Skip to content

GitLab

Commit e288451f authored by BRAMAS Berenger's avatar BRAMAS Berenger
Browse files

update and add SCALFMM_NO_COMMUTE_LAST_L2L env variable to control the last L2L commute or not

parent a348fd97
Hide whitespace changes
Inline Side-by-side
Showing with 102 additions and 40 deletions
Src/GroupTree/Core/FGroupTaskDepAlgorithm.hpp
View file @ e288451f
......@@ -34,14 +34,14 @@
#ifdef OPENMP_SUPPORT_PRIORITY
#define priority_if_supported(x) priority(x)
enum FGroupTaskDepAlgorithm_Priorities{
FGroupTaskDepAlgorithm_Prio_P2M = 5,
FGroupTaskDepAlgorithm_Prio_M2M = 4,
FGroupTaskDepAlgorithm_Prio_M2L_High = 3,
FGroupTaskDepAlgorithm_Prio_L2L = 2,
FGroupTaskDepAlgorithm_Prio_P2P_Big = 1,
FGroupTaskDepAlgorithm_Prio_M2L = 0,
FGroupTaskDepAlgorithm_Prio_L2P = -1,
FGroupTaskDepAlgorithm_Prio_P2P_Small = -2
FGroupTaskDepAlgorithm_Prio_P2M = 9,
FGroupTaskDepAlgorithm_Prio_M2M = 8,
FGroupTaskDepAlgorithm_Prio_M2L_High = 7,
FGroupTaskDepAlgorithm_Prio_L2L = 6,
FGroupTaskDepAlgorithm_Prio_P2P_Big = 5,
FGroupTaskDepAlgorithm_Prio_M2L = 4,
FGroupTaskDepAlgorithm_Prio_L2P = 3,
FGroupTaskDepAlgorithm_Prio_P2P_Small = 2
};
#else
#define priority_if_supported(x)
......@@ -65,6 +65,7 @@ protected:
const int MaxThreads; //< The number of threads
OctreeClass*const tree; //< The Tree
KernelClass** kernels; //< The kernels
const bool noCommuteAtLastLevel;
#ifdef SCALFMM_TIME_OMPTASKS
FTaskTimer taskTimeRecorder;
......@@ -72,7 +73,8 @@ protected:
public:
FGroupTaskDepAlgorithm(OctreeClass*const inTree, KernelClass* inKernels, const int inMaxThreads = -1)
: MaxThreads(inMaxThreads==-1?omp_get_max_threads():inMaxThreads), tree(inTree), kernels(nullptr)
: MaxThreads(inMaxThreads==-1?omp_get_max_threads():inMaxThreads), tree(inTree), kernels(nullptr),
noCommuteAtLastLevel(getenv("SCALFMM_NO_COMMUTE_LAST_L2L") != NULL && getenv("SCALFMM_NO_COMMUTE_LAST_L2L")[0] != '0'?true:false)
#ifdef SCALFMM_TIME_OMPTASKS
, taskTimeRecorder(MaxThreads)
#endif
......@@ -103,6 +105,7 @@ public:
taskTimeRecorder.init(omp_get_thread_num());
}
#endif
FLOG(FLog::Controller << "SCALFMM_NO_COMMUTE_LAST_L2L " << noCommuteAtLastLevel << "\n");
}
~FGroupTaskDepAlgorithm(){
......@@ -627,45 +630,90 @@ protected:
subCellGroup = (*iterChildCells);
subCellLocalGroupsLocal = (*iterChildCells)->getRawLocalBuffer();
#pragma omp task default(none) firstprivate(idxLevel, currentCells, cellLocals, subCellGroup, subCellLocalGroupsLocal) depend(commute_if_supported: subCellLocalGroupsLocal[0]) depend(in: cellLocals[0]) priority_if_supported(FGroupTaskDepAlgorithm_Prio_L2L)
{
KernelClass*const kernel = kernels[omp_get_thread_num()];
if(noCommuteAtLastLevel == false || idxLevel != FAbstractAlgorithm::lowerWorkingLevel - 2){
#pragma omp task default(none) firstprivate(idxLevel, currentCells, cellLocals, subCellGroup, subCellLocalGroupsLocal) depend(commute_if_supported: subCellLocalGroupsLocal[0]) depend(in: cellLocals[0]) priority_if_supported(FGroupTaskDepAlgorithm_Prio_L2L)
{
KernelClass*const kernel = kernels[omp_get_thread_num()];
const MortonIndex firstParent = FMath::Max(currentCells->getStartingIndex(), subCellGroup->getStartingIndex()>>3);
const MortonIndex lastParent = FMath::Min(currentCells->getEndingIndex()-1, (subCellGroup->getEndingIndex()-1)>>3);
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(omp_get_thread_num(), &taskTimeRecorder, ((lastParent * 20) + idxLevel) * 8 + 4, "L2L"));
const MortonIndex firstParent = FMath::Max(currentCells->getStartingIndex(), subCellGroup->getStartingIndex()>>3);
const MortonIndex lastParent = FMath::Min(currentCells->getEndingIndex()-1, (subCellGroup->getEndingIndex()-1)>>3);
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(omp_get_thread_num(), &taskTimeRecorder, ((lastParent * 20) + idxLevel) * 8 + 4, "L2L"));
int idxParentCell = currentCells->getCellIndex(firstParent);
FAssertLF(idxParentCell != -1);
int idxParentCell = currentCells->getCellIndex(firstParent);
FAssertLF(idxParentCell != -1);
int idxChildCell = subCellGroup->getFistChildIdx(firstParent);
FAssertLF(idxChildCell != -1);
CellClass childData[8];
int idxChildCell = subCellGroup->getFistChildIdx(firstParent);
FAssertLF(idxChildCell != -1);
CellClass childData[8];
while(true){
CellClass cell = currentCells->getDownCell(idxParentCell);
FAssertLF(cell.getMortonIndex() == currentCells->getCellMortonIndex(idxParentCell));
CellClass* child[8] = {nullptr,nullptr,nullptr,nullptr,nullptr,nullptr,nullptr,nullptr};
while(true){
CellClass cell = currentCells->getDownCell(idxParentCell);
FAssertLF(cell.getMortonIndex() == currentCells->getCellMortonIndex(idxParentCell));
CellClass* child[8] = {nullptr,nullptr,nullptr,nullptr,nullptr,nullptr,nullptr,nullptr};
FAssertLF(cell.getMortonIndex() == (subCellGroup->getCellMortonIndex(idxChildCell)>>3));
FAssertLF(cell.getMortonIndex() == (subCellGroup->getCellMortonIndex(idxChildCell)>>3));
do{
const int idxChild = ((subCellGroup->getCellMortonIndex(idxChildCell)) & 7);
FAssertLF(child[idxChild] == nullptr);
childData[idxChild] = subCellGroup->getDownCell(idxChildCell);
FAssertLF(subCellGroup->getCellMortonIndex(idxChildCell) == childData[idxChild].getMortonIndex());
child[idxChild] = &childData[idxChild];
do{
const int idxChild = ((subCellGroup->getCellMortonIndex(idxChildCell)) & 7);
FAssertLF(child[idxChild] == nullptr);
childData[idxChild] = subCellGroup->getDownCell(idxChildCell);
FAssertLF(subCellGroup->getCellMortonIndex(idxChildCell) == childData[idxChild].getMortonIndex());
child[idxChild] = &childData[idxChild];
idxChildCell += 1;
}while(idxChildCell != subCellGroup->getNumberOfCellsInBlock() && cell.getMortonIndex() == (subCellGroup->getCellMortonIndex(idxChildCell)>>3));
idxChildCell += 1;
}while(idxChildCell != subCellGroup->getNumberOfCellsInBlock() && cell.getMortonIndex() == (subCellGroup->getCellMortonIndex(idxChildCell)>>3));
kernel->L2L(&cell, child, idxLevel);
kernel->L2L(&cell, child, idxLevel);
if(currentCells->getCellMortonIndex(idxParentCell) == lastParent){
break;
if(currentCells->getCellMortonIndex(idxParentCell) == lastParent){
break;
}
idxParentCell += 1;
}
}
}
else{
#pragma omp task default(none) firstprivate(idxLevel, currentCells, cellLocals, subCellGroup, subCellLocalGroupsLocal) depend(inout: subCellLocalGroupsLocal[0]) depend(in: cellLocals[0]) priority_if_supported(FGroupTaskDepAlgorithm_Prio_L2L)
{
KernelClass*const kernel = kernels[omp_get_thread_num()];
idxParentCell += 1;
const MortonIndex firstParent = FMath::Max(currentCells->getStartingIndex(), subCellGroup->getStartingIndex()>>3);
const MortonIndex lastParent = FMath::Min(currentCells->getEndingIndex()-1, (subCellGroup->getEndingIndex()-1)>>3);
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(omp_get_thread_num(), &taskTimeRecorder, ((lastParent * 20) + idxLevel) * 8 + 4, "L2L"));
int idxParentCell = currentCells->getCellIndex(firstParent);
FAssertLF(idxParentCell != -1);
int idxChildCell = subCellGroup->getFistChildIdx(firstParent);
FAssertLF(idxChildCell != -1);
CellClass childData[8];
while(true){
CellClass cell = currentCells->getDownCell(idxParentCell);
FAssertLF(cell.getMortonIndex() == currentCells->getCellMortonIndex(idxParentCell));
CellClass* child[8] = {nullptr,nullptr,nullptr,nullptr,nullptr,nullptr,nullptr,nullptr};
FAssertLF(cell.getMortonIndex() == (subCellGroup->getCellMortonIndex(idxChildCell)>>3));
do{
const int idxChild = ((subCellGroup->getCellMortonIndex(idxChildCell)) & 7);
FAssertLF(child[idxChild] == nullptr);
childData[idxChild] = subCellGroup->getDownCell(idxChildCell);
FAssertLF(subCellGroup->getCellMortonIndex(idxChildCell) == childData[idxChild].getMortonIndex());
child[idxChild] = &childData[idxChild];
idxChildCell += 1;
}while(idxChildCell != subCellGroup->getNumberOfCellsInBlock() && cell.getMortonIndex() == (subCellGroup->getCellMortonIndex(idxChildCell)>>3));
kernel->L2L(&cell, child, idxLevel);
if(currentCells->getCellMortonIndex(idxParentCell) == lastParent){
break;
}
idxParentCell += 1;
}
}
}
......
Src/GroupTree/Core/FGroupTaskStarpuAlgorithm.hpp
View file @ e288451f
......@@ -109,6 +109,8 @@ protected:
starpu_codelet p2p_redux_read;
#endif
const bool noCommuteAtLastLevel;
#ifdef STARPU_USE_CPU
StarPUCpuWrapperClass cpuWrapper;
#endif
......@@ -140,7 +142,8 @@ protected:
public:
FGroupTaskStarPUAlgorithm(OctreeClass*const inTree, KernelClass* inKernels)
: tree(inTree), originalCpuKernel(inKernels),
cellHandles(nullptr),
cellHandles(nullptr),
noCommuteAtLastLevel(getenv("SCALFMM_NO_COMMUTE_LAST_L2L") != NULL && getenv("SCALFMM_NO_COMMUTE_LAST_L2L")[0] != '0'?true:false),
#ifdef STARPU_USE_CPU
cpuWrapper(tree->getHeight()),
#endif
......@@ -212,6 +215,7 @@ public:
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max CUDA " << starpu_cuda_worker_get_count() << ")\n");
#endif
FLOG(FLog::Controller << "SCALFMM_NO_COMMUTE_LAST_L2L " << noCommuteAtLastLevel << "\n");
buildTaskNames();
}
......@@ -1118,7 +1122,12 @@ protected:
task->dyn_handles[3] = cellHandles[idxLevel+1][idxSubGroup].down;
// put the right codelet
task->cl = (idxLevel == FAbstractAlgorithm::lowerWorkingLevel - 2 ? &l2l_cl_nocommute : &l2l_cl);
if(noCommuteAtLastLevel){
task->cl = (idxLevel == FAbstractAlgorithm::lowerWorkingLevel - 2 ? &l2l_cl_nocommute : &l2l_cl);
}
else{
task->cl = &l2l_cl;
}
// put args values
char *arg_buffer;
size_t arg_buffer_size;
......@@ -1149,7 +1158,12 @@ protected:
task->dyn_handles[3] = cellHandles[idxLevel+1][idxSubGroup].down;
// put the right codelet
task->cl = (idxLevel == FAbstractAlgorithm::lowerWorkingLevel - 2 ? &l2l_cl_nocommute : &l2l_cl);
if(noCommuteAtLastLevel){
task->cl = (idxLevel == FAbstractAlgorithm::lowerWorkingLevel - 2 ? &l2l_cl_nocommute : &l2l_cl);
}
else{
task->cl = &l2l_cl;
}
// put args values
char *arg_buffer;
size_t arg_buffer_size;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment