typedef WilsonImpl<vComplexF, TwoIndexAntiSymmetricRepresentation, CoeffReal > WilsonTwoIndexAntiSymmetricImplF; // Float
typedef WilsonImpl<vComplexD, TwoIndexAntiSymmetricRepresentation, CoeffReal > WilsonTwoIndexAntiSymmetricImplD; // Double
//sp 2n
typedef WilsonImpl<vComplex, SpFundamentalRepresentation, CoeffReal > SpWilsonImplR; // Real.. whichever prec
typedef WilsonImpl<vComplexF, SpFundamentalRepresentation, CoeffReal > SpWilsonImplF; // Float
typedef WilsonImpl<vComplexD, SpFundamentalRepresentation, CoeffReal > SpWilsonImplD; // Double
typedef WilsonImpl<vComplex, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplR; // Real.. whichever prec
typedef WilsonImpl<vComplexF, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplF; // Float
typedef WilsonImpl<vComplexD, SpTwoIndexAntiSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexAntiSymmetricImplD; // Double
typedef WilsonImpl<vComplex, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplR; // Real.. whichever prec
typedef WilsonImpl<vComplexF, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplF; // Float
typedef WilsonImpl<vComplexD, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonTwoIndexSymmetricImplD; // Double
typedef WilsonImpl<vComplex, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplR; // Real.. whichever prec // adj = 2indx symmetric for Sp(2N)
typedef WilsonImpl<vComplexF, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplF; // Float // adj = 2indx symmetric for Sp(2N)
typedef WilsonImpl<vComplexD, SpTwoIndexSymmetricRepresentation, CoeffReal > SpWilsonAdjImplD; // Double // adj = 2indx symmetric for Sp(2N)
NAMESPACE_END(Grid);

7
Grid/qcd/action/fermion/WilsonKernels.h
View File

@ -52,13 +52,6 @@ public:
typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;
public:
#ifdef GRID_SYCL
#define SYCL_HACK
#endif
#ifdef SYCL_HACK
static void HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor *buf,
int ss,int sU,const SiteSpinor *in, SiteSpinor *out);
#endif
static void DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,

139
Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h
View File

@ -152,58 +152,6 @@ void CayleyFermion5D<Impl>::DminusDag(const FermionField &psi, FermionField &chi
}
}
template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
{
this->Report();
Coordinate latt = GridDefaultLatt();
RealD volume = this->Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = this->_FourDimGrid->_Nprocessors;
if ( M5Dcalls > 0 ) {
std::cout << GridLogMessage << "#### M5D calls report " << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D Number of M5D Calls : " << M5Dcalls << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls : " << M5Dtime / M5Dcalls << " us" << std::endl;
// Flops = 10.0*(Nc*Ns) *Ls*vol
RealD mflops = 10.0*(Nc*Ns)*volume*M5Dcalls/M5Dtime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
// Bytes = sizeof(Real) * (Nc*Ns*Nreim) * Ls * vol * (read+write) (/2 for red black counting)
// read = 2 ( psi[ss+s+1] and psi[ss+s-1] count as 1 )
// write = 1
RealD Gbytes = sizeof(Real) * (Nc*Ns*2) * volume * 3 /2. * 1.e-9;
std::cout << GridLogMessage << "Average bandwidth (GB/s) : " << Gbytes/M5Dtime*M5Dcalls*1.e6 << std::endl;
}
if ( MooeeInvCalls > 0 ) {
std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls : " << MooeeInvCalls << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
#ifdef GRID_CUDA
RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
#else
// Flops = MADD * Ls *Ls *4dvol * spin/colour/complex
RealD mflops = 2.0*24*this->Ls*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
#endif
}
}
template<class Impl> void CayleyFermion5D<Impl>::CayleyZeroCounters(void)
{
this->ZeroCounters();
M5Dflops=0;
M5Dcalls=0;
M5Dtime=0;
MooeeInvFlops=0;
MooeeInvCalls=0;
MooeeInvTime=0;
}
template<class Impl>
void CayleyFermion5D<Impl>::M5D (const FermionField &psi, FermionField &chi)
{
@ -646,7 +594,6 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
assert(mass_plus == mass_minus);
RealD mass = mass_plus;
#if (!defined(GRID_HIP))
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
@ -765,7 +712,7 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
else q_out += C;
}
#endif
}
template <class Impl>
@ -832,7 +779,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
}
#endif
#if (!defined(GRID_HIP))
int tshift = (mu == Nd-1) ? 1 : 0;
unsigned int LLt = GridDefaultLatt()[Tp];
////////////////////////////////////////////////
@ -952,7 +898,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
InsertSlice(L_Q, q_out, s , 0);
}
#endif
}
#undef Pp
#undef Pm
@ -960,88 +905,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
#undef TopRowWithSource
#if 0
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInternalCompute(int dag, int inv,
Vector<iSinglet<Simd> > & Matp,
Vector<iSinglet<Simd> > & Matm)
{
int Ls=this->Ls;
GridBase *grid = this->FermionRedBlackGrid();
int LLs = grid->_rdimensions[0];
if ( LLs == Ls ) {
return; // Not vectorised in 5th direction
}
Eigen::MatrixXcd Pplus = Eigen::MatrixXcd::Zero(Ls,Ls);
Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
for(int s=0;s<Ls;s++){
Pplus(s,s) = bee[s];
Pminus(s,s)= bee[s];
}
for(int s=0;s<Ls-1;s++){
Pminus(s,s+1) = -cee[s];
}
for(int s=0;s<Ls-1;s++){
Pplus(s+1,s) = -cee[s+1];
}
Pplus (0,Ls-1) = mass*cee[0];
Pminus(Ls-1,0) = mass*cee[Ls-1];
Eigen::MatrixXcd PplusMat ;
Eigen::MatrixXcd PminusMat;
if ( inv ) {
PplusMat =Pplus.inverse();
PminusMat=Pminus.inverse();
} else {
PplusMat =Pplus;
PminusMat=Pminus;
}
if(dag){
PplusMat.adjointInPlace();
PminusMat.adjointInPlace();
}
typedef typename SiteHalfSpinor::scalar_type scalar_type;
const int Nsimd=Simd::Nsimd();
Matp.resize(Ls*LLs);
Matm.resize(Ls*LLs);
for(int s2=0;s2<Ls;s2++){
for(int s1=0;s1<LLs;s1++){
int istride = LLs;
int ostride = 1;
Simd Vp;
Simd Vm;
scalar_type *sp = (scalar_type *)&Vp;
scalar_type *sm = (scalar_type *)&Vm;
for(int l=0;l<Nsimd;l++){
if ( switcheroo<Coeff_t>::iscomplex() ) {
sp[l] = PplusMat (l*istride+s1*ostride,s2);
sm[l] = PminusMat(l*istride+s1*ostride,s2);
} else {
// if real
scalar_type tmp;
tmp = PplusMat (l*istride+s1*ostride,s2);
sp[l] = scalar_type(tmp.real(),tmp.real());
tmp = PminusMat(l*istride+s1*ostride,s2);
sm[l] = scalar_type(tmp.real(),tmp.real());
}
}
Matp[LLs*s2+s1] = Vp;
Matm[LLs*s2+s1] = Vm;
}}
}
#endif
NAMESPACE_END(Grid);

51
Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h
View File

@ -63,23 +63,18 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
// 10 = 3 complex mult + 2 complex add
// Flops = 10.0*(Nc*Ns) *Ls*vol (/2 for red black counting)
M5Dcalls++;
M5Dtime-=usecond();
uint64_t nloop = grid->oSites()/Ls;
uint64_t nloop = grid->oSites();
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss= sss*Ls;
uint64_t s = sss%Ls;
uint64_t ss= sss-s;
typedef decltype(coalescedRead(psi[0])) spinor;
spinor tmp1, tmp2;
for(int s=0;s<Ls;s++){
uint64_t idx_u = ss+((s+1)%Ls);
uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5m(tmp1,psi(idx_u));
spProj5p(tmp2,psi(idx_l));
coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
}
uint64_t idx_u = ss+((s+1)%Ls);
uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5m(tmp1,psi(idx_u));
spProj5p(tmp2,psi(idx_l));
coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
});
M5Dtime+=usecond();
}
template<class Impl>
@ -105,23 +100,18 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
int Ls=this->Ls;
// Flops = 6.0*(Nc*Ns) *Ls*vol
M5Dcalls++;
M5Dtime-=usecond();
uint64_t nloop = grid->oSites()/Ls;
uint64_t nloop = grid->oSites();
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss=sss*Ls;
uint64_t s = sss%Ls;
uint64_t ss= sss-s;
typedef decltype(coalescedRead(psi[0])) spinor;
spinor tmp1,tmp2;
for(int s=0;s<Ls;s++){
uint64_t idx_u = ss+((s+1)%Ls);
uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5p(tmp1,psi(idx_u));
spProj5m(tmp2,psi(idx_l));
coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
}
uint64_t idx_u = ss+((s+1)%Ls);
uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5p(tmp1,psi(idx_u));
spProj5m(tmp2,psi(idx_l));
coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
});
M5Dtime+=usecond();
}
template<class Impl>
@ -142,8 +132,6 @@ CayleyFermion5D<Impl>::MooeeInv (const FermionField &psi_i, FermionField &chi
auto pleem = & leem[0];
auto pueem = & ueem[0];
MooeeInvCalls++;
MooeeInvTime-=usecond();
uint64_t nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss=sss*Ls;
@ -180,8 +168,6 @@ CayleyFermion5D<Impl>::MooeeInv (const FermionField &psi_i, FermionField &chi
coalescedWrite(chi[ss+s],res);
}
});
MooeeInvTime+=usecond();
}
@ -204,10 +190,6 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
assert(psi.Checkerboard() == psi.Checkerboard());
MooeeInvCalls++;
MooeeInvTime-=usecond();
uint64_t nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss=sss*Ls;
@ -244,7 +226,6 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
coalescedWrite(chi[ss+s],res);
}
});
MooeeInvTime+=usecond();
}

13
Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h
View File

@ -94,10 +94,6 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
d_p[ss] = diag[s];
}}
M5Dcalls++;
M5Dtime-=usecond();
assert(Nc==3);
thread_loop( (int ss=0;ss<grid->oSites();ss+=LLs),{ // adds LLs
@ -198,7 +194,6 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
}
#endif
});
M5Dtime+=usecond();
}
template<class Impl>
@ -242,8 +237,6 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
d_p[ss] = diag[s];
}}
M5Dcalls++;
M5Dtime-=usecond();
thread_loop( (int ss=0;ss<grid->oSites();ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
@ -339,7 +332,6 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
}
#endif
});
M5Dtime+=usecond();
}
@ -813,9 +805,6 @@ CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,
}
assert(_Matp->size()==Ls*LLs);
MooeeInvCalls++;
MooeeInvTime-=usecond();
if ( switcheroo<Coeff_t>::iscomplex() ) {
thread_loop( (auto site=0;site<vol;site++),{
MooeeInternalZAsm(psi,chi,LLs,site,*_Matp,*_Matm);
@ -825,7 +814,7 @@ CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,
MooeeInternalAsm(psi,chi,LLs,site,*_Matp,*_Matm);
});
}
MooeeInvTime+=usecond();
}
NAMESPACE_END(Grid);

12
Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionCache.h
View File

@ -54,8 +54,6 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
auto pupper = &upper[0];
auto plower = &lower[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
auto nloop=grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
@ -71,7 +69,6 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
}
});
this->M5Dtime += usecond();
}
template<class Impl>
@ -91,8 +88,6 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
auto plower = &lower[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
auto nloop=grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
@ -108,7 +103,6 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
}
});
this->M5Dtime += usecond();
}
template<class Impl>
@ -127,8 +121,6 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
auto pleem = & this->leem[0];
auto pueem = & this->ueem[0];
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
uint64_t nloop=grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss=sss*Ls;
@ -164,7 +156,6 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
coalescedWrite(chi[ss+s],res);
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
@ -185,8 +176,6 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
assert(psi.Checkerboard() == psi.Checkerboard());
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
auto nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss=sss*Ls;
@ -223,7 +212,6 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
}
});
this->MooeeInvTime += usecond();
}
NAMESPACE_END(Grid);

75
Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h
View File

@ -298,45 +298,33 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
int LLs = in.Grid()->_rdimensions[0];
int len = U.Grid()->oSites();
DhopFaceTime-=usecond();
st.Prepare();
st.HaloGather(in,compressor);
DhopFaceTime+=usecond();
DhopCommTime -=usecond();
std::vector<std::vector<CommsRequest_t> > requests;
st.CommunicateBegin(requests);
// st.HaloExchangeOptGather(in,compressor); // Wilson compressor
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Remove explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
DhopComputeTime-=usecond();
{
int interior=1;
int exterior=0;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime+=usecond();
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
DhopComputeTime2-=usecond();
{
int interior=0;
int exterior=1;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime2+=usecond();
}
template<class Impl>
@ -347,22 +335,14 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
Compressor compressor;
int LLs = in.Grid()->_rdimensions[0];
//double t1=usecond();
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
st.HaloExchange(in,compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
{
int interior=1;
int exterior=1;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
}
/*CHANGE END*/
@ -371,7 +351,6 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=1;
conformable(in.Grid(),FermionRedBlackGrid()); // verifies half grid
conformable(in.Grid(),out.Grid()); // drops the cb check
@ -383,7 +362,6 @@ void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionFie
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=1;
conformable(in.Grid(),FermionRedBlackGrid()); // verifies half grid
conformable(in.Grid(),out.Grid()); // drops the cb check
@ -395,7 +373,6 @@ void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionFie
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=2;
conformable(in.Grid(),FermionGrid()); // verifies full grid
conformable(in.Grid(),out.Grid());
@ -404,58 +381,6 @@ void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField
DhopInternal(Stencil,Lebesgue,Umu,UUUmu,in,out,dag);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::Report(void)
{
Coordinate latt = GridDefaultLatt();
RealD volume = Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_FourDimGrid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DhopFaceTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
/////////////////////////////////////////////////////////////////////////
// Implement the general interface. Here we use SAME mass on all slices
/////////////////////////////////////////////////////////////////////////

79
Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermionImplementation.h
View File

@ -334,7 +334,6 @@ void ImprovedStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionF
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=2;
conformable(in.Grid(), _grid); // verifies full grid
conformable(in.Grid(), out.Grid());
@ -346,7 +345,6 @@ void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=1;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
@ -359,7 +357,6 @@ void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=1;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
@ -418,47 +415,33 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
Compressor compressor;
int len = U.Grid()->oSites();
DhopTotalTime -= usecond();
DhopFaceTime -= usecond();
st.Prepare();
st.HaloGather(in,compressor);
DhopFaceTime += usecond();
DhopCommTime -=usecond();
std::vector<std::vector<CommsRequest_t> > requests;
st.CommunicateBegin(requests);
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);
DhopFaceTime+= usecond();
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Removed explicit thread comms
//////////////////////////////////////////////////////////////////////////////////////////////////////
DhopComputeTime -= usecond();
{
int interior=1;
int exterior=0;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
// First to enter, last to leave timing
DhopFaceTime -= usecond();
st.CommsMerge(compressor);
DhopFaceTime -= usecond();
DhopComputeTime2 -= usecond();
{
int interior=0;
int exterior=1;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime2 += usecond();
}
@ -471,78 +454,16 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Le
{
assert((dag == DaggerNo) || (dag == DaggerYes));
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
Compressor compressor;
st.HaloExchange(in, compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
{
int interior=1;
int exterior=1;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
};
////////////////////////////////////////////////////////////////
// Reporting
////////////////////////////////////////////////////////////////
template<class Impl>
void ImprovedStaggeredFermion<Impl>::Report(void)
{
Coordinate latt = _grid->GlobalDimensions();
RealD volume = 1; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_grid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void ImprovedStaggeredFermion<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DhopFaceTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
////////////////////////////////////////////////////////
// Conserved current - not yet implemented.
////////////////////////////////////////////////////////

32
Grid/qcd/action/fermion/implementation/MobiusEOFAFermionCache.h
View File

@ -55,9 +55,6 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
auto plower = &lower[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
int nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss = sss*Ls;
@ -73,7 +70,6 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
}
});
this->M5Dtime += usecond();
}
template<class Impl>
@ -99,9 +95,6 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
auto pshift_coeffs = &shift_coeffs[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
int nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss = sss*Ls;
@ -122,7 +115,6 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
}
});
this->M5Dtime += usecond();
}
template<class Impl>
@ -143,9 +135,6 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
auto plower = &lower[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
int nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(), {
uint64_t ss = sss*Ls;
@ -161,8 +150,6 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
}
});
this->M5Dtime += usecond();
}
template<class Impl>
@ -186,9 +173,6 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
auto pshift_coeffs = &shift_coeffs[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
auto pm = this->pm;
int nloop = grid->oSites()/Ls;
@ -217,7 +201,6 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
}
});
this->M5Dtime += usecond();
}
template<class Impl>
@ -237,9 +220,6 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
if(this->shift != 0.0){ MooeeInv_shift(psi_i,chi_i); return; }
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
int nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss=sss*Ls;
@ -277,7 +257,6 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
@ -297,8 +276,6 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
auto pueem= & this->ueem[0];
auto pMooeeInv_shift_lc = &MooeeInv_shift_lc[0];
auto pMooeeInv_shift_norm = &MooeeInv_shift_norm[0];
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
int nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
@ -343,7 +320,6 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
@ -363,9 +339,6 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
auto pleem= & this->leem[0];
auto pueem= & this->ueem[0];
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
int nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss=sss*Ls;
@ -402,7 +375,6 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
coalescedWrite(chi[ss+s],res);
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
@ -423,9 +395,6 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
auto pMooeeInvDag_shift_lc = &MooeeInvDag_shift_lc[0];
auto pMooeeInvDag_shift_norm = &MooeeInvDag_shift_norm[0];
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
int nloop = grid->oSites()/Ls;
accelerator_for(sss,nloop,Simd::Nsimd(),{
uint64_t ss=sss*Ls;
@ -469,7 +438,6 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
}
});
this->MooeeInvTime += usecond();
}
NAMESPACE_END(Grid);

79
Grid/qcd/action/fermion/implementation/NaiveStaggeredFermionImplementation.h
View File

@ -263,7 +263,6 @@ void NaiveStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionFiel
template <class Impl>
void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=2;
conformable(in.Grid(), _grid); // verifies full grid
conformable(in.Grid(), out.Grid());
@ -275,7 +274,6 @@ void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=1;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
@ -288,7 +286,6 @@ void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &o
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=1;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
@ -345,47 +342,33 @@ void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, L
Compressor compressor;
int len = U.Grid()->oSites();
DhopTotalTime -= usecond();
DhopFaceTime -= usecond();
st.Prepare();
st.HaloGather(in,compressor);
DhopFaceTime += usecond();
DhopCommTime -=usecond();
std::vector<std::vector<CommsRequest_t> > requests;
st.CommunicateBegin(requests);
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);
DhopFaceTime+= usecond();
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Removed explicit thread comms
//////////////////////////////////////////////////////////////////////////////////////////////////////
DhopComputeTime -= usecond();
{
int interior=1;
int exterior=0;
Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
// First to enter, last to leave timing
DhopFaceTime -= usecond();
st.CommsMerge(compressor);
DhopFaceTime -= usecond();
DhopComputeTime2 -= usecond();
{
int interior=0;
int exterior=1;
Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
}
DhopComputeTime2 += usecond();
}
template <class Impl>
@ -396,78 +379,16 @@ void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Lebes
{
assert((dag == DaggerNo) || (dag == DaggerYes));
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
Compressor compressor;
st.HaloExchange(in, compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
{
int interior=1;
int exterior=1;
Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
};
////////////////////////////////////////////////////////////////
// Reporting
////////////////////////////////////////////////////////////////
template<class Impl>
void NaiveStaggeredFermion<Impl>::Report(void)
{
Coordinate latt = _grid->GlobalDimensions();
RealD volume = 1; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_grid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "NaiveStaggeredFermion StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "NaiveStaggeredFermion StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void NaiveStaggeredFermion<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DhopFaceTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
////////////////////////////////////////////////////////
// Conserved current - not yet implemented.
////////////////////////////////////////////////////////

292
Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
View File

@ -60,8 +60,13 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
UmuOdd (_FourDimRedBlackGrid),
Lebesgue(_FourDimGrid),
LebesgueEvenOdd(_FourDimRedBlackGrid),
_tmp(&FiveDimRedBlackGrid)
_tmp(&FiveDimRedBlackGrid),
Dirichlet(0)
{
Stencil.lo = &Lebesgue;
StencilEven.lo = &LebesgueEvenOdd;
StencilOdd.lo = &LebesgueEvenOdd;
// some assertions
assert(FiveDimGrid._ndimension==5);
assert(FourDimGrid._ndimension==4);
@ -91,6 +96,19 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
assert(FourDimRedBlackGrid._simd_layout[d] ==FourDimGrid._simd_layout[d]);
}
if ( p.dirichlet.size() == Nd+1) {
Coordinate block = p.dirichlet;
if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
Dirichlet = 1;
std::cout << GridLogMessage << " WilsonFermion: non-trivial Dirichlet condition "<< block << std::endl;
std::cout << GridLogMessage << " WilsonFermion: partial Dirichlet "<< p.partialDirichlet << std::endl;
Block = block;
}
} else {
Coordinate block(Nd+1,0);
Block = block;
}
if (Impl::LsVectorised) {
int nsimd = Simd::Nsimd();
@ -125,99 +143,38 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
StencilEven.BuildSurfaceList(LLs,vol4);
StencilOdd.BuildSurfaceList(LLs,vol4);
// std::cout << GridLogMessage << " SurfaceLists "<< Stencil.surface_list.size()
// <<" " << StencilEven.surface_list.size()<<std::endl;
}
template<class Impl>
void WilsonFermion5D<Impl>::Report(void)
{
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
RealD volume = Ls;
Coordinate latt = _FourDimGrid->GlobalDimensions();
for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
if ( DhopCalls > 0 ) {
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Number of DhopEO Calls : " << DhopCalls << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D TotalTime /Calls : " << DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D CommTime /Calls : " << DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D FaceTime /Calls : " << DhopFaceTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D ComputeTime1/Calls : " << DhopComputeTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D ComputeTime2/Calls : " << DhopComputeTime2/ DhopCalls << " us" << std::endl;
// Average the compute time
_FourDimGrid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1344*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1344*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
}
if ( DerivCalls > 0 ) {
std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Number of Deriv Calls : " <<DerivCalls <<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D CommTime/Calls : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D ComputeTime/Calls : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Dhop ComputeTime/Calls : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NP << std::endl;
RealD Fullmflops = 144*volume*DerivCalls/(DerivDhopComputeTime+DerivCommTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NP << std::endl; }
if (DerivCalls > 0 || DhopCalls > 0){
std::cout << GridLogMessage << "WilsonFermion5D Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "WilsonFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd" <<std::endl; StencilOdd.Report();
}
if ( DhopCalls > 0){
std::cout << GridLogMessage << "WilsonFermion5D Stencil Reporti()" <<std::endl; Stencil.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion5D StencilEven Reporti()"<<std::endl; StencilEven.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd Reporti()" <<std::endl; StencilOdd.Reporti(DhopCalls);
}
}
template<class Impl>
void WilsonFermion5D<Impl>::ZeroCounters(void) {
DhopCalls = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DhopComputeTime2= 0;
DhopFaceTime = 0;
DhopTotalTime = 0;
DerivCalls = 0;
DerivCommTime = 0;
DerivComputeTime = 0;
DerivDhopComputeTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
Stencil.ZeroCountersi();
StencilEven.ZeroCountersi();
StencilOdd.ZeroCountersi();
}
template<class Impl>
void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
{
GaugeField HUmu(_Umu.Grid());
HUmu = _Umu*(-0.5);
if ( Dirichlet ) {
if ( this->Params.partialDirichlet ) {
std::cout << GridLogMessage << " partialDirichlet BCs " <<Block<<std::endl;
} else {
std::cout << GridLogMessage << " FULL Dirichlet BCs " <<Block<<std::endl;
}
std:: cout << GridLogMessage << "Checking block size multiple of rank boundaries for Dirichlet"<<std::endl;
for(int d=0;d<Nd;d++) {
int GaugeBlock = Block[d+1];
int ldim=GaugeGrid()->LocalDimensions()[d];
if (GaugeBlock) assert( (GaugeBlock%ldim)==0);
}
if (!this->Params.partialDirichlet) {
std::cout << GridLogMessage << " Dirichlet filtering gauge field BCs block " <<Block<<std::endl;
Coordinate GaugeBlock(Nd);
for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
DirichletFilter<GaugeField> Filter(GaugeBlock);
Filter.applyFilter(HUmu);
} else {
std::cout << GridLogMessage << " Dirichlet "<< Dirichlet << " NOT filtered gauge field" <<std::endl;
}
}
Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
pickCheckerboard(Even,UmuEven,Umu);
pickCheckerboard(Odd ,UmuOdd,Umu);
@ -259,7 +216,6 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
const FermionField &B,
int dag)
{
DerivCalls++;
assert((dag==DaggerNo) ||(dag==DaggerYes));
conformable(st.Grid(),A.Grid());
@ -270,15 +226,12 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
FermionField Btilde(B.Grid());
FermionField Atilde(B.Grid());
DerivCommTime-=usecond();
st.HaloExchange(B,compressor);
DerivCommTime+=usecond();
Atilde=A;
int LLs = B.Grid()->_rdimensions[0];
DerivComputeTime-=usecond();
for (int mu = 0; mu < Nd; mu++) {
////////////////////////////////////////////////////////////////////////
// Flip gamma if dag
@ -290,8 +243,6 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
// Call the single hop
////////////////////////
DerivDhopComputeTime -= usecond();
int Usites = U.Grid()->oSites();
Kernels::DhopDirKernel(st, U, st.CommBuf(), Ls, Usites, B, Btilde, mu,gamma);
@ -299,10 +250,8 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
////////////////////////////
// spin trace outer product
////////////////////////////
DerivDhopComputeTime += usecond();
Impl::InsertForce5D(mat, Btilde, Atilde, mu);
}
DerivComputeTime += usecond();
}
template<class Impl>
@ -360,12 +309,10 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
DhopTotalTime-=usecond();
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,in,out,dag);
else
DhopInternalSerialComms(st,lo,U,in,out,dag);
DhopTotalTime+=usecond();
}
@ -374,6 +321,7 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
GRID_TRACE("DhopInternalOverlappedComms");
Compressor compressor(dag);
int LLs = in.Grid()->_rdimensions[0];
@ -382,53 +330,57 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
/////////////////////////////
// Start comms // Gather intranode and extra node differentiated??
/////////////////////////////
DhopFaceTime-=usecond();
st.HaloExchangeOptGather(in,compressor);
DhopFaceTime+=usecond();
DhopCommTime -=usecond();
{
GRID_TRACE("Gather");
st.HaloExchangeOptGather(in,compressor); // Put the barrier in the routine
}
std::vector<std::vector<CommsRequest_t> > requests;
auto id=traceStart("Communicate overlapped");
st.CommunicateBegin(requests);
/////////////////////////////
// Overlap with comms
/////////////////////////////
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
{
GRID_TRACE("MergeSHM");
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
}
/////////////////////////////
// do the compute interior
/////////////////////////////
int Opt = WilsonKernelsStatic::Opt; // Why pass this. Kernels should know
DhopComputeTime-=usecond();
if (dag == DaggerYes) {
GRID_TRACE("DhopDagInterior");
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
} else {
GRID_TRACE("DhopInterior");
Kernels::DhopKernel (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
}
DhopComputeTime+=usecond();
/////////////////////////////
// Complete comms
/////////////////////////////
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
traceStop(id);
/////////////////////////////
// do the compute exterior
/////////////////////////////
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
{
GRID_TRACE("Merge");
st.CommsMerge(compressor);
}
DhopComputeTime2-=usecond();
if (dag == DaggerYes) {
GRID_TRACE("DhopDagExterior");
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
} else {
GRID_TRACE("DhopExterior");
Kernels::DhopKernel (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
}
DhopComputeTime2+=usecond();
}
@ -438,29 +390,30 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOr
const FermionField &in,
FermionField &out,int dag)
{
GRID_TRACE("DhopInternalSerialComms");
Compressor compressor(dag);
int LLs = in.Grid()->_rdimensions[0];
{
GRID_TRACE("HaloExchange");
st.HaloExchangeOpt(in,compressor);
}
DhopCommTime-=usecond();
st.HaloExchangeOpt(in,compressor);
DhopCommTime+=usecond();
DhopComputeTime-=usecond();
int Opt = WilsonKernelsStatic::Opt;
if (dag == DaggerYes) {
GRID_TRACE("DhopDag");
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
} else {
GRID_TRACE("Dhop");
Kernels::DhopKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
}
DhopComputeTime+=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
{
DhopCalls++;
conformable(in.Grid(),FermionRedBlackGrid()); // verifies half grid
conformable(in.Grid(),out.Grid()); // drops the cb check
@ -472,7 +425,6 @@ void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int
template<class Impl>
void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
{
DhopCalls++;
conformable(in.Grid(),FermionRedBlackGrid()); // verifies half grid
conformable(in.Grid(),out.Grid()); // drops the cb check
@ -484,7 +436,6 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
template<class Impl>
void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=2;
conformable(in.Grid(),FermionGrid()); // verifies full grid
conformable(in.Grid(),out.Grid());
@ -539,12 +490,17 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
LatComplex sk(_grid); sk = Zero();
LatComplex sk2(_grid); sk2= Zero();
LatComplex W(_grid); W= Zero();
LatComplex a(_grid); a= Zero();
LatComplex one (_grid); one = ScalComplex(1.0,0.0);
LatComplex cosha(_grid);
LatComplex kmu(_grid);
LatComplex Wea(_grid);
LatComplex Wema(_grid);
LatComplex ea(_grid);
LatComplex ema(_grid);
LatComplex eaLs(_grid);
LatComplex emaLs(_grid);
LatComplex ea2Ls(_grid);
LatComplex ema2Ls(_grid);
LatComplex sinha(_grid);
LatComplex sinhaLs(_grid);
LatComplex coshaLs(_grid);
@ -579,39 +535,29 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
////////////////////////////////////////////
cosha = (one + W*W + sk) / (abs(W)*2.0);
// FIXME Need a Lattice acosh
{
autoView(cosha_v,cosha,CpuRead);
autoView(a_v,a,CpuWrite);
for(int idx=0;idx<_grid->lSites();idx++){
Coordinate lcoor(Nd);
Tcomplex cc;
// RealD sgn;
_grid->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(cc,cosha_v,lcoor);
assert((double)real(cc)>=1.0);
assert(fabs((double)imag(cc))<=1.0e-15);
cc = ScalComplex(::acosh(real(cc)),0.0);
pokeLocalSite(cc,a_v,lcoor);
}
}
Wea = ( exp( a) * abs(W) );
Wema= ( exp(-a) * abs(W) );
sinha = 0.5*(exp( a) - exp(-a));
sinhaLs = 0.5*(exp( a*Ls) - exp(-a*Ls));
coshaLs = 0.5*(exp( a*Ls) + exp(-a*Ls));
ea = (cosha + sqrt(cosha*cosha-one));
ema= (cosha - sqrt(cosha*cosha-one));
eaLs = pow(ea,Ls);
emaLs= pow(ema,Ls);
ea2Ls = pow(ea,2.0*Ls);
ema2Ls= pow(ema,2.0*Ls);
Wea= abs(W) * ea;
Wema= abs(W) * ema;
// a=log(ea);
sinha = 0.5*(ea - ema);
sinhaLs = 0.5*(eaLs-emaLs);
coshaLs = 0.5*(eaLs+emaLs);
A = one / (abs(W) * sinha * 2.0) * one / (sinhaLs * 2.0);
F = exp( a*Ls) * (one - Wea + (Wema - one) * mass*mass);
F = F + exp(-a*Ls) * (Wema - one + (one - Wea) * mass*mass);
F = eaLs * (one - Wea + (Wema - one) * mass*mass);
F = F + emaLs * (Wema - one + (one - Wea) * mass*mass);
F = F - abs(W) * sinha * 4.0 * mass;
Bpp = (A/F) * (exp(-a*Ls*2.0) - one) * (one - Wema) * (one - mass*mass * one);
Bmm = (A/F) * (one - exp(a*Ls*2.0)) * (one - Wea) * (one - mass*mass * one);
App = (A/F) * (exp(-a*Ls*2.0) - one) * exp(-a) * (exp(-a) - abs(W)) * (one - mass*mass * one);
Amm = (A/F) * (one - exp(a*Ls*2.0)) * exp(a) * (exp(a) - abs(W)) * (one - mass*mass * one);
Bpp = (A/F) * (ema2Ls - one) * (one - Wema) * (one - mass*mass * one);
Bmm = (A/F) * (one - ea2Ls) * (one - Wea) * (one - mass*mass * one);
App = (A/F) * (ema2Ls - one) * ema * (ema - abs(W)) * (one - mass*mass * one);
Amm = (A/F) * (one - ea2Ls) * ea * (ea - abs(W)) * (one - mass*mass * one);
ABpm = (A/F) * abs(W) * sinha * 2.0 * (one + mass * coshaLs * 2.0 + mass*mass * one);
//P+ source, P- source
@ -634,29 +580,29 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
buf1_4d = Zero();
ExtractSlice(buf1_4d, PRsource, (tt-1), 0);
//G(s,t)
bufR_4d = bufR_4d + A * exp(a*Ls) * exp(-a*f) * signW * buf1_4d + A * exp(-a*Ls) * exp(a*f) * signW * buf1_4d;
bufR_4d = bufR_4d + A * eaLs * pow(ema,f) * signW * buf1_4d + A * emaLs * pow(ea,f) * signW * buf1_4d;
//A++*exp(a(s+t))
bufR_4d = bufR_4d + App * exp(a*ss) * exp(a*tt) * signW * buf1_4d ;
bufR_4d = bufR_4d + App * pow(ea,ss) * pow(ea,tt) * signW * buf1_4d ;
//A+-*exp(a(s-t))
bufR_4d = bufR_4d + ABpm * exp(a*ss) * exp(-a*tt) * signW * buf1_4d ;
bufR_4d = bufR_4d + ABpm * pow(ea,ss) * pow(ema,tt) * signW * buf1_4d ;
//A-+*exp(a(-s+t))
bufR_4d = bufR_4d + ABpm * exp(-a*ss) * exp(a*tt) * signW * buf1_4d ;
bufR_4d = bufR_4d + ABpm * pow(ema,ss) * pow(ea,tt) * signW * buf1_4d ;
//A--*exp(a(-s-t))
bufR_4d = bufR_4d + Amm * exp(-a*ss) * exp(-a*tt) * signW * buf1_4d ;
bufR_4d = bufR_4d + Amm * pow(ema,ss) * pow(ema,tt) * signW * buf1_4d ;
//GL
buf2_4d = Zero();
ExtractSlice(buf2_4d, PLsource, (tt-1), 0);
//G(s,t)
bufL_4d = bufL_4d + A * exp(a*Ls) * exp(-a*f) * signW * buf2_4d + A * exp(-a*Ls) * exp(a*f) * signW * buf2_4d;
bufL_4d = bufL_4d + A * eaLs * pow(ema,f) * signW * buf2_4d + A * emaLs * pow(ea,f) * signW * buf2_4d;
//B++*exp(a(s+t))
bufL_4d = bufL_4d + Bpp * exp(a*ss) * exp(a*tt) * signW * buf2_4d ;
bufL_4d = bufL_4d + Bpp * pow(ea,ss) * pow(ea,tt) * signW * buf2_4d ;
//B+-*exp(a(s-t))
bufL_4d = bufL_4d + ABpm * exp(a*ss) * exp(-a*tt) * signW * buf2_4d ;
bufL_4d = bufL_4d + ABpm * pow(ea,ss) * pow(ema,tt) * signW * buf2_4d ;
//B-+*exp(a(-s+t))
bufL_4d = bufL_4d + ABpm * exp(-a*ss) * exp(a*tt) * signW * buf2_4d ;
bufL_4d = bufL_4d + ABpm * pow(ema,ss) * pow(ea,tt) * signW * buf2_4d ;
//B--*exp(a(-s-t))
bufL_4d = bufL_4d + Bmm * exp(-a*ss) * exp(-a*tt) * signW * buf2_4d ;
bufL_4d = bufL_4d + Bmm * pow(ema,ss) * pow(ema,tt) * signW * buf2_4d ;
}
InsertSlice(bufR_4d, GR, (ss-1), 0);
InsertSlice(bufL_4d, GL, (ss-1), 0);
@ -775,28 +721,12 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt(FermionField &out,const Fe
W = one - M5 + sk2;
////////////////////////////////////////////
// Cosh alpha -> alpha
// Cosh alpha -> exp(+/- alpha)
////////////////////////////////////////////
cosha = (one + W*W + sk) / (abs(W)*2.0);
// FIXME Need a Lattice acosh
{
autoView(cosha_v,cosha,CpuRead);
autoView(a_v,a,CpuWrite);
for(int idx=0;idx<_grid->lSites();idx++){
Coordinate lcoor(Nd);
Tcomplex cc;
// RealD sgn;
_grid->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(cc,cosha_v,lcoor);
assert((double)real(cc)>=1.0);
assert(fabs((double)imag(cc))<=1.0e-15);
cc = ScalComplex(::acosh(real(cc)),0.0);
pokeLocalSite(cc,a_v,lcoor);
}}
Wea = ( exp( a) * abs(W) );
Wema= ( exp(-a) * abs(W) );
Wea = abs(W)*(cosha + sqrt(cosha*cosha-one));
Wema= abs(W)*(cosha - sqrt(cosha*cosha-one));
num = num + ( one - Wema ) * mass * in;
denom= ( Wea - one ) + mass*mass * (one - Wema);

147
Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
View File

@ -60,6 +60,9 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
_tmp(&Hgrid),
anisotropyCoeff(anis)
{
Stencil.lo = &Lebesgue;
StencilEven.lo = &LebesgueEvenOdd;
StencilOdd.lo = &LebesgueEvenOdd;
// Allocate the required comms buffer
ImportGauge(_Umu);
if (anisotropyCoeff.isAnisotropic){
@ -76,91 +79,6 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
StencilOdd.BuildSurfaceList(1,vol4);
}
template<class Impl>
void WilsonFermion<Impl>::Report(void)
{
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
RealD volume = 1;
Coordinate latt = _grid->GlobalDimensions();
for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
if ( DhopCalls > 0 ) {
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "WilsonFermion Number of DhopEO Calls : " << DhopCalls << std::endl;
std::cout << GridLogMessage << "WilsonFermion TotalTime /Calls : " << DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion CommTime /Calls : " << DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion FaceTime /Calls : " << DhopFaceTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion ComputeTime1/Calls : " << DhopComputeTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion ComputeTime2/Calls : " << DhopComputeTime2/ DhopCalls << " us" << std::endl;
// Average the compute time
_grid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1320*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1320*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
}
if ( DerivCalls > 0 ) {
std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
std::cout << GridLogMessage << "WilsonFermion Number of Deriv Calls : " <<DerivCalls <<std::endl;
std::cout << GridLogMessage << "WilsonFermion CommTime/Calls : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
std::cout << GridLogMessage << "WilsonFermion ComputeTime/Calls : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
std::cout << GridLogMessage << "WilsonFermion Dhop ComputeTime/Calls : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
// how to count flops here?
RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
std::cout << GridLogMessage << "Average mflops/s per call ? : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node ? : " << mflops/NP << std::endl;
// how to count flops here?
RealD Fullmflops = 144*volume*DerivCalls/(DerivDhopComputeTime+DerivCommTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) ? : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full) ? : " << Fullmflops/NP << std::endl; }
if (DerivCalls > 0 || DhopCalls > 0){
std::cout << GridLogMessage << "WilsonFermion Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "WilsonFermion StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "WilsonFermion StencilOdd" <<std::endl; StencilOdd.Report();
}
if ( DhopCalls > 0){
std::cout << GridLogMessage << "WilsonFermion Stencil Reporti()" <<std::endl; Stencil.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion StencilEven Reporti()"<<std::endl; StencilEven.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion StencilOdd Reporti()" <<std::endl; StencilOdd.Reporti(DhopCalls);
}
}
template<class Impl>
void WilsonFermion<Impl>::ZeroCounters(void) {
DhopCalls = 0; // ok
DhopCommTime = 0;
DhopComputeTime = 0;
DhopComputeTime2= 0;
DhopFaceTime = 0;
DhopTotalTime = 0;
DerivCalls = 0; // ok
DerivCommTime = 0;
DerivComputeTime = 0;
DerivDhopComputeTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
Stencil.ZeroCountersi();
StencilEven.ZeroCountersi();
StencilOdd.ZeroCountersi();
}
template <class Impl>
void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
{
@ -320,7 +238,6 @@ template <class Impl>
void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
GaugeField &mat, const FermionField &A,
const FermionField &B, int dag) {
DerivCalls++;
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor(dag);
@ -329,11 +246,8 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
FermionField Atilde(B.Grid());
Atilde = A;
DerivCommTime-=usecond();
st.HaloExchange(B, compressor);
DerivCommTime+=usecond();
DerivComputeTime-=usecond();
for (int mu = 0; mu < Nd; mu++) {
////////////////////////////////////////////////////////////////////////
// Flip gamma (1+g)<->(1-g) if dag
@ -341,7 +255,6 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
int gamma = mu;
if (!dag) gamma += Nd;
DerivDhopComputeTime -= usecond();
int Ls=1;
Kernels::DhopDirKernel(st, U, st.CommBuf(), Ls, B.Grid()->oSites(), B, Btilde, mu, gamma);
@ -349,9 +262,7 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
// spin trace outer product
//////////////////////////////////////////////////
Impl::InsertForce4D(mat, Btilde, Atilde, mu);
DerivDhopComputeTime += usecond();
}
DerivComputeTime += usecond();
}
template <class Impl>
@ -398,7 +309,6 @@ void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, co
template <class Impl>
void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=2;
conformable(in.Grid(), _grid); // verifies full grid
conformable(in.Grid(), out.Grid());
@ -410,7 +320,6 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
template <class Impl>
void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
{
DhopCalls++;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
@ -423,7 +332,6 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
template <class Impl>
void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
{
DhopCalls++;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
@ -488,14 +396,12 @@ void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
const FermionField &in,
FermionField &out, int dag)
{
DhopTotalTime-=usecond();
#ifdef GRID_OMP
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,in,out,dag);
else
#endif
DhopInternalSerial(st,lo,U,in,out,dag);
DhopTotalTime+=usecond();
}
template <class Impl>
@ -504,6 +410,7 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
const FermionField &in,
FermionField &out, int dag)
{
GRID_TRACE("DhopOverlapped");
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor(dag);
@ -514,53 +421,55 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
/////////////////////////////
std::vector<std::vector<CommsRequest_t> > requests;
st.Prepare();
DhopFaceTime-=usecond();
st.HaloGather(in,compressor);
DhopFaceTime+=usecond();
{
GRID_TRACE("Gather");
st.HaloGather(in,compressor);
}
DhopCommTime -=usecond();
tracePush("Communication");
st.CommunicateBegin(requests);
/////////////////////////////
// Overlap with comms
/////////////////////////////
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);
DhopFaceTime+=usecond();
{
GRID_TRACE("MergeSHM");
st.CommsMergeSHM(compressor);
}
/////////////////////////////
// do the compute interior
/////////////////////////////
int Opt = WilsonKernelsStatic::Opt;
DhopComputeTime-=usecond();
if (dag == DaggerYes) {
GRID_TRACE("DhopDagInterior");
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
} else {
GRID_TRACE("DhopInterior");
Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
}
DhopComputeTime+=usecond();
/////////////////////////////
// Complete comms
/////////////////////////////
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
tracePop("Communication");
{
GRID_TRACE("Merge");
st.CommsMerge(compressor);
}
/////////////////////////////
// do the compute exterior
/////////////////////////////
DhopComputeTime2-=usecond();
if (dag == DaggerYes) {
GRID_TRACE("DhopDagExterior");
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
} else {
GRID_TRACE("DhopExterior");
Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
}
DhopComputeTime2+=usecond();
};
@ -570,20 +479,22 @@ void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
const FermionField &in,
FermionField &out, int dag)
{
GRID_TRACE("DhopSerial");
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor(dag);
DhopCommTime-=usecond();
st.HaloExchange(in, compressor);
DhopCommTime+=usecond();
{
GRID_TRACE("HaloExchange");
st.HaloExchange(in, compressor);
}
DhopComputeTime-=usecond();
int Opt = WilsonKernelsStatic::Opt;
if (dag == DaggerYes) {
GRID_TRACE("DhopDag");
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
} else {
GRID_TRACE("Dhop");
Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out);
}
DhopComputeTime+=usecond();
};
/*Change ends */

73
Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
View File

@ -72,20 +72,15 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
if (SE->_is_local) { \
int perm= SE->_permute; \
auto tmp = coalescedReadPermute(in[SE->_offset],ptype,perm,lane); \
spProj(chi,tmp); \
} else if ( st.same_node[Dir] ) { \
chi = coalescedRead(buf[SE->_offset],lane); \
} \
acceleratorSynchronise(); \
if (SE->_is_local || st.same_node[Dir] ) { \
Impl::multLink(Uchi, U[sU], chi, Dir, SE, st); \
Recon(result, Uchi); \
} \
spProj(chi,tmp); \
Impl::multLink(Uchi, U[sU], chi, Dir, SE, st); \
Recon(result, Uchi); \
} \
acceleratorSynchronise();
#define GENERIC_STENCIL_LEG_EXT(Dir,spProj,Recon) \
SE = st.GetEntry(ptype, Dir, sF); \
if ((!SE->_is_local) && (!st.same_node[Dir]) ) { \
if (!SE->_is_local ) { \
auto chi = coalescedRead(buf[SE->_offset],lane); \
Impl::multLink(Uchi, U[sU], chi, Dir, SE, st); \
Recon(result, Uchi); \
@ -416,19 +411,6 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
#undef LoopBody
}
#define KERNEL_CALL_TMP(A) \
const uint64_t NN = Nsite*Ls; \
auto U_p = & U_v[0]; \
auto in_p = & in_v[0]; \
auto out_p = & out_v[0]; \
auto st_p = st_v._entries_p; \
auto st_perm = st_v._permute_type; \
accelerator_forNB( ss, NN, Simd::Nsimd(), { \
int sF = ss; \
int sU = ss/Ls; \
WilsonKernels<Impl>::A(st_perm,st_p,U_p,buf,sF,sU,in_p,out_p); \
}); \
accelerator_barrier();
#define KERNEL_CALLNB(A) \
const uint64_t NN = Nsite*Ls; \
@ -440,12 +422,34 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
#define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
#define KERNEL_CALL_EXT(A) \
const uint64_t sz = st.surface_list.size(); \
auto ptr = &st.surface_list[0]; \
accelerator_forNB( ss, sz, Simd::Nsimd(), { \
int sF = ptr[ss]; \
int sU = sF/Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v); \
}); \
accelerator_barrier();
#define ASM_CALL(A) \
thread_for( ss, Nsite, { \
thread_for( sss, Nsite, { \
int ss = st.lo->Reorder(sss); \
int sU = ss; \
int sF = ss*Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v); \
});
#define ASM_CALL_SLICE(A) \
auto grid = in.Grid() ; \
int nt = grid->LocalDimensions()[4]; \
int nxyz = Nsite/nt ; \
for(int t=0;t<nt;t++){ \
thread_for( sss, nxyz, { \
int ss = t*nxyz+sss; \
int sU = ss; \
int sF = ss*Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v); \
});}
template <class Impl>
void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
@ -459,11 +463,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
if( interior && exterior ) {
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSite); return;}
#ifdef SYCL_HACK
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl); return; }
#else
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite); return;}
#endif
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSite); return;}
#endif
@ -474,8 +474,10 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteInt); return;}
#endif
} else if( exterior ) {
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteExt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt); return;}
// dependent on result of merge
acceleratorFenceComputeStream();
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL_EXT(GenericDhopSiteExt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteExt); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteExt); return;}
#endif
@ -498,21 +500,20 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDag); return;}
#endif
acceleratorFenceComputeStream();
} else if( interior ) {
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt); return;}
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALLNB(GenericDhopSiteDagInt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteDagInt); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDagInt); return;}
#endif
} else if( exterior ) {
// Dependent on result of merge
acceleratorFenceComputeStream();
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt); return;}
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL_EXT(GenericDhopSiteDagExt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_EXT(HandDhopSiteDagExt); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDagExt); return;}
#endif
acceleratorFenceComputeStream();
}
assert(0 && " Kernel optimisation case not covered ");
}

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonCloverFermionInstantiationSpWilsonImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonCloverFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonFermionInstantiationSpWilsonImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonKernelsInstantiationSpWilsonImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplD/WilsonTMFermionInstantiationSpWilsonImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonTMFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplD/impl.h Normal file
View File

@ -0,0 +1 @@
#define IMPLEMENTATION SpWilsonImplD

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonCloverFermionInstantiationSpWilsonImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonCloverFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonFermionInstantiationSpWilsonImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonKernelsInstantiationSpWilsonImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplF/WilsonTMFermionInstantiationSpWilsonImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonTMFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonImplF/impl.h Normal file
View File

@ -0,0 +1 @@
#define IMPLEMENTATION SpWilsonImplF

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonCloverFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonCloverFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/WilsonTMFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonTMFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplD/impl.h Normal file
View File

@ -0,0 +1 @@
#define IMPLEMENTATION SpWilsonTwoIndexAntiSymmetricImplD

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonCloverFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonCloverFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/WilsonTMFermionInstantiationSpWilsonTwoIndexAntiSymmetricImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonTMFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexAntiSymmetricImplF/impl.h Normal file
View File

@ -0,0 +1 @@
#define IMPLEMENTATION SpWilsonTwoIndexAntiSymmetricImplF

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonCloverFermionInstantiationSpWilsonTwoIndexSymmetricImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonCloverFermionInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/SpWilsonTwoIndexSymmetricImplD/WilsonFermionInstantiationSpWilsonTwoIndexSymmetricImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonFermionInstantiation.cc.master

Some files were not shown because too many files have changed in this diff Show More