1 #if !defined(__CUSPARSEMATIMPL) 2 #define __CUSPARSEMATIMPL 3 4 #include <../src/vec/vec/impls/seq/seqcusp/cuspvecimpl.h> 5 6 #if CUDA_VERSION>=4020 7 #include <cusparse_v2.h> 8 #else 9 #include <cusparse.h> 10 #endif 11 12 #include <algorithm> 13 #include <vector> 14 15 /* Single instance of the cusparse handle for the class. */ 16 MatCUSPARSEStorageFormat cusparseMatSolveStorageFormat=MAT_CUSPARSE_CSR; 17 18 #if defined(PETSC_USE_COMPLEX) 19 #if defined(PETSC_USE_REAL_SINGLE) 20 #define cusparse_solve cusparseCcsrsv_solve 21 #define cusparse_analysis cusparseCcsrsv_analysis 22 #define cusparse_csr_spmv cusparseCcsrmv 23 #define cusparse_csr2csc cusparseCcsr2csc 24 #define cusparse_hyb_spmv cusparseChybmv 25 #define cusparse_csr2hyb cusparseCcsr2hyb 26 #define cusparse_hyb2csr cusparseChyb2csr 27 cuFloatComplex ALPHA = {1.0f, 0.0f}; 28 cuFloatComplex BETA = {0.0f, 0.0f}; 29 #elif defined(PETSC_USE_REAL_DOUBLE) 30 #define cusparse_solve cusparseZcsrsv_solve 31 #define cusparse_analysis cusparseZcsrsv_analysis 32 #define cusparse_csr_spmv cusparseZcsrmv 33 #define cusparse_csr2csc cusparseZcsr2csc 34 #define cusparse_hyb_spmv cusparseZhybmv 35 #define cusparse_csr2hyb cusparseZcsr2hyb 36 #define cusparse_hyb2csr cusparseZhyb2csr 37 cuDoubleComplex ALPHA = {1.0, 0.0}; 38 cuDoubleComplex BETA = {0.0, 0.0}; 39 #endif 40 #else 41 PetscScalar ALPHA = 1.0; 42 PetscScalar BETA = 0.0; 43 #if defined(PETSC_USE_REAL_SINGLE) 44 #define cusparse_solve cusparseScsrsv_solve 45 #define cusparse_analysis cusparseScsrsv_analysis 46 #define cusparse_csr_spmv cusparseScsrmv 47 #define cusparse_csr2csc cusparseScsr2csc 48 #define cusparse_hyb_spmv cusparseShybmv 49 #define cusparse_csr2hyb cusparseScsr2hyb 50 #define cusparse_hyb2csr cusparseShyb2csr 51 #elif defined(PETSC_USE_REAL_DOUBLE) 52 #define cusparse_solve cusparseDcsrsv_solve 53 #define cusparse_analysis cusparseDcsrsv_analysis 54 #define cusparse_csr_spmv cusparseDcsrmv 55 #define cusparse_csr2csc cusparseDcsr2csc 56 #define cusparse_hyb_spmv cusparseDhybmv 57 #define cusparse_csr2hyb cusparseDcsr2hyb 58 #define cusparse_hyb2csr cusparseDhyb2csr 59 #endif 60 #endif 61 62 #define THRUSTINTARRAY32 thrust::device_vector<int> 63 #define THRUSTINTARRAY thrust::device_vector<PetscInt> 64 #define THRUSTARRAY thrust::device_vector<PetscScalar> 65 66 /* A CSR matrix structure */ 67 struct CsrMatrix { 68 PetscInt num_rows; 69 PetscInt num_cols; 70 PetscInt num_entries; 71 THRUSTINTARRAY32 *row_offsets; 72 THRUSTINTARRAY32 *column_indices; 73 THRUSTARRAY *values; 74 }; 75 76 //#define CUSPMATRIXCSR32 cusp::csr_matrix<int,PetscScalar,cusp::device_memory> 77 78 /* This is struct holding the relevant data needed to a MatSolve */ 79 struct Mat_SeqAIJCUSPARSETriFactorStruct { 80 /* Data needed for triangular solve */ 81 cusparseMatDescr_t descr; 82 cusparseSolveAnalysisInfo_t solveInfo; 83 cusparseOperation_t solveOp; 84 CsrMatrix *csrMat; 85 }; 86 87 /* This is struct holding the relevant data needed to a MatMult */ 88 struct Mat_SeqAIJCUSPARSEMultStruct { 89 void *mat; /* opaque pointer to a matrix. This could be either a cusparseHybMat_t or a CsrMatrix */ 90 cusparseMatDescr_t descr; /* Data needed to describe the matrix for a multiply */ 91 THRUSTINTARRAY *cprowIndices; /* compressed row indices used in the parallel SpMV */ 92 PetscScalar *alpha; /* pointer to a device "scalar" storing the alpha parameter in the SpMV */ 93 PetscScalar *beta; /* pointer to a device "scalar" storing the beta parameter in the SpMV */ 94 }; 95 96 /* This is a larger struct holding all the triangular factors for a solve, transpose solve, and 97 any indices used in a reordering */ 98 struct Mat_SeqAIJCUSPARSETriFactors { 99 Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorPtr; /* pointer for lower triangular (factored matrix) on GPU */ 100 Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorPtr; /* pointer for upper triangular (factored matrix) on GPU */ 101 Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorPtrTranspose; /* pointer for lower triangular (factored matrix) on GPU for the transpose (useful for BiCG) */ 102 Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorPtrTranspose; /* pointer for upper triangular (factored matrix) on GPU for the transpose (useful for BiCG)*/ 103 THRUSTINTARRAY *rpermIndices; /* indices used for any reordering */ 104 THRUSTINTARRAY *cpermIndices; /* indices used for any reordering */ 105 THRUSTARRAY *workVector; 106 MatCUSPARSEStorageFormat format; /* the storage format for the matrix on the device */ 107 cusparseHandle_t handle; /* a handle to the cusparse library */ 108 PetscInt nnz; /* number of nonzeros ... need this for accurate logging between ICC and ILU */ 109 }; 110 111 /* This is a larger struct holding all the matrices for a SpMV, and SpMV Tranpose */ 112 struct Mat_SeqAIJCUSPARSE { 113 Mat_SeqAIJCUSPARSEMultStruct *mat; /* pointer to the matrix on the GPU */ 114 Mat_SeqAIJCUSPARSEMultStruct *matTranspose; /* pointer to the matrix on the GPU (for the transpose ... useful for BiCG) */ 115 THRUSTARRAY *workVector; /*pointer to a workvector to which we can copy the relevant indices of a vector we want to multiply */ 116 PetscInt nonzerorow; /* number of nonzero rows ... used in the flop calculations */ 117 MatCUSPARSEStorageFormat format; /* the storage format for the matrix on the device */ 118 cudaStream_t stream; /* a stream for the parallel SpMV ... this is not owned and should not be deleted */ 119 cusparseHandle_t handle; /* a handle to the cusparse library ... this may not be owned (if we're working in parallel i.e. multiGPUs) */ 120 }; 121 122 PETSC_INTERN PetscErrorCode MatCUSPARSECopyToGPU(Mat); 123 PETSC_INTERN PetscErrorCode MatCUSPARSESetStream(Mat, const cudaStream_t stream); 124 PETSC_INTERN PetscErrorCode MatCUSPARSESetHandle(Mat, const cusparseHandle_t handle); 125 PETSC_INTERN PetscErrorCode MatCUSPARSEClearHandle(Mat); 126 #endif 127