1519f805aSKarl Rupp #if !defined(__CUSPARSEMATIMPL) 29ae82921SPaul Mullowney #define __CUSPARSEMATIMPL 39ae82921SPaul Mullowney 49ae82921SPaul Mullowney #include <../src/vec/vec/impls/seq/seqcusp/cuspvecimpl.h> 59ae82921SPaul Mullowney 69ae82921SPaul Mullowney #include <cusparse_v2.h> 79ae82921SPaul Mullowney 89ae82921SPaul Mullowney /* New Way */ 99ae82921SPaul Mullowney #include "tx_sparse_interface.h" 109ae82921SPaul Mullowney 119ae82921SPaul Mullowney #include <algorithm> 129ae82921SPaul Mullowney #include <vector> 139ae82921SPaul Mullowney #include <string> 149ae82921SPaul Mullowney #include <thrust/sort.h> 159ae82921SPaul Mullowney #include <thrust/fill.h> 169ae82921SPaul Mullowney 1757eb53cdSKarl Rupp /* Single instance of the cusparse handle for the class. */ 189ae82921SPaul Mullowney cusparseHandle_t MAT_cusparseHandle=0; 199ae82921SPaul Mullowney 20e057df02SPaul Mullowney MatCUSPARSEStorageFormat cusparseMatSolveStorageFormat=MAT_CUSPARSE_CSR; 21e057df02SPaul Mullowney 229ae82921SPaul Mullowney struct Mat_SeqAIJCUSPARSETriFactors { 239ae82921SPaul Mullowney GPU_Matrix_Ifc *loTriFactorPtr; /* pointer for lower triangular (factored matrix) on GPU */ 249ae82921SPaul Mullowney GPU_Matrix_Ifc *upTriFactorPtr; /* pointer for upper triangular (factored matrix) on GPU */ 259ae82921SPaul Mullowney CUSPARRAY *tempvec; 26e057df02SPaul Mullowney MatCUSPARSEStorageFormat format; /* the storage format for the matrix on the device */ 27*bda325fcSPaul Mullowney PetscBool hasTranspose; /* boolean describing whether a transpose has been built or not */ 289ae82921SPaul Mullowney }; 299ae82921SPaul Mullowney 309ae82921SPaul Mullowney struct Mat_SeqAIJCUSPARSE { 319ae82921SPaul Mullowney GPU_Matrix_Ifc *mat; /* pointer to the matrix on the GPU */ 329ae82921SPaul Mullowney CUSPARRAY *tempvec; /*pointer to a workvector to which we can copy the relevant indices of a vector we want to multiply */ 339ae82921SPaul Mullowney PetscInt nonzerorow; /* number of nonzero rows ... used in the flop calculations */ 34e057df02SPaul Mullowney MatCUSPARSEStorageFormat format; /* the storage format for the matrix on the device */ 35*bda325fcSPaul Mullowney PetscBool hasTranspose; /* boolean describing whether a transpose has been built or not */ 369ae82921SPaul Mullowney }; 379ae82921SPaul Mullowney 389ae82921SPaul Mullowney extern PetscErrorCode MatCUSPARSECopyToGPU(Mat); 3957eb53cdSKarl Rupp /* extern PetscErrorCode MatGetFactor_seqaij_cusparse(Mat,MatFactorType,Mat*); */ 4057eb53cdSKarl Rupp /* extern PetscErrorCode MatFactorGetSolverPackage_seqaij_cusparse(Mat,const MatSolverPackage *); */ 4157eb53cdSKarl Rupp /* extern PetscErrorCode MatCUSPARSECopyFromGPU(Mat, CUSPMATRIX *); */ 429ae82921SPaul Mullowney #endif 43