Commit bf9031a9 authored by BRAMAS Berenger's avatar BRAMAS Berenger

add priority

parent d904be81
......@@ -53,6 +53,8 @@ public:
}
}
rebuildInteractions();
FLOG(FLog::Controller << "FGroupTaskAlgorithm (Max Thread " << MaxThreads << ")\n");
}
......@@ -63,13 +65,7 @@ public:
delete[] kernels;
}
protected:
/**
* Runs the complete algorithm.
*/
void executeCore(const unsigned operationsToProceed) override {
FLOG( FLog::Controller << "\tStart FGroupTaskAlgorithm\n" );
void rebuildInteractions(){
#pragma omp parallel num_threads(MaxThreads)
{
#pragma omp single nowait
......@@ -78,6 +74,14 @@ protected:
buildExternalInteractionVecs();
}
}
}
protected:
/**
* Runs the complete algorithm.
*/
void executeCore(const unsigned operationsToProceed) override {
FLOG( FLog::Controller << "\tStart FGroupTaskAlgorithm\n" );
#pragma omp parallel num_threads(MaxThreads)
{
......
......@@ -58,9 +58,13 @@ protected:
FTaskTimer taskTimeRecorder;
#endif
// Used with OPENMP_SUPPORT_PRIORITY
size_t p2pPrioCriteria;
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), p2pPrioCriteria(0)
#ifdef SCALFMM_TIME_OMPTASKS
, taskTimeRecorder(MaxThreads)
#endif
......@@ -81,6 +85,8 @@ public:
}
}
rebuildInteractions();
FLOG(FLog::Controller << "FGroupTaskDepAlgorithm (Max Thread " << MaxThreads << ")\n");
}
......@@ -91,13 +97,7 @@ public:
delete[] kernels;
}
protected:
/**
* Runs the complete algorithm.
*/
void executeCore(const unsigned operationsToProceed) override {
FLOG( FLog::Controller << "\tStart FGroupTaskDepAlgorithm\n" );
void rebuildInteractions(){
#pragma omp parallel num_threads(MaxThreads)
{
#pragma omp single nowait
......@@ -107,6 +107,25 @@ protected:
}
}
#ifdef OPENMP_SUPPORT_PRIORITY
size_t nbLeaves = 0;
size_t nbParticles = 0;
for(int idxGroup = 0 ; idxGroup < tree->getNbParticleGroup() ; ++idxGroup){
ParticleGroupClass* containers = tree->getParticleGroup(idxGroup);
nbLeaves += containers->getNumberOfLeavesInBlock();
nbParticles += containers->getNbParticlesInGroup();
}
p2pPrioCriteria = (nbParticles/nbLeaves);
#endif
}
protected:
/**
* Runs the complete algorithm.
*/
void executeCore(const unsigned operationsToProceed) override {
FLOG( FLog::Controller << "\tStart FGroupTaskDepAlgorithm\n" );
FTIME_TASKS(taskTimeRecorder.start());
#pragma omp parallel num_threads(MaxThreads)
......@@ -318,7 +337,7 @@ protected:
ParticleGroupClass* containers = tree->getParticleGroup(idxGroup);
#pragma omp task default(shared) firstprivate(leafCells, cellPoles, containers) depend(inout: cellPoles[0]) priority_if_supported(1)
#pragma omp task default(shared) firstprivate(leafCells, cellPoles, containers) depend(inout: cellPoles[0]) priority_if_supported(-4)
{
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(&taskTimeRecorder, (leafCells->getStartingIndex() << 16) | (0<<8) | 0, "P2M"));
KernelClass*const kernel = kernels[omp_get_thread_num()];
......@@ -375,7 +394,7 @@ protected:
FAssertLF( nbSubCellGroups <= 9 );
}
#pragma omp task default(none) firstprivate(idxLevel, currentCells, cellPoles, subCellGroups, subCellGroupPoles, nbSubCellGroups) depend(inout: cellPoles[0]) depend(in: subCellGroupPoles[0][0], subCellGroupPoles[1][0], subCellGroupPoles[2][0], subCellGroupPoles[3][0], subCellGroupPoles[4][0], subCellGroupPoles[5][0], subCellGroupPoles[6][0], subCellGroupPoles[7][0], subCellGroupPoles[8][0]) priority_if_supported(1)
#pragma omp task default(none) firstprivate(idxLevel, currentCells, cellPoles, subCellGroups, subCellGroupPoles, nbSubCellGroups) depend(inout: cellPoles[0]) depend(in: subCellGroupPoles[0][0], subCellGroupPoles[1][0], subCellGroupPoles[2][0], subCellGroupPoles[3][0], subCellGroupPoles[4][0], subCellGroupPoles[5][0], subCellGroupPoles[6][0], subCellGroupPoles[7][0], subCellGroupPoles[8][0]) priority_if_supported(-3)
{
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(&taskTimeRecorder, (currentCells->getStartingIndex() << 16) | (idxLevel<<8) | 1, "M2M"));
KernelClass*const kernel = kernels[omp_get_thread_num()];
......@@ -433,7 +452,7 @@ protected:
PoleCellClass* cellPoles = currentCells->getRawMultipoleBuffer();
LocalCellClass* cellLocals = currentCells->getRawLocalBuffer();
#pragma omp task default(none) firstprivate(currentCells, cellPoles, cellLocals, idxLevel) depend(commute_if_supported: cellLocals[0]) depend(in: cellPoles[0]) priority_if_supported(1)
#pragma omp task default(none) firstprivate(currentCells, cellPoles, cellLocals, idxLevel) depend(commute_if_supported: cellLocals[0]) depend(in: cellPoles[0]) priority_if_supported(idxLevel==FAbstractAlgorithm::lowerWorkingLevel-1?0:-2)
{
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(&taskTimeRecorder, (currentCells->getStartingIndex() << 16) | (idxLevel<<8) | 2, "M2L"));
const MortonIndex blockStartIdx = currentCells->getStartingIndex();
......@@ -497,7 +516,7 @@ protected:
LocalCellClass* cellOtherLocals = cellsOther->getRawLocalBuffer();
const std::vector<OutOfBlockInteraction>* outsideInteractions = &(*currentInteractions).interactions;
#pragma omp task default(none) firstprivate(currentCells, cellPoles, cellLocals, outsideInteractions, cellsOther, cellOtherPoles, cellOtherLocals, idxLevel) depend(commute_if_supported: cellLocals[0], cellOtherLocals[0]) depend(in: cellPoles[0], cellOtherPoles[0]) priority_if_supported(1)
#pragma omp task default(none) firstprivate(currentCells, cellPoles, cellLocals, outsideInteractions, cellsOther, cellOtherPoles, cellOtherLocals, idxLevel) depend(commute_if_supported: cellLocals[0], cellOtherLocals[0]) depend(in: cellPoles[0], cellOtherPoles[0]) priority_if_supported(idxLevel==FAbstractAlgorithm::lowerWorkingLevel-1?0:-2)
{
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(&taskTimeRecorder, (cellsOther->getStartingIndex()) << 50 | (currentCells->getStartingIndex() << 16) | (idxLevel<<8) | 3, "M2L ext"));
KernelClass*const kernel = kernels[omp_get_thread_num()];
......@@ -578,7 +597,7 @@ protected:
FAssertLF( nbSubCellGroups <= 9 );
}
#pragma omp task default(none) firstprivate(idxLevel, currentCells, cellLocals, subCellGroups, subCellLocalGroupsLocal, nbSubCellGroups) depend(commute_if_supported: subCellLocalGroupsLocal[0][0], subCellLocalGroupsLocal[1][0], subCellLocalGroupsLocal[2][0], subCellLocalGroupsLocal[3][0], subCellLocalGroupsLocal[4][0], subCellLocalGroupsLocal[5][0], subCellLocalGroupsLocal[6][0], subCellLocalGroupsLocal[7][0], subCellLocalGroupsLocal[8][0]) depend(in: cellLocals[0]) priority_if_supported(1)
#pragma omp task default(none) firstprivate(idxLevel, currentCells, cellLocals, subCellGroups, subCellLocalGroupsLocal, nbSubCellGroups) depend(commute_if_supported: subCellLocalGroupsLocal[0][0], subCellLocalGroupsLocal[1][0], subCellLocalGroupsLocal[2][0], subCellLocalGroupsLocal[3][0], subCellLocalGroupsLocal[4][0], subCellLocalGroupsLocal[5][0], subCellLocalGroupsLocal[6][0], subCellLocalGroupsLocal[7][0], subCellLocalGroupsLocal[8][0]) depend(in: cellLocals[0]) priority_if_supported(-1)
{
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(&taskTimeRecorder, (currentCells->getStartingIndex() << 16) | (idxLevel<<8) | 4, "L2L"));
KernelClass*const kernel = kernels[omp_get_thread_num()];
......@@ -631,7 +650,7 @@ protected:
ParticleGroupClass* containers = (*iterParticles);
unsigned char* containersDown = containers->getRawAttributesBuffer();
#pragma omp task default(none) firstprivate(containers, containersDown) depend(commute_if_supported: containersDown[0]) priority_if_supported(1)
#pragma omp task default(none) firstprivate(containers, containersDown) depend(commute_if_supported: containersDown[0]) priority_if_supported((containers->getNbParticlesInGroup()/containers->getNumberOfLeavesInBlock())>size_t(p2pPrioCriteria*1.1)?-5:2)
{
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(&taskTimeRecorder, (containers->getStartingIndex() << 16) | (0<<8) | 5, "P2P"));
const MortonIndex blockStartIdx = containers->getStartingIndex();
......@@ -690,7 +709,7 @@ protected:
unsigned char* containersOtherDown = containersOther->getRawAttributesBuffer();
const std::vector<OutOfBlockInteraction>* outsideInteractions = &(*currentInteractions).interactions;
#pragma omp task default(none) firstprivate(containers, containersDown, containersOther, containersOtherDown, outsideInteractions) depend(commute_if_supported: containersOtherDown[0], containersDown[0]) priority_if_supported(1)
#pragma omp task default(none) firstprivate(containers, containersDown, containersOther, containersOtherDown, outsideInteractions) depend(commute_if_supported: containersOtherDown[0], containersDown[0]) priority_if_supported((containers->getNbParticlesInGroup()/containers->getNumberOfLeavesInBlock())>size_t(p2pPrioCriteria*1.1)?-5:2)
{
FTIME_TASKS(FTaskTimer::ScopeEvent taskTime(&taskTimeRecorder, (containersOther->getStartingIndex()) << 50 | (containers->getStartingIndex() << 16) | (0<<8) | 6, "P2P ext"));
KernelClass*const kernel = kernels[omp_get_thread_num()];
......
......@@ -174,6 +174,7 @@ public:
cellHandles = new std::vector<CellHandles>[tree->getHeight()];
initCodelet();
rebuildInteractions();
#ifdef STARPU_SUPPORT_ARBITER
arbiterPole = starpu_arbiter_create();
......@@ -237,6 +238,14 @@ public:
starpu_shutdown();
}
void rebuildInteractions(){
#pragma omp parallel
#pragma omp single
buildExternalInteractionVecs();
buildHandles();
}
protected:
/**
* Runs the complete algorithm.
......@@ -245,12 +254,6 @@ protected:
FLOG( FLog::Controller << "\tStart FGroupTaskStarPUAlgorithm\n" );
const bool directOnly = (tree->getHeight() <= 2);
#pragma omp parallel
#pragma omp single
buildExternalInteractionVecs();
buildHandles();
starpu_resume();
if(operationsToProceed & FFmmP2M && !directOnly) bottomPass();
......
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