Actual source code: aijcusparse.cu

  1: /*
  2:   Defines the basic matrix operations for the AIJ (compressed row)
  3:   matrix storage format using the CUSPARSE library,
  4: */
  5: #define PETSC_SKIP_IMMINTRIN_H_CUDAWORKAROUND 1

  7: #include <petscconf.h>
  8: #include <../src/mat/impls/aij/seq/aij.h>
  9: #include <../src/mat/impls/sbaij/seq/sbaij.h>
 10: #include <../src/vec/vec/impls/dvecimpl.h>
 11: #include <petsc/private/vecimpl.h>
 12: #undef VecType
 13: #include <../src/mat/impls/aij/seq/seqcusparse/cusparsematimpl.h>
 14: #include <thrust/adjacent_difference.h>
 15: #if PETSC_CPP_VERSION >= 14
 16:   #define PETSC_HAVE_THRUST_ASYNC 1
 17:   // thrust::for_each(thrust::cuda::par.on()) requires C++14
 18:   #include <thrust/async/for_each.h>
 19: #endif
 20: #include <thrust/iterator/constant_iterator.h>
 21: #include <thrust/remove.h>
 22: #include <thrust/sort.h>
 23: #include <thrust/unique.h>

 25: const char *const MatCUSPARSEStorageFormats[] = {"CSR", "ELL", "HYB", "MatCUSPARSEStorageFormat", "MAT_CUSPARSE_", 0};
 26: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
 27: /* The following are copied from cusparse.h in CUDA-11.0. In MatCUSPARSESpMVAlgorithms[] etc, we copy them in
 28:     0-based integer value order, since we want to use PetscOptionsEnum() to parse user command line options for them.

 30:   typedef enum {
 31:       CUSPARSE_MV_ALG_DEFAULT = 0,
 32:       CUSPARSE_COOMV_ALG      = 1,
 33:       CUSPARSE_CSRMV_ALG1     = 2,
 34:       CUSPARSE_CSRMV_ALG2     = 3
 35:   } cusparseSpMVAlg_t;

 37:   typedef enum {
 38:       CUSPARSE_MM_ALG_DEFAULT     CUSPARSE_DEPRECATED_ENUM(CUSPARSE_SPMM_ALG_DEFAULT) = 0,
 39:       CUSPARSE_COOMM_ALG1         CUSPARSE_DEPRECATED_ENUM(CUSPARSE_SPMM_COO_ALG1)    = 1,
 40:       CUSPARSE_COOMM_ALG2         CUSPARSE_DEPRECATED_ENUM(CUSPARSE_SPMM_COO_ALG2)    = 2,
 41:       CUSPARSE_COOMM_ALG3         CUSPARSE_DEPRECATED_ENUM(CUSPARSE_SPMM_COO_ALG3)    = 3,
 42:       CUSPARSE_CSRMM_ALG1         CUSPARSE_DEPRECATED_ENUM(CUSPARSE_SPMM_CSR_ALG1)    = 4,
 43:       CUSPARSE_SPMM_ALG_DEFAULT = 0,
 44:       CUSPARSE_SPMM_COO_ALG1    = 1,
 45:       CUSPARSE_SPMM_COO_ALG2    = 2,
 46:       CUSPARSE_SPMM_COO_ALG3    = 3,
 47:       CUSPARSE_SPMM_COO_ALG4    = 5,
 48:       CUSPARSE_SPMM_CSR_ALG1    = 4,
 49:       CUSPARSE_SPMM_CSR_ALG2    = 6,
 50:   } cusparseSpMMAlg_t;

 52:   typedef enum {
 53:       CUSPARSE_CSR2CSC_ALG1 = 1, // faster than V2 (in general), deterministic
 54:       CUSPARSE_CSR2CSC_ALG2 = 2  // low memory requirement, non-deterministic
 55:   } cusparseCsr2CscAlg_t;
 56:   */
 57: const char *const MatCUSPARSESpMVAlgorithms[]    = {"MV_ALG_DEFAULT", "COOMV_ALG", "CSRMV_ALG1", "CSRMV_ALG2", "cusparseSpMVAlg_t", "CUSPARSE_", 0};
 58: const char *const MatCUSPARSESpMMAlgorithms[]    = {"ALG_DEFAULT", "COO_ALG1", "COO_ALG2", "COO_ALG3", "CSR_ALG1", "COO_ALG4", "CSR_ALG2", "cusparseSpMMAlg_t", "CUSPARSE_SPMM_", 0};
 59: const char *const MatCUSPARSECsr2CscAlgorithms[] = {"INVALID" /*cusparse does not have enum 0! We created one*/, "ALG1", "ALG2", "cusparseCsr2CscAlg_t", "CUSPARSE_CSR2CSC_", 0};
 60: #endif

 62: static PetscErrorCode MatICCFactorSymbolic_SeqAIJCUSPARSE(Mat, Mat, IS, const MatFactorInfo *);
 63: static PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJCUSPARSE(Mat, Mat, IS, const MatFactorInfo *);
 64: static PetscErrorCode MatCholeskyFactorNumeric_SeqAIJCUSPARSE(Mat, Mat, const MatFactorInfo *);
 65: static PetscErrorCode MatILUFactorSymbolic_SeqAIJCUSPARSE(Mat, Mat, IS, IS, const MatFactorInfo *);
 66: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
 67: static PetscErrorCode MatSolve_SeqAIJCUSPARSE(Mat, Vec, Vec);
 68: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE(Mat, Vec, Vec);
 69: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_NaturalOrdering(Mat, Vec, Vec);
 70: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering(Mat, Vec, Vec);
 71: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSETriFactorStruct **);
 72: #endif
 73: static PetscErrorCode MatSetFromOptions_SeqAIJCUSPARSE(Mat, PetscOptionItems *PetscOptionsObject);
 74: static PetscErrorCode MatAXPY_SeqAIJCUSPARSE(Mat, PetscScalar, Mat, MatStructure);
 75: static PetscErrorCode MatScale_SeqAIJCUSPARSE(Mat, PetscScalar);
 76: static PetscErrorCode MatMult_SeqAIJCUSPARSE(Mat, Vec, Vec);
 77: static PetscErrorCode MatMultAdd_SeqAIJCUSPARSE(Mat, Vec, Vec, Vec);
 78: static PetscErrorCode MatMultTranspose_SeqAIJCUSPARSE(Mat, Vec, Vec);
 79: static PetscErrorCode MatMultTransposeAdd_SeqAIJCUSPARSE(Mat, Vec, Vec, Vec);
 80: static PetscErrorCode MatMultHermitianTranspose_SeqAIJCUSPARSE(Mat, Vec, Vec);
 81: static PetscErrorCode MatMultHermitianTransposeAdd_SeqAIJCUSPARSE(Mat, Vec, Vec, Vec);
 82: static PetscErrorCode MatMultAddKernel_SeqAIJCUSPARSE(Mat, Vec, Vec, Vec, PetscBool, PetscBool);

 84: static PetscErrorCode CsrMatrix_Destroy(CsrMatrix **);
 85: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSEMultStruct **, MatCUSPARSEStorageFormat);
 86: static PetscErrorCode MatSeqAIJCUSPARSETriFactors_Destroy(Mat_SeqAIJCUSPARSETriFactors **);
 87: static PetscErrorCode MatSeqAIJCUSPARSE_Destroy(Mat);

 89: static PetscErrorCode MatSeqAIJCUSPARSECopyFromGPU(Mat);
 90: static PetscErrorCode MatSeqAIJCUSPARSEInvalidateTranspose(Mat, PetscBool);

 92: static PetscErrorCode MatSeqAIJCopySubArray_SeqAIJCUSPARSE(Mat, PetscInt, const PetscInt[], PetscScalar[]);
 93: static PetscErrorCode MatSetPreallocationCOO_SeqAIJCUSPARSE(Mat, PetscCount, PetscInt[], PetscInt[]);
 94: static PetscErrorCode MatSetValuesCOO_SeqAIJCUSPARSE(Mat, const PetscScalar[], InsertMode);

 96: PETSC_INTERN PetscErrorCode MatCUSPARSESetFormat_SeqAIJCUSPARSE(Mat A, MatCUSPARSEFormatOperation op, MatCUSPARSEStorageFormat format)
 97: {
 98:   Mat_SeqAIJCUSPARSE *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;

100:   PetscFunctionBegin;
101:   switch (op) {
102:   case MAT_CUSPARSE_MULT:
103:     cusparsestruct->format = format;
104:     break;
105:   case MAT_CUSPARSE_ALL:
106:     cusparsestruct->format = format;
107:     break;
108:   default:
109:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unsupported operation %d for MatCUSPARSEFormatOperation. MAT_CUSPARSE_MULT and MAT_CUSPARSE_ALL are currently supported.", op);
110:   }
111:   PetscFunctionReturn(PETSC_SUCCESS);
112: }

114: /*@
115:   MatCUSPARSESetFormat - Sets the storage format of `MATSEQCUSPARSE` matrices for a particular
116:   operation. Only the `MatMult()` operation can use different GPU storage formats

118:   Not Collective

120:   Input Parameters:
121: + A      - Matrix of type `MATSEQAIJCUSPARSE`
122: . op     - `MatCUSPARSEFormatOperation`. `MATSEQAIJCUSPARSE` matrices support `MAT_CUSPARSE_MULT` and `MAT_CUSPARSE_ALL`.
123:         `MATMPIAIJCUSPARSE` matrices support `MAT_CUSPARSE_MULT_DIAG`,`MAT_CUSPARSE_MULT_OFFDIAG`, and `MAT_CUSPARSE_ALL`.
124: - format - `MatCUSPARSEStorageFormat` (one of `MAT_CUSPARSE_CSR`, `MAT_CUSPARSE_ELL`, `MAT_CUSPARSE_HYB`.)

126:   Level: intermediate

128: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJCUSPARSE`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
129: @*/
130: PetscErrorCode MatCUSPARSESetFormat(Mat A, MatCUSPARSEFormatOperation op, MatCUSPARSEStorageFormat format)
131: {
132:   PetscFunctionBegin;
134:   PetscTryMethod(A, "MatCUSPARSESetFormat_C", (Mat, MatCUSPARSEFormatOperation, MatCUSPARSEStorageFormat), (A, op, format));
135:   PetscFunctionReturn(PETSC_SUCCESS);
136: }

138: PETSC_INTERN PetscErrorCode MatCUSPARSESetUseCPUSolve_SeqAIJCUSPARSE(Mat A, PetscBool use_cpu)
139: {
140:   Mat_SeqAIJCUSPARSE *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;

142:   PetscFunctionBegin;
143:   cusparsestruct->use_cpu_solve = use_cpu;
144:   PetscFunctionReturn(PETSC_SUCCESS);
145: }

147: /*@
148:   MatCUSPARSESetUseCPUSolve - Sets to use CPU `MatSolve()`.

150:   Input Parameters:
151: + A       - Matrix of type `MATSEQAIJCUSPARSE`
152: - use_cpu - set flag for using the built-in CPU `MatSolve()`

154:   Level: intermediate

156:   Note:
157:   The cuSparse LU solver currently computes the factors with the built-in CPU method
158:   and moves the factors to the GPU for the solve. We have observed better performance keeping the data on the CPU and computing the solve there.
159:   This method to specify if the solve is done on the CPU or GPU (GPU is the default).

161: .seealso: [](ch_matrices), `Mat`, `MatSolve()`, `MATSEQAIJCUSPARSE`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
162: @*/
163: PetscErrorCode MatCUSPARSESetUseCPUSolve(Mat A, PetscBool use_cpu)
164: {
165:   PetscFunctionBegin;
167:   PetscTryMethod(A, "MatCUSPARSESetUseCPUSolve_C", (Mat, PetscBool), (A, use_cpu));
168:   PetscFunctionReturn(PETSC_SUCCESS);
169: }

171: static PetscErrorCode MatSetOption_SeqAIJCUSPARSE(Mat A, MatOption op, PetscBool flg)
172: {
173:   PetscFunctionBegin;
174:   switch (op) {
175:   case MAT_FORM_EXPLICIT_TRANSPOSE:
176:     /* need to destroy the transpose matrix if present to prevent from logic errors if flg is set to true later */
177:     if (A->form_explicit_transpose && !flg) PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_TRUE));
178:     A->form_explicit_transpose = flg;
179:     break;
180:   default:
181:     PetscCall(MatSetOption_SeqAIJ(A, op, flg));
182:     break;
183:   }
184:   PetscFunctionReturn(PETSC_SUCCESS);
185: }

187: static PetscErrorCode MatSetFromOptions_SeqAIJCUSPARSE(Mat A, PetscOptionItems *PetscOptionsObject)
188: {
189:   MatCUSPARSEStorageFormat format;
190:   PetscBool                flg;
191:   Mat_SeqAIJCUSPARSE      *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;

193:   PetscFunctionBegin;
194:   PetscOptionsHeadBegin(PetscOptionsObject, "SeqAIJCUSPARSE options");
195:   if (A->factortype == MAT_FACTOR_NONE) {
196:     PetscCall(PetscOptionsEnum("-mat_cusparse_mult_storage_format", "sets storage format of (seq)aijcusparse gpu matrices for SpMV", "MatCUSPARSESetFormat", MatCUSPARSEStorageFormats, (PetscEnum)cusparsestruct->format, (PetscEnum *)&format, &flg));
197:     if (flg) PetscCall(MatCUSPARSESetFormat(A, MAT_CUSPARSE_MULT, format));

199:     PetscCall(PetscOptionsEnum("-mat_cusparse_storage_format", "sets storage format of (seq)aijcusparse gpu matrices for SpMV and TriSolve", "MatCUSPARSESetFormat", MatCUSPARSEStorageFormats, (PetscEnum)cusparsestruct->format, (PetscEnum *)&format, &flg));
200:     if (flg) PetscCall(MatCUSPARSESetFormat(A, MAT_CUSPARSE_ALL, format));
201:     PetscCall(PetscOptionsBool("-mat_cusparse_use_cpu_solve", "Use CPU (I)LU solve", "MatCUSPARSESetUseCPUSolve", cusparsestruct->use_cpu_solve, &cusparsestruct->use_cpu_solve, &flg));
202:     if (flg) PetscCall(MatCUSPARSESetUseCPUSolve(A, cusparsestruct->use_cpu_solve));
203: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
204:     PetscCall(PetscOptionsEnum("-mat_cusparse_spmv_alg", "sets cuSPARSE algorithm used in sparse-mat dense-vector multiplication (SpMV)", "cusparseSpMVAlg_t", MatCUSPARSESpMVAlgorithms, (PetscEnum)cusparsestruct->spmvAlg, (PetscEnum *)&cusparsestruct->spmvAlg, &flg));
205:     /* If user did use this option, check its consistency with cuSPARSE, since PetscOptionsEnum() sets enum values based on their position in MatCUSPARSESpMVAlgorithms[] */
206:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
207:     PetscCheck(!flg || CUSPARSE_SPMV_CSR_ALG1 == 2, PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSPARSE enum cusparseSpMVAlg_t has been changed but PETSc has not been updated accordingly");
208:   #else
209:     PetscCheck(!flg || CUSPARSE_CSRMV_ALG1 == 2, PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSPARSE enum cusparseSpMVAlg_t has been changed but PETSc has not been updated accordingly");
210:   #endif
211:     PetscCall(PetscOptionsEnum("-mat_cusparse_spmm_alg", "sets cuSPARSE algorithm used in sparse-mat dense-mat multiplication (SpMM)", "cusparseSpMMAlg_t", MatCUSPARSESpMMAlgorithms, (PetscEnum)cusparsestruct->spmmAlg, (PetscEnum *)&cusparsestruct->spmmAlg, &flg));
212:     PetscCheck(!flg || CUSPARSE_SPMM_CSR_ALG1 == 4, PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSPARSE enum cusparseSpMMAlg_t has been changed but PETSc has not been updated accordingly");

214:     PetscCall(
215:       PetscOptionsEnum("-mat_cusparse_csr2csc_alg", "sets cuSPARSE algorithm used in converting CSR matrices to CSC matrices", "cusparseCsr2CscAlg_t", MatCUSPARSECsr2CscAlgorithms, (PetscEnum)cusparsestruct->csr2cscAlg, (PetscEnum *)&cusparsestruct->csr2cscAlg, &flg));
216:     PetscCheck(!flg || CUSPARSE_CSR2CSC_ALG1 == 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSPARSE enum cusparseCsr2CscAlg_t has been changed but PETSc has not been updated accordingly");
217: #endif
218:   }
219:   PetscOptionsHeadEnd();
220:   PetscFunctionReturn(PETSC_SUCCESS);
221: }

223: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
224: static PetscErrorCode MatSeqAIJCUSPARSEBuildFactoredMatrix_LU(Mat A)
225: {
226:   Mat_SeqAIJ                   *a  = static_cast<Mat_SeqAIJ *>(A->data);
227:   PetscInt                      m  = A->rmap->n;
228:   Mat_SeqAIJCUSPARSETriFactors *fs = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
229:   const PetscInt               *Ai = a->i, *Aj = a->j, *Adiag = a->diag;
230:   const MatScalar              *Aa = a->a;
231:   PetscInt                     *Mi, *Mj, Mnz;
232:   PetscScalar                  *Ma;

234:   PetscFunctionBegin;
235:   if (A->offloadmask == PETSC_OFFLOAD_CPU) { // A's latest factors are on CPU
236:     if (!fs->csrRowPtr) {                    // Is't the first time to do the setup? Use csrRowPtr since it is not null even when m=0
237:       // Re-arrange the (skewed) factored matrix and put the result into M, a regular csr matrix on host
238:       Mnz = (Ai[m] - Ai[0]) + (Adiag[0] - Adiag[m]); // Lnz (without the unit diagonal) + Unz (with the non-unit diagonal)
239:       PetscCall(PetscMalloc1(m + 1, &Mi));
240:       PetscCall(PetscMalloc1(Mnz, &Mj)); // Mj is temp
241:       PetscCall(PetscMalloc1(Mnz, &Ma));
242:       Mi[0] = 0;
243:       for (PetscInt i = 0; i < m; i++) {
244:         PetscInt llen = Ai[i + 1] - Ai[i];
245:         PetscInt ulen = Adiag[i] - Adiag[i + 1];
246:         PetscCall(PetscArraycpy(Mj + Mi[i], Aj + Ai[i], llen));                           // entries of L
247:         Mj[Mi[i] + llen] = i;                                                             // diagonal entry
248:         PetscCall(PetscArraycpy(Mj + Mi[i] + llen + 1, Aj + Adiag[i + 1] + 1, ulen - 1)); // entries of U on the right of the diagonal
249:         Mi[i + 1] = Mi[i] + llen + ulen;
250:       }
251:       // Copy M (L,U) from host to device
252:       PetscCallCUDA(cudaMalloc(&fs->csrRowPtr, sizeof(*fs->csrRowPtr) * (m + 1)));
253:       PetscCallCUDA(cudaMalloc(&fs->csrColIdx, sizeof(*fs->csrColIdx) * Mnz));
254:       PetscCallCUDA(cudaMalloc(&fs->csrVal, sizeof(*fs->csrVal) * Mnz));
255:       PetscCallCUDA(cudaMemcpy(fs->csrRowPtr, Mi, sizeof(*fs->csrRowPtr) * (m + 1), cudaMemcpyHostToDevice));
256:       PetscCallCUDA(cudaMemcpy(fs->csrColIdx, Mj, sizeof(*fs->csrColIdx) * Mnz, cudaMemcpyHostToDevice));

258:       // Create descriptors for L, U. See https://docs.nvidia.com/cuda/cusparse/index.html#cusparseDiagType_t
259:       // cusparseDiagType_t: This type indicates if the matrix diagonal entries are unity. The diagonal elements are always
260:       // assumed to be present, but if CUSPARSE_DIAG_TYPE_UNIT is passed to an API routine, then the routine assumes that
261:       // all diagonal entries are unity and will not read or modify those entries. Note that in this case the routine
262:       // assumes the diagonal entries are equal to one, regardless of what those entries are actually set to in memory.
263:       cusparseFillMode_t        fillMode  = CUSPARSE_FILL_MODE_LOWER;
264:       cusparseDiagType_t        diagType  = CUSPARSE_DIAG_TYPE_UNIT;
265:       const cusparseIndexType_t indexType = PetscDefined(USE_64BIT_INDICES) ? CUSPARSE_INDEX_64I : CUSPARSE_INDEX_32I;

267:       PetscCallCUSPARSE(cusparseCreateCsr(&fs->spMatDescr_L, m, m, Mnz, fs->csrRowPtr, fs->csrColIdx, fs->csrVal, indexType, indexType, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype));
268:       PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
269:       PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

271:       fillMode = CUSPARSE_FILL_MODE_UPPER;
272:       diagType = CUSPARSE_DIAG_TYPE_NON_UNIT;
273:       PetscCallCUSPARSE(cusparseCreateCsr(&fs->spMatDescr_U, m, m, Mnz, fs->csrRowPtr, fs->csrColIdx, fs->csrVal, indexType, indexType, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype));
274:       PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
275:       PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

277:       // Allocate work vectors in SpSv
278:       PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(*fs->X) * m));
279:       PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(*fs->Y) * m));

281:       PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_X, m, fs->X, cusparse_scalartype));
282:       PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_Y, m, fs->Y, cusparse_scalartype));

284:       // Query buffer sizes for SpSV and then allocate buffers, temporarily assuming opA = CUSPARSE_OPERATION_NON_TRANSPOSE
285:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_L));
286:       PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L, &fs->spsvBufferSize_L));
287:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_U));
288:       PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_U, &fs->spsvBufferSize_U));
289:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U));
290:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L));

292:       // Record for reuse
293:       fs->csrRowPtr_h = Mi;
294:       fs->csrVal_h    = Ma;
295:       PetscCall(PetscFree(Mj));
296:     }
297:     // Copy the value
298:     Mi  = fs->csrRowPtr_h;
299:     Ma  = fs->csrVal_h;
300:     Mnz = Mi[m];
301:     for (PetscInt i = 0; i < m; i++) {
302:       PetscInt llen = Ai[i + 1] - Ai[i];
303:       PetscInt ulen = Adiag[i] - Adiag[i + 1];
304:       PetscCall(PetscArraycpy(Ma + Mi[i], Aa + Ai[i], llen));                           // entries of L
305:       Ma[Mi[i] + llen] = (MatScalar)1.0 / Aa[Adiag[i]];                                 // recover the diagonal entry
306:       PetscCall(PetscArraycpy(Ma + Mi[i] + llen + 1, Aa + Adiag[i + 1] + 1, ulen - 1)); // entries of U on the right of the diagonal
307:     }
308:     PetscCallCUDA(cudaMemcpy(fs->csrVal, Ma, sizeof(*Ma) * Mnz, cudaMemcpyHostToDevice));

310:     // Do cusparseSpSV_analysis(), which is numeric and requires valid and up-to-date matrix values
311:     PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L, fs->spsvBuffer_L));

313:     PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_U, fs->spsvBuffer_U));

315:     // L, U values have changed, reset the flag to indicate we need to redo cusparseSpSV_analysis() for transpose solve
316:     fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;
317:   }
318:   PetscFunctionReturn(PETSC_SUCCESS);
319: }
320: #else
321: static PetscErrorCode MatSeqAIJCUSPARSEBuildILULowerTriMatrix(Mat A)
322: {
323:   Mat_SeqAIJ                        *a                  = (Mat_SeqAIJ *)A->data;
324:   PetscInt                           n                  = A->rmap->n;
325:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
326:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
327:   const PetscInt                    *ai = a->i, *aj = a->j, *vi;
328:   const MatScalar                   *aa = a->a, *v;
329:   PetscInt                          *AiLo, *AjLo;
330:   PetscInt                           i, nz, nzLower, offset, rowOffset;

332:   PetscFunctionBegin;
333:   if (!n) PetscFunctionReturn(PETSC_SUCCESS);
334:   if (A->offloadmask == PETSC_OFFLOAD_UNALLOCATED || A->offloadmask == PETSC_OFFLOAD_CPU) {
335:     try {
336:       /* first figure out the number of nonzeros in the lower triangular matrix including 1's on the diagonal. */
337:       nzLower = n + ai[n] - ai[1];
338:       if (!loTriFactor) {
339:         PetscScalar *AALo;

341:         PetscCallCUDA(cudaMallocHost((void **)&AALo, nzLower * sizeof(PetscScalar)));

343:         /* Allocate Space for the lower triangular matrix */
344:         PetscCallCUDA(cudaMallocHost((void **)&AiLo, (n + 1) * sizeof(PetscInt)));
345:         PetscCallCUDA(cudaMallocHost((void **)&AjLo, nzLower * sizeof(PetscInt)));

347:         /* Fill the lower triangular matrix */
348:         AiLo[0]   = (PetscInt)0;
349:         AiLo[n]   = nzLower;
350:         AjLo[0]   = (PetscInt)0;
351:         AALo[0]   = (MatScalar)1.0;
352:         v         = aa;
353:         vi        = aj;
354:         offset    = 1;
355:         rowOffset = 1;
356:         for (i = 1; i < n; i++) {
357:           nz = ai[i + 1] - ai[i];
358:           /* additional 1 for the term on the diagonal */
359:           AiLo[i] = rowOffset;
360:           rowOffset += nz + 1;

362:           PetscCall(PetscArraycpy(&AjLo[offset], vi, nz));
363:           PetscCall(PetscArraycpy(&AALo[offset], v, nz));

365:           offset += nz;
366:           AjLo[offset] = (PetscInt)i;
367:           AALo[offset] = (MatScalar)1.0;
368:           offset += 1;

370:           v += nz;
371:           vi += nz;
372:         }

374:         /* allocate space for the triangular factor information */
375:         PetscCall(PetscNew(&loTriFactor));
376:         loTriFactor->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;
377:         /* Create the matrix description */
378:         PetscCallCUSPARSE(cusparseCreateMatDescr(&loTriFactor->descr));
379:         PetscCallCUSPARSE(cusparseSetMatIndexBase(loTriFactor->descr, CUSPARSE_INDEX_BASE_ZERO));
380:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
381:         PetscCallCUSPARSE(cusparseSetMatType(loTriFactor->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
382:   #else
383:         PetscCallCUSPARSE(cusparseSetMatType(loTriFactor->descr, CUSPARSE_MATRIX_TYPE_TRIANGULAR));
384:   #endif
385:         PetscCallCUSPARSE(cusparseSetMatFillMode(loTriFactor->descr, CUSPARSE_FILL_MODE_LOWER));
386:         PetscCallCUSPARSE(cusparseSetMatDiagType(loTriFactor->descr, CUSPARSE_DIAG_TYPE_UNIT));

388:         /* set the operation */
389:         loTriFactor->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

391:         /* set the matrix */
392:         loTriFactor->csrMat              = new CsrMatrix;
393:         loTriFactor->csrMat->num_rows    = n;
394:         loTriFactor->csrMat->num_cols    = n;
395:         loTriFactor->csrMat->num_entries = nzLower;

397:         loTriFactor->csrMat->row_offsets = new THRUSTINTARRAY32(n + 1);
398:         loTriFactor->csrMat->row_offsets->assign(AiLo, AiLo + n + 1);

400:         loTriFactor->csrMat->column_indices = new THRUSTINTARRAY32(nzLower);
401:         loTriFactor->csrMat->column_indices->assign(AjLo, AjLo + nzLower);

403:         loTriFactor->csrMat->values = new THRUSTARRAY(nzLower);
404:         loTriFactor->csrMat->values->assign(AALo, AALo + nzLower);

406:         /* Create the solve analysis information */
407:         PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
408:         PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&loTriFactor->solveInfo));
409:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
410:         PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, loTriFactor->solveOp, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_entries, loTriFactor->descr, loTriFactor->csrMat->values->data().get(),
411:                                                   loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, &loTriFactor->solveBufferSize));
412:         PetscCallCUDA(cudaMalloc(&loTriFactor->solveBuffer, loTriFactor->solveBufferSize));
413:   #endif

415:         /* perform the solve analysis */
416:         PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, loTriFactor->solveOp, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_entries, loTriFactor->descr, loTriFactor->csrMat->values->data().get(),
417:                                                   loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, loTriFactor->solvePolicy, loTriFactor->solveBuffer));
418:         PetscCallCUDA(WaitForCUDA());
419:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

421:         /* assign the pointer */
422:         ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->loTriFactorPtr = loTriFactor;
423:         loTriFactor->AA_h                                          = AALo;
424:         PetscCallCUDA(cudaFreeHost(AiLo));
425:         PetscCallCUDA(cudaFreeHost(AjLo));
426:         PetscCall(PetscLogCpuToGpu((n + 1 + nzLower) * sizeof(int) + nzLower * sizeof(PetscScalar)));
427:       } else { /* update values only */
428:         if (!loTriFactor->AA_h) PetscCallCUDA(cudaMallocHost((void **)&loTriFactor->AA_h, nzLower * sizeof(PetscScalar)));
429:         /* Fill the lower triangular matrix */
430:         loTriFactor->AA_h[0] = 1.0;
431:         v                    = aa;
432:         vi                   = aj;
433:         offset               = 1;
434:         for (i = 1; i < n; i++) {
435:           nz = ai[i + 1] - ai[i];
436:           PetscCall(PetscArraycpy(&loTriFactor->AA_h[offset], v, nz));
437:           offset += nz;
438:           loTriFactor->AA_h[offset] = 1.0;
439:           offset += 1;
440:           v += nz;
441:         }
442:         loTriFactor->csrMat->values->assign(loTriFactor->AA_h, loTriFactor->AA_h + nzLower);
443:         PetscCall(PetscLogCpuToGpu(nzLower * sizeof(PetscScalar)));
444:       }
445:     } catch (char *ex) {
446:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
447:     }
448:   }
449:   PetscFunctionReturn(PETSC_SUCCESS);
450: }

452: static PetscErrorCode MatSeqAIJCUSPARSEBuildILUUpperTriMatrix(Mat A)
453: {
454:   Mat_SeqAIJ                        *a                  = (Mat_SeqAIJ *)A->data;
455:   PetscInt                           n                  = A->rmap->n;
456:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
457:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
458:   const PetscInt                    *aj = a->j, *adiag = a->diag, *vi;
459:   const MatScalar                   *aa = a->a, *v;
460:   PetscInt                          *AiUp, *AjUp;
461:   PetscInt                           i, nz, nzUpper, offset;

463:   PetscFunctionBegin;
464:   if (!n) PetscFunctionReturn(PETSC_SUCCESS);
465:   if (A->offloadmask == PETSC_OFFLOAD_UNALLOCATED || A->offloadmask == PETSC_OFFLOAD_CPU) {
466:     try {
467:       /* next, figure out the number of nonzeros in the upper triangular matrix. */
468:       nzUpper = adiag[0] - adiag[n];
469:       if (!upTriFactor) {
470:         PetscScalar *AAUp;

472:         PetscCallCUDA(cudaMallocHost((void **)&AAUp, nzUpper * sizeof(PetscScalar)));

474:         /* Allocate Space for the upper triangular matrix */
475:         PetscCallCUDA(cudaMallocHost((void **)&AiUp, (n + 1) * sizeof(PetscInt)));
476:         PetscCallCUDA(cudaMallocHost((void **)&AjUp, nzUpper * sizeof(PetscInt)));

478:         /* Fill the upper triangular matrix */
479:         AiUp[0] = (PetscInt)0;
480:         AiUp[n] = nzUpper;
481:         offset  = nzUpper;
482:         for (i = n - 1; i >= 0; i--) {
483:           v  = aa + adiag[i + 1] + 1;
484:           vi = aj + adiag[i + 1] + 1;

486:           /* number of elements NOT on the diagonal */
487:           nz = adiag[i] - adiag[i + 1] - 1;

489:           /* decrement the offset */
490:           offset -= (nz + 1);

492:           /* first, set the diagonal elements */
493:           AjUp[offset] = (PetscInt)i;
494:           AAUp[offset] = (MatScalar)1. / v[nz];
495:           AiUp[i]      = AiUp[i + 1] - (nz + 1);

497:           PetscCall(PetscArraycpy(&AjUp[offset + 1], vi, nz));
498:           PetscCall(PetscArraycpy(&AAUp[offset + 1], v, nz));
499:         }

501:         /* allocate space for the triangular factor information */
502:         PetscCall(PetscNew(&upTriFactor));
503:         upTriFactor->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

505:         /* Create the matrix description */
506:         PetscCallCUSPARSE(cusparseCreateMatDescr(&upTriFactor->descr));
507:         PetscCallCUSPARSE(cusparseSetMatIndexBase(upTriFactor->descr, CUSPARSE_INDEX_BASE_ZERO));
508:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
509:         PetscCallCUSPARSE(cusparseSetMatType(upTriFactor->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
510:   #else
511:         PetscCallCUSPARSE(cusparseSetMatType(upTriFactor->descr, CUSPARSE_MATRIX_TYPE_TRIANGULAR));
512:   #endif
513:         PetscCallCUSPARSE(cusparseSetMatFillMode(upTriFactor->descr, CUSPARSE_FILL_MODE_UPPER));
514:         PetscCallCUSPARSE(cusparseSetMatDiagType(upTriFactor->descr, CUSPARSE_DIAG_TYPE_NON_UNIT));

516:         /* set the operation */
517:         upTriFactor->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

519:         /* set the matrix */
520:         upTriFactor->csrMat              = new CsrMatrix;
521:         upTriFactor->csrMat->num_rows    = n;
522:         upTriFactor->csrMat->num_cols    = n;
523:         upTriFactor->csrMat->num_entries = nzUpper;

525:         upTriFactor->csrMat->row_offsets = new THRUSTINTARRAY32(n + 1);
526:         upTriFactor->csrMat->row_offsets->assign(AiUp, AiUp + n + 1);

528:         upTriFactor->csrMat->column_indices = new THRUSTINTARRAY32(nzUpper);
529:         upTriFactor->csrMat->column_indices->assign(AjUp, AjUp + nzUpper);

531:         upTriFactor->csrMat->values = new THRUSTARRAY(nzUpper);
532:         upTriFactor->csrMat->values->assign(AAUp, AAUp + nzUpper);

534:         /* Create the solve analysis information */
535:         PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
536:         PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&upTriFactor->solveInfo));
537:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
538:         PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, upTriFactor->solveOp, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_entries, upTriFactor->descr, upTriFactor->csrMat->values->data().get(),
539:                                                   upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, &upTriFactor->solveBufferSize));
540:         PetscCallCUDA(cudaMalloc(&upTriFactor->solveBuffer, upTriFactor->solveBufferSize));
541:   #endif

543:         /* perform the solve analysis */
544:         PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, upTriFactor->solveOp, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_entries, upTriFactor->descr, upTriFactor->csrMat->values->data().get(),
545:                                                   upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, upTriFactor->solvePolicy, upTriFactor->solveBuffer));

547:         PetscCallCUDA(WaitForCUDA());
548:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

550:         /* assign the pointer */
551:         ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->upTriFactorPtr = upTriFactor;
552:         upTriFactor->AA_h                                          = AAUp;
553:         PetscCallCUDA(cudaFreeHost(AiUp));
554:         PetscCallCUDA(cudaFreeHost(AjUp));
555:         PetscCall(PetscLogCpuToGpu((n + 1 + nzUpper) * sizeof(int) + nzUpper * sizeof(PetscScalar)));
556:       } else {
557:         if (!upTriFactor->AA_h) PetscCallCUDA(cudaMallocHost((void **)&upTriFactor->AA_h, nzUpper * sizeof(PetscScalar)));
558:         /* Fill the upper triangular matrix */
559:         offset = nzUpper;
560:         for (i = n - 1; i >= 0; i--) {
561:           v = aa + adiag[i + 1] + 1;

563:           /* number of elements NOT on the diagonal */
564:           nz = adiag[i] - adiag[i + 1] - 1;

566:           /* decrement the offset */
567:           offset -= (nz + 1);

569:           /* first, set the diagonal elements */
570:           upTriFactor->AA_h[offset] = 1. / v[nz];
571:           PetscCall(PetscArraycpy(&upTriFactor->AA_h[offset + 1], v, nz));
572:         }
573:         upTriFactor->csrMat->values->assign(upTriFactor->AA_h, upTriFactor->AA_h + nzUpper);
574:         PetscCall(PetscLogCpuToGpu(nzUpper * sizeof(PetscScalar)));
575:       }
576:     } catch (char *ex) {
577:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
578:     }
579:   }
580:   PetscFunctionReturn(PETSC_SUCCESS);
581: }
582: #endif

584: static PetscErrorCode MatSeqAIJCUSPARSEILUAnalysisAndCopyToGPU(Mat A)
585: {
586:   Mat_SeqAIJ                   *a                  = (Mat_SeqAIJ *)A->data;
587:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
588:   IS                            isrow = a->row, iscol = a->icol;
589:   PetscBool                     row_identity, col_identity;
590:   PetscInt                      n = A->rmap->n;

592:   PetscFunctionBegin;
593:   PetscCheck(cusparseTriFactors, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing cusparseTriFactors");
594: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
595:   PetscCall(MatSeqAIJCUSPARSEBuildFactoredMatrix_LU(A));
596: #else
597:   PetscCall(MatSeqAIJCUSPARSEBuildILULowerTriMatrix(A));
598:   PetscCall(MatSeqAIJCUSPARSEBuildILUUpperTriMatrix(A));
599:   if (!cusparseTriFactors->workVector) cusparseTriFactors->workVector = new THRUSTARRAY(n);
600: #endif

602:   cusparseTriFactors->nnz = a->nz;

604:   A->offloadmask = PETSC_OFFLOAD_BOTH; // factored matrix is sync'ed to GPU
605:   /* lower triangular indices */
606:   PetscCall(ISIdentity(isrow, &row_identity));
607:   if (!row_identity && !cusparseTriFactors->rpermIndices) {
608:     const PetscInt *r;

610:     PetscCall(ISGetIndices(isrow, &r));
611:     cusparseTriFactors->rpermIndices = new THRUSTINTARRAY(n);
612:     cusparseTriFactors->rpermIndices->assign(r, r + n);
613:     PetscCall(ISRestoreIndices(isrow, &r));
614:     PetscCall(PetscLogCpuToGpu(n * sizeof(PetscInt)));
615:   }

617:   /* upper triangular indices */
618:   PetscCall(ISIdentity(iscol, &col_identity));
619:   if (!col_identity && !cusparseTriFactors->cpermIndices) {
620:     const PetscInt *c;

622:     PetscCall(ISGetIndices(iscol, &c));
623:     cusparseTriFactors->cpermIndices = new THRUSTINTARRAY(n);
624:     cusparseTriFactors->cpermIndices->assign(c, c + n);
625:     PetscCall(ISRestoreIndices(iscol, &c));
626:     PetscCall(PetscLogCpuToGpu(n * sizeof(PetscInt)));
627:   }
628:   PetscFunctionReturn(PETSC_SUCCESS);
629: }

631: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
632: static PetscErrorCode MatSeqAIJCUSPARSEBuildFactoredMatrix_Cheolesky(Mat A)
633: {
634:   Mat_SeqAIJ                   *a  = static_cast<Mat_SeqAIJ *>(A->data);
635:   PetscInt                      m  = A->rmap->n;
636:   Mat_SeqAIJCUSPARSETriFactors *fs = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
637:   const PetscInt               *Ai = a->i, *Aj = a->j, *Adiag = a->diag;
638:   const MatScalar              *Aa = a->a;
639:   PetscInt                     *Mj, Mnz;
640:   PetscScalar                  *Ma, *D;

642:   PetscFunctionBegin;
643:   if (A->offloadmask == PETSC_OFFLOAD_CPU) { // A's latest factors are on CPU
644:     if (!fs->csrRowPtr) {                    // Is't the first time to do the setup? Use csrRowPtr since it is not null even m=0
645:       // Re-arrange the (skewed) factored matrix and put the result into M, a regular csr matrix on host.
646:       // See comments at MatICCFactorSymbolic_SeqAIJ() on the layout of the factored matrix (U) on host.
647:       Mnz = Ai[m]; // Unz (with the unit diagonal)
648:       PetscCall(PetscMalloc1(Mnz, &Ma));
649:       PetscCall(PetscMalloc1(Mnz, &Mj)); // Mj[] is temp
650:       PetscCall(PetscMalloc1(m, &D));    // the diagonal
651:       for (PetscInt i = 0; i < m; i++) {
652:         PetscInt ulen = Ai[i + 1] - Ai[i];
653:         Mj[Ai[i]]     = i;                                              // diagonal entry
654:         PetscCall(PetscArraycpy(Mj + Ai[i] + 1, Aj + Ai[i], ulen - 1)); // entries of U on the right of the diagonal
655:       }
656:       // Copy M (U) from host to device
657:       PetscCallCUDA(cudaMalloc(&fs->csrRowPtr, sizeof(*fs->csrRowPtr) * (m + 1)));
658:       PetscCallCUDA(cudaMalloc(&fs->csrColIdx, sizeof(*fs->csrColIdx) * Mnz));
659:       PetscCallCUDA(cudaMalloc(&fs->csrVal, sizeof(*fs->csrVal) * Mnz));
660:       PetscCallCUDA(cudaMalloc(&fs->diag, sizeof(*fs->diag) * m));
661:       PetscCallCUDA(cudaMemcpy(fs->csrRowPtr, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyHostToDevice));
662:       PetscCallCUDA(cudaMemcpy(fs->csrColIdx, Mj, sizeof(*Mj) * Mnz, cudaMemcpyHostToDevice));

664:       // Create descriptors for L, U. See https://docs.nvidia.com/cuda/cusparse/index.html#cusparseDiagType_t
665:       // cusparseDiagType_t: This type indicates if the matrix diagonal entries are unity. The diagonal elements are always
666:       // assumed to be present, but if CUSPARSE_DIAG_TYPE_UNIT is passed to an API routine, then the routine assumes that
667:       // all diagonal entries are unity and will not read or modify those entries. Note that in this case the routine
668:       // assumes the diagonal entries are equal to one, regardless of what those entries are actually set to in memory.
669:       cusparseFillMode_t        fillMode  = CUSPARSE_FILL_MODE_UPPER;
670:       cusparseDiagType_t        diagType  = CUSPARSE_DIAG_TYPE_UNIT; // U is unit diagonal
671:       const cusparseIndexType_t indexType = PetscDefined(USE_64BIT_INDICES) ? CUSPARSE_INDEX_64I : CUSPARSE_INDEX_32I;

673:       PetscCallCUSPARSE(cusparseCreateCsr(&fs->spMatDescr_U, m, m, Mnz, fs->csrRowPtr, fs->csrColIdx, fs->csrVal, indexType, indexType, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype));
674:       PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
675:       PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

677:       // Allocate work vectors in SpSv
678:       PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(*fs->X) * m));
679:       PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(*fs->Y) * m));

681:       PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_X, m, fs->X, cusparse_scalartype));
682:       PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_Y, m, fs->Y, cusparse_scalartype));

684:       // Query buffer sizes for SpSV and then allocate buffers
685:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_U));
686:       PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_U, &fs->spsvBufferSize_U));
687:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U));

689:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Ut)); // Ut solve uses the same matrix (spMatDescr_U), but different descr and buffer
690:       PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, CUSPARSE_OPERATION_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_Ut, &fs->spsvBufferSize_Ut));
691:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Ut, fs->spsvBufferSize_Ut));

693:       // Record for reuse
694:       fs->csrVal_h = Ma;
695:       fs->diag_h   = D;
696:       PetscCall(PetscFree(Mj));
697:     }
698:     // Copy the value
699:     Ma  = fs->csrVal_h;
700:     D   = fs->diag_h;
701:     Mnz = Ai[m];
702:     for (PetscInt i = 0; i < m; i++) {
703:       D[i]      = Aa[Adiag[i]];   // actually Aa[Adiag[i]] is the inverse of the diagonal
704:       Ma[Ai[i]] = (MatScalar)1.0; // set the unit diagonal, which is cosmetic since cusparse does not really read it given CUSPARSE_DIAG_TYPE_UNIT
705:       for (PetscInt k = 0; k < Ai[i + 1] - Ai[i] - 1; k++) Ma[Ai[i] + 1 + k] = -Aa[Ai[i] + k];
706:     }
707:     PetscCallCUDA(cudaMemcpy(fs->csrVal, Ma, sizeof(*Ma) * Mnz, cudaMemcpyHostToDevice));
708:     PetscCallCUDA(cudaMemcpy(fs->diag, D, sizeof(*D) * m, cudaMemcpyHostToDevice));

710:     // Do cusparseSpSV_analysis(), which is numeric and requires valid and up-to-date matrix values
711:     PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_U, fs->spsvBuffer_U));
712:     PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, CUSPARSE_OPERATION_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_Ut, fs->spsvBuffer_Ut));
713:   }
714:   PetscFunctionReturn(PETSC_SUCCESS);
715: }

717: // Solve Ut D U x = b
718: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_Cholesky(Mat A, Vec b, Vec x)
719: {
720:   Mat_SeqAIJCUSPARSETriFactors         *fs  = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
721:   Mat_SeqAIJ                           *aij = static_cast<Mat_SeqAIJ *>(A->data);
722:   const PetscScalar                    *barray;
723:   PetscScalar                          *xarray;
724:   thrust::device_ptr<const PetscScalar> bGPU;
725:   thrust::device_ptr<PetscScalar>       xGPU;
726:   const cusparseSpSVAlg_t               alg = CUSPARSE_SPSV_ALG_DEFAULT;
727:   PetscInt                              m   = A->rmap->n;

729:   PetscFunctionBegin;
730:   PetscCall(PetscLogGpuTimeBegin());
731:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
732:   PetscCall(VecCUDAGetArrayRead(b, &barray));
733:   xGPU = thrust::device_pointer_cast(xarray);
734:   bGPU = thrust::device_pointer_cast(barray);

736:   // Reorder b with the row permutation if needed, and wrap the result in fs->X
737:   if (fs->rpermIndices) {
738:     PetscCallThrust(thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(bGPU, fs->rpermIndices->begin()), thrust::make_permutation_iterator(bGPU, fs->rpermIndices->end()), thrust::device_pointer_cast(fs->X)));
739:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
740:   } else {
741:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
742:   }

744:   // Solve Ut Y = X
745:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
746:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, CUSPARSE_OPERATION_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_Ut));

748:   // Solve diag(D) Z = Y. Actually just do Y = Y*D since D is already inverted in MatCholeskyFactorNumeric_SeqAIJ().
749:   // It is basically a vector element-wise multiplication, but cublas does not have it!
750:   PetscCallThrust(thrust::transform(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::device_pointer_cast(fs->Y), thrust::device_pointer_cast(fs->Y + m), thrust::device_pointer_cast(fs->diag), thrust::device_pointer_cast(fs->Y), thrust::multiplies<PetscScalar>()));

752:   // Solve U X = Y
753:   if (fs->cpermIndices) { // if need to permute, we need to use the intermediate buffer X
754:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
755:   } else {
756:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
757:   }
758:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_Y, fs->dnVecDescr_X, cusparse_scalartype, alg, fs->spsvDescr_U));

760:   // Reorder X with the column permutation if needed, and put the result back to x
761:   if (fs->cpermIndices) {
762:     PetscCallThrust(thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X), fs->cpermIndices->begin()),
763:                                  thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X + m), fs->cpermIndices->end()), xGPU));
764:   }

766:   PetscCall(VecCUDARestoreArrayRead(b, &barray));
767:   PetscCall(VecCUDARestoreArrayWrite(x, &xarray));
768:   PetscCall(PetscLogGpuTimeEnd());
769:   PetscCall(PetscLogGpuFlops(4.0 * aij->nz - A->rmap->n));
770:   PetscFunctionReturn(PETSC_SUCCESS);
771: }
772: #else
773: static PetscErrorCode MatSeqAIJCUSPARSEBuildICCTriMatrices(Mat A)
774: {
775:   Mat_SeqAIJ                        *a                  = (Mat_SeqAIJ *)A->data;
776:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
777:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
778:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
779:   PetscInt                          *AiUp, *AjUp;
780:   PetscScalar                       *AAUp;
781:   PetscScalar                       *AALo;
782:   PetscInt                           nzUpper = a->nz, n = A->rmap->n, i, offset, nz, j;
783:   Mat_SeqSBAIJ                      *b  = (Mat_SeqSBAIJ *)A->data;
784:   const PetscInt                    *ai = b->i, *aj = b->j, *vj;
785:   const MatScalar                   *aa = b->a, *v;

787:   PetscFunctionBegin;
788:   if (!n) PetscFunctionReturn(PETSC_SUCCESS);
789:   if (A->offloadmask == PETSC_OFFLOAD_UNALLOCATED || A->offloadmask == PETSC_OFFLOAD_CPU) {
790:     try {
791:       PetscCallCUDA(cudaMallocHost((void **)&AAUp, nzUpper * sizeof(PetscScalar)));
792:       PetscCallCUDA(cudaMallocHost((void **)&AALo, nzUpper * sizeof(PetscScalar)));
793:       if (!upTriFactor && !loTriFactor) {
794:         /* Allocate Space for the upper triangular matrix */
795:         PetscCallCUDA(cudaMallocHost((void **)&AiUp, (n + 1) * sizeof(PetscInt)));
796:         PetscCallCUDA(cudaMallocHost((void **)&AjUp, nzUpper * sizeof(PetscInt)));

798:         /* Fill the upper triangular matrix */
799:         AiUp[0] = (PetscInt)0;
800:         AiUp[n] = nzUpper;
801:         offset  = 0;
802:         for (i = 0; i < n; i++) {
803:           /* set the pointers */
804:           v  = aa + ai[i];
805:           vj = aj + ai[i];
806:           nz = ai[i + 1] - ai[i] - 1; /* exclude diag[i] */

808:           /* first, set the diagonal elements */
809:           AjUp[offset] = (PetscInt)i;
810:           AAUp[offset] = (MatScalar)1.0 / v[nz];
811:           AiUp[i]      = offset;
812:           AALo[offset] = (MatScalar)1.0 / v[nz];

814:           offset += 1;
815:           if (nz > 0) {
816:             PetscCall(PetscArraycpy(&AjUp[offset], vj, nz));
817:             PetscCall(PetscArraycpy(&AAUp[offset], v, nz));
818:             for (j = offset; j < offset + nz; j++) {
819:               AAUp[j] = -AAUp[j];
820:               AALo[j] = AAUp[j] / v[nz];
821:             }
822:             offset += nz;
823:           }
824:         }

826:         /* allocate space for the triangular factor information */
827:         PetscCall(PetscNew(&upTriFactor));
828:         upTriFactor->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

830:         /* Create the matrix description */
831:         PetscCallCUSPARSE(cusparseCreateMatDescr(&upTriFactor->descr));
832:         PetscCallCUSPARSE(cusparseSetMatIndexBase(upTriFactor->descr, CUSPARSE_INDEX_BASE_ZERO));
833:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
834:         PetscCallCUSPARSE(cusparseSetMatType(upTriFactor->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
835:   #else
836:         PetscCallCUSPARSE(cusparseSetMatType(upTriFactor->descr, CUSPARSE_MATRIX_TYPE_TRIANGULAR));
837:   #endif
838:         PetscCallCUSPARSE(cusparseSetMatFillMode(upTriFactor->descr, CUSPARSE_FILL_MODE_UPPER));
839:         PetscCallCUSPARSE(cusparseSetMatDiagType(upTriFactor->descr, CUSPARSE_DIAG_TYPE_UNIT));

841:         /* set the matrix */
842:         upTriFactor->csrMat              = new CsrMatrix;
843:         upTriFactor->csrMat->num_rows    = A->rmap->n;
844:         upTriFactor->csrMat->num_cols    = A->cmap->n;
845:         upTriFactor->csrMat->num_entries = a->nz;

847:         upTriFactor->csrMat->row_offsets = new THRUSTINTARRAY32(A->rmap->n + 1);
848:         upTriFactor->csrMat->row_offsets->assign(AiUp, AiUp + A->rmap->n + 1);

850:         upTriFactor->csrMat->column_indices = new THRUSTINTARRAY32(a->nz);
851:         upTriFactor->csrMat->column_indices->assign(AjUp, AjUp + a->nz);

853:         upTriFactor->csrMat->values = new THRUSTARRAY(a->nz);
854:         upTriFactor->csrMat->values->assign(AAUp, AAUp + a->nz);

856:         /* set the operation */
857:         upTriFactor->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

859:         /* Create the solve analysis information */
860:         PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
861:         PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&upTriFactor->solveInfo));
862:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
863:         PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, upTriFactor->solveOp, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_entries, upTriFactor->descr, upTriFactor->csrMat->values->data().get(),
864:                                                   upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, &upTriFactor->solveBufferSize));
865:         PetscCallCUDA(cudaMalloc(&upTriFactor->solveBuffer, upTriFactor->solveBufferSize));
866:   #endif

868:         /* perform the solve analysis */
869:         PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, upTriFactor->solveOp, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_entries, upTriFactor->descr, upTriFactor->csrMat->values->data().get(),
870:                                                   upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, upTriFactor->solvePolicy, upTriFactor->solveBuffer));

872:         PetscCallCUDA(WaitForCUDA());
873:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

875:         /* assign the pointer */
876:         ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->upTriFactorPtr = upTriFactor;

878:         /* allocate space for the triangular factor information */
879:         PetscCall(PetscNew(&loTriFactor));
880:         loTriFactor->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

882:         /* Create the matrix description */
883:         PetscCallCUSPARSE(cusparseCreateMatDescr(&loTriFactor->descr));
884:         PetscCallCUSPARSE(cusparseSetMatIndexBase(loTriFactor->descr, CUSPARSE_INDEX_BASE_ZERO));
885:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
886:         PetscCallCUSPARSE(cusparseSetMatType(loTriFactor->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
887:   #else
888:         PetscCallCUSPARSE(cusparseSetMatType(loTriFactor->descr, CUSPARSE_MATRIX_TYPE_TRIANGULAR));
889:   #endif
890:         PetscCallCUSPARSE(cusparseSetMatFillMode(loTriFactor->descr, CUSPARSE_FILL_MODE_UPPER));
891:         PetscCallCUSPARSE(cusparseSetMatDiagType(loTriFactor->descr, CUSPARSE_DIAG_TYPE_NON_UNIT));

893:         /* set the operation */
894:         loTriFactor->solveOp = CUSPARSE_OPERATION_TRANSPOSE;

896:         /* set the matrix */
897:         loTriFactor->csrMat              = new CsrMatrix;
898:         loTriFactor->csrMat->num_rows    = A->rmap->n;
899:         loTriFactor->csrMat->num_cols    = A->cmap->n;
900:         loTriFactor->csrMat->num_entries = a->nz;

902:         loTriFactor->csrMat->row_offsets = new THRUSTINTARRAY32(A->rmap->n + 1);
903:         loTriFactor->csrMat->row_offsets->assign(AiUp, AiUp + A->rmap->n + 1);

905:         loTriFactor->csrMat->column_indices = new THRUSTINTARRAY32(a->nz);
906:         loTriFactor->csrMat->column_indices->assign(AjUp, AjUp + a->nz);

908:         loTriFactor->csrMat->values = new THRUSTARRAY(a->nz);
909:         loTriFactor->csrMat->values->assign(AALo, AALo + a->nz);

911:         /* Create the solve analysis information */
912:         PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
913:         PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&loTriFactor->solveInfo));
914:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
915:         PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, loTriFactor->solveOp, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_entries, loTriFactor->descr, loTriFactor->csrMat->values->data().get(),
916:                                                   loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, &loTriFactor->solveBufferSize));
917:         PetscCallCUDA(cudaMalloc(&loTriFactor->solveBuffer, loTriFactor->solveBufferSize));
918:   #endif

920:         /* perform the solve analysis */
921:         PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, loTriFactor->solveOp, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_entries, loTriFactor->descr, loTriFactor->csrMat->values->data().get(),
922:                                                   loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, loTriFactor->solvePolicy, loTriFactor->solveBuffer));

924:         PetscCallCUDA(WaitForCUDA());
925:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

927:         /* assign the pointer */
928:         ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->loTriFactorPtr = loTriFactor;

930:         PetscCall(PetscLogCpuToGpu(2 * (((A->rmap->n + 1) + (a->nz)) * sizeof(int) + (a->nz) * sizeof(PetscScalar))));
931:         PetscCallCUDA(cudaFreeHost(AiUp));
932:         PetscCallCUDA(cudaFreeHost(AjUp));
933:       } else {
934:         /* Fill the upper triangular matrix */
935:         offset = 0;
936:         for (i = 0; i < n; i++) {
937:           /* set the pointers */
938:           v  = aa + ai[i];
939:           nz = ai[i + 1] - ai[i] - 1; /* exclude diag[i] */

941:           /* first, set the diagonal elements */
942:           AAUp[offset] = 1.0 / v[nz];
943:           AALo[offset] = 1.0 / v[nz];

945:           offset += 1;
946:           if (nz > 0) {
947:             PetscCall(PetscArraycpy(&AAUp[offset], v, nz));
948:             for (j = offset; j < offset + nz; j++) {
949:               AAUp[j] = -AAUp[j];
950:               AALo[j] = AAUp[j] / v[nz];
951:             }
952:             offset += nz;
953:           }
954:         }
955:         PetscCheck(upTriFactor, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing cusparseTriFactors");
956:         PetscCheck(loTriFactor, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing cusparseTriFactors");
957:         upTriFactor->csrMat->values->assign(AAUp, AAUp + a->nz);
958:         loTriFactor->csrMat->values->assign(AALo, AALo + a->nz);
959:         PetscCall(PetscLogCpuToGpu(2 * (a->nz) * sizeof(PetscScalar)));
960:       }
961:       PetscCallCUDA(cudaFreeHost(AAUp));
962:       PetscCallCUDA(cudaFreeHost(AALo));
963:     } catch (char *ex) {
964:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
965:     }
966:   }
967:   PetscFunctionReturn(PETSC_SUCCESS);
968: }
969: #endif

971: static PetscErrorCode MatSeqAIJCUSPARSEICCAnalysisAndCopyToGPU(Mat A)
972: {
973:   Mat_SeqAIJ                   *a                  = (Mat_SeqAIJ *)A->data;
974:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
975:   IS                            ip                 = a->row;
976:   PetscBool                     perm_identity;
977:   PetscInt                      n = A->rmap->n;

979:   PetscFunctionBegin;
980:   PetscCheck(cusparseTriFactors, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing cusparseTriFactors");

982: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
983:   PetscCall(MatSeqAIJCUSPARSEBuildFactoredMatrix_Cheolesky(A));
984: #else
985:   PetscCall(MatSeqAIJCUSPARSEBuildICCTriMatrices(A));
986:   if (!cusparseTriFactors->workVector) cusparseTriFactors->workVector = new THRUSTARRAY(n);
987: #endif
988:   cusparseTriFactors->nnz = (a->nz - n) * 2 + n;

990:   A->offloadmask = PETSC_OFFLOAD_BOTH;

992:   /* lower triangular indices */
993:   PetscCall(ISIdentity(ip, &perm_identity));
994:   if (!perm_identity) {
995:     IS              iip;
996:     const PetscInt *irip, *rip;

998:     PetscCall(ISInvertPermutation(ip, PETSC_DECIDE, &iip));
999:     PetscCall(ISGetIndices(iip, &irip));
1000:     PetscCall(ISGetIndices(ip, &rip));
1001:     cusparseTriFactors->rpermIndices = new THRUSTINTARRAY(n);
1002:     cusparseTriFactors->rpermIndices->assign(rip, rip + n);
1003:     cusparseTriFactors->cpermIndices = new THRUSTINTARRAY(n);
1004:     cusparseTriFactors->cpermIndices->assign(irip, irip + n);
1005:     PetscCall(ISRestoreIndices(iip, &irip));
1006:     PetscCall(ISDestroy(&iip));
1007:     PetscCall(ISRestoreIndices(ip, &rip));
1008:     PetscCall(PetscLogCpuToGpu(2. * n * sizeof(PetscInt)));
1009:   }
1010:   PetscFunctionReturn(PETSC_SUCCESS);
1011: }

1013: static PetscErrorCode MatCholeskyFactorNumeric_SeqAIJCUSPARSE(Mat B, Mat A, const MatFactorInfo *info)
1014: {
1015:   PetscFunctionBegin;
1016:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
1017:   PetscCall(MatCholeskyFactorNumeric_SeqAIJ(B, A, info));
1018:   B->offloadmask = PETSC_OFFLOAD_CPU;

1020: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
1021:   B->ops->solve          = MatSolve_SeqAIJCUSPARSE_Cholesky;
1022:   B->ops->solvetranspose = MatSolve_SeqAIJCUSPARSE_Cholesky;
1023: #else
1024:   /* determine which version of MatSolve needs to be used. */
1025:   Mat_SeqAIJ *b  = (Mat_SeqAIJ *)B->data;
1026:   IS          ip = b->row;
1027:   PetscBool   perm_identity;

1029:   PetscCall(ISIdentity(ip, &perm_identity));
1030:   if (perm_identity) {
1031:     B->ops->solve          = MatSolve_SeqAIJCUSPARSE_NaturalOrdering;
1032:     B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering;
1033:   } else {
1034:     B->ops->solve          = MatSolve_SeqAIJCUSPARSE;
1035:     B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE;
1036:   }
1037: #endif
1038:   B->ops->matsolve          = NULL;
1039:   B->ops->matsolvetranspose = NULL;

1041:   /* get the triangular factors */
1042:   PetscCall(MatSeqAIJCUSPARSEICCAnalysisAndCopyToGPU(B));
1043:   PetscFunctionReturn(PETSC_SUCCESS);
1044: }

1046: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
1047: static PetscErrorCode MatSeqAIJCUSPARSEAnalyzeTransposeForSolve(Mat A)
1048: {
1049:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1050:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
1051:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
1052:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorT;
1053:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorT;
1054:   cusparseIndexBase_t                indexBase;
1055:   cusparseMatrixType_t               matrixType;
1056:   cusparseFillMode_t                 fillMode;
1057:   cusparseDiagType_t                 diagType;

1059:   PetscFunctionBegin;
1060:   /* allocate space for the transpose of the lower triangular factor */
1061:   PetscCall(PetscNew(&loTriFactorT));
1062:   loTriFactorT->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

1064:   /* set the matrix descriptors of the lower triangular factor */
1065:   matrixType = cusparseGetMatType(loTriFactor->descr);
1066:   indexBase  = cusparseGetMatIndexBase(loTriFactor->descr);
1067:   fillMode   = cusparseGetMatFillMode(loTriFactor->descr) == CUSPARSE_FILL_MODE_UPPER ? CUSPARSE_FILL_MODE_LOWER : CUSPARSE_FILL_MODE_UPPER;
1068:   diagType   = cusparseGetMatDiagType(loTriFactor->descr);

1070:   /* Create the matrix description */
1071:   PetscCallCUSPARSE(cusparseCreateMatDescr(&loTriFactorT->descr));
1072:   PetscCallCUSPARSE(cusparseSetMatIndexBase(loTriFactorT->descr, indexBase));
1073:   PetscCallCUSPARSE(cusparseSetMatType(loTriFactorT->descr, matrixType));
1074:   PetscCallCUSPARSE(cusparseSetMatFillMode(loTriFactorT->descr, fillMode));
1075:   PetscCallCUSPARSE(cusparseSetMatDiagType(loTriFactorT->descr, diagType));

1077:   /* set the operation */
1078:   loTriFactorT->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

1080:   /* allocate GPU space for the CSC of the lower triangular factor*/
1081:   loTriFactorT->csrMat                 = new CsrMatrix;
1082:   loTriFactorT->csrMat->num_rows       = loTriFactor->csrMat->num_cols;
1083:   loTriFactorT->csrMat->num_cols       = loTriFactor->csrMat->num_rows;
1084:   loTriFactorT->csrMat->num_entries    = loTriFactor->csrMat->num_entries;
1085:   loTriFactorT->csrMat->row_offsets    = new THRUSTINTARRAY32(loTriFactorT->csrMat->num_rows + 1);
1086:   loTriFactorT->csrMat->column_indices = new THRUSTINTARRAY32(loTriFactorT->csrMat->num_entries);
1087:   loTriFactorT->csrMat->values         = new THRUSTARRAY(loTriFactorT->csrMat->num_entries);

1089:   /* compute the transpose of the lower triangular factor, i.e. the CSC */
1090:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1091:   PetscCallCUSPARSE(cusparseCsr2cscEx2_bufferSize(cusparseTriFactors->handle, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_cols, loTriFactor->csrMat->num_entries, loTriFactor->csrMat->values->data().get(),
1092:                                                   loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactorT->csrMat->values->data().get(), loTriFactorT->csrMat->row_offsets->data().get(),
1093:                                                   loTriFactorT->csrMat->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, CUSPARSE_CSR2CSC_ALG1, &loTriFactor->csr2cscBufferSize));
1094:   PetscCallCUDA(cudaMalloc(&loTriFactor->csr2cscBuffer, loTriFactor->csr2cscBufferSize));
1095:   #endif

1097:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));
1098:   {
1099:     // there is no clean way to have PetscCallCUSPARSE wrapping this function...
1100:     auto stat = cusparse_csr2csc(cusparseTriFactors->handle, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_cols, loTriFactor->csrMat->num_entries, loTriFactor->csrMat->values->data().get(), loTriFactor->csrMat->row_offsets->data().get(),
1101:                                  loTriFactor->csrMat->column_indices->data().get(), loTriFactorT->csrMat->values->data().get(),
1102:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1103:                                  loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, CUSPARSE_CSR2CSC_ALG1, loTriFactor->csr2cscBuffer);
1104:   #else
1105:                                  loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->csrMat->row_offsets->data().get(), CUSPARSE_ACTION_NUMERIC, indexBase);
1106:   #endif
1107:     PetscCallCUSPARSE(stat);
1108:   }

1110:   PetscCallCUDA(WaitForCUDA());
1111:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));

1113:   /* Create the solve analysis information */
1114:   PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
1115:   PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&loTriFactorT->solveInfo));
1116:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
1117:   PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, loTriFactorT->solveOp, loTriFactorT->csrMat->num_rows, loTriFactorT->csrMat->num_entries, loTriFactorT->descr, loTriFactorT->csrMat->values->data().get(),
1118:                                             loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->solveInfo, &loTriFactorT->solveBufferSize));
1119:   PetscCallCUDA(cudaMalloc(&loTriFactorT->solveBuffer, loTriFactorT->solveBufferSize));
1120:   #endif

1122:   /* perform the solve analysis */
1123:   PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, loTriFactorT->solveOp, loTriFactorT->csrMat->num_rows, loTriFactorT->csrMat->num_entries, loTriFactorT->descr, loTriFactorT->csrMat->values->data().get(),
1124:                                             loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->solveInfo, loTriFactorT->solvePolicy, loTriFactorT->solveBuffer));

1126:   PetscCallCUDA(WaitForCUDA());
1127:   PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

1129:   /* assign the pointer */
1130:   ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->loTriFactorPtrTranspose = loTriFactorT;

1132:   /*********************************************/
1133:   /* Now the Transpose of the Upper Tri Factor */
1134:   /*********************************************/

1136:   /* allocate space for the transpose of the upper triangular factor */
1137:   PetscCall(PetscNew(&upTriFactorT));
1138:   upTriFactorT->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

1140:   /* set the matrix descriptors of the upper triangular factor */
1141:   matrixType = cusparseGetMatType(upTriFactor->descr);
1142:   indexBase  = cusparseGetMatIndexBase(upTriFactor->descr);
1143:   fillMode   = cusparseGetMatFillMode(upTriFactor->descr) == CUSPARSE_FILL_MODE_UPPER ? CUSPARSE_FILL_MODE_LOWER : CUSPARSE_FILL_MODE_UPPER;
1144:   diagType   = cusparseGetMatDiagType(upTriFactor->descr);

1146:   /* Create the matrix description */
1147:   PetscCallCUSPARSE(cusparseCreateMatDescr(&upTriFactorT->descr));
1148:   PetscCallCUSPARSE(cusparseSetMatIndexBase(upTriFactorT->descr, indexBase));
1149:   PetscCallCUSPARSE(cusparseSetMatType(upTriFactorT->descr, matrixType));
1150:   PetscCallCUSPARSE(cusparseSetMatFillMode(upTriFactorT->descr, fillMode));
1151:   PetscCallCUSPARSE(cusparseSetMatDiagType(upTriFactorT->descr, diagType));

1153:   /* set the operation */
1154:   upTriFactorT->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

1156:   /* allocate GPU space for the CSC of the upper triangular factor*/
1157:   upTriFactorT->csrMat                 = new CsrMatrix;
1158:   upTriFactorT->csrMat->num_rows       = upTriFactor->csrMat->num_cols;
1159:   upTriFactorT->csrMat->num_cols       = upTriFactor->csrMat->num_rows;
1160:   upTriFactorT->csrMat->num_entries    = upTriFactor->csrMat->num_entries;
1161:   upTriFactorT->csrMat->row_offsets    = new THRUSTINTARRAY32(upTriFactorT->csrMat->num_rows + 1);
1162:   upTriFactorT->csrMat->column_indices = new THRUSTINTARRAY32(upTriFactorT->csrMat->num_entries);
1163:   upTriFactorT->csrMat->values         = new THRUSTARRAY(upTriFactorT->csrMat->num_entries);

1165:   /* compute the transpose of the upper triangular factor, i.e. the CSC */
1166:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1167:   PetscCallCUSPARSE(cusparseCsr2cscEx2_bufferSize(cusparseTriFactors->handle, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_cols, upTriFactor->csrMat->num_entries, upTriFactor->csrMat->values->data().get(),
1168:                                                   upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactorT->csrMat->values->data().get(), upTriFactorT->csrMat->row_offsets->data().get(),
1169:                                                   upTriFactorT->csrMat->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, CUSPARSE_CSR2CSC_ALG1, &upTriFactor->csr2cscBufferSize));
1170:   PetscCallCUDA(cudaMalloc(&upTriFactor->csr2cscBuffer, upTriFactor->csr2cscBufferSize));
1171:   #endif

1173:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));
1174:   {
1175:     // there is no clean way to have PetscCallCUSPARSE wrapping this function...
1176:     auto stat = cusparse_csr2csc(cusparseTriFactors->handle, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_cols, upTriFactor->csrMat->num_entries, upTriFactor->csrMat->values->data().get(), upTriFactor->csrMat->row_offsets->data().get(),
1177:                                  upTriFactor->csrMat->column_indices->data().get(), upTriFactorT->csrMat->values->data().get(),
1178:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1179:                                  upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, CUSPARSE_CSR2CSC_ALG1, upTriFactor->csr2cscBuffer);
1180:   #else
1181:                                  upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->csrMat->row_offsets->data().get(), CUSPARSE_ACTION_NUMERIC, indexBase);
1182:   #endif
1183:     PetscCallCUSPARSE(stat);
1184:   }

1186:   PetscCallCUDA(WaitForCUDA());
1187:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));

1189:   /* Create the solve analysis information */
1190:   PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
1191:   PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&upTriFactorT->solveInfo));
1192:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
1193:   PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, upTriFactorT->solveOp, upTriFactorT->csrMat->num_rows, upTriFactorT->csrMat->num_entries, upTriFactorT->descr, upTriFactorT->csrMat->values->data().get(),
1194:                                             upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->solveInfo, &upTriFactorT->solveBufferSize));
1195:   PetscCallCUDA(cudaMalloc(&upTriFactorT->solveBuffer, upTriFactorT->solveBufferSize));
1196:   #endif

1198:   /* perform the solve analysis */
1199:   /* christ, would it have killed you to put this stuff in a function????????? */
1200:   PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, upTriFactorT->solveOp, upTriFactorT->csrMat->num_rows, upTriFactorT->csrMat->num_entries, upTriFactorT->descr, upTriFactorT->csrMat->values->data().get(),
1201:                                             upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->solveInfo, upTriFactorT->solvePolicy, upTriFactorT->solveBuffer));

1203:   PetscCallCUDA(WaitForCUDA());
1204:   PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

1206:   /* assign the pointer */
1207:   ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->upTriFactorPtrTranspose = upTriFactorT;
1208:   PetscFunctionReturn(PETSC_SUCCESS);
1209: }
1210: #endif

1212: struct PetscScalarToPetscInt {
1213:   __host__ __device__ PetscInt operator()(PetscScalar s) { return (PetscInt)PetscRealPart(s); }
1214: };

1216: static PetscErrorCode MatSeqAIJCUSPARSEFormExplicitTranspose(Mat A)
1217: {
1218:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
1219:   Mat_SeqAIJCUSPARSEMultStruct *matstruct, *matstructT;
1220:   Mat_SeqAIJ                   *a = (Mat_SeqAIJ *)A->data;
1221:   cusparseStatus_t              stat;
1222:   cusparseIndexBase_t           indexBase;

1224:   PetscFunctionBegin;
1225:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
1226:   matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
1227:   PetscCheck(matstruct, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing mat struct");
1228:   matstructT = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->matTranspose;
1229:   PetscCheck(!A->transupdated || matstructT, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing matTranspose struct");
1230:   if (A->transupdated) PetscFunctionReturn(PETSC_SUCCESS);
1231:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));
1232:   PetscCall(PetscLogGpuTimeBegin());
1233:   if (cusparsestruct->format != MAT_CUSPARSE_CSR) PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_TRUE));
1234:   if (!cusparsestruct->matTranspose) { /* create cusparse matrix */
1235:     matstructT = new Mat_SeqAIJCUSPARSEMultStruct;
1236:     PetscCallCUSPARSE(cusparseCreateMatDescr(&matstructT->descr));
1237:     indexBase = cusparseGetMatIndexBase(matstruct->descr);
1238:     PetscCallCUSPARSE(cusparseSetMatIndexBase(matstructT->descr, indexBase));
1239:     PetscCallCUSPARSE(cusparseSetMatType(matstructT->descr, CUSPARSE_MATRIX_TYPE_GENERAL));

1241:     /* set alpha and beta */
1242:     PetscCallCUDA(cudaMalloc((void **)&matstructT->alpha_one, sizeof(PetscScalar)));
1243:     PetscCallCUDA(cudaMalloc((void **)&matstructT->beta_zero, sizeof(PetscScalar)));
1244:     PetscCallCUDA(cudaMalloc((void **)&matstructT->beta_one, sizeof(PetscScalar)));
1245:     PetscCallCUDA(cudaMemcpy(matstructT->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
1246:     PetscCallCUDA(cudaMemcpy(matstructT->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
1247:     PetscCallCUDA(cudaMemcpy(matstructT->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));

1249:     if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
1250:       CsrMatrix *matrixT      = new CsrMatrix;
1251:       matstructT->mat         = matrixT;
1252:       matrixT->num_rows       = A->cmap->n;
1253:       matrixT->num_cols       = A->rmap->n;
1254:       matrixT->num_entries    = a->nz;
1255:       matrixT->row_offsets    = new THRUSTINTARRAY32(matrixT->num_rows + 1);
1256:       matrixT->column_indices = new THRUSTINTARRAY32(a->nz);
1257:       matrixT->values         = new THRUSTARRAY(a->nz);

1259:       if (!cusparsestruct->rowoffsets_gpu) cusparsestruct->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
1260:       cusparsestruct->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);

1262: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1263:   #if PETSC_PKG_CUDA_VERSION_GE(11, 2, 1)
1264:       stat = cusparseCreateCsr(&matstructT->matDescr, matrixT->num_rows, matrixT->num_cols, matrixT->num_entries, matrixT->row_offsets->data().get(), matrixT->column_indices->data().get(), matrixT->values->data().get(), CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, /* row offset, col idx type due to THRUSTINTARRAY32 */
1265:                                indexBase, cusparse_scalartype);
1266:       PetscCallCUSPARSE(stat);
1267:   #else
1268:       /* cusparse-11.x returns errors with zero-sized matrices until 11.2.1,
1269:            see https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#cusparse-11.2.1

1271:            I don't know what a proper value should be for matstructT->matDescr with empty matrices, so I just set
1272:            it to NULL to blow it up if one relies on it. Per https://docs.nvidia.com/cuda/cusparse/index.html#csr2cscEx2,
1273:            when nnz = 0, matrixT->row_offsets[] should be filled with indexBase. So I also set it accordingly.
1274:         */
1275:       if (matrixT->num_entries) {
1276:         stat = cusparseCreateCsr(&matstructT->matDescr, matrixT->num_rows, matrixT->num_cols, matrixT->num_entries, matrixT->row_offsets->data().get(), matrixT->column_indices->data().get(), matrixT->values->data().get(), CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, indexBase, cusparse_scalartype);
1277:         PetscCallCUSPARSE(stat);

1279:       } else {
1280:         matstructT->matDescr = NULL;
1281:         matrixT->row_offsets->assign(matrixT->row_offsets->size(), indexBase);
1282:       }
1283:   #endif
1284: #endif
1285:     } else if (cusparsestruct->format == MAT_CUSPARSE_ELL || cusparsestruct->format == MAT_CUSPARSE_HYB) {
1286: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1287:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
1288: #else
1289:       CsrMatrix *temp  = new CsrMatrix;
1290:       CsrMatrix *tempT = new CsrMatrix;
1291:       /* First convert HYB to CSR */
1292:       temp->num_rows       = A->rmap->n;
1293:       temp->num_cols       = A->cmap->n;
1294:       temp->num_entries    = a->nz;
1295:       temp->row_offsets    = new THRUSTINTARRAY32(A->rmap->n + 1);
1296:       temp->column_indices = new THRUSTINTARRAY32(a->nz);
1297:       temp->values         = new THRUSTARRAY(a->nz);

1299:       stat = cusparse_hyb2csr(cusparsestruct->handle, matstruct->descr, (cusparseHybMat_t)matstruct->mat, temp->values->data().get(), temp->row_offsets->data().get(), temp->column_indices->data().get());
1300:       PetscCallCUSPARSE(stat);

1302:       /* Next, convert CSR to CSC (i.e. the matrix transpose) */
1303:       tempT->num_rows       = A->rmap->n;
1304:       tempT->num_cols       = A->cmap->n;
1305:       tempT->num_entries    = a->nz;
1306:       tempT->row_offsets    = new THRUSTINTARRAY32(A->rmap->n + 1);
1307:       tempT->column_indices = new THRUSTINTARRAY32(a->nz);
1308:       tempT->values         = new THRUSTARRAY(a->nz);

1310:       stat = cusparse_csr2csc(cusparsestruct->handle, temp->num_rows, temp->num_cols, temp->num_entries, temp->values->data().get(), temp->row_offsets->data().get(), temp->column_indices->data().get(), tempT->values->data().get(),
1311:                               tempT->column_indices->data().get(), tempT->row_offsets->data().get(), CUSPARSE_ACTION_NUMERIC, indexBase);
1312:       PetscCallCUSPARSE(stat);

1314:       /* Last, convert CSC to HYB */
1315:       cusparseHybMat_t hybMat;
1316:       PetscCallCUSPARSE(cusparseCreateHybMat(&hybMat));
1317:       cusparseHybPartition_t partition = cusparsestruct->format == MAT_CUSPARSE_ELL ? CUSPARSE_HYB_PARTITION_MAX : CUSPARSE_HYB_PARTITION_AUTO;
1318:       stat                             = cusparse_csr2hyb(cusparsestruct->handle, A->rmap->n, A->cmap->n, matstructT->descr, tempT->values->data().get(), tempT->row_offsets->data().get(), tempT->column_indices->data().get(), hybMat, 0, partition);
1319:       PetscCallCUSPARSE(stat);

1321:       /* assign the pointer */
1322:       matstructT->mat = hybMat;
1323:       A->transupdated = PETSC_TRUE;
1324:       /* delete temporaries */
1325:       if (tempT) {
1326:         if (tempT->values) delete (THRUSTARRAY *)tempT->values;
1327:         if (tempT->column_indices) delete (THRUSTINTARRAY32 *)tempT->column_indices;
1328:         if (tempT->row_offsets) delete (THRUSTINTARRAY32 *)tempT->row_offsets;
1329:         delete (CsrMatrix *)tempT;
1330:       }
1331:       if (temp) {
1332:         if (temp->values) delete (THRUSTARRAY *)temp->values;
1333:         if (temp->column_indices) delete (THRUSTINTARRAY32 *)temp->column_indices;
1334:         if (temp->row_offsets) delete (THRUSTINTARRAY32 *)temp->row_offsets;
1335:         delete (CsrMatrix *)temp;
1336:       }
1337: #endif
1338:     }
1339:   }
1340:   if (cusparsestruct->format == MAT_CUSPARSE_CSR) { /* transpose mat struct may be already present, update data */
1341:     CsrMatrix *matrix  = (CsrMatrix *)matstruct->mat;
1342:     CsrMatrix *matrixT = (CsrMatrix *)matstructT->mat;
1343:     PetscCheck(matrix, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrix");
1344:     PetscCheck(matrix->row_offsets, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrix rows");
1345:     PetscCheck(matrix->column_indices, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrix cols");
1346:     PetscCheck(matrix->values, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrix values");
1347:     PetscCheck(matrixT, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrixT");
1348:     PetscCheck(matrixT->row_offsets, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrixT rows");
1349:     PetscCheck(matrixT->column_indices, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrixT cols");
1350:     PetscCheck(matrixT->values, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrixT values");
1351:     if (!cusparsestruct->rowoffsets_gpu) { /* this may be absent when we did not construct the transpose with csr2csc */
1352:       cusparsestruct->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
1353:       cusparsestruct->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);
1354:       PetscCall(PetscLogCpuToGpu((A->rmap->n + 1) * sizeof(PetscInt)));
1355:     }
1356:     if (!cusparsestruct->csr2csc_i) {
1357:       THRUSTARRAY csr2csc_a(matrix->num_entries);
1358:       PetscCallThrust(thrust::sequence(thrust::device, csr2csc_a.begin(), csr2csc_a.end(), 0.0));

1360:       indexBase = cusparseGetMatIndexBase(matstruct->descr);
1361: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1362:       void  *csr2cscBuffer;
1363:       size_t csr2cscBufferSize;
1364:       stat = cusparseCsr2cscEx2_bufferSize(cusparsestruct->handle, A->rmap->n, A->cmap->n, matrix->num_entries, matrix->values->data().get(), cusparsestruct->rowoffsets_gpu->data().get(), matrix->column_indices->data().get(), matrixT->values->data().get(),
1365:                                            matrixT->row_offsets->data().get(), matrixT->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, cusparsestruct->csr2cscAlg, &csr2cscBufferSize);
1366:       PetscCallCUSPARSE(stat);
1367:       PetscCallCUDA(cudaMalloc(&csr2cscBuffer, csr2cscBufferSize));
1368: #endif

1370:       if (matrix->num_entries) {
1371:         /* When there are no nonzeros, this routine mistakenly returns CUSPARSE_STATUS_INVALID_VALUE in
1372:            mat_tests-ex62_15_mpiaijcusparse on ranks 0 and 2 with CUDA-11. But CUDA-10 is OK.
1373:            I checked every parameters and they were just fine. I have no clue why cusparse complains.

1375:            Per https://docs.nvidia.com/cuda/cusparse/index.html#csr2cscEx2, when nnz = 0, matrixT->row_offsets[]
1376:            should be filled with indexBase. So I just take a shortcut here.
1377:         */
1378:         stat = cusparse_csr2csc(cusparsestruct->handle, A->rmap->n, A->cmap->n, matrix->num_entries, csr2csc_a.data().get(), cusparsestruct->rowoffsets_gpu->data().get(), matrix->column_indices->data().get(), matrixT->values->data().get(),
1379: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1380:                                 matrixT->row_offsets->data().get(), matrixT->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, cusparsestruct->csr2cscAlg, csr2cscBuffer);
1381:         PetscCallCUSPARSE(stat);
1382: #else
1383:                                 matrixT->column_indices->data().get(), matrixT->row_offsets->data().get(), CUSPARSE_ACTION_NUMERIC, indexBase);
1384:         PetscCallCUSPARSE(stat);
1385: #endif
1386:       } else {
1387:         matrixT->row_offsets->assign(matrixT->row_offsets->size(), indexBase);
1388:       }

1390:       cusparsestruct->csr2csc_i = new THRUSTINTARRAY(matrix->num_entries);
1391:       PetscCallThrust(thrust::transform(thrust::device, matrixT->values->begin(), matrixT->values->end(), cusparsestruct->csr2csc_i->begin(), PetscScalarToPetscInt()));
1392: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1393:       PetscCallCUDA(cudaFree(csr2cscBuffer));
1394: #endif
1395:     }
1396:     PetscCallThrust(
1397:       thrust::copy(thrust::device, thrust::make_permutation_iterator(matrix->values->begin(), cusparsestruct->csr2csc_i->begin()), thrust::make_permutation_iterator(matrix->values->begin(), cusparsestruct->csr2csc_i->end()), matrixT->values->begin()));
1398:   }
1399:   PetscCall(PetscLogGpuTimeEnd());
1400:   PetscCall(PetscLogEventEnd(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));
1401:   /* the compressed row indices is not used for matTranspose */
1402:   matstructT->cprowIndices = NULL;
1403:   /* assign the pointer */
1404:   ((Mat_SeqAIJCUSPARSE *)A->spptr)->matTranspose = matstructT;
1405:   A->transupdated                                = PETSC_TRUE;
1406:   PetscFunctionReturn(PETSC_SUCCESS);
1407: }

1409: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
1410: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_LU(Mat A, Vec b, Vec x)
1411: {
1412:   const PetscScalar                    *barray;
1413:   PetscScalar                          *xarray;
1414:   thrust::device_ptr<const PetscScalar> bGPU;
1415:   thrust::device_ptr<PetscScalar>       xGPU;
1416:   Mat_SeqAIJCUSPARSETriFactors         *fs  = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
1417:   const Mat_SeqAIJ                     *aij = static_cast<Mat_SeqAIJ *>(A->data);
1418:   const cusparseOperation_t             op  = CUSPARSE_OPERATION_NON_TRANSPOSE;
1419:   const cusparseSpSVAlg_t               alg = CUSPARSE_SPSV_ALG_DEFAULT;
1420:   PetscInt                              m   = A->rmap->n;

1422:   PetscFunctionBegin;
1423:   PetscCall(PetscLogGpuTimeBegin());
1424:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
1425:   PetscCall(VecCUDAGetArrayRead(b, &barray));
1426:   xGPU = thrust::device_pointer_cast(xarray);
1427:   bGPU = thrust::device_pointer_cast(barray);

1429:   // Reorder b with the row permutation if needed, and wrap the result in fs->X
1430:   if (fs->rpermIndices) {
1431:     PetscCallThrust(thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(bGPU, fs->rpermIndices->begin()), thrust::make_permutation_iterator(bGPU, fs->rpermIndices->end()), thrust::device_pointer_cast(fs->X)));
1432:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
1433:   } else {
1434:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
1435:   }

1437:   // Solve L Y = X
1438:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
1439:   // Note that cusparseSpSV_solve() secretly uses the external buffer used in cusparseSpSV_analysis()!
1440:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, op, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_L));

1442:   // Solve U X = Y
1443:   if (fs->cpermIndices) {
1444:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
1445:   } else {
1446:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
1447:   }
1448:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, op, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_Y, fs->dnVecDescr_X, cusparse_scalartype, alg, fs->spsvDescr_U));

1450:   // Reorder X with the column permutation if needed, and put the result back to x
1451:   if (fs->cpermIndices) {
1452:     PetscCallThrust(thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X), fs->cpermIndices->begin()),
1453:                                  thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X + m), fs->cpermIndices->end()), xGPU));
1454:   }
1455:   PetscCall(VecCUDARestoreArrayRead(b, &barray));
1456:   PetscCall(VecCUDARestoreArrayWrite(x, &xarray));
1457:   PetscCall(PetscLogGpuTimeEnd());
1458:   PetscCall(PetscLogGpuFlops(2.0 * aij->nz - m));
1459:   PetscFunctionReturn(PETSC_SUCCESS);
1460: }

1462: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE_LU(Mat A, Vec b, Vec x)
1463: {
1464:   Mat_SeqAIJCUSPARSETriFactors         *fs  = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
1465:   Mat_SeqAIJ                           *aij = static_cast<Mat_SeqAIJ *>(A->data);
1466:   const PetscScalar                    *barray;
1467:   PetscScalar                          *xarray;
1468:   thrust::device_ptr<const PetscScalar> bGPU;
1469:   thrust::device_ptr<PetscScalar>       xGPU;
1470:   const cusparseOperation_t             opA = CUSPARSE_OPERATION_TRANSPOSE;
1471:   const cusparseSpSVAlg_t               alg = CUSPARSE_SPSV_ALG_DEFAULT;
1472:   PetscInt                              m   = A->rmap->n;

1474:   PetscFunctionBegin;
1475:   PetscCall(PetscLogGpuTimeBegin());
1476:   if (!fs->createdTransposeSpSVDescr) { // Call MatSolveTranspose() for the first time
1477:     PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Lt));
1478:     PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, /* The matrix is still L. We only do transpose solve with it */
1479:                                               fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_Lt, &fs->spsvBufferSize_Lt));

1481:     PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Ut));
1482:     PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_Ut, &fs->spsvBufferSize_Ut));
1483:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Lt, fs->spsvBufferSize_Lt));
1484:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Ut, fs->spsvBufferSize_Ut));
1485:     fs->createdTransposeSpSVDescr = PETSC_TRUE;
1486:   }

1488:   if (!fs->updatedTransposeSpSVAnalysis) {
1489:     PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_Lt, fs->spsvBuffer_Lt));

1491:     PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_Ut, fs->spsvBuffer_Ut));
1492:     fs->updatedTransposeSpSVAnalysis = PETSC_TRUE;
1493:   }

1495:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
1496:   PetscCall(VecCUDAGetArrayRead(b, &barray));
1497:   xGPU = thrust::device_pointer_cast(xarray);
1498:   bGPU = thrust::device_pointer_cast(barray);

1500:   // Reorder b with the row permutation if needed, and wrap the result in fs->X
1501:   if (fs->rpermIndices) {
1502:     PetscCallThrust(thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(bGPU, fs->rpermIndices->begin()), thrust::make_permutation_iterator(bGPU, fs->rpermIndices->end()), thrust::device_pointer_cast(fs->X)));
1503:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
1504:   } else {
1505:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
1506:   }

1508:   // Solve Ut Y = X
1509:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
1510:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_Ut));

1512:   // Solve Lt X = Y
1513:   if (fs->cpermIndices) { // if need to permute, we need to use the intermediate buffer X
1514:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
1515:   } else {
1516:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
1517:   }
1518:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_Y, fs->dnVecDescr_X, cusparse_scalartype, alg, fs->spsvDescr_Lt));

1520:   // Reorder X with the column permutation if needed, and put the result back to x
1521:   if (fs->cpermIndices) {
1522:     PetscCallThrust(thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X), fs->cpermIndices->begin()),
1523:                                  thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X + m), fs->cpermIndices->end()), xGPU));
1524:   }

1526:   PetscCall(VecCUDARestoreArrayRead(b, &barray));
1527:   PetscCall(VecCUDARestoreArrayWrite(x, &xarray));
1528:   PetscCall(PetscLogGpuTimeEnd());
1529:   PetscCall(PetscLogGpuFlops(2.0 * aij->nz - A->rmap->n));
1530:   PetscFunctionReturn(PETSC_SUCCESS);
1531: }
1532: #else
1533: /* Why do we need to analyze the transposed matrix again? Can't we just use op(A) = CUSPARSE_OPERATION_TRANSPOSE in MatSolve_SeqAIJCUSPARSE? */
1534: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE(Mat A, Vec bb, Vec xx)
1535: {
1536:   PetscInt                              n = xx->map->n;
1537:   const PetscScalar                    *barray;
1538:   PetscScalar                          *xarray;
1539:   thrust::device_ptr<const PetscScalar> bGPU;
1540:   thrust::device_ptr<PetscScalar>       xGPU;
1541:   Mat_SeqAIJCUSPARSETriFactors         *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1542:   Mat_SeqAIJCUSPARSETriFactorStruct    *loTriFactorT       = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtrTranspose;
1543:   Mat_SeqAIJCUSPARSETriFactorStruct    *upTriFactorT       = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtrTranspose;
1544:   THRUSTARRAY                          *tempGPU            = (THRUSTARRAY *)cusparseTriFactors->workVector;

1546:   PetscFunctionBegin;
1547:   /* Analyze the matrix and create the transpose ... on the fly */
1548:   if (!loTriFactorT && !upTriFactorT) {
1549:     PetscCall(MatSeqAIJCUSPARSEAnalyzeTransposeForSolve(A));
1550:     loTriFactorT = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtrTranspose;
1551:     upTriFactorT = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtrTranspose;
1552:   }

1554:   /* Get the GPU pointers */
1555:   PetscCall(VecCUDAGetArrayWrite(xx, &xarray));
1556:   PetscCall(VecCUDAGetArrayRead(bb, &barray));
1557:   xGPU = thrust::device_pointer_cast(xarray);
1558:   bGPU = thrust::device_pointer_cast(barray);

1560:   PetscCall(PetscLogGpuTimeBegin());
1561:   /* First, reorder with the row permutation */
1562:   thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(bGPU, cusparseTriFactors->rpermIndices->begin()), thrust::make_permutation_iterator(bGPU + n, cusparseTriFactors->rpermIndices->end()), xGPU);

1564:   /* First, solve U */
1565:   PetscCallCUSPARSE(cusparseXcsrsv_solve(cusparseTriFactors->handle, upTriFactorT->solveOp, upTriFactorT->csrMat->num_rows, upTriFactorT->csrMat->num_entries, &PETSC_CUSPARSE_ONE, upTriFactorT->descr, upTriFactorT->csrMat->values->data().get(),
1566:                                          upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->solveInfo, xarray, tempGPU->data().get(), upTriFactorT->solvePolicy, upTriFactorT->solveBuffer));

1568:   /* Then, solve L */
1569:   PetscCallCUSPARSE(cusparseXcsrsv_solve(cusparseTriFactors->handle, loTriFactorT->solveOp, loTriFactorT->csrMat->num_rows, loTriFactorT->csrMat->num_entries, &PETSC_CUSPARSE_ONE, loTriFactorT->descr, loTriFactorT->csrMat->values->data().get(),
1570:                                          loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->solveInfo, tempGPU->data().get(), xarray, loTriFactorT->solvePolicy, loTriFactorT->solveBuffer));

1572:   /* Last, copy the solution, xGPU, into a temporary with the column permutation ... can't be done in place. */
1573:   thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(xGPU, cusparseTriFactors->cpermIndices->begin()), thrust::make_permutation_iterator(xGPU + n, cusparseTriFactors->cpermIndices->end()), tempGPU->begin());

1575:   /* Copy the temporary to the full solution. */
1576:   thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), tempGPU->begin(), tempGPU->end(), xGPU);

1578:   /* restore */
1579:   PetscCall(VecCUDARestoreArrayRead(bb, &barray));
1580:   PetscCall(VecCUDARestoreArrayWrite(xx, &xarray));
1581:   PetscCall(PetscLogGpuTimeEnd());
1582:   PetscCall(PetscLogGpuFlops(2.0 * cusparseTriFactors->nnz - A->cmap->n));
1583:   PetscFunctionReturn(PETSC_SUCCESS);
1584: }

1586: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering(Mat A, Vec bb, Vec xx)
1587: {
1588:   const PetscScalar                 *barray;
1589:   PetscScalar                       *xarray;
1590:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1591:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorT       = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtrTranspose;
1592:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorT       = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtrTranspose;
1593:   THRUSTARRAY                       *tempGPU            = (THRUSTARRAY *)cusparseTriFactors->workVector;

1595:   PetscFunctionBegin;
1596:   /* Analyze the matrix and create the transpose ... on the fly */
1597:   if (!loTriFactorT && !upTriFactorT) {
1598:     PetscCall(MatSeqAIJCUSPARSEAnalyzeTransposeForSolve(A));
1599:     loTriFactorT = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtrTranspose;
1600:     upTriFactorT = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtrTranspose;
1601:   }

1603:   /* Get the GPU pointers */
1604:   PetscCall(VecCUDAGetArrayWrite(xx, &xarray));
1605:   PetscCall(VecCUDAGetArrayRead(bb, &barray));

1607:   PetscCall(PetscLogGpuTimeBegin());
1608:   /* First, solve U */
1609:   PetscCallCUSPARSE(cusparseXcsrsv_solve(cusparseTriFactors->handle, upTriFactorT->solveOp, upTriFactorT->csrMat->num_rows, upTriFactorT->csrMat->num_entries, &PETSC_CUSPARSE_ONE, upTriFactorT->descr, upTriFactorT->csrMat->values->data().get(),
1610:                                          upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->solveInfo, barray, tempGPU->data().get(), upTriFactorT->solvePolicy, upTriFactorT->solveBuffer));

1612:   /* Then, solve L */
1613:   PetscCallCUSPARSE(cusparseXcsrsv_solve(cusparseTriFactors->handle, loTriFactorT->solveOp, loTriFactorT->csrMat->num_rows, loTriFactorT->csrMat->num_entries, &PETSC_CUSPARSE_ONE, loTriFactorT->descr, loTriFactorT->csrMat->values->data().get(),
1614:                                          loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->solveInfo, tempGPU->data().get(), xarray, loTriFactorT->solvePolicy, loTriFactorT->solveBuffer));

1616:   /* restore */
1617:   PetscCall(VecCUDARestoreArrayRead(bb, &barray));
1618:   PetscCall(VecCUDARestoreArrayWrite(xx, &xarray));
1619:   PetscCall(PetscLogGpuTimeEnd());
1620:   PetscCall(PetscLogGpuFlops(2.0 * cusparseTriFactors->nnz - A->cmap->n));
1621:   PetscFunctionReturn(PETSC_SUCCESS);
1622: }

1624: static PetscErrorCode MatSolve_SeqAIJCUSPARSE(Mat A, Vec bb, Vec xx)
1625: {
1626:   const PetscScalar                    *barray;
1627:   PetscScalar                          *xarray;
1628:   thrust::device_ptr<const PetscScalar> bGPU;
1629:   thrust::device_ptr<PetscScalar>       xGPU;
1630:   Mat_SeqAIJCUSPARSETriFactors         *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1631:   Mat_SeqAIJCUSPARSETriFactorStruct    *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
1632:   Mat_SeqAIJCUSPARSETriFactorStruct    *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
1633:   THRUSTARRAY                          *tempGPU            = (THRUSTARRAY *)cusparseTriFactors->workVector;

1635:   PetscFunctionBegin;
1636:   /* Get the GPU pointers */
1637:   PetscCall(VecCUDAGetArrayWrite(xx, &xarray));
1638:   PetscCall(VecCUDAGetArrayRead(bb, &barray));
1639:   xGPU = thrust::device_pointer_cast(xarray);
1640:   bGPU = thrust::device_pointer_cast(barray);

1642:   PetscCall(PetscLogGpuTimeBegin());
1643:   /* First, reorder with the row permutation */
1644:   thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(bGPU, cusparseTriFactors->rpermIndices->begin()), thrust::make_permutation_iterator(bGPU, cusparseTriFactors->rpermIndices->end()), tempGPU->begin());

1646:   /* Next, solve L */
1647:   PetscCallCUSPARSE(cusparseXcsrsv_solve(cusparseTriFactors->handle, loTriFactor->solveOp, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_entries, &PETSC_CUSPARSE_ONE, loTriFactor->descr, loTriFactor->csrMat->values->data().get(),
1648:                                          loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, tempGPU->data().get(), xarray, loTriFactor->solvePolicy, loTriFactor->solveBuffer));

1650:   /* Then, solve U */
1651:   PetscCallCUSPARSE(cusparseXcsrsv_solve(cusparseTriFactors->handle, upTriFactor->solveOp, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_entries, &PETSC_CUSPARSE_ONE, upTriFactor->descr, upTriFactor->csrMat->values->data().get(),
1652:                                          upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, xarray, tempGPU->data().get(), upTriFactor->solvePolicy, upTriFactor->solveBuffer));

1654:   /* Last, reorder with the column permutation */
1655:   thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(tempGPU->begin(), cusparseTriFactors->cpermIndices->begin()), thrust::make_permutation_iterator(tempGPU->begin(), cusparseTriFactors->cpermIndices->end()), xGPU);

1657:   PetscCall(VecCUDARestoreArrayRead(bb, &barray));
1658:   PetscCall(VecCUDARestoreArrayWrite(xx, &xarray));
1659:   PetscCall(PetscLogGpuTimeEnd());
1660:   PetscCall(PetscLogGpuFlops(2.0 * cusparseTriFactors->nnz - A->cmap->n));
1661:   PetscFunctionReturn(PETSC_SUCCESS);
1662: }

1664: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_NaturalOrdering(Mat A, Vec bb, Vec xx)
1665: {
1666:   const PetscScalar                 *barray;
1667:   PetscScalar                       *xarray;
1668:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1669:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
1670:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
1671:   THRUSTARRAY                       *tempGPU            = (THRUSTARRAY *)cusparseTriFactors->workVector;

1673:   PetscFunctionBegin;
1674:   /* Get the GPU pointers */
1675:   PetscCall(VecCUDAGetArrayWrite(xx, &xarray));
1676:   PetscCall(VecCUDAGetArrayRead(bb, &barray));

1678:   PetscCall(PetscLogGpuTimeBegin());
1679:   /* First, solve L */
1680:   PetscCallCUSPARSE(cusparseXcsrsv_solve(cusparseTriFactors->handle, loTriFactor->solveOp, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_entries, &PETSC_CUSPARSE_ONE, loTriFactor->descr, loTriFactor->csrMat->values->data().get(),
1681:                                          loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, barray, tempGPU->data().get(), loTriFactor->solvePolicy, loTriFactor->solveBuffer));

1683:   /* Next, solve U */
1684:   PetscCallCUSPARSE(cusparseXcsrsv_solve(cusparseTriFactors->handle, upTriFactor->solveOp, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_entries, &PETSC_CUSPARSE_ONE, upTriFactor->descr, upTriFactor->csrMat->values->data().get(),
1685:                                          upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, tempGPU->data().get(), xarray, upTriFactor->solvePolicy, upTriFactor->solveBuffer));

1687:   PetscCall(VecCUDARestoreArrayRead(bb, &barray));
1688:   PetscCall(VecCUDARestoreArrayWrite(xx, &xarray));
1689:   PetscCall(PetscLogGpuTimeEnd());
1690:   PetscCall(PetscLogGpuFlops(2.0 * cusparseTriFactors->nnz - A->cmap->n));
1691:   PetscFunctionReturn(PETSC_SUCCESS);
1692: }
1693: #endif

1695: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
1696: static PetscErrorCode MatILUFactorNumeric_SeqAIJCUSPARSE_ILU0(Mat fact, Mat A, const MatFactorInfo *)
1697: {
1698:   Mat_SeqAIJCUSPARSETriFactors *fs    = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1699:   Mat_SeqAIJ                   *aij   = (Mat_SeqAIJ *)fact->data;
1700:   Mat_SeqAIJCUSPARSE           *Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
1701:   CsrMatrix                    *Acsr;
1702:   PetscInt                      m, nz;
1703:   PetscBool                     flg;

1705:   PetscFunctionBegin;
1706:   if (PetscDefined(USE_DEBUG)) {
1707:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
1708:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
1709:   }

1711:   /* Copy A's value to fact */
1712:   m  = fact->rmap->n;
1713:   nz = aij->nz;
1714:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
1715:   Acsr = (CsrMatrix *)Acusp->mat->mat;
1716:   PetscCallCUDA(cudaMemcpyAsync(fs->csrVal, Acsr->values->data().get(), sizeof(PetscScalar) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

1718:   PetscCall(PetscLogGpuTimeBegin());
1719:   /* Factorize fact inplace */
1720:   if (m)
1721:     PetscCallCUSPARSE(cusparseXcsrilu02(fs->handle, m, nz, /* cusparseXcsrilu02 errors out with empty matrices (m=0) */
1722:                                         fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ilu0Info_M, fs->policy_M, fs->factBuffer_M));
1723:   if (PetscDefined(USE_DEBUG)) {
1724:     int              numerical_zero;
1725:     cusparseStatus_t status;
1726:     status = cusparseXcsrilu02_zeroPivot(fs->handle, fs->ilu0Info_M, &numerical_zero);
1727:     PetscAssert(CUSPARSE_STATUS_ZERO_PIVOT != status, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "Numerical zero pivot detected in csrilu02: A(%d,%d) is zero", numerical_zero, numerical_zero);
1728:   }

1730:   /* cusparseSpSV_analysis() is numeric, i.e., it requires valid matrix values, therefore, we do it after cusparseXcsrilu02()
1731:      See discussion at https://github.com/NVIDIA/CUDALibrarySamples/issues/78
1732:   */
1733:   PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L, fs->spsvBuffer_L));

1735:   PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_U, fs->spsvBuffer_U));

1737:   /* L, U values have changed, reset the flag to indicate we need to redo cusparseSpSV_analysis() for transpose solve */
1738:   fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;

1740:   fact->offloadmask            = PETSC_OFFLOAD_GPU;
1741:   fact->ops->solve             = MatSolve_SeqAIJCUSPARSE_LU; // spMatDescr_L/U uses 32-bit indices, but cusparseSpSV_solve() supports both 32 and 64. The info is encoded in cusparseSpMatDescr_t.
1742:   fact->ops->solvetranspose    = MatSolveTranspose_SeqAIJCUSPARSE_LU;
1743:   fact->ops->matsolve          = NULL;
1744:   fact->ops->matsolvetranspose = NULL;
1745:   PetscCall(PetscLogGpuTimeEnd());
1746:   PetscCall(PetscLogGpuFlops(fs->numericFactFlops));
1747:   PetscFunctionReturn(PETSC_SUCCESS);
1748: }

1750: static PetscErrorCode MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0(Mat fact, Mat A, IS, IS, const MatFactorInfo *info)
1751: {
1752:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1753:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
1754:   PetscInt                      m, nz;

1756:   PetscFunctionBegin;
1757:   if (PetscDefined(USE_DEBUG)) {
1758:     PetscInt  i;
1759:     PetscBool flg, missing;

1761:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
1762:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
1763:     PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Must be square matrix, rows %" PetscInt_FMT " columns %" PetscInt_FMT, A->rmap->n, A->cmap->n);
1764:     PetscCall(MatMissingDiagonal(A, &missing, &i));
1765:     PetscCheck(!missing, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry %" PetscInt_FMT, i);
1766:   }

1768:   /* Free the old stale stuff */
1769:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&fs));

1771:   /* Copy over A's meta data to fact. Note that we also allocated fact's i,j,a on host,
1772:      but they will not be used. Allocate them just for easy debugging.
1773:    */
1774:   PetscCall(MatDuplicateNoCreate_SeqAIJ(fact, A, MAT_DO_NOT_COPY_VALUES, PETSC_TRUE /*malloc*/));

1776:   fact->offloadmask            = PETSC_OFFLOAD_BOTH;
1777:   fact->factortype             = MAT_FACTOR_ILU;
1778:   fact->info.factor_mallocs    = 0;
1779:   fact->info.fill_ratio_given  = info->fill;
1780:   fact->info.fill_ratio_needed = 1.0;

1782:   aij->row = NULL;
1783:   aij->col = NULL;

1785:   /* ====================================================================== */
1786:   /* Copy A's i, j to fact and also allocate the value array of fact.       */
1787:   /* We'll do in-place factorization on fact                                */
1788:   /* ====================================================================== */
1789:   const int *Ai, *Aj;

1791:   m  = fact->rmap->n;
1792:   nz = aij->nz;

1794:   PetscCallCUDA(cudaMalloc((void **)&fs->csrRowPtr32, sizeof(*fs->csrRowPtr32) * (m + 1)));
1795:   PetscCallCUDA(cudaMalloc((void **)&fs->csrColIdx32, sizeof(*fs->csrColIdx32) * nz));
1796:   PetscCallCUDA(cudaMalloc((void **)&fs->csrVal, sizeof(*fs->csrVal) * nz));
1797:   PetscCall(MatSeqAIJCUSPARSEGetIJ(A, PETSC_FALSE, &Ai, &Aj)); /* Do not use compressed Ai.  The returned Ai, Aj are 32-bit */
1798:   PetscCallCUDA(cudaMemcpyAsync(fs->csrRowPtr32, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));
1799:   PetscCallCUDA(cudaMemcpyAsync(fs->csrColIdx32, Aj, sizeof(*Aj) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

1801:   /* ====================================================================== */
1802:   /* Create descriptors for M, L, U                                         */
1803:   /* ====================================================================== */
1804:   cusparseFillMode_t fillMode;
1805:   cusparseDiagType_t diagType;

1807:   PetscCallCUSPARSE(cusparseCreateMatDescr(&fs->matDescr_M));
1808:   PetscCallCUSPARSE(cusparseSetMatIndexBase(fs->matDescr_M, CUSPARSE_INDEX_BASE_ZERO));
1809:   PetscCallCUSPARSE(cusparseSetMatType(fs->matDescr_M, CUSPARSE_MATRIX_TYPE_GENERAL));

1811:   /* https://docs.nvidia.com/cuda/cusparse/index.html#cusparseDiagType_t
1812:     cusparseDiagType_t: This type indicates if the matrix diagonal entries are unity. The diagonal elements are always
1813:     assumed to be present, but if CUSPARSE_DIAG_TYPE_UNIT is passed to an API routine, then the routine assumes that
1814:     all diagonal entries are unity and will not read or modify those entries. Note that in this case the routine
1815:     assumes the diagonal entries are equal to one, regardless of what those entries are actually set to in memory.
1816:   */
1817:   fillMode = CUSPARSE_FILL_MODE_LOWER;
1818:   diagType = CUSPARSE_DIAG_TYPE_UNIT;
1819:   PetscCallCUSPARSE(cusparseCreateCsr(&fs->spMatDescr_L, m, m, nz, fs->csrRowPtr32, fs->csrColIdx32, fs->csrVal, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype));
1820:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
1821:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

1823:   fillMode = CUSPARSE_FILL_MODE_UPPER;
1824:   diagType = CUSPARSE_DIAG_TYPE_NON_UNIT;
1825:   PetscCallCUSPARSE(cusparseCreateCsr(&fs->spMatDescr_U, m, m, nz, fs->csrRowPtr32, fs->csrColIdx32, fs->csrVal, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype));
1826:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
1827:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

1829:   /* ========================================================================= */
1830:   /* Query buffer sizes for csrilu0, SpSV and allocate buffers                 */
1831:   /* ========================================================================= */
1832:   PetscCallCUSPARSE(cusparseCreateCsrilu02Info(&fs->ilu0Info_M));
1833:   if (m)
1834:     PetscCallCUSPARSE(cusparseXcsrilu02_bufferSize(fs->handle, m, nz, /* cusparseXcsrilu02 errors out with empty matrices (m=0) */
1835:                                                    fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ilu0Info_M, &fs->factBufferSize_M));

1837:   PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(PetscScalar) * m));
1838:   PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(PetscScalar) * m));

1840:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_X, m, fs->X, cusparse_scalartype));
1841:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_Y, m, fs->Y, cusparse_scalartype));

1843:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_L));
1844:   PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L, &fs->spsvBufferSize_L));

1846:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_U));
1847:   PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_U, &fs->spsvBufferSize_U));

1849:   /* From my experiment with the example at https://github.com/NVIDIA/CUDALibrarySamples/tree/master/cuSPARSE/bicgstab,
1850:      and discussion at https://github.com/NVIDIA/CUDALibrarySamples/issues/77,
1851:      spsvBuffer_L/U can not be shared (i.e., the same) for our case, but factBuffer_M can share with either of spsvBuffer_L/U.
1852:      To save memory, we make factBuffer_M share with the bigger of spsvBuffer_L/U.
1853:    */
1854:   if (fs->spsvBufferSize_L > fs->spsvBufferSize_U) {
1855:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_L, (size_t)fs->factBufferSize_M)));
1856:     fs->spsvBuffer_L = fs->factBuffer_M;
1857:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U));
1858:   } else {
1859:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_U, (size_t)fs->factBufferSize_M)));
1860:     fs->spsvBuffer_U = fs->factBuffer_M;
1861:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L));
1862:   }

1864:   /* ========================================================================== */
1865:   /* Perform analysis of ilu0 on M, SpSv on L and U                             */
1866:   /* The lower(upper) triangular part of M has the same sparsity pattern as L(U)*/
1867:   /* ========================================================================== */
1868:   int              structural_zero;
1869:   cusparseStatus_t status;

1871:   fs->policy_M = CUSPARSE_SOLVE_POLICY_USE_LEVEL;
1872:   if (m)
1873:     PetscCallCUSPARSE(cusparseXcsrilu02_analysis(fs->handle, m, nz, /* cusparseXcsrilu02 errors out with empty matrices (m=0) */
1874:                                                  fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ilu0Info_M, fs->policy_M, fs->factBuffer_M));
1875:   if (PetscDefined(USE_DEBUG)) {
1876:     /* Function cusparseXcsrilu02_zeroPivot() is a blocking call. It calls cudaDeviceSynchronize() to make sure all previous kernels are done. */
1877:     status = cusparseXcsrilu02_zeroPivot(fs->handle, fs->ilu0Info_M, &structural_zero);
1878:     PetscCheck(CUSPARSE_STATUS_ZERO_PIVOT != status, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "Structural zero pivot detected in csrilu02: A(%d,%d) is missing", structural_zero, structural_zero);
1879:   }

1881:   /* Estimate FLOPs of the numeric factorization */
1882:   {
1883:     Mat_SeqAIJ    *Aseq = (Mat_SeqAIJ *)A->data;
1884:     PetscInt      *Ai, *Adiag, nzRow, nzLeft;
1885:     PetscLogDouble flops = 0.0;

1887:     PetscCall(MatMarkDiagonal_SeqAIJ(A));
1888:     Ai    = Aseq->i;
1889:     Adiag = Aseq->diag;
1890:     for (PetscInt i = 0; i < m; i++) {
1891:       if (Ai[i] < Adiag[i] && Adiag[i] < Ai[i + 1]) { /* There are nonzeros left to the diagonal of row i */
1892:         nzRow  = Ai[i + 1] - Ai[i];
1893:         nzLeft = Adiag[i] - Ai[i];
1894:         /* We want to eliminate nonzeros left to the diagonal one by one. Assume each time, nonzeros right
1895:           and include the eliminated one will be updated, which incurs a multiplication and an addition.
1896:         */
1897:         nzLeft = (nzRow - 1) / 2;
1898:         flops += nzLeft * (2.0 * nzRow - nzLeft + 1);
1899:       }
1900:     }
1901:     fs->numericFactFlops = flops;
1902:   }
1903:   fact->ops->lufactornumeric = MatILUFactorNumeric_SeqAIJCUSPARSE_ILU0;
1904:   PetscFunctionReturn(PETSC_SUCCESS);
1905: }

1907: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_ICC0(Mat fact, Vec b, Vec x)
1908: {
1909:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1910:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
1911:   const PetscScalar            *barray;
1912:   PetscScalar                  *xarray;

1914:   PetscFunctionBegin;
1915:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
1916:   PetscCall(VecCUDAGetArrayRead(b, &barray));
1917:   PetscCall(PetscLogGpuTimeBegin());

1919:   /* Solve L*y = b */
1920:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
1921:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
1922:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, /* L Y = X */
1923:                                        fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L));

1925:   /* Solve Lt*x = y */
1926:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
1927:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, CUSPARSE_OPERATION_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, /* Lt X = Y */
1928:                                        fs->dnVecDescr_Y, fs->dnVecDescr_X, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_Lt));

1930:   PetscCall(VecCUDARestoreArrayRead(b, &barray));
1931:   PetscCall(VecCUDARestoreArrayWrite(x, &xarray));

1933:   PetscCall(PetscLogGpuTimeEnd());
1934:   PetscCall(PetscLogGpuFlops(2.0 * aij->nz - fact->rmap->n));
1935:   PetscFunctionReturn(PETSC_SUCCESS);
1936: }

1938: static PetscErrorCode MatICCFactorNumeric_SeqAIJCUSPARSE_ICC0(Mat fact, Mat A, const MatFactorInfo *)
1939: {
1940:   Mat_SeqAIJCUSPARSETriFactors *fs    = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1941:   Mat_SeqAIJ                   *aij   = (Mat_SeqAIJ *)fact->data;
1942:   Mat_SeqAIJCUSPARSE           *Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
1943:   CsrMatrix                    *Acsr;
1944:   PetscInt                      m, nz;
1945:   PetscBool                     flg;

1947:   PetscFunctionBegin;
1948:   if (PetscDefined(USE_DEBUG)) {
1949:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
1950:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
1951:   }

1953:   /* Copy A's value to fact */
1954:   m  = fact->rmap->n;
1955:   nz = aij->nz;
1956:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
1957:   Acsr = (CsrMatrix *)Acusp->mat->mat;
1958:   PetscCallCUDA(cudaMemcpyAsync(fs->csrVal, Acsr->values->data().get(), sizeof(PetscScalar) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

1960:   /* Factorize fact inplace */
1961:   /* https://docs.nvidia.com/cuda/cusparse/index.html#csric02_solve
1962:      Function csric02() only takes the lower triangular part of matrix A to perform factorization.
1963:      The matrix type must be CUSPARSE_MATRIX_TYPE_GENERAL, the fill mode and diagonal type are ignored,
1964:      and the strictly upper triangular part is ignored and never touched. It does not matter if A is Hermitian or not.
1965:      In other words, from the point of view of csric02() A is Hermitian and only the lower triangular part is provided.
1966:    */
1967:   if (m) PetscCallCUSPARSE(cusparseXcsric02(fs->handle, m, nz, fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ic0Info_M, fs->policy_M, fs->factBuffer_M));
1968:   if (PetscDefined(USE_DEBUG)) {
1969:     int              numerical_zero;
1970:     cusparseStatus_t status;
1971:     status = cusparseXcsric02_zeroPivot(fs->handle, fs->ic0Info_M, &numerical_zero);
1972:     PetscAssert(CUSPARSE_STATUS_ZERO_PIVOT != status, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "Numerical zero pivot detected in csric02: A(%d,%d) is zero", numerical_zero, numerical_zero);
1973:   }

1975:   PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L, fs->spsvBuffer_L));

1977:   /* Note that cusparse reports this error if we use double and CUSPARSE_OPERATION_CONJUGATE_TRANSPOSE
1978:     ** On entry to cusparseSpSV_analysis(): conjugate transpose (opA) is not supported for matA data type, current -> CUDA_R_64F
1979:   */
1980:   PetscCallCUSPARSE(cusparseSpSV_analysis(fs->handle, CUSPARSE_OPERATION_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_Lt, fs->spsvBuffer_Lt));

1982:   fact->offloadmask            = PETSC_OFFLOAD_GPU;
1983:   fact->ops->solve             = MatSolve_SeqAIJCUSPARSE_ICC0;
1984:   fact->ops->solvetranspose    = MatSolve_SeqAIJCUSPARSE_ICC0;
1985:   fact->ops->matsolve          = NULL;
1986:   fact->ops->matsolvetranspose = NULL;
1987:   PetscCall(PetscLogGpuFlops(fs->numericFactFlops));
1988:   PetscFunctionReturn(PETSC_SUCCESS);
1989: }

1991: static PetscErrorCode MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0(Mat fact, Mat A, IS, const MatFactorInfo *info)
1992: {
1993:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1994:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
1995:   PetscInt                      m, nz;

1997:   PetscFunctionBegin;
1998:   if (PetscDefined(USE_DEBUG)) {
1999:     PetscInt  i;
2000:     PetscBool flg, missing;

2002:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2003:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
2004:     PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Must be square matrix, rows %" PetscInt_FMT " columns %" PetscInt_FMT, A->rmap->n, A->cmap->n);
2005:     PetscCall(MatMissingDiagonal(A, &missing, &i));
2006:     PetscCheck(!missing, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry %" PetscInt_FMT, i);
2007:   }

2009:   /* Free the old stale stuff */
2010:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&fs));

2012:   /* Copy over A's meta data to fact. Note that we also allocated fact's i,j,a on host,
2013:      but they will not be used. Allocate them just for easy debugging.
2014:    */
2015:   PetscCall(MatDuplicateNoCreate_SeqAIJ(fact, A, MAT_DO_NOT_COPY_VALUES, PETSC_TRUE /*malloc*/));

2017:   fact->offloadmask            = PETSC_OFFLOAD_BOTH;
2018:   fact->factortype             = MAT_FACTOR_ICC;
2019:   fact->info.factor_mallocs    = 0;
2020:   fact->info.fill_ratio_given  = info->fill;
2021:   fact->info.fill_ratio_needed = 1.0;

2023:   aij->row = NULL;
2024:   aij->col = NULL;

2026:   /* ====================================================================== */
2027:   /* Copy A's i, j to fact and also allocate the value array of fact.       */
2028:   /* We'll do in-place factorization on fact                                */
2029:   /* ====================================================================== */
2030:   const int *Ai, *Aj;

2032:   m  = fact->rmap->n;
2033:   nz = aij->nz;

2035:   PetscCallCUDA(cudaMalloc((void **)&fs->csrRowPtr32, sizeof(*fs->csrRowPtr32) * (m + 1)));
2036:   PetscCallCUDA(cudaMalloc((void **)&fs->csrColIdx32, sizeof(*fs->csrColIdx32) * nz));
2037:   PetscCallCUDA(cudaMalloc((void **)&fs->csrVal, sizeof(PetscScalar) * nz));
2038:   PetscCall(MatSeqAIJCUSPARSEGetIJ(A, PETSC_FALSE, &Ai, &Aj)); /* Do not use compressed Ai */
2039:   PetscCallCUDA(cudaMemcpyAsync(fs->csrRowPtr32, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));
2040:   PetscCallCUDA(cudaMemcpyAsync(fs->csrColIdx32, Aj, sizeof(*Aj) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

2042:   /* ====================================================================== */
2043:   /* Create mat descriptors for M, L                                        */
2044:   /* ====================================================================== */
2045:   cusparseFillMode_t fillMode;
2046:   cusparseDiagType_t diagType;

2048:   PetscCallCUSPARSE(cusparseCreateMatDescr(&fs->matDescr_M));
2049:   PetscCallCUSPARSE(cusparseSetMatIndexBase(fs->matDescr_M, CUSPARSE_INDEX_BASE_ZERO));
2050:   PetscCallCUSPARSE(cusparseSetMatType(fs->matDescr_M, CUSPARSE_MATRIX_TYPE_GENERAL));

2052:   /* https://docs.nvidia.com/cuda/cusparse/index.html#cusparseDiagType_t
2053:     cusparseDiagType_t: This type indicates if the matrix diagonal entries are unity. The diagonal elements are always
2054:     assumed to be present, but if CUSPARSE_DIAG_TYPE_UNIT is passed to an API routine, then the routine assumes that
2055:     all diagonal entries are unity and will not read or modify those entries. Note that in this case the routine
2056:     assumes the diagonal entries are equal to one, regardless of what those entries are actually set to in memory.
2057:   */
2058:   fillMode = CUSPARSE_FILL_MODE_LOWER;
2059:   diagType = CUSPARSE_DIAG_TYPE_NON_UNIT;
2060:   PetscCallCUSPARSE(cusparseCreateCsr(&fs->spMatDescr_L, m, m, nz, fs->csrRowPtr32, fs->csrColIdx32, fs->csrVal, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype));
2061:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
2062:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

2064:   /* ========================================================================= */
2065:   /* Query buffer sizes for csric0, SpSV of L and Lt, and allocate buffers     */
2066:   /* ========================================================================= */
2067:   PetscCallCUSPARSE(cusparseCreateCsric02Info(&fs->ic0Info_M));
2068:   if (m) PetscCallCUSPARSE(cusparseXcsric02_bufferSize(fs->handle, m, nz, fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ic0Info_M, &fs->factBufferSize_M));

2070:   PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(PetscScalar) * m));
2071:   PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(PetscScalar) * m));

2073:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_X, m, fs->X, cusparse_scalartype));
2074:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_Y, m, fs->Y, cusparse_scalartype));

2076:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_L));
2077:   PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L, &fs->spsvBufferSize_L));

2079:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Lt));
2080:   PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, CUSPARSE_OPERATION_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_Lt, &fs->spsvBufferSize_Lt));

2082:   /* To save device memory, we make the factorization buffer share with one of the solver buffer.
2083:      See also comments in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0().
2084:    */
2085:   if (fs->spsvBufferSize_L > fs->spsvBufferSize_Lt) {
2086:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_L, (size_t)fs->factBufferSize_M)));
2087:     fs->spsvBuffer_L = fs->factBuffer_M;
2088:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Lt, fs->spsvBufferSize_Lt));
2089:   } else {
2090:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_Lt, (size_t)fs->factBufferSize_M)));
2091:     fs->spsvBuffer_Lt = fs->factBuffer_M;
2092:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L));
2093:   }

2095:   /* ========================================================================== */
2096:   /* Perform analysis of ic0 on M                                               */
2097:   /* The lower triangular part of M has the same sparsity pattern as L          */
2098:   /* ========================================================================== */
2099:   int              structural_zero;
2100:   cusparseStatus_t status;

2102:   fs->policy_M = CUSPARSE_SOLVE_POLICY_USE_LEVEL;
2103:   if (m) PetscCallCUSPARSE(cusparseXcsric02_analysis(fs->handle, m, nz, fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ic0Info_M, fs->policy_M, fs->factBuffer_M));
2104:   if (PetscDefined(USE_DEBUG)) {
2105:     /* Function cusparseXcsric02_zeroPivot() is a blocking call. It calls cudaDeviceSynchronize() to make sure all previous kernels are done. */
2106:     status = cusparseXcsric02_zeroPivot(fs->handle, fs->ic0Info_M, &structural_zero);
2107:     PetscCheck(CUSPARSE_STATUS_ZERO_PIVOT != status, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "Structural zero pivot detected in csric02: A(%d,%d) is missing", structural_zero, structural_zero);
2108:   }

2110:   /* Estimate FLOPs of the numeric factorization */
2111:   {
2112:     Mat_SeqAIJ    *Aseq = (Mat_SeqAIJ *)A->data;
2113:     PetscInt      *Ai, nzRow, nzLeft;
2114:     PetscLogDouble flops = 0.0;

2116:     Ai = Aseq->i;
2117:     for (PetscInt i = 0; i < m; i++) {
2118:       nzRow = Ai[i + 1] - Ai[i];
2119:       if (nzRow > 1) {
2120:         /* We want to eliminate nonzeros left to the diagonal one by one. Assume each time, nonzeros right
2121:           and include the eliminated one will be updated, which incurs a multiplication and an addition.
2122:         */
2123:         nzLeft = (nzRow - 1) / 2;
2124:         flops += nzLeft * (2.0 * nzRow - nzLeft + 1);
2125:       }
2126:     }
2127:     fs->numericFactFlops = flops;
2128:   }
2129:   fact->ops->choleskyfactornumeric = MatICCFactorNumeric_SeqAIJCUSPARSE_ICC0;
2130:   PetscFunctionReturn(PETSC_SUCCESS);
2131: }
2132: #endif

2134: static PetscErrorCode MatLUFactorNumeric_SeqAIJCUSPARSE(Mat B, Mat A, const MatFactorInfo *info)
2135: {
2136:   // use_cpu_solve is a field in Mat_SeqAIJCUSPARSE. B, a factored matrix, uses Mat_SeqAIJCUSPARSETriFactors.
2137:   Mat_SeqAIJCUSPARSE *cusparsestruct = static_cast<Mat_SeqAIJCUSPARSE *>(A->spptr);

2139:   PetscFunctionBegin;
2140:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2141:   PetscCall(MatLUFactorNumeric_SeqAIJ(B, A, info));
2142:   B->offloadmask = PETSC_OFFLOAD_CPU;

2144:   if (!cusparsestruct->use_cpu_solve) {
2145: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2146:     B->ops->solve          = MatSolve_SeqAIJCUSPARSE_LU;
2147:     B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE_LU;
2148: #else
2149:     /* determine which version of MatSolve needs to be used. */
2150:     Mat_SeqAIJ *b     = (Mat_SeqAIJ *)B->data;
2151:     IS          isrow = b->row, iscol = b->col;
2152:     PetscBool   row_identity, col_identity;

2154:     PetscCall(ISIdentity(isrow, &row_identity));
2155:     PetscCall(ISIdentity(iscol, &col_identity));
2156:     if (row_identity && col_identity) {
2157:       B->ops->solve          = MatSolve_SeqAIJCUSPARSE_NaturalOrdering;
2158:       B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering;
2159:     } else {
2160:       B->ops->solve          = MatSolve_SeqAIJCUSPARSE;
2161:       B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE;
2162:     }
2163: #endif
2164:   }
2165:   B->ops->matsolve          = NULL;
2166:   B->ops->matsolvetranspose = NULL;

2168:   /* get the triangular factors */
2169:   if (!cusparsestruct->use_cpu_solve) PetscCall(MatSeqAIJCUSPARSEILUAnalysisAndCopyToGPU(B));
2170:   PetscFunctionReturn(PETSC_SUCCESS);
2171: }

2173: static PetscErrorCode MatLUFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS isrow, IS iscol, const MatFactorInfo *info)
2174: {
2175:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(B->spptr);

2177:   PetscFunctionBegin;
2178:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2179:   PetscCall(MatLUFactorSymbolic_SeqAIJ(B, A, isrow, iscol, info));
2180:   B->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJCUSPARSE;
2181:   PetscFunctionReturn(PETSC_SUCCESS);
2182: }

2184: static PetscErrorCode MatILUFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS isrow, IS iscol, const MatFactorInfo *info)
2185: {
2186:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2188:   PetscFunctionBegin;
2189: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2190:   PetscBool row_identity = PETSC_FALSE, col_identity = PETSC_FALSE;
2191:   if (cusparseTriFactors->factorizeOnDevice) {
2192:     PetscCall(ISIdentity(isrow, &row_identity));
2193:     PetscCall(ISIdentity(iscol, &col_identity));
2194:   }
2195:   if (!info->levels && row_identity && col_identity) {
2196:     PetscCall(MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0(B, A, isrow, iscol, info));
2197:   } else
2198: #endif
2199:   {
2200:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2201:     PetscCall(MatILUFactorSymbolic_SeqAIJ(B, A, isrow, iscol, info));
2202:     B->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJCUSPARSE;
2203:   }
2204:   PetscFunctionReturn(PETSC_SUCCESS);
2205: }

2207: static PetscErrorCode MatICCFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS perm, const MatFactorInfo *info)
2208: {
2209:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2211:   PetscFunctionBegin;
2212: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2213:   PetscBool perm_identity = PETSC_FALSE;
2214:   if (cusparseTriFactors->factorizeOnDevice) PetscCall(ISIdentity(perm, &perm_identity));
2215:   if (!info->levels && perm_identity) {
2216:     PetscCall(MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0(B, A, perm, info));
2217:   } else
2218: #endif
2219:   {
2220:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2221:     PetscCall(MatICCFactorSymbolic_SeqAIJ(B, A, perm, info));
2222:     B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJCUSPARSE;
2223:   }
2224:   PetscFunctionReturn(PETSC_SUCCESS);
2225: }

2227: static PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS perm, const MatFactorInfo *info)
2228: {
2229:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2231:   PetscFunctionBegin;
2232:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2233:   PetscCall(MatCholeskyFactorSymbolic_SeqAIJ(B, A, perm, info));
2234:   B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJCUSPARSE;
2235:   PetscFunctionReturn(PETSC_SUCCESS);
2236: }

2238: static PetscErrorCode MatFactorGetSolverType_seqaij_cusparse(Mat, MatSolverType *type)
2239: {
2240:   PetscFunctionBegin;
2241:   *type = MATSOLVERCUSPARSE;
2242:   PetscFunctionReturn(PETSC_SUCCESS);
2243: }

2245: /*MC
2246:   MATSOLVERCUSPARSE = "cusparse" - A matrix type providing triangular solvers for seq matrices
2247:   on a single GPU of type, `MATSEQAIJCUSPARSE`. Currently supported
2248:   algorithms are ILU(k) and ICC(k). Typically, deeper factorizations (larger k) results in poorer
2249:   performance in the triangular solves. Full LU, and Cholesky decompositions can be solved through the
2250:   CuSPARSE triangular solve algorithm. However, the performance can be quite poor and thus these
2251:   algorithms are not recommended. This class does NOT support direct solver operations.

2253:   Level: beginner

2255: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJCUSPARSE`, `PCFactorSetMatSolverType()`, `MatSolverType`, `MatCreateSeqAIJCUSPARSE()`,
2256:           `MATAIJCUSPARSE`, `MatCreateAIJCUSPARSE()`, `MatCUSPARSESetFormat()`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
2257: M*/

2259: PETSC_EXTERN PetscErrorCode MatGetFactor_seqaijcusparse_cusparse(Mat A, MatFactorType ftype, Mat *B)
2260: {
2261:   PetscInt  n = A->rmap->n;
2262:   PetscBool factOnDevice, factOnHost;
2263:   char     *prefix;
2264:   char      factPlace[32] = "device"; /* the default */

2266:   PetscFunctionBegin;
2267:   PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
2268:   PetscCall(MatSetSizes(*B, n, n, n, n));
2269:   (*B)->factortype = ftype; // factortype makes MatSetType() allocate spptr of type Mat_SeqAIJCUSPARSETriFactors
2270:   PetscCall(MatSetType(*B, MATSEQAIJCUSPARSE));

2272:   prefix = (*B)->factorprefix ? (*B)->factorprefix : ((PetscObject)A)->prefix;
2273:   PetscOptionsBegin(PetscObjectComm((PetscObject)*B), prefix, "MatGetFactor", "Mat");
2274:   PetscCall(PetscOptionsString("-mat_factor_bind_factorization", "Do matrix factorization on host or device when possible", "MatGetFactor", NULL, factPlace, sizeof(factPlace), NULL));
2275:   PetscOptionsEnd();
2276:   PetscCall(PetscStrcasecmp("device", factPlace, &factOnDevice));
2277:   PetscCall(PetscStrcasecmp("host", factPlace, &factOnHost));
2278:   PetscCheck(factOnDevice || factOnHost, PetscObjectComm((PetscObject)*B), PETSC_ERR_ARG_OUTOFRANGE, "Wrong option %s to -mat_factor_bind_factorization <string>. Only host and device are allowed", factPlace);
2279:   ((Mat_SeqAIJCUSPARSETriFactors *)(*B)->spptr)->factorizeOnDevice = factOnDevice;

2281:   if (A->boundtocpu && A->bindingpropagates) PetscCall(MatBindToCPU(*B, PETSC_TRUE));
2282:   if (ftype == MAT_FACTOR_LU || ftype == MAT_FACTOR_ILU || ftype == MAT_FACTOR_ILUDT) {
2283:     PetscCall(MatSetBlockSizesFromMats(*B, A, A));
2284:     if (!A->boundtocpu) {
2285:       (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJCUSPARSE;
2286:       (*B)->ops->lufactorsymbolic  = MatLUFactorSymbolic_SeqAIJCUSPARSE;
2287:     } else {
2288:       (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJ;
2289:       (*B)->ops->lufactorsymbolic  = MatLUFactorSymbolic_SeqAIJ;
2290:     }
2291:     PetscCall(PetscStrallocpy(MATORDERINGND, (char **)&(*B)->preferredordering[MAT_FACTOR_LU]));
2292:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ILU]));
2293:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ILUDT]));
2294:   } else if (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) {
2295:     if (!A->boundtocpu) {
2296:       (*B)->ops->iccfactorsymbolic      = MatICCFactorSymbolic_SeqAIJCUSPARSE;
2297:       (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJCUSPARSE;
2298:     } else {
2299:       (*B)->ops->iccfactorsymbolic      = MatICCFactorSymbolic_SeqAIJ;
2300:       (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJ;
2301:     }
2302:     PetscCall(PetscStrallocpy(MATORDERINGND, (char **)&(*B)->preferredordering[MAT_FACTOR_CHOLESKY]));
2303:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ICC]));
2304:   } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Factor type not supported for CUSPARSE Matrix Types");

2306:   PetscCall(MatSeqAIJSetPreallocation(*B, MAT_SKIP_ALLOCATION, NULL));
2307:   (*B)->canuseordering = PETSC_TRUE;
2308:   PetscCall(PetscObjectComposeFunction((PetscObject)*B, "MatFactorGetSolverType_C", MatFactorGetSolverType_seqaij_cusparse));
2309:   PetscFunctionReturn(PETSC_SUCCESS);
2310: }

2312: static PetscErrorCode MatSeqAIJCUSPARSECopyFromGPU(Mat A)
2313: {
2314:   Mat_SeqAIJ         *a    = (Mat_SeqAIJ *)A->data;
2315:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2316: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2317:   Mat_SeqAIJCUSPARSETriFactors *fs = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
2318: #endif

2320:   PetscFunctionBegin;
2321:   if (A->offloadmask == PETSC_OFFLOAD_GPU) {
2322:     PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyFromGPU, A, 0, 0, 0));
2323:     if (A->factortype == MAT_FACTOR_NONE) {
2324:       CsrMatrix *matrix = (CsrMatrix *)cusp->mat->mat;
2325:       PetscCallCUDA(cudaMemcpy(a->a, matrix->values->data().get(), a->nz * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
2326:     }
2327: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2328:     else if (fs->csrVal) {
2329:       /* We have a factorized matrix on device and are able to copy it to host */
2330:       PetscCallCUDA(cudaMemcpy(a->a, fs->csrVal, a->nz * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
2331:     }
2332: #endif
2333:     else
2334:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for copying this type of factorized matrix from device to host");
2335:     PetscCall(PetscLogGpuToCpu(a->nz * sizeof(PetscScalar)));
2336:     PetscCall(PetscLogEventEnd(MAT_CUSPARSECopyFromGPU, A, 0, 0, 0));
2337:     A->offloadmask = PETSC_OFFLOAD_BOTH;
2338:   }
2339:   PetscFunctionReturn(PETSC_SUCCESS);
2340: }

2342: static PetscErrorCode MatSeqAIJGetArray_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2343: {
2344:   PetscFunctionBegin;
2345:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2346:   *array = ((Mat_SeqAIJ *)A->data)->a;
2347:   PetscFunctionReturn(PETSC_SUCCESS);
2348: }

2350: static PetscErrorCode MatSeqAIJRestoreArray_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2351: {
2352:   PetscFunctionBegin;
2353:   A->offloadmask = PETSC_OFFLOAD_CPU;
2354:   *array         = NULL;
2355:   PetscFunctionReturn(PETSC_SUCCESS);
2356: }

2358: static PetscErrorCode MatSeqAIJGetArrayRead_SeqAIJCUSPARSE(Mat A, const PetscScalar *array[])
2359: {
2360:   PetscFunctionBegin;
2361:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2362:   *array = ((Mat_SeqAIJ *)A->data)->a;
2363:   PetscFunctionReturn(PETSC_SUCCESS);
2364: }

2366: static PetscErrorCode MatSeqAIJRestoreArrayRead_SeqAIJCUSPARSE(Mat, const PetscScalar *array[])
2367: {
2368:   PetscFunctionBegin;
2369:   *array = NULL;
2370:   PetscFunctionReturn(PETSC_SUCCESS);
2371: }

2373: static PetscErrorCode MatSeqAIJGetArrayWrite_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2374: {
2375:   PetscFunctionBegin;
2376:   *array = ((Mat_SeqAIJ *)A->data)->a;
2377:   PetscFunctionReturn(PETSC_SUCCESS);
2378: }

2380: static PetscErrorCode MatSeqAIJRestoreArrayWrite_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2381: {
2382:   PetscFunctionBegin;
2383:   A->offloadmask = PETSC_OFFLOAD_CPU;
2384:   *array         = NULL;
2385:   PetscFunctionReturn(PETSC_SUCCESS);
2386: }

2388: static PetscErrorCode MatSeqAIJGetCSRAndMemType_SeqAIJCUSPARSE(Mat A, const PetscInt **i, const PetscInt **j, PetscScalar **a, PetscMemType *mtype)
2389: {
2390:   Mat_SeqAIJCUSPARSE *cusp;
2391:   CsrMatrix          *matrix;

2393:   PetscFunctionBegin;
2394:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
2395:   PetscCheck(A->factortype == MAT_FACTOR_NONE, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2396:   cusp = static_cast<Mat_SeqAIJCUSPARSE *>(A->spptr);
2397:   PetscCheck(cusp != NULL, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "cusp is NULL");
2398:   matrix = (CsrMatrix *)cusp->mat->mat;

2400:   if (i) {
2401: #if !defined(PETSC_USE_64BIT_INDICES)
2402:     *i = matrix->row_offsets->data().get();
2403: #else
2404:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSparse does not supported 64-bit indices");
2405: #endif
2406:   }
2407:   if (j) {
2408: #if !defined(PETSC_USE_64BIT_INDICES)
2409:     *j = matrix->column_indices->data().get();
2410: #else
2411:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSparse does not supported 64-bit indices");
2412: #endif
2413:   }
2414:   if (a) *a = matrix->values->data().get();
2415:   if (mtype) *mtype = PETSC_MEMTYPE_CUDA;
2416:   PetscFunctionReturn(PETSC_SUCCESS);
2417: }

2419: PETSC_INTERN PetscErrorCode MatSeqAIJCUSPARSECopyToGPU(Mat A)
2420: {
2421:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
2422:   Mat_SeqAIJCUSPARSEMultStruct *matstruct      = cusparsestruct->mat;
2423:   Mat_SeqAIJ                   *a              = (Mat_SeqAIJ *)A->data;
2424:   PetscInt                      m              = A->rmap->n, *ii, *ridx, tmp;
2425:   cusparseStatus_t              stat;
2426:   PetscBool                     both = PETSC_TRUE;

2428:   PetscFunctionBegin;
2429:   PetscCheck(!A->boundtocpu, PETSC_COMM_SELF, PETSC_ERR_GPU, "Cannot copy to GPU");
2430:   if (A->offloadmask == PETSC_OFFLOAD_UNALLOCATED || A->offloadmask == PETSC_OFFLOAD_CPU) {
2431:     if (A->nonzerostate == cusparsestruct->nonzerostate && cusparsestruct->format == MAT_CUSPARSE_CSR) { /* Copy values only */
2432:       CsrMatrix *matrix;
2433:       matrix = (CsrMatrix *)cusparsestruct->mat->mat;

2435:       PetscCheck(!a->nz || a->a, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR values");
2436:       PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2437:       matrix->values->assign(a->a, a->a + a->nz);
2438:       PetscCallCUDA(WaitForCUDA());
2439:       PetscCall(PetscLogCpuToGpu(a->nz * sizeof(PetscScalar)));
2440:       PetscCall(PetscLogEventEnd(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2441:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
2442:     } else {
2443:       PetscInt nnz;
2444:       PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2445:       PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusparsestruct->mat, cusparsestruct->format));
2446:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_TRUE));
2447:       delete cusparsestruct->workVector;
2448:       delete cusparsestruct->rowoffsets_gpu;
2449:       cusparsestruct->workVector     = NULL;
2450:       cusparsestruct->rowoffsets_gpu = NULL;
2451:       try {
2452:         if (a->compressedrow.use) {
2453:           m    = a->compressedrow.nrows;
2454:           ii   = a->compressedrow.i;
2455:           ridx = a->compressedrow.rindex;
2456:         } else {
2457:           m    = A->rmap->n;
2458:           ii   = a->i;
2459:           ridx = NULL;
2460:         }
2461:         PetscCheck(ii, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR row data");
2462:         if (!a->a) {
2463:           nnz  = ii[m];
2464:           both = PETSC_FALSE;
2465:         } else nnz = a->nz;
2466:         PetscCheck(!nnz || a->j, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR column data");

2468:         /* create cusparse matrix */
2469:         cusparsestruct->nrows = m;
2470:         matstruct             = new Mat_SeqAIJCUSPARSEMultStruct;
2471:         PetscCallCUSPARSE(cusparseCreateMatDescr(&matstruct->descr));
2472:         PetscCallCUSPARSE(cusparseSetMatIndexBase(matstruct->descr, CUSPARSE_INDEX_BASE_ZERO));
2473:         PetscCallCUSPARSE(cusparseSetMatType(matstruct->descr, CUSPARSE_MATRIX_TYPE_GENERAL));

2475:         PetscCallCUDA(cudaMalloc((void **)&matstruct->alpha_one, sizeof(PetscScalar)));
2476:         PetscCallCUDA(cudaMalloc((void **)&matstruct->beta_zero, sizeof(PetscScalar)));
2477:         PetscCallCUDA(cudaMalloc((void **)&matstruct->beta_one, sizeof(PetscScalar)));
2478:         PetscCallCUDA(cudaMemcpy(matstruct->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2479:         PetscCallCUDA(cudaMemcpy(matstruct->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2480:         PetscCallCUDA(cudaMemcpy(matstruct->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2481:         PetscCallCUSPARSE(cusparseSetPointerMode(cusparsestruct->handle, CUSPARSE_POINTER_MODE_DEVICE));

2483:         /* Build a hybrid/ellpack matrix if this option is chosen for the storage */
2484:         if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
2485:           /* set the matrix */
2486:           CsrMatrix *mat   = new CsrMatrix;
2487:           mat->num_rows    = m;
2488:           mat->num_cols    = A->cmap->n;
2489:           mat->num_entries = nnz;
2490:           PetscCallCXX(mat->row_offsets = new THRUSTINTARRAY32(m + 1));
2491:           mat->row_offsets->assign(ii, ii + m + 1);

2493:           PetscCallCXX(mat->column_indices = new THRUSTINTARRAY32(nnz));
2494:           mat->column_indices->assign(a->j, a->j + nnz);

2496:           PetscCallCXX(mat->values = new THRUSTARRAY(nnz));
2497:           if (a->a) mat->values->assign(a->a, a->a + nnz);

2499:           /* assign the pointer */
2500:           matstruct->mat = mat;
2501: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2502:           if (mat->num_rows) { /* cusparse errors on empty matrices! */
2503:             stat = cusparseCreateCsr(&matstruct->matDescr, mat->num_rows, mat->num_cols, mat->num_entries, mat->row_offsets->data().get(), mat->column_indices->data().get(), mat->values->data().get(), CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, /* row offset, col idx types due to THRUSTINTARRAY32 */
2504:                                      CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype);
2505:             PetscCallCUSPARSE(stat);
2506:           }
2507: #endif
2508:         } else if (cusparsestruct->format == MAT_CUSPARSE_ELL || cusparsestruct->format == MAT_CUSPARSE_HYB) {
2509: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2510:           SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
2511: #else
2512:           CsrMatrix *mat   = new CsrMatrix;
2513:           mat->num_rows    = m;
2514:           mat->num_cols    = A->cmap->n;
2515:           mat->num_entries = nnz;
2516:           PetscCallCXX(mat->row_offsets = new THRUSTINTARRAY32(m + 1));
2517:           mat->row_offsets->assign(ii, ii + m + 1);

2519:           PetscCallCXX(mat->column_indices = new THRUSTINTARRAY32(nnz));
2520:           mat->column_indices->assign(a->j, a->j + nnz);

2522:           PetscCallCXX(mat->values = new THRUSTARRAY(nnz));
2523:           if (a->a) mat->values->assign(a->a, a->a + nnz);

2525:           cusparseHybMat_t hybMat;
2526:           PetscCallCUSPARSE(cusparseCreateHybMat(&hybMat));
2527:           cusparseHybPartition_t partition = cusparsestruct->format == MAT_CUSPARSE_ELL ? CUSPARSE_HYB_PARTITION_MAX : CUSPARSE_HYB_PARTITION_AUTO;
2528:           stat                             = cusparse_csr2hyb(cusparsestruct->handle, mat->num_rows, mat->num_cols, matstruct->descr, mat->values->data().get(), mat->row_offsets->data().get(), mat->column_indices->data().get(), hybMat, 0, partition);
2529:           PetscCallCUSPARSE(stat);
2530:           /* assign the pointer */
2531:           matstruct->mat = hybMat;

2533:           if (mat) {
2534:             if (mat->values) delete (THRUSTARRAY *)mat->values;
2535:             if (mat->column_indices) delete (THRUSTINTARRAY32 *)mat->column_indices;
2536:             if (mat->row_offsets) delete (THRUSTINTARRAY32 *)mat->row_offsets;
2537:             delete (CsrMatrix *)mat;
2538:           }
2539: #endif
2540:         }

2542:         /* assign the compressed row indices */
2543:         if (a->compressedrow.use) {
2544:           PetscCallCXX(cusparsestruct->workVector = new THRUSTARRAY(m));
2545:           PetscCallCXX(matstruct->cprowIndices = new THRUSTINTARRAY(m));
2546:           matstruct->cprowIndices->assign(ridx, ridx + m);
2547:           tmp = m;
2548:         } else {
2549:           cusparsestruct->workVector = NULL;
2550:           matstruct->cprowIndices    = NULL;
2551:           tmp                        = 0;
2552:         }
2553:         PetscCall(PetscLogCpuToGpu(((m + 1) + (a->nz)) * sizeof(int) + tmp * sizeof(PetscInt) + (3 + (a->nz)) * sizeof(PetscScalar)));

2555:         /* assign the pointer */
2556:         cusparsestruct->mat = matstruct;
2557:       } catch (char *ex) {
2558:         SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
2559:       }
2560:       PetscCallCUDA(WaitForCUDA());
2561:       PetscCall(PetscLogEventEnd(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2562:       cusparsestruct->nonzerostate = A->nonzerostate;
2563:     }
2564:     if (both) A->offloadmask = PETSC_OFFLOAD_BOTH;
2565:   }
2566:   PetscFunctionReturn(PETSC_SUCCESS);
2567: }

2569: struct VecCUDAPlusEquals {
2570:   template <typename Tuple>
2571:   __host__ __device__ void operator()(Tuple t)
2572:   {
2573:     thrust::get<1>(t) = thrust::get<1>(t) + thrust::get<0>(t);
2574:   }
2575: };

2577: struct VecCUDAEquals {
2578:   template <typename Tuple>
2579:   __host__ __device__ void operator()(Tuple t)
2580:   {
2581:     thrust::get<1>(t) = thrust::get<0>(t);
2582:   }
2583: };

2585: struct VecCUDAEqualsReverse {
2586:   template <typename Tuple>
2587:   __host__ __device__ void operator()(Tuple t)
2588:   {
2589:     thrust::get<0>(t) = thrust::get<1>(t);
2590:   }
2591: };

2593: struct MatMatCusparse {
2594:   PetscBool      cisdense;
2595:   PetscScalar   *Bt;
2596:   Mat            X;
2597:   PetscBool      reusesym; /* Cusparse does not have split symbolic and numeric phases for sparse matmat operations */
2598:   PetscLogDouble flops;
2599:   CsrMatrix     *Bcsr;

2601: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2602:   cusparseSpMatDescr_t matSpBDescr;
2603:   PetscBool            initialized; /* C = alpha op(A) op(B) + beta C */
2604:   cusparseDnMatDescr_t matBDescr;
2605:   cusparseDnMatDescr_t matCDescr;
2606:   PetscInt             Blda, Clda; /* Record leading dimensions of B and C here to detect changes*/
2607:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2608:   void *dBuffer4;
2609:   void *dBuffer5;
2610:   #endif
2611:   size_t                mmBufferSize;
2612:   void                 *mmBuffer;
2613:   void                 *mmBuffer2; /* SpGEMM WorkEstimation buffer */
2614:   cusparseSpGEMMDescr_t spgemmDesc;
2615: #endif
2616: };

2618: static PetscErrorCode MatDestroy_MatMatCusparse(void *data)
2619: {
2620:   MatMatCusparse *mmdata = (MatMatCusparse *)data;

2622:   PetscFunctionBegin;
2623:   PetscCallCUDA(cudaFree(mmdata->Bt));
2624:   delete mmdata->Bcsr;
2625: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2626:   if (mmdata->matSpBDescr) PetscCallCUSPARSE(cusparseDestroySpMat(mmdata->matSpBDescr));
2627:   if (mmdata->matBDescr) PetscCallCUSPARSE(cusparseDestroyDnMat(mmdata->matBDescr));
2628:   if (mmdata->matCDescr) PetscCallCUSPARSE(cusparseDestroyDnMat(mmdata->matCDescr));
2629:   if (mmdata->spgemmDesc) PetscCallCUSPARSE(cusparseSpGEMM_destroyDescr(mmdata->spgemmDesc));
2630:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2631:   if (mmdata->dBuffer4) PetscCallCUDA(cudaFree(mmdata->dBuffer4));
2632:   if (mmdata->dBuffer5) PetscCallCUDA(cudaFree(mmdata->dBuffer5));
2633:   #endif
2634:   if (mmdata->mmBuffer) PetscCallCUDA(cudaFree(mmdata->mmBuffer));
2635:   if (mmdata->mmBuffer2) PetscCallCUDA(cudaFree(mmdata->mmBuffer2));
2636: #endif
2637:   PetscCall(MatDestroy(&mmdata->X));
2638:   PetscCall(PetscFree(data));
2639:   PetscFunctionReturn(PETSC_SUCCESS);
2640: }

2642: #include <../src/mat/impls/dense/seq/dense.h>

2644: static PetscErrorCode MatProductNumeric_SeqAIJCUSPARSE_SeqDENSECUDA(Mat C)
2645: {
2646:   Mat_Product                  *product = C->product;
2647:   Mat                           A, B;
2648:   PetscInt                      m, n, blda, clda;
2649:   PetscBool                     flg, biscuda;
2650:   Mat_SeqAIJCUSPARSE           *cusp;
2651:   cusparseStatus_t              stat;
2652:   cusparseOperation_t           opA;
2653:   const PetscScalar            *barray;
2654:   PetscScalar                  *carray;
2655:   MatMatCusparse               *mmdata;
2656:   Mat_SeqAIJCUSPARSEMultStruct *mat;
2657:   CsrMatrix                    *csrmat;

2659:   PetscFunctionBegin;
2660:   MatCheckProduct(C, 1);
2661:   PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Product data empty");
2662:   mmdata = (MatMatCusparse *)product->data;
2663:   A      = product->A;
2664:   B      = product->B;
2665:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2666:   PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Not for type %s", ((PetscObject)A)->type_name);
2667:   /* currently CopyToGpu does not copy if the matrix is bound to CPU
2668:      Instead of silently accepting the wrong answer, I prefer to raise the error */
2669:   PetscCheck(!A->boundtocpu, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Cannot bind to CPU a CUSPARSE matrix between MatProductSymbolic and MatProductNumeric phases");
2670:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
2671:   cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2672:   switch (product->type) {
2673:   case MATPRODUCT_AB:
2674:   case MATPRODUCT_PtAP:
2675:     mat = cusp->mat;
2676:     opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
2677:     m   = A->rmap->n;
2678:     n   = B->cmap->n;
2679:     break;
2680:   case MATPRODUCT_AtB:
2681:     if (!A->form_explicit_transpose) {
2682:       mat = cusp->mat;
2683:       opA = CUSPARSE_OPERATION_TRANSPOSE;
2684:     } else {
2685:       PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
2686:       mat = cusp->matTranspose;
2687:       opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
2688:     }
2689:     m = A->cmap->n;
2690:     n = B->cmap->n;
2691:     break;
2692:   case MATPRODUCT_ABt:
2693:   case MATPRODUCT_RARt:
2694:     mat = cusp->mat;
2695:     opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
2696:     m   = A->rmap->n;
2697:     n   = B->rmap->n;
2698:     break;
2699:   default:
2700:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Unsupported product type %s", MatProductTypes[product->type]);
2701:   }
2702:   PetscCheck(mat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing Mat_SeqAIJCUSPARSEMultStruct");
2703:   csrmat = (CsrMatrix *)mat->mat;
2704:   /* if the user passed a CPU matrix, copy the data to the GPU */
2705:   PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQDENSECUDA, &biscuda));
2706:   if (!biscuda) PetscCall(MatConvert(B, MATSEQDENSECUDA, MAT_INPLACE_MATRIX, &B));
2707:   PetscCall(MatDenseGetArrayReadAndMemType(B, &barray, nullptr));

2709:   PetscCall(MatDenseGetLDA(B, &blda));
2710:   if (product->type == MATPRODUCT_RARt || product->type == MATPRODUCT_PtAP) {
2711:     PetscCall(MatDenseGetArrayWriteAndMemType(mmdata->X, &carray, nullptr));
2712:     PetscCall(MatDenseGetLDA(mmdata->X, &clda));
2713:   } else {
2714:     PetscCall(MatDenseGetArrayWriteAndMemType(C, &carray, nullptr));
2715:     PetscCall(MatDenseGetLDA(C, &clda));
2716:   }

2718:   PetscCall(PetscLogGpuTimeBegin());
2719: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2720:   cusparseOperation_t opB = (product->type == MATPRODUCT_ABt || product->type == MATPRODUCT_RARt) ? CUSPARSE_OPERATION_TRANSPOSE : CUSPARSE_OPERATION_NON_TRANSPOSE;
2721:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
2722:   cusparseSpMatDescr_t &matADescr = mat->matDescr_SpMM[opA];
2723:   #else
2724:   cusparseSpMatDescr_t &matADescr = mat->matDescr;
2725:   #endif

2727:   /* (re)allocate mmBuffer if not initialized or LDAs are different */
2728:   if (!mmdata->initialized || mmdata->Blda != blda || mmdata->Clda != clda) {
2729:     size_t mmBufferSize;
2730:     if (mmdata->initialized && mmdata->Blda != blda) {
2731:       PetscCallCUSPARSE(cusparseDestroyDnMat(mmdata->matBDescr));
2732:       mmdata->matBDescr = NULL;
2733:     }
2734:     if (!mmdata->matBDescr) {
2735:       PetscCallCUSPARSE(cusparseCreateDnMat(&mmdata->matBDescr, B->rmap->n, B->cmap->n, blda, (void *)barray, cusparse_scalartype, CUSPARSE_ORDER_COL));
2736:       mmdata->Blda = blda;
2737:     }

2739:     if (mmdata->initialized && mmdata->Clda != clda) {
2740:       PetscCallCUSPARSE(cusparseDestroyDnMat(mmdata->matCDescr));
2741:       mmdata->matCDescr = NULL;
2742:     }
2743:     if (!mmdata->matCDescr) { /* matCDescr is for C or mmdata->X */
2744:       PetscCallCUSPARSE(cusparseCreateDnMat(&mmdata->matCDescr, m, n, clda, (void *)carray, cusparse_scalartype, CUSPARSE_ORDER_COL));
2745:       mmdata->Clda = clda;
2746:     }

2748:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0) // tested up to 12.6.0
2749:     if (matADescr) {
2750:       PetscCallCUSPARSE(cusparseDestroySpMat(matADescr)); // Because I find I could not reuse matADescr. It could be a cusparse bug
2751:       matADescr = NULL;
2752:     }
2753:   #endif

2755:     if (!matADescr) {
2756:       stat = cusparseCreateCsr(&matADescr, csrmat->num_rows, csrmat->num_cols, csrmat->num_entries, csrmat->row_offsets->data().get(), csrmat->column_indices->data().get(), csrmat->values->data().get(), CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, /* row offset, col idx types due to THRUSTINTARRAY32 */
2757:                                CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype);
2758:       PetscCallCUSPARSE(stat);
2759:     }

2761:     PetscCallCUSPARSE(cusparseSpMM_bufferSize(cusp->handle, opA, opB, mat->alpha_one, matADescr, mmdata->matBDescr, mat->beta_zero, mmdata->matCDescr, cusparse_scalartype, cusp->spmmAlg, &mmBufferSize));

2763:     if ((mmdata->mmBuffer && mmdata->mmBufferSize < mmBufferSize) || !mmdata->mmBuffer) {
2764:       PetscCallCUDA(cudaFree(mmdata->mmBuffer));
2765:       PetscCallCUDA(cudaMalloc(&mmdata->mmBuffer, mmBufferSize));
2766:       mmdata->mmBufferSize = mmBufferSize;
2767:     }

2769:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0) // the _preprocess was added in 11.2.1, but petsc worked without it until 12.4.0
2770:     PetscCallCUSPARSE(cusparseSpMM_preprocess(cusp->handle, opA, opB, mat->alpha_one, matADescr, mmdata->matBDescr, mat->beta_zero, mmdata->matCDescr, cusparse_scalartype, cusp->spmmAlg, mmdata->mmBuffer));
2771:   #endif

2773:     mmdata->initialized = PETSC_TRUE;
2774:   } else {
2775:     /* to be safe, always update pointers of the mats */
2776:     PetscCallCUSPARSE(cusparseSpMatSetValues(matADescr, csrmat->values->data().get()));
2777:     PetscCallCUSPARSE(cusparseDnMatSetValues(mmdata->matBDescr, (void *)barray));
2778:     PetscCallCUSPARSE(cusparseDnMatSetValues(mmdata->matCDescr, (void *)carray));
2779:   }

2781:   /* do cusparseSpMM, which supports transpose on B */
2782:   PetscCallCUSPARSE(cusparseSpMM(cusp->handle, opA, opB, mat->alpha_one, matADescr, mmdata->matBDescr, mat->beta_zero, mmdata->matCDescr, cusparse_scalartype, cusp->spmmAlg, mmdata->mmBuffer));
2783: #else
2784:   PetscInt k;
2785:   /* cusparseXcsrmm does not support transpose on B */
2786:   if (product->type == MATPRODUCT_ABt || product->type == MATPRODUCT_RARt) {
2787:     cublasHandle_t cublasv2handle;
2788:     cublasStatus_t cerr;

2790:     PetscCall(PetscCUBLASGetHandle(&cublasv2handle));
2791:     cerr = cublasXgeam(cublasv2handle, CUBLAS_OP_T, CUBLAS_OP_T, B->cmap->n, B->rmap->n, &PETSC_CUSPARSE_ONE, barray, blda, &PETSC_CUSPARSE_ZERO, barray, blda, mmdata->Bt, B->cmap->n);
2792:     PetscCallCUBLAS(cerr);
2793:     blda = B->cmap->n;
2794:     k    = B->cmap->n;
2795:   } else {
2796:     k = B->rmap->n;
2797:   }

2799:   /* perform the MatMat operation, op(A) is m x k, op(B) is k x n */
2800:   stat = cusparse_csr_spmm(cusp->handle, opA, m, n, k, csrmat->num_entries, mat->alpha_one, mat->descr, csrmat->values->data().get(), csrmat->row_offsets->data().get(), csrmat->column_indices->data().get(), mmdata->Bt ? mmdata->Bt : barray, blda, mat->beta_zero, carray, clda);
2801:   PetscCallCUSPARSE(stat);
2802: #endif
2803:   PetscCall(PetscLogGpuTimeEnd());
2804:   PetscCall(PetscLogGpuFlops(n * 2.0 * csrmat->num_entries));
2805:   PetscCall(MatDenseRestoreArrayReadAndMemType(B, &barray));
2806:   if (product->type == MATPRODUCT_RARt) {
2807:     PetscCall(MatDenseRestoreArrayWriteAndMemType(mmdata->X, &carray));
2808:     PetscCall(MatMatMultNumeric_SeqDenseCUDA_SeqDenseCUDA_Internal(B, mmdata->X, C, PETSC_FALSE, PETSC_FALSE));
2809:   } else if (product->type == MATPRODUCT_PtAP) {
2810:     PetscCall(MatDenseRestoreArrayWriteAndMemType(mmdata->X, &carray));
2811:     PetscCall(MatMatMultNumeric_SeqDenseCUDA_SeqDenseCUDA_Internal(B, mmdata->X, C, PETSC_TRUE, PETSC_FALSE));
2812:   } else {
2813:     PetscCall(MatDenseRestoreArrayWriteAndMemType(C, &carray));
2814:   }
2815:   if (mmdata->cisdense) PetscCall(MatConvert(C, MATSEQDENSE, MAT_INPLACE_MATRIX, &C));
2816:   if (!biscuda) PetscCall(MatConvert(B, MATSEQDENSE, MAT_INPLACE_MATRIX, &B));
2817:   PetscFunctionReturn(PETSC_SUCCESS);
2818: }

2820: static PetscErrorCode MatProductSymbolic_SeqAIJCUSPARSE_SeqDENSECUDA(Mat C)
2821: {
2822:   Mat_Product        *product = C->product;
2823:   Mat                 A, B;
2824:   PetscInt            m, n;
2825:   PetscBool           cisdense, flg;
2826:   MatMatCusparse     *mmdata;
2827:   Mat_SeqAIJCUSPARSE *cusp;

2829:   PetscFunctionBegin;
2830:   MatCheckProduct(C, 1);
2831:   PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Product data not empty");
2832:   A = product->A;
2833:   B = product->B;
2834:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2835:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for type %s", ((PetscObject)A)->type_name);
2836:   cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2837:   PetscCheck(cusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2838:   switch (product->type) {
2839:   case MATPRODUCT_AB:
2840:     m = A->rmap->n;
2841:     n = B->cmap->n;
2842:     break;
2843:   case MATPRODUCT_AtB:
2844:     m = A->cmap->n;
2845:     n = B->cmap->n;
2846:     break;
2847:   case MATPRODUCT_ABt:
2848:     m = A->rmap->n;
2849:     n = B->rmap->n;
2850:     break;
2851:   case MATPRODUCT_PtAP:
2852:     m = B->cmap->n;
2853:     n = B->cmap->n;
2854:     break;
2855:   case MATPRODUCT_RARt:
2856:     m = B->rmap->n;
2857:     n = B->rmap->n;
2858:     break;
2859:   default:
2860:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Unsupported product type %s", MatProductTypes[product->type]);
2861:   }
2862:   PetscCall(MatSetSizes(C, m, n, m, n));
2863:   /* if C is of type MATSEQDENSE (CPU), perform the operation on the GPU and then copy on the CPU */
2864:   PetscCall(PetscObjectTypeCompare((PetscObject)C, MATSEQDENSE, &cisdense));
2865:   PetscCall(MatSetType(C, MATSEQDENSECUDA));

2867:   /* product data */
2868:   PetscCall(PetscNew(&mmdata));
2869:   mmdata->cisdense = cisdense;
2870: #if PETSC_PKG_CUDA_VERSION_LT(11, 0, 0)
2871:   /* cusparseXcsrmm does not support transpose on B, so we allocate buffer to store B^T */
2872:   if (product->type == MATPRODUCT_ABt || product->type == MATPRODUCT_RARt) PetscCallCUDA(cudaMalloc((void **)&mmdata->Bt, (size_t)B->rmap->n * (size_t)B->cmap->n * sizeof(PetscScalar)));
2873: #endif
2874:   /* for these products we need intermediate storage */
2875:   if (product->type == MATPRODUCT_RARt || product->type == MATPRODUCT_PtAP) {
2876:     PetscCall(MatCreate(PetscObjectComm((PetscObject)C), &mmdata->X));
2877:     PetscCall(MatSetType(mmdata->X, MATSEQDENSECUDA));
2878:     if (product->type == MATPRODUCT_RARt) { /* do not preallocate, since the first call to MatDenseCUDAGetArray will preallocate on the GPU for us */
2879:       PetscCall(MatSetSizes(mmdata->X, A->rmap->n, B->rmap->n, A->rmap->n, B->rmap->n));
2880:     } else {
2881:       PetscCall(MatSetSizes(mmdata->X, A->rmap->n, B->cmap->n, A->rmap->n, B->cmap->n));
2882:     }
2883:   }
2884:   C->product->data    = mmdata;
2885:   C->product->destroy = MatDestroy_MatMatCusparse;

2887:   C->ops->productnumeric = MatProductNumeric_SeqAIJCUSPARSE_SeqDENSECUDA;
2888:   PetscFunctionReturn(PETSC_SUCCESS);
2889: }

2891: static PetscErrorCode MatProductNumeric_SeqAIJCUSPARSE_SeqAIJCUSPARSE(Mat C)
2892: {
2893:   Mat_Product                  *product = C->product;
2894:   Mat                           A, B;
2895:   Mat_SeqAIJCUSPARSE           *Acusp, *Bcusp, *Ccusp;
2896:   Mat_SeqAIJ                   *c = (Mat_SeqAIJ *)C->data;
2897:   Mat_SeqAIJCUSPARSEMultStruct *Amat, *Bmat, *Cmat;
2898:   CsrMatrix                    *Acsr, *Bcsr, *Ccsr;
2899:   PetscBool                     flg;
2900:   cusparseStatus_t              stat;
2901:   MatProductType                ptype;
2902:   MatMatCusparse               *mmdata;
2903: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2904:   cusparseSpMatDescr_t BmatSpDescr;
2905: #endif
2906:   cusparseOperation_t opA = CUSPARSE_OPERATION_NON_TRANSPOSE, opB = CUSPARSE_OPERATION_NON_TRANSPOSE; /* cuSPARSE spgemm doesn't support transpose yet */

2908:   PetscFunctionBegin;
2909:   MatCheckProduct(C, 1);
2910:   PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Product data empty");
2911:   PetscCall(PetscObjectTypeCompare((PetscObject)C, MATSEQAIJCUSPARSE, &flg));
2912:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for C of type %s", ((PetscObject)C)->type_name);
2913:   mmdata = (MatMatCusparse *)C->product->data;
2914:   A      = product->A;
2915:   B      = product->B;
2916:   if (mmdata->reusesym) { /* this happens when api_user is true, meaning that the matrix values have been already computed in the MatProductSymbolic phase */
2917:     mmdata->reusesym = PETSC_FALSE;
2918:     Ccusp            = (Mat_SeqAIJCUSPARSE *)C->spptr;
2919:     PetscCheck(Ccusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2920:     Cmat = Ccusp->mat;
2921:     PetscCheck(Cmat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing C mult struct for product type %s", MatProductTypes[C->product->type]);
2922:     Ccsr = (CsrMatrix *)Cmat->mat;
2923:     PetscCheck(Ccsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing C CSR struct");
2924:     goto finalize;
2925:   }
2926:   if (!c->nz) goto finalize;
2927:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2928:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for type %s", ((PetscObject)A)->type_name);
2929:   PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJCUSPARSE, &flg));
2930:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for B of type %s", ((PetscObject)B)->type_name);
2931:   PetscCheck(!A->boundtocpu, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONG, "Cannot bind to CPU a CUSPARSE matrix between MatProductSymbolic and MatProductNumeric phases");
2932:   PetscCheck(!B->boundtocpu, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONG, "Cannot bind to CPU a CUSPARSE matrix between MatProductSymbolic and MatProductNumeric phases");
2933:   Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2934:   Bcusp = (Mat_SeqAIJCUSPARSE *)B->spptr;
2935:   Ccusp = (Mat_SeqAIJCUSPARSE *)C->spptr;
2936:   PetscCheck(Acusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2937:   PetscCheck(Bcusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2938:   PetscCheck(Ccusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2939:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
2940:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));

2942:   ptype = product->type;
2943:   if (A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
2944:     ptype = MATPRODUCT_AB;
2945:     PetscCheck(product->symbolic_used_the_fact_A_is_symmetric, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Symbolic should have been built using the fact that A is symmetric");
2946:   }
2947:   if (B->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_ABt) {
2948:     ptype = MATPRODUCT_AB;
2949:     PetscCheck(product->symbolic_used_the_fact_B_is_symmetric, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Symbolic should have been built using the fact that B is symmetric");
2950:   }
2951:   switch (ptype) {
2952:   case MATPRODUCT_AB:
2953:     Amat = Acusp->mat;
2954:     Bmat = Bcusp->mat;
2955:     break;
2956:   case MATPRODUCT_AtB:
2957:     Amat = Acusp->matTranspose;
2958:     Bmat = Bcusp->mat;
2959:     break;
2960:   case MATPRODUCT_ABt:
2961:     Amat = Acusp->mat;
2962:     Bmat = Bcusp->matTranspose;
2963:     break;
2964:   default:
2965:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Unsupported product type %s", MatProductTypes[product->type]);
2966:   }
2967:   Cmat = Ccusp->mat;
2968:   PetscCheck(Amat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing A mult struct for product type %s", MatProductTypes[ptype]);
2969:   PetscCheck(Bmat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing B mult struct for product type %s", MatProductTypes[ptype]);
2970:   PetscCheck(Cmat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing C mult struct for product type %s", MatProductTypes[ptype]);
2971:   Acsr = (CsrMatrix *)Amat->mat;
2972:   Bcsr = mmdata->Bcsr ? mmdata->Bcsr : (CsrMatrix *)Bmat->mat; /* B may be in compressed row storage */
2973:   Ccsr = (CsrMatrix *)Cmat->mat;
2974:   PetscCheck(Acsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing A CSR struct");
2975:   PetscCheck(Bcsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing B CSR struct");
2976:   PetscCheck(Ccsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing C CSR struct");
2977:   PetscCall(PetscLogGpuTimeBegin());
2978: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2979:   BmatSpDescr = mmdata->Bcsr ? mmdata->matSpBDescr : Bmat->matDescr; /* B may be in compressed row storage */
2980:   PetscCallCUSPARSE(cusparseSetPointerMode(Ccusp->handle, CUSPARSE_POINTER_MODE_DEVICE));
2981:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2982:   stat = cusparseSpGEMMreuse_compute(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc);
2983:   PetscCallCUSPARSE(stat);
2984:   #else
2985:   stat = cusparseSpGEMM_compute(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &mmdata->mmBufferSize, mmdata->mmBuffer);
2986:   PetscCallCUSPARSE(stat);
2987:   stat = cusparseSpGEMM_copy(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc);
2988:   PetscCallCUSPARSE(stat);
2989:   #endif
2990: #else
2991:   stat = cusparse_csr_spgemm(Ccusp->handle, opA, opB, Acsr->num_rows, Bcsr->num_cols, Acsr->num_cols, Amat->descr, Acsr->num_entries, Acsr->values->data().get(), Acsr->row_offsets->data().get(), Acsr->column_indices->data().get(), Bmat->descr, Bcsr->num_entries,
2992:                              Bcsr->values->data().get(), Bcsr->row_offsets->data().get(), Bcsr->column_indices->data().get(), Cmat->descr, Ccsr->values->data().get(), Ccsr->row_offsets->data().get(), Ccsr->column_indices->data().get());
2993:   PetscCallCUSPARSE(stat);
2994: #endif
2995:   PetscCall(PetscLogGpuFlops(mmdata->flops));
2996:   PetscCallCUDA(WaitForCUDA());
2997:   PetscCall(PetscLogGpuTimeEnd());
2998:   C->offloadmask = PETSC_OFFLOAD_GPU;
2999: finalize:
3000:   /* shorter version of MatAssemblyEnd_SeqAIJ */
3001:   PetscCall(PetscInfo(C, "Matrix size: %" PetscInt_FMT " X %" PetscInt_FMT "; storage space: 0 unneeded,%" PetscInt_FMT " used\n", C->rmap->n, C->cmap->n, c->nz));
3002:   PetscCall(PetscInfo(C, "Number of mallocs during MatSetValues() is 0\n"));
3003:   PetscCall(PetscInfo(C, "Maximum nonzeros in any row is %" PetscInt_FMT "\n", c->rmax));
3004:   c->reallocs = 0;
3005:   C->info.mallocs += 0;
3006:   C->info.nz_unneeded = 0;
3007:   C->assembled = C->was_assembled = PETSC_TRUE;
3008:   C->num_ass++;
3009:   PetscFunctionReturn(PETSC_SUCCESS);
3010: }

3012: static PetscErrorCode MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE(Mat C)
3013: {
3014:   Mat_Product                  *product = C->product;
3015:   Mat                           A, B;
3016:   Mat_SeqAIJCUSPARSE           *Acusp, *Bcusp, *Ccusp;
3017:   Mat_SeqAIJ                   *a, *b, *c;
3018:   Mat_SeqAIJCUSPARSEMultStruct *Amat, *Bmat, *Cmat;
3019:   CsrMatrix                    *Acsr, *Bcsr, *Ccsr;
3020:   PetscInt                      i, j, m, n, k;
3021:   PetscBool                     flg;
3022:   cusparseStatus_t              stat;
3023:   MatProductType                ptype;
3024:   MatMatCusparse               *mmdata;
3025:   PetscLogDouble                flops;
3026:   PetscBool                     biscompressed, ciscompressed;
3027: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3028:   int64_t              C_num_rows1, C_num_cols1, C_nnz1;
3029:   cusparseSpMatDescr_t BmatSpDescr;
3030: #else
3031:   int cnz;
3032: #endif
3033:   cusparseOperation_t opA = CUSPARSE_OPERATION_NON_TRANSPOSE, opB = CUSPARSE_OPERATION_NON_TRANSPOSE; /* cuSPARSE spgemm doesn't support transpose yet */

3035:   PetscFunctionBegin;
3036:   MatCheckProduct(C, 1);
3037:   PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Product data not empty");
3038:   A = product->A;
3039:   B = product->B;
3040:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
3041:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for type %s", ((PetscObject)A)->type_name);
3042:   PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJCUSPARSE, &flg));
3043:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for B of type %s", ((PetscObject)B)->type_name);
3044:   a = (Mat_SeqAIJ *)A->data;
3045:   b = (Mat_SeqAIJ *)B->data;
3046:   /* product data */
3047:   PetscCall(PetscNew(&mmdata));
3048:   C->product->data    = mmdata;
3049:   C->product->destroy = MatDestroy_MatMatCusparse;

3051:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
3052:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));
3053:   Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr; /* Access spptr after MatSeqAIJCUSPARSECopyToGPU, not before */
3054:   Bcusp = (Mat_SeqAIJCUSPARSE *)B->spptr;
3055:   PetscCheck(Acusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
3056:   PetscCheck(Bcusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");

3058:   ptype = product->type;
3059:   if (A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
3060:     ptype                                          = MATPRODUCT_AB;
3061:     product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
3062:   }
3063:   if (B->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_ABt) {
3064:     ptype                                          = MATPRODUCT_AB;
3065:     product->symbolic_used_the_fact_B_is_symmetric = PETSC_TRUE;
3066:   }
3067:   biscompressed = PETSC_FALSE;
3068:   ciscompressed = PETSC_FALSE;
3069:   switch (ptype) {
3070:   case MATPRODUCT_AB:
3071:     m    = A->rmap->n;
3072:     n    = B->cmap->n;
3073:     k    = A->cmap->n;
3074:     Amat = Acusp->mat;
3075:     Bmat = Bcusp->mat;
3076:     if (a->compressedrow.use) ciscompressed = PETSC_TRUE;
3077:     if (b->compressedrow.use) biscompressed = PETSC_TRUE;
3078:     break;
3079:   case MATPRODUCT_AtB:
3080:     m = A->cmap->n;
3081:     n = B->cmap->n;
3082:     k = A->rmap->n;
3083:     PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
3084:     Amat = Acusp->matTranspose;
3085:     Bmat = Bcusp->mat;
3086:     if (b->compressedrow.use) biscompressed = PETSC_TRUE;
3087:     break;
3088:   case MATPRODUCT_ABt:
3089:     m = A->rmap->n;
3090:     n = B->rmap->n;
3091:     k = A->cmap->n;
3092:     PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(B));
3093:     Amat = Acusp->mat;
3094:     Bmat = Bcusp->matTranspose;
3095:     if (a->compressedrow.use) ciscompressed = PETSC_TRUE;
3096:     break;
3097:   default:
3098:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Unsupported product type %s", MatProductTypes[product->type]);
3099:   }

3101:   /* create cusparse matrix */
3102:   PetscCall(MatSetSizes(C, m, n, m, n));
3103:   PetscCall(MatSetType(C, MATSEQAIJCUSPARSE));
3104:   c     = (Mat_SeqAIJ *)C->data;
3105:   Ccusp = (Mat_SeqAIJCUSPARSE *)C->spptr;
3106:   Cmat  = new Mat_SeqAIJCUSPARSEMultStruct;
3107:   Ccsr  = new CsrMatrix;

3109:   c->compressedrow.use = ciscompressed;
3110:   if (c->compressedrow.use) { /* if a is in compressed row, than c will be in compressed row format */
3111:     c->compressedrow.nrows = a->compressedrow.nrows;
3112:     PetscCall(PetscMalloc2(c->compressedrow.nrows + 1, &c->compressedrow.i, c->compressedrow.nrows, &c->compressedrow.rindex));
3113:     PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, c->compressedrow.nrows));
3114:     Ccusp->workVector  = new THRUSTARRAY(c->compressedrow.nrows);
3115:     Cmat->cprowIndices = new THRUSTINTARRAY(c->compressedrow.nrows);
3116:     Cmat->cprowIndices->assign(c->compressedrow.rindex, c->compressedrow.rindex + c->compressedrow.nrows);
3117:   } else {
3118:     c->compressedrow.nrows  = 0;
3119:     c->compressedrow.i      = NULL;
3120:     c->compressedrow.rindex = NULL;
3121:     Ccusp->workVector       = NULL;
3122:     Cmat->cprowIndices      = NULL;
3123:   }
3124:   Ccusp->nrows      = ciscompressed ? c->compressedrow.nrows : m;
3125:   Ccusp->mat        = Cmat;
3126:   Ccusp->mat->mat   = Ccsr;
3127:   Ccsr->num_rows    = Ccusp->nrows;
3128:   Ccsr->num_cols    = n;
3129:   Ccsr->row_offsets = new THRUSTINTARRAY32(Ccusp->nrows + 1);
3130:   PetscCallCUSPARSE(cusparseCreateMatDescr(&Cmat->descr));
3131:   PetscCallCUSPARSE(cusparseSetMatIndexBase(Cmat->descr, CUSPARSE_INDEX_BASE_ZERO));
3132:   PetscCallCUSPARSE(cusparseSetMatType(Cmat->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
3133:   PetscCallCUDA(cudaMalloc((void **)&Cmat->alpha_one, sizeof(PetscScalar)));
3134:   PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_zero, sizeof(PetscScalar)));
3135:   PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_one, sizeof(PetscScalar)));
3136:   PetscCallCUDA(cudaMemcpy(Cmat->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
3137:   PetscCallCUDA(cudaMemcpy(Cmat->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
3138:   PetscCallCUDA(cudaMemcpy(Cmat->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
3139:   if (!Ccsr->num_rows || !Ccsr->num_cols || !a->nz || !b->nz) { /* cusparse raise errors in different calls when matrices have zero rows/columns! */
3140:     PetscCallThrust(thrust::fill(thrust::device, Ccsr->row_offsets->begin(), Ccsr->row_offsets->end(), 0));
3141:     c->nz                = 0;
3142:     Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
3143:     Ccsr->values         = new THRUSTARRAY(c->nz);
3144:     goto finalizesym;
3145:   }

3147:   PetscCheck(Amat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing A mult struct for product type %s", MatProductTypes[ptype]);
3148:   PetscCheck(Bmat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing B mult struct for product type %s", MatProductTypes[ptype]);
3149:   Acsr = (CsrMatrix *)Amat->mat;
3150:   if (!biscompressed) {
3151:     Bcsr = (CsrMatrix *)Bmat->mat;
3152: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3153:     BmatSpDescr = Bmat->matDescr;
3154: #endif
3155:   } else { /* we need to use row offsets for the full matrix */
3156:     CsrMatrix *cBcsr     = (CsrMatrix *)Bmat->mat;
3157:     Bcsr                 = new CsrMatrix;
3158:     Bcsr->num_rows       = B->rmap->n;
3159:     Bcsr->num_cols       = cBcsr->num_cols;
3160:     Bcsr->num_entries    = cBcsr->num_entries;
3161:     Bcsr->column_indices = cBcsr->column_indices;
3162:     Bcsr->values         = cBcsr->values;
3163:     if (!Bcusp->rowoffsets_gpu) {
3164:       Bcusp->rowoffsets_gpu = new THRUSTINTARRAY32(B->rmap->n + 1);
3165:       Bcusp->rowoffsets_gpu->assign(b->i, b->i + B->rmap->n + 1);
3166:       PetscCall(PetscLogCpuToGpu((B->rmap->n + 1) * sizeof(PetscInt)));
3167:     }
3168:     Bcsr->row_offsets = Bcusp->rowoffsets_gpu;
3169:     mmdata->Bcsr      = Bcsr;
3170: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3171:     if (Bcsr->num_rows && Bcsr->num_cols) {
3172:       stat = cusparseCreateCsr(&mmdata->matSpBDescr, Bcsr->num_rows, Bcsr->num_cols, Bcsr->num_entries, Bcsr->row_offsets->data().get(), Bcsr->column_indices->data().get(), Bcsr->values->data().get(), CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype);
3173:       PetscCallCUSPARSE(stat);
3174:     }
3175:     BmatSpDescr = mmdata->matSpBDescr;
3176: #endif
3177:   }
3178:   PetscCheck(Acsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing A CSR struct");
3179:   PetscCheck(Bcsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing B CSR struct");
3180:   /* precompute flops count */
3181:   if (ptype == MATPRODUCT_AB) {
3182:     for (i = 0, flops = 0; i < A->rmap->n; i++) {
3183:       const PetscInt st = a->i[i];
3184:       const PetscInt en = a->i[i + 1];
3185:       for (j = st; j < en; j++) {
3186:         const PetscInt brow = a->j[j];
3187:         flops += 2. * (b->i[brow + 1] - b->i[brow]);
3188:       }
3189:     }
3190:   } else if (ptype == MATPRODUCT_AtB) {
3191:     for (i = 0, flops = 0; i < A->rmap->n; i++) {
3192:       const PetscInt anzi = a->i[i + 1] - a->i[i];
3193:       const PetscInt bnzi = b->i[i + 1] - b->i[i];
3194:       flops += (2. * anzi) * bnzi;
3195:     }
3196:   } else { /* TODO */
3197:     flops = 0.;
3198:   }

3200:   mmdata->flops = flops;
3201:   PetscCall(PetscLogGpuTimeBegin());

3203: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3204:   PetscCallCUSPARSE(cusparseSetPointerMode(Ccusp->handle, CUSPARSE_POINTER_MODE_DEVICE));
3205:   // cuda-12.2 requires non-null csrRowOffsets
3206:   stat = cusparseCreateCsr(&Cmat->matDescr, Ccsr->num_rows, Ccsr->num_cols, 0, Ccsr->row_offsets->data().get(), NULL, NULL, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype);
3207:   PetscCallCUSPARSE(stat);
3208:   PetscCallCUSPARSE(cusparseSpGEMM_createDescr(&mmdata->spgemmDesc));
3209:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
3210:   {
3211:     /* cusparseSpGEMMreuse has more reasonable APIs than cusparseSpGEMM, so we prefer to use it.
3212:      We follow the sample code at https://github.com/NVIDIA/CUDALibrarySamples/blob/master/cuSPARSE/spgemm_reuse
3213:   */
3214:     void *dBuffer1 = NULL;
3215:     void *dBuffer2 = NULL;
3216:     void *dBuffer3 = NULL;
3217:     /* dBuffer4, dBuffer5 are needed by cusparseSpGEMMreuse_compute, and therefore are stored in mmdata */
3218:     size_t bufferSize1 = 0;
3219:     size_t bufferSize2 = 0;
3220:     size_t bufferSize3 = 0;
3221:     size_t bufferSize4 = 0;
3222:     size_t bufferSize5 = 0;

3224:     /* ask bufferSize1 bytes for external memory */
3225:     stat = cusparseSpGEMMreuse_workEstimation(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize1, NULL);
3226:     PetscCallCUSPARSE(stat);
3227:     PetscCallCUDA(cudaMalloc((void **)&dBuffer1, bufferSize1));
3228:     /* inspect the matrices A and B to understand the memory requirement for the next step */
3229:     stat = cusparseSpGEMMreuse_workEstimation(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize1, dBuffer1);
3230:     PetscCallCUSPARSE(stat);

3232:     stat = cusparseSpGEMMreuse_nnz(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize2, NULL, &bufferSize3, NULL, &bufferSize4, NULL);
3233:     PetscCallCUSPARSE(stat);
3234:     PetscCallCUDA(cudaMalloc((void **)&dBuffer2, bufferSize2));
3235:     PetscCallCUDA(cudaMalloc((void **)&dBuffer3, bufferSize3));
3236:     PetscCallCUDA(cudaMalloc((void **)&mmdata->dBuffer4, bufferSize4));
3237:     stat = cusparseSpGEMMreuse_nnz(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize2, dBuffer2, &bufferSize3, dBuffer3, &bufferSize4, mmdata->dBuffer4);
3238:     PetscCallCUSPARSE(stat);
3239:     PetscCallCUDA(cudaFree(dBuffer1));
3240:     PetscCallCUDA(cudaFree(dBuffer2));

3242:     /* get matrix C non-zero entries C_nnz1 */
3243:     PetscCallCUSPARSE(cusparseSpMatGetSize(Cmat->matDescr, &C_num_rows1, &C_num_cols1, &C_nnz1));
3244:     c->nz = (PetscInt)C_nnz1;
3245:     /* allocate matrix C */
3246:     Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
3247:     PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3248:     Ccsr->values = new THRUSTARRAY(c->nz);
3249:     PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3250:     /* update matC with the new pointers */
3251:     stat = cusparseCsrSetPointers(Cmat->matDescr, Ccsr->row_offsets->data().get(), Ccsr->column_indices->data().get(), Ccsr->values->data().get());
3252:     PetscCallCUSPARSE(stat);

3254:     stat = cusparseSpGEMMreuse_copy(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize5, NULL);
3255:     PetscCallCUSPARSE(stat);
3256:     PetscCallCUDA(cudaMalloc((void **)&mmdata->dBuffer5, bufferSize5));
3257:     stat = cusparseSpGEMMreuse_copy(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize5, mmdata->dBuffer5);
3258:     PetscCallCUSPARSE(stat);
3259:     PetscCallCUDA(cudaFree(dBuffer3));
3260:     stat = cusparseSpGEMMreuse_compute(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc);
3261:     PetscCallCUSPARSE(stat);
3262:     PetscCall(PetscInfo(C, "Buffer sizes for type %s, result %" PetscInt_FMT " x %" PetscInt_FMT " (k %" PetscInt_FMT ", nzA %" PetscInt_FMT ", nzB %" PetscInt_FMT ", nzC %" PetscInt_FMT ") are: %ldKB %ldKB\n", MatProductTypes[ptype], m, n, k, a->nz, b->nz, c->nz, bufferSize4 / 1024, bufferSize5 / 1024));
3263:   }
3264:   #else
3265:   size_t bufSize2;
3266:   /* ask bufferSize bytes for external memory */
3267:   stat = cusparseSpGEMM_workEstimation(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufSize2, NULL);
3268:   PetscCallCUSPARSE(stat);
3269:   PetscCallCUDA(cudaMalloc((void **)&mmdata->mmBuffer2, bufSize2));
3270:   /* inspect the matrices A and B to understand the memory requirement for the next step */
3271:   stat = cusparseSpGEMM_workEstimation(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufSize2, mmdata->mmBuffer2);
3272:   PetscCallCUSPARSE(stat);
3273:   /* ask bufferSize again bytes for external memory */
3274:   stat = cusparseSpGEMM_compute(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &mmdata->mmBufferSize, NULL);
3275:   PetscCallCUSPARSE(stat);
3276:   /* The CUSPARSE documentation is not clear, nor the API
3277:      We need both buffers to perform the operations properly!
3278:      mmdata->mmBuffer2 does not appear anywhere in the compute/copy API
3279:      it only appears for the workEstimation stuff, but it seems it is needed in compute, so probably the address
3280:      is stored in the descriptor! What a messy API... */
3281:   PetscCallCUDA(cudaMalloc((void **)&mmdata->mmBuffer, mmdata->mmBufferSize));
3282:   /* compute the intermediate product of A * B */
3283:   stat = cusparseSpGEMM_compute(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &mmdata->mmBufferSize, mmdata->mmBuffer);
3284:   PetscCallCUSPARSE(stat);
3285:   /* get matrix C non-zero entries C_nnz1 */
3286:   PetscCallCUSPARSE(cusparseSpMatGetSize(Cmat->matDescr, &C_num_rows1, &C_num_cols1, &C_nnz1));
3287:   c->nz = (PetscInt)C_nnz1;
3288:   PetscCall(PetscInfo(C, "Buffer sizes for type %s, result %" PetscInt_FMT " x %" PetscInt_FMT " (k %" PetscInt_FMT ", nzA %" PetscInt_FMT ", nzB %" PetscInt_FMT ", nzC %" PetscInt_FMT ") are: %ldKB %ldKB\n", MatProductTypes[ptype], m, n, k, a->nz, b->nz, c->nz, bufSize2 / 1024,
3289:                       mmdata->mmBufferSize / 1024));
3290:   Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
3291:   PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3292:   Ccsr->values = new THRUSTARRAY(c->nz);
3293:   PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3294:   stat = cusparseCsrSetPointers(Cmat->matDescr, Ccsr->row_offsets->data().get(), Ccsr->column_indices->data().get(), Ccsr->values->data().get());
3295:   PetscCallCUSPARSE(stat);
3296:   stat = cusparseSpGEMM_copy(Ccusp->handle, opA, opB, Cmat->alpha_one, Amat->matDescr, BmatSpDescr, Cmat->beta_zero, Cmat->matDescr, cusparse_scalartype, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc);
3297:   PetscCallCUSPARSE(stat);
3298:   #endif // PETSC_PKG_CUDA_VERSION_GE(11,4,0)
3299: #else
3300:   PetscCallCUSPARSE(cusparseSetPointerMode(Ccusp->handle, CUSPARSE_POINTER_MODE_HOST));
3301:   stat = cusparseXcsrgemmNnz(Ccusp->handle, opA, opB, Acsr->num_rows, Bcsr->num_cols, Acsr->num_cols, Amat->descr, Acsr->num_entries, Acsr->row_offsets->data().get(), Acsr->column_indices->data().get(), Bmat->descr, Bcsr->num_entries,
3302:                              Bcsr->row_offsets->data().get(), Bcsr->column_indices->data().get(), Cmat->descr, Ccsr->row_offsets->data().get(), &cnz);
3303:   PetscCallCUSPARSE(stat);
3304:   c->nz                = cnz;
3305:   Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
3306:   PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3307:   Ccsr->values = new THRUSTARRAY(c->nz);
3308:   PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */

3310:   PetscCallCUSPARSE(cusparseSetPointerMode(Ccusp->handle, CUSPARSE_POINTER_MODE_DEVICE));
3311:   /* with the old gemm interface (removed from 11.0 on) we cannot compute the symbolic factorization only.
3312:      I have tried using the gemm2 interface (alpha * A * B + beta * D), which allows to do symbolic by passing NULL for values, but it seems quite buggy when
3313:      D is NULL, despite the fact that CUSPARSE documentation claims it is supported! */
3314:   stat = cusparse_csr_spgemm(Ccusp->handle, opA, opB, Acsr->num_rows, Bcsr->num_cols, Acsr->num_cols, Amat->descr, Acsr->num_entries, Acsr->values->data().get(), Acsr->row_offsets->data().get(), Acsr->column_indices->data().get(), Bmat->descr, Bcsr->num_entries,
3315:                              Bcsr->values->data().get(), Bcsr->row_offsets->data().get(), Bcsr->column_indices->data().get(), Cmat->descr, Ccsr->values->data().get(), Ccsr->row_offsets->data().get(), Ccsr->column_indices->data().get());
3316:   PetscCallCUSPARSE(stat);
3317: #endif
3318:   PetscCall(PetscLogGpuFlops(mmdata->flops));
3319:   PetscCall(PetscLogGpuTimeEnd());
3320: finalizesym:
3321:   c->singlemalloc = PETSC_FALSE;
3322:   c->free_a       = PETSC_TRUE;
3323:   c->free_ij      = PETSC_TRUE;
3324:   PetscCall(PetscMalloc1(m + 1, &c->i));
3325:   PetscCall(PetscMalloc1(c->nz, &c->j));
3326:   if (PetscDefined(USE_64BIT_INDICES)) { /* 32 to 64-bit conversion on the GPU and then copy to host (lazy) */
3327:     PetscInt      *d_i = c->i;
3328:     THRUSTINTARRAY ii(Ccsr->row_offsets->size());
3329:     THRUSTINTARRAY jj(Ccsr->column_indices->size());
3330:     ii = *Ccsr->row_offsets;
3331:     jj = *Ccsr->column_indices;
3332:     if (ciscompressed) d_i = c->compressedrow.i;
3333:     PetscCallCUDA(cudaMemcpy(d_i, ii.data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
3334:     PetscCallCUDA(cudaMemcpy(c->j, jj.data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
3335:   } else {
3336:     PetscInt *d_i = c->i;
3337:     if (ciscompressed) d_i = c->compressedrow.i;
3338:     PetscCallCUDA(cudaMemcpy(d_i, Ccsr->row_offsets->data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
3339:     PetscCallCUDA(cudaMemcpy(c->j, Ccsr->column_indices->data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
3340:   }
3341:   if (ciscompressed) { /* need to expand host row offsets */
3342:     PetscInt r = 0;
3343:     c->i[0]    = 0;
3344:     for (k = 0; k < c->compressedrow.nrows; k++) {
3345:       const PetscInt next = c->compressedrow.rindex[k];
3346:       const PetscInt old  = c->compressedrow.i[k];
3347:       for (; r < next; r++) c->i[r + 1] = old;
3348:     }
3349:     for (; r < m; r++) c->i[r + 1] = c->compressedrow.i[c->compressedrow.nrows];
3350:   }
3351:   PetscCall(PetscLogGpuToCpu((Ccsr->column_indices->size() + Ccsr->row_offsets->size()) * sizeof(PetscInt)));
3352:   PetscCall(PetscMalloc1(m, &c->ilen));
3353:   PetscCall(PetscMalloc1(m, &c->imax));
3354:   c->maxnz         = c->nz;
3355:   c->nonzerorowcnt = 0;
3356:   c->rmax          = 0;
3357:   for (k = 0; k < m; k++) {
3358:     const PetscInt nn = c->i[k + 1] - c->i[k];
3359:     c->ilen[k] = c->imax[k] = nn;
3360:     c->nonzerorowcnt += (PetscInt) !!nn;
3361:     c->rmax = PetscMax(c->rmax, nn);
3362:   }
3363:   PetscCall(MatMarkDiagonal_SeqAIJ(C));
3364:   PetscCall(PetscMalloc1(c->nz, &c->a));
3365:   Ccsr->num_entries = c->nz;

3367:   C->nonzerostate++;
3368:   PetscCall(PetscLayoutSetUp(C->rmap));
3369:   PetscCall(PetscLayoutSetUp(C->cmap));
3370:   Ccusp->nonzerostate = C->nonzerostate;
3371:   C->offloadmask      = PETSC_OFFLOAD_UNALLOCATED;
3372:   C->preallocated     = PETSC_TRUE;
3373:   C->assembled        = PETSC_FALSE;
3374:   C->was_assembled    = PETSC_FALSE;
3375:   if (product->api_user && A->offloadmask == PETSC_OFFLOAD_BOTH && B->offloadmask == PETSC_OFFLOAD_BOTH) { /* flag the matrix C values as computed, so that the numeric phase will only call MatAssembly */
3376:     mmdata->reusesym = PETSC_TRUE;
3377:     C->offloadmask   = PETSC_OFFLOAD_GPU;
3378:   }
3379:   C->ops->productnumeric = MatProductNumeric_SeqAIJCUSPARSE_SeqAIJCUSPARSE;
3380:   PetscFunctionReturn(PETSC_SUCCESS);
3381: }

3383: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ_SeqDense(Mat);

3385: /* handles sparse or dense B */
3386: static PetscErrorCode MatProductSetFromOptions_SeqAIJCUSPARSE(Mat mat)
3387: {
3388:   Mat_Product *product = mat->product;
3389:   PetscBool    isdense = PETSC_FALSE, Biscusp = PETSC_FALSE, Ciscusp = PETSC_TRUE;

3391:   PetscFunctionBegin;
3392:   MatCheckProduct(mat, 1);
3393:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)product->B, MATSEQDENSE, &isdense));
3394:   if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, MATSEQAIJCUSPARSE, &Biscusp));
3395:   if (product->type == MATPRODUCT_ABC) {
3396:     Ciscusp = PETSC_FALSE;
3397:     if (!product->C->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->C, MATSEQAIJCUSPARSE, &Ciscusp));
3398:   }
3399:   if (Biscusp && Ciscusp) { /* we can always select the CPU backend */
3400:     PetscBool usecpu = PETSC_FALSE;
3401:     switch (product->type) {
3402:     case MATPRODUCT_AB:
3403:       if (product->api_user) {
3404:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
3405:         PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
3406:         PetscOptionsEnd();
3407:       } else {
3408:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
3409:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
3410:         PetscOptionsEnd();
3411:       }
3412:       break;
3413:     case MATPRODUCT_AtB:
3414:       if (product->api_user) {
3415:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
3416:         PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
3417:         PetscOptionsEnd();
3418:       } else {
3419:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
3420:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
3421:         PetscOptionsEnd();
3422:       }
3423:       break;
3424:     case MATPRODUCT_PtAP:
3425:       if (product->api_user) {
3426:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
3427:         PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
3428:         PetscOptionsEnd();
3429:       } else {
3430:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
3431:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
3432:         PetscOptionsEnd();
3433:       }
3434:       break;
3435:     case MATPRODUCT_RARt:
3436:       if (product->api_user) {
3437:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatRARt", "Mat");
3438:         PetscCall(PetscOptionsBool("-matrart_backend_cpu", "Use CPU code", "MatRARt", usecpu, &usecpu, NULL));
3439:         PetscOptionsEnd();
3440:       } else {
3441:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_RARt", "Mat");
3442:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatRARt", usecpu, &usecpu, NULL));
3443:         PetscOptionsEnd();
3444:       }
3445:       break;
3446:     case MATPRODUCT_ABC:
3447:       if (product->api_user) {
3448:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMatMult", "Mat");
3449:         PetscCall(PetscOptionsBool("-matmatmatmult_backend_cpu", "Use CPU code", "MatMatMatMult", usecpu, &usecpu, NULL));
3450:         PetscOptionsEnd();
3451:       } else {
3452:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_ABC", "Mat");
3453:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMatMult", usecpu, &usecpu, NULL));
3454:         PetscOptionsEnd();
3455:       }
3456:       break;
3457:     default:
3458:       break;
3459:     }
3460:     if (usecpu) Biscusp = Ciscusp = PETSC_FALSE;
3461:   }
3462:   /* dispatch */
3463:   if (isdense) {
3464:     switch (product->type) {
3465:     case MATPRODUCT_AB:
3466:     case MATPRODUCT_AtB:
3467:     case MATPRODUCT_ABt:
3468:     case MATPRODUCT_PtAP:
3469:     case MATPRODUCT_RARt:
3470:       if (product->A->boundtocpu) {
3471:         PetscCall(MatProductSetFromOptions_SeqAIJ_SeqDense(mat));
3472:       } else {
3473:         mat->ops->productsymbolic = MatProductSymbolic_SeqAIJCUSPARSE_SeqDENSECUDA;
3474:       }
3475:       break;
3476:     case MATPRODUCT_ABC:
3477:       mat->ops->productsymbolic = MatProductSymbolic_ABC_Basic;
3478:       break;
3479:     default:
3480:       break;
3481:     }
3482:   } else if (Biscusp && Ciscusp) {
3483:     switch (product->type) {
3484:     case MATPRODUCT_AB:
3485:     case MATPRODUCT_AtB:
3486:     case MATPRODUCT_ABt:
3487:       mat->ops->productsymbolic = MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE;
3488:       break;
3489:     case MATPRODUCT_PtAP:
3490:     case MATPRODUCT_RARt:
3491:     case MATPRODUCT_ABC:
3492:       mat->ops->productsymbolic = MatProductSymbolic_ABC_Basic;
3493:       break;
3494:     default:
3495:       break;
3496:     }
3497:   } else { /* fallback for AIJ */
3498:     PetscCall(MatProductSetFromOptions_SeqAIJ(mat));
3499:   }
3500:   PetscFunctionReturn(PETSC_SUCCESS);
3501: }

3503: static PetscErrorCode MatMult_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3504: {
3505:   PetscFunctionBegin;
3506:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_FALSE, PETSC_FALSE));
3507:   PetscFunctionReturn(PETSC_SUCCESS);
3508: }

3510: static PetscErrorCode MatMultAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3511: {
3512:   PetscFunctionBegin;
3513:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_FALSE, PETSC_FALSE));
3514:   PetscFunctionReturn(PETSC_SUCCESS);
3515: }

3517: static PetscErrorCode MatMultHermitianTranspose_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3518: {
3519:   PetscFunctionBegin;
3520:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_TRUE, PETSC_TRUE));
3521:   PetscFunctionReturn(PETSC_SUCCESS);
3522: }

3524: static PetscErrorCode MatMultHermitianTransposeAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3525: {
3526:   PetscFunctionBegin;
3527:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_TRUE, PETSC_TRUE));
3528:   PetscFunctionReturn(PETSC_SUCCESS);
3529: }

3531: static PetscErrorCode MatMultTranspose_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3532: {
3533:   PetscFunctionBegin;
3534:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_TRUE, PETSC_FALSE));
3535:   PetscFunctionReturn(PETSC_SUCCESS);
3536: }

3538: __global__ static void ScatterAdd(PetscInt n, PetscInt *idx, const PetscScalar *x, PetscScalar *y)
3539: {
3540:   int i = blockIdx.x * blockDim.x + threadIdx.x;
3541:   if (i < n) y[idx[i]] += x[i];
3542: }

3544: /* z = op(A) x + y. If trans & !herm, op = ^T; if trans & herm, op = ^H; if !trans, op = no-op */
3545: static PetscErrorCode MatMultAddKernel_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz, PetscBool trans, PetscBool herm)
3546: {
3547:   Mat_SeqAIJ                   *a              = (Mat_SeqAIJ *)A->data;
3548:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
3549:   Mat_SeqAIJCUSPARSEMultStruct *matstruct;
3550:   PetscScalar                  *xarray, *zarray, *dptr, *beta, *xptr;
3551:   cusparseOperation_t           opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
3552:   PetscBool                     compressed;
3553: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3554:   PetscInt nx, ny;
3555: #endif

3557:   PetscFunctionBegin;
3558:   PetscCheck(!herm || trans, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Hermitian and not transpose not supported");
3559:   if (!a->nz) {
3560:     if (yy) PetscCall(VecSeq_CUDA::Copy(yy, zz));
3561:     else PetscCall(VecSeq_CUDA::Set(zz, 0));
3562:     PetscFunctionReturn(PETSC_SUCCESS);
3563:   }
3564:   /* The line below is necessary due to the operations that modify the matrix on the CPU (axpy, scale, etc) */
3565:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
3566:   if (!trans) {
3567:     matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
3568:     PetscCheck(matstruct, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "SeqAIJCUSPARSE does not have a 'mat' (need to fix)");
3569:   } else {
3570:     if (herm || !A->form_explicit_transpose) {
3571:       opA       = herm ? CUSPARSE_OPERATION_CONJUGATE_TRANSPOSE : CUSPARSE_OPERATION_TRANSPOSE;
3572:       matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
3573:     } else {
3574:       if (!cusparsestruct->matTranspose) PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
3575:       matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->matTranspose;
3576:     }
3577:   }
3578:   /* Does the matrix use compressed rows (i.e., drop zero rows)? */
3579:   compressed = matstruct->cprowIndices ? PETSC_TRUE : PETSC_FALSE;

3581:   try {
3582:     PetscCall(VecCUDAGetArrayRead(xx, (const PetscScalar **)&xarray));
3583:     if (yy == zz) PetscCall(VecCUDAGetArray(zz, &zarray)); /* read & write zz, so need to get up-to-date zarray on GPU */
3584:     else PetscCall(VecCUDAGetArrayWrite(zz, &zarray));     /* write zz, so no need to init zarray on GPU */

3586:     PetscCall(PetscLogGpuTimeBegin());
3587:     if (opA == CUSPARSE_OPERATION_NON_TRANSPOSE) {
3588:       /* z = A x + beta y.
3589:          If A is compressed (with less rows), then Ax is shorter than the full z, so we need a work vector to store Ax.
3590:          When A is non-compressed, and z = y, we can set beta=1 to compute y = Ax + y in one call.
3591:       */
3592:       xptr = xarray;
3593:       dptr = compressed ? cusparsestruct->workVector->data().get() : zarray;
3594:       beta = (yy == zz && !compressed) ? matstruct->beta_one : matstruct->beta_zero;
3595: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3596:       /* Get length of x, y for y=Ax. ny might be shorter than the work vector's allocated length, since the work vector is
3597:           allocated to accommodate different uses. So we get the length info directly from mat.
3598:        */
3599:       if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3600:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3601:         nx             = mat->num_cols; // since y = Ax
3602:         ny             = mat->num_rows;
3603:       }
3604: #endif
3605:     } else {
3606:       /* z = A^T x + beta y
3607:          If A is compressed, then we need a work vector as the shorter version of x to compute A^T x.
3608:          Note A^Tx is of full length, so we set beta to 1.0 if y exists.
3609:        */
3610:       xptr = compressed ? cusparsestruct->workVector->data().get() : xarray;
3611:       dptr = zarray;
3612:       beta = yy ? matstruct->beta_one : matstruct->beta_zero;
3613:       if (compressed) { /* Scatter x to work vector */
3614:         thrust::device_ptr<PetscScalar> xarr = thrust::device_pointer_cast(xarray);

3616:         thrust::for_each(
3617: #if PetscDefined(HAVE_THRUST_ASYNC)
3618:           thrust::cuda::par.on(PetscDefaultCudaStream),
3619: #endif
3620:           thrust::make_zip_iterator(thrust::make_tuple(cusparsestruct->workVector->begin(), thrust::make_permutation_iterator(xarr, matstruct->cprowIndices->begin()))),
3621:           thrust::make_zip_iterator(thrust::make_tuple(cusparsestruct->workVector->begin(), thrust::make_permutation_iterator(xarr, matstruct->cprowIndices->begin()))) + matstruct->cprowIndices->size(), VecCUDAEqualsReverse());
3622:       }
3623: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3624:       if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3625:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3626:         nx             = mat->num_rows; // since y = A^T x
3627:         ny             = mat->num_cols;
3628:       }
3629: #endif
3630:     }

3632:     /* csr_spmv does y = alpha op(A) x + beta y */
3633:     if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3634: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3635:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
3636:       cusparseSpMatDescr_t &matDescr = matstruct->matDescr_SpMV[opA]; // All opA's should use the same matDescr, but the cusparse issue/bug (#212) after 12.4 forced us to create a new one for each opA.
3637:   #else
3638:       cusparseSpMatDescr_t &matDescr = matstruct->matDescr;
3639:   #endif

3641:       PetscCheck(opA >= 0 && opA <= 2, PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSPARSE ABI on cusparseOperation_t has changed and PETSc has not been updated accordingly");
3642:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
3643:       if (!matDescr) {
3644:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3645:         PetscCallCUSPARSE(cusparseCreateCsr(&matDescr, mat->num_rows, mat->num_cols, mat->num_entries, mat->row_offsets->data().get(), mat->column_indices->data().get(), mat->values->data().get(), CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype));
3646:       }
3647:   #endif

3649:       if (!matstruct->cuSpMV[opA].initialized) { /* built on demand */
3650:         PetscCallCUSPARSE(cusparseCreateDnVec(&matstruct->cuSpMV[opA].vecXDescr, nx, xptr, cusparse_scalartype));
3651:         PetscCallCUSPARSE(cusparseCreateDnVec(&matstruct->cuSpMV[opA].vecYDescr, ny, dptr, cusparse_scalartype));
3652:         PetscCallCUSPARSE(
3653:           cusparseSpMV_bufferSize(cusparsestruct->handle, opA, matstruct->alpha_one, matDescr, matstruct->cuSpMV[opA].vecXDescr, beta, matstruct->cuSpMV[opA].vecYDescr, cusparse_scalartype, cusparsestruct->spmvAlg, &matstruct->cuSpMV[opA].spmvBufferSize));
3654:         PetscCallCUDA(cudaMalloc(&matstruct->cuSpMV[opA].spmvBuffer, matstruct->cuSpMV[opA].spmvBufferSize));
3655:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0) // cusparseSpMV_preprocess is added in 12.4
3656:         PetscCallCUSPARSE(
3657:           cusparseSpMV_preprocess(cusparsestruct->handle, opA, matstruct->alpha_one, matDescr, matstruct->cuSpMV[opA].vecXDescr, beta, matstruct->cuSpMV[opA].vecYDescr, cusparse_scalartype, cusparsestruct->spmvAlg, matstruct->cuSpMV[opA].spmvBuffer));
3658:   #endif
3659:         matstruct->cuSpMV[opA].initialized = PETSC_TRUE;
3660:       } else {
3661:         /* x, y's value pointers might change between calls, but their shape is kept, so we just update pointers */
3662:         PetscCallCUSPARSE(cusparseDnVecSetValues(matstruct->cuSpMV[opA].vecXDescr, xptr));
3663:         PetscCallCUSPARSE(cusparseDnVecSetValues(matstruct->cuSpMV[opA].vecYDescr, dptr));
3664:       }

3666:       PetscCallCUSPARSE(cusparseSpMV(cusparsestruct->handle, opA, matstruct->alpha_one, matDescr, matstruct->cuSpMV[opA].vecXDescr, beta, matstruct->cuSpMV[opA].vecYDescr, cusparse_scalartype, cusparsestruct->spmvAlg, matstruct->cuSpMV[opA].spmvBuffer));
3667: #else
3668:       CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3669:       PetscCallCUSPARSE(cusparse_csr_spmv(cusparsestruct->handle, opA, mat->num_rows, mat->num_cols, mat->num_entries, matstruct->alpha_one, matstruct->descr, mat->values->data().get(), mat->row_offsets->data().get(), mat->column_indices->data().get(), xptr, beta, dptr));
3670: #endif
3671:     } else {
3672:       if (cusparsestruct->nrows) {
3673: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3674:         SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
3675: #else
3676:         cusparseHybMat_t hybMat = (cusparseHybMat_t)matstruct->mat;
3677:         PetscCallCUSPARSE(cusparse_hyb_spmv(cusparsestruct->handle, opA, matstruct->alpha_one, matstruct->descr, hybMat, xptr, beta, dptr));
3678: #endif
3679:       }
3680:     }
3681:     PetscCall(PetscLogGpuTimeEnd());

3683:     if (opA == CUSPARSE_OPERATION_NON_TRANSPOSE) {
3684:       if (yy) {                                      /* MatMultAdd: zz = A*xx + yy */
3685:         if (compressed) {                            /* A is compressed. We first copy yy to zz, then ScatterAdd the work vector to zz */
3686:           PetscCall(VecSeq_CUDA::Copy(yy, zz));      /* zz = yy */
3687:         } else if (zz != yy) {                       /* A is not compressed. zz already contains A*xx, and we just need to add yy */
3688:           PetscCall(VecSeq_CUDA::AXPY(zz, 1.0, yy)); /* zz += yy */
3689:         }
3690:       } else if (compressed) { /* MatMult: zz = A*xx. A is compressed, so we zero zz first, then ScatterAdd the work vector to zz */
3691:         PetscCall(VecSeq_CUDA::Set(zz, 0));
3692:       }

3694:       /* ScatterAdd the result from work vector into the full vector when A is compressed */
3695:       if (compressed) {
3696:         PetscCall(PetscLogGpuTimeBegin());
3697:         /* I wanted to make this for_each asynchronous but failed. thrust::async::for_each() returns an event (internally registered)
3698:            and in the destructor of the scope, it will call cudaStreamSynchronize() on this stream. One has to store all events to
3699:            prevent that. So I just add a ScatterAdd kernel.
3700:          */
3701: #if 0
3702:         thrust::device_ptr<PetscScalar> zptr = thrust::device_pointer_cast(zarray);
3703:         thrust::async::for_each(thrust::cuda::par.on(cusparsestruct->stream),
3704:                          thrust::make_zip_iterator(thrust::make_tuple(cusparsestruct->workVector->begin(), thrust::make_permutation_iterator(zptr, matstruct->cprowIndices->begin()))),
3705:                          thrust::make_zip_iterator(thrust::make_tuple(cusparsestruct->workVector->begin(), thrust::make_permutation_iterator(zptr, matstruct->cprowIndices->begin()))) + matstruct->cprowIndices->size(),
3706:                          VecCUDAPlusEquals());
3707: #else
3708:         PetscInt n = matstruct->cprowIndices->size();
3709:         ScatterAdd<<<(n + 255) / 256, 256, 0, PetscDefaultCudaStream>>>(n, matstruct->cprowIndices->data().get(), cusparsestruct->workVector->data().get(), zarray);
3710: #endif
3711:         PetscCall(PetscLogGpuTimeEnd());
3712:       }
3713:     } else {
3714:       if (yy && yy != zz) PetscCall(VecSeq_CUDA::AXPY(zz, 1.0, yy)); /* zz += yy */
3715:     }
3716:     PetscCall(VecCUDARestoreArrayRead(xx, (const PetscScalar **)&xarray));
3717:     if (yy == zz) PetscCall(VecCUDARestoreArray(zz, &zarray));
3718:     else PetscCall(VecCUDARestoreArrayWrite(zz, &zarray));
3719:   } catch (char *ex) {
3720:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
3721:   }
3722:   if (yy) {
3723:     PetscCall(PetscLogGpuFlops(2.0 * a->nz));
3724:   } else {
3725:     PetscCall(PetscLogGpuFlops(2.0 * a->nz - a->nonzerorowcnt));
3726:   }
3727:   PetscFunctionReturn(PETSC_SUCCESS);
3728: }

3730: static PetscErrorCode MatMultTransposeAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3731: {
3732:   PetscFunctionBegin;
3733:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_TRUE, PETSC_FALSE));
3734:   PetscFunctionReturn(PETSC_SUCCESS);
3735: }

3737: static PetscErrorCode MatAssemblyEnd_SeqAIJCUSPARSE(Mat A, MatAssemblyType mode)
3738: {
3739:   PetscFunctionBegin;
3740:   PetscCall(MatAssemblyEnd_SeqAIJ(A, mode));
3741:   PetscFunctionReturn(PETSC_SUCCESS);
3742: }

3744: /*@
3745:   MatCreateSeqAIJCUSPARSE - Creates a sparse matrix in `MATAIJCUSPARSE` (compressed row) format
3746:   (the default parallel PETSc format).

3748:   Collective

3750:   Input Parameters:
3751: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3752: . m    - number of rows
3753: . n    - number of columns
3754: . nz   - number of nonzeros per row (same for all rows), ignored if `nnz` is provide
3755: - nnz  - array containing the number of nonzeros in the various rows (possibly different for each row) or `NULL`

3757:   Output Parameter:
3758: . A - the matrix

3760:   Level: intermediate

3762:   Notes:
3763:   This matrix will ultimately pushed down to NVIDIA GPUs and use the CuSPARSE library for
3764:   calculations. For good matrix assembly performance the user should preallocate the matrix
3765:   storage by setting the parameter `nz` (or the array `nnz`).

3767:   It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
3768:   MatXXXXSetPreallocation() paradgm instead of this routine directly.
3769:   [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]

3771:   The AIJ format, also called
3772:   compressed row storage, is fully compatible with standard Fortran
3773:   storage.  That is, the stored row and column indices can begin at
3774:   either one (as in Fortran) or zero.

3776:   Specify the preallocated storage with either nz or nnz (not both).
3777:   Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3778:   allocation.

3780: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJCUSPARSE`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MATAIJCUSPARSE`
3781: @*/
3782: PetscErrorCode MatCreateSeqAIJCUSPARSE(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3783: {
3784:   PetscFunctionBegin;
3785:   PetscCall(MatCreate(comm, A));
3786:   PetscCall(MatSetSizes(*A, m, n, m, n));
3787:   PetscCall(MatSetType(*A, MATSEQAIJCUSPARSE));
3788:   PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, (PetscInt *)nnz));
3789:   PetscFunctionReturn(PETSC_SUCCESS);
3790: }

3792: static PetscErrorCode MatDestroy_SeqAIJCUSPARSE(Mat A)
3793: {
3794:   PetscFunctionBegin;
3795:   if (A->factortype == MAT_FACTOR_NONE) {
3796:     PetscCall(MatSeqAIJCUSPARSE_Destroy(A));
3797:   } else {
3798:     PetscCall(MatSeqAIJCUSPARSETriFactors_Destroy((Mat_SeqAIJCUSPARSETriFactors **)&A->spptr));
3799:   }
3800:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", NULL));
3801:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatCUSPARSESetFormat_C", NULL));
3802:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatCUSPARSESetUseCPUSolve_C", NULL));
3803:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", NULL));
3804:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", NULL));
3805:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", NULL));
3806:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
3807:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
3808:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
3809:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijcusparse_hypre_C", NULL));
3810:   PetscCall(MatDestroy_SeqAIJ(A));
3811:   PetscFunctionReturn(PETSC_SUCCESS);
3812: }

3814: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
3815: static PetscErrorCode       MatBindToCPU_SeqAIJCUSPARSE(Mat, PetscBool);
3816: static PetscErrorCode       MatDuplicate_SeqAIJCUSPARSE(Mat A, MatDuplicateOption cpvalues, Mat *B)
3817: {
3818:   PetscFunctionBegin;
3819:   PetscCall(MatDuplicate_SeqAIJ(A, cpvalues, B));
3820:   PetscCall(MatConvert_SeqAIJ_SeqAIJCUSPARSE(*B, MATSEQAIJCUSPARSE, MAT_INPLACE_MATRIX, B));
3821:   PetscFunctionReturn(PETSC_SUCCESS);
3822: }

3824: static PetscErrorCode MatAXPY_SeqAIJCUSPARSE(Mat Y, PetscScalar a, Mat X, MatStructure str)
3825: {
3826:   Mat_SeqAIJ         *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data;
3827:   Mat_SeqAIJCUSPARSE *cy;
3828:   Mat_SeqAIJCUSPARSE *cx;
3829:   PetscScalar        *ay;
3830:   const PetscScalar  *ax;
3831:   CsrMatrix          *csry, *csrx;

3833:   PetscFunctionBegin;
3834:   cy = (Mat_SeqAIJCUSPARSE *)Y->spptr;
3835:   cx = (Mat_SeqAIJCUSPARSE *)X->spptr;
3836:   if (X->ops->axpy != Y->ops->axpy) {
3837:     PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(Y, PETSC_FALSE));
3838:     PetscCall(MatAXPY_SeqAIJ(Y, a, X, str));
3839:     PetscFunctionReturn(PETSC_SUCCESS);
3840:   }
3841:   /* if we are here, it means both matrices are bound to GPU */
3842:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(Y));
3843:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(X));
3844:   PetscCheck(cy->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)Y), PETSC_ERR_GPU, "only MAT_CUSPARSE_CSR supported");
3845:   PetscCheck(cx->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)X), PETSC_ERR_GPU, "only MAT_CUSPARSE_CSR supported");
3846:   csry = (CsrMatrix *)cy->mat->mat;
3847:   csrx = (CsrMatrix *)cx->mat->mat;
3848:   /* see if we can turn this into a cublas axpy */
3849:   if (str != SAME_NONZERO_PATTERN && x->nz == y->nz && !x->compressedrow.use && !y->compressedrow.use) {
3850:     bool eq = thrust::equal(thrust::device, csry->row_offsets->begin(), csry->row_offsets->end(), csrx->row_offsets->begin());
3851:     if (eq) eq = thrust::equal(thrust::device, csry->column_indices->begin(), csry->column_indices->end(), csrx->column_indices->begin());
3852:     if (eq) str = SAME_NONZERO_PATTERN;
3853:   }
3854:   /* spgeam is buggy with one column */
3855:   if (Y->cmap->n == 1 && str != SAME_NONZERO_PATTERN) str = DIFFERENT_NONZERO_PATTERN;

3857:   if (str == SUBSET_NONZERO_PATTERN) {
3858:     PetscScalar b = 1.0;
3859: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3860:     size_t bufferSize;
3861:     void  *buffer;
3862: #endif

3864:     PetscCall(MatSeqAIJCUSPARSEGetArrayRead(X, &ax));
3865:     PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3866:     PetscCallCUSPARSE(cusparseSetPointerMode(cy->handle, CUSPARSE_POINTER_MODE_HOST));
3867: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3868:     PetscCallCUSPARSE(cusparse_csr_spgeam_bufferSize(cy->handle, Y->rmap->n, Y->cmap->n, &a, cx->mat->descr, x->nz, ax, csrx->row_offsets->data().get(), csrx->column_indices->data().get(), &b, cy->mat->descr, y->nz, ay, csry->row_offsets->data().get(),
3869:                                                      csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get(), &bufferSize));
3870:     PetscCallCUDA(cudaMalloc(&buffer, bufferSize));
3871:     PetscCall(PetscLogGpuTimeBegin());
3872:     PetscCallCUSPARSE(cusparse_csr_spgeam(cy->handle, Y->rmap->n, Y->cmap->n, &a, cx->mat->descr, x->nz, ax, csrx->row_offsets->data().get(), csrx->column_indices->data().get(), &b, cy->mat->descr, y->nz, ay, csry->row_offsets->data().get(),
3873:                                           csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get(), buffer));
3874:     PetscCall(PetscLogGpuFlops(x->nz + y->nz));
3875:     PetscCall(PetscLogGpuTimeEnd());
3876:     PetscCallCUDA(cudaFree(buffer));
3877: #else
3878:     PetscCall(PetscLogGpuTimeBegin());
3879:     PetscCallCUSPARSE(cusparse_csr_spgeam(cy->handle, Y->rmap->n, Y->cmap->n, &a, cx->mat->descr, x->nz, ax, csrx->row_offsets->data().get(), csrx->column_indices->data().get(), &b, cy->mat->descr, y->nz, ay, csry->row_offsets->data().get(),
3880:                                           csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get()));
3881:     PetscCall(PetscLogGpuFlops(x->nz + y->nz));
3882:     PetscCall(PetscLogGpuTimeEnd());
3883: #endif
3884:     PetscCallCUSPARSE(cusparseSetPointerMode(cy->handle, CUSPARSE_POINTER_MODE_DEVICE));
3885:     PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(X, &ax));
3886:     PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3887:     PetscCall(MatSeqAIJInvalidateDiagonal(Y));
3888:   } else if (str == SAME_NONZERO_PATTERN) {
3889:     cublasHandle_t cublasv2handle;
3890:     PetscBLASInt   one = 1, bnz = 1;

3892:     PetscCall(MatSeqAIJCUSPARSEGetArrayRead(X, &ax));
3893:     PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3894:     PetscCall(PetscCUBLASGetHandle(&cublasv2handle));
3895:     PetscCall(PetscBLASIntCast(x->nz, &bnz));
3896:     PetscCall(PetscLogGpuTimeBegin());
3897:     PetscCallCUBLAS(cublasXaxpy(cublasv2handle, bnz, &a, ax, one, ay, one));
3898:     PetscCall(PetscLogGpuFlops(2.0 * bnz));
3899:     PetscCall(PetscLogGpuTimeEnd());
3900:     PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(X, &ax));
3901:     PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3902:     PetscCall(MatSeqAIJInvalidateDiagonal(Y));
3903:   } else {
3904:     PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(Y, PETSC_FALSE));
3905:     PetscCall(MatAXPY_SeqAIJ(Y, a, X, str));
3906:   }
3907:   PetscFunctionReturn(PETSC_SUCCESS);
3908: }

3910: static PetscErrorCode MatScale_SeqAIJCUSPARSE(Mat Y, PetscScalar a)
3911: {
3912:   Mat_SeqAIJ    *y = (Mat_SeqAIJ *)Y->data;
3913:   PetscScalar   *ay;
3914:   cublasHandle_t cublasv2handle;
3915:   PetscBLASInt   one = 1, bnz = 1;

3917:   PetscFunctionBegin;
3918:   PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3919:   PetscCall(PetscCUBLASGetHandle(&cublasv2handle));
3920:   PetscCall(PetscBLASIntCast(y->nz, &bnz));
3921:   PetscCall(PetscLogGpuTimeBegin());
3922:   PetscCallCUBLAS(cublasXscal(cublasv2handle, bnz, &a, ay, one));
3923:   PetscCall(PetscLogGpuFlops(bnz));
3924:   PetscCall(PetscLogGpuTimeEnd());
3925:   PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3926:   PetscCall(MatSeqAIJInvalidateDiagonal(Y));
3927:   PetscFunctionReturn(PETSC_SUCCESS);
3928: }

3930: static PetscErrorCode MatZeroEntries_SeqAIJCUSPARSE(Mat A)
3931: {
3932:   PetscBool   both = PETSC_FALSE;
3933:   Mat_SeqAIJ *a    = (Mat_SeqAIJ *)A->data;

3935:   PetscFunctionBegin;
3936:   if (A->factortype == MAT_FACTOR_NONE) {
3937:     Mat_SeqAIJCUSPARSE *spptr = (Mat_SeqAIJCUSPARSE *)A->spptr;
3938:     if (spptr->mat) {
3939:       CsrMatrix *matrix = (CsrMatrix *)spptr->mat->mat;
3940:       if (matrix->values) {
3941:         both = PETSC_TRUE;
3942:         thrust::fill(thrust::device, matrix->values->begin(), matrix->values->end(), 0.);
3943:       }
3944:     }
3945:     if (spptr->matTranspose) {
3946:       CsrMatrix *matrix = (CsrMatrix *)spptr->matTranspose->mat;
3947:       if (matrix->values) thrust::fill(thrust::device, matrix->values->begin(), matrix->values->end(), 0.);
3948:     }
3949:   }
3950:   PetscCall(PetscArrayzero(a->a, a->i[A->rmap->n]));
3951:   PetscCall(MatSeqAIJInvalidateDiagonal(A));
3952:   if (both) A->offloadmask = PETSC_OFFLOAD_BOTH;
3953:   else A->offloadmask = PETSC_OFFLOAD_CPU;
3954:   PetscFunctionReturn(PETSC_SUCCESS);
3955: }

3957: static PetscErrorCode MatBindToCPU_SeqAIJCUSPARSE(Mat A, PetscBool flg)
3958: {
3959:   Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;

3961:   PetscFunctionBegin;
3962:   if (A->factortype != MAT_FACTOR_NONE) {
3963:     A->boundtocpu = flg;
3964:     PetscFunctionReturn(PETSC_SUCCESS);
3965:   }
3966:   if (flg) {
3967:     PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));

3969:     A->ops->scale                     = MatScale_SeqAIJ;
3970:     A->ops->axpy                      = MatAXPY_SeqAIJ;
3971:     A->ops->zeroentries               = MatZeroEntries_SeqAIJ;
3972:     A->ops->mult                      = MatMult_SeqAIJ;
3973:     A->ops->multadd                   = MatMultAdd_SeqAIJ;
3974:     A->ops->multtranspose             = MatMultTranspose_SeqAIJ;
3975:     A->ops->multtransposeadd          = MatMultTransposeAdd_SeqAIJ;
3976:     A->ops->multhermitiantranspose    = NULL;
3977:     A->ops->multhermitiantransposeadd = NULL;
3978:     A->ops->productsetfromoptions     = MatProductSetFromOptions_SeqAIJ;
3979:     PetscCall(PetscMemzero(a->ops, sizeof(Mat_SeqAIJOps)));
3980:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", NULL));
3981:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", NULL));
3982:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", NULL));
3983:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
3984:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
3985:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", NULL));
3986:   } else {
3987:     A->ops->scale                     = MatScale_SeqAIJCUSPARSE;
3988:     A->ops->axpy                      = MatAXPY_SeqAIJCUSPARSE;
3989:     A->ops->zeroentries               = MatZeroEntries_SeqAIJCUSPARSE;
3990:     A->ops->mult                      = MatMult_SeqAIJCUSPARSE;
3991:     A->ops->multadd                   = MatMultAdd_SeqAIJCUSPARSE;
3992:     A->ops->multtranspose             = MatMultTranspose_SeqAIJCUSPARSE;
3993:     A->ops->multtransposeadd          = MatMultTransposeAdd_SeqAIJCUSPARSE;
3994:     A->ops->multhermitiantranspose    = MatMultHermitianTranspose_SeqAIJCUSPARSE;
3995:     A->ops->multhermitiantransposeadd = MatMultHermitianTransposeAdd_SeqAIJCUSPARSE;
3996:     A->ops->productsetfromoptions     = MatProductSetFromOptions_SeqAIJCUSPARSE;
3997:     a->ops->getarray                  = MatSeqAIJGetArray_SeqAIJCUSPARSE;
3998:     a->ops->restorearray              = MatSeqAIJRestoreArray_SeqAIJCUSPARSE;
3999:     a->ops->getarrayread              = MatSeqAIJGetArrayRead_SeqAIJCUSPARSE;
4000:     a->ops->restorearrayread          = MatSeqAIJRestoreArrayRead_SeqAIJCUSPARSE;
4001:     a->ops->getarraywrite             = MatSeqAIJGetArrayWrite_SeqAIJCUSPARSE;
4002:     a->ops->restorearraywrite         = MatSeqAIJRestoreArrayWrite_SeqAIJCUSPARSE;
4003:     a->ops->getcsrandmemtype          = MatSeqAIJGetCSRAndMemType_SeqAIJCUSPARSE;

4005:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", MatSeqAIJCopySubArray_SeqAIJCUSPARSE));
4006:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4007:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4008:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJCUSPARSE));
4009:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJCUSPARSE));
4010:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4011:   }
4012:   A->boundtocpu = flg;
4013:   if (flg && a->inode.size) {
4014:     a->inode.use = PETSC_TRUE;
4015:   } else {
4016:     a->inode.use = PETSC_FALSE;
4017:   }
4018:   PetscFunctionReturn(PETSC_SUCCESS);
4019: }

4021: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat A, MatType, MatReuse reuse, Mat *newmat)
4022: {
4023:   Mat B;

4025:   PetscFunctionBegin;
4026:   PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUDA)); /* first use of CUSPARSE may be via MatConvert */
4027:   if (reuse == MAT_INITIAL_MATRIX) {
4028:     PetscCall(MatDuplicate(A, MAT_COPY_VALUES, newmat));
4029:   } else if (reuse == MAT_REUSE_MATRIX) {
4030:     PetscCall(MatCopy(A, *newmat, SAME_NONZERO_PATTERN));
4031:   }
4032:   B = *newmat;

4034:   PetscCall(PetscFree(B->defaultvectype));
4035:   PetscCall(PetscStrallocpy(VECCUDA, &B->defaultvectype));

4037:   if (reuse != MAT_REUSE_MATRIX && !B->spptr) {
4038:     if (B->factortype == MAT_FACTOR_NONE) {
4039:       Mat_SeqAIJCUSPARSE *spptr;
4040:       PetscCall(PetscNew(&spptr));
4041:       PetscCallCUSPARSE(cusparseCreate(&spptr->handle));
4042:       PetscCallCUSPARSE(cusparseSetStream(spptr->handle, PetscDefaultCudaStream));
4043:       spptr->format = MAT_CUSPARSE_CSR;
4044: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4045:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
4046:       spptr->spmvAlg = CUSPARSE_SPMV_CSR_ALG1; /* default, since we only support csr */
4047:   #else
4048:       spptr->spmvAlg = CUSPARSE_CSRMV_ALG1; /* default, since we only support csr */
4049:   #endif
4050:       spptr->spmmAlg    = CUSPARSE_SPMM_CSR_ALG1; /* default, only support column-major dense matrix B */
4051:       spptr->csr2cscAlg = CUSPARSE_CSR2CSC_ALG1;
4052: #endif
4053:       B->spptr = spptr;
4054:     } else {
4055:       Mat_SeqAIJCUSPARSETriFactors *spptr;

4057:       PetscCall(PetscNew(&spptr));
4058:       PetscCallCUSPARSE(cusparseCreate(&spptr->handle));
4059:       PetscCallCUSPARSE(cusparseSetStream(spptr->handle, PetscDefaultCudaStream));
4060:       B->spptr = spptr;
4061:     }
4062:     B->offloadmask = PETSC_OFFLOAD_UNALLOCATED;
4063:   }
4064:   B->ops->assemblyend    = MatAssemblyEnd_SeqAIJCUSPARSE;
4065:   B->ops->destroy        = MatDestroy_SeqAIJCUSPARSE;
4066:   B->ops->setoption      = MatSetOption_SeqAIJCUSPARSE;
4067:   B->ops->setfromoptions = MatSetFromOptions_SeqAIJCUSPARSE;
4068:   B->ops->bindtocpu      = MatBindToCPU_SeqAIJCUSPARSE;
4069:   B->ops->duplicate      = MatDuplicate_SeqAIJCUSPARSE;

4071:   PetscCall(MatBindToCPU_SeqAIJCUSPARSE(B, PETSC_FALSE));
4072:   PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJCUSPARSE));
4073:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatCUSPARSESetFormat_C", MatCUSPARSESetFormat_SeqAIJCUSPARSE));
4074: #if defined(PETSC_HAVE_HYPRE)
4075:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaijcusparse_hypre_C", MatConvert_AIJ_HYPRE));
4076: #endif
4077:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatCUSPARSESetUseCPUSolve_C", MatCUSPARSESetUseCPUSolve_SeqAIJCUSPARSE));
4078:   PetscFunctionReturn(PETSC_SUCCESS);
4079: }

4081: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJCUSPARSE(Mat B)
4082: {
4083:   PetscFunctionBegin;
4084:   PetscCall(MatCreate_SeqAIJ(B));
4085:   PetscCall(MatConvert_SeqAIJ_SeqAIJCUSPARSE(B, MATSEQAIJCUSPARSE, MAT_INPLACE_MATRIX, &B));
4086:   PetscFunctionReturn(PETSC_SUCCESS);
4087: }

4089: /*MC
4090:    MATSEQAIJCUSPARSE - MATAIJCUSPARSE = "(seq)aijcusparse" - A matrix type to be used for sparse matrices.

4092:    A matrix type whose data resides on NVIDIA GPUs. These matrices can be in either
4093:    CSR, ELL, or Hybrid format.
4094:    All matrix calculations are performed on NVIDIA GPUs using the CuSPARSE library.

4096:    Options Database Keys:
4097: +  -mat_type aijcusparse - sets the matrix type to "seqaijcusparse" during a call to `MatSetFromOptions()`
4098: .  -mat_cusparse_storage_format csr - sets the storage format of matrices (for `MatMult()` and factors in `MatSolve()`).
4099:                                       Other options include ell (ellpack) or hyb (hybrid).
4100: .  -mat_cusparse_mult_storage_format csr - sets the storage format of matrices (for `MatMult()`). Other options include ell (ellpack) or hyb (hybrid).
4101: -  -mat_cusparse_use_cpu_solve - Do `MatSolve()` on CPU

4103:   Level: beginner

4105: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJCUSPARSE()`, `MatCUSPARSESetUseCPUSolve()`, `MATAIJCUSPARSE`, `MatCreateAIJCUSPARSE()`, `MatCUSPARSESetFormat()`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
4106: M*/

4108: PETSC_EXTERN PetscErrorCode MatSolverTypeRegister_CUSPARSE(void)
4109: {
4110:   PetscFunctionBegin;
4111:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_LU, MatGetFactor_seqaijcusparse_cusparse));
4112:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_CHOLESKY, MatGetFactor_seqaijcusparse_cusparse));
4113:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_ILU, MatGetFactor_seqaijcusparse_cusparse));
4114:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_ICC, MatGetFactor_seqaijcusparse_cusparse));
4115:   PetscFunctionReturn(PETSC_SUCCESS);
4116: }

4118: static PetscErrorCode MatSeqAIJCUSPARSE_Destroy(Mat mat)
4119: {
4120:   Mat_SeqAIJCUSPARSE *cusp = static_cast<Mat_SeqAIJCUSPARSE *>(mat->spptr);

4122:   PetscFunctionBegin;
4123:   if (cusp) {
4124:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->mat, cusp->format));
4125:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->matTranspose, cusp->format));
4126:     delete cusp->workVector;
4127:     delete cusp->rowoffsets_gpu;
4128:     delete cusp->csr2csc_i;
4129:     delete cusp->coords;
4130:     if (cusp->handle) PetscCallCUSPARSE(cusparseDestroy(cusp->handle));
4131:     PetscCall(PetscFree(mat->spptr));
4132:   }
4133:   PetscFunctionReturn(PETSC_SUCCESS);
4134: }

4136: static PetscErrorCode CsrMatrix_Destroy(CsrMatrix **mat)
4137: {
4138:   PetscFunctionBegin;
4139:   if (*mat) {
4140:     delete (*mat)->values;
4141:     delete (*mat)->column_indices;
4142:     delete (*mat)->row_offsets;
4143:     delete *mat;
4144:     *mat = 0;
4145:   }
4146:   PetscFunctionReturn(PETSC_SUCCESS);
4147: }

4149: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
4150: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSETriFactorStruct **trifactor)
4151: {
4152:   PetscFunctionBegin;
4153:   if (*trifactor) {
4154:     if ((*trifactor)->descr) PetscCallCUSPARSE(cusparseDestroyMatDescr((*trifactor)->descr));
4155:     if ((*trifactor)->solveInfo) PetscCallCUSPARSE(cusparseDestroyCsrsvInfo((*trifactor)->solveInfo));
4156:     PetscCall(CsrMatrix_Destroy(&(*trifactor)->csrMat));
4157:     if ((*trifactor)->solveBuffer) PetscCallCUDA(cudaFree((*trifactor)->solveBuffer));
4158:     if ((*trifactor)->AA_h) PetscCallCUDA(cudaFreeHost((*trifactor)->AA_h));
4159:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4160:     if ((*trifactor)->csr2cscBuffer) PetscCallCUDA(cudaFree((*trifactor)->csr2cscBuffer));
4161:   #endif
4162:     PetscCall(PetscFree(*trifactor));
4163:   }
4164:   PetscFunctionReturn(PETSC_SUCCESS);
4165: }
4166: #endif

4168: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSEMultStruct **matstruct, MatCUSPARSEStorageFormat format)
4169: {
4170:   CsrMatrix *mat;

4172:   PetscFunctionBegin;
4173:   if (*matstruct) {
4174:     if ((*matstruct)->mat) {
4175:       if (format == MAT_CUSPARSE_ELL || format == MAT_CUSPARSE_HYB) {
4176: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4177:         SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
4178: #else
4179:         cusparseHybMat_t hybMat = (cusparseHybMat_t)(*matstruct)->mat;
4180:         PetscCallCUSPARSE(cusparseDestroyHybMat(hybMat));
4181: #endif
4182:       } else {
4183:         mat = (CsrMatrix *)(*matstruct)->mat;
4184:         PetscCall(CsrMatrix_Destroy(&mat));
4185:       }
4186:     }
4187:     if ((*matstruct)->descr) PetscCallCUSPARSE(cusparseDestroyMatDescr((*matstruct)->descr));
4188:     delete (*matstruct)->cprowIndices;
4189:     if ((*matstruct)->alpha_one) PetscCallCUDA(cudaFree((*matstruct)->alpha_one));
4190:     if ((*matstruct)->beta_zero) PetscCallCUDA(cudaFree((*matstruct)->beta_zero));
4191:     if ((*matstruct)->beta_one) PetscCallCUDA(cudaFree((*matstruct)->beta_one));

4193: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4194:     Mat_SeqAIJCUSPARSEMultStruct *mdata = *matstruct;
4195:     if (mdata->matDescr) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr));

4197:     for (int i = 0; i < 3; i++) {
4198:       if (mdata->cuSpMV[i].initialized) {
4199:         PetscCallCUDA(cudaFree(mdata->cuSpMV[i].spmvBuffer));
4200:         PetscCallCUSPARSE(cusparseDestroyDnVec(mdata->cuSpMV[i].vecXDescr));
4201:         PetscCallCUSPARSE(cusparseDestroyDnVec(mdata->cuSpMV[i].vecYDescr));
4202:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
4203:         if (mdata->matDescr_SpMV[i]) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr_SpMV[i]));
4204:         if (mdata->matDescr_SpMM[i]) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr_SpMM[i]));
4205:   #endif
4206:       }
4207:     }
4208: #endif
4209:     delete *matstruct;
4210:     *matstruct = NULL;
4211:   }
4212:   PetscFunctionReturn(PETSC_SUCCESS);
4213: }

4215: PetscErrorCode MatSeqAIJCUSPARSETriFactors_Reset(Mat_SeqAIJCUSPARSETriFactors_p *trifactors)
4216: {
4217:   Mat_SeqAIJCUSPARSETriFactors *fs = *trifactors;

4219:   PetscFunctionBegin;
4220:   if (fs) {
4221: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
4222:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->loTriFactorPtr));
4223:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->upTriFactorPtr));
4224:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->loTriFactorPtrTranspose));
4225:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->upTriFactorPtrTranspose));
4226:     delete fs->workVector;
4227:     fs->workVector = NULL;
4228: #endif
4229:     delete fs->rpermIndices;
4230:     delete fs->cpermIndices;
4231:     fs->rpermIndices  = NULL;
4232:     fs->cpermIndices  = NULL;
4233:     fs->init_dev_prop = PETSC_FALSE;
4234: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
4235:     PetscCallCUDA(cudaFree(fs->csrRowPtr));
4236:     PetscCallCUDA(cudaFree(fs->csrColIdx));
4237:     PetscCallCUDA(cudaFree(fs->csrRowPtr32));
4238:     PetscCallCUDA(cudaFree(fs->csrColIdx32));
4239:     PetscCallCUDA(cudaFree(fs->csrVal));
4240:     PetscCallCUDA(cudaFree(fs->diag));
4241:     PetscCallCUDA(cudaFree(fs->X));
4242:     PetscCallCUDA(cudaFree(fs->Y));
4243:     // PetscCallCUDA(cudaFree(fs->factBuffer_M)); /* No needed since factBuffer_M shares with one of spsvBuffer_L/U */
4244:     PetscCallCUDA(cudaFree(fs->spsvBuffer_L));
4245:     PetscCallCUDA(cudaFree(fs->spsvBuffer_U));
4246:     PetscCallCUDA(cudaFree(fs->spsvBuffer_Lt));
4247:     PetscCallCUDA(cudaFree(fs->spsvBuffer_Ut));
4248:     PetscCallCUSPARSE(cusparseDestroyMatDescr(fs->matDescr_M));
4249:     PetscCallCUSPARSE(cusparseDestroySpMat(fs->spMatDescr_L));
4250:     PetscCallCUSPARSE(cusparseDestroySpMat(fs->spMatDescr_U));
4251:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_L));
4252:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_Lt));
4253:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_U));
4254:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_Ut));
4255:     PetscCallCUSPARSE(cusparseDestroyDnVec(fs->dnVecDescr_X));
4256:     PetscCallCUSPARSE(cusparseDestroyDnVec(fs->dnVecDescr_Y));
4257:     PetscCallCUSPARSE(cusparseDestroyCsrilu02Info(fs->ilu0Info_M));
4258:     PetscCallCUSPARSE(cusparseDestroyCsric02Info(fs->ic0Info_M));
4259:     PetscCall(PetscFree(fs->csrRowPtr_h));
4260:     PetscCall(PetscFree(fs->csrVal_h));
4261:     PetscCall(PetscFree(fs->diag_h));
4262:     fs->createdTransposeSpSVDescr    = PETSC_FALSE;
4263:     fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;
4264: #endif
4265:   }
4266:   PetscFunctionReturn(PETSC_SUCCESS);
4267: }

4269: static PetscErrorCode MatSeqAIJCUSPARSETriFactors_Destroy(Mat_SeqAIJCUSPARSETriFactors **trifactors)
4270: {
4271:   PetscFunctionBegin;
4272:   if (*trifactors) {
4273:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(trifactors));
4274:     PetscCallCUSPARSE(cusparseDestroy((*trifactors)->handle));
4275:     PetscCall(PetscFree(*trifactors));
4276:   }
4277:   PetscFunctionReturn(PETSC_SUCCESS);
4278: }

4280: struct IJCompare {
4281:   __host__ __device__ inline bool operator()(const thrust::tuple<PetscInt, PetscInt> &t1, const thrust::tuple<PetscInt, PetscInt> &t2)
4282:   {
4283:     if (thrust::get<0>(t1) < thrust::get<0>(t2)) return true;
4284:     if (thrust::get<0>(t1) == thrust::get<0>(t2)) return thrust::get<1>(t1) < thrust::get<1>(t2);
4285:     return false;
4286:   }
4287: };

4289: static PetscErrorCode MatSeqAIJCUSPARSEInvalidateTranspose(Mat A, PetscBool destroy)
4290: {
4291:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;

4293:   PetscFunctionBegin;
4294:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4295:   if (!cusp) PetscFunctionReturn(PETSC_SUCCESS);
4296:   if (destroy) {
4297:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->matTranspose, cusp->format));
4298:     delete cusp->csr2csc_i;
4299:     cusp->csr2csc_i = NULL;
4300:   }
4301:   A->transupdated = PETSC_FALSE;
4302:   PetscFunctionReturn(PETSC_SUCCESS);
4303: }

4305: static PetscErrorCode MatCOOStructDestroy_SeqAIJCUSPARSE(void *data)
4306: {
4307:   MatCOOStruct_SeqAIJ *coo = (MatCOOStruct_SeqAIJ *)data;

4309:   PetscFunctionBegin;
4310:   PetscCallCUDA(cudaFree(coo->perm));
4311:   PetscCallCUDA(cudaFree(coo->jmap));
4312:   PetscCall(PetscFree(coo));
4313:   PetscFunctionReturn(PETSC_SUCCESS);
4314: }

4316: static PetscErrorCode MatSetPreallocationCOO_SeqAIJCUSPARSE(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
4317: {
4318:   PetscBool            dev_ij = PETSC_FALSE;
4319:   PetscMemType         mtype  = PETSC_MEMTYPE_HOST;
4320:   PetscInt            *i, *j;
4321:   PetscContainer       container_h, container_d;
4322:   MatCOOStruct_SeqAIJ *coo_h, *coo_d;

4324:   PetscFunctionBegin;
4325:   // The two MatResetPreallocationCOO_* must be done in order. The former relies on values that might be destroyed by the latter
4326:   PetscCall(PetscGetMemType(coo_i, &mtype));
4327:   if (PetscMemTypeDevice(mtype)) {
4328:     dev_ij = PETSC_TRUE;
4329:     PetscCall(PetscMalloc2(coo_n, &i, coo_n, &j));
4330:     PetscCallCUDA(cudaMemcpy(i, coo_i, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4331:     PetscCallCUDA(cudaMemcpy(j, coo_j, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4332:   } else {
4333:     i = coo_i;
4334:     j = coo_j;
4335:   }

4337:   PetscCall(MatSetPreallocationCOO_SeqAIJ(mat, coo_n, i, j));
4338:   if (dev_ij) PetscCall(PetscFree2(i, j));
4339:   mat->offloadmask = PETSC_OFFLOAD_CPU;
4340:   // Create the GPU memory
4341:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(mat));

4343:   // Copy the COO struct to device
4344:   PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container_h));
4345:   PetscCall(PetscContainerGetPointer(container_h, (void **)&coo_h));
4346:   PetscCall(PetscMalloc1(1, &coo_d));
4347:   *coo_d = *coo_h; // do a shallow copy and then amend some fields that need to be different
4348:   PetscCallCUDA(cudaMalloc((void **)&coo_d->jmap, (coo_h->nz + 1) * sizeof(PetscCount)));
4349:   PetscCallCUDA(cudaMemcpy(coo_d->jmap, coo_h->jmap, (coo_h->nz + 1) * sizeof(PetscCount), cudaMemcpyHostToDevice));
4350:   PetscCallCUDA(cudaMalloc((void **)&coo_d->perm, coo_h->Atot * sizeof(PetscCount)));
4351:   PetscCallCUDA(cudaMemcpy(coo_d->perm, coo_h->perm, coo_h->Atot * sizeof(PetscCount), cudaMemcpyHostToDevice));

4353:   // Put the COO struct in a container and then attach that to the matrix
4354:   PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container_d));
4355:   PetscCall(PetscContainerSetPointer(container_d, coo_d));
4356:   PetscCall(PetscContainerSetUserDestroy(container_d, MatCOOStructDestroy_SeqAIJCUSPARSE));
4357:   PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Device", (PetscObject)container_d));
4358:   PetscCall(PetscContainerDestroy(&container_d));
4359:   PetscFunctionReturn(PETSC_SUCCESS);
4360: }

4362: __global__ static void MatAddCOOValues(const PetscScalar kv[], PetscCount nnz, const PetscCount jmap[], const PetscCount perm[], InsertMode imode, PetscScalar a[])
4363: {
4364:   PetscCount       i         = blockIdx.x * blockDim.x + threadIdx.x;
4365:   const PetscCount grid_size = gridDim.x * blockDim.x;
4366:   for (; i < nnz; i += grid_size) {
4367:     PetscScalar sum = 0.0;
4368:     for (PetscCount k = jmap[i]; k < jmap[i + 1]; k++) sum += kv[perm[k]];
4369:     a[i] = (imode == INSERT_VALUES ? 0.0 : a[i]) + sum;
4370:   }
4371: }

4373: static PetscErrorCode MatSetValuesCOO_SeqAIJCUSPARSE(Mat A, const PetscScalar v[], InsertMode imode)
4374: {
4375:   Mat_SeqAIJ          *seq  = (Mat_SeqAIJ *)A->data;
4376:   Mat_SeqAIJCUSPARSE  *dev  = (Mat_SeqAIJCUSPARSE *)A->spptr;
4377:   PetscCount           Annz = seq->nz;
4378:   PetscMemType         memtype;
4379:   const PetscScalar   *v1 = v;
4380:   PetscScalar         *Aa;
4381:   PetscContainer       container;
4382:   MatCOOStruct_SeqAIJ *coo;

4384:   PetscFunctionBegin;
4385:   if (!dev->mat) PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));

4387:   PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Device", (PetscObject *)&container));
4388:   PetscCall(PetscContainerGetPointer(container, (void **)&coo));

4390:   PetscCall(PetscGetMemType(v, &memtype));
4391:   if (PetscMemTypeHost(memtype)) { /* If user gave v[] in host, we might need to copy it to device if any */
4392:     PetscCallCUDA(cudaMalloc((void **)&v1, coo->n * sizeof(PetscScalar)));
4393:     PetscCallCUDA(cudaMemcpy((void *)v1, v, coo->n * sizeof(PetscScalar), cudaMemcpyHostToDevice));
4394:   }

4396:   if (imode == INSERT_VALUES) PetscCall(MatSeqAIJCUSPARSEGetArrayWrite(A, &Aa));
4397:   else PetscCall(MatSeqAIJCUSPARSEGetArray(A, &Aa));

4399:   PetscCall(PetscLogGpuTimeBegin());
4400:   if (Annz) {
4401:     MatAddCOOValues<<<(Annz + 255) / 256, 256>>>(v1, Annz, coo->jmap, coo->perm, imode, Aa);
4402:     PetscCallCUDA(cudaPeekAtLastError());
4403:   }
4404:   PetscCall(PetscLogGpuTimeEnd());

4406:   if (imode == INSERT_VALUES) PetscCall(MatSeqAIJCUSPARSERestoreArrayWrite(A, &Aa));
4407:   else PetscCall(MatSeqAIJCUSPARSERestoreArray(A, &Aa));

4409:   if (PetscMemTypeHost(memtype)) PetscCallCUDA(cudaFree((void *)v1));
4410:   PetscFunctionReturn(PETSC_SUCCESS);
4411: }

4413: /*@C
4414:   MatSeqAIJCUSPARSEGetIJ - returns the device row storage `i` and `j` indices for `MATSEQAIJCUSPARSE` matrices.

4416:   Not Collective

4418:   Input Parameters:
4419: + A          - the matrix
4420: - compressed - `PETSC_TRUE` or `PETSC_FALSE` indicating the matrix data structure should be always returned in compressed form

4422:   Output Parameters:
4423: + i - the CSR row pointers
4424: - j - the CSR column indices

4426:   Level: developer

4428:   Note:
4429:   When compressed is true, the CSR structure does not contain empty rows

4431: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSERestoreIJ()`, `MatSeqAIJCUSPARSEGetArrayRead()`
4432: @*/
4433: PetscErrorCode MatSeqAIJCUSPARSEGetIJ(Mat A, PetscBool compressed, const int **i, const int **j)
4434: {
4435:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4436:   CsrMatrix          *csr;
4437:   Mat_SeqAIJ         *a = (Mat_SeqAIJ *)A->data;

4439:   PetscFunctionBegin;
4441:   if (!i || !j) PetscFunctionReturn(PETSC_SUCCESS);
4442:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4443:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4444:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4445:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4446:   csr = (CsrMatrix *)cusp->mat->mat;
4447:   if (i) {
4448:     if (!compressed && a->compressedrow.use) { /* need full row offset */
4449:       if (!cusp->rowoffsets_gpu) {
4450:         cusp->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
4451:         cusp->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);
4452:         PetscCall(PetscLogCpuToGpu((A->rmap->n + 1) * sizeof(PetscInt)));
4453:       }
4454:       *i = cusp->rowoffsets_gpu->data().get();
4455:     } else *i = csr->row_offsets->data().get();
4456:   }
4457:   if (j) *j = csr->column_indices->data().get();
4458:   PetscFunctionReturn(PETSC_SUCCESS);
4459: }

4461: /*@C
4462:   MatSeqAIJCUSPARSERestoreIJ - restore the device row storage `i` and `j` indices obtained with `MatSeqAIJCUSPARSEGetIJ()`

4464:   Not Collective

4466:   Input Parameters:
4467: + A          - the matrix
4468: . compressed - `PETSC_TRUE` or `PETSC_FALSE` indicating the matrix data structure should be always returned in compressed form
4469: . i          - the CSR row pointers
4470: - j          - the CSR column indices

4472:   Level: developer

4474: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetIJ()`
4475: @*/
4476: PetscErrorCode MatSeqAIJCUSPARSERestoreIJ(Mat A, PetscBool compressed, const int **i, const int **j)
4477: {
4478:   PetscFunctionBegin;
4480:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4481:   if (i) *i = NULL;
4482:   if (j) *j = NULL;
4483:   (void)compressed;
4484:   PetscFunctionReturn(PETSC_SUCCESS);
4485: }

4487: /*@C
4488:   MatSeqAIJCUSPARSEGetArrayRead - gives read-only access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix is stored

4490:   Not Collective

4492:   Input Parameter:
4493: . A - a `MATSEQAIJCUSPARSE` matrix

4495:   Output Parameter:
4496: . a - pointer to the device data

4498:   Level: developer

4500:   Note:
4501:   May trigger host-device copies if up-to-date matrix data is on host

4503: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`, `MatSeqAIJCUSPARSEGetArrayWrite()`, `MatSeqAIJCUSPARSERestoreArrayRead()`
4504: @*/
4505: PetscErrorCode MatSeqAIJCUSPARSEGetArrayRead(Mat A, const PetscScalar **a)
4506: {
4507:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4508:   CsrMatrix          *csr;

4510:   PetscFunctionBegin;
4512:   PetscAssertPointer(a, 2);
4513:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4514:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4515:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4516:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4517:   csr = (CsrMatrix *)cusp->mat->mat;
4518:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4519:   *a = csr->values->data().get();
4520:   PetscFunctionReturn(PETSC_SUCCESS);
4521: }

4523: /*@C
4524:   MatSeqAIJCUSPARSERestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJCUSPARSEGetArrayRead()`

4526:   Not Collective

4528:   Input Parameters:
4529: + A - a `MATSEQAIJCUSPARSE` matrix
4530: - a - pointer to the device data

4532:   Level: developer

4534: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayRead()`
4535: @*/
4536: PetscErrorCode MatSeqAIJCUSPARSERestoreArrayRead(Mat A, const PetscScalar **a)
4537: {
4538:   PetscFunctionBegin;
4540:   PetscAssertPointer(a, 2);
4541:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4542:   *a = NULL;
4543:   PetscFunctionReturn(PETSC_SUCCESS);
4544: }

4546: /*@C
4547:   MatSeqAIJCUSPARSEGetArray - gives read-write access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix is stored

4549:   Not Collective

4551:   Input Parameter:
4552: . A - a `MATSEQAIJCUSPARSE` matrix

4554:   Output Parameter:
4555: . a - pointer to the device data

4557:   Level: developer

4559:   Note:
4560:   May trigger host-device copies if up-to-date matrix data is on host

4562: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayRead()`, `MatSeqAIJCUSPARSEGetArrayWrite()`, `MatSeqAIJCUSPARSERestoreArray()`
4563: @*/
4564: PetscErrorCode MatSeqAIJCUSPARSEGetArray(Mat A, PetscScalar **a)
4565: {
4566:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4567:   CsrMatrix          *csr;

4569:   PetscFunctionBegin;
4571:   PetscAssertPointer(a, 2);
4572:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4573:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4574:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4575:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4576:   csr = (CsrMatrix *)cusp->mat->mat;
4577:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4578:   *a             = csr->values->data().get();
4579:   A->offloadmask = PETSC_OFFLOAD_GPU;
4580:   PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
4581:   PetscFunctionReturn(PETSC_SUCCESS);
4582: }
4583: /*@C
4584:   MatSeqAIJCUSPARSERestoreArray - restore the read-write access array obtained from `MatSeqAIJCUSPARSEGetArray()`

4586:   Not Collective

4588:   Input Parameters:
4589: + A - a `MATSEQAIJCUSPARSE` matrix
4590: - a - pointer to the device data

4592:   Level: developer

4594: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`
4595: @*/
4596: PetscErrorCode MatSeqAIJCUSPARSERestoreArray(Mat A, PetscScalar **a)
4597: {
4598:   PetscFunctionBegin;
4600:   PetscAssertPointer(a, 2);
4601:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4602:   PetscCall(MatSeqAIJInvalidateDiagonal(A));
4603:   PetscCall(PetscObjectStateIncrease((PetscObject)A));
4604:   *a = NULL;
4605:   PetscFunctionReturn(PETSC_SUCCESS);
4606: }

4608: /*@C
4609:   MatSeqAIJCUSPARSEGetArrayWrite - gives write access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix is stored

4611:   Not Collective

4613:   Input Parameter:
4614: . A - a `MATSEQAIJCUSPARSE` matrix

4616:   Output Parameter:
4617: . a - pointer to the device data

4619:   Level: developer

4621:   Note:
4622:   Does not trigger host-device copies and flags data validity on the GPU

4624: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`, `MatSeqAIJCUSPARSEGetArrayRead()`, `MatSeqAIJCUSPARSERestoreArrayWrite()`
4625: @*/
4626: PetscErrorCode MatSeqAIJCUSPARSEGetArrayWrite(Mat A, PetscScalar **a)
4627: {
4628:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4629:   CsrMatrix          *csr;

4631:   PetscFunctionBegin;
4633:   PetscAssertPointer(a, 2);
4634:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4635:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4636:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4637:   csr = (CsrMatrix *)cusp->mat->mat;
4638:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4639:   *a             = csr->values->data().get();
4640:   A->offloadmask = PETSC_OFFLOAD_GPU;
4641:   PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
4642:   PetscFunctionReturn(PETSC_SUCCESS);
4643: }

4645: /*@C
4646:   MatSeqAIJCUSPARSERestoreArrayWrite - restore the write-only access array obtained from `MatSeqAIJCUSPARSEGetArrayWrite()`

4648:   Not Collective

4650:   Input Parameters:
4651: + A - a `MATSEQAIJCUSPARSE` matrix
4652: - a - pointer to the device data

4654:   Level: developer

4656: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayWrite()`
4657: @*/
4658: PetscErrorCode MatSeqAIJCUSPARSERestoreArrayWrite(Mat A, PetscScalar **a)
4659: {
4660:   PetscFunctionBegin;
4662:   PetscAssertPointer(a, 2);
4663:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4664:   PetscCall(MatSeqAIJInvalidateDiagonal(A));
4665:   PetscCall(PetscObjectStateIncrease((PetscObject)A));
4666:   *a = NULL;
4667:   PetscFunctionReturn(PETSC_SUCCESS);
4668: }

4670: struct IJCompare4 {
4671:   __host__ __device__ inline bool operator()(const thrust::tuple<int, int, PetscScalar, int> &t1, const thrust::tuple<int, int, PetscScalar, int> &t2)
4672:   {
4673:     if (thrust::get<0>(t1) < thrust::get<0>(t2)) return true;
4674:     if (thrust::get<0>(t1) == thrust::get<0>(t2)) return thrust::get<1>(t1) < thrust::get<1>(t2);
4675:     return false;
4676:   }
4677: };

4679: struct Shift {
4680:   int _shift;

4682:   Shift(int shift) : _shift(shift) { }
4683:   __host__ __device__ inline int operator()(const int &c) { return c + _shift; }
4684: };

4686: /* merges two SeqAIJCUSPARSE matrices A, B by concatenating their rows. [A';B']' operation in MATLAB notation */
4687: PetscErrorCode MatSeqAIJCUSPARSEMergeMats(Mat A, Mat B, MatReuse reuse, Mat *C)
4688: {
4689:   Mat_SeqAIJ                   *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data, *c;
4690:   Mat_SeqAIJCUSPARSE           *Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr, *Bcusp = (Mat_SeqAIJCUSPARSE *)B->spptr, *Ccusp;
4691:   Mat_SeqAIJCUSPARSEMultStruct *Cmat;
4692:   CsrMatrix                    *Acsr, *Bcsr, *Ccsr;
4693:   PetscInt                      Annz, Bnnz;
4694:   cusparseStatus_t              stat;
4695:   PetscInt                      i, m, n, zero = 0;

4697:   PetscFunctionBegin;
4700:   PetscAssertPointer(C, 4);
4701:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4702:   PetscCheckTypeName(B, MATSEQAIJCUSPARSE);
4703:   PetscCheck(A->rmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Invalid number or rows %" PetscInt_FMT " != %" PetscInt_FMT, A->rmap->n, B->rmap->n);
4704:   PetscCheck(reuse != MAT_INPLACE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_INPLACE_MATRIX not supported");
4705:   PetscCheck(Acusp->format != MAT_CUSPARSE_ELL && Acusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4706:   PetscCheck(Bcusp->format != MAT_CUSPARSE_ELL && Bcusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4707:   if (reuse == MAT_INITIAL_MATRIX) {
4708:     m = A->rmap->n;
4709:     n = A->cmap->n + B->cmap->n;
4710:     PetscCall(MatCreate(PETSC_COMM_SELF, C));
4711:     PetscCall(MatSetSizes(*C, m, n, m, n));
4712:     PetscCall(MatSetType(*C, MATSEQAIJCUSPARSE));
4713:     c                       = (Mat_SeqAIJ *)(*C)->data;
4714:     Ccusp                   = (Mat_SeqAIJCUSPARSE *)(*C)->spptr;
4715:     Cmat                    = new Mat_SeqAIJCUSPARSEMultStruct;
4716:     Ccsr                    = new CsrMatrix;
4717:     Cmat->cprowIndices      = NULL;
4718:     c->compressedrow.use    = PETSC_FALSE;
4719:     c->compressedrow.nrows  = 0;
4720:     c->compressedrow.i      = NULL;
4721:     c->compressedrow.rindex = NULL;
4722:     Ccusp->workVector       = NULL;
4723:     Ccusp->nrows            = m;
4724:     Ccusp->mat              = Cmat;
4725:     Ccusp->mat->mat         = Ccsr;
4726:     Ccsr->num_rows          = m;
4727:     Ccsr->num_cols          = n;
4728:     PetscCallCUSPARSE(cusparseCreateMatDescr(&Cmat->descr));
4729:     PetscCallCUSPARSE(cusparseSetMatIndexBase(Cmat->descr, CUSPARSE_INDEX_BASE_ZERO));
4730:     PetscCallCUSPARSE(cusparseSetMatType(Cmat->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
4731:     PetscCallCUDA(cudaMalloc((void **)&Cmat->alpha_one, sizeof(PetscScalar)));
4732:     PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_zero, sizeof(PetscScalar)));
4733:     PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_one, sizeof(PetscScalar)));
4734:     PetscCallCUDA(cudaMemcpy(Cmat->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4735:     PetscCallCUDA(cudaMemcpy(Cmat->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4736:     PetscCallCUDA(cudaMemcpy(Cmat->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4737:     PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4738:     PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));
4739:     PetscCheck(Acusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4740:     PetscCheck(Bcusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");

4742:     Acsr                 = (CsrMatrix *)Acusp->mat->mat;
4743:     Bcsr                 = (CsrMatrix *)Bcusp->mat->mat;
4744:     Annz                 = (PetscInt)Acsr->column_indices->size();
4745:     Bnnz                 = (PetscInt)Bcsr->column_indices->size();
4746:     c->nz                = Annz + Bnnz;
4747:     Ccsr->row_offsets    = new THRUSTINTARRAY32(m + 1);
4748:     Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
4749:     Ccsr->values         = new THRUSTARRAY(c->nz);
4750:     Ccsr->num_entries    = c->nz;
4751:     Ccusp->coords        = new THRUSTINTARRAY(c->nz);
4752:     if (c->nz) {
4753:       auto              Acoo = new THRUSTINTARRAY32(Annz);
4754:       auto              Bcoo = new THRUSTINTARRAY32(Bnnz);
4755:       auto              Ccoo = new THRUSTINTARRAY32(c->nz);
4756:       THRUSTINTARRAY32 *Aroff, *Broff;

4758:       if (a->compressedrow.use) { /* need full row offset */
4759:         if (!Acusp->rowoffsets_gpu) {
4760:           Acusp->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
4761:           Acusp->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);
4762:           PetscCall(PetscLogCpuToGpu((A->rmap->n + 1) * sizeof(PetscInt)));
4763:         }
4764:         Aroff = Acusp->rowoffsets_gpu;
4765:       } else Aroff = Acsr->row_offsets;
4766:       if (b->compressedrow.use) { /* need full row offset */
4767:         if (!Bcusp->rowoffsets_gpu) {
4768:           Bcusp->rowoffsets_gpu = new THRUSTINTARRAY32(B->rmap->n + 1);
4769:           Bcusp->rowoffsets_gpu->assign(b->i, b->i + B->rmap->n + 1);
4770:           PetscCall(PetscLogCpuToGpu((B->rmap->n + 1) * sizeof(PetscInt)));
4771:         }
4772:         Broff = Bcusp->rowoffsets_gpu;
4773:       } else Broff = Bcsr->row_offsets;
4774:       PetscCall(PetscLogGpuTimeBegin());
4775:       stat = cusparseXcsr2coo(Acusp->handle, Aroff->data().get(), Annz, m, Acoo->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4776:       PetscCallCUSPARSE(stat);
4777:       stat = cusparseXcsr2coo(Bcusp->handle, Broff->data().get(), Bnnz, m, Bcoo->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4778:       PetscCallCUSPARSE(stat);
4779:       /* Issues when using bool with large matrices on SUMMIT 10.2.89 */
4780:       auto Aperm = thrust::make_constant_iterator(1);
4781:       auto Bperm = thrust::make_constant_iterator(0);
4782: #if PETSC_PKG_CUDA_VERSION_GE(10, 0, 0)
4783:       auto Bcib = thrust::make_transform_iterator(Bcsr->column_indices->begin(), Shift(A->cmap->n));
4784:       auto Bcie = thrust::make_transform_iterator(Bcsr->column_indices->end(), Shift(A->cmap->n));
4785: #else
4786:       /* there are issues instantiating the merge operation using a transform iterator for the columns of B */
4787:       auto Bcib = Bcsr->column_indices->begin();
4788:       auto Bcie = Bcsr->column_indices->end();
4789:       thrust::transform(Bcib, Bcie, Bcib, Shift(A->cmap->n));
4790: #endif
4791:       auto wPerm = new THRUSTINTARRAY32(Annz + Bnnz);
4792:       auto Azb   = thrust::make_zip_iterator(thrust::make_tuple(Acoo->begin(), Acsr->column_indices->begin(), Acsr->values->begin(), Aperm));
4793:       auto Aze   = thrust::make_zip_iterator(thrust::make_tuple(Acoo->end(), Acsr->column_indices->end(), Acsr->values->end(), Aperm));
4794:       auto Bzb   = thrust::make_zip_iterator(thrust::make_tuple(Bcoo->begin(), Bcib, Bcsr->values->begin(), Bperm));
4795:       auto Bze   = thrust::make_zip_iterator(thrust::make_tuple(Bcoo->end(), Bcie, Bcsr->values->end(), Bperm));
4796:       auto Czb   = thrust::make_zip_iterator(thrust::make_tuple(Ccoo->begin(), Ccsr->column_indices->begin(), Ccsr->values->begin(), wPerm->begin()));
4797:       auto p1    = Ccusp->coords->begin();
4798:       auto p2    = Ccusp->coords->begin();
4799:       thrust::advance(p2, Annz);
4800:       PetscCallThrust(thrust::merge(thrust::device, Azb, Aze, Bzb, Bze, Czb, IJCompare4()));
4801: #if PETSC_PKG_CUDA_VERSION_LT(10, 0, 0)
4802:       thrust::transform(Bcib, Bcie, Bcib, Shift(-A->cmap->n));
4803: #endif
4804:       auto cci = thrust::make_counting_iterator(zero);
4805:       auto cce = thrust::make_counting_iterator(c->nz);
4806: #if 0 //Errors on SUMMIT cuda 11.1.0
4807:       PetscCallThrust(thrust::partition_copy(thrust::device,cci,cce,wPerm->begin(),p1,p2,thrust::identity<int>()));
4808: #else
4809:       auto pred = thrust::identity<int>();
4810:       PetscCallThrust(thrust::copy_if(thrust::device, cci, cce, wPerm->begin(), p1, pred));
4811:       PetscCallThrust(thrust::remove_copy_if(thrust::device, cci, cce, wPerm->begin(), p2, pred));
4812: #endif
4813:       stat = cusparseXcoo2csr(Ccusp->handle, Ccoo->data().get(), c->nz, m, Ccsr->row_offsets->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4814:       PetscCallCUSPARSE(stat);
4815:       PetscCall(PetscLogGpuTimeEnd());
4816:       delete wPerm;
4817:       delete Acoo;
4818:       delete Bcoo;
4819:       delete Ccoo;
4820: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4821:       stat = cusparseCreateCsr(&Cmat->matDescr, Ccsr->num_rows, Ccsr->num_cols, Ccsr->num_entries, Ccsr->row_offsets->data().get(), Ccsr->column_indices->data().get(), Ccsr->values->data().get(), CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype);
4822:       PetscCallCUSPARSE(stat);
4823: #endif
4824:       if (A->form_explicit_transpose && B->form_explicit_transpose) { /* if A and B have the transpose, generate C transpose too */
4825:         PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
4826:         PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(B));
4827:         PetscBool                     AT = Acusp->matTranspose ? PETSC_TRUE : PETSC_FALSE, BT = Bcusp->matTranspose ? PETSC_TRUE : PETSC_FALSE;
4828:         Mat_SeqAIJCUSPARSEMultStruct *CmatT = new Mat_SeqAIJCUSPARSEMultStruct;
4829:         CsrMatrix                    *CcsrT = new CsrMatrix;
4830:         CsrMatrix                    *AcsrT = AT ? (CsrMatrix *)Acusp->matTranspose->mat : NULL;
4831:         CsrMatrix                    *BcsrT = BT ? (CsrMatrix *)Bcusp->matTranspose->mat : NULL;

4833:         (*C)->form_explicit_transpose = PETSC_TRUE;
4834:         (*C)->transupdated            = PETSC_TRUE;
4835:         Ccusp->rowoffsets_gpu         = NULL;
4836:         CmatT->cprowIndices           = NULL;
4837:         CmatT->mat                    = CcsrT;
4838:         CcsrT->num_rows               = n;
4839:         CcsrT->num_cols               = m;
4840:         CcsrT->num_entries            = c->nz;

4842:         CcsrT->row_offsets    = new THRUSTINTARRAY32(n + 1);
4843:         CcsrT->column_indices = new THRUSTINTARRAY32(c->nz);
4844:         CcsrT->values         = new THRUSTARRAY(c->nz);

4846:         PetscCall(PetscLogGpuTimeBegin());
4847:         auto rT = CcsrT->row_offsets->begin();
4848:         if (AT) {
4849:           rT = thrust::copy(AcsrT->row_offsets->begin(), AcsrT->row_offsets->end(), rT);
4850:           thrust::advance(rT, -1);
4851:         }
4852:         if (BT) {
4853:           auto titb = thrust::make_transform_iterator(BcsrT->row_offsets->begin(), Shift(a->nz));
4854:           auto tite = thrust::make_transform_iterator(BcsrT->row_offsets->end(), Shift(a->nz));
4855:           thrust::copy(titb, tite, rT);
4856:         }
4857:         auto cT = CcsrT->column_indices->begin();
4858:         if (AT) cT = thrust::copy(AcsrT->column_indices->begin(), AcsrT->column_indices->end(), cT);
4859:         if (BT) thrust::copy(BcsrT->column_indices->begin(), BcsrT->column_indices->end(), cT);
4860:         auto vT = CcsrT->values->begin();
4861:         if (AT) vT = thrust::copy(AcsrT->values->begin(), AcsrT->values->end(), vT);
4862:         if (BT) thrust::copy(BcsrT->values->begin(), BcsrT->values->end(), vT);
4863:         PetscCall(PetscLogGpuTimeEnd());

4865:         PetscCallCUSPARSE(cusparseCreateMatDescr(&CmatT->descr));
4866:         PetscCallCUSPARSE(cusparseSetMatIndexBase(CmatT->descr, CUSPARSE_INDEX_BASE_ZERO));
4867:         PetscCallCUSPARSE(cusparseSetMatType(CmatT->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
4868:         PetscCallCUDA(cudaMalloc((void **)&CmatT->alpha_one, sizeof(PetscScalar)));
4869:         PetscCallCUDA(cudaMalloc((void **)&CmatT->beta_zero, sizeof(PetscScalar)));
4870:         PetscCallCUDA(cudaMalloc((void **)&CmatT->beta_one, sizeof(PetscScalar)));
4871:         PetscCallCUDA(cudaMemcpy(CmatT->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4872:         PetscCallCUDA(cudaMemcpy(CmatT->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4873:         PetscCallCUDA(cudaMemcpy(CmatT->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4874: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4875:         stat = cusparseCreateCsr(&CmatT->matDescr, CcsrT->num_rows, CcsrT->num_cols, CcsrT->num_entries, CcsrT->row_offsets->data().get(), CcsrT->column_indices->data().get(), CcsrT->values->data().get(), CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype);
4876:         PetscCallCUSPARSE(stat);
4877: #endif
4878:         Ccusp->matTranspose = CmatT;
4879:       }
4880:     }

4882:     c->singlemalloc = PETSC_FALSE;
4883:     c->free_a       = PETSC_TRUE;
4884:     c->free_ij      = PETSC_TRUE;
4885:     PetscCall(PetscMalloc1(m + 1, &c->i));
4886:     PetscCall(PetscMalloc1(c->nz, &c->j));
4887:     if (PetscDefined(USE_64BIT_INDICES)) { /* 32 to 64-bit conversion on the GPU and then copy to host (lazy) */
4888:       THRUSTINTARRAY ii(Ccsr->row_offsets->size());
4889:       THRUSTINTARRAY jj(Ccsr->column_indices->size());
4890:       ii = *Ccsr->row_offsets;
4891:       jj = *Ccsr->column_indices;
4892:       PetscCallCUDA(cudaMemcpy(c->i, ii.data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4893:       PetscCallCUDA(cudaMemcpy(c->j, jj.data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4894:     } else {
4895:       PetscCallCUDA(cudaMemcpy(c->i, Ccsr->row_offsets->data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4896:       PetscCallCUDA(cudaMemcpy(c->j, Ccsr->column_indices->data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4897:     }
4898:     PetscCall(PetscLogGpuToCpu((Ccsr->column_indices->size() + Ccsr->row_offsets->size()) * sizeof(PetscInt)));
4899:     PetscCall(PetscMalloc1(m, &c->ilen));
4900:     PetscCall(PetscMalloc1(m, &c->imax));
4901:     c->maxnz         = c->nz;
4902:     c->nonzerorowcnt = 0;
4903:     c->rmax          = 0;
4904:     for (i = 0; i < m; i++) {
4905:       const PetscInt nn = c->i[i + 1] - c->i[i];
4906:       c->ilen[i] = c->imax[i] = nn;
4907:       c->nonzerorowcnt += (PetscInt) !!nn;
4908:       c->rmax = PetscMax(c->rmax, nn);
4909:     }
4910:     PetscCall(MatMarkDiagonal_SeqAIJ(*C));
4911:     PetscCall(PetscMalloc1(c->nz, &c->a));
4912:     (*C)->nonzerostate++;
4913:     PetscCall(PetscLayoutSetUp((*C)->rmap));
4914:     PetscCall(PetscLayoutSetUp((*C)->cmap));
4915:     Ccusp->nonzerostate = (*C)->nonzerostate;
4916:     (*C)->preallocated  = PETSC_TRUE;
4917:   } else {
4918:     PetscCheck((*C)->rmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Invalid number or rows %" PetscInt_FMT " != %" PetscInt_FMT, (*C)->rmap->n, B->rmap->n);
4919:     c = (Mat_SeqAIJ *)(*C)->data;
4920:     if (c->nz) {
4921:       Ccusp = (Mat_SeqAIJCUSPARSE *)(*C)->spptr;
4922:       PetscCheck(Ccusp->coords, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing coords");
4923:       PetscCheck(Ccusp->format != MAT_CUSPARSE_ELL && Ccusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4924:       PetscCheck(Ccusp->nonzerostate == (*C)->nonzerostate, PETSC_COMM_SELF, PETSC_ERR_COR, "Wrong nonzerostate");
4925:       PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4926:       PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));
4927:       PetscCheck(Acusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4928:       PetscCheck(Bcusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4929:       Acsr = (CsrMatrix *)Acusp->mat->mat;
4930:       Bcsr = (CsrMatrix *)Bcusp->mat->mat;
4931:       Ccsr = (CsrMatrix *)Ccusp->mat->mat;
4932:       PetscCheck(Acsr->num_entries == (PetscInt)Acsr->values->size(), PETSC_COMM_SELF, PETSC_ERR_COR, "A nnz %" PetscInt_FMT " != %" PetscInt_FMT, Acsr->num_entries, (PetscInt)Acsr->values->size());
4933:       PetscCheck(Bcsr->num_entries == (PetscInt)Bcsr->values->size(), PETSC_COMM_SELF, PETSC_ERR_COR, "B nnz %" PetscInt_FMT " != %" PetscInt_FMT, Bcsr->num_entries, (PetscInt)Bcsr->values->size());
4934:       PetscCheck(Ccsr->num_entries == (PetscInt)Ccsr->values->size(), PETSC_COMM_SELF, PETSC_ERR_COR, "C nnz %" PetscInt_FMT " != %" PetscInt_FMT, Ccsr->num_entries, (PetscInt)Ccsr->values->size());
4935:       PetscCheck(Ccsr->num_entries == Acsr->num_entries + Bcsr->num_entries, PETSC_COMM_SELF, PETSC_ERR_COR, "C nnz %" PetscInt_FMT " != %" PetscInt_FMT " + %" PetscInt_FMT, Ccsr->num_entries, Acsr->num_entries, Bcsr->num_entries);
4936:       PetscCheck(Ccusp->coords->size() == Ccsr->values->size(), PETSC_COMM_SELF, PETSC_ERR_COR, "permSize %" PetscInt_FMT " != %" PetscInt_FMT, (PetscInt)Ccusp->coords->size(), (PetscInt)Ccsr->values->size());
4937:       auto pmid = Ccusp->coords->begin();
4938:       thrust::advance(pmid, Acsr->num_entries);
4939:       PetscCall(PetscLogGpuTimeBegin());
4940:       auto zibait = thrust::make_zip_iterator(thrust::make_tuple(Acsr->values->begin(), thrust::make_permutation_iterator(Ccsr->values->begin(), Ccusp->coords->begin())));
4941:       auto zieait = thrust::make_zip_iterator(thrust::make_tuple(Acsr->values->end(), thrust::make_permutation_iterator(Ccsr->values->begin(), pmid)));
4942:       thrust::for_each(zibait, zieait, VecCUDAEquals());
4943:       auto zibbit = thrust::make_zip_iterator(thrust::make_tuple(Bcsr->values->begin(), thrust::make_permutation_iterator(Ccsr->values->begin(), pmid)));
4944:       auto ziebit = thrust::make_zip_iterator(thrust::make_tuple(Bcsr->values->end(), thrust::make_permutation_iterator(Ccsr->values->begin(), Ccusp->coords->end())));
4945:       thrust::for_each(zibbit, ziebit, VecCUDAEquals());
4946:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(*C, PETSC_FALSE));
4947:       if (A->form_explicit_transpose && B->form_explicit_transpose && (*C)->form_explicit_transpose) {
4948:         PetscCheck(Ccusp->matTranspose, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing transpose Mat_SeqAIJCUSPARSEMultStruct");
4949:         PetscBool  AT = Acusp->matTranspose ? PETSC_TRUE : PETSC_FALSE, BT = Bcusp->matTranspose ? PETSC_TRUE : PETSC_FALSE;
4950:         CsrMatrix *AcsrT = AT ? (CsrMatrix *)Acusp->matTranspose->mat : NULL;
4951:         CsrMatrix *BcsrT = BT ? (CsrMatrix *)Bcusp->matTranspose->mat : NULL;
4952:         CsrMatrix *CcsrT = (CsrMatrix *)Ccusp->matTranspose->mat;
4953:         auto       vT    = CcsrT->values->begin();
4954:         if (AT) vT = thrust::copy(AcsrT->values->begin(), AcsrT->values->end(), vT);
4955:         if (BT) thrust::copy(BcsrT->values->begin(), BcsrT->values->end(), vT);
4956:         (*C)->transupdated = PETSC_TRUE;
4957:       }
4958:       PetscCall(PetscLogGpuTimeEnd());
4959:     }
4960:   }
4961:   PetscCall(PetscObjectStateIncrease((PetscObject)*C));
4962:   (*C)->assembled     = PETSC_TRUE;
4963:   (*C)->was_assembled = PETSC_FALSE;
4964:   (*C)->offloadmask   = PETSC_OFFLOAD_GPU;
4965:   PetscFunctionReturn(PETSC_SUCCESS);
4966: }

4968: static PetscErrorCode MatSeqAIJCopySubArray_SeqAIJCUSPARSE(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
4969: {
4970:   bool               dmem;
4971:   const PetscScalar *av;

4973:   PetscFunctionBegin;
4974:   dmem = isCudaMem(v);
4975:   PetscCall(MatSeqAIJCUSPARSEGetArrayRead(A, &av));
4976:   if (n && idx) {
4977:     THRUSTINTARRAY widx(n);
4978:     widx.assign(idx, idx + n);
4979:     PetscCall(PetscLogCpuToGpu(n * sizeof(PetscInt)));

4981:     THRUSTARRAY                    *w = NULL;
4982:     thrust::device_ptr<PetscScalar> dv;
4983:     if (dmem) {
4984:       dv = thrust::device_pointer_cast(v);
4985:     } else {
4986:       w  = new THRUSTARRAY(n);
4987:       dv = w->data();
4988:     }
4989:     thrust::device_ptr<const PetscScalar> dav = thrust::device_pointer_cast(av);

4991:     auto zibit = thrust::make_zip_iterator(thrust::make_tuple(thrust::make_permutation_iterator(dav, widx.begin()), dv));
4992:     auto zieit = thrust::make_zip_iterator(thrust::make_tuple(thrust::make_permutation_iterator(dav, widx.end()), dv + n));
4993:     thrust::for_each(zibit, zieit, VecCUDAEquals());
4994:     if (w) PetscCallCUDA(cudaMemcpy(v, w->data().get(), n * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
4995:     delete w;
4996:   } else {
4997:     PetscCallCUDA(cudaMemcpy(v, av, n * sizeof(PetscScalar), dmem ? cudaMemcpyDeviceToDevice : cudaMemcpyDeviceToHost));
4998:   }
4999:   if (!dmem) PetscCall(PetscLogCpuToGpu(n * sizeof(PetscScalar)));
5000:   PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(A, &av));
5001:   PetscFunctionReturn(PETSC_SUCCESS);
5002: }