diff --git a/Batch/run-benchs.sh b/Batch/run-benchs.sh
index 9ea8c9bc5ef1ccd2006d6f7b31a6ddd09a0f0823..3340dd6e81fba908778e1dfd7bbbc322bb4c3075 100644
--- a/Batch/run-benchs.sh
+++ b/Batch/run-benchs.sh
@@ -54,18 +54,18 @@ function main(){
NB_LOOPS=10
# AXPY
- "$RUN_DIR/Benchmark/axpy/axpy" --lp=$NB_LOOPS --minnbb=16 --maxnbb=256 --minbs=128 --maxbs=65536 --gputh=256 --od="$results_dir" >> "$results_dir/output_axpy.txt"
+ "$RUN_DIR/Benchmark/axpy/axpy" --lp=$NB_LOOPS --minnbb=16 --maxnbb=256 --minbs=128 --maxbs=131072 --gputh=256 --od="$results_dir" >> "$results_dir/output_axpy.txt"
remove_core_files
# Cholesky/gemm
- "$RUN_DIR/Benchmark/cholesky_gemm/cholesky" --lp=$NB_LOOPS --minms=4096 --maxms=8192 --minbs=128 --maxbs=512 --od="$results_dir" >> "$results_dir/output_cholesky.txt"
+ "$RUN_DIR/Benchmark/cholesky_gemm/cholesky" --lp=$NB_LOOPS --minms=4096 --maxms=16384 --minbs=128 --maxbs=512 --od="$results_dir" >> "$results_dir/output_cholesky.txt"
remove_core_files
- "$RUN_DIR/Benchmark/cholesky_gemm/gemm" --lp=$NB_LOOPS --minms=4096 --maxms=8192 --minbs=128 --maxbs=512 --od="$results_dir" >> "$results_dir/output_gemm.txt"
+ "$RUN_DIR/Benchmark/cholesky_gemm/gemm" --lp=$NB_LOOPS --minms=4096 --maxms=16384 --minbs=128 --maxbs=512 --od="$results_dir" >> "$results_dir/output_gemm.txt"
remove_core_files
# Particles
- "$RUN_DIR/Benchmark/particles/particles-simu" --lp=$NB_LOOPS --minp=500 --maxp=8000 --minnbgroups=128 --maxnbgroups=512 --od="$results_dir" >> "$results_dir/output_particles.txt"
+ "$RUN_DIR/Benchmark/particles/particles-simu" --lp=$NB_LOOPS --minp=500 --maxp=15000 --minnbgroups=128 --maxnbgroups=1024 --od="$results_dir" >> "$results_dir/output_particles.txt"
remove_core_files
}
diff --git a/Benchmark/axpy/axpy.cpp b/Benchmark/axpy/axpy.cpp
index 77e17c9c2e90a04a51c689f247fb7c2377953d9d..448bd33c1cc5116a96f86f6097efe713bc08ee2e 100644
--- a/Benchmark/axpy/axpy.cpp
+++ b/Benchmark/axpy/axpy.cpp
@@ -77,7 +77,7 @@ __global__ void cu_axpy(int n, NumType a, NumType *x, NumType *y, NumType *out)
template <int MaxNbDevices, const bool FavorLocality>
-auto BenchmarkTest(const int NbLoops, const int nbGpu, const int nbblocks, const int blocksize, const int gpunbthreads,
+auto BenchmarkTest(const int NbLoops, const int nbCpu, const int nbGpu, const int nbblocks, const int blocksize, const int gpunbthreads,
const bool useMultiPrioScheduler){
std::vector<Vector<float>> x(nbblocks, Vector<float>(blocksize, 1));
std::vector<Vector<float>> y(nbblocks, Vector<float>(blocksize, 1));
@@ -92,9 +92,9 @@ auto BenchmarkTest(const int NbLoops, const int nbGpu, const int nbblocks, const
else{
scheduler = std::unique_ptr<SpAbstractScheduler>(new SpMultiPrioScheduler<MaxNbDevices,FavorLocality>(nbGpu*SpCudaUtils::GetDefaultNbStreams()));
}
- SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuGpuWorkers(SpUtils::DefaultNumThreads(), nbGpu), std::move(scheduler));
+ SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuGpuWorkers(nbCpu, nbGpu), std::move(scheduler));
#else
- SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuWorkers());
+ SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuWorkers(nbCpu));
#endif
std::vector<double> minMaxAvg(3);
@@ -252,7 +252,9 @@ int main(int argc, char** argv){
std::make_tuple(true, false, true)};
for(auto useMultiprioAndPairs: schedPairConf){
- for(int idxGpu = 0 ; idxGpu <= nbGpus ; ++idxGpu){
+ for(int idxGpu = -1 ; idxGpu <= nbGpus ; ++idxGpu){
+ const int nbCpus = (idxGpu == -1 ? 1 : SpUtils::DefaultNumThreads());
+ const int nbGpus = (idxGpu == -1 ? 0 : idxGpu);
const bool useMultiprio = std::get<0>(useMultiprioAndPairs);
const bool usePrioPairs = std::get<1>(useMultiprioAndPairs);
const bool useLocality = std::get<2>(useMultiprioAndPairs);
@@ -262,8 +264,8 @@ int main(int argc, char** argv){
std::cout << " - NbBlocks = " << idxNbBlocks << " BlockSize = " << idxSize << std::endl;
const auto minMaxAvg = (useLocality ?
- BenchmarkTest<8,true>(NbLoops, idxGpu, idxNbBlocks, idxSize, gpunbthreads, useMultiprio)
- : BenchmarkTest<8,false>(NbLoops, idxGpu, idxNbBlocks, idxSize, gpunbthreads, useMultiprio));
+ BenchmarkTest<8,true>(NbLoops, nbCpus, nbGpus, idxNbBlocks, idxSize, gpunbthreads, useMultiprio)
+ : BenchmarkTest<8,false>(NbLoops, nbCpus, nbGpus, idxNbBlocks, idxSize, gpunbthreads, useMultiprio));
std::cout << " - Duration = " << minMaxAvg[0] << " " << minMaxAvg[1] << " " << minMaxAvg[2] << std::endl;
std::cout << " - Transfers = " << minMaxAvg[3] << " " << minMaxAvg[4] << " " << minMaxAvg[5] << " " << minMaxAvg[6] << " " << minMaxAvg[7] << std::endl;
std::cout << " - End" << std::endl;
@@ -280,16 +282,19 @@ int main(int argc, char** argv){
return 1;
}
- file << "NbGpu,NbBlocks,BlockSize,Multiprio,PrioPair,FavorLocality,MinDuration,MaxDuration,AvgDuration,TotalTransfer,MaxTransfer,DeviceToHostTransfer,HostToDeviceTransfer,DeviceToDeviceTransfer" << std::endl;
+ file << "NbCpus,NbGpu,NbBlocks,BlockSize,Multiprio,PrioPair,FavorLocality,MinDuration,MaxDuration,AvgDuration,TotalTransfer,MaxTransfer,DeviceToHostTransfer,HostToDeviceTransfer,DeviceToDeviceTransfer" << std::endl;
int idxDuration = 0;
for(auto useMultiprioAndPairs: schedPairConf){
const bool useMultiprio = std::get<0>(useMultiprioAndPairs);
const bool usePrioPairs = std::get<1>(useMultiprioAndPairs);
const bool useLocality = std::get<2>(useMultiprioAndPairs);
- for(int idxGpu = 0 ; idxGpu <= nbGpus ; ++idxGpu){
+ for(int idxGpu = -1 ; idxGpu <= nbGpus ; ++idxGpu){
for(int idxNbBlocks = minnbblocks ; idxNbBlocks <= maxnbblocks ; idxNbBlocks *= 2){
for(int idxSize = minblocksize ; idxSize <= maxblocksize ; idxSize *= 2){
- file << idxGpu << "," << idxNbBlocks << "," << idxSize << ","
+ const int nbCpus = (idxGpu == -1 ? 1 : SpUtils::DefaultNumThreads());
+ const int nbGpus = (idxGpu == -1 ? 0 : idxGpu);
+
+ file << nbCpus << "," << nbGpus << "," << idxNbBlocks << "," << idxSize << ","
<< (useMultiprio?"TRUE":"FALSE") << ","
<< (usePrioPairs?"TRUE":"FALSE") << ","
<< (useLocality?"TRUE":"FALSE") << ","
diff --git a/Benchmark/cholesky_gemm/cholesky.cpp b/Benchmark/cholesky_gemm/cholesky.cpp
index 95f13d158d9a310a0b90b90ef24fc512be565e58..6e85cb174a53327b43766a09401c95f479b6ca43 100644
--- a/Benchmark/cholesky_gemm/cholesky.cpp
+++ b/Benchmark/cholesky_gemm/cholesky.cpp
@@ -58,7 +58,7 @@ thread_local HipHandles handle;
template <int MaxNbDevices, const bool FavorLocality>
auto choleskyFactorization(const int NbLoops, SpBlas::Block blocksInput[], const int inMatrixDim, const int inBlockDim,
- const int nbGpu, const bool useMultiPrioScheduler, const bool usePrioPairs){
+ const int nbCpu, const int nbGpu, const bool useMultiPrioScheduler, const bool usePrioPairs){
const int nbBlocks = (inMatrixDim+inBlockDim-1)/inBlockDim;
#if defined(SPECX_COMPILE_WITH_CUDA) || defined(SPECX_COMPILE_WITH_HIP)
@@ -69,9 +69,9 @@ auto choleskyFactorization(const int NbLoops, SpBlas::Block blocksInput[], const
else{
scheduler = std::unique_ptr<SpAbstractScheduler>(new SpMultiPrioScheduler<MaxNbDevices,FavorLocality>(nbGpu*SpCudaUtils::GetDefaultNbStreams()));
}
- SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuGpuWorkers(SpUtils::DefaultNumThreads(), nbGpu), std::move(scheduler));
+ SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuGpuWorkers(nbCpu, nbGpu), std::move(scheduler));
#else
- SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuWorkers());
+ SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuWorkers(nbCpu));
#endif
#ifdef SPECX_COMPILE_WITH_CUDA
@@ -434,9 +434,12 @@ int main(int argc, char** argv){
std::make_tuple(true, false, true)};
for(auto useMultiprioAndPairs: schedPairConf){
- for(int idxGpu = 0 ; idxGpu <= nbGpus ; ++idxGpu){
+ for(int idxGpu = -1 ; idxGpu <= nbGpus ; ++idxGpu){
for(int BlockSize = MinBlockSize ; BlockSize <= MaxBlockSize ; BlockSize *= 2){
for(int MatrixSize = MinMatrixSize ; MatrixSize <= MaxMatrixSize ; MatrixSize *= 2){
+ const int nbCpus = (idxGpu == -1 ? 1 : SpUtils::DefaultNumThreads());
+ const int nbGpus = (idxGpu == -1 ? 0 : idxGpu);
+
const bool useMultiprio = std::get<0>(useMultiprioAndPairs);
const bool usePrioPairs = std::get<1>(useMultiprioAndPairs);
const bool useLocality = std::get<2>(useMultiprioAndPairs);
@@ -468,9 +471,9 @@ int main(int argc, char** argv){
/////////////////////////////////////////////////////////
const auto minMaxAvg = (useLocality ?
choleskyFactorization<8,true>(NbLoops, blocks.get(), MatrixSize, BlockSize,
- idxGpu, useMultiprio, usePrioPairs) :
+ nbCpus, nbGpus, useMultiprio, usePrioPairs) :
choleskyFactorization<8,false>(NbLoops, blocks.get(), MatrixSize, BlockSize,
- idxGpu, useMultiprio, usePrioPairs));
+ nbCpus, nbGpus, useMultiprio, usePrioPairs));
allDurations.push_back(minMaxAvg);
std::cout << " - Duration = " << minMaxAvg[0] << " " << minMaxAvg[1] << " " << minMaxAvg[2] << std::endl;
std::cout << " - Transfers = " << minMaxAvg[3] << " " << minMaxAvg[4] << " " << minMaxAvg[5] << " " << minMaxAvg[6] << " " << minMaxAvg[7] << std::endl;
@@ -500,17 +503,20 @@ int main(int argc, char** argv){
return 1;
}
- file << "NbGpu,MatrixSize,BlockSize,Multiprio,PrioPair,FavorLocality,MinDuration,MaxDuration,AvgDuration,TotalTransfer,MaxTransfer,DeviceToHostTransfer,HostToDeviceTransfer,DeviceToDeviceTransfer" << std::endl;
+ file << "NbCpu,NbGpu,MatrixSize,BlockSize,Multiprio,PrioPair,FavorLocality,MinDuration,MaxDuration,AvgDuration,TotalTransfer,MaxTransfer,DeviceToHostTransfer,HostToDeviceTransfer,DeviceToDeviceTransfer" << std::endl;
int idxDuration = 0;
for(auto useMultiprioAndPairs: schedPairConf){
- for(int idxGpu = 0 ; idxGpu <= nbGpus ; ++idxGpu){
+ for(int idxGpu = -1 ; idxGpu <= nbGpus ; ++idxGpu){
for(int BlockSize = MinBlockSize ; BlockSize <= MaxBlockSize ; BlockSize *= 2){
for(int MatrixSize = MinMatrixSize ; MatrixSize <= MaxMatrixSize ; MatrixSize *= 2){
+ const int nbCpus = (idxGpu == -1 ? 1 : SpUtils::DefaultNumThreads());
+ const int nbGpus = (idxGpu == -1 ? 0 : idxGpu);
+
const bool useMultiprio = std::get<0>(useMultiprioAndPairs);
const bool usePrioPairs = std::get<1>(useMultiprioAndPairs);
const bool useLocality = std::get<2>(useMultiprioAndPairs);
- file << idxGpu << "," << MatrixSize << "," << BlockSize << ","
+ file << nbCpus << "," << nbGpus << "," << MatrixSize << "," << BlockSize << ","
<< (useMultiprio?"TRUE":"FALSE") << ","
<< (usePrioPairs?"TRUE":"FALSE") << ","
<< (useLocality?"TRUE":"FALSE") << ","
diff --git a/Benchmark/cholesky_gemm/gemm.cpp b/Benchmark/cholesky_gemm/gemm.cpp
index 0b62f36516166fc239988fbf8d18e2cd59676c34..49f7b34db0b54797240c18637808922e690385ab 100644
--- a/Benchmark/cholesky_gemm/gemm.cpp
+++ b/Benchmark/cholesky_gemm/gemm.cpp
@@ -46,7 +46,7 @@ thread_local hipblasHandle_t handle;
template <int MaxNbDevices, const bool FavorLocality>
auto gemm(const int NbLoops, SpBlas::Block blocksC[], const SpBlas::Block blocksA[], const SpBlas::Block blocksB[],
const int inMatrixDim, const int inBlockDim,
- const int nbGpu, const bool useMultiPrioScheduler){
+ const int nbCpus, const int nbGpu, const bool useMultiPrioScheduler){
const int nbBlocks = (inMatrixDim+inBlockDim-1)/inBlockDim;
const bool exportTrace = false;
@@ -58,9 +58,9 @@ auto gemm(const int NbLoops, SpBlas::Block blocksC[], const SpBlas::Block blocks
else{
scheduler = std::unique_ptr<SpAbstractScheduler>(new SpMultiPrioScheduler<MaxNbDevices,FavorLocality>(nbGpu*SpCudaUtils::GetDefaultNbStreams()));
}
- SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuGpuWorkers(SpUtils::DefaultNumThreads(), nbGpu), std::move(scheduler));
+ SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuGpuWorkers(nbCpus, nbGpu), std::move(scheduler));
#else
- SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuWorkers());
+ SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuWorkers(nbCpus));
#endif
#ifdef SPECX_COMPILE_WITH_CUDA
@@ -268,7 +268,10 @@ int main(int argc, char** argv){
for(auto useMultiprioAndPairs: schedPairConf){
for(int BlockSize = MinBlockSize ; BlockSize <= MaxBlockSize ; BlockSize *= 2){
for(int MatrixSize = MinMatrixSize ; MatrixSize <= MaxMatrixSize ; MatrixSize *= 2){
- for(int idxGpu = 0 ; idxGpu <= nbGpus ; ++idxGpu){
+ for(int idxGpu = -1 ; idxGpu <= nbGpus ; ++idxGpu){
+ const int nbCpus = (idxGpu == -1 ? 1 : SpUtils::DefaultNumThreads());
+ const int nbGpus = (idxGpu == -1 ? 0 : idxGpu);
+
const bool useMultiprio = std::get<0>(useMultiprioAndPairs);
const bool useLocality = std::get<1>(useMultiprioAndPairs);
@@ -301,9 +304,9 @@ int main(int argc, char** argv){
auto blocksC = SpBlas::matrixToBlock(matrixC.get(), MatrixSize, BlockSize);
const auto minMaxAvg = (useLocality ?
gemm<8,true>(NbLoops, blocksC.get(), blocksA.get(), blocksB.get(),
- MatrixSize, BlockSize, idxGpu, useMultiprio):
+ MatrixSize, BlockSize, nbCpus, nbGpus, useMultiprio):
gemm<8,false>(NbLoops, blocksC.get(), blocksA.get(), blocksB.get(),
- MatrixSize, BlockSize, idxGpu, useMultiprio));
+ MatrixSize, BlockSize, nbCpus, nbGpus, useMultiprio));
allDurations.push_back(minMaxAvg);
std::cout << " - Duration = " << minMaxAvg[0] << " " << minMaxAvg[1] << " " << minMaxAvg[2] << std::endl;
std::cout << " - Transfers = " << minMaxAvg[3] << " " << minMaxAvg[4] << " " << minMaxAvg[5] << " " << minMaxAvg[6] << " " << minMaxAvg[7] << std::endl;
@@ -333,16 +336,19 @@ int main(int argc, char** argv){
return 1;
}
- file << "NbGpu,MatrixSize,BlockSize,Multiprio,PrioPair,FavorLocality,MinDuration,MaxDuration,AvgDuration,TotalTransfer,MaxTransfer,DeviceToHostTransfer,HostToDeviceTransfer,DeviceToDeviceTransfer" << std::endl;
+ file << "NbCpu,NbGpu,MatrixSize,BlockSize,Multiprio,PrioPair,FavorLocality,MinDuration,MaxDuration,AvgDuration,TotalTransfer,MaxTransfer,DeviceToHostTransfer,HostToDeviceTransfer,DeviceToDeviceTransfer" << std::endl;
int idxDuration = 0;
for(auto useMultiprioAndPairs: schedPairConf){
for(int BlockSize = MinBlockSize ; BlockSize <= MaxBlockSize ; BlockSize *= 2){
for(int MatrixSize = MinMatrixSize ; MatrixSize <= MaxMatrixSize ; MatrixSize *= 2){
- for(int idxGpu = 0 ; idxGpu <= nbGpus ; ++idxGpu){
+ for(int idxGpu = -1 ; idxGpu <= nbGpus ; ++idxGpu){
+ const int nbCpus = (idxGpu == -1 ? 1 : SpUtils::DefaultNumThreads());
+ const int nbGpus = (idxGpu == -1 ? 0 : idxGpu);
+
const bool useMultiprio = std::get<0>(useMultiprioAndPairs);
const bool useLocality = std::get<1>(useMultiprioAndPairs);
- file << idxGpu << "," << MatrixSize << "," << BlockSize << ","
+ file << nbCpus << "," << nbGpus << "," << MatrixSize << "," << BlockSize << ","
<< (useMultiprio?"TRUE":"FALSE") << ","
<< "FALSE" << ","
<< (useLocality?"TRUE":"FALSE") << ","
diff --git a/Benchmark/particles/particles-simu.cpp b/Benchmark/particles/particles-simu.cpp
index 54cc36c941d4509ee965ac5d89f05767787ab208..4360854c7d0380c5830ec74a388dcb57141fee50 100644
--- a/Benchmark/particles/particles-simu.cpp
+++ b/Benchmark/particles/particles-simu.cpp
@@ -704,7 +704,7 @@ auto GetPriority(const bool usePrioPairs, const int maxInteractions, const int m
template <int MaxNbDevices, const bool FavorLocality>
auto BenchCore( const int NbLoops, const int MinPartsPerGroup, const int MaxPartsPerGroup,
- const int NbGroups, const int nbGpu, const bool useMultiPrioScheduler,
+ const int NbGroups, const int nbCpu, const int nbGpu, const bool useMultiPrioScheduler,
const bool usePrioPairs, const TuneResult& inKernelConfig,
const int maxInteractions, const int minInteractions){
@@ -727,9 +727,9 @@ auto BenchCore( const int NbLoops, const int MinPartsPerGroup, const int MaxPart
else{
scheduler = std::unique_ptr<SpAbstractScheduler>(new SpMultiPrioScheduler<MaxNbDevices,FavorLocality>(nbGpu*SpCudaUtils::GetDefaultNbStreams()));
}
- SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuGpuWorkers(SpUtils::DefaultNumThreads(), nbGpu), std::move(scheduler));
+ SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuGpuWorkers(nbCpu, nbGpu), std::move(scheduler));
#else
- SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuWorkers());
+ SpComputeEngine ce(SpWorkerTeamBuilder::TeamOfCpuWorkers(nbCpu));
#endif
std::vector<double> minMaxAvg(3);
@@ -917,8 +917,10 @@ void BenchmarkTest(int argc, char** argv, const TuneResult& inKernelConfig){
std::make_tuple(true, false, true)};
for(auto useMultiprioAndPairs: schedPairConf){
- for(int idxGpu = 0 ; idxGpu <= nbGpus ; ++idxGpu){
+ for(int idxGpu = -1 ; idxGpu <= nbGpus ; ++idxGpu){
for(int idxBlock = MinNbGroups ; idxBlock <= MaxNbGroups ; idxBlock *= 2){
+ const int nbCpus = (idxGpu == -1 ? 1 : SpUtils::DefaultNumThreads());
+ const int nbGpus = (idxGpu == -1 ? 0 : idxGpu);
const bool useMultiprio = std::get<0>(useMultiprioAndPairs);
const bool usePrioPairs = std::get<1>(useMultiprioAndPairs);
const bool useLocality = std::get<2>(useMultiprioAndPairs);
@@ -929,10 +931,10 @@ void BenchmarkTest(int argc, char** argv, const TuneResult& inKernelConfig){
<< " Favor Loc = " << useLocality << std::endl;
const auto minMaxAvg = (useLocality ?
BenchCore<8,true>(NbLoops, MinPartsPerGroup,
- MaxPartsPerGroup, idxBlock, idxGpu, useMultiprio, usePrioPairs, inKernelConfig,
+ MaxPartsPerGroup, idxBlock, nbCpus, nbGpus, useMultiprio, usePrioPairs, inKernelConfig,
maxInteractions, minInteractions):
BenchCore<8,false>(NbLoops, MinPartsPerGroup,
- MaxPartsPerGroup, idxBlock, idxGpu, useMultiprio, usePrioPairs, inKernelConfig,
+ MaxPartsPerGroup, idxBlock, nbCpus, nbGpus, useMultiprio, usePrioPairs, inKernelConfig,
maxInteractions, minInteractions));
allDurations.push_back(minMaxAvg);
std::cout << " - Min = " << minMaxAvg[0] << " Max = " << minMaxAvg[1] << " Avg = " << minMaxAvg[2] << std::endl;
@@ -947,11 +949,13 @@ void BenchmarkTest(int argc, char** argv, const TuneResult& inKernelConfig){
return;
}
- file << "NbGpu,BlockSize,Multiprio,Multiprio,PrioPair,MinDuration,MaxDuration,AvgDuration,TotalTransfer,MaxTransfer,DeviceToHostTransfer,HostToDeviceTransfer,DeviceToDeviceTransfer" << std::endl;
+ file << "NbCpu,NbGpu,BlockSize,Multiprio,Multiprio,PrioPair,MinDuration,MaxDuration,AvgDuration,TotalTransfer,MaxTransfer,DeviceToHostTransfer,HostToDeviceTransfer,DeviceToDeviceTransfer" << std::endl;
int idxDuration = 0;
for(auto useMultiprioAndPairs: schedPairConf){
- for(int idxGpu = 0 ; idxGpu <= nbGpus ; ++idxGpu){
+ for(int idxGpu = -1 ; idxGpu <= nbGpus ; ++idxGpu){
for(int idxBlock = MinNbGroups ; idxBlock <= MaxNbGroups ; idxBlock *= 2){
+ const int nbCpus = (idxGpu == -1 ? 1 : SpUtils::DefaultNumThreads());
+ const int nbGpus = (idxGpu == -1 ? 0 : idxGpu);
const bool useMultiprio = std::get<0>(useMultiprioAndPairs);
const bool usePrioPairs = std::get<1>(useMultiprioAndPairs);
const bool useLocality = std::get<2>(useMultiprioAndPairs);