OilfieldToolsNet/opm-simulators
4
0
Fork 0
mirror of https://github.com/OPM/opm-simulators.git synced 2025-02-25 18:55:30 -06:00
Code Issues Packages Projects Releases Wiki Activity

add support for single thread copy

Browse Source
This commit is contained in:
Razvan Nane 2024-04-12 20:17:38 +02:00
parent b0157def17
commit cc1dfca9e0
3 changed files with 26 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
opm/simulators/linalg/ISTLSolverBda.cpp
View File

@ -45,6 +45,7 @@
#if HAVE_OPENMP
#include <thread>
#include <omp.h>
std::shared_ptr<std::thread> copyThread;
#endif // HAVE_OPENMP
@ -113,15 +114,21 @@ apply(Vector& rhs,
}
#endif
if (numJacobiBlocks_ > 1) {
bool use_multithreading = false;
#if HAVE_OPENMP
use_multithreading = omp_get_max_threads() > 1;
#endif // HAVE_OPENMP
if (numJacobiBlocks_ > 1) {
if(use_multithreading) {
//NOTE: copyThread can safely write to jacMat because in solve_system both matrix and *blockJacobiForGPUILU0_ diagonal entries
//are checked and potentially overwritten in replaceZeroDiagonal() by mainThread. However, no matter the thread writing sequence,
//the final entry in jacMat is correct.
copyThread = std::make_shared<std::thread>([&](){this->copyMatToBlockJac(matrix, *blockJacobiForGPUILU0_);});
#else
}
else {
this->copyMatToBlockJac(matrix, *blockJacobiForGPUILU0_);
#endif
}
// Const_cast needed since the CUDA stuff overwrites values for better matrix condition..
bridge_->solve_system(&matrix, blockJacobiForGPUILU0_.get(),

3
opm/simulators/linalg/bda/cuda/cusparseSolverBackend.cu
View File

@ -40,6 +40,7 @@
#if HAVE_OPENMP
#include <thread>
#include <omp.h>
extern std::shared_ptr<std::thread> copyThread;
#endif // HAVE_OPENMP
@ -328,6 +329,7 @@ void cusparseSolverBackend<block_size>::copy_system_to_gpu(std::shared_ptr<Block
cudaMemcpyAsync(d_bVals, matrix->nnzValues, nnz * sizeof(double), cudaMemcpyHostToDevice, stream);
if (useJacMatrix) {
#if HAVE_OPENMP
if(omp_get_max_threads() > 1)
copyThread->join();
#endif
cudaMemcpyAsync(d_mVals, jacMatrix->nnzValues, nnzbs_prec * block_size * block_size * sizeof(double), cudaMemcpyHostToDevice, stream);
@ -372,6 +374,7 @@ void cusparseSolverBackend<block_size>::update_system_on_gpu(std::shared_ptr<Blo
cudaMemcpyAsync(d_bVals, matrix->nnzValues, nnz * sizeof(double), cudaMemcpyHostToDevice, stream);
if (useJacMatrix) {
#if HAVE_OPENMP
if(omp_get_max_threads() > 1)
copyThread->join();
#endif
cudaMemcpyAsync(d_mVals, jacMatrix->nnzValues, nnzbs_prec * block_size * block_size * sizeof(double), cudaMemcpyHostToDevice, stream);

3
opm/simulators/linalg/bda/rocsparseSolverBackend.cpp
View File

@ -89,6 +89,7 @@
#if HAVE_OPENMP
#include <thread>
#include <omp.h>
extern std::shared_ptr<std::thread> copyThread;
#endif //HAVE_OPENMP
@ -441,6 +442,7 @@ void rocsparseSolverBackend<block_size>::copy_system_to_gpu(double *b) {
if (useJacMatrix) {
#if HAVE_OPENMP
if(omp_get_max_threads() > 1)
copyThread->join();
#endif
HIP_CHECK(hipMemcpyAsync(d_Mrows, jacMat->rowPointers, sizeof(rocsparse_int) * (Nb + 1), hipMemcpyHostToDevice, stream));
@ -472,6 +474,7 @@ void rocsparseSolverBackend<block_size>::update_system_on_gpu(double *b) {
if (useJacMatrix) {
#if HAVE_OPENMP
if (omp_get_max_threads() > 1)
copyThread->join();
#endif
HIP_CHECK(hipMemcpyAsync(d_Mvals, jacMat->nnzValues, sizeof(double) * nnzbs_prec * block_size * block_size, hipMemcpyHostToDevice, stream));