1 #if !defined(CUSPARSEMATIMPL) 2 #define CUSPARSEMATIMPL 3 4 #include <petscpkg_version.h> 5 #include <petsc/private/cudavecimpl.h> 6 #include <petscaijdevice.h> 7 8 #include <cusparse_v2.h> 9 10 #include <algorithm> 11 #include <vector> 12 13 #include <thrust/device_vector.h> 14 #include <thrust/device_ptr.h> 15 #include <thrust/device_malloc_allocator.h> 16 #include <thrust/transform.h> 17 #include <thrust/functional.h> 18 #include <thrust/sequence.h> 19 #include <thrust/system/system_error.h> 20 21 #define PetscStackCallThrust(body) do { \ 22 try { \ 23 body; \ 24 } catch(thrust::system_error& e) { \ 25 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in Thrust %s",e.what());\ 26 } \ 27 } while (0) 28 29 #if defined(PETSC_USE_COMPLEX) 30 #if defined(PETSC_USE_REAL_SINGLE) 31 const cuComplex PETSC_CUSPARSE_ONE = {1.0f, 0.0f}; 32 const cuComplex PETSC_CUSPARSE_ZERO = {0.0f, 0.0f}; 33 #elif defined(PETSC_USE_REAL_DOUBLE) 34 const cuDoubleComplex PETSC_CUSPARSE_ONE = {1.0, 0.0}; 35 const cuDoubleComplex PETSC_CUSPARSE_ZERO = {0.0, 0.0}; 36 #endif 37 #else 38 const PetscScalar PETSC_CUSPARSE_ONE = 1.0; 39 const PetscScalar PETSC_CUSPARSE_ZERO = 0.0; 40 #endif 41 42 #if PETSC_PKG_CUDA_VERSION_GE(9,0,0) 43 #define cusparse_create_analysis_info cusparseCreateCsrsv2Info 44 #define cusparse_destroy_analysis_info cusparseDestroyCsrsv2Info 45 #if defined(PETSC_USE_COMPLEX) 46 #if defined(PETSC_USE_REAL_SINGLE) 47 #define cusparse_get_svbuffsize(a,b,c,d,e,f,g,h,i,j) cusparseCcsrsv2_bufferSize(a,b,c,d,e,(cuComplex*)(f),g,h,i,j) 48 #define cusparse_analysis(a,b,c,d,e,f,g,h,i,j,k) cusparseCcsrsv2_analysis(a,b,c,d,e,(const cuComplex*)(f),g,h,i,j,k) 49 #define cusparse_solve(a,b,c,d,e,f,g,h,i,j,k,l,m,n) cusparseCcsrsv2_solve(a,b,c,d,(const cuComplex*)(e),f,(const cuComplex*)(g),h,i,j,(const cuComplex*)(k),(cuComplex*)(l),m,n) 50 #elif defined(PETSC_USE_REAL_DOUBLE) 51 #define cusparse_get_svbuffsize(a,b,c,d,e,f,g,h,i,j) cusparseZcsrsv2_bufferSize(a,b,c,d,e,(cuDoubleComplex*)(f),g,h,i,j) 52 #define cusparse_analysis(a,b,c,d,e,f,g,h,i,j,k) cusparseZcsrsv2_analysis(a,b,c,d,e,(const cuDoubleComplex*)(f),g,h,i,j,k) 53 #define cusparse_solve(a,b,c,d,e,f,g,h,i,j,k,l,m,n) cusparseZcsrsv2_solve(a,b,c,d,(const cuDoubleComplex*)(e),f,(const cuDoubleComplex*)(g),h,i,j,(const cuDoubleComplex*)(k),(cuDoubleComplex*)(l),m,n) 54 #endif 55 #else /* not complex */ 56 #if defined(PETSC_USE_REAL_SINGLE) 57 #define cusparse_get_svbuffsize cusparseScsrsv2_bufferSize 58 #define cusparse_analysis cusparseScsrsv2_analysis 59 #define cusparse_solve cusparseScsrsv2_solve 60 #elif defined(PETSC_USE_REAL_DOUBLE) 61 #define cusparse_get_svbuffsize cusparseDcsrsv2_bufferSize 62 #define cusparse_analysis cusparseDcsrsv2_analysis 63 #define cusparse_solve cusparseDcsrsv2_solve 64 #endif 65 #endif 66 #else 67 #define cusparse_create_analysis_info cusparseCreateSolveAnalysisInfo 68 #define cusparse_destroy_analysis_info cusparseDestroySolveAnalysisInfo 69 #if defined(PETSC_USE_COMPLEX) 70 #if defined(PETSC_USE_REAL_SINGLE) 71 #define cusparse_solve(a,b,c,d,e,f,g,h,i,j,k) cusparseCcsrsv_solve((a),(b),(c),(cuComplex*)(d),(e),(cuComplex*)(f),(g),(h),(i),(cuComplex*)(j),(cuComplex*)(k)) 72 #define cusparse_analysis(a,b,c,d,e,f,g,h,i) cusparseCcsrsv_analysis((a),(b),(c),(d),(e),(cuComplex*)(f),(g),(h),(i)) 73 #elif defined(PETSC_USE_REAL_DOUBLE) 74 #define cusparse_solve(a,b,c,d,e,f,g,h,i,j,k) cusparseZcsrsv_solve((a),(b),(c),(cuDoubleComplex*)(d),(e),(cuDoubleComplex*)(f),(g),(h),(i),(cuDoubleComplex*)(j),(cuDoubleComplex*)(k)) 75 #define cusparse_analysis(a,b,c,d,e,f,g,h,i) cusparseZcsrsv_analysis((a),(b),(c),(d),(e),(cuDoubleComplex*)(f),(g),(h),(i)) 76 #endif 77 #else /* not complex */ 78 #if defined(PETSC_USE_REAL_SINGLE) 79 #define cusparse_solve cusparseScsrsv_solve 80 #define cusparse_analysis cusparseScsrsv_analysis 81 #elif defined(PETSC_USE_REAL_DOUBLE) 82 #define cusparse_solve cusparseDcsrsv_solve 83 #define cusparse_analysis cusparseDcsrsv_analysis 84 #endif 85 #endif 86 #endif 87 88 #if PETSC_PKG_CUDA_VERSION_GE(11,0,0) 89 #define cusparse_csr2csc cusparseCsr2cscEx2 90 #if defined(PETSC_USE_COMPLEX) 91 #if defined(PETSC_USE_REAL_SINGLE) 92 #define cusparse_scalartype CUDA_C_32F 93 #define cusparse_csr_spgeam(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t) cusparseCcsrgeam2(a,b,c,(cuComplex*)d,e,f,(cuComplex*)g,h,i,(cuComplex*)j,k,l,(cuComplex*)m,n,o,p,(cuComplex*)q,r,s,t) 94 #define cusparse_csr_spgeam_bufferSize(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t) cusparseCcsrgeam2_bufferSizeExt(a,b,c,(cuComplex*)d,e,f,(cuComplex*)g,h,i,(cuComplex*)j,k,l,(cuComplex*)m,n,o,p,(cuComplex*)q,r,s,t) 95 #elif defined(PETSC_USE_REAL_DOUBLE) 96 #define cusparse_scalartype CUDA_C_64F 97 #define cusparse_csr_spgeam(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t) cusparseZcsrgeam2(a,b,c,(cuDoubleComplex*)d,e,f,(cuDoubleComplex*)g,h,i,(cuDoubleComplex*)j,k,l,(cuDoubleComplex*)m,n,o,p,(cuDoubleComplex*)q,r,s,t) 98 #define cusparse_csr_spgeam_bufferSize(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t) cusparseZcsrgeam2_bufferSizeExt(a,b,c,(cuDoubleComplex*)d,e,f,(cuDoubleComplex*)g,h,i,(cuDoubleComplex*)j,k,l,(cuDoubleComplex*)m,n,o,p,(cuDoubleComplex*)q,r,s,t) 99 #endif 100 #else /* not complex */ 101 #if defined(PETSC_USE_REAL_SINGLE) 102 #define cusparse_scalartype CUDA_R_32F 103 #define cusparse_csr_spgeam cusparseScsrgeam2 104 #define cusparse_csr_spgeam_bufferSize cusparseScsrgeam2_bufferSizeExt 105 #elif defined(PETSC_USE_REAL_DOUBLE) 106 #define cusparse_scalartype CUDA_R_64F 107 #define cusparse_csr_spgeam cusparseDcsrgeam2 108 #define cusparse_csr_spgeam_bufferSize cusparseDcsrgeam2_bufferSizeExt 109 #endif 110 #endif 111 #else 112 #if defined(PETSC_USE_COMPLEX) 113 #if defined(PETSC_USE_REAL_SINGLE) 114 #define cusparse_csr_spmv(a,b,c,d,e,f,g,h,i,j,k,l,m) cusparseCcsrmv((a),(b),(c),(d),(e),(cuComplex*)(f),(g),(cuComplex*)(h),(i),(j),(cuComplex*)(k),(cuComplex*)(l),(cuComplex*)(m)) 115 #define cusparse_csr_spmm(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p) cusparseCcsrmm((a),(b),(c),(d),(e),(f),(cuComplex*)(g),(h),(cuComplex*)(i),(j),(k),(cuComplex*)(l),(m),(cuComplex*)(n),(cuComplex*)(o),(p)) 116 #define cusparse_csr2csc(a,b,c,d,e,f,g,h,i,j,k,l) cusparseCcsr2csc((a),(b),(c),(d),(cuComplex*)(e),(f),(g),(cuComplex*)(h),(i),(j),(k),(l)) 117 #define cusparse_hyb_spmv(a,b,c,d,e,f,g,h) cusparseChybmv((a),(b),(cuComplex*)(c),(d),(e),(cuComplex*)(f),(cuComplex*)(g),(cuComplex*)(h)) 118 #define cusparse_csr2hyb(a,b,c,d,e,f,g,h,i,j) cusparseCcsr2hyb((a),(b),(c),(d),(cuComplex*)(e),(f),(g),(h),(i),(j)) 119 #define cusparse_hyb2csr(a,b,c,d,e,f) cusparseChyb2csr((a),(b),(c),(cuComplex*)(d),(e),(f)) 120 #define cusparse_csr_spgemm(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t) cusparseCcsrgemm(a,b,c,d,e,f,g,h,(cuComplex*)i,j,k,l,m,(cuComplex*)n,o,p,q,(cuComplex*)r,s,t) 121 #define cusparse_csr_spgeam(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s) cusparseCcsrgeam(a,b,c,(cuComplex*)d,e,f,(cuComplex*)g,h,i,(cuComplex*)j,k,l,(cuComplex*)m,n,o,p,(cuComplex*)q,r,s) 122 #elif defined(PETSC_USE_REAL_DOUBLE) 123 #define cusparse_csr_spmv(a,b,c,d,e,f,g,h,i,j,k,l,m) cusparseZcsrmv((a),(b),(c),(d),(e),(cuDoubleComplex*)(f),(g),(cuDoubleComplex*)(h),(i),(j),(cuDoubleComplex*)(k),(cuDoubleComplex*)(l),(cuDoubleComplex*)(m)) 124 #define cusparse_csr_spmm(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p) cusparseZcsrmm((a),(b),(c),(d),(e),(f),(cuDoubleComplex*)(g),(h),(cuDoubleComplex*)(i),(j),(k),(cuDoubleComplex*)(l),(m),(cuDoubleComplex*)(n),(cuDoubleComplex*)(o),(p)) 125 #define cusparse_csr2csc(a,b,c,d,e,f,g,h,i,j,k,l) cusparseZcsr2csc((a),(b),(c),(d),(cuDoubleComplex*)(e),(f),(g),(cuDoubleComplex*)(h),(i),(j),(k),(l)) 126 #define cusparse_hyb_spmv(a,b,c,d,e,f,g,h) cusparseZhybmv((a),(b),(cuDoubleComplex*)(c),(d),(e),(cuDoubleComplex*)(f),(cuDoubleComplex*)(g),(cuDoubleComplex*)(h)) 127 #define cusparse_csr2hyb(a,b,c,d,e,f,g,h,i,j) cusparseZcsr2hyb((a),(b),(c),(d),(cuDoubleComplex*)(e),(f),(g),(h),(i),(j)) 128 #define cusparse_hyb2csr(a,b,c,d,e,f) cusparseZhyb2csr((a),(b),(c),(cuDoubleComplex*)(d),(e),(f)) 129 #define cusparse_csr_spgemm(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t) cusparseZcsrgemm(a,b,c,d,e,f,g,h,(cuDoubleComplex*)i,j,k,l,m,(cuDoubleComplex*)n,o,p,q,(cuDoubleComplex*)r,s,t) 130 #define cusparse_csr_spgeam(a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s) cusparseZcsrgeam(a,b,c,(cuDoubleComplex*)d,e,f,(cuDoubleComplex*)g,h,i,(cuDoubleComplex*)j,k,l,(cuDoubleComplex*)m,n,o,p,(cuDoubleComplex*)q,r,s) 131 #endif 132 #else 133 #if defined(PETSC_USE_REAL_SINGLE) 134 #define cusparse_csr_spmv cusparseScsrmv 135 #define cusparse_csr_spmm cusparseScsrmm 136 #define cusparse_csr2csc cusparseScsr2csc 137 #define cusparse_hyb_spmv cusparseShybmv 138 #define cusparse_csr2hyb cusparseScsr2hyb 139 #define cusparse_hyb2csr cusparseShyb2csr 140 #define cusparse_csr_spgemm cusparseScsrgemm 141 #define cusparse_csr_spgeam cusparseScsrgeam 142 #elif defined(PETSC_USE_REAL_DOUBLE) 143 #define cusparse_csr_spmv cusparseDcsrmv 144 #define cusparse_csr_spmm cusparseDcsrmm 145 #define cusparse_csr2csc cusparseDcsr2csc 146 #define cusparse_hyb_spmv cusparseDhybmv 147 #define cusparse_csr2hyb cusparseDcsr2hyb 148 #define cusparse_hyb2csr cusparseDhyb2csr 149 #define cusparse_csr_spgemm cusparseDcsrgemm 150 #define cusparse_csr_spgeam cusparseDcsrgeam 151 #endif 152 #endif 153 #endif 154 155 #define THRUSTINTARRAY32 thrust::device_vector<int> 156 #define THRUSTINTARRAY thrust::device_vector<PetscInt> 157 #define THRUSTARRAY thrust::device_vector<PetscScalar> 158 159 /* A CSR matrix structure */ 160 struct CsrMatrix { 161 PetscInt num_rows; 162 PetscInt num_cols; 163 PetscInt num_entries; 164 THRUSTINTARRAY32 *row_offsets; 165 THRUSTINTARRAY32 *column_indices; 166 THRUSTARRAY *values; 167 }; 168 169 /* This is struct holding the relevant data needed to a MatSolve */ 170 struct Mat_SeqAIJCUSPARSETriFactorStruct { 171 /* Data needed for triangular solve */ 172 cusparseMatDescr_t descr; 173 cusparseOperation_t solveOp; 174 CsrMatrix *csrMat; 175 #if PETSC_PKG_CUDA_VERSION_GE(9,0,0) 176 csrsv2Info_t solveInfo; 177 #else 178 cusparseSolveAnalysisInfo_t solveInfo; 179 #endif 180 cusparseSolvePolicy_t solvePolicy; /* whether level information is generated and used */ 181 int solveBufferSize; 182 void *solveBuffer; 183 size_t csr2cscBufferSize; /* to transpose the triangular factor (only used for CUDA >= 11.0) */ 184 void *csr2cscBuffer; 185 PetscScalar *AA_h; /* managed host buffer for moving values to the GPU */ 186 }; 187 188 /* This is a larger struct holding all the triangular factors for a solve, transpose solve, and any indices used in a reordering */ 189 struct Mat_SeqAIJCUSPARSETriFactors { 190 Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorPtr; /* pointer for lower triangular (factored matrix) on GPU */ 191 Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorPtr; /* pointer for upper triangular (factored matrix) on GPU */ 192 Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorPtrTranspose; /* pointer for lower triangular (factored matrix) on GPU for the transpose (useful for BiCG) */ 193 Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorPtrTranspose; /* pointer for upper triangular (factored matrix) on GPU for the transpose (useful for BiCG)*/ 194 THRUSTINTARRAY *rpermIndices; /* indices used for any reordering */ 195 THRUSTINTARRAY *cpermIndices; /* indices used for any reordering */ 196 THRUSTARRAY *workVector; 197 cusparseHandle_t handle; /* a handle to the cusparse library */ 198 PetscInt nnz; /* number of nonzeros ... need this for accurate logging between ICC and ILU */ 199 PetscScalar *a_band_d; /* GPU data for banded CSR LU factorization matrix diag(L)=1 */ 200 int *i_band_d; /* this could be optimized away */ 201 cudaDeviceProp dev_prop; 202 PetscBool init_dev_prop; 203 }; 204 205 struct Mat_CusparseSpMV { 206 PetscBool initialized; /* Don't rely on spmvBuffer != NULL to test if the struct is initialized, */ 207 size_t spmvBufferSize; /* since I'm not sure if smvBuffer can be NULL even after cusparseSpMV_bufferSize() */ 208 void *spmvBuffer; 209 #if PETSC_PKG_CUDA_VERSION_GE(11,0,0) /* these are present from CUDA 10.1, but PETSc code makes use of them from CUDA 11 on */ 210 cusparseDnVecDescr_t vecXDescr,vecYDescr; /* descriptor for the dense vectors in y=op(A)x */ 211 #endif 212 }; 213 214 /* This is struct holding the relevant data needed to a MatMult */ 215 struct Mat_SeqAIJCUSPARSEMultStruct { 216 void *mat; /* opaque pointer to a matrix. This could be either a cusparseHybMat_t or a CsrMatrix */ 217 cusparseMatDescr_t descr; /* Data needed to describe the matrix for a multiply */ 218 THRUSTINTARRAY *cprowIndices; /* compressed row indices used in the parallel SpMV */ 219 PetscScalar *alpha_one; /* pointer to a device "scalar" storing the alpha parameter in the SpMV */ 220 PetscScalar *beta_zero; /* pointer to a device "scalar" storing the beta parameter in the SpMV as zero*/ 221 PetscScalar *beta_one; /* pointer to a device "scalar" storing the beta parameter in the SpMV as one */ 222 #if PETSC_PKG_CUDA_VERSION_GE(11,0,0) 223 cusparseSpMatDescr_t matDescr; /* descriptor for the matrix, used by SpMV and SpMM */ 224 Mat_CusparseSpMV cuSpMV[3]; /* different Mat_CusparseSpMV structs for non-transpose, transpose, conj-transpose */ 225 Mat_SeqAIJCUSPARSEMultStruct() : matDescr(NULL) { 226 for (int i=0; i<3; i++) cuSpMV[i].initialized = PETSC_FALSE; 227 } 228 #endif 229 }; 230 231 /* This is a larger struct holding all the matrices for a SpMV, and SpMV Transpose */ 232 struct Mat_SeqAIJCUSPARSE { 233 Mat_SeqAIJCUSPARSEMultStruct *mat; /* pointer to the matrix on the GPU */ 234 Mat_SeqAIJCUSPARSEMultStruct *matTranspose; /* pointer to the matrix on the GPU (for the transpose ... useful for BiCG) */ 235 THRUSTARRAY *workVector; /* pointer to a workvector to which we can copy the relevant indices of a vector we want to multiply */ 236 THRUSTINTARRAY32 *rowoffsets_gpu; /* rowoffsets on GPU in non-compressed-row format. It is used to convert CSR to CSC */ 237 PetscInt nrows; /* number of rows of the matrix seen by GPU */ 238 MatCUSPARSEStorageFormat format; /* the storage format for the matrix on the device */ 239 PetscBool use_cpu_solve; /* Use AIJ_Seq (I)LU solve */ 240 cudaStream_t stream; /* a stream for the parallel SpMV ... this is not owned and should not be deleted */ 241 cusparseHandle_t handle; /* a handle to the cusparse library ... this may not be owned (if we're working in parallel i.e. multiGPUs) */ 242 PetscObjectState nonzerostate; /* track nonzero state to possibly recreate the GPU matrix */ 243 #if PETSC_PKG_CUDA_VERSION_GE(11,0,0) 244 size_t csr2cscBufferSize; /* stuff used to compute the matTranspose above */ 245 void *csr2cscBuffer; /* This is used as a C struct and is calloc'ed by PetscNewLog() */ 246 cusparseCsr2CscAlg_t csr2cscAlg; /* algorithms can be selected from command line options */ 247 cusparseSpMVAlg_t spmvAlg; 248 cusparseSpMMAlg_t spmmAlg; 249 #endif 250 THRUSTINTARRAY *csr2csc_i; 251 PetscSplitCSRDataStructure deviceMat; /* Matrix on device for, eg, assembly */ 252 THRUSTINTARRAY *cooPerm; /* permutation array that sorts the input coo entris by row and col */ 253 THRUSTINTARRAY *cooPerm_a; /* ordered array that indicate i-th nonzero (after sorting) is the j-th unique nonzero */ 254 }; 255 256 PETSC_INTERN PetscErrorCode MatSeqAIJCUSPARSECopyToGPU(Mat); 257 PETSC_INTERN PetscErrorCode MatCUSPARSESetStream(Mat, const cudaStream_t stream); 258 PETSC_INTERN PetscErrorCode MatCUSPARSESetHandle(Mat, const cusparseHandle_t handle); 259 PETSC_INTERN PetscErrorCode MatCUSPARSEClearHandle(Mat); 260 PETSC_INTERN PetscErrorCode MatSetPreallocationCOO_SeqAIJCUSPARSE(Mat,PetscCount,const PetscInt[],const PetscInt[]); 261 PETSC_INTERN PetscErrorCode MatSetValuesCOO_SeqAIJCUSPARSE(Mat,const PetscScalar[],InsertMode); 262 PETSC_INTERN PetscErrorCode MatSeqAIJCUSPARSEMergeMats(Mat,Mat,MatReuse,Mat*); 263 PETSC_INTERN PetscErrorCode MatSeqAIJCUSPARSETriFactors_Reset(Mat_SeqAIJCUSPARSETriFactors_p*); 264 265 static inline bool isCudaMem(const void *data) 266 { 267 cudaError_t cerr; 268 struct cudaPointerAttributes attr; 269 enum cudaMemoryType mtype; 270 cerr = cudaPointerGetAttributes(&attr,data); /* Do not check error since before CUDA 11.0, passing a host pointer returns cudaErrorInvalidValue */ 271 cudaGetLastError(); /* Reset the last error */ 272 #if (CUDART_VERSION < 10000) 273 mtype = attr.memoryType; 274 #else 275 mtype = attr.type; 276 #endif 277 if (cerr == cudaSuccess && mtype == cudaMemoryTypeDevice) return true; 278 else return false; 279 } 280 281 #endif 282