From dc1809b4bf2e3b52b153ebee47fd972d665e3b46 Mon Sep 17 00:00:00 2001
From: Berenger Bramas <Berenger.Bramas@inria.fr>
Date: Mon, 17 Aug 2015 08:46:57 +0200
Subject: [PATCH] update the transfer pass with or without finalize
---
Src/Components/FAbstractKernels.hpp | 2 +-
Src/Core/FFmmAlgorithmOmp4.hpp | 2 +
Src/Core/FFmmAlgorithmSectionTask.hpp | 54 +++++++++++++++++--
Src/Core/FFmmAlgorithmTask.hpp | 29 +++++-----
Src/Core/FFmmAlgorithmThreadProc.hpp | 2 +-
Src/Core/FFmmAlgorithmThreadProcPeriodic.hpp | 2 +-
.../Spherical/FSphericalBlockBlasKernel.hpp | 2 +-
7 files changed, 73 insertions(+), 20 deletions(-)
diff --git a/Src/Components/FAbstractKernels.hpp b/Src/Components/FAbstractKernels.hpp
index 9238a9fbd..323f63b07 100644
--- a/Src/Components/FAbstractKernels.hpp
+++ b/Src/Components/FAbstractKernels.hpp
@@ -77,7 +77,7 @@ public:
*
* @return false
*/
- bool needFinishedM2LEvent(){
+ constexpr static bool NeedFinishedM2LEvent(){
return false;
}
/** This method can be optionally inherited
diff --git a/Src/Core/FFmmAlgorithmOmp4.hpp b/Src/Core/FFmmAlgorithmOmp4.hpp
index e7d7e80e0..de1cf3639 100644
--- a/Src/Core/FFmmAlgorithmOmp4.hpp
+++ b/Src/Core/FFmmAlgorithmOmp4.hpp
@@ -86,6 +86,8 @@ public:
FLOG(FLog::Controller << "FFmmAlgorithmOmp4 (Max Thread " << omp_get_max_threads() << ")\n");
+ FAssertLF(KernelClass::NeedFinishedM2LEvent() == false, "FFmmAlgorithmOmp4 cannot notify for M2L at level ending.");
+
#ifdef OPENMP_SUPPORT_PRIORITY
size_t nbLeaves = 0;
size_t nbParticles = 0;
diff --git a/Src/Core/FFmmAlgorithmSectionTask.hpp b/Src/Core/FFmmAlgorithmSectionTask.hpp
index 79863e0a0..7324a8aac 100644
--- a/Src/Core/FFmmAlgorithmSectionTask.hpp
+++ b/Src/Core/FFmmAlgorithmSectionTask.hpp
@@ -190,7 +190,7 @@ protected:
do{
// We need the current cell and the child
// child is an array (of 8 child) that may be null
- #pragma omp task firstprivate(octreeIterator) shared(idxLevel)
+ #pragma omp task firstprivate(octreeIterator,idxLevel)
{
kernels[omp_get_thread_num()]->M2M( octreeIterator.getCurrentCell() , octreeIterator.getCurrentChild(), idxLevel);
}
@@ -218,7 +218,12 @@ protected:
eztrace_start();
#endif
- this->transferPassWithFinalize() ;
+ if(KernelClass::NeedFinishedM2LEvent()){
+ this->transferPassWithFinalize() ;
+ }
+ else{
+ this->transferPassWithoutFinalize() ;
+ }
#ifdef SCALFMM_USE_EZTRACE
eztrace_stop();
#endif
@@ -246,7 +251,7 @@ protected:
do{
const int counter = tree->getInteractionNeighbors(neighbors, octreeIterator.getCurrentGlobalCoordinate(), idxLevel, separationCriteria);
if(counter){
- #pragma omp task firstprivate(octreeIterator, neighbors, counter) shared(idxLevel)
+ #pragma omp task firstprivate(octreeIterator, neighbors, counter,idxLevel)
{
kernels[omp_get_thread_num()]->M2L( octreeIterator.getCurrentCell() , neighbors, counter, idxLevel);
}
@@ -273,6 +278,47 @@ protected:
FLOG( FLog::Controller << "\tFinished (@Downward Pass (M2L) = " << counterTime.tacAndElapsed() << " s)\n" );
}
+ void transferPassWithoutFinalize(){
+ FLOG( FLog::Controller.write("\tStart Downward Pass (M2L)\n").write(FLog::Flush); );
+ FLOG(FTic counterTime);
+
+ const CellClass* neighbors[343];
+
+ typename OctreeClass::Iterator octreeIterator(tree);
+ octreeIterator.moveDown();
+
+ for(int idxLevel = 2 ; idxLevel < FAbstractAlgorithm::upperWorkingLevel ; ++idxLevel){
+ octreeIterator.moveDown();
+ }
+
+ typename OctreeClass::Iterator avoidGotoLeftIterator(octreeIterator);
+ //
+ // for each levels
+ for(int idxLevel = FAbstractAlgorithm::upperWorkingLevel ; idxLevel < FAbstractAlgorithm::lowerWorkingLevel ; ++idxLevel ){
+ FLOG(FTic counterTimeLevel);
+ const int separationCriteria = (idxLevel != FAbstractAlgorithm::lowerWorkingLevel-1 ? 1 : leafLevelSeparationCriteria);
+ // for each cells
+ do{
+ const int counter = tree->getInteractionNeighbors(neighbors, octreeIterator.getCurrentGlobalCoordinate(), idxLevel, separationCriteria);
+ if(counter){
+ #pragma omp task firstprivate(octreeIterator, neighbors, counter,idxLevel)
+ {
+ kernels[omp_get_thread_num()]->M2L( octreeIterator.getCurrentCell() , neighbors, counter, idxLevel);
+ }
+ }
+
+ } while(octreeIterator.moveRight());
+
+ avoidGotoLeftIterator.moveDown();
+ octreeIterator = avoidGotoLeftIterator;
+
+ }
+
+ #pragma omp taskwait
+
+ FLOG( FLog::Controller << "\tFinished (@Downward Pass (M2L) = " << counterTime.tacAndElapsed() << " s)\n" );
+ }
+
/////////////////////////////////////////////////////////////////////////////
// Downward
/////////////////////////////////////////////////////////////////////////////
@@ -296,7 +342,7 @@ protected:
FLOG(FTic counterTimeLevel);
// for each cells
do{
- #pragma omp task firstprivate(octreeIterator) shared(idxLevel)
+ #pragma omp task firstprivate(octreeIterator,idxLevel)
{
kernels[omp_get_thread_num()]->L2L( octreeIterator.getCurrentCell() , octreeIterator.getCurrentChild(), idxLevel);
}
diff --git a/Src/Core/FFmmAlgorithmTask.hpp b/Src/Core/FFmmAlgorithmTask.hpp
index 88b11dbd8..eb6487954 100644
--- a/Src/Core/FFmmAlgorithmTask.hpp
+++ b/Src/Core/FFmmAlgorithmTask.hpp
@@ -209,18 +209,23 @@ protected:
/////////////////////////////////////////////////////////////////////////////
/** M2L */
- void transferPass(){
- #ifdef SCALFMM_USE_EZTRACE
-
- eztrace_start();
-#endif
- this->transferPassWithFinalize() ;
-#ifdef SCALFMM_USE_EZTRACE
- eztrace_stop();
-#endif
- }
-
- void transferPassWithOutFinalize(){
+ void transferPass(){
+ #ifdef SCALFMM_USE_EZTRACE
+
+ eztrace_start();
+ #endif
+ if(KernelClass::NeedFinishedM2LEvent()){
+ this->transferPassWithFinalize() ;
+ }
+ else{
+ this->transferPassWithoutFinalize() ;
+ }
+ #ifdef SCALFMM_USE_EZTRACE
+ eztrace_stop();
+ #endif
+ }
+
+ void transferPassWithoutFinalize(){
FLOG( FLog::Controller.write("\tStart Downward Pass (M2L)\n").write(FLog::Flush); );
FLOG(FTic counterTime);
diff --git a/Src/Core/FFmmAlgorithmThreadProc.hpp b/Src/Core/FFmmAlgorithmThreadProc.hpp
index dee7f4cfc..af16d7b8d 100644
--- a/Src/Core/FFmmAlgorithmThreadProc.hpp
+++ b/Src/Core/FFmmAlgorithmThreadProc.hpp
@@ -859,7 +859,7 @@ protected:
#pragma omp taskwait
for(int idxThread = 0 ; idxThread < omp_get_num_threads() ; ++idxThread){
- #pragma omp task default(none) firstprivate(idxThread) shared(idxLevel)
+ #pragma omp task default(none) firstprivate(idxThread,idxLevel)
{
kernels[idxThread]->finishedLevelM2L(idxLevel);
}
diff --git a/Src/Core/FFmmAlgorithmThreadProcPeriodic.hpp b/Src/Core/FFmmAlgorithmThreadProcPeriodic.hpp
index ce9690d55..e6fef2877 100644
--- a/Src/Core/FFmmAlgorithmThreadProcPeriodic.hpp
+++ b/Src/Core/FFmmAlgorithmThreadProcPeriodic.hpp
@@ -960,7 +960,7 @@ protected:
#pragma omp taskwait
for(int idxThread = 0 ; idxThread < omp_get_num_threads() ; ++idxThread){
- #pragma omp task default(none) firstprivate(idxThread) shared(idxLevel)
+ #pragma omp task default(none) firstprivate(idxThread,idxLevel)
{
kernels[idxThread]->finishedLevelM2L(fackLevel);
}
diff --git a/Src/Kernels/Spherical/FSphericalBlockBlasKernel.hpp b/Src/Kernels/Spherical/FSphericalBlockBlasKernel.hpp
index cf5a4654c..c64b7b72b 100644
--- a/Src/Kernels/Spherical/FSphericalBlockBlasKernel.hpp
+++ b/Src/Kernels/Spherical/FSphericalBlockBlasKernel.hpp
@@ -189,7 +189,7 @@ public:
}
}
/** the needFinishedM2LEvent method is needed */
- bool needFinishedM2LEvent(){
+ constexpr static bool NeedFinishedM2LEvent(){
return true ;
}
/** Do we have some computation to do in the buffers */
--
GitLab