OilfieldToolsNet/opm-simulators
4
0
Fork 0
mirror of https://github.com/OPM/opm-simulators.git synced 2025-02-02 09:50:30 -06:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #5256 from BigDataAccelerate/overlap_copy_jacmat_gpu

Browse Source 
Overlap creation of jacobian matrix with GPU data transfers
This commit is contained in:
Bård Skaflestad 2024-05-24 13:11:27 +02:00 committed by GitHub
commit 6f094c558a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 88 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
opm/simulators/linalg/ISTLSolverBda.cpp
View File

@ -43,6 +43,13 @@
#include <opm/grid/polyhedralgrid.hh>
#if HAVE_OPENMP
#include <thread>
#include <omp.h>
std::shared_ptr<std::thread> copyThread;
#endif // HAVE_OPENMP
namespace Opm {
namespace detail {
@ -107,14 +114,28 @@ apply(Vector& rhs,
}
#endif
bool use_multithreading = false;
#if HAVE_OPENMP
use_multithreading = omp_get_max_threads() > 1;
#endif // HAVE_OPENMP
if (numJacobiBlocks_ > 1) {
this->copyMatToBlockJac(matrix, *blockJacobiForGPUILU0_);
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 {
this->copyMatToBlockJac(matrix, *blockJacobiForGPUILU0_);
}
// Const_cast needed since the CUDA stuff overwrites values for better matrix condition..
bridge_->solve_system(&matrix, blockJacobiForGPUILU0_.get(),
numJacobiBlocks_, rhs, *wellContribs, result);
}
else
bridge_->solve_system(&matrix, &matrix,
bridge_->solve_system(&matrix, &matrix,
numJacobiBlocks_, rhs, *wellContribs, result);
if (result.converged) {
// get result vector x from non-Dune backend, iff solve was successful

42
opm/simulators/linalg/bda/cuda/cusparseSolverBackend.cu
View File

@ -38,6 +38,12 @@
// otherwise, the nonzeroes of the matrix are assumed to be in a contiguous array, and a single GPU memcpy is enough
#define COPY_ROW_BY_ROW 0
#if HAVE_OPENMP
#include <thread>
#include <omp.h>
extern std::shared_ptr<std::thread> copyThread;
#endif // HAVE_OPENMP
namespace Opm
{
namespace Accelerator
@ -306,6 +312,11 @@ template <unsigned int block_size>
void cusparseSolverBackend<block_size>::copy_system_to_gpu(std::shared_ptr<BlockedMatrix> matrix, double *b, std::shared_ptr<BlockedMatrix> jacMatrix) {
Timer t;
cudaMemcpyAsync(d_bCols, matrix->colIndices, nnzb * sizeof(int), cudaMemcpyHostToDevice, stream);
cudaMemcpyAsync(d_bRows, matrix->rowPointers, (Nb + 1) * sizeof(int), cudaMemcpyHostToDevice, stream);
cudaMemcpyAsync(d_b, b, N * sizeof(double), cudaMemcpyHostToDevice, stream);
cudaMemsetAsync(d_x, 0, sizeof(double) * N, stream);
#if COPY_ROW_BY_ROW
int sum = 0;
for (int i = 0; i < Nb; ++i) {
@ -317,25 +328,27 @@ void cusparseSolverBackend<block_size>::copy_system_to_gpu(std::shared_ptr<Block
#else
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);
} else {
cudaMemcpyAsync(d_mVals, d_bVals, nnz * sizeof(double), cudaMemcpyDeviceToDevice, stream);
}
#endif
cudaMemcpyAsync(d_bCols, matrix->colIndices, nnzb * sizeof(int), cudaMemcpyHostToDevice, stream);
cudaMemcpyAsync(d_bRows, matrix->rowPointers, (Nb + 1) * sizeof(int), cudaMemcpyHostToDevice, stream);
if (useJacMatrix) {
cudaMemcpyAsync(d_mCols, jacMatrix->colIndices, nnzbs_prec * sizeof(int), cudaMemcpyHostToDevice, stream);
cudaMemcpyAsync(d_mRows, jacMatrix->rowPointers, (Nb + 1) * sizeof(int), cudaMemcpyHostToDevice, stream);
}
cudaMemcpyAsync(d_b, b, N * sizeof(double), cudaMemcpyHostToDevice, stream);
cudaMemsetAsync(d_x, 0, sizeof(double) * N, stream);
if (verbosity > 2) {
if (verbosity >= 3) {
cudaStreamSynchronize(stream);
c_copy += t.stop();
std::ostringstream out;
out << "cusparseSolver::copy_system_to_gpu(): " << t.stop() << " s";
out << "---cusparseSolver::copy_system_to_gpu(): " << t.elapsed() << " s";
OpmLog::info(out.str());
}
} // end copy_system_to_gpu()
@ -346,6 +359,9 @@ template <unsigned int block_size>
void cusparseSolverBackend<block_size>::update_system_on_gpu(std::shared_ptr<BlockedMatrix> matrix, double *b, std::shared_ptr<BlockedMatrix> jacMatrix) {
Timer t;
cudaMemcpyAsync(d_b, b, N * sizeof(double), cudaMemcpyHostToDevice, stream);
cudaMemsetAsync(d_x, 0, sizeof(double) * N, stream);
#if COPY_ROW_BY_ROW
int sum = 0;
for (int i = 0; i < Nb; ++i) {
@ -357,19 +373,23 @@ void cusparseSolverBackend<block_size>::update_system_on_gpu(std::shared_ptr<Blo
#else
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);
} else {
cudaMemcpyAsync(d_mVals, d_bVals, nnz * sizeof(double), cudaMemcpyDeviceToDevice, stream);
}
#endif
cudaMemcpyAsync(d_b, b, N * sizeof(double), cudaMemcpyHostToDevice, stream);
cudaMemsetAsync(d_x, 0, sizeof(double) * N, stream);
if (verbosity > 2) {
if (verbosity >= 3) {
cudaStreamSynchronize(stream);
c_copy += t.stop();
std::ostringstream out;
out << "cusparseSolver::update_system_on_gpu(): " << t.stop() << " s";
out << "-----cusparseSolver::update_system_on_gpu(): " << t.elapsed() << " s\n";
out << "---cusparseSolver::cum copy: " << c_copy << " s";
OpmLog::info(out.str());
}
} // end update_system_on_gpu()

2
opm/simulators/linalg/bda/cuda/cusparseSolverBackend.hpp
View File

@ -71,6 +71,8 @@ private:
bool useJacMatrix = false;
int nnzbs_prec; // number of nonzero blocks in the matrix for preconditioner
// could be jacMatrix or matrix
double c_copy = 0.0; // cummulative timer measuring the total time it takes to transfer the data to the GPU
/// Solve linear system using ilu0-bicgstab
/// \param[in] wellContribs contains all WellContributions, to apply them separately, instead of adding them to matrix A

38
opm/simulators/linalg/bda/rocsparseSolverBackend.cpp
View File

@ -87,6 +87,12 @@
#include <cstddef>
#if HAVE_OPENMP
#include <thread>
#include <omp.h>
extern std::shared_ptr<std::thread> copyThread;
#endif //HAVE_OPENMP
namespace Opm
{
namespace Accelerator
@ -431,21 +437,29 @@ void rocsparseSolverBackend<block_size>::copy_system_to_gpu(double *b) {
HIP_CHECK(hipMemcpyAsync(d_Arows, mat->rowPointers, sizeof(rocsparse_int) * (Nb + 1), hipMemcpyHostToDevice, stream));
HIP_CHECK(hipMemcpyAsync(d_Acols, mat->colIndices, sizeof(rocsparse_int) * nnzb, hipMemcpyHostToDevice, stream));
HIP_CHECK(hipMemcpyAsync(d_Avals, mat->nnzValues, sizeof(double) * nnz, hipMemcpyHostToDevice, stream));
HIP_CHECK(hipMemsetAsync(d_x, 0, sizeof(double) * N, stream));
HIP_CHECK(hipMemcpyAsync(d_b, b, sizeof(double) * N, hipMemcpyHostToDevice, stream));
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));
HIP_CHECK(hipMemcpyAsync(d_Mcols, jacMat->colIndices, sizeof(rocsparse_int) * nnzbs_prec, hipMemcpyHostToDevice, stream));
HIP_CHECK(hipMemcpyAsync(d_Mvals, jacMat->nnzValues, sizeof(double) * nnzbs_prec * block_size * block_size, hipMemcpyHostToDevice, stream));
} else {
HIP_CHECK(hipMemcpyAsync(d_Mvals, d_Avals, sizeof(double) * nnz, hipMemcpyDeviceToDevice, stream));
}
HIP_CHECK(hipMemsetAsync(d_x, 0, sizeof(double) * N, stream));
HIP_CHECK(hipMemcpyAsync(d_b, b, sizeof(double) * N, hipMemcpyHostToDevice, stream));
if (verbosity >= 3) {
HIP_CHECK(hipStreamSynchronize(stream));
c_copy += t.stop();
std::ostringstream out;
out << "rocsparseSolver::copy_system_to_gpu(): " << t.stop() << " s";
OpmLog::info(out.str());
out << "-----rocsparseSolver::copy_system_to_gpu(): " << t.elapsed() << " s\n";
out << "---rocsparseSolver::cum copy: " << c_copy << " s";
OpmLog::info(out.str());
}
} // end copy_system_to_gpu()
@ -455,18 +469,26 @@ void rocsparseSolverBackend<block_size>::update_system_on_gpu(double *b) {
Timer t;
HIP_CHECK(hipMemcpyAsync(d_Avals, mat->nnzValues, sizeof(double) * nnz, hipMemcpyHostToDevice, stream));
HIP_CHECK(hipMemsetAsync(d_x, 0, sizeof(double) * N, stream));
HIP_CHECK(hipMemcpyAsync(d_b, b, sizeof(double) * N, hipMemcpyHostToDevice, stream));
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));
} else {
HIP_CHECK(hipMemcpyAsync(d_Mvals, d_Avals, sizeof(double) * nnz, hipMemcpyDeviceToDevice, stream));
}
HIP_CHECK(hipMemsetAsync(d_x, 0, sizeof(double) * N, stream));
HIP_CHECK(hipMemcpyAsync(d_b, b, sizeof(double) * N, hipMemcpyHostToDevice, stream));
if (verbosity >= 3) {
HIP_CHECK(hipStreamSynchronize(stream));
c_copy += t.stop();
std::ostringstream out;
out << "rocsparseSolver::update_system_on_gpu(): " << t.stop() << " s";
out << "-----rocsparseSolver::update_system_on_gpu(): " << t.elapsed() << " s\n";
out << "---rocsparseSolver::cum copy: " << c_copy << " s";
OpmLog::info(out.str());
}
} // end update_system_on_gpu()

2
opm/simulators/linalg/bda/rocsparseSolverBackend.hpp
View File

@ -55,6 +55,8 @@ class rocsparseSolverBackend : public BdaSolver<block_size>
private:
double c_copy = 0.0; // cummulative timer measuring the total time it takes to transfer the data to the GPU
bool useJacMatrix = false;
bool analysis_done = false;