portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-04 05:54:32 +00:00
Code Releases Activity

Merge branch 'feature/dirichlet' of https://github.com/paboyle/Grid into feature/dirichlet

Browse Source
This commit is contained in:
Peter Boyle
2023-04-03 18:26:11 -04:00
parent 6252ffaf76 58e020b62a
commit fc4db5e963
39 changed files with 820 additions and 115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
Grid/algorithms/Algorithms.h
View File

@@ -55,6 +55,7 @@ NAMESPACE_CHECK(BiCGSTAB);
#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>

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

@@ -191,7 +191,7 @@ public:
std::cout << GridLogMessage << "\tAxpyNorm " << AxpyNormTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
std::cout << GridLogDebug << "\tMobius flop rate " << DwfFlops/ usecs<< " Gflops " <<std::endl;
if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);

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

50
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
@@ -222,8 +240,11 @@ 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 +253,12 @@ 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 +293,11 @@ 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 +306,6 @@ 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

37
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];
@@ -122,6 +138,25 @@ private:
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
//////////////////////////////////////////////////////////////////////

3
Grid/communicator/Communicator_mpi3.cc
View File

@@ -400,9 +400,6 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
{
acceleratorCopySynchronise();
StencilBarrier();// Synch shared memory on a single nodes
int nreq=list.size();
if (nreq==0) return;

5
Grid/communicator/SharedMemoryMPI.cc
View File

@@ -37,10 +37,11 @@ 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
#endif
NAMESPACE_BEGIN(Grid);
#define header "SharedMemoryMpi: "
/*Construct from an MPI communicator*/

40
Grid/cshift/Cshift_common.h
View File

@@ -297,6 +297,30 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
}
}
#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
template <typename T>
T iDivUp(T a, T b) // Round a / b to nearest higher integer value
{ return (a % b != 0) ? (a / b + 1) : (a / b); }
template <typename T>
__global__ void populate_Cshift_table(T* vector, T lo, T ro, T e1, T e2, T stride)
{
int idx = blockIdx.x*blockDim.x + threadIdx.x;
if (idx >= e1*e2) return;
int n, b, o;
n = idx / e2;
b = idx % e2;
o = n*stride + b;
vector[2*idx + 0] = lo + o;
vector[2*idx + 1] = ro + o;
}
#endif
//////////////////////////////////////////////////////
// local to node block strided copies
//////////////////////////////////////////////////////
@@ -321,12 +345,20 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
int ent=0;
if(cbmask == 0x3 ){
#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
ent = e1*e2;
dim3 blockSize(acceleratorThreads());
dim3 gridSize(iDivUp((unsigned int)ent, blockSize.x));
populate_Cshift_table<<<gridSize, blockSize>>>(&Cshift_table[0].first, lo, ro, e1, e2, stride);
accelerator_barrier();
#else
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*stride+b;
Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
}
}
#endif
} else {
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
@@ -377,11 +409,19 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
int ent=0;
if ( cbmask == 0x3 ) {
#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
ent = e1*e2;
dim3 blockSize(acceleratorThreads());
dim3 gridSize(iDivUp((unsigned int)ent, blockSize.x));
populate_Cshift_table<<<gridSize, blockSize>>>(&Cshift_table[0].first, lo, ro, e1, e2, stride);
accelerator_barrier();
#else
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*stride;
Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
}}
#endif
} else {
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){

25
Grid/lattice/Lattice_reduction.h
View File

@@ -153,33 +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;
}

13
Grid/lattice/Lattice_reduction_gpu.h
View File

@@ -211,25 +211,22 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
assert(ok);
Integer smemSize = numThreads * sizeof(sobj);
// UVM seems to be buggy under later CUDA drivers
// This fails on A100 and driver 5.30.02 / CUDA 12.1
// Fails with multiple NVCC versions back to 11.4,
// which worked with earlier drivers.
// Not sure which driver had first fail and this bears checking
// Is awkward as must install multiple driver versions
// 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 >>>(lat, buffer_v, size);
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];
sobj result;
reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
reduceKernel<<< numBlocks, numThreads, smemSize, computeStream >>>(lat, buffer_v, size);
accelerator_barrier();
result = *buffer_v;
#endif

11
Grid/lattice/Lattice_rng.h
View File

@@ -440,17 +440,8 @@ public:
_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
#if 1
assert(rank == _grid->ThisRank() );
#else
//
if (rank != _grid->ThisRank() ){
std::cout <<"rank "<<rank<<" _grid->ThisRank() "<<_grid->ThisRank()<< std::endl;
// exit(-42);
// assert(0);
}
#endif
assert(rank == _grid->ThisRank() );
int l_idx=generator_idx(o_idx,i_idx);
_generators[l_idx] = master_engine;

49
Grid/lattice/Lattice_transfer.h
View File

@@ -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>

11
Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
View File

@@ -463,11 +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(HandDhopSiteSycl); return; }
#else
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite); return;}
#endif
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSite); return;}
#endif
@@ -478,6 +474,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteInt); return;}
#endif
} else if( exterior ) {
acceleratorFenceComputeStream();
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteExt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt); return;}
#ifndef GRID_CUDA
@@ -502,10 +499,9 @@ 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
@@ -516,7 +512,6 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
#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/WilsonImplD2/CayleyFermion5DInstantiationWilsonImplD2.cc
View File

@@ -1 +0,0 @@
../CayleyFermion5DInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/ContinuedFractionFermion5DInstantiationWilsonImplD2.cc
View File

@@ -1 +0,0 @@
../ContinuedFractionFermion5DInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/DomainWallEOFAFermionInstantiationWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/MobiusEOFAFermionInstantiationWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/PartialFractionFermion5DInstantiationWilsonImplD2.cc
View File

@@ -1 +0,0 @@
../PartialFractionFermion5DInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonCloverFermionInstantiationWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonFermion5DInstantiationWilsonImplD2.cc
View File

@@ -1 +0,0 @@
../WilsonFermion5DInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonFermionInstantiationWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonKernelsInstantiationWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/WilsonTMFermionInstantiationWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/WilsonImplD2/impl.h
View File

@@ -1 +0,0 @@
#define IMPLEMENTATION WilsonImplD2

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/CayleyFermion5DInstantiationZWilsonImplD2.cc
View File

@@ -1 +0,0 @@
../CayleyFermion5DInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/ContinuedFractionFermion5DInstantiationZWilsonImplD2.cc
View File

@@ -1 +0,0 @@
../ContinuedFractionFermion5DInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/DomainWallEOFAFermionInstantiationZWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/MobiusEOFAFermionInstantiationZWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/PartialFractionFermion5DInstantiationZWilsonImplD2.cc
View File

@@ -1 +0,0 @@
../PartialFractionFermion5DInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/WilsonFermion5DInstantiationZWilsonImplD2.cc
View File

@@ -1 +0,0 @@
../WilsonFermion5DInstantiation.cc.master

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/WilsonKernelsInstantiationZWilsonImplD2.cc
View File

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

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD2/impl.h
View File

@@ -1 +0,0 @@
#define IMPLEMENTATION ZWilsonImplD2

13
Grid/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h
View File

@@ -112,40 +112,27 @@ NAMESPACE_BEGIN(Grid);
// NumOp == V
// DenOp == M
//
AUDIT();
FermionField etaOdd (NumOp.FermionRedBlackGrid());
FermionField etaEven(NumOp.FermionRedBlackGrid());
FermionField tmp (NumOp.FermionRedBlackGrid());
AUDIT();
pickCheckerboard(Even,etaEven,eta);
AUDIT();
pickCheckerboard(Odd,etaOdd,eta);
AUDIT();
NumOp.ImportGauge(U);
AUDIT();
DenOp.ImportGauge(U);
std::cout << " TwoFlavourRefresh: Imported gauge "<<std::endl;
AUDIT();
SchurDifferentiableOperator<Impl> Mpc(DenOp);
AUDIT();
SchurDifferentiableOperator<Impl> Vpc(NumOp);
AUDIT();
std::cout << " TwoFlavourRefresh: Diff ops "<<std::endl;
AUDIT();
// Odd det factors
Mpc.MpcDag(etaOdd,PhiOdd);
AUDIT();
std::cout << " TwoFlavourRefresh: MpcDag "<<std::endl;
tmp=Zero();
AUDIT();
std::cout << " TwoFlavourRefresh: Zero() guess "<<std::endl;
AUDIT();
HeatbathSolver(Vpc,PhiOdd,tmp);
AUDIT();
std::cout << " TwoFlavourRefresh: Heatbath solver "<<std::endl;
Vpc.Mpc(tmp,PhiOdd);
std::cout << " TwoFlavourRefresh: Mpc "<<std::endl;

21
Grid/qcd/hmc/integrators/Integrator.h
View File

@@ -134,14 +134,12 @@ protected:
double start_force = usecond();
std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] before"<<std::endl;
AUDIT();
as[level].actions.at(a)->deriv_timer_start();
as[level].actions.at(a)->deriv(Us, force); // deriv should NOT include Ta
as[level].actions.at(a)->deriv_timer_stop();
std::cout << GridLogMessage << "AuditForce["<<level<<"]["<<a<<"] after"<<std::endl;
AUDIT();
std::cout << GridLogIntegrator << "Smearing (on/off): " << as[level].actions.at(a)->is_smeared << std::endl;
auto name = as[level].actions.at(a)->action_name();
@@ -382,12 +380,12 @@ public:
Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] before"<<std::endl;
AUDIT();
as[level].actions.at(actionID)->refresh_timer_start();
as[level].actions.at(actionID)->refresh(Us, sRNG, pRNG);
as[level].actions.at(actionID)->refresh_timer_stop();
std::cout << GridLogMessage << "AuditRefresh["<<level<<"]["<<actionID<<"] after"<<std::endl;
AUDIT();
}
// Refresh the higher representation actions
@@ -424,7 +422,7 @@ public:
// Actions
for (int level = 0; level < as.size(); ++level) {
for (int actionID = 0; actionID < as[level].actions.size(); ++actionID) {
AUDIT();
// get gauge field from the SmearingPolicy and
// based on the boolean is_smeared in actionID
Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
@@ -434,7 +432,7 @@ public:
as[level].actions.at(actionID)->S_timer_stop();
std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
H += Hterm;
AUDIT();
}
as[level].apply(S_hireps, Representations, level, H);
}
@@ -447,9 +445,9 @@ public:
void operator()(std::vector<Action<FieldType>*> repr_set, Repr& Rep, int level, RealD& H) {
for (int a = 0; a < repr_set.size(); ++a) {
AUDIT();
RealD Hterm = repr_set.at(a)->Sinitial(Rep.U);
AUDIT();
std::cout << GridLogMessage << "Sinitial Level " << level << " term " << a << " H Hirep = " << Hterm << std::endl;
H += Hterm;
@@ -474,10 +472,10 @@ public:
Field& Us = Smearer.get_U(as[level].actions.at(actionID)->is_smeared);
std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] action eval " << std::endl;
as[level].actions.at(actionID)->S_timer_start();
AUDIT();
Hterm = as[level].actions.at(actionID)->Sinitial(Us);
as[level].actions.at(actionID)->S_timer_stop();
AUDIT();
std::cout << GridLogMessage << "S [" << level << "][" << actionID << "] H = " << Hterm << std::endl;
H += Hterm;
}
@@ -490,7 +488,6 @@ public:
void integrate(Field& U)
{
AUDIT();
// reset the clocks
t_U = 0;
for (int level = 0; level < as.size(); ++level) {
@@ -508,10 +505,8 @@ public:
assert(fabs(t_U - t_P[level]) < 1.0e-6); // must be the same
std::cout << GridLogIntegrator << " times[" << level << "]= " << t_P[level] << " " << t_U << std::endl;
}
AUDIT();
FieldImplementation::Project(U);
AUDIT();
// and that we indeed got to the end of the trajectory
assert(fabs(t_U - Params.trajL) < 1.0e-6);

12
Grid/stencil/Stencil.h
View File

@@ -434,7 +434,6 @@ public:
////////////////////////////////////////////////////////////////////////
void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
{
accelerator_barrier();
for(int i=0;i<Packets.size();i++){
_grid->StencilSendToRecvFromBegin(MpiReqs,
Packets[i].send_buf,
@@ -443,7 +442,6 @@ public:
Packets[i].from_rank,Packets[i].do_recv,
Packets[i].xbytes,Packets[i].rbytes,i);
}
_grid->StencilBarrier();// Synch shared memory on a single nodes
}
void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
@@ -452,6 +450,9 @@ public:
if ( this->partialDirichlet ) DslashLogPartial();
else if ( this->fullDirichlet ) DslashLogDirichlet();
else DslashLogFull();
acceleratorCopySynchronise();
// Everyone agrees we are all done
_grid->StencilBarrier();
}
////////////////////////////////////////////////////////////////////////
// Blocking send and receive. Either sequential or parallel.
@@ -529,7 +530,6 @@ public:
{
_grid->StencilBarrier();// Synch shared memory on a single nodes
// conformable(source.Grid(),_grid);
assert(source.Grid()==_grid);
u_comm_offset=0;
@@ -655,8 +655,8 @@ public:
CommsMerge(decompress,Mergers,Decompressions);
}
template<class decompressor> void CommsMergeSHM(decompressor decompress) {
_grid->StencilBarrier();// Synch shared memory on a single nodes
CommsMerge(decompress,MergersSHM,DecompressionsSHM);
assert(MergersSHM.size()==0);
assert(DecompressionsSHM.size()==0);
}
template<class decompressor>
@@ -665,9 +665,11 @@ public:
for(int i=0;i<mm.size();i++){
decompressor::MergeFace(decompress,mm[i]);
}
if ( mm.size() ) acceleratorFenceComputeStream();
for(int i=0;i<dd.size();i++){
decompressor::DecompressFace(decompress,dd[i]);
}
if ( dd.size() ) acceleratorFenceComputeStream();
}
////////////////////////////////////////
// Set up routines

387
benchmarks/Benchmark_dwf_fp32_paranoid.cc Normal file
View File

@@ -0,0 +1,387 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./benchmarks/Benchmark_dwf.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid.h>
#ifdef GRID_CUDA
#define CUDA_PROFILE
#endif
#ifdef CUDA_PROFILE
#include <cuda_profiler_api.h>
#endif
using namespace std;
using namespace Grid;
template<class d>
struct scal {
d internal;
};
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT
};
int main (int argc, char ** argv)
{
Grid_init(&argc,&argv);
int threads = GridThread::GetThreads();
Coordinate latt4 = GridDefaultLatt();
int Ls=16;
for(int i=0;i<argc;i++)
if(std::string(argv[i]) == "-Ls"){
std::stringstream ss(argv[i+1]); ss >> Ls;
}
GridLogLayout();
long unsigned int single_site_flops = 8*Nc*(7+16*Nc);
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
std::cout << GridLogMessage << "Making s innermost grids"<<std::endl;
GridCartesian * sUGrid = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(),GridDefaultMpi());
GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
GridCartesian * sFGrid = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
std::vector<int> seeds4({1,2,3,4});
std::vector<int> seeds5({5,6,7,8});
std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
GridParallelRNG RNG4(UGrid); RNG4.SeedUniqueString(std::string("The 4D RNG"));
std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
GridParallelRNG RNG5(FGrid); RNG5.SeedUniqueString(std::string("The 5D RNG"));
std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
LatticeFermionF src (FGrid); random(RNG5,src);
LatticeFermionF src1 (FGrid); random(RNG5,src1);
#if 0
src = Zero();
{
Coordinate origin({0,0,0,latt4[2]-1,0});
SpinColourVectorF tmp;
tmp=Zero();
tmp()(0)(0)=Complex(-2.0,0.0);
std::cout << " source site 0 " << tmp<<std::endl;
pokeSite(tmp,src,origin);
}
#else
RealD N2 = 1.0/::sqrt(norm2(src));
src = src*N2;
#endif
LatticeFermionF result(FGrid); result=Zero();
LatticeFermionF ref(FGrid); ref=Zero();
LatticeFermionF tmp(FGrid);
LatticeFermionF err(FGrid);
std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
LatticeGaugeFieldF Umu(UGrid);
SU<Nc>::HotConfiguration(RNG4,Umu);
std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
#if 0
Umu=1.0;
for(int mu=0;mu<Nd;mu++){
LatticeColourMatrixF ttmp(UGrid);
ttmp = PeekIndex<LorentzIndex>(Umu,mu);
// if (mu !=2 ) ttmp = 0;
// ttmp = ttmp* pow(10.0,mu);
PokeIndex<LorentzIndex>(Umu,ttmp,mu);
}
std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
#endif
////////////////////////////////////
// Naive wilson implementation
////////////////////////////////////
// replicate across fifth dimension
// LatticeGaugeFieldF Umu5d(FGrid);
std::vector<LatticeColourMatrixF> U(4,UGrid);
for(int mu=0;mu<Nd;mu++){
U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
}
std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
if (1)
{
ref = Zero();
for(int mu=0;mu<Nd;mu++){
tmp = Cshift(src,mu+1,1);
{
autoView( tmp_v , tmp , CpuWrite);
autoView( U_v , U[mu] , CpuRead);
for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
tmp_v[Ls*ss+s] = U_v[ss]*tmp_v[Ls*ss+s];
}
}
}
ref=ref + tmp - Gamma(Gmu[mu])*tmp;
{
autoView( tmp_v , tmp , CpuWrite);
autoView( U_v , U[mu] , CpuRead);
autoView( src_v, src , CpuRead);
for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
}
}
}
tmp =Cshift(tmp,mu+1,-1);
ref=ref + tmp + Gamma(Gmu[mu])*tmp;
}
ref = -0.5*ref;
}
RealD mass=0.1;
RealD M5 =1.8;
RealD NP = UGrid->_Nprocessors;
RealD NN = UGrid->NodeCount();
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
std::cout << GridLogMessage<< "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" <<std::endl;
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionR::Dhop "<<std::endl;
std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplexF::Nsimd()<<std::endl;
std::cout << GridLogMessage<< "* VComplexF size is "<<sizeof(vComplexF)<< " B"<<std::endl;
if ( sizeof(RealF)==4 ) std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
if ( sizeof(RealF)==8 ) std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
#ifdef GRID_OMP
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
#endif
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3 WilsonKernels" <<std::endl;
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
int ncall =100;
if (1) {
FGrid->Barrier();
Dw.Dhop(src,result,0);
std::cout<<GridLogMessage<<"Called warmup"<<std::endl;
double t0=usecond();
for(int i=0;i<ncall;i++){
Dw.Dhop(src1,result,0);
Dw.Dhop(src,result,0);
err = ref-result;
std::cout<<GridLogMessage << "norm diff "<< norm2(err)<<std::endl;
assert (norm2(err)< 1.0e-4 );
}
double t1=usecond();
FGrid->Barrier();
double volume=Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
double flops=single_site_flops*volume*ncall;
auto nsimd = vComplex::Nsimd();
auto simdwidth = sizeof(vComplex);
// RF: Nd Wilson * Ls, Nd gauge * Ls, Nc colors
double data_rf = volume * ((2*Nd+1)*Nd*Nc + 2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
// mem: Nd Wilson * Ls, Nd gauge, Nc colors
double data_mem = (volume * (2*Nd+1)*Nd*Nc + (volume/Ls) *2*Nd*Nc*Nc) * simdwidth / nsimd * ncall / (1024.*1024.*1024.);
std::cout<<GridLogMessage << "Called Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
// std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
// std::cout<<GridLogMessage << "norm ref "<< norm2(ref)<<std::endl;
std::cout<<GridLogMessage << "mflop/s = "<< flops/(t1-t0)<<std::endl;
std::cout<<GridLogMessage << "mflop/s per rank = "<< flops/(t1-t0)/NP<<std::endl;
std::cout<<GridLogMessage << "mflop/s per node = "<< flops/(t1-t0)/NN<<std::endl;
std::cout<<GridLogMessage << "RF GiB/s (base 2) = "<< 1000000. * data_rf/((t1-t0))<<std::endl;
std::cout<<GridLogMessage << "mem GiB/s (base 2) = "<< 1000000. * data_mem/((t1-t0))<<std::endl;
err = ref-result;
std::cout<<GridLogMessage << "norm diff "<< norm2(err)<<std::endl;
//exit(0);
if(( norm2(err)>1.0e-4) ) {
/*
std::cout << "RESULT\n " << result<<std::endl;
std::cout << "REF \n " << ref <<std::endl;
std::cout << "ERR \n " << err <<std::endl;
*/
std::cout<<GridLogMessage << "WRONG RESULT" << std::endl;
FGrid->Barrier();
exit(-1);
}
assert (norm2(err)< 1.0e-4 );
}
if (1)
{ // Naive wilson dag implementation
ref = Zero();
for(int mu=0;mu<Nd;mu++){
// ref = src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
tmp = Cshift(src,mu+1,1);
{
autoView( ref_v, ref, CpuWrite);
autoView( tmp_v, tmp, CpuRead);
autoView( U_v , U[mu] , CpuRead);
for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
int i=s+Ls*ss;
ref_v[i]+= U_v[ss]*(tmp_v[i] + Gamma(Gmu[mu])*tmp_v[i]); ;
}
}
}
{
autoView( tmp_v , tmp , CpuWrite);
autoView( U_v , U[mu] , CpuRead);
autoView( src_v, src , CpuRead);
for(int ss=0;ss<U[mu].Grid()->oSites();ss++){
for(int s=0;s<Ls;s++){
tmp_v[Ls*ss+s] = adj(U_v[ss])*src_v[Ls*ss+s];
}
}
}
// tmp =adj(U[mu])*src;
tmp =Cshift(tmp,mu+1,-1);
{
autoView( ref_v, ref, CpuWrite);
autoView( tmp_v, tmp, CpuRead);
for(int i=0;i<ref_v.size();i++){
ref_v[i]+= tmp_v[i] - Gamma(Gmu[mu])*tmp_v[i]; ;
}
}
}
ref = -0.5*ref;
}
// dump=1;
Dw.Dhop(src,result,1);
std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
std::cout<<GridLogMessage << "Called DwDag"<<std::endl;
std::cout<<GridLogMessage << "norm dag result "<< norm2(result)<<std::endl;
std::cout<<GridLogMessage << "norm dag ref "<< norm2(ref)<<std::endl;
err = ref-result;
std::cout<<GridLogMessage << "norm dag diff "<< norm2(err)<<std::endl;
if((norm2(err)>1.0e-4)){
/*
std::cout<< "DAG RESULT\n " <<ref << std::endl;
std::cout<< "DAG sRESULT\n " <<result << std::endl;
std::cout<< "DAG ERR \n " << err <<std::endl;
*/
}
LatticeFermionF src_e (FrbGrid);
LatticeFermionF src_o (FrbGrid);
LatticeFermionF r_e (FrbGrid);
LatticeFermionF r_o (FrbGrid);
LatticeFermionF r_eo (FGrid);
std::cout<<GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec"<<std::endl;
pickCheckerboard(Even,src_e,src);
pickCheckerboard(Odd,src_o,src);
std::cout<<GridLogMessage << "src_e"<<norm2(src_e)<<std::endl;
std::cout<<GridLogMessage << "src_o"<<norm2(src_o)<<std::endl;
// S-direction is INNERMOST and takes no part in the parity.
std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
std::cout << GridLogMessage<< "* Benchmarking DomainWallFermionF::DhopEO "<<std::endl;
std::cout << GridLogMessage<< "* Vectorising space-time by "<<vComplexF::Nsimd()<<std::endl;
if ( sizeof(RealF)==4 ) std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
if ( sizeof(RealF)==8 ) std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
#ifdef GRID_OMP
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
#endif
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3 WilsonKernels" <<std::endl;
std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
{
FGrid->Barrier();
Dw.DhopEO(src_o,r_e,DaggerNo);
double t0=usecond();
for(int i=0;i<ncall;i++){
#ifdef CUDA_PROFILE
if(i==10) cudaProfilerStart();
#endif
Dw.DhopEO(src_o,r_e,DaggerNo);
#ifdef CUDA_PROFILE
if(i==20) cudaProfilerStop();
#endif
}
double t1=usecond();
FGrid->Barrier();
double volume=Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
double flops=(single_site_flops*volume*ncall)/2.0;
std::cout<<GridLogMessage << "Deo mflop/s = "<< flops/(t1-t0)<<std::endl;
std::cout<<GridLogMessage << "Deo mflop/s per rank "<< flops/(t1-t0)/NP<<std::endl;
std::cout<<GridLogMessage << "Deo mflop/s per node "<< flops/(t1-t0)/NN<<std::endl;
}
Dw.DhopEO(src_o,r_e,DaggerNo);
Dw.DhopOE(src_e,r_o,DaggerNo);
Dw.Dhop (src ,result,DaggerNo);
std::cout<<GridLogMessage << "r_e"<<norm2(r_e)<<std::endl;
std::cout<<GridLogMessage << "r_o"<<norm2(r_o)<<std::endl;
std::cout<<GridLogMessage << "res"<<norm2(result)<<std::endl;
setCheckerboard(r_eo,r_o);
setCheckerboard(r_eo,r_e);
err = r_eo-result;
std::cout<<GridLogMessage << "norm diff "<< norm2(err)<<std::endl;
if((norm2(err)>1.0e-4)){
/*
std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
std::cout<< "Deo REF\n " <<result << std::endl;
std::cout<< "Deo ERR \n " << err <<std::endl;
*/
}
pickCheckerboard(Even,src_e,err);
pickCheckerboard(Odd,src_o,err);
std::cout<<GridLogMessage << "norm diff even "<< norm2(src_e)<<std::endl;
std::cout<<GridLogMessage << "norm diff odd "<< norm2(src_o)<<std::endl;
assert(norm2(src_e)<1.0e-4);
assert(norm2(src_o)<1.0e-4);
Grid_finalize();
exit(0);
}

2
systems/PVC/benchmarks/run-1tile.sh
View File

@@ -4,7 +4,7 @@
#SBATCH -p QZ1J-ICX-PVC
##SBATCH -p QZ1J-SPR-PVC-2C
source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
#source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
export NT=8

12
systems/PVC/benchmarks/run-2tile-mpi.sh
View File

@@ -4,7 +4,7 @@
#SBATCH -p QZ1J-ICX-PVC
source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
#source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
export NT=16
@@ -19,11 +19,15 @@ export SYCL_DEVICE_FILTER=gpu,level_zero
export I_MPI_OFFLOAD_CELL=tile
export EnableImplicitScaling=0
export EnableWalkerPartition=0
export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
#export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
#export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0
#mpiexec -launcher ssh -n 1 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 0 > 1tile.log
for i in 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
do
mpiexec -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT --shm-mpi 0 --device-mem 32768 > 1.1.1.2.log$i
mpiexec -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --shm-mpi 0 --device-mem 32768 > 2.1.1.1.log$i
done
mpiexec -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 0

7
systems/PVC/benchmarks/wrap.sh
View File

@@ -5,10 +5,5 @@ export ZE_AFFINITY_MASK=0.$MPI_LOCALRANKID
echo Ranke $MPI_LOCALRANKID ZE_AFFINITY_MASK is $ZE_AFFINITY_MASK
#if [ $MPI_LOCALRANKID = "0" ]
#then
# ~psteinbr/build_pti/ze_tracer -c $@
# onetrace --chrome-kernel-timeline $@
#else
$@
#fi