xref: /petsc/src/mat/impls/aij/seq/seqcusparse/cusparsematimpl.h (revision fe998a80077c9ee0917a39496df43fc256e1b478)
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 #if defined(PETSC_USE_COMPLEX)
16 #if defined(PETSC_USE_REAL_SINGLE)
17 #define cusparse_solve    cusparseCcsrsv_solve
18 #define cusparse_analysis cusparseCcsrsv_analysis
19 #define cusparse_csr_spmv cusparseCcsrmv
20 #define cusparse_csr2csc  cusparseCcsr2csc
21 #define cusparse_hyb_spmv cusparseChybmv
22 #define cusparse_csr2hyb  cusparseCcsr2hyb
23 #define cusparse_hyb2csr  cusparseChyb2csr
24 cuFloatComplex ALPHA = {1.0f, 0.0f};
25 cuFloatComplex BETA  = {0.0f, 0.0f};
26 #elif defined(PETSC_USE_REAL_DOUBLE)
27 #define cusparse_solve    cusparseZcsrsv_solve
28 #define cusparse_analysis cusparseZcsrsv_analysis
29 #define cusparse_csr_spmv cusparseZcsrmv
30 #define cusparse_csr2csc  cusparseZcsr2csc
31 #define cusparse_hyb_spmv cusparseZhybmv
32 #define cusparse_csr2hyb  cusparseZcsr2hyb
33 #define cusparse_hyb2csr  cusparseZhyb2csr
34 cuDoubleComplex ALPHA = {1.0, 0.0};
35 cuDoubleComplex BETA  = {0.0, 0.0};
36 #endif
37 #else
38 PetscScalar ALPHA = 1.0;
39 PetscScalar BETA  = 0.0;
40 #if defined(PETSC_USE_REAL_SINGLE)
41 #define cusparse_solve    cusparseScsrsv_solve
42 #define cusparse_analysis cusparseScsrsv_analysis
43 #define cusparse_csr_spmv cusparseScsrmv
44 #define cusparse_csr2csc  cusparseScsr2csc
45 #define cusparse_hyb_spmv cusparseShybmv
46 #define cusparse_csr2hyb  cusparseScsr2hyb
47 #define cusparse_hyb2csr  cusparseShyb2csr
48 #elif defined(PETSC_USE_REAL_DOUBLE)
49 #define cusparse_solve    cusparseDcsrsv_solve
50 #define cusparse_analysis cusparseDcsrsv_analysis
51 #define cusparse_csr_spmv cusparseDcsrmv
52 #define cusparse_csr2csc  cusparseDcsr2csc
53 #define cusparse_hyb_spmv cusparseDhybmv
54 #define cusparse_csr2hyb  cusparseDcsr2hyb
55 #define cusparse_hyb2csr  cusparseDhyb2csr
56 #endif
57 #endif
58 
59 #define THRUSTINTARRAY32 thrust::device_vector<int>
60 #define THRUSTINTARRAY thrust::device_vector<PetscInt>
61 #define THRUSTARRAY thrust::device_vector<PetscScalar>
62 
63 /* A CSR matrix structure */
64 struct CsrMatrix {
65   PetscInt         num_rows;
66   PetscInt         num_cols;
67   PetscInt         num_entries;
68   THRUSTINTARRAY32 *row_offsets;
69   THRUSTINTARRAY32 *column_indices;
70   THRUSTARRAY      *values;
71 };
72 
73 //#define CUSPMATRIXCSR32 cusp::csr_matrix<int,PetscScalar,cusp::device_memory>
74 
75 /* This is struct holding the relevant data needed to a MatSolve */
76 struct Mat_SeqAIJCUSPARSETriFactorStruct {
77   /* Data needed for triangular solve */
78   cusparseMatDescr_t          descr;
79   cusparseSolveAnalysisInfo_t solveInfo;
80   cusparseOperation_t         solveOp;
81   CsrMatrix                   *csrMat;
82 };
83 
84 /* This is struct holding the relevant data needed to a MatMult */
85 struct Mat_SeqAIJCUSPARSEMultStruct {
86   void               *mat;  /* opaque pointer to a matrix. This could be either a cusparseHybMat_t or a CsrMatrix */
87   cusparseMatDescr_t descr; /* Data needed to describe the matrix for a multiply */
88   THRUSTINTARRAY     *cprowIndices;   /* compressed row indices used in the parallel SpMV */
89   PetscScalar        *alpha; /* pointer to a device "scalar" storing the alpha parameter in the SpMV */
90   PetscScalar        *beta; /* pointer to a device "scalar" storing the beta parameter in the SpMV */
91 };
92 
93 /* This is a larger struct holding all the triangular factors for a solve, transpose solve, and
94  any indices used in a reordering */
95 struct Mat_SeqAIJCUSPARSETriFactors {
96   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorPtr; /* pointer for lower triangular (factored matrix) on GPU */
97   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorPtr; /* pointer for upper triangular (factored matrix) on GPU */
98   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorPtrTranspose; /* pointer for lower triangular (factored matrix) on GPU for the transpose (useful for BiCG) */
99   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorPtrTranspose; /* pointer for upper triangular (factored matrix) on GPU for the transpose (useful for BiCG)*/
100   THRUSTINTARRAY                    *rpermIndices;  /* indices used for any reordering */
101   THRUSTINTARRAY                    *cpermIndices;  /* indices used for any reordering */
102   THRUSTARRAY                       *workVector;
103   cusparseHandle_t                  handle;   /* a handle to the cusparse library */
104   PetscInt                          nnz;      /* number of nonzeros ... need this for accurate logging between ICC and ILU */
105 };
106 
107 /* This is a larger struct holding all the matrices for a SpMV, and SpMV Tranpose */
108 struct Mat_SeqAIJCUSPARSE {
109   Mat_SeqAIJCUSPARSEMultStruct *mat; /* pointer to the matrix on the GPU */
110   Mat_SeqAIJCUSPARSEMultStruct *matTranspose; /* pointer to the matrix on the GPU (for the transpose ... useful for BiCG) */
111   THRUSTARRAY                  *workVector; /*pointer to a workvector to which we can copy the relevant indices of a vector we want to multiply */
112   PetscInt                     nonzerorow; /* number of nonzero rows ... used in the flop calculations */
113   MatCUSPARSEStorageFormat     format;   /* the storage format for the matrix on the device */
114   cudaStream_t                 stream;   /* a stream for the parallel SpMV ... this is not owned and should not be deleted */
115   cusparseHandle_t             handle;   /* a handle to the cusparse library ... this may not be owned (if we're working in parallel i.e. multiGPUs) */
116 };
117 
118 PETSC_INTERN PetscErrorCode MatCUSPARSECopyToGPU(Mat);
119 PETSC_INTERN PetscErrorCode MatCUSPARSESetStream(Mat, const cudaStream_t stream);
120 PETSC_INTERN PetscErrorCode MatCUSPARSESetHandle(Mat, const cusparseHandle_t handle);
121 PETSC_INTERN PetscErrorCode MatCUSPARSEClearHandle(Mat);
122 #endif
123