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

First commit

Browse Source
This commit is contained in:
Jose Eduardo Bueno
2020-09-24 16:34:46 -03:00
parent c43648076f
commit 1ba81cece2
10 changed files with 317 additions and 258 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
opm/simulators/linalg/bda/openclSolverBackend.hpp
View File

@@ -21,11 +21,10 @@
#define OPM_OPENCLSOLVER_BACKEND_HEADER_INCLUDED
#include <opm/simulators/linalg/bda/opencl.hpp>
#include <opm/simulators/linalg/bda/BdaResult.hpp>
#include <opm/simulators/linalg/bda/BdaSolver.hpp>
#include <opm/simulators/linalg/bda/WellContributions.hpp>
#include <opm/simulators/linalg/bda/WellContributionsOCLContainer.hpp>
#include <opm/simulators/linalg/bda/BILU0.hpp>
namespace bda
@@ -35,9 +34,9 @@ namespace bda
template <unsigned int block_size>
class openclSolverBackend : public BdaSolver<block_size>
{
typedef BdaSolver<block_size> Base;
typedef BILU0<block_size> Preconditioner;
typedef WellContributionsOCLContainer WContainer;
using Base::N;
using Base::Nb;
@@ -51,7 +50,6 @@ class openclSolverBackend : public BdaSolver<block_size>
using Base::initialized;
private:
double *rb = nullptr; // reordered b vector, the matrix is reordered, so b must also be
double *vals_contiguous = nullptr; // only used if COPY_ROW_BY_ROW is true in openclSolverBackend.cpp
@@ -64,13 +62,6 @@ private:
cl::Buffer d_tmp; // used as tmp GPU buffer for dot() and norm()
double *tmp = nullptr; // used as tmp CPU buffer for dot() and norm()
unsigned int num_blocks, dim_weqs, dim_wells, num_std_wells;
unsigned int *h_val_pointers;
int *h_Ccols, *h_Bcols;
double *h_Cnnzs, *h_Dnnzs, *h_Bnnzs;
cl::Buffer d_Cnnzs, d_Dnnzs, d_Bnnzs;
cl::Buffer d_Ccols, d_Bcols, d_val_pointers;
// shared pointers are also passed to other objects
cl::Program program;
std::shared_ptr<cl::Context> context;
@@ -82,9 +73,13 @@ private:
std::unique_ptr<cl::make_kernel<cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int, cl::Buffer&, cl::Buffer&, const unsigned int, cl::LocalSpaceArg> > spmv_blocked_k;
std::shared_ptr<cl::make_kernel<cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int, cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int, const unsigned int, cl::LocalSpaceArg> > ILU_apply1_k;
std::shared_ptr<cl::make_kernel<cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int, cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int, const unsigned int, cl::LocalSpaceArg> > ILU_apply2_k;
std::shared_ptr<cl::make_kernel<cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&, const unsigned int, const unsigned int, cl::Buffer&, cl::LocalSpaceArg, cl::LocalSpaceArg, cl::LocalSpaceArg> > add_well_contributions_k;
std::shared_ptr<cl::make_kernel<cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
cl::Buffer&, cl::Buffer&, cl::Buffer&, cl::Buffer&,
const unsigned int, const unsigned int, cl::Buffer&,
cl::LocalSpaceArg, cl::LocalSpaceArg, cl::LocalSpaceArg> > stdwell_apply_k;
Preconditioner *prec = nullptr; // only supported preconditioner is BILU0
WContainer *wcontainer = nullptr;
int *toOrder = nullptr, *fromOrder = nullptr; // BILU0 reorders rows of the matrix via these mappings
std::unique_ptr<BlockedMatrix<block_size> > mat = nullptr; // original matrix
BlockedMatrix<block_size> *rmat = nullptr; // reordered matrix, used for spmv
@@ -148,17 +143,13 @@ private:
/// \param[in] vals array of nonzeroes, each block is stored row-wise and contiguous, contains nnz values
/// \param[in] rows array of rowPointers, contains N/dim+1 values
/// \param[in] cols array of columnIndices, contains nnz values
void initialize(int N, int nnz, int dim, double *vals, int *rows, int *cols);
void initialize_wellContribs(WellContributions& wellContribs);
void initialize(int N, int nnz, int dim, double *vals, int *rows, int *cols, WellContributions &wellContribs);
/// Clean memory
void finalize();
/// Copy linear system to GPU
void copy_system_to_gpu();
void copy_wells_to_gpu(WellContributions& wellContribs);
void copy_system_to_gpu(WellContributions &wellContribs);
/// Reorder the linear system so it corresponds with the coloring
/// \param[in] vals array of nonzeroes, each block is stored row-wise and contiguous, contains nnz values
@@ -166,7 +157,7 @@ private:
void update_system(double *vals, double *b);
/// Update linear system on GPU, don't copy rowpointers and colindices, they stay the same
void update_system_on_gpu();
void update_system_on_gpu(WellContributions &wellContribs);
/// Analyse sparsity pattern to extract parallelism
/// \return true iff analysis was successful
@@ -179,7 +170,7 @@ private:
/// Solve linear system
/// \param[in] wellContribs WellContributions, to apply them separately, instead of adding them to matrix A
/// \param[inout] res summary of solver result
void solve_system(BdaResult &res);
void solve_system(WellContributions &wellContribs, BdaResult &res);
public: