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: #endif
 19: #include <thrust/iterator/constant_iterator.h>
 20: #include <thrust/remove.h>
 21: #include <thrust/sort.h>
 22: #include <thrust/unique.h>
 23: #if PETSC_PKG_CUDA_VERSION_GE(12, 9, 0) && !PetscDefined(HAVE_THRUST)
 24:   #include <cuda/std/functional>
 25: #endif

 27: const char *const MatCUSPARSEStorageFormats[] = {"CSR", "ELL", "HYB", "MatCUSPARSEStorageFormat", "MAT_CUSPARSE_", 0};
 28: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
 29: /*
 30:   The following are copied from cusparse.h in CUDA-11.0. In MatCUSPARSESpMVAlgorithms[] etc, we copy them in
 31:   0-based integer value order, since we want to use PetscOptionsEnum() to parse user command line options for them.
 32: */
 33: const char *const MatCUSPARSESpMVAlgorithms[]    = {"MV_ALG_DEFAULT", "COOMV_ALG", "CSRMV_ALG1", "CSRMV_ALG2", "cusparseSpMVAlg_t", "CUSPARSE_", 0};
 34: const char *const MatCUSPARSESpMMAlgorithms[]    = {"ALG_DEFAULT", "COO_ALG1", "COO_ALG2", "COO_ALG3", "CSR_ALG1", "COO_ALG4", "CSR_ALG2", "cusparseSpMMAlg_t", "CUSPARSE_SPMM_", 0};
 35: const char *const MatCUSPARSECsr2CscAlgorithms[] = {"INVALID" /*cusparse does not have enum 0! We created one*/, "ALG1", "ALG2", "cusparseCsr2CscAlg_t", "CUSPARSE_CSR2CSC_", 0};
 36: #endif

 38: static PetscErrorCode MatICCFactorSymbolic_SeqAIJCUSPARSE(Mat, Mat, IS, const MatFactorInfo *);
 39: static PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJCUSPARSE(Mat, Mat, IS, const MatFactorInfo *);
 40: static PetscErrorCode MatCholeskyFactorNumeric_SeqAIJCUSPARSE(Mat, Mat, const MatFactorInfo *);
 41: static PetscErrorCode MatILUFactorSymbolic_SeqAIJCUSPARSE(Mat, Mat, IS, IS, const MatFactorInfo *);
 42: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
 43: static PetscErrorCode MatSolve_SeqAIJCUSPARSE(Mat, Vec, Vec);
 44: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE(Mat, Vec, Vec);
 45: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_NaturalOrdering(Mat, Vec, Vec);
 46: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering(Mat, Vec, Vec);
 47: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSETriFactorStruct **);
 48: #endif
 49: static PetscErrorCode MatSetFromOptions_SeqAIJCUSPARSE(Mat, PetscOptionItems PetscOptionsObject);
 50: static PetscErrorCode MatAXPY_SeqAIJCUSPARSE(Mat, PetscScalar, Mat, MatStructure);
 51: static PetscErrorCode MatScale_SeqAIJCUSPARSE(Mat, PetscScalar);
 52: static PetscErrorCode MatMult_SeqAIJCUSPARSE(Mat, Vec, Vec);
 53: static PetscErrorCode MatMultAdd_SeqAIJCUSPARSE(Mat, Vec, Vec, Vec);
 54: static PetscErrorCode MatMultTranspose_SeqAIJCUSPARSE(Mat, Vec, Vec);
 55: static PetscErrorCode MatMultTransposeAdd_SeqAIJCUSPARSE(Mat, Vec, Vec, Vec);
 56: static PetscErrorCode MatMultHermitianTranspose_SeqAIJCUSPARSE(Mat, Vec, Vec);
 57: static PetscErrorCode MatMultHermitianTransposeAdd_SeqAIJCUSPARSE(Mat, Vec, Vec, Vec);
 58: static PetscErrorCode MatMultAddKernel_SeqAIJCUSPARSE(Mat, Vec, Vec, Vec, PetscBool, PetscBool);

 60: static PetscErrorCode CsrMatrix_Destroy(CsrMatrix **);
 61: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSEMultStruct **, MatCUSPARSEStorageFormat);
 62: static PetscErrorCode MatSeqAIJCUSPARSETriFactors_Destroy(Mat_SeqAIJCUSPARSETriFactors **);
 63: static PetscErrorCode MatSeqAIJCUSPARSE_Destroy(Mat);

 65: static PetscErrorCode MatSeqAIJCUSPARSECopyFromGPU(Mat);
 66: static PetscErrorCode MatSeqAIJCUSPARSEInvalidateTranspose(Mat, PetscBool);

 68: static PetscErrorCode MatSeqAIJCopySubArray_SeqAIJCUSPARSE(Mat, PetscInt, const PetscInt[], PetscScalar[]);
 69: static PetscErrorCode MatSetPreallocationCOO_SeqAIJCUSPARSE(Mat, PetscCount, PetscInt[], PetscInt[]);
 70: static PetscErrorCode MatSetValuesCOO_SeqAIJCUSPARSE(Mat, const PetscScalar[], InsertMode);

 72: PETSC_INTERN PetscErrorCode MatCUSPARSESetFormat_SeqAIJCUSPARSE(Mat A, MatCUSPARSEFormatOperation op, MatCUSPARSEStorageFormat format)
 73: {
 74:   Mat_SeqAIJCUSPARSE *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;

 76:   PetscFunctionBegin;
 77:   switch (op) {
 78:   case MAT_CUSPARSE_MULT:
 79:     cusparsestruct->format = format;
 80:     break;
 81:   case MAT_CUSPARSE_ALL:
 82:     cusparsestruct->format = format;
 83:     break;
 84:   default:
 85:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unsupported operation %d for MatCUSPARSEFormatOperation. MAT_CUSPARSE_MULT and MAT_CUSPARSE_ALL are currently supported.", op);
 86:   }
 87:   PetscFunctionReturn(PETSC_SUCCESS);
 88: }

 90: /*@
 91:   MatCUSPARSESetFormat - Sets the storage format of `MATSEQCUSPARSE` matrices for a particular
 92:   operation. Only the `MatMult()` operation can use different GPU storage formats

 94:   Not Collective

 96:   Input Parameters:
 97: + A      - Matrix of type `MATSEQAIJCUSPARSE`
 98: . op     - `MatCUSPARSEFormatOperation`. `MATSEQAIJCUSPARSE` matrices support `MAT_CUSPARSE_MULT` and `MAT_CUSPARSE_ALL`.
 99:            `MATMPIAIJCUSPARSE` matrices support `MAT_CUSPARSE_MULT_DIAG`,`MAT_CUSPARSE_MULT_OFFDIAG`, and `MAT_CUSPARSE_ALL`.
100: - format - `MatCUSPARSEStorageFormat` (one of `MAT_CUSPARSE_CSR`, `MAT_CUSPARSE_ELL`, `MAT_CUSPARSE_HYB`.)

102:   Level: intermediate

104: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJCUSPARSE`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
105: @*/
106: PetscErrorCode MatCUSPARSESetFormat(Mat A, MatCUSPARSEFormatOperation op, MatCUSPARSEStorageFormat format)
107: {
108:   PetscFunctionBegin;
110:   PetscTryMethod(A, "MatCUSPARSESetFormat_C", (Mat, MatCUSPARSEFormatOperation, MatCUSPARSEStorageFormat), (A, op, format));
111:   PetscFunctionReturn(PETSC_SUCCESS);
112: }

114: PETSC_INTERN PetscErrorCode MatCUSPARSESetUseCPUSolve_SeqAIJCUSPARSE(Mat A, PetscBool use_cpu)
115: {
116:   Mat_SeqAIJCUSPARSE *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;

118:   PetscFunctionBegin;
119:   cusparsestruct->use_cpu_solve = use_cpu;
120:   PetscFunctionReturn(PETSC_SUCCESS);
121: }

123: /*@
124:   MatCUSPARSESetUseCPUSolve - Sets to use CPU `MatSolve()`.

126:   Input Parameters:
127: + A       - Matrix of type `MATSEQAIJCUSPARSE`
128: - use_cpu - set flag for using the built-in CPU `MatSolve()`

130:   Level: intermediate

132:   Note:
133:   The NVIDIA cuSPARSE LU solver currently computes the factors with the built-in CPU method
134:   and moves the factors to the GPU for the solve. We have observed better performance keeping the data on the CPU and performing the solve there.
135:   This method to specify if the solve is done on the CPU or GPU (GPU is the default).

137: .seealso: [](ch_matrices), `Mat`, `MatSolve()`, `MATSEQAIJCUSPARSE`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
138: @*/
139: PetscErrorCode MatCUSPARSESetUseCPUSolve(Mat A, PetscBool use_cpu)
140: {
141:   PetscFunctionBegin;
143:   PetscTryMethod(A, "MatCUSPARSESetUseCPUSolve_C", (Mat, PetscBool), (A, use_cpu));
144:   PetscFunctionReturn(PETSC_SUCCESS);
145: }

147: static PetscErrorCode MatSetOption_SeqAIJCUSPARSE(Mat A, MatOption op, PetscBool flg)
148: {
149:   PetscFunctionBegin;
150:   switch (op) {
151:   case MAT_FORM_EXPLICIT_TRANSPOSE:
152:     /* need to destroy the transpose matrix if present to prevent from logic errors if flg is set to true later */
153:     if (A->form_explicit_transpose && !flg) PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_TRUE));
154:     A->form_explicit_transpose = flg;
155:     break;
156:   default:
157:     PetscCall(MatSetOption_SeqAIJ(A, op, flg));
158:     break;
159:   }
160:   PetscFunctionReturn(PETSC_SUCCESS);
161: }

163: static PetscErrorCode MatSetFromOptions_SeqAIJCUSPARSE(Mat A, PetscOptionItems PetscOptionsObject)
164: {
165:   MatCUSPARSEStorageFormat format;
166:   PetscBool                flg;
167:   Mat_SeqAIJCUSPARSE      *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;

169:   PetscFunctionBegin;
170:   PetscOptionsHeadBegin(PetscOptionsObject, "SeqAIJCUSPARSE options");
171:   if (A->factortype == MAT_FACTOR_NONE) {
172:     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));
173:     if (flg) PetscCall(MatCUSPARSESetFormat(A, MAT_CUSPARSE_MULT, format));

175:     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));
176:     if (flg) PetscCall(MatCUSPARSESetFormat(A, MAT_CUSPARSE_ALL, format));
177:     PetscCall(PetscOptionsBool("-mat_cusparse_use_cpu_solve", "Use CPU (I)LU solve", "MatCUSPARSESetUseCPUSolve", cusparsestruct->use_cpu_solve, &cusparsestruct->use_cpu_solve, &flg));
178:     if (flg) PetscCall(MatCUSPARSESetUseCPUSolve(A, cusparsestruct->use_cpu_solve));
179: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
180:     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));
181:     /* If user did use this option, check its consistency with cuSPARSE, since PetscOptionsEnum() sets enum values based on their position in MatCUSPARSESpMVAlgorithms[] */
182:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
183:     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");
184:   #else
185:     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");
186:   #endif
187:     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));
188:     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");

190:     PetscCall(
191:       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));
192:     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");
193: #endif
194:   }
195:   PetscOptionsHeadEnd();
196:   PetscFunctionReturn(PETSC_SUCCESS);
197: }

199: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
200: static PetscErrorCode MatSeqAIJCUSPARSEBuildFactoredMatrix_LU(Mat A)
201: {
202:   Mat_SeqAIJ                   *a  = static_cast<Mat_SeqAIJ *>(A->data);
203:   PetscInt                      m  = A->rmap->n;
204:   Mat_SeqAIJCUSPARSETriFactors *fs = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
205:   const PetscInt               *Ai = a->i, *Aj = a->j, *adiag;
206:   const MatScalar              *Aa = a->a;
207:   PetscInt                     *Mi, *Mj, Mnz;
208:   PetscScalar                  *Ma;

210:   PetscFunctionBegin;
211:   PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &adiag, NULL));
212:   if (A->offloadmask == PETSC_OFFLOAD_CPU) { // A's latest factors are on CPU
213:     if (!fs->csrRowPtr) {                    // Is't the first time to do the setup? Use csrRowPtr since it is not null even when m=0
214:       // Re-arrange the (skewed) factored matrix and put the result into M, a regular csr matrix on host
215:       Mnz = (Ai[m] - Ai[0]) + (adiag[0] - adiag[m]); // Lnz (without the unit diagonal) + Unz (with the non-unit diagonal)
216:       PetscCall(PetscMalloc1(m + 1, &Mi));
217:       PetscCall(PetscMalloc1(Mnz, &Mj)); // Mj is temp
218:       PetscCall(PetscMalloc1(Mnz, &Ma));
219:       Mi[0] = 0;
220:       for (PetscInt i = 0; i < m; i++) {
221:         PetscInt llen = Ai[i + 1] - Ai[i];
222:         PetscInt ulen = adiag[i] - adiag[i + 1];
223:         PetscCall(PetscArraycpy(Mj + Mi[i], Aj + Ai[i], llen));                           // entries of L
224:         Mj[Mi[i] + llen] = i;                                                             // diagonal entry
225:         PetscCall(PetscArraycpy(Mj + Mi[i] + llen + 1, Aj + adiag[i + 1] + 1, ulen - 1)); // entries of U on the right of the diagonal
226:         Mi[i + 1] = Mi[i] + llen + ulen;
227:       }
228:       // Copy M (L,U) from host to device
229:       PetscCallCUDA(cudaMalloc(&fs->csrRowPtr, sizeof(*fs->csrRowPtr) * (m + 1)));
230:       PetscCallCUDA(cudaMalloc(&fs->csrColIdx, sizeof(*fs->csrColIdx) * Mnz));
231:       PetscCallCUDA(cudaMalloc(&fs->csrVal, sizeof(*fs->csrVal) * Mnz));
232:       PetscCallCUDA(cudaMemcpy(fs->csrRowPtr, Mi, sizeof(*fs->csrRowPtr) * (m + 1), cudaMemcpyHostToDevice));
233:       PetscCallCUDA(cudaMemcpy(fs->csrColIdx, Mj, sizeof(*fs->csrColIdx) * Mnz, cudaMemcpyHostToDevice));

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

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

248:       fillMode = CUSPARSE_FILL_MODE_UPPER;
249:       diagType = CUSPARSE_DIAG_TYPE_NON_UNIT;
250:       PetscCallCUSPARSE(cusparseCreateCsr(&fs->spMatDescr_U, m, m, Mnz, fs->csrRowPtr, fs->csrColIdx, fs->csrVal, indexType, indexType, CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype));
251:       PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
252:       PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

254:       // Allocate work vectors in SpSv
255:       PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(*fs->X) * m));
256:       PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(*fs->Y) * m));

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

261:       // Query buffer sizes for SpSV and then allocate buffers, temporarily assuming opA = CUSPARSE_OPERATION_NON_TRANSPOSE
262:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_L));
263:       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));
264:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_U));
265:       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));
266:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U));
267:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L));

269:       // Record for reuse
270:       fs->csrRowPtr_h = Mi;
271:       fs->csrVal_h    = Ma;
272:       PetscCall(PetscFree(Mj));
273:     }
274:     // Copy the value
275:     Mi  = fs->csrRowPtr_h;
276:     Ma  = fs->csrVal_h;
277:     Mnz = Mi[m];
278:     for (PetscInt i = 0; i < m; i++) {
279:       PetscInt llen = Ai[i + 1] - Ai[i];
280:       PetscInt ulen = adiag[i] - adiag[i + 1];
281:       PetscCall(PetscArraycpy(Ma + Mi[i], Aa + Ai[i], llen));                           // entries of L
282:       Ma[Mi[i] + llen] = (MatScalar)1.0 / Aa[adiag[i]];                                 // recover the diagonal entry
283:       PetscCall(PetscArraycpy(Ma + Mi[i] + llen + 1, Aa + adiag[i + 1] + 1, ulen - 1)); // entries of U on the right of the diagonal
284:     }
285:     PetscCallCUDA(cudaMemcpy(fs->csrVal, Ma, sizeof(*Ma) * Mnz, cudaMemcpyHostToDevice));

287:   #if PETSC_PKG_CUDA_VERSION_GE(12, 1, 1)
288:     if (fs->updatedSpSVAnalysis) { // have done cusparseSpSV_analysis before, and only matrix values changed?
289:       // Otherwise cusparse would error out: "On entry to cusparseSpSV_updateMatrix() parameter number 3 (newValues) had an illegal value: NULL pointer"
290:       if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_L, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
291:       if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_U, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
292:     } else
293:   #endif
294:     {
295:       // Do cusparseSpSV_analysis(), which is numeric and requires valid and up-to-date matrix values
296:       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));

298:       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));
299:       fs->updatedSpSVAnalysis          = PETSC_TRUE;
300:       fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;
301:     }
302:   }
303:   PetscFunctionReturn(PETSC_SUCCESS);
304: }
305: #else
306: static PetscErrorCode MatSeqAIJCUSPARSEBuildILULowerTriMatrix(Mat A)
307: {
308:   Mat_SeqAIJ                        *a                  = (Mat_SeqAIJ *)A->data;
309:   PetscInt                           n                  = A->rmap->n;
310:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
311:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
312:   const PetscInt                    *ai = a->i, *aj = a->j, *vi;
313:   const MatScalar                   *aa = a->a, *v;
314:   PetscInt                          *AiLo, *AjLo;
315:   PetscInt                           i, nz, nzLower, offset, rowOffset;

317:   PetscFunctionBegin;
318:   if (!n) PetscFunctionReturn(PETSC_SUCCESS);
319:   if (A->offloadmask == PETSC_OFFLOAD_UNALLOCATED || A->offloadmask == PETSC_OFFLOAD_CPU) {
320:     try {
321:       /* first figure out the number of nonzeros in the lower triangular matrix including 1's on the diagonal. */
322:       nzLower = n + ai[n] - ai[1];
323:       if (!loTriFactor) {
324:         PetscScalar *AALo;

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

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

332:         /* Fill the lower triangular matrix */
333:         AiLo[0]   = (PetscInt)0;
334:         AiLo[n]   = nzLower;
335:         AjLo[0]   = (PetscInt)0;
336:         AALo[0]   = (MatScalar)1.0;
337:         v         = aa;
338:         vi        = aj;
339:         offset    = 1;
340:         rowOffset = 1;
341:         for (i = 1; i < n; i++) {
342:           nz = ai[i + 1] - ai[i];
343:           /* additional 1 for the term on the diagonal */
344:           AiLo[i] = rowOffset;
345:           rowOffset += nz + 1;

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

350:           offset += nz;
351:           AjLo[offset] = (PetscInt)i;
352:           AALo[offset] = (MatScalar)1.0;
353:           offset += 1;

355:           v += nz;
356:           vi += nz;
357:         }

359:         /* allocate space for the triangular factor information */
360:         PetscCall(PetscNew(&loTriFactor));
361:         loTriFactor->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;
362:         /* Create the matrix description */
363:         PetscCallCUSPARSE(cusparseCreateMatDescr(&loTriFactor->descr));
364:         PetscCallCUSPARSE(cusparseSetMatIndexBase(loTriFactor->descr, CUSPARSE_INDEX_BASE_ZERO));
365:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
366:         PetscCallCUSPARSE(cusparseSetMatType(loTriFactor->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
367:   #else
368:         PetscCallCUSPARSE(cusparseSetMatType(loTriFactor->descr, CUSPARSE_MATRIX_TYPE_TRIANGULAR));
369:   #endif
370:         PetscCallCUSPARSE(cusparseSetMatFillMode(loTriFactor->descr, CUSPARSE_FILL_MODE_LOWER));
371:         PetscCallCUSPARSE(cusparseSetMatDiagType(loTriFactor->descr, CUSPARSE_DIAG_TYPE_UNIT));

373:         /* set the operation */
374:         loTriFactor->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

376:         /* set the matrix */
377:         loTriFactor->csrMat              = new CsrMatrix;
378:         loTriFactor->csrMat->num_rows    = n;
379:         loTriFactor->csrMat->num_cols    = n;
380:         loTriFactor->csrMat->num_entries = nzLower;

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

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

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

391:         /* Create the solve analysis information */
392:         PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
393:         PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&loTriFactor->solveInfo));
394:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
395:         PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, loTriFactor->solveOp, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_entries, loTriFactor->descr, loTriFactor->csrMat->values->data().get(),
396:                                                   loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, &loTriFactor->solveBufferSize));
397:         PetscCallCUDA(cudaMalloc(&loTriFactor->solveBuffer, loTriFactor->solveBufferSize));
398:   #endif

400:         /* perform the solve analysis */
401:         PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, loTriFactor->solveOp, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_entries, loTriFactor->descr, loTriFactor->csrMat->values->data().get(),
402:                                                   loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, loTriFactor->solvePolicy, loTriFactor->solveBuffer));
403:         PetscCallCUDA(WaitForCUDA());
404:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

406:         /* assign the pointer */
407:         ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->loTriFactorPtr = loTriFactor;
408:         loTriFactor->AA_h                                          = AALo;
409:         PetscCallCUDA(cudaFreeHost(AiLo));
410:         PetscCallCUDA(cudaFreeHost(AjLo));
411:         PetscCall(PetscLogCpuToGpu((n + 1 + nzLower) * sizeof(int) + nzLower * sizeof(PetscScalar)));
412:       } else { /* update values only */
413:         if (!loTriFactor->AA_h) PetscCallCUDA(cudaMallocHost((void **)&loTriFactor->AA_h, nzLower * sizeof(PetscScalar)));
414:         /* Fill the lower triangular matrix */
415:         loTriFactor->AA_h[0] = 1.0;
416:         v                    = aa;
417:         vi                   = aj;
418:         offset               = 1;
419:         for (i = 1; i < n; i++) {
420:           nz = ai[i + 1] - ai[i];
421:           PetscCall(PetscArraycpy(&loTriFactor->AA_h[offset], v, nz));
422:           offset += nz;
423:           loTriFactor->AA_h[offset] = 1.0;
424:           offset += 1;
425:           v += nz;
426:         }
427:         loTriFactor->csrMat->values->assign(loTriFactor->AA_h, loTriFactor->AA_h + nzLower);
428:         PetscCall(PetscLogCpuToGpu(nzLower * sizeof(PetscScalar)));
429:       }
430:     } catch (char *ex) {
431:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
432:     }
433:   }
434:   PetscFunctionReturn(PETSC_SUCCESS);
435: }

437: static PetscErrorCode MatSeqAIJCUSPARSEBuildILUUpperTriMatrix(Mat A)
438: {
439:   Mat_SeqAIJ                        *a                  = (Mat_SeqAIJ *)A->data;
440:   PetscInt                           n                  = A->rmap->n;
441:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
442:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
443:   const PetscInt                    *aj                 = a->j, *adiag, *vi;
444:   const MatScalar                   *aa                 = a->a, *v;
445:   PetscInt                          *AiUp, *AjUp;
446:   PetscInt                           i, nz, nzUpper, offset;

448:   PetscFunctionBegin;
449:   if (!n) PetscFunctionReturn(PETSC_SUCCESS);
450:   PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &adiag, NULL));
451:   if (A->offloadmask == PETSC_OFFLOAD_UNALLOCATED || A->offloadmask == PETSC_OFFLOAD_CPU) {
452:     try {
453:       /* next, figure out the number of nonzeros in the upper triangular matrix. */
454:       nzUpper = adiag[0] - adiag[n];
455:       if (!upTriFactor) {
456:         PetscScalar *AAUp;

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

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

464:         /* Fill the upper triangular matrix */
465:         AiUp[0] = (PetscInt)0;
466:         AiUp[n] = nzUpper;
467:         offset  = nzUpper;
468:         for (i = n - 1; i >= 0; i--) {
469:           v  = aa + adiag[i + 1] + 1;
470:           vi = aj + adiag[i + 1] + 1;

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

475:           /* decrement the offset */
476:           offset -= (nz + 1);

478:           /* first, set the diagonal elements */
479:           AjUp[offset] = (PetscInt)i;
480:           AAUp[offset] = (MatScalar)1. / v[nz];
481:           AiUp[i]      = AiUp[i + 1] - (nz + 1);

483:           PetscCall(PetscArraycpy(&AjUp[offset + 1], vi, nz));
484:           PetscCall(PetscArraycpy(&AAUp[offset + 1], v, nz));
485:         }

487:         /* allocate space for the triangular factor information */
488:         PetscCall(PetscNew(&upTriFactor));
489:         upTriFactor->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

491:         /* Create the matrix description */
492:         PetscCallCUSPARSE(cusparseCreateMatDescr(&upTriFactor->descr));
493:         PetscCallCUSPARSE(cusparseSetMatIndexBase(upTriFactor->descr, CUSPARSE_INDEX_BASE_ZERO));
494:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
495:         PetscCallCUSPARSE(cusparseSetMatType(upTriFactor->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
496:   #else
497:         PetscCallCUSPARSE(cusparseSetMatType(upTriFactor->descr, CUSPARSE_MATRIX_TYPE_TRIANGULAR));
498:   #endif
499:         PetscCallCUSPARSE(cusparseSetMatFillMode(upTriFactor->descr, CUSPARSE_FILL_MODE_UPPER));
500:         PetscCallCUSPARSE(cusparseSetMatDiagType(upTriFactor->descr, CUSPARSE_DIAG_TYPE_NON_UNIT));

502:         /* set the operation */
503:         upTriFactor->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

505:         /* set the matrix */
506:         upTriFactor->csrMat              = new CsrMatrix;
507:         upTriFactor->csrMat->num_rows    = n;
508:         upTriFactor->csrMat->num_cols    = n;
509:         upTriFactor->csrMat->num_entries = nzUpper;

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

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

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

520:         /* Create the solve analysis information */
521:         PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
522:         PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&upTriFactor->solveInfo));
523:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
524:         PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, upTriFactor->solveOp, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_entries, upTriFactor->descr, upTriFactor->csrMat->values->data().get(),
525:                                                   upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, &upTriFactor->solveBufferSize));
526:         PetscCallCUDA(cudaMalloc(&upTriFactor->solveBuffer, upTriFactor->solveBufferSize));
527:   #endif

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

533:         PetscCallCUDA(WaitForCUDA());
534:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

536:         /* assign the pointer */
537:         ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->upTriFactorPtr = upTriFactor;
538:         upTriFactor->AA_h                                          = AAUp;
539:         PetscCallCUDA(cudaFreeHost(AiUp));
540:         PetscCallCUDA(cudaFreeHost(AjUp));
541:         PetscCall(PetscLogCpuToGpu((n + 1 + nzUpper) * sizeof(int) + nzUpper * sizeof(PetscScalar)));
542:       } else {
543:         if (!upTriFactor->AA_h) PetscCallCUDA(cudaMallocHost((void **)&upTriFactor->AA_h, nzUpper * sizeof(PetscScalar)));
544:         /* Fill the upper triangular matrix */
545:         offset = nzUpper;
546:         for (i = n - 1; i >= 0; i--) {
547:           v = aa + adiag[i + 1] + 1;

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

552:           /* decrement the offset */
553:           offset -= (nz + 1);

555:           /* first, set the diagonal elements */
556:           upTriFactor->AA_h[offset] = 1. / v[nz];
557:           PetscCall(PetscArraycpy(&upTriFactor->AA_h[offset + 1], v, nz));
558:         }
559:         upTriFactor->csrMat->values->assign(upTriFactor->AA_h, upTriFactor->AA_h + nzUpper);
560:         PetscCall(PetscLogCpuToGpu(nzUpper * sizeof(PetscScalar)));
561:       }
562:     } catch (char *ex) {
563:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
564:     }
565:   }
566:   PetscFunctionReturn(PETSC_SUCCESS);
567: }
568: #endif

570: static PetscErrorCode MatSeqAIJCUSPARSEILUAnalysisAndCopyToGPU(Mat A)
571: {
572:   Mat_SeqAIJ                   *a                  = (Mat_SeqAIJ *)A->data;
573:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
574:   IS                            isrow = a->row, isicol = a->icol;
575:   PetscBool                     row_identity, col_identity;
576:   PetscInt                      n = A->rmap->n;

578:   PetscFunctionBegin;
579:   PetscCheck(cusparseTriFactors, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing cusparseTriFactors");
580: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
581:   PetscCall(MatSeqAIJCUSPARSEBuildFactoredMatrix_LU(A));
582: #else
583:   PetscCall(MatSeqAIJCUSPARSEBuildILULowerTriMatrix(A));
584:   PetscCall(MatSeqAIJCUSPARSEBuildILUUpperTriMatrix(A));
585:   if (!cusparseTriFactors->workVector) cusparseTriFactors->workVector = new THRUSTARRAY(n);
586: #endif

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

590:   A->offloadmask = PETSC_OFFLOAD_BOTH; // factored matrix is sync'ed to GPU
591:   /* lower triangular indices */
592:   PetscCall(ISIdentity(isrow, &row_identity));
593:   if (!row_identity && !cusparseTriFactors->rpermIndices) {
594:     const PetscInt *r;

596:     PetscCall(ISGetIndices(isrow, &r));
597:     cusparseTriFactors->rpermIndices = new THRUSTINTARRAY(n);
598:     cusparseTriFactors->rpermIndices->assign(r, r + n);
599:     PetscCall(ISRestoreIndices(isrow, &r));
600:     PetscCall(PetscLogCpuToGpu(n * sizeof(PetscInt)));
601:   }

603:   /* upper triangular indices */
604:   PetscCall(ISIdentity(isicol, &col_identity));
605:   if (!col_identity && !cusparseTriFactors->cpermIndices) {
606:     const PetscInt *c;

608:     PetscCall(ISGetIndices(isicol, &c));
609:     cusparseTriFactors->cpermIndices = new THRUSTINTARRAY(n);
610:     cusparseTriFactors->cpermIndices->assign(c, c + n);
611:     PetscCall(ISRestoreIndices(isicol, &c));
612:     PetscCall(PetscLogCpuToGpu(n * sizeof(PetscInt)));
613:   }
614:   PetscFunctionReturn(PETSC_SUCCESS);
615: }

617: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
618: static PetscErrorCode MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky(Mat A)
619: {
620:   Mat_SeqAIJ                   *a  = static_cast<Mat_SeqAIJ *>(A->data);
621:   PetscInt                      m  = A->rmap->n;
622:   Mat_SeqAIJCUSPARSETriFactors *fs = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
623:   const PetscInt               *Ai = a->i, *Aj = a->j, *adiag;
624:   const MatScalar              *Aa = a->a;
625:   PetscInt                     *Mj, Mnz;
626:   PetscScalar                  *Ma, *D;

628:   PetscFunctionBegin;
629:   PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &adiag, NULL));
630:   if (A->offloadmask == PETSC_OFFLOAD_CPU) { // A's latest factors are on CPU
631:     if (!fs->csrRowPtr) {                    // Is't the first time to do the setup? Use csrRowPtr since it is not null even m=0
632:       // Re-arrange the (skewed) factored matrix and put the result into M, a regular csr matrix on host.
633:       // See comments at MatICCFactorSymbolic_SeqAIJ() on the layout of the factored matrix (U) on host.
634:       Mnz = Ai[m]; // Unz (with the unit diagonal)
635:       PetscCall(PetscMalloc1(Mnz, &Ma));
636:       PetscCall(PetscMalloc1(Mnz, &Mj)); // Mj[] is temp
637:       PetscCall(PetscMalloc1(m, &D));    // the diagonal
638:       for (PetscInt i = 0; i < m; i++) {
639:         PetscInt ulen = Ai[i + 1] - Ai[i];
640:         Mj[Ai[i]]     = i;                                              // diagonal entry
641:         PetscCall(PetscArraycpy(Mj + Ai[i] + 1, Aj + Ai[i], ulen - 1)); // entries of U on the right of the diagonal
642:       }
643:       // Copy M (U) from host to device
644:       PetscCallCUDA(cudaMalloc(&fs->csrRowPtr, sizeof(*fs->csrRowPtr) * (m + 1)));
645:       PetscCallCUDA(cudaMalloc(&fs->csrColIdx, sizeof(*fs->csrColIdx) * Mnz));
646:       PetscCallCUDA(cudaMalloc(&fs->csrVal, sizeof(*fs->csrVal) * Mnz));
647:       PetscCallCUDA(cudaMalloc(&fs->diag, sizeof(*fs->diag) * m));
648:       PetscCallCUDA(cudaMemcpy(fs->csrRowPtr, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyHostToDevice));
649:       PetscCallCUDA(cudaMemcpy(fs->csrColIdx, Mj, sizeof(*Mj) * Mnz, cudaMemcpyHostToDevice));

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

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

664:       // Allocate work vectors in SpSv
665:       PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(*fs->X) * m));
666:       PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(*fs->Y) * m));

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

671:       // Query buffer sizes for SpSV and then allocate buffers
672:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_U));
673:       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));
674:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U));

676:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Ut)); // Ut solve uses the same matrix (spMatDescr_U), but different descr and buffer
677:       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));
678:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Ut, fs->spsvBufferSize_Ut));

680:       // Record for reuse
681:       fs->csrVal_h = Ma;
682:       fs->diag_h   = D;
683:       PetscCall(PetscFree(Mj));
684:     }
685:     // Copy the value
686:     Ma  = fs->csrVal_h;
687:     D   = fs->diag_h;
688:     Mnz = Ai[m];
689:     for (PetscInt i = 0; i < m; i++) {
690:       D[i]      = Aa[adiag[i]];   // actually Aa[adiag[i]] is the inverse of the diagonal
691:       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
692:       for (PetscInt k = 0; k < Ai[i + 1] - Ai[i] - 1; k++) Ma[Ai[i] + 1 + k] = -Aa[Ai[i] + k];
693:     }
694:     PetscCallCUDA(cudaMemcpy(fs->csrVal, Ma, sizeof(*Ma) * Mnz, cudaMemcpyHostToDevice));
695:     PetscCallCUDA(cudaMemcpy(fs->diag, D, sizeof(*D) * m, cudaMemcpyHostToDevice));

697:   #if PETSC_PKG_CUDA_VERSION_GE(12, 1, 1)
698:     if (fs->updatedSpSVAnalysis) {
699:       if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_U, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
700:       if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_Ut, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
701:     } else
702:   #endif
703:     {
704:       // Do cusparseSpSV_analysis(), which is numeric and requires valid and up-to-date matrix values
705:       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));
706:       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));
707:       fs->updatedSpSVAnalysis = PETSC_TRUE;
708:     }
709:   }
710:   PetscFunctionReturn(PETSC_SUCCESS);
711: }

713: // Solve Ut D U x = b
714: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_Cholesky(Mat A, Vec b, Vec x)
715: {
716:   Mat_SeqAIJCUSPARSETriFactors         *fs  = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
717:   Mat_SeqAIJ                           *aij = static_cast<Mat_SeqAIJ *>(A->data);
718:   const PetscScalar                    *barray;
719:   PetscScalar                          *xarray;
720:   thrust::device_ptr<const PetscScalar> bGPU;
721:   thrust::device_ptr<PetscScalar>       xGPU;
722:   const cusparseSpSVAlg_t               alg = CUSPARSE_SPSV_ALG_DEFAULT;
723:   PetscInt                              m   = A->rmap->n;

725:   PetscFunctionBegin;
726:   PetscCall(PetscLogGpuTimeBegin());
727:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
728:   PetscCall(VecCUDAGetArrayRead(b, &barray));
729:   xGPU = thrust::device_pointer_cast(xarray);
730:   bGPU = thrust::device_pointer_cast(barray);

732:   // Reorder b with the row permutation if needed, and wrap the result in fs->X
733:   if (fs->rpermIndices) {
734:     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)));
735:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
736:   } else {
737:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
738:   }

740:   // Solve Ut Y = X
741:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
742:   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));

744:   // Solve diag(D) Z = Y. Actually just do Y = Y*D since D is already inverted in MatCholeskyFactorNumeric_SeqAIJ().
745:   // It is basically a vector element-wise multiplication, but cublas does not have it!
746:   #if CCCL_VERSION >= 3001000
747:   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), cuda::std::multiplies<PetscScalar>()));
748:   #else
749:   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>()));
750:   #endif

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_Cholesky(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:   #if PETSC_PKG_CUDA_VERSION_GE(12, 1, 1)
1731:   if (fs->updatedSpSVAnalysis) {
1732:     if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_L, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
1733:     if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_U, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
1734:   } else
1735:   #endif
1736:   {
1737:     /* cusparseSpSV_analysis() is numeric, i.e., it requires valid matrix values, therefore, we do it after cusparseXcsrilu02()
1738:      See discussion at https://github.com/NVIDIA/CUDALibrarySamples/issues/78
1739:     */
1740:     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));

1742:     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));

1744:     fs->updatedSpSVAnalysis = PETSC_TRUE;
1745:     /* L, U values have changed, reset the flag to indicate we need to redo cusparseSpSV_analysis() for transpose solve */
1746:     fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;
1747:   }

1749:   fact->offloadmask            = PETSC_OFFLOAD_GPU;
1750:   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.
1751:   fact->ops->solvetranspose    = MatSolveTranspose_SeqAIJCUSPARSE_LU;
1752:   fact->ops->matsolve          = NULL;
1753:   fact->ops->matsolvetranspose = NULL;
1754:   PetscCall(PetscLogGpuTimeEnd());
1755:   PetscCall(PetscLogGpuFlops(fs->numericFactFlops));
1756:   PetscFunctionReturn(PETSC_SUCCESS);
1757: }

1759: static PetscErrorCode MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0(Mat fact, Mat A, IS, IS, const MatFactorInfo *info)
1760: {
1761:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1762:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
1763:   PetscInt                      m, nz;

1765:   PetscFunctionBegin;
1766:   if (PetscDefined(USE_DEBUG)) {
1767:     PetscBool flg, diagDense;

1769:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
1770:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
1771:     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);
1772:     PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, NULL, &diagDense));
1773:     PetscCheck(diagDense, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing a diagonal entry");
1774:   }

1776:   /* Free the old stale stuff */
1777:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&fs));

1779:   /* Copy over A's meta data to fact. Note that we also allocated fact's i,j,a on host,
1780:      but they will not be used. Allocate them just for easy debugging.
1781:    */
1782:   PetscCall(MatDuplicateNoCreate_SeqAIJ(fact, A, MAT_DO_NOT_COPY_VALUES, PETSC_TRUE /*malloc*/));

1784:   fact->offloadmask            = PETSC_OFFLOAD_BOTH;
1785:   fact->factortype             = MAT_FACTOR_ILU;
1786:   fact->info.factor_mallocs    = 0;
1787:   fact->info.fill_ratio_given  = info->fill;
1788:   fact->info.fill_ratio_needed = 1.0;

1790:   aij->row = NULL;
1791:   aij->col = NULL;

1793:   /* ====================================================================== */
1794:   /* Copy A's i, j to fact and also allocate the value array of fact.       */
1795:   /* We'll do in-place factorization on fact                                */
1796:   /* ====================================================================== */
1797:   const int *Ai, *Aj;

1799:   m  = fact->rmap->n;
1800:   nz = aij->nz;

1802:   PetscCallCUDA(cudaMalloc((void **)&fs->csrRowPtr32, sizeof(*fs->csrRowPtr32) * (m + 1)));
1803:   PetscCallCUDA(cudaMalloc((void **)&fs->csrColIdx32, sizeof(*fs->csrColIdx32) * nz));
1804:   PetscCallCUDA(cudaMalloc((void **)&fs->csrVal, sizeof(*fs->csrVal) * nz));
1805:   PetscCall(MatSeqAIJCUSPARSEGetIJ(A, PETSC_FALSE, &Ai, &Aj)); /* Do not use compressed Ai.  The returned Ai, Aj are 32-bit */
1806:   PetscCallCUDA(cudaMemcpyAsync(fs->csrRowPtr32, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));
1807:   PetscCallCUDA(cudaMemcpyAsync(fs->csrColIdx32, Aj, sizeof(*Aj) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

1809:   /* ====================================================================== */
1810:   /* Create descriptors for M, L, U                                         */
1811:   /* ====================================================================== */
1812:   cusparseFillMode_t fillMode;
1813:   cusparseDiagType_t diagType;

1815:   PetscCallCUSPARSE(cusparseCreateMatDescr(&fs->matDescr_M));
1816:   PetscCallCUSPARSE(cusparseSetMatIndexBase(fs->matDescr_M, CUSPARSE_INDEX_BASE_ZERO));
1817:   PetscCallCUSPARSE(cusparseSetMatType(fs->matDescr_M, CUSPARSE_MATRIX_TYPE_GENERAL));

1819:   /* https://docs.nvidia.com/cuda/cusparse/index.html#cusparseDiagType_t
1820:     cusparseDiagType_t: This type indicates if the matrix diagonal entries are unity. The diagonal elements are always
1821:     assumed to be present, but if CUSPARSE_DIAG_TYPE_UNIT is passed to an API routine, then the routine assumes that
1822:     all diagonal entries are unity and will not read or modify those entries. Note that in this case the routine
1823:     assumes the diagonal entries are equal to one, regardless of what those entries are actually set to in memory.
1824:   */
1825:   fillMode = CUSPARSE_FILL_MODE_LOWER;
1826:   diagType = CUSPARSE_DIAG_TYPE_UNIT;
1827:   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));
1828:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
1829:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

1831:   fillMode = CUSPARSE_FILL_MODE_UPPER;
1832:   diagType = CUSPARSE_DIAG_TYPE_NON_UNIT;
1833:   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));
1834:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
1835:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

1837:   /* ========================================================================= */
1838:   /* Query buffer sizes for csrilu0, SpSV and allocate buffers                 */
1839:   /* ========================================================================= */
1840:   PetscCallCUSPARSE(cusparseCreateCsrilu02Info(&fs->ilu0Info_M));
1841:   if (m)
1842:     PetscCallCUSPARSE(cusparseXcsrilu02_bufferSize(fs->handle, m, nz, /* cusparseXcsrilu02 errors out with empty matrices (m=0) */
1843:                                                    fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ilu0Info_M, &fs->factBufferSize_M));

1845:   PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(PetscScalar) * m));
1846:   PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(PetscScalar) * m));

1848:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_X, m, fs->X, cusparse_scalartype));
1849:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_Y, m, fs->Y, cusparse_scalartype));

1851:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_L));
1852:   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));

1854:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_U));
1855:   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));

1857:   /* From my experiment with the example at https://github.com/NVIDIA/CUDALibrarySamples/tree/master/cuSPARSE/bicgstab,
1858:      and discussion at https://github.com/NVIDIA/CUDALibrarySamples/issues/77,
1859:      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.
1860:      To save memory, we make factBuffer_M share with the bigger of spsvBuffer_L/U.
1861:    */
1862:   if (fs->spsvBufferSize_L > fs->spsvBufferSize_U) {
1863:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_L, (size_t)fs->factBufferSize_M)));
1864:     fs->spsvBuffer_L = fs->factBuffer_M;
1865:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U));
1866:   } else {
1867:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_U, (size_t)fs->factBufferSize_M)));
1868:     fs->spsvBuffer_U = fs->factBuffer_M;
1869:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L));
1870:   }

1872:   /* ========================================================================== */
1873:   /* Perform analysis of ilu0 on M, SpSv on L and U                             */
1874:   /* The lower(upper) triangular part of M has the same sparsity pattern as L(U)*/
1875:   /* ========================================================================== */
1876:   int              structural_zero;
1877:   cusparseStatus_t status;

1879:   fs->policy_M = CUSPARSE_SOLVE_POLICY_USE_LEVEL;
1880:   if (m)
1881:     PetscCallCUSPARSE(cusparseXcsrilu02_analysis(fs->handle, m, nz, /* cusparseXcsrilu02 errors out with empty matrices (m=0) */
1882:                                                  fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ilu0Info_M, fs->policy_M, fs->factBuffer_M));
1883:   if (PetscDefined(USE_DEBUG)) {
1884:     /* cusparseXcsrilu02_zeroPivot() is a blocking call. It calls cudaDeviceSynchronize() to make sure all previous kernels are done. */
1885:     status = cusparseXcsrilu02_zeroPivot(fs->handle, fs->ilu0Info_M, &structural_zero);
1886:     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);
1887:   }

1889:   /* Estimate FLOPs of the numeric factorization */
1890:   {
1891:     Mat_SeqAIJ     *Aseq = (Mat_SeqAIJ *)A->data;
1892:     PetscInt       *Ai, nzRow, nzLeft;
1893:     const PetscInt *adiag;
1894:     PetscLogDouble  flops = 0.0;

1896:     PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &adiag, NULL));
1897:     Ai = Aseq->i;
1898:     for (PetscInt i = 0; i < m; i++) {
1899:       if (Ai[i] < adiag[i] && adiag[i] < Ai[i + 1]) { /* There are nonzeros left to the diagonal of row i */
1900:         nzRow  = Ai[i + 1] - Ai[i];
1901:         nzLeft = adiag[i] - Ai[i];
1902:         /* We want to eliminate nonzeros left to the diagonal one by one. Assume each time, nonzeros right
1903:           and include the eliminated one will be updated, which incurs a multiplication and an addition.
1904:         */
1905:         nzLeft = (nzRow - 1) / 2;
1906:         flops += nzLeft * (2.0 * nzRow - nzLeft + 1);
1907:       }
1908:     }
1909:     fs->numericFactFlops = flops;
1910:   }
1911:   fact->ops->lufactornumeric = MatILUFactorNumeric_SeqAIJCUSPARSE_ILU0;
1912:   PetscFunctionReturn(PETSC_SUCCESS);
1913: }

1915: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_ICC0(Mat fact, Vec b, Vec x)
1916: {
1917:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1918:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
1919:   const PetscScalar            *barray;
1920:   PetscScalar                  *xarray;

1922:   PetscFunctionBegin;
1923:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
1924:   PetscCall(VecCUDAGetArrayRead(b, &barray));
1925:   PetscCall(PetscLogGpuTimeBegin());

1927:   /* Solve L*y = b */
1928:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
1929:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
1930:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, /* L Y = X */
1931:                                        fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L));

1933:   /* Solve Lt*x = y */
1934:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
1935:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, CUSPARSE_OPERATION_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, /* Lt X = Y */
1936:                                        fs->dnVecDescr_Y, fs->dnVecDescr_X, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_Lt));

1938:   PetscCall(VecCUDARestoreArrayRead(b, &barray));
1939:   PetscCall(VecCUDARestoreArrayWrite(x, &xarray));

1941:   PetscCall(PetscLogGpuTimeEnd());
1942:   PetscCall(PetscLogGpuFlops(2.0 * aij->nz - fact->rmap->n));
1943:   PetscFunctionReturn(PETSC_SUCCESS);
1944: }

1946: static PetscErrorCode MatICCFactorNumeric_SeqAIJCUSPARSE_ICC0(Mat fact, Mat A, const MatFactorInfo *)
1947: {
1948:   Mat_SeqAIJCUSPARSETriFactors *fs    = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1949:   Mat_SeqAIJ                   *aij   = (Mat_SeqAIJ *)fact->data;
1950:   Mat_SeqAIJCUSPARSE           *Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
1951:   CsrMatrix                    *Acsr;
1952:   PetscInt                      m, nz;
1953:   PetscBool                     flg;

1955:   PetscFunctionBegin;
1956:   if (PetscDefined(USE_DEBUG)) {
1957:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
1958:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
1959:   }

1961:   /* Copy A's value to fact */
1962:   m  = fact->rmap->n;
1963:   nz = aij->nz;
1964:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
1965:   Acsr = (CsrMatrix *)Acusp->mat->mat;
1966:   PetscCallCUDA(cudaMemcpyAsync(fs->csrVal, Acsr->values->data().get(), sizeof(PetscScalar) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

1968:   /* Factorize fact inplace */
1969:   /* https://docs.nvidia.com/cuda/cusparse/index.html#csric02_solve
1970:      csric02() only takes the lower triangular part of matrix A to perform factorization.
1971:      The matrix type must be CUSPARSE_MATRIX_TYPE_GENERAL, the fill mode and diagonal type are ignored,
1972:      and the strictly upper triangular part is ignored and never touched. It does not matter if A is Hermitian or not.
1973:      In other words, from the point of view of csric02() A is Hermitian and only the lower triangular part is provided.
1974:    */
1975:   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));
1976:   if (PetscDefined(USE_DEBUG)) {
1977:     int              numerical_zero;
1978:     cusparseStatus_t status;
1979:     status = cusparseXcsric02_zeroPivot(fs->handle, fs->ic0Info_M, &numerical_zero);
1980:     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);
1981:   }

1983:   #if PETSC_PKG_CUDA_VERSION_GE(12, 1, 1)
1984:   if (fs->updatedSpSVAnalysis) {
1985:     if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_L, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
1986:     if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_Lt, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
1987:   } else
1988:   #endif
1989:   {
1990:     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));

1992:     /* Note that cusparse reports this error if we use double and CUSPARSE_OPERATION_CONJUGATE_TRANSPOSE
1993:     ** On entry to cusparseSpSV_analysis(): conjugate transpose (opA) is not supported for matA data type, current -> CUDA_R_64F
1994:   */
1995:     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));
1996:     fs->updatedSpSVAnalysis = PETSC_TRUE;
1997:   }

1999:   fact->offloadmask            = PETSC_OFFLOAD_GPU;
2000:   fact->ops->solve             = MatSolve_SeqAIJCUSPARSE_ICC0;
2001:   fact->ops->solvetranspose    = MatSolve_SeqAIJCUSPARSE_ICC0;
2002:   fact->ops->matsolve          = NULL;
2003:   fact->ops->matsolvetranspose = NULL;
2004:   PetscCall(PetscLogGpuFlops(fs->numericFactFlops));
2005:   PetscFunctionReturn(PETSC_SUCCESS);
2006: }

2008: static PetscErrorCode MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0(Mat fact, Mat A, IS, const MatFactorInfo *info)
2009: {
2010:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
2011:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
2012:   PetscInt                      m, nz;

2014:   PetscFunctionBegin;
2015:   if (PetscDefined(USE_DEBUG)) {
2016:     PetscBool flg, diagDense;

2018:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2019:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
2020:     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);
2021:     PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, NULL, &diagDense));
2022:     PetscCheck(diagDense, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entries");
2023:   }

2025:   /* Free the old stale stuff */
2026:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&fs));

2028:   /* Copy over A's meta data to fact. Note that we also allocated fact's i,j,a on host,
2029:      but they will not be used. Allocate them just for easy debugging.
2030:    */
2031:   PetscCall(MatDuplicateNoCreate_SeqAIJ(fact, A, MAT_DO_NOT_COPY_VALUES, PETSC_TRUE /*malloc*/));

2033:   fact->offloadmask            = PETSC_OFFLOAD_BOTH;
2034:   fact->factortype             = MAT_FACTOR_ICC;
2035:   fact->info.factor_mallocs    = 0;
2036:   fact->info.fill_ratio_given  = info->fill;
2037:   fact->info.fill_ratio_needed = 1.0;

2039:   aij->row = NULL;
2040:   aij->col = NULL;

2042:   /* ====================================================================== */
2043:   /* Copy A's i, j to fact and also allocate the value array of fact.       */
2044:   /* We'll do in-place factorization on fact                                */
2045:   /* ====================================================================== */
2046:   const int *Ai, *Aj;

2048:   m  = fact->rmap->n;
2049:   nz = aij->nz;

2051:   PetscCallCUDA(cudaMalloc((void **)&fs->csrRowPtr32, sizeof(*fs->csrRowPtr32) * (m + 1)));
2052:   PetscCallCUDA(cudaMalloc((void **)&fs->csrColIdx32, sizeof(*fs->csrColIdx32) * nz));
2053:   PetscCallCUDA(cudaMalloc((void **)&fs->csrVal, sizeof(PetscScalar) * nz));
2054:   PetscCall(MatSeqAIJCUSPARSEGetIJ(A, PETSC_FALSE, &Ai, &Aj)); /* Do not use compressed Ai */
2055:   PetscCallCUDA(cudaMemcpyAsync(fs->csrRowPtr32, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));
2056:   PetscCallCUDA(cudaMemcpyAsync(fs->csrColIdx32, Aj, sizeof(*Aj) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

2058:   /* ====================================================================== */
2059:   /* Create mat descriptors for M, L                                        */
2060:   /* ====================================================================== */
2061:   cusparseFillMode_t fillMode;
2062:   cusparseDiagType_t diagType;

2064:   PetscCallCUSPARSE(cusparseCreateMatDescr(&fs->matDescr_M));
2065:   PetscCallCUSPARSE(cusparseSetMatIndexBase(fs->matDescr_M, CUSPARSE_INDEX_BASE_ZERO));
2066:   PetscCallCUSPARSE(cusparseSetMatType(fs->matDescr_M, CUSPARSE_MATRIX_TYPE_GENERAL));

2068:   /* https://docs.nvidia.com/cuda/cusparse/index.html#cusparseDiagType_t
2069:     cusparseDiagType_t: This type indicates if the matrix diagonal entries are unity. The diagonal elements are always
2070:     assumed to be present, but if CUSPARSE_DIAG_TYPE_UNIT is passed to an API routine, then the routine assumes that
2071:     all diagonal entries are unity and will not read or modify those entries. Note that in this case the routine
2072:     assumes the diagonal entries are equal to one, regardless of what those entries are actually set to in memory.
2073:   */
2074:   fillMode = CUSPARSE_FILL_MODE_LOWER;
2075:   diagType = CUSPARSE_DIAG_TYPE_NON_UNIT;
2076:   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));
2077:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
2078:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

2080:   /* ========================================================================= */
2081:   /* Query buffer sizes for csric0, SpSV of L and Lt, and allocate buffers     */
2082:   /* ========================================================================= */
2083:   PetscCallCUSPARSE(cusparseCreateCsric02Info(&fs->ic0Info_M));
2084:   if (m) PetscCallCUSPARSE(cusparseXcsric02_bufferSize(fs->handle, m, nz, fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ic0Info_M, &fs->factBufferSize_M));

2086:   PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(PetscScalar) * m));
2087:   PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(PetscScalar) * m));

2089:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_X, m, fs->X, cusparse_scalartype));
2090:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_Y, m, fs->Y, cusparse_scalartype));

2092:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_L));
2093:   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));

2095:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Lt));
2096:   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));

2098:   /* To save device memory, we make the factorization buffer share with one of the solver buffer.
2099:      See also comments in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0().
2100:    */
2101:   if (fs->spsvBufferSize_L > fs->spsvBufferSize_Lt) {
2102:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_L, (size_t)fs->factBufferSize_M)));
2103:     fs->spsvBuffer_L = fs->factBuffer_M;
2104:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Lt, fs->spsvBufferSize_Lt));
2105:   } else {
2106:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_Lt, (size_t)fs->factBufferSize_M)));
2107:     fs->spsvBuffer_Lt = fs->factBuffer_M;
2108:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L));
2109:   }

2111:   /* ========================================================================== */
2112:   /* Perform analysis of ic0 on M                                               */
2113:   /* The lower triangular part of M has the same sparsity pattern as L          */
2114:   /* ========================================================================== */
2115:   int              structural_zero;
2116:   cusparseStatus_t status;

2118:   fs->policy_M = CUSPARSE_SOLVE_POLICY_USE_LEVEL;
2119:   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));
2120:   if (PetscDefined(USE_DEBUG)) {
2121:     /* cusparseXcsric02_zeroPivot() is a blocking call. It calls cudaDeviceSynchronize() to make sure all previous kernels are done. */
2122:     status = cusparseXcsric02_zeroPivot(fs->handle, fs->ic0Info_M, &structural_zero);
2123:     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);
2124:   }

2126:   /* Estimate FLOPs of the numeric factorization */
2127:   {
2128:     Mat_SeqAIJ    *Aseq = (Mat_SeqAIJ *)A->data;
2129:     PetscInt      *Ai, nzRow, nzLeft;
2130:     PetscLogDouble flops = 0.0;

2132:     Ai = Aseq->i;
2133:     for (PetscInt i = 0; i < m; i++) {
2134:       nzRow = Ai[i + 1] - Ai[i];
2135:       if (nzRow > 1) {
2136:         /* We want to eliminate nonzeros left to the diagonal one by one. Assume each time, nonzeros right
2137:           and include the eliminated one will be updated, which incurs a multiplication and an addition.
2138:         */
2139:         nzLeft = (nzRow - 1) / 2;
2140:         flops += nzLeft * (2.0 * nzRow - nzLeft + 1);
2141:       }
2142:     }
2143:     fs->numericFactFlops = flops;
2144:   }
2145:   fact->ops->choleskyfactornumeric = MatICCFactorNumeric_SeqAIJCUSPARSE_ICC0;
2146:   PetscFunctionReturn(PETSC_SUCCESS);
2147: }
2148: #endif

2150: static PetscErrorCode MatLUFactorNumeric_SeqAIJCUSPARSE(Mat B, Mat A, const MatFactorInfo *info)
2151: {
2152:   // use_cpu_solve is a field in Mat_SeqAIJCUSPARSE. B, a factored matrix, uses Mat_SeqAIJCUSPARSETriFactors.
2153:   Mat_SeqAIJCUSPARSE *cusparsestruct = static_cast<Mat_SeqAIJCUSPARSE *>(A->spptr);

2155:   PetscFunctionBegin;
2156:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2157:   PetscCall(MatLUFactorNumeric_SeqAIJ(B, A, info));
2158:   B->offloadmask = PETSC_OFFLOAD_CPU;

2160:   if (!cusparsestruct->use_cpu_solve) {
2161: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2162:     B->ops->solve          = MatSolve_SeqAIJCUSPARSE_LU;
2163:     B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE_LU;
2164: #else
2165:     /* determine which version of MatSolve needs to be used. */
2166:     Mat_SeqAIJ *b     = (Mat_SeqAIJ *)B->data;
2167:     IS          isrow = b->row, iscol = b->col;
2168:     PetscBool   row_identity, col_identity;

2170:     PetscCall(ISIdentity(isrow, &row_identity));
2171:     PetscCall(ISIdentity(iscol, &col_identity));
2172:     if (row_identity && col_identity) {
2173:       B->ops->solve          = MatSolve_SeqAIJCUSPARSE_NaturalOrdering;
2174:       B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering;
2175:     } else {
2176:       B->ops->solve          = MatSolve_SeqAIJCUSPARSE;
2177:       B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE;
2178:     }
2179: #endif
2180:   }
2181:   B->ops->matsolve          = NULL;
2182:   B->ops->matsolvetranspose = NULL;

2184:   /* get the triangular factors */
2185:   if (!cusparsestruct->use_cpu_solve) PetscCall(MatSeqAIJCUSPARSEILUAnalysisAndCopyToGPU(B));
2186:   PetscFunctionReturn(PETSC_SUCCESS);
2187: }

2189: static PetscErrorCode MatLUFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS isrow, IS iscol, const MatFactorInfo *info)
2190: {
2191:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(B->spptr);

2193:   PetscFunctionBegin;
2194:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2195:   PetscCall(MatLUFactorSymbolic_SeqAIJ(B, A, isrow, iscol, info));
2196:   B->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJCUSPARSE;
2197:   PetscFunctionReturn(PETSC_SUCCESS);
2198: }

2200: static PetscErrorCode MatILUFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS isrow, IS iscol, const MatFactorInfo *info)
2201: {
2202:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2204:   PetscFunctionBegin;
2205: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2206:   PetscBool row_identity = PETSC_FALSE, col_identity = PETSC_FALSE;
2207:   if (!info->factoronhost) {
2208:     PetscCall(ISIdentity(isrow, &row_identity));
2209:     PetscCall(ISIdentity(iscol, &col_identity));
2210:   }
2211:   if (!info->levels && row_identity && col_identity) {
2212:     PetscCall(MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0(B, A, isrow, iscol, info));
2213:   } else
2214: #endif
2215:   {
2216:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2217:     PetscCall(MatILUFactorSymbolic_SeqAIJ(B, A, isrow, iscol, info));
2218:     B->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJCUSPARSE;
2219:   }
2220:   PetscFunctionReturn(PETSC_SUCCESS);
2221: }

2223: static PetscErrorCode MatICCFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS perm, const MatFactorInfo *info)
2224: {
2225:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2227:   PetscFunctionBegin;
2228: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2229:   PetscBool perm_identity = PETSC_FALSE;
2230:   if (!info->factoronhost) PetscCall(ISIdentity(perm, &perm_identity));
2231:   if (!info->levels && perm_identity) {
2232:     PetscCall(MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0(B, A, perm, info));
2233:   } else
2234: #endif
2235:   {
2236:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2237:     PetscCall(MatICCFactorSymbolic_SeqAIJ(B, A, perm, info));
2238:     B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJCUSPARSE;
2239:   }
2240:   PetscFunctionReturn(PETSC_SUCCESS);
2241: }

2243: static PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS perm, const MatFactorInfo *info)
2244: {
2245:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2247:   PetscFunctionBegin;
2248:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2249:   PetscCall(MatCholeskyFactorSymbolic_SeqAIJ(B, A, perm, info));
2250:   B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJCUSPARSE;
2251:   PetscFunctionReturn(PETSC_SUCCESS);
2252: }

2254: static PetscErrorCode MatFactorGetSolverType_seqaij_cusparse(Mat, MatSolverType *type)
2255: {
2256:   PetscFunctionBegin;
2257:   *type = MATSOLVERCUSPARSE;
2258:   PetscFunctionReturn(PETSC_SUCCESS);
2259: }

2261: /*MC
2262:   MATSOLVERCUSPARSE = "cusparse" - A matrix type providing triangular solvers for seq matrices
2263:   on a single GPU of type, `MATSEQAIJCUSPARSE`. Currently supported
2264:   algorithms are ILU(k) and ICC(k). Typically, deeper factorizations (larger k) results in poorer
2265:   performance in the triangular solves. Full LU, and Cholesky decompositions can be solved through the
2266:   CuSPARSE triangular solve algorithm. However, the performance can be quite poor and thus these
2267:   algorithms are not recommended. This class does NOT support direct solver operations.

2269:   Level: beginner

2271: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJCUSPARSE`, `PCFactorSetMatSolverType()`, `MatSolverType`, `MatCreateSeqAIJCUSPARSE()`,
2272:           `MATAIJCUSPARSE`, `MatCreateAIJCUSPARSE()`, `MatCUSPARSESetFormat()`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
2273: M*/

2275: PETSC_EXTERN PetscErrorCode MatGetFactor_seqaijcusparse_cusparse(Mat A, MatFactorType ftype, Mat *B)
2276: {
2277:   PetscInt n = A->rmap->n;

2279:   PetscFunctionBegin;
2280:   PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
2281:   PetscCall(MatSetSizes(*B, n, n, n, n));
2282:   (*B)->factortype = ftype; // factortype makes MatSetType() allocate spptr of type Mat_SeqAIJCUSPARSETriFactors
2283:   PetscCall(MatSetType(*B, MATSEQAIJCUSPARSE));

2285:   if (A->boundtocpu && A->bindingpropagates) PetscCall(MatBindToCPU(*B, PETSC_TRUE));
2286:   if (ftype == MAT_FACTOR_LU || ftype == MAT_FACTOR_ILU || ftype == MAT_FACTOR_ILUDT) {
2287:     PetscCall(MatSetBlockSizesFromMats(*B, A, A));
2288:     if (!A->boundtocpu) {
2289:       (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJCUSPARSE;
2290:       (*B)->ops->lufactorsymbolic  = MatLUFactorSymbolic_SeqAIJCUSPARSE;
2291:     } else {
2292:       (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJ;
2293:       (*B)->ops->lufactorsymbolic  = MatLUFactorSymbolic_SeqAIJ;
2294:     }
2295:     PetscCall(PetscStrallocpy(MATORDERINGND, (char **)&(*B)->preferredordering[MAT_FACTOR_LU]));
2296:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ILU]));
2297:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ILUDT]));
2298:   } else if (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) {
2299:     if (!A->boundtocpu) {
2300:       (*B)->ops->iccfactorsymbolic      = MatICCFactorSymbolic_SeqAIJCUSPARSE;
2301:       (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJCUSPARSE;
2302:     } else {
2303:       (*B)->ops->iccfactorsymbolic      = MatICCFactorSymbolic_SeqAIJ;
2304:       (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJ;
2305:     }
2306:     PetscCall(PetscStrallocpy(MATORDERINGND, (char **)&(*B)->preferredordering[MAT_FACTOR_CHOLESKY]));
2307:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ICC]));
2308:   } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Factor type not supported for CUSPARSE Matrix Types");

2310:   PetscCall(MatSeqAIJSetPreallocation(*B, MAT_SKIP_ALLOCATION, NULL));
2311:   (*B)->canuseordering = PETSC_TRUE;
2312:   PetscCall(PetscObjectComposeFunction((PetscObject)*B, "MatFactorGetSolverType_C", MatFactorGetSolverType_seqaij_cusparse));
2313:   PetscFunctionReturn(PETSC_SUCCESS);
2314: }

2316: static PetscErrorCode MatSeqAIJCUSPARSECopyFromGPU(Mat A)
2317: {
2318:   Mat_SeqAIJ         *a    = (Mat_SeqAIJ *)A->data;
2319:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2320: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2321:   Mat_SeqAIJCUSPARSETriFactors *fs = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
2322: #endif

2324:   PetscFunctionBegin;
2325:   if (A->offloadmask == PETSC_OFFLOAD_GPU) {
2326:     PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyFromGPU, A, 0, 0, 0));
2327:     if (A->factortype == MAT_FACTOR_NONE) {
2328:       CsrMatrix *matrix = (CsrMatrix *)cusp->mat->mat;
2329:       PetscCallCUDA(cudaMemcpy(a->a, matrix->values->data().get(), a->nz * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
2330:     }
2331: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2332:     else if (fs->csrVal) {
2333:       /* We have a factorized matrix on device and are able to copy it to host */
2334:       PetscCallCUDA(cudaMemcpy(a->a, fs->csrVal, a->nz * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
2335:     }
2336: #endif
2337:     else
2338:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for copying this type of factorized matrix from device to host");
2339:     PetscCall(PetscLogGpuToCpu(a->nz * sizeof(PetscScalar)));
2340:     PetscCall(PetscLogEventEnd(MAT_CUSPARSECopyFromGPU, A, 0, 0, 0));
2341:     A->offloadmask = PETSC_OFFLOAD_BOTH;
2342:   }
2343:   PetscFunctionReturn(PETSC_SUCCESS);
2344: }

2346: static PetscErrorCode MatSeqAIJGetArray_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2347: {
2348:   PetscFunctionBegin;
2349:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2350:   *array = ((Mat_SeqAIJ *)A->data)->a;
2351:   PetscFunctionReturn(PETSC_SUCCESS);
2352: }

2354: static PetscErrorCode MatSeqAIJRestoreArray_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2355: {
2356:   PetscFunctionBegin;
2357:   A->offloadmask = PETSC_OFFLOAD_CPU;
2358:   *array         = NULL;
2359:   PetscFunctionReturn(PETSC_SUCCESS);
2360: }

2362: static PetscErrorCode MatSeqAIJGetArrayRead_SeqAIJCUSPARSE(Mat A, const PetscScalar *array[])
2363: {
2364:   PetscFunctionBegin;
2365:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2366:   *array = ((Mat_SeqAIJ *)A->data)->a;
2367:   PetscFunctionReturn(PETSC_SUCCESS);
2368: }

2370: static PetscErrorCode MatSeqAIJRestoreArrayRead_SeqAIJCUSPARSE(Mat, const PetscScalar *array[])
2371: {
2372:   PetscFunctionBegin;
2373:   *array = NULL;
2374:   PetscFunctionReturn(PETSC_SUCCESS);
2375: }

2377: static PetscErrorCode MatSeqAIJGetArrayWrite_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2378: {
2379:   PetscFunctionBegin;
2380:   *array = ((Mat_SeqAIJ *)A->data)->a;
2381:   PetscFunctionReturn(PETSC_SUCCESS);
2382: }

2384: static PetscErrorCode MatSeqAIJRestoreArrayWrite_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2385: {
2386:   PetscFunctionBegin;
2387:   A->offloadmask = PETSC_OFFLOAD_CPU;
2388:   *array         = NULL;
2389:   PetscFunctionReturn(PETSC_SUCCESS);
2390: }

2392: static PetscErrorCode MatSeqAIJGetCSRAndMemType_SeqAIJCUSPARSE(Mat A, const PetscInt **i, const PetscInt **j, PetscScalar **a, PetscMemType *mtype)
2393: {
2394:   Mat_SeqAIJCUSPARSE *cusp;
2395:   CsrMatrix          *matrix;

2397:   PetscFunctionBegin;
2398:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
2399:   PetscCheck(A->factortype == MAT_FACTOR_NONE, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2400:   cusp = static_cast<Mat_SeqAIJCUSPARSE *>(A->spptr);
2401:   PetscCheck(cusp != NULL, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "cusp is NULL");
2402:   matrix = (CsrMatrix *)cusp->mat->mat;

2404:   if (i) {
2405: #if !defined(PETSC_USE_64BIT_INDICES)
2406:     *i = matrix->row_offsets->data().get();
2407: #else
2408:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSparse does not supported 64-bit indices");
2409: #endif
2410:   }
2411:   if (j) {
2412: #if !defined(PETSC_USE_64BIT_INDICES)
2413:     *j = matrix->column_indices->data().get();
2414: #else
2415:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSparse does not supported 64-bit indices");
2416: #endif
2417:   }
2418:   if (a) *a = matrix->values->data().get();
2419:   if (mtype) *mtype = PETSC_MEMTYPE_CUDA;
2420:   PetscFunctionReturn(PETSC_SUCCESS);
2421: }

2423: PETSC_INTERN PetscErrorCode MatSeqAIJCUSPARSECopyToGPU(Mat A)
2424: {
2425:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
2426:   Mat_SeqAIJCUSPARSEMultStruct *matstruct      = cusparsestruct->mat;
2427:   Mat_SeqAIJ                   *a              = (Mat_SeqAIJ *)A->data;
2428:   PetscInt                      m              = A->rmap->n, *ii, *ridx, tmp;
2429:   cusparseStatus_t              stat;
2430:   PetscBool                     both = PETSC_TRUE;

2432:   PetscFunctionBegin;
2433:   PetscCheck(!A->boundtocpu, PETSC_COMM_SELF, PETSC_ERR_GPU, "Cannot copy to GPU");
2434:   if (A->offloadmask == PETSC_OFFLOAD_UNALLOCATED || A->offloadmask == PETSC_OFFLOAD_CPU) {
2435:     if (A->nonzerostate == cusparsestruct->nonzerostate && cusparsestruct->format == MAT_CUSPARSE_CSR) { /* Copy values only */
2436:       CsrMatrix *matrix;
2437:       matrix = (CsrMatrix *)cusparsestruct->mat->mat;

2439:       PetscCheck(!a->nz || a->a, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR values");
2440:       PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2441:       matrix->values->assign(a->a, a->a + a->nz);
2442:       PetscCallCUDA(WaitForCUDA());
2443:       PetscCall(PetscLogCpuToGpu(a->nz * sizeof(PetscScalar)));
2444:       PetscCall(PetscLogEventEnd(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2445:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
2446:     } else {
2447:       PetscInt nnz;
2448:       PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2449:       PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusparsestruct->mat, cusparsestruct->format));
2450:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_TRUE));
2451:       delete cusparsestruct->workVector;
2452:       delete cusparsestruct->rowoffsets_gpu;
2453:       cusparsestruct->workVector     = NULL;
2454:       cusparsestruct->rowoffsets_gpu = NULL;
2455:       try {
2456:         if (a->compressedrow.use) {
2457:           m    = a->compressedrow.nrows;
2458:           ii   = a->compressedrow.i;
2459:           ridx = a->compressedrow.rindex;
2460:         } else {
2461:           m    = A->rmap->n;
2462:           ii   = a->i;
2463:           ridx = NULL;
2464:         }
2465:         PetscCheck(ii, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR row data");
2466:         if (!a->a) {
2467:           nnz  = ii[m];
2468:           both = PETSC_FALSE;
2469:         } else nnz = a->nz;
2470:         PetscCheck(!nnz || a->j, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR column data");

2472:         /* create cusparse matrix */
2473:         cusparsestruct->nrows = m;
2474:         matstruct             = new Mat_SeqAIJCUSPARSEMultStruct;
2475:         PetscCallCUSPARSE(cusparseCreateMatDescr(&matstruct->descr));
2476:         PetscCallCUSPARSE(cusparseSetMatIndexBase(matstruct->descr, CUSPARSE_INDEX_BASE_ZERO));
2477:         PetscCallCUSPARSE(cusparseSetMatType(matstruct->descr, CUSPARSE_MATRIX_TYPE_GENERAL));

2479:         PetscCallCUDA(cudaMalloc((void **)&matstruct->alpha_one, sizeof(PetscScalar)));
2480:         PetscCallCUDA(cudaMalloc((void **)&matstruct->beta_zero, sizeof(PetscScalar)));
2481:         PetscCallCUDA(cudaMalloc((void **)&matstruct->beta_one, sizeof(PetscScalar)));
2482:         PetscCallCUDA(cudaMemcpy(matstruct->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2483:         PetscCallCUDA(cudaMemcpy(matstruct->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2484:         PetscCallCUDA(cudaMemcpy(matstruct->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2485:         PetscCallCUSPARSE(cusparseSetPointerMode(cusparsestruct->handle, CUSPARSE_POINTER_MODE_DEVICE));

2487:         /* Build a hybrid/ellpack matrix if this option is chosen for the storage */
2488:         if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
2489:           /* set the matrix */
2490:           CsrMatrix *mat   = new CsrMatrix;
2491:           mat->num_rows    = m;
2492:           mat->num_cols    = A->cmap->n;
2493:           mat->num_entries = nnz;
2494:           PetscCallCXX(mat->row_offsets = new THRUSTINTARRAY32(m + 1));
2495:           mat->row_offsets->assign(ii, ii + m + 1);
2496:           PetscCallCXX(mat->column_indices = new THRUSTINTARRAY32(nnz));
2497:           mat->column_indices->assign(a->j, a->j + nnz);

2499:           PetscCallCXX(mat->values = new THRUSTARRAY(nnz));
2500:           if (a->a) mat->values->assign(a->a, a->a + nnz);

2502:           /* assign the pointer */
2503:           matstruct->mat = mat;
2504: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2505:           if (mat->num_rows) { /* cusparse errors on empty matrices! */
2506:             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 */
2507:                                      CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype);
2508:             PetscCallCUSPARSE(stat);
2509:           }
2510: #endif
2511:         } else if (cusparsestruct->format == MAT_CUSPARSE_ELL || cusparsestruct->format == MAT_CUSPARSE_HYB) {
2512: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2513:           SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
2514: #else
2515:           CsrMatrix *mat   = new CsrMatrix;
2516:           mat->num_rows    = m;
2517:           mat->num_cols    = A->cmap->n;
2518:           mat->num_entries = nnz;
2519:           PetscCallCXX(mat->row_offsets = new THRUSTINTARRAY32(m + 1));
2520:           mat->row_offsets->assign(ii, ii + m + 1);

2522:           PetscCallCXX(mat->column_indices = new THRUSTINTARRAY32(nnz));
2523:           mat->column_indices->assign(a->j, a->j + nnz);

2525:           PetscCallCXX(mat->values = new THRUSTARRAY(nnz));
2526:           if (a->a) mat->values->assign(a->a, a->a + nnz);

2528:           cusparseHybMat_t hybMat;
2529:           PetscCallCUSPARSE(cusparseCreateHybMat(&hybMat));
2530:           cusparseHybPartition_t partition = cusparsestruct->format == MAT_CUSPARSE_ELL ? CUSPARSE_HYB_PARTITION_MAX : CUSPARSE_HYB_PARTITION_AUTO;
2531:           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);
2532:           PetscCallCUSPARSE(stat);
2533:           /* assign the pointer */
2534:           matstruct->mat = hybMat;

2536:           if (mat) {
2537:             if (mat->values) delete (THRUSTARRAY *)mat->values;
2538:             if (mat->column_indices) delete (THRUSTINTARRAY32 *)mat->column_indices;
2539:             if (mat->row_offsets) delete (THRUSTINTARRAY32 *)mat->row_offsets;
2540:             delete (CsrMatrix *)mat;
2541:           }
2542: #endif
2543:         }

2545:         /* assign the compressed row indices */
2546:         if (a->compressedrow.use) {
2547:           PetscCallCXX(cusparsestruct->workVector = new THRUSTARRAY(m));
2548:           PetscCallCXX(matstruct->cprowIndices = new THRUSTINTARRAY(m));
2549:           matstruct->cprowIndices->assign(ridx, ridx + m);
2550:           tmp = m;
2551:         } else {
2552:           cusparsestruct->workVector = NULL;
2553:           matstruct->cprowIndices    = NULL;
2554:           tmp                        = 0;
2555:         }
2556:         PetscCall(PetscLogCpuToGpu(((m + 1) + (a->nz)) * sizeof(int) + tmp * sizeof(PetscInt) + (3 + (a->nz)) * sizeof(PetscScalar)));

2558:         /* assign the pointer */
2559:         cusparsestruct->mat = matstruct;
2560:       } catch (char *ex) {
2561:         SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
2562:       }
2563:       PetscCallCUDA(WaitForCUDA());
2564:       PetscCall(PetscLogEventEnd(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2565:       cusparsestruct->nonzerostate = A->nonzerostate;
2566:     }
2567:     if (both) A->offloadmask = PETSC_OFFLOAD_BOTH;
2568:   }
2569:   PetscFunctionReturn(PETSC_SUCCESS);
2570: }

2572: struct VecCUDAPlusEquals {
2573:   template <typename Tuple>
2574:   __host__ __device__ void operator()(Tuple t)
2575:   {
2576:     thrust::get<1>(t) = thrust::get<1>(t) + thrust::get<0>(t);
2577:   }
2578: };

2580: struct VecCUDAEquals {
2581:   template <typename Tuple>
2582:   __host__ __device__ void operator()(Tuple t)
2583:   {
2584:     thrust::get<1>(t) = thrust::get<0>(t);
2585:   }
2586: };

2588: struct VecCUDAEqualsReverse {
2589:   template <typename Tuple>
2590:   __host__ __device__ void operator()(Tuple t)
2591:   {
2592:     thrust::get<0>(t) = thrust::get<1>(t);
2593:   }
2594: };

2596: struct MatProductCtx_MatMatCusparse {
2597:   PetscBool      cisdense;
2598:   PetscScalar   *Bt;
2599:   Mat            X;
2600:   PetscBool      reusesym; /* Cusparse does not have split symbolic and numeric phases for sparse matmat operations */
2601:   PetscLogDouble flops;
2602:   CsrMatrix     *Bcsr;

2604: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2605:   cusparseSpMatDescr_t matSpBDescr;
2606:   PetscBool            initialized; /* C = alpha op(A) op(B) + beta C */
2607:   cusparseDnMatDescr_t matBDescr;
2608:   cusparseDnMatDescr_t matCDescr;
2609:   PetscInt             Blda, Clda; /* Record leading dimensions of B and C here to detect changes*/
2610:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2611:   void *dBuffer4;
2612:   void *dBuffer5;
2613:   #endif
2614:   size_t                mmBufferSize;
2615:   void                 *mmBuffer;
2616:   void                 *mmBuffer2; /* SpGEMM WorkEstimation buffer */
2617:   cusparseSpGEMMDescr_t spgemmDesc;
2618: #endif
2619: };

2621: static PetscErrorCode MatProductCtxDestroy_MatMatCusparse(void **data)
2622: {
2623:   MatProductCtx_MatMatCusparse *mmdata = *(MatProductCtx_MatMatCusparse **)data;

2625:   PetscFunctionBegin;
2626:   PetscCallCUDA(cudaFree(mmdata->Bt));
2627:   delete mmdata->Bcsr;
2628: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2629:   if (mmdata->matSpBDescr) PetscCallCUSPARSE(cusparseDestroySpMat(mmdata->matSpBDescr));
2630:   if (mmdata->matBDescr) PetscCallCUSPARSE(cusparseDestroyDnMat(mmdata->matBDescr));
2631:   if (mmdata->matCDescr) PetscCallCUSPARSE(cusparseDestroyDnMat(mmdata->matCDescr));
2632:   if (mmdata->spgemmDesc) PetscCallCUSPARSE(cusparseSpGEMM_destroyDescr(mmdata->spgemmDesc));
2633:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2634:   if (mmdata->dBuffer4) PetscCallCUDA(cudaFree(mmdata->dBuffer4));
2635:   if (mmdata->dBuffer5) PetscCallCUDA(cudaFree(mmdata->dBuffer5));
2636:   #endif
2637:   if (mmdata->mmBuffer) PetscCallCUDA(cudaFree(mmdata->mmBuffer));
2638:   if (mmdata->mmBuffer2) PetscCallCUDA(cudaFree(mmdata->mmBuffer2));
2639: #endif
2640:   PetscCall(MatDestroy(&mmdata->X));
2641:   PetscCall(PetscFree(*data));
2642:   PetscFunctionReturn(PETSC_SUCCESS);
2643: }

2645: #include <../src/mat/impls/dense/seq/dense.h>

2647: static PetscErrorCode MatProductNumeric_SeqAIJCUSPARSE_SeqDENSECUDA(Mat C)
2648: {
2649:   Mat_Product                  *product = C->product;
2650:   Mat                           A, B;
2651:   PetscInt                      m, n, blda, clda;
2652:   PetscBool                     flg, biscuda;
2653:   Mat_SeqAIJCUSPARSE           *cusp;
2654:   cusparseStatus_t              stat;
2655:   cusparseOperation_t           opA;
2656:   const PetscScalar            *barray;
2657:   PetscScalar                  *carray;
2658:   MatProductCtx_MatMatCusparse *mmdata;
2659:   Mat_SeqAIJCUSPARSEMultStruct *mat;
2660:   CsrMatrix                    *csrmat;

2662:   PetscFunctionBegin;
2663:   MatCheckProduct(C, 1);
2664:   PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Product data empty");
2665:   mmdata = (MatProductCtx_MatMatCusparse *)product->data;
2666:   A      = product->A;
2667:   B      = product->B;
2668:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2669:   PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Not for type %s", ((PetscObject)A)->type_name);
2670:   /* currently CopyToGpu does not copy if the matrix is bound to CPU
2671:      Instead of silently accepting the wrong answer, I prefer to raise the error */
2672:   PetscCheck(!A->boundtocpu, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Cannot bind to CPU a CUSPARSE matrix between MatProductSymbolic and MatProductNumeric phases");
2673:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
2674:   cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2675:   switch (product->type) {
2676:   case MATPRODUCT_AB:
2677:   case MATPRODUCT_PtAP:
2678:     mat = cusp->mat;
2679:     opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
2680:     m   = A->rmap->n;
2681:     n   = B->cmap->n;
2682:     break;
2683:   case MATPRODUCT_AtB:
2684:     if (!A->form_explicit_transpose) {
2685:       mat = cusp->mat;
2686:       opA = CUSPARSE_OPERATION_TRANSPOSE;
2687:     } else {
2688:       PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
2689:       mat = cusp->matTranspose;
2690:       opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
2691:     }
2692:     m = A->cmap->n;
2693:     n = B->cmap->n;
2694:     break;
2695:   case MATPRODUCT_ABt:
2696:   case MATPRODUCT_RARt:
2697:     mat = cusp->mat;
2698:     opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
2699:     m   = A->rmap->n;
2700:     n   = B->rmap->n;
2701:     break;
2702:   default:
2703:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Unsupported product type %s", MatProductTypes[product->type]);
2704:   }
2705:   PetscCheck(mat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing Mat_SeqAIJCUSPARSEMultStruct");
2706:   csrmat = (CsrMatrix *)mat->mat;
2707:   /* if the user passed a CPU matrix, copy the data to the GPU */
2708:   PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQDENSECUDA, &biscuda));
2709:   if (!biscuda) PetscCall(MatConvert(B, MATSEQDENSECUDA, MAT_INPLACE_MATRIX, &B));
2710:   PetscCall(MatDenseGetArrayReadAndMemType(B, &barray, nullptr));

2712:   PetscCall(MatDenseGetLDA(B, &blda));
2713:   if (product->type == MATPRODUCT_RARt || product->type == MATPRODUCT_PtAP) {
2714:     PetscCall(MatDenseGetArrayWriteAndMemType(mmdata->X, &carray, nullptr));
2715:     PetscCall(MatDenseGetLDA(mmdata->X, &clda));
2716:   } else {
2717:     PetscCall(MatDenseGetArrayWriteAndMemType(C, &carray, nullptr));
2718:     PetscCall(MatDenseGetLDA(C, &clda));
2719:   }

2721:   PetscCall(PetscLogGpuTimeBegin());
2722: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2723:   cusparseOperation_t opB = (product->type == MATPRODUCT_ABt || product->type == MATPRODUCT_RARt) ? CUSPARSE_OPERATION_TRANSPOSE : CUSPARSE_OPERATION_NON_TRANSPOSE;
2724:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
2725:   cusparseSpMatDescr_t &matADescr = mat->matDescr_SpMM[opA];
2726:   #else
2727:   cusparseSpMatDescr_t &matADescr = mat->matDescr;
2728:   #endif

2730:   /* (re)allocate mmBuffer if not initialized or LDAs are different */
2731:   if (!mmdata->initialized || mmdata->Blda != blda || mmdata->Clda != clda) {
2732:     size_t mmBufferSize;
2733:     if (mmdata->initialized && mmdata->Blda != blda) {
2734:       PetscCallCUSPARSE(cusparseDestroyDnMat(mmdata->matBDescr));
2735:       mmdata->matBDescr = NULL;
2736:     }
2737:     if (!mmdata->matBDescr) {
2738:       PetscCallCUSPARSE(cusparseCreateDnMat(&mmdata->matBDescr, B->rmap->n, B->cmap->n, blda, (void *)barray, cusparse_scalartype, CUSPARSE_ORDER_COL));
2739:       mmdata->Blda = blda;
2740:     }

2742:     if (mmdata->initialized && mmdata->Clda != clda) {
2743:       PetscCallCUSPARSE(cusparseDestroyDnMat(mmdata->matCDescr));
2744:       mmdata->matCDescr = NULL;
2745:     }
2746:     if (!mmdata->matCDescr) { /* matCDescr is for C or mmdata->X */
2747:       PetscCallCUSPARSE(cusparseCreateDnMat(&mmdata->matCDescr, m, n, clda, (void *)carray, cusparse_scalartype, CUSPARSE_ORDER_COL));
2748:       mmdata->Clda = clda;
2749:     }

2751:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0) // tested up to 12.6.0
2752:     if (matADescr) {
2753:       PetscCallCUSPARSE(cusparseDestroySpMat(matADescr)); // Because I find I could not reuse matADescr. It could be a cusparse bug
2754:       matADescr = NULL;
2755:     }
2756:   #endif

2758:     if (!matADescr) {
2759:       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 */
2760:                                CUSPARSE_INDEX_BASE_ZERO, cusparse_scalartype);
2761:       PetscCallCUSPARSE(stat);
2762:     }

2764:     PetscCallCUSPARSE(cusparseSpMM_bufferSize(cusp->handle, opA, opB, mat->alpha_one, matADescr, mmdata->matBDescr, mat->beta_zero, mmdata->matCDescr, cusparse_scalartype, cusp->spmmAlg, &mmBufferSize));

2766:     if ((mmdata->mmBuffer && mmdata->mmBufferSize < mmBufferSize) || !mmdata->mmBuffer) {
2767:       PetscCallCUDA(cudaFree(mmdata->mmBuffer));
2768:       PetscCallCUDA(cudaMalloc(&mmdata->mmBuffer, mmBufferSize));
2769:       mmdata->mmBufferSize = mmBufferSize;
2770:     }

2772:   #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
2773:     PetscCallCUSPARSE(cusparseSpMM_preprocess(cusp->handle, opA, opB, mat->alpha_one, matADescr, mmdata->matBDescr, mat->beta_zero, mmdata->matCDescr, cusparse_scalartype, cusp->spmmAlg, mmdata->mmBuffer));
2774:   #endif

2776:     mmdata->initialized = PETSC_TRUE;
2777:   } else {
2778:     /* to be safe, always update pointers of the mats */
2779:     PetscCallCUSPARSE(cusparseSpMatSetValues(matADescr, csrmat->values->data().get()));
2780:     PetscCallCUSPARSE(cusparseDnMatSetValues(mmdata->matBDescr, (void *)barray));
2781:     PetscCallCUSPARSE(cusparseDnMatSetValues(mmdata->matCDescr, (void *)carray));
2782:   }

2784:   /* do cusparseSpMM, which supports transpose on B */
2785:   PetscCallCUSPARSE(cusparseSpMM(cusp->handle, opA, opB, mat->alpha_one, matADescr, mmdata->matBDescr, mat->beta_zero, mmdata->matCDescr, cusparse_scalartype, cusp->spmmAlg, mmdata->mmBuffer));
2786: #else
2787:   PetscInt k;
2788:   /* cusparseXcsrmm does not support transpose on B */
2789:   if (product->type == MATPRODUCT_ABt || product->type == MATPRODUCT_RARt) {
2790:     cublasHandle_t cublasv2handle;
2791:     cublasStatus_t cerr;

2793:     PetscCall(PetscCUBLASGetHandle(&cublasv2handle));
2794:     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);
2795:     PetscCallCUBLAS(cerr);
2796:     blda = B->cmap->n;
2797:     k    = B->cmap->n;
2798:   } else {
2799:     k = B->rmap->n;
2800:   }

2802:   /* perform the MatMat operation, op(A) is m x k, op(B) is k x n */
2803:   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);
2804:   PetscCallCUSPARSE(stat);
2805: #endif
2806:   PetscCall(PetscLogGpuTimeEnd());
2807:   PetscCall(PetscLogGpuFlops(n * 2.0 * csrmat->num_entries));
2808:   PetscCall(MatDenseRestoreArrayReadAndMemType(B, &barray));
2809:   if (product->type == MATPRODUCT_RARt) {
2810:     PetscCall(MatDenseRestoreArrayWriteAndMemType(mmdata->X, &carray));
2811:     PetscCall(MatMatMultNumeric_SeqDenseCUDA_SeqDenseCUDA_Internal(B, mmdata->X, C, PETSC_FALSE, PETSC_FALSE));
2812:   } else if (product->type == MATPRODUCT_PtAP) {
2813:     PetscCall(MatDenseRestoreArrayWriteAndMemType(mmdata->X, &carray));
2814:     PetscCall(MatMatMultNumeric_SeqDenseCUDA_SeqDenseCUDA_Internal(B, mmdata->X, C, PETSC_TRUE, PETSC_FALSE));
2815:   } else {
2816:     PetscCall(MatDenseRestoreArrayWriteAndMemType(C, &carray));
2817:   }
2818:   if (mmdata->cisdense) PetscCall(MatConvert(C, MATSEQDENSE, MAT_INPLACE_MATRIX, &C));
2819:   if (!biscuda) PetscCall(MatConvert(B, MATSEQDENSE, MAT_INPLACE_MATRIX, &B));
2820:   PetscFunctionReturn(PETSC_SUCCESS);
2821: }

2823: static PetscErrorCode MatProductSymbolic_SeqAIJCUSPARSE_SeqDENSECUDA(Mat C)
2824: {
2825:   Mat_Product                  *product = C->product;
2826:   Mat                           A, B;
2827:   PetscInt                      m, n;
2828:   PetscBool                     cisdense, flg;
2829:   MatProductCtx_MatMatCusparse *mmdata;
2830:   Mat_SeqAIJCUSPARSE           *cusp;

2832:   PetscFunctionBegin;
2833:   MatCheckProduct(C, 1);
2834:   PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Product data not empty");
2835:   A = product->A;
2836:   B = product->B;
2837:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2838:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for type %s", ((PetscObject)A)->type_name);
2839:   cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2840:   PetscCheck(cusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2841:   switch (product->type) {
2842:   case MATPRODUCT_AB:
2843:     m = A->rmap->n;
2844:     n = B->cmap->n;
2845:     PetscCall(MatSetBlockSizesFromMats(C, A, B));
2846:     break;
2847:   case MATPRODUCT_AtB:
2848:     m = A->cmap->n;
2849:     n = B->cmap->n;
2850:     if (A->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->rmap, A->cmap->bs));
2851:     if (B->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->cmap, B->cmap->bs));
2852:     break;
2853:   case MATPRODUCT_ABt:
2854:     m = A->rmap->n;
2855:     n = B->rmap->n;
2856:     if (A->rmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->rmap, A->rmap->bs));
2857:     if (B->rmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->cmap, B->rmap->bs));
2858:     break;
2859:   case MATPRODUCT_PtAP:
2860:     m = B->cmap->n;
2861:     n = B->cmap->n;
2862:     if (B->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->rmap, B->cmap->bs));
2863:     if (B->cmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->cmap, B->cmap->bs));
2864:     break;
2865:   case MATPRODUCT_RARt:
2866:     m = B->rmap->n;
2867:     n = B->rmap->n;
2868:     if (B->rmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->rmap, B->rmap->bs));
2869:     if (B->rmap->bs > 0) PetscCall(PetscLayoutSetBlockSize(C->cmap, B->rmap->bs));
2870:     break;
2871:   default:
2872:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Unsupported product type %s", MatProductTypes[product->type]);
2873:   }
2874:   PetscCall(MatSetSizes(C, m, n, m, n));
2875:   /* if C is of type MATSEQDENSE (CPU), perform the operation on the GPU and then copy on the CPU */
2876:   PetscCall(PetscObjectTypeCompare((PetscObject)C, MATSEQDENSE, &cisdense));
2877:   PetscCall(MatSetType(C, MATSEQDENSECUDA));

2879:   /* product data */
2880:   PetscCall(PetscNew(&mmdata));
2881:   mmdata->cisdense = cisdense;
2882: #if PETSC_PKG_CUDA_VERSION_LT(11, 0, 0)
2883:   /* cusparseXcsrmm does not support transpose on B, so we allocate buffer to store B^T */
2884:   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)));
2885: #endif
2886:   /* for these products we need intermediate storage */
2887:   if (product->type == MATPRODUCT_RARt || product->type == MATPRODUCT_PtAP) {
2888:     PetscCall(MatCreate(PetscObjectComm((PetscObject)C), &mmdata->X));
2889:     PetscCall(MatSetType(mmdata->X, MATSEQDENSECUDA));
2890:     if (product->type == MATPRODUCT_RARt) { /* do not preallocate, since the first call to MatDenseCUDAGetArray will preallocate on the GPU for us */
2891:       PetscCall(MatSetSizes(mmdata->X, A->rmap->n, B->rmap->n, A->rmap->n, B->rmap->n));
2892:     } else {
2893:       PetscCall(MatSetSizes(mmdata->X, A->rmap->n, B->cmap->n, A->rmap->n, B->cmap->n));
2894:     }
2895:   }
2896:   C->product->data    = mmdata;
2897:   C->product->destroy = MatProductCtxDestroy_MatMatCusparse;

2899:   C->ops->productnumeric = MatProductNumeric_SeqAIJCUSPARSE_SeqDENSECUDA;
2900:   PetscFunctionReturn(PETSC_SUCCESS);
2901: }

2903: static PetscErrorCode MatProductNumeric_SeqAIJCUSPARSE_SeqAIJCUSPARSE(Mat C)
2904: {
2905:   Mat_Product                  *product = C->product;
2906:   Mat                           A, B;
2907:   Mat_SeqAIJCUSPARSE           *Acusp, *Bcusp, *Ccusp;
2908:   Mat_SeqAIJ                   *c = (Mat_SeqAIJ *)C->data;
2909:   Mat_SeqAIJCUSPARSEMultStruct *Amat, *Bmat, *Cmat;
2910:   CsrMatrix                    *Acsr, *Bcsr, *Ccsr;
2911:   PetscBool                     flg;
2912:   cusparseStatus_t              stat;
2913:   MatProductType                ptype;
2914:   MatProductCtx_MatMatCusparse *mmdata;
2915: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2916:   cusparseSpMatDescr_t BmatSpDescr;
2917: #endif
2918:   cusparseOperation_t opA = CUSPARSE_OPERATION_NON_TRANSPOSE, opB = CUSPARSE_OPERATION_NON_TRANSPOSE; /* cuSPARSE spgemm doesn't support transpose yet */

2920:   PetscFunctionBegin;
2921:   MatCheckProduct(C, 1);
2922:   PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Product data empty");
2923:   PetscCall(PetscObjectTypeCompare((PetscObject)C, MATSEQAIJCUSPARSE, &flg));
2924:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for C of type %s", ((PetscObject)C)->type_name);
2925:   mmdata = (MatProductCtx_MatMatCusparse *)C->product->data;
2926:   A      = product->A;
2927:   B      = product->B;
2928:   if (mmdata->reusesym) { /* this happens when api_user is true, meaning that the matrix values have been already computed in the MatProductSymbolic phase */
2929:     mmdata->reusesym = PETSC_FALSE;
2930:     Ccusp            = (Mat_SeqAIJCUSPARSE *)C->spptr;
2931:     PetscCheck(Ccusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2932:     Cmat = Ccusp->mat;
2933:     PetscCheck(Cmat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing C mult struct for product type %s", MatProductTypes[C->product->type]);
2934:     Ccsr = (CsrMatrix *)Cmat->mat;
2935:     PetscCheck(Ccsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing C CSR struct");
2936:     goto finalize;
2937:   }
2938:   if (!c->nz) goto finalize;
2939:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2940:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for type %s", ((PetscObject)A)->type_name);
2941:   PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJCUSPARSE, &flg));
2942:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for B of type %s", ((PetscObject)B)->type_name);
2943:   PetscCheck(!A->boundtocpu, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONG, "Cannot bind to CPU a CUSPARSE matrix between MatProductSymbolic and MatProductNumeric phases");
2944:   PetscCheck(!B->boundtocpu, PetscObjectComm((PetscObject)C), PETSC_ERR_ARG_WRONG, "Cannot bind to CPU a CUSPARSE matrix between MatProductSymbolic and MatProductNumeric phases");
2945:   Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2946:   Bcusp = (Mat_SeqAIJCUSPARSE *)B->spptr;
2947:   Ccusp = (Mat_SeqAIJCUSPARSE *)C->spptr;
2948:   PetscCheck(Acusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2949:   PetscCheck(Bcusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2950:   PetscCheck(Ccusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
2951:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
2952:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));

2954:   ptype = product->type;
2955:   if (A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
2956:     ptype = MATPRODUCT_AB;
2957:     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");
2958:   }
2959:   if (B->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_ABt) {
2960:     ptype = MATPRODUCT_AB;
2961:     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");
2962:   }
2963:   switch (ptype) {
2964:   case MATPRODUCT_AB:
2965:     Amat = Acusp->mat;
2966:     Bmat = Bcusp->mat;
2967:     break;
2968:   case MATPRODUCT_AtB:
2969:     Amat = Acusp->matTranspose;
2970:     Bmat = Bcusp->mat;
2971:     break;
2972:   case MATPRODUCT_ABt:
2973:     Amat = Acusp->mat;
2974:     Bmat = Bcusp->matTranspose;
2975:     break;
2976:   default:
2977:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Unsupported product type %s", MatProductTypes[product->type]);
2978:   }
2979:   Cmat = Ccusp->mat;
2980:   PetscCheck(Amat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing A mult struct for product type %s", MatProductTypes[ptype]);
2981:   PetscCheck(Bmat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing B mult struct for product type %s", MatProductTypes[ptype]);
2982:   PetscCheck(Cmat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing C mult struct for product type %s", MatProductTypes[ptype]);
2983:   Acsr = (CsrMatrix *)Amat->mat;
2984:   Bcsr = mmdata->Bcsr ? mmdata->Bcsr : (CsrMatrix *)Bmat->mat; /* B may be in compressed row storage */
2985:   Ccsr = (CsrMatrix *)Cmat->mat;
2986:   PetscCheck(Acsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing A CSR struct");
2987:   PetscCheck(Bcsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing B CSR struct");
2988:   PetscCheck(Ccsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing C CSR struct");
2989:   PetscCall(PetscLogGpuTimeBegin());
2990: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
2991:   BmatSpDescr = mmdata->Bcsr ? mmdata->matSpBDescr : Bmat->matDescr; /* B may be in compressed row storage */
2992:   PetscCallCUSPARSE(cusparseSetPointerMode(Ccusp->handle, CUSPARSE_POINTER_MODE_DEVICE));
2993:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2994:   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);
2995:   PetscCallCUSPARSE(stat);
2996:   #else
2997:   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);
2998:   PetscCallCUSPARSE(stat);
2999:   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);
3000:   PetscCallCUSPARSE(stat);
3001:   #endif
3002: #else
3003:   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,
3004:                              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());
3005:   PetscCallCUSPARSE(stat);
3006: #endif
3007:   PetscCall(PetscLogGpuFlops(mmdata->flops));
3008:   PetscCallCUDA(WaitForCUDA());
3009:   PetscCall(PetscLogGpuTimeEnd());
3010:   C->offloadmask = PETSC_OFFLOAD_GPU;
3011: finalize:
3012:   /* shorter version of MatAssemblyEnd_SeqAIJ */
3013:   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));
3014:   PetscCall(PetscInfo(C, "Number of mallocs during MatSetValues() is 0\n"));
3015:   PetscCall(PetscInfo(C, "Maximum nonzeros in any row is %" PetscInt_FMT "\n", c->rmax));
3016:   c->reallocs = 0;
3017:   C->info.mallocs += 0;
3018:   C->info.nz_unneeded = 0;
3019:   C->assembled = C->was_assembled = PETSC_TRUE;
3020:   C->num_ass++;
3021:   PetscFunctionReturn(PETSC_SUCCESS);
3022: }

3024: static PetscErrorCode MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE(Mat C)
3025: {
3026:   Mat_Product                  *product = C->product;
3027:   Mat                           A, B;
3028:   Mat_SeqAIJCUSPARSE           *Acusp, *Bcusp, *Ccusp;
3029:   Mat_SeqAIJ                   *a, *b, *c;
3030:   Mat_SeqAIJCUSPARSEMultStruct *Amat, *Bmat, *Cmat;
3031:   CsrMatrix                    *Acsr, *Bcsr, *Ccsr;
3032:   PetscInt                      i, j, m, n, k;
3033:   PetscBool                     flg;
3034:   cusparseStatus_t              stat;
3035:   MatProductType                ptype;
3036:   MatProductCtx_MatMatCusparse *mmdata;
3037:   PetscLogDouble                flops;
3038:   PetscBool                     biscompressed, ciscompressed;
3039: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3040:   int64_t              C_num_rows1, C_num_cols1, C_nnz1;
3041:   cusparseSpMatDescr_t BmatSpDescr;
3042: #else
3043:   int cnz;
3044: #endif
3045:   cusparseOperation_t opA = CUSPARSE_OPERATION_NON_TRANSPOSE, opB = CUSPARSE_OPERATION_NON_TRANSPOSE; /* cuSPARSE spgemm doesn't support transpose yet */

3047:   PetscFunctionBegin;
3048:   MatCheckProduct(C, 1);
3049:   PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Product data not empty");
3050:   A = product->A;
3051:   B = product->B;
3052:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
3053:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for type %s", ((PetscObject)A)->type_name);
3054:   PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJCUSPARSE, &flg));
3055:   PetscCheck(flg, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Not for B of type %s", ((PetscObject)B)->type_name);
3056:   a = (Mat_SeqAIJ *)A->data;
3057:   b = (Mat_SeqAIJ *)B->data;
3058:   /* product data */
3059:   PetscCall(PetscNew(&mmdata));
3060:   C->product->data    = mmdata;
3061:   C->product->destroy = MatProductCtxDestroy_MatMatCusparse;

3063:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
3064:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));
3065:   Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr; /* Access spptr after MatSeqAIJCUSPARSECopyToGPU, not before */
3066:   Bcusp = (Mat_SeqAIJCUSPARSE *)B->spptr;
3067:   PetscCheck(Acusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");
3068:   PetscCheck(Bcusp->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Only for MAT_CUSPARSE_CSR format");

3070:   ptype = product->type;
3071:   if (A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
3072:     ptype                                          = MATPRODUCT_AB;
3073:     product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
3074:   }
3075:   if (B->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_ABt) {
3076:     ptype                                          = MATPRODUCT_AB;
3077:     product->symbolic_used_the_fact_B_is_symmetric = PETSC_TRUE;
3078:   }
3079:   biscompressed = PETSC_FALSE;
3080:   ciscompressed = PETSC_FALSE;
3081:   switch (ptype) {
3082:   case MATPRODUCT_AB:
3083:     m    = A->rmap->n;
3084:     n    = B->cmap->n;
3085:     k    = A->cmap->n;
3086:     Amat = Acusp->mat;
3087:     Bmat = Bcusp->mat;
3088:     if (a->compressedrow.use) ciscompressed = PETSC_TRUE;
3089:     if (b->compressedrow.use) biscompressed = PETSC_TRUE;
3090:     break;
3091:   case MATPRODUCT_AtB:
3092:     m = A->cmap->n;
3093:     n = B->cmap->n;
3094:     k = A->rmap->n;
3095:     PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
3096:     Amat = Acusp->matTranspose;
3097:     Bmat = Bcusp->mat;
3098:     if (b->compressedrow.use) biscompressed = PETSC_TRUE;
3099:     break;
3100:   case MATPRODUCT_ABt:
3101:     m = A->rmap->n;
3102:     n = B->rmap->n;
3103:     k = A->cmap->n;
3104:     PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(B));
3105:     Amat = Acusp->mat;
3106:     Bmat = Bcusp->matTranspose;
3107:     if (a->compressedrow.use) ciscompressed = PETSC_TRUE;
3108:     break;
3109:   default:
3110:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Unsupported product type %s", MatProductTypes[product->type]);
3111:   }

3113:   /* create cusparse matrix */
3114:   PetscCall(MatSetSizes(C, m, n, m, n));
3115:   PetscCall(MatSetType(C, MATSEQAIJCUSPARSE));
3116:   c     = (Mat_SeqAIJ *)C->data;
3117:   Ccusp = (Mat_SeqAIJCUSPARSE *)C->spptr;
3118:   Cmat  = new Mat_SeqAIJCUSPARSEMultStruct;
3119:   Ccsr  = new CsrMatrix;

3121:   c->compressedrow.use = ciscompressed;
3122:   if (c->compressedrow.use) { /* if a is in compressed row, than c will be in compressed row format */
3123:     c->compressedrow.nrows = a->compressedrow.nrows;
3124:     PetscCall(PetscMalloc2(c->compressedrow.nrows + 1, &c->compressedrow.i, c->compressedrow.nrows, &c->compressedrow.rindex));
3125:     PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, c->compressedrow.nrows));
3126:     Ccusp->workVector  = new THRUSTARRAY(c->compressedrow.nrows);
3127:     Cmat->cprowIndices = new THRUSTINTARRAY(c->compressedrow.nrows);
3128:     Cmat->cprowIndices->assign(c->compressedrow.rindex, c->compressedrow.rindex + c->compressedrow.nrows);
3129:   } else {
3130:     c->compressedrow.nrows  = 0;
3131:     c->compressedrow.i      = NULL;
3132:     c->compressedrow.rindex = NULL;
3133:     Ccusp->workVector       = NULL;
3134:     Cmat->cprowIndices      = NULL;
3135:   }
3136:   Ccusp->nrows      = ciscompressed ? c->compressedrow.nrows : m;
3137:   Ccusp->mat        = Cmat;
3138:   Ccusp->mat->mat   = Ccsr;
3139:   Ccsr->num_rows    = Ccusp->nrows;
3140:   Ccsr->num_cols    = n;
3141:   Ccsr->row_offsets = new THRUSTINTARRAY32(Ccusp->nrows + 1);
3142:   PetscCallCUSPARSE(cusparseCreateMatDescr(&Cmat->descr));
3143:   PetscCallCUSPARSE(cusparseSetMatIndexBase(Cmat->descr, CUSPARSE_INDEX_BASE_ZERO));
3144:   PetscCallCUSPARSE(cusparseSetMatType(Cmat->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
3145:   PetscCallCUDA(cudaMalloc((void **)&Cmat->alpha_one, sizeof(PetscScalar)));
3146:   PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_zero, sizeof(PetscScalar)));
3147:   PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_one, sizeof(PetscScalar)));
3148:   PetscCallCUDA(cudaMemcpy(Cmat->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
3149:   PetscCallCUDA(cudaMemcpy(Cmat->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
3150:   PetscCallCUDA(cudaMemcpy(Cmat->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
3151:   if (!Ccsr->num_rows || !Ccsr->num_cols || !a->nz || !b->nz) { /* cusparse raise errors in different calls when matrices have zero rows/columns! */
3152:     PetscCallThrust(thrust::fill(thrust::device, Ccsr->row_offsets->begin(), Ccsr->row_offsets->end(), 0));
3153:     c->nz                = 0;
3154:     Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
3155:     Ccsr->values         = new THRUSTARRAY(c->nz);
3156:     goto finalizesym;
3157:   }

3159:   PetscCheck(Amat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing A mult struct for product type %s", MatProductTypes[ptype]);
3160:   PetscCheck(Bmat, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing B mult struct for product type %s", MatProductTypes[ptype]);
3161:   Acsr = (CsrMatrix *)Amat->mat;
3162:   if (!biscompressed) {
3163:     Bcsr = (CsrMatrix *)Bmat->mat;
3164: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3165:     BmatSpDescr = Bmat->matDescr;
3166: #endif
3167:   } else { /* we need to use row offsets for the full matrix */
3168:     CsrMatrix *cBcsr     = (CsrMatrix *)Bmat->mat;
3169:     Bcsr                 = new CsrMatrix;
3170:     Bcsr->num_rows       = B->rmap->n;
3171:     Bcsr->num_cols       = cBcsr->num_cols;
3172:     Bcsr->num_entries    = cBcsr->num_entries;
3173:     Bcsr->column_indices = cBcsr->column_indices;
3174:     Bcsr->values         = cBcsr->values;
3175:     if (!Bcusp->rowoffsets_gpu) {
3176:       Bcusp->rowoffsets_gpu = new THRUSTINTARRAY32(B->rmap->n + 1);
3177:       Bcusp->rowoffsets_gpu->assign(b->i, b->i + B->rmap->n + 1);
3178:       PetscCall(PetscLogCpuToGpu((B->rmap->n + 1) * sizeof(PetscInt)));
3179:     }
3180:     Bcsr->row_offsets = Bcusp->rowoffsets_gpu;
3181:     mmdata->Bcsr      = Bcsr;
3182: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3183:     if (Bcsr->num_rows && Bcsr->num_cols) {
3184:       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);
3185:       PetscCallCUSPARSE(stat);
3186:     }
3187:     BmatSpDescr = mmdata->matSpBDescr;
3188: #endif
3189:   }
3190:   PetscCheck(Acsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing A CSR struct");
3191:   PetscCheck(Bcsr, PetscObjectComm((PetscObject)C), PETSC_ERR_GPU, "Missing B CSR struct");
3192:   /* precompute flops count */
3193:   if (ptype == MATPRODUCT_AB) {
3194:     for (i = 0, flops = 0; i < A->rmap->n; i++) {
3195:       const PetscInt st = a->i[i];
3196:       const PetscInt en = a->i[i + 1];
3197:       for (j = st; j < en; j++) {
3198:         const PetscInt brow = a->j[j];
3199:         flops += 2. * (b->i[brow + 1] - b->i[brow]);
3200:       }
3201:     }
3202:   } else if (ptype == MATPRODUCT_AtB) {
3203:     for (i = 0, flops = 0; i < A->rmap->n; i++) {
3204:       const PetscInt anzi = a->i[i + 1] - a->i[i];
3205:       const PetscInt bnzi = b->i[i + 1] - b->i[i];
3206:       flops += (2. * anzi) * bnzi;
3207:     }
3208:   } else { /* TODO */
3209:     flops = 0.;
3210:   }

3212:   mmdata->flops = flops;
3213:   PetscCall(PetscLogGpuTimeBegin());

3215: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3216:   PetscCallCUSPARSE(cusparseSetPointerMode(Ccusp->handle, CUSPARSE_POINTER_MODE_DEVICE));
3217:   // cuda-12.2 requires non-null csrRowOffsets
3218:   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);
3219:   PetscCallCUSPARSE(stat);
3220:   PetscCallCUSPARSE(cusparseSpGEMM_createDescr(&mmdata->spgemmDesc));
3221:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
3222:   {
3223:     /* cusparseSpGEMMreuse has more reasonable APIs than cusparseSpGEMM, so we prefer to use it.
3224:      We follow the sample code at https://github.com/NVIDIA/CUDALibrarySamples/blob/master/cuSPARSE/spgemm_reuse
3225:   */
3226:     void *dBuffer1 = NULL;
3227:     void *dBuffer2 = NULL;
3228:     void *dBuffer3 = NULL;
3229:     /* dBuffer4, dBuffer5 are needed by cusparseSpGEMMreuse_compute, and therefore are stored in mmdata */
3230:     size_t bufferSize1 = 0;
3231:     size_t bufferSize2 = 0;
3232:     size_t bufferSize3 = 0;
3233:     size_t bufferSize4 = 0;
3234:     size_t bufferSize5 = 0;

3236:     /* ask bufferSize1 bytes for external memory */
3237:     stat = cusparseSpGEMMreuse_workEstimation(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize1, NULL);
3238:     PetscCallCUSPARSE(stat);
3239:     PetscCallCUDA(cudaMalloc((void **)&dBuffer1, bufferSize1));
3240:     /* inspect the matrices A and B to understand the memory requirement for the next step */
3241:     stat = cusparseSpGEMMreuse_workEstimation(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize1, dBuffer1);
3242:     PetscCallCUSPARSE(stat);

3244:     stat = cusparseSpGEMMreuse_nnz(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize2, NULL, &bufferSize3, NULL, &bufferSize4, NULL);
3245:     PetscCallCUSPARSE(stat);
3246:     PetscCallCUDA(cudaMalloc((void **)&dBuffer2, bufferSize2));
3247:     PetscCallCUDA(cudaMalloc((void **)&dBuffer3, bufferSize3));
3248:     PetscCallCUDA(cudaMalloc((void **)&mmdata->dBuffer4, bufferSize4));
3249:     stat = cusparseSpGEMMreuse_nnz(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize2, dBuffer2, &bufferSize3, dBuffer3, &bufferSize4, mmdata->dBuffer4);
3250:     PetscCallCUSPARSE(stat);
3251:     PetscCallCUDA(cudaFree(dBuffer1));
3252:     PetscCallCUDA(cudaFree(dBuffer2));

3254:     /* get matrix C non-zero entries C_nnz1 */
3255:     PetscCallCUSPARSE(cusparseSpMatGetSize(Cmat->matDescr, &C_num_rows1, &C_num_cols1, &C_nnz1));
3256:     c->nz = (PetscInt)C_nnz1;
3257:     /* allocate matrix C */
3258:     Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
3259:     PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3260:     Ccsr->values = new THRUSTARRAY(c->nz);
3261:     PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3262:     /* update matC with the new pointers */
3263:     stat = cusparseCsrSetPointers(Cmat->matDescr, Ccsr->row_offsets->data().get(), Ccsr->column_indices->data().get(), Ccsr->values->data().get());
3264:     PetscCallCUSPARSE(stat);

3266:     stat = cusparseSpGEMMreuse_copy(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize5, NULL);
3267:     PetscCallCUSPARSE(stat);
3268:     PetscCallCUDA(cudaMalloc((void **)&mmdata->dBuffer5, bufferSize5));
3269:     stat = cusparseSpGEMMreuse_copy(Ccusp->handle, opA, opB, Amat->matDescr, BmatSpDescr, Cmat->matDescr, CUSPARSE_SPGEMM_DEFAULT, mmdata->spgemmDesc, &bufferSize5, mmdata->dBuffer5);
3270:     PetscCallCUSPARSE(stat);
3271:     PetscCallCUDA(cudaFree(dBuffer3));
3272:     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);
3273:     PetscCallCUSPARSE(stat);
3274:     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));
3275:   }
3276:   #else
3277:   size_t bufSize2;
3278:   /* ask bufferSize bytes for external memory */
3279:   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);
3280:   PetscCallCUSPARSE(stat);
3281:   PetscCallCUDA(cudaMalloc((void **)&mmdata->mmBuffer2, bufSize2));
3282:   /* inspect the matrices A and B to understand the memory requirement for the next step */
3283:   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);
3284:   PetscCallCUSPARSE(stat);
3285:   /* ask bufferSize again bytes for external memory */
3286:   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);
3287:   PetscCallCUSPARSE(stat);
3288:   /* The CUSPARSE documentation is not clear, nor the API
3289:      We need both buffers to perform the operations properly!
3290:      mmdata->mmBuffer2 does not appear anywhere in the compute/copy API
3291:      it only appears for the workEstimation stuff, but it seems it is needed in compute, so probably the address
3292:      is stored in the descriptor! What a messy API... */
3293:   PetscCallCUDA(cudaMalloc((void **)&mmdata->mmBuffer, mmdata->mmBufferSize));
3294:   /* compute the intermediate product of A * B */
3295:   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);
3296:   PetscCallCUSPARSE(stat);
3297:   /* get matrix C non-zero entries C_nnz1 */
3298:   PetscCallCUSPARSE(cusparseSpMatGetSize(Cmat->matDescr, &C_num_rows1, &C_num_cols1, &C_nnz1));
3299:   c->nz = (PetscInt)C_nnz1;
3300:   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,
3301:                       mmdata->mmBufferSize / 1024));
3302:   Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
3303:   PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3304:   Ccsr->values = new THRUSTARRAY(c->nz);
3305:   PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3306:   stat = cusparseCsrSetPointers(Cmat->matDescr, Ccsr->row_offsets->data().get(), Ccsr->column_indices->data().get(), Ccsr->values->data().get());
3307:   PetscCallCUSPARSE(stat);
3308:   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);
3309:   PetscCallCUSPARSE(stat);
3310:   #endif // PETSC_PKG_CUDA_VERSION_GE(11,4,0)
3311: #else
3312:   PetscCallCUSPARSE(cusparseSetPointerMode(Ccusp->handle, CUSPARSE_POINTER_MODE_HOST));
3313:   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,
3314:                              Bcsr->row_offsets->data().get(), Bcsr->column_indices->data().get(), Cmat->descr, Ccsr->row_offsets->data().get(), &cnz);
3315:   PetscCallCUSPARSE(stat);
3316:   c->nz                = cnz;
3317:   Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
3318:   PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */
3319:   Ccsr->values = new THRUSTARRAY(c->nz);
3320:   PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */

3322:   PetscCallCUSPARSE(cusparseSetPointerMode(Ccusp->handle, CUSPARSE_POINTER_MODE_DEVICE));
3323:   /* with the old gemm interface (removed from 11.0 on) we cannot compute the symbolic factorization only.
3324:      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
3325:      D is NULL, despite the fact that CUSPARSE documentation claims it is supported! */
3326:   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,
3327:                              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());
3328:   PetscCallCUSPARSE(stat);
3329: #endif
3330:   PetscCall(PetscLogGpuFlops(mmdata->flops));
3331:   PetscCall(PetscLogGpuTimeEnd());
3332: finalizesym:
3333:   c->free_a = PETSC_TRUE;
3334:   PetscCall(PetscShmgetAllocateArray(c->nz, sizeof(PetscInt), (void **)&c->j));
3335:   PetscCall(PetscShmgetAllocateArray(m + 1, sizeof(PetscInt), (void **)&c->i));
3336:   c->free_ij = PETSC_TRUE;
3337:   if (PetscDefined(USE_64BIT_INDICES)) { /* 32 to 64-bit conversion on the GPU and then copy to host (lazy) */
3338:     PetscInt      *d_i = c->i;
3339:     THRUSTINTARRAY ii(Ccsr->row_offsets->size());
3340:     THRUSTINTARRAY jj(Ccsr->column_indices->size());
3341:     ii = *Ccsr->row_offsets;
3342:     jj = *Ccsr->column_indices;
3343:     if (ciscompressed) d_i = c->compressedrow.i;
3344:     PetscCallCUDA(cudaMemcpy(d_i, ii.data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
3345:     PetscCallCUDA(cudaMemcpy(c->j, jj.data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
3346:   } else {
3347:     PetscInt *d_i = c->i;
3348:     if (ciscompressed) d_i = c->compressedrow.i;
3349:     PetscCallCUDA(cudaMemcpy(d_i, Ccsr->row_offsets->data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
3350:     PetscCallCUDA(cudaMemcpy(c->j, Ccsr->column_indices->data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
3351:   }
3352:   if (ciscompressed) { /* need to expand host row offsets */
3353:     PetscInt r = 0;
3354:     c->i[0]    = 0;
3355:     for (k = 0; k < c->compressedrow.nrows; k++) {
3356:       const PetscInt next = c->compressedrow.rindex[k];
3357:       const PetscInt old  = c->compressedrow.i[k];
3358:       for (; r < next; r++) c->i[r + 1] = old;
3359:     }
3360:     for (; r < m; r++) c->i[r + 1] = c->compressedrow.i[c->compressedrow.nrows];
3361:   }
3362:   PetscCall(PetscLogGpuToCpu((Ccsr->column_indices->size() + Ccsr->row_offsets->size()) * sizeof(PetscInt)));
3363:   PetscCall(PetscMalloc1(m, &c->ilen));
3364:   PetscCall(PetscMalloc1(m, &c->imax));
3365:   c->maxnz         = c->nz;
3366:   c->nonzerorowcnt = 0;
3367:   c->rmax          = 0;
3368:   for (k = 0; k < m; k++) {
3369:     const PetscInt nn = c->i[k + 1] - c->i[k];
3370:     c->ilen[k] = c->imax[k] = nn;
3371:     c->nonzerorowcnt += (PetscInt)!!nn;
3372:     c->rmax = PetscMax(c->rmax, nn);
3373:   }
3374:   PetscCall(PetscMalloc1(c->nz, &c->a));
3375:   Ccsr->num_entries = c->nz;

3377:   C->nonzerostate++;
3378:   PetscCall(PetscLayoutSetUp(C->rmap));
3379:   PetscCall(PetscLayoutSetUp(C->cmap));
3380:   Ccusp->nonzerostate = C->nonzerostate;
3381:   C->offloadmask      = PETSC_OFFLOAD_UNALLOCATED;
3382:   C->preallocated     = PETSC_TRUE;
3383:   C->assembled        = PETSC_FALSE;
3384:   C->was_assembled    = PETSC_FALSE;
3385:   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 */
3386:     mmdata->reusesym = PETSC_TRUE;
3387:     C->offloadmask   = PETSC_OFFLOAD_GPU;
3388:   }
3389:   C->ops->productnumeric = MatProductNumeric_SeqAIJCUSPARSE_SeqAIJCUSPARSE;
3390:   PetscFunctionReturn(PETSC_SUCCESS);
3391: }

3393: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ_SeqDense(Mat);

3395: /* handles sparse or dense B */
3396: static PetscErrorCode MatProductSetFromOptions_SeqAIJCUSPARSE(Mat mat)
3397: {
3398:   Mat_Product *product = mat->product;
3399:   PetscBool    isdense = PETSC_FALSE, Biscusp = PETSC_FALSE, Ciscusp = PETSC_TRUE;

3401:   PetscFunctionBegin;
3402:   MatCheckProduct(mat, 1);
3403:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)product->B, MATSEQDENSE, &isdense));
3404:   if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, MATSEQAIJCUSPARSE, &Biscusp));
3405:   if (product->type == MATPRODUCT_ABC) {
3406:     Ciscusp = PETSC_FALSE;
3407:     if (!product->C->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->C, MATSEQAIJCUSPARSE, &Ciscusp));
3408:   }
3409:   if (Biscusp && Ciscusp) { /* we can always select the CPU backend */
3410:     PetscBool usecpu = PETSC_FALSE;
3411:     switch (product->type) {
3412:     case MATPRODUCT_AB:
3413:       if (product->api_user) {
3414:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
3415:         PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
3416:         PetscOptionsEnd();
3417:       } else {
3418:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
3419:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
3420:         PetscOptionsEnd();
3421:       }
3422:       break;
3423:     case MATPRODUCT_AtB:
3424:       if (product->api_user) {
3425:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
3426:         PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
3427:         PetscOptionsEnd();
3428:       } else {
3429:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
3430:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
3431:         PetscOptionsEnd();
3432:       }
3433:       break;
3434:     case MATPRODUCT_PtAP:
3435:       if (product->api_user) {
3436:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
3437:         PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
3438:         PetscOptionsEnd();
3439:       } else {
3440:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
3441:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
3442:         PetscOptionsEnd();
3443:       }
3444:       break;
3445:     case MATPRODUCT_RARt:
3446:       if (product->api_user) {
3447:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatRARt", "Mat");
3448:         PetscCall(PetscOptionsBool("-matrart_backend_cpu", "Use CPU code", "MatRARt", usecpu, &usecpu, NULL));
3449:         PetscOptionsEnd();
3450:       } else {
3451:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_RARt", "Mat");
3452:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatRARt", usecpu, &usecpu, NULL));
3453:         PetscOptionsEnd();
3454:       }
3455:       break;
3456:     case MATPRODUCT_ABC:
3457:       if (product->api_user) {
3458:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMatMult", "Mat");
3459:         PetscCall(PetscOptionsBool("-matmatmatmult_backend_cpu", "Use CPU code", "MatMatMatMult", usecpu, &usecpu, NULL));
3460:         PetscOptionsEnd();
3461:       } else {
3462:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_ABC", "Mat");
3463:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMatMult", usecpu, &usecpu, NULL));
3464:         PetscOptionsEnd();
3465:       }
3466:       break;
3467:     default:
3468:       break;
3469:     }
3470:     if (usecpu) Biscusp = Ciscusp = PETSC_FALSE;
3471:   }
3472:   /* dispatch */
3473:   if (isdense) {
3474:     switch (product->type) {
3475:     case MATPRODUCT_AB:
3476:     case MATPRODUCT_AtB:
3477:     case MATPRODUCT_ABt:
3478:     case MATPRODUCT_PtAP:
3479:     case MATPRODUCT_RARt:
3480:       if (product->A->boundtocpu) {
3481:         PetscCall(MatProductSetFromOptions_SeqAIJ_SeqDense(mat));
3482:       } else {
3483:         mat->ops->productsymbolic = MatProductSymbolic_SeqAIJCUSPARSE_SeqDENSECUDA;
3484:       }
3485:       break;
3486:     case MATPRODUCT_ABC:
3487:       mat->ops->productsymbolic = MatProductSymbolic_ABC_Basic;
3488:       break;
3489:     default:
3490:       break;
3491:     }
3492:   } else if (Biscusp && Ciscusp) {
3493:     switch (product->type) {
3494:     case MATPRODUCT_AB:
3495:     case MATPRODUCT_AtB:
3496:     case MATPRODUCT_ABt:
3497:       mat->ops->productsymbolic = MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE;
3498:       break;
3499:     case MATPRODUCT_PtAP:
3500:     case MATPRODUCT_RARt:
3501:     case MATPRODUCT_ABC:
3502:       mat->ops->productsymbolic = MatProductSymbolic_ABC_Basic;
3503:       break;
3504:     default:
3505:       break;
3506:     }
3507:   } else { /* fallback for AIJ */
3508:     PetscCall(MatProductSetFromOptions_SeqAIJ(mat));
3509:   }
3510:   PetscFunctionReturn(PETSC_SUCCESS);
3511: }

3513: static PetscErrorCode MatMult_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3514: {
3515:   PetscFunctionBegin;
3516:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_FALSE, PETSC_FALSE));
3517:   PetscFunctionReturn(PETSC_SUCCESS);
3518: }

3520: static PetscErrorCode MatMultAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3521: {
3522:   PetscFunctionBegin;
3523:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_FALSE, PETSC_FALSE));
3524:   PetscFunctionReturn(PETSC_SUCCESS);
3525: }

3527: static PetscErrorCode MatMultHermitianTranspose_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3528: {
3529:   PetscFunctionBegin;
3530:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_TRUE, PETSC_TRUE));
3531:   PetscFunctionReturn(PETSC_SUCCESS);
3532: }

3534: static PetscErrorCode MatMultHermitianTransposeAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3535: {
3536:   PetscFunctionBegin;
3537:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_TRUE, PETSC_TRUE));
3538:   PetscFunctionReturn(PETSC_SUCCESS);
3539: }

3541: static PetscErrorCode MatMultTranspose_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3542: {
3543:   PetscFunctionBegin;
3544:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_TRUE, PETSC_FALSE));
3545:   PetscFunctionReturn(PETSC_SUCCESS);
3546: }

3548: __global__ static void ScatterAdd(PetscInt n, PetscInt *idx, const PetscScalar *x, PetscScalar *y)
3549: {
3550:   int i = blockIdx.x * blockDim.x + threadIdx.x;
3551:   if (i < n) y[idx[i]] += x[i];
3552: }

3554: /* z = op(A) x + y. If trans & !herm, op = ^T; if trans & herm, op = ^H; if !trans, op = no-op */
3555: static PetscErrorCode MatMultAddKernel_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz, PetscBool trans, PetscBool herm)
3556: {
3557:   Mat_SeqAIJ                   *a              = (Mat_SeqAIJ *)A->data;
3558:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
3559:   Mat_SeqAIJCUSPARSEMultStruct *matstruct;
3560:   PetscScalar                  *xarray, *zarray, *dptr, *beta, *xptr;
3561:   cusparseOperation_t           opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
3562:   PetscBool                     compressed;
3563: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3564:   PetscInt nx, ny;
3565: #endif

3567:   PetscFunctionBegin;
3568:   PetscCheck(!herm || trans, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Hermitian and not transpose not supported");
3569:   if (!a->nz) {
3570:     if (yy) PetscCall(VecSeq_CUDA::Copy(yy, zz));
3571:     else PetscCall(VecSeq_CUDA::Set(zz, 0));
3572:     PetscFunctionReturn(PETSC_SUCCESS);
3573:   }
3574:   /* The line below is necessary due to the operations that modify the matrix on the CPU (axpy, scale, etc) */
3575:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
3576:   if (!trans) {
3577:     matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
3578:     PetscCheck(matstruct, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "SeqAIJCUSPARSE does not have a 'mat' (need to fix)");
3579:   } else {
3580:     if (herm || !A->form_explicit_transpose) {
3581:       opA       = herm ? CUSPARSE_OPERATION_CONJUGATE_TRANSPOSE : CUSPARSE_OPERATION_TRANSPOSE;
3582:       matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
3583:     } else {
3584:       if (!cusparsestruct->matTranspose) PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
3585:       matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->matTranspose;
3586:     }
3587:   }
3588:   /* Does the matrix use compressed rows (i.e., drop zero rows)? */
3589:   compressed = matstruct->cprowIndices ? PETSC_TRUE : PETSC_FALSE;

3591:   try {
3592:     PetscCall(VecCUDAGetArrayRead(xx, (const PetscScalar **)&xarray));
3593:     if (yy == zz) PetscCall(VecCUDAGetArray(zz, &zarray)); /* read & write zz, so need to get up-to-date zarray on GPU */
3594:     else PetscCall(VecCUDAGetArrayWrite(zz, &zarray));     /* write zz, so no need to init zarray on GPU */

3596:     PetscCall(PetscLogGpuTimeBegin());
3597:     if (opA == CUSPARSE_OPERATION_NON_TRANSPOSE) {
3598:       /* z = A x + beta y.
3599:          If A is compressed (with less rows), then Ax is shorter than the full z, so we need a work vector to store Ax.
3600:          When A is non-compressed, and z = y, we can set beta=1 to compute y = Ax + y in one call.
3601:       */
3602:       xptr = xarray;
3603:       dptr = compressed ? cusparsestruct->workVector->data().get() : zarray;
3604:       beta = (yy == zz && !compressed) ? matstruct->beta_one : matstruct->beta_zero;
3605: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3606:       /* Get length of x, y for y=Ax. ny might be shorter than the work vector's allocated length, since the work vector is
3607:           allocated to accommodate different uses. So we get the length info directly from mat.
3608:        */
3609:       if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3610:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3611:         nx             = mat->num_cols; // since y = Ax
3612:         ny             = mat->num_rows;
3613:       }
3614: #endif
3615:     } else {
3616:       /* z = A^T x + beta y
3617:          If A is compressed, then we need a work vector as the shorter version of x to compute A^T x.
3618:          Note A^Tx is of full length, so we set beta to 1.0 if y exists.
3619:        */
3620:       xptr = compressed ? cusparsestruct->workVector->data().get() : xarray;
3621:       dptr = zarray;
3622:       beta = yy ? matstruct->beta_one : matstruct->beta_zero;
3623:       if (compressed) { /* Scatter x to work vector */
3624:         thrust::device_ptr<PetscScalar> xarr = thrust::device_pointer_cast(xarray);

3626:         thrust::for_each(
3627: #if PetscDefined(HAVE_THRUST_ASYNC)
3628:           thrust::cuda::par.on(PetscDefaultCudaStream),
3629: #endif
3630:           thrust::make_zip_iterator(thrust::make_tuple(cusparsestruct->workVector->begin(), thrust::make_permutation_iterator(xarr, matstruct->cprowIndices->begin()))),
3631:           thrust::make_zip_iterator(thrust::make_tuple(cusparsestruct->workVector->begin(), thrust::make_permutation_iterator(xarr, matstruct->cprowIndices->begin()))) + matstruct->cprowIndices->size(), VecCUDAEqualsReverse());
3632:       }
3633: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3634:       if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3635:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3636:         nx             = mat->num_rows; // since y = A^T x
3637:         ny             = mat->num_cols;
3638:       }
3639: #endif
3640:     }

3642:     /* csr_spmv does y = alpha op(A) x + beta y */
3643:     if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3644: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3645:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
3646:       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.
3647:   #else
3648:       cusparseSpMatDescr_t &matDescr = matstruct->matDescr;
3649:   #endif

3651:       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");
3652:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
3653:       if (!matDescr) {
3654:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3655:         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));
3656:       }
3657:   #endif

3659:       if (!matstruct->cuSpMV[opA].initialized) { /* built on demand */
3660:         PetscCallCUSPARSE(cusparseCreateDnVec(&matstruct->cuSpMV[opA].vecXDescr, nx, xptr, cusparse_scalartype));
3661:         PetscCallCUSPARSE(cusparseCreateDnVec(&matstruct->cuSpMV[opA].vecYDescr, ny, dptr, cusparse_scalartype));
3662:         PetscCallCUSPARSE(
3663:           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));
3664:         PetscCallCUDA(cudaMalloc(&matstruct->cuSpMV[opA].spmvBuffer, matstruct->cuSpMV[opA].spmvBufferSize));
3665:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0) // cusparseSpMV_preprocess is added in 12.4
3666:         PetscCallCUSPARSE(
3667:           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));
3668:   #endif
3669:         matstruct->cuSpMV[opA].initialized = PETSC_TRUE;
3670:       } else {
3671:         /* x, y's value pointers might change between calls, but their shape is kept, so we just update pointers */
3672:         PetscCallCUSPARSE(cusparseDnVecSetValues(matstruct->cuSpMV[opA].vecXDescr, xptr));
3673:         PetscCallCUSPARSE(cusparseDnVecSetValues(matstruct->cuSpMV[opA].vecYDescr, dptr));
3674:       }

3676:       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));
3677: #else
3678:       CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3679:       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));
3680: #endif
3681:     } else {
3682:       if (cusparsestruct->nrows) {
3683: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3684:         SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
3685: #else
3686:         cusparseHybMat_t hybMat = (cusparseHybMat_t)matstruct->mat;
3687:         PetscCallCUSPARSE(cusparse_hyb_spmv(cusparsestruct->handle, opA, matstruct->alpha_one, matstruct->descr, hybMat, xptr, beta, dptr));
3688: #endif
3689:       }
3690:     }
3691:     PetscCall(PetscLogGpuTimeEnd());

3693:     if (opA == CUSPARSE_OPERATION_NON_TRANSPOSE) {
3694:       if (yy) {                                      /* MatMultAdd: zz = A*xx + yy */
3695:         if (compressed) {                            /* A is compressed. We first copy yy to zz, then ScatterAdd the work vector to zz */
3696:           PetscCall(VecSeq_CUDA::Copy(yy, zz));      /* zz = yy */
3697:         } else if (zz != yy) {                       /* A is not compressed. zz already contains A*xx, and we just need to add yy */
3698:           PetscCall(VecSeq_CUDA::AXPY(zz, 1.0, yy)); /* zz += yy */
3699:         }
3700:       } else if (compressed) { /* MatMult: zz = A*xx. A is compressed, so we zero zz first, then ScatterAdd the work vector to zz */
3701:         PetscCall(VecSeq_CUDA::Set(zz, 0));
3702:       }

3704:       /* ScatterAdd the result from work vector into the full vector when A is compressed */
3705:       if (compressed) {
3706:         PetscCall(PetscLogGpuTimeBegin());
3707:         PetscInt n = (PetscInt)matstruct->cprowIndices->size();
3708:         ScatterAdd<<<(int)((n + 255) / 256), 256, 0, PetscDefaultCudaStream>>>(n, matstruct->cprowIndices->data().get(), cusparsestruct->workVector->data().get(), zarray);
3709:         PetscCall(PetscLogGpuTimeEnd());
3710:       }
3711:     } else {
3712:       if (yy && yy != zz) PetscCall(VecSeq_CUDA::AXPY(zz, 1.0, yy)); /* zz += yy */
3713:     }
3714:     PetscCall(VecCUDARestoreArrayRead(xx, (const PetscScalar **)&xarray));
3715:     if (yy == zz) PetscCall(VecCUDARestoreArray(zz, &zarray));
3716:     else PetscCall(VecCUDARestoreArrayWrite(zz, &zarray));
3717:   } catch (char *ex) {
3718:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
3719:   }
3720:   if (yy) {
3721:     PetscCall(PetscLogGpuFlops(2.0 * a->nz));
3722:   } else {
3723:     PetscCall(PetscLogGpuFlops(2.0 * a->nz - a->nonzerorowcnt));
3724:   }
3725:   PetscFunctionReturn(PETSC_SUCCESS);
3726: }

3728: static PetscErrorCode MatMultTransposeAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3729: {
3730:   PetscFunctionBegin;
3731:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_TRUE, PETSC_FALSE));
3732:   PetscFunctionReturn(PETSC_SUCCESS);
3733: }

3735: PETSC_INTERN PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A, Vec xx);

3737: __global__ static void GetDiagonal_CSR(const int *row, const int *col, const PetscScalar *val, const PetscInt len, PetscScalar *diag)
3738: {
3739:   const size_t x = blockIdx.x * blockDim.x + threadIdx.x;

3741:   if (x < len) {
3742:     const PetscInt rowx = row[x], num_non0_row = row[x + 1] - rowx;
3743:     PetscScalar    d = 0.0;

3745:     for (PetscInt i = 0; i < num_non0_row; i++) {
3746:       if (col[i + rowx] == x) {
3747:         d = val[i + rowx];
3748:         break;
3749:       }
3750:     }
3751:     diag[x] = d;
3752:   }
3753: }

3755: static PetscErrorCode MatGetDiagonal_SeqAIJCUSPARSE(Mat A, Vec diag)
3756: {
3757:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
3758:   Mat_SeqAIJCUSPARSEMultStruct *matstruct      = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
3759:   PetscScalar                  *darray;

3761:   PetscFunctionBegin;
3762:   if (A->offloadmask == PETSC_OFFLOAD_BOTH || A->offloadmask == PETSC_OFFLOAD_GPU) {
3763:     PetscInt   n   = A->rmap->n;
3764:     CsrMatrix *mat = (CsrMatrix *)matstruct->mat;

3766:     PetscCheck(cusparsestruct->format == MAT_CUSPARSE_CSR, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only CSR format supported");
3767:     if (n > 0) {
3768:       PetscCall(VecCUDAGetArrayWrite(diag, &darray));
3769:       GetDiagonal_CSR<<<(int)((n + 255) / 256), 256, 0, PetscDefaultCudaStream>>>(mat->row_offsets->data().get(), mat->column_indices->data().get(), mat->values->data().get(), n, darray);
3770:       PetscCallCUDA(cudaPeekAtLastError());
3771:       PetscCall(VecCUDARestoreArrayWrite(diag, &darray));
3772:     }
3773:   } else PetscCall(MatGetDiagonal_SeqAIJ(A, diag));
3774:   PetscFunctionReturn(PETSC_SUCCESS);
3775: }

3777: static PetscErrorCode MatAssemblyEnd_SeqAIJCUSPARSE(Mat A, MatAssemblyType mode)
3778: {
3779:   PetscFunctionBegin;
3780:   PetscCall(MatAssemblyEnd_SeqAIJ(A, mode));
3781:   PetscFunctionReturn(PETSC_SUCCESS);
3782: }

3784: /*@
3785:   MatCreateSeqAIJCUSPARSE - Creates a sparse matrix in `MATAIJCUSPARSE` (compressed row) format for use on NVIDIA GPUs

3787:   Collective

3789:   Input Parameters:
3790: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3791: . m    - number of rows
3792: . n    - number of columns
3793: . nz   - number of nonzeros per row (same for all rows), ignored if `nnz` is provide
3794: - nnz  - array containing the number of nonzeros in the various rows (possibly different for each row) or `NULL`

3796:   Output Parameter:
3797: . A - the matrix

3799:   Level: intermediate

3801:   Notes:
3802:   This matrix will ultimately pushed down to NVIDIA GPUs and use the CuSPARSE library for
3803:   calculations. For good matrix assembly performance the user should preallocate the matrix
3804:   storage by setting the parameter `nz` (or the array `nnz`).

3806:   It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
3807:   MatXXXXSetPreallocation() paradgm instead of this routine directly.
3808:   [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]

3810:   The AIJ format, also called
3811:   compressed row storage, is fully compatible with standard Fortran
3812:   storage.  That is, the stored row and column indices can begin at
3813:   either one (as in Fortran) or zero.

3815:   Specify the preallocated storage with either nz or nnz (not both).
3816:   Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3817:   allocation.

3819:   When working with matrices for GPUs, it is often better to use the `MatSetPreallocationCOO()` and `MatSetValuesCOO()` paradigm rather than using this routine and `MatSetValues()`

3821: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJCUSPARSE`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MATAIJCUSPARSE`,
3822:           `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
3823: @*/
3824: PetscErrorCode MatCreateSeqAIJCUSPARSE(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3825: {
3826:   PetscFunctionBegin;
3827:   PetscCall(MatCreate(comm, A));
3828:   PetscCall(MatSetSizes(*A, m, n, m, n));
3829:   PetscCall(MatSetType(*A, MATSEQAIJCUSPARSE));
3830:   PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, (PetscInt *)nnz));
3831:   PetscFunctionReturn(PETSC_SUCCESS);
3832: }

3834: static PetscErrorCode MatDestroy_SeqAIJCUSPARSE(Mat A)
3835: {
3836:   PetscFunctionBegin;
3837:   if (A->factortype == MAT_FACTOR_NONE) {
3838:     PetscCall(MatSeqAIJCUSPARSE_Destroy(A));
3839:   } else {
3840:     PetscCall(MatSeqAIJCUSPARSETriFactors_Destroy((Mat_SeqAIJCUSPARSETriFactors **)&A->spptr));
3841:   }
3842:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", NULL));
3843:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatCUSPARSESetFormat_C", NULL));
3844:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatCUSPARSESetUseCPUSolve_C", NULL));
3845:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", NULL));
3846:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", NULL));
3847:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", NULL));
3848:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
3849:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
3850:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
3851:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijcusparse_hypre_C", NULL));
3852:   PetscCall(MatDestroy_SeqAIJ(A));
3853:   PetscFunctionReturn(PETSC_SUCCESS);
3854: }

3856: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
3857: static PetscErrorCode       MatBindToCPU_SeqAIJCUSPARSE(Mat, PetscBool);
3858: static PetscErrorCode       MatDuplicate_SeqAIJCUSPARSE(Mat A, MatDuplicateOption cpvalues, Mat *B)
3859: {
3860:   PetscFunctionBegin;
3861:   PetscCall(MatDuplicate_SeqAIJ(A, cpvalues, B));
3862:   PetscCall(MatConvert_SeqAIJ_SeqAIJCUSPARSE(*B, MATSEQAIJCUSPARSE, MAT_INPLACE_MATRIX, B));
3863:   PetscFunctionReturn(PETSC_SUCCESS);
3864: }

3866: static PetscErrorCode MatAXPY_SeqAIJCUSPARSE(Mat Y, PetscScalar a, Mat X, MatStructure str)
3867: {
3868:   Mat_SeqAIJ         *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data;
3869:   Mat_SeqAIJCUSPARSE *cy;
3870:   Mat_SeqAIJCUSPARSE *cx;
3871:   PetscScalar        *ay;
3872:   const PetscScalar  *ax;
3873:   CsrMatrix          *csry, *csrx;

3875:   PetscFunctionBegin;
3876:   cy = (Mat_SeqAIJCUSPARSE *)Y->spptr;
3877:   cx = (Mat_SeqAIJCUSPARSE *)X->spptr;
3878:   if (X->ops->axpy != Y->ops->axpy) {
3879:     PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(Y, PETSC_FALSE));
3880:     PetscCall(MatAXPY_SeqAIJ(Y, a, X, str));
3881:     PetscFunctionReturn(PETSC_SUCCESS);
3882:   }
3883:   /* if we are here, it means both matrices are bound to GPU */
3884:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(Y));
3885:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(X));
3886:   PetscCheck(cy->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)Y), PETSC_ERR_GPU, "only MAT_CUSPARSE_CSR supported");
3887:   PetscCheck(cx->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)X), PETSC_ERR_GPU, "only MAT_CUSPARSE_CSR supported");
3888:   csry = (CsrMatrix *)cy->mat->mat;
3889:   csrx = (CsrMatrix *)cx->mat->mat;
3890:   /* see if we can turn this into a cublas axpy */
3891:   if (str != SAME_NONZERO_PATTERN && x->nz == y->nz && !x->compressedrow.use && !y->compressedrow.use) {
3892:     bool eq = thrust::equal(thrust::device, csry->row_offsets->begin(), csry->row_offsets->end(), csrx->row_offsets->begin());
3893:     if (eq) eq = thrust::equal(thrust::device, csry->column_indices->begin(), csry->column_indices->end(), csrx->column_indices->begin());
3894:     if (eq) str = SAME_NONZERO_PATTERN;
3895:   }
3896:   /* spgeam is buggy with one column */
3897:   if (Y->cmap->n == 1 && str != SAME_NONZERO_PATTERN) str = DIFFERENT_NONZERO_PATTERN;

3899:   if (str == SUBSET_NONZERO_PATTERN) {
3900:     PetscScalar b = 1.0;
3901: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3902:     size_t bufferSize;
3903:     void  *buffer;
3904: #endif

3906:     PetscCall(MatSeqAIJCUSPARSEGetArrayRead(X, &ax));
3907:     PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3908:     PetscCallCUSPARSE(cusparseSetPointerMode(cy->handle, CUSPARSE_POINTER_MODE_HOST));
3909: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3910:     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(),
3911:                                                      csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get(), &bufferSize));
3912:     PetscCallCUDA(cudaMalloc(&buffer, bufferSize));
3913:     PetscCall(PetscLogGpuTimeBegin());
3914:     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(),
3915:                                           csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get(), buffer));
3916:     PetscCall(PetscLogGpuFlops(x->nz + y->nz));
3917:     PetscCall(PetscLogGpuTimeEnd());
3918:     PetscCallCUDA(cudaFree(buffer));
3919: #else
3920:     PetscCall(PetscLogGpuTimeBegin());
3921:     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(),
3922:                                           csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get()));
3923:     PetscCall(PetscLogGpuFlops(x->nz + y->nz));
3924:     PetscCall(PetscLogGpuTimeEnd());
3925: #endif
3926:     PetscCallCUSPARSE(cusparseSetPointerMode(cy->handle, CUSPARSE_POINTER_MODE_DEVICE));
3927:     PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(X, &ax));
3928:     PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3929:   } else if (str == SAME_NONZERO_PATTERN) {
3930:     cublasHandle_t cublasv2handle;
3931:     PetscBLASInt   one = 1, bnz = 1;

3933:     PetscCall(MatSeqAIJCUSPARSEGetArrayRead(X, &ax));
3934:     PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3935:     PetscCall(PetscCUBLASGetHandle(&cublasv2handle));
3936:     PetscCall(PetscBLASIntCast(x->nz, &bnz));
3937:     PetscCall(PetscLogGpuTimeBegin());
3938:     PetscCallCUBLAS(cublasXaxpy(cublasv2handle, bnz, &a, ax, one, ay, one));
3939:     PetscCall(PetscLogGpuFlops(2.0 * bnz));
3940:     PetscCall(PetscLogGpuTimeEnd());
3941:     PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(X, &ax));
3942:     PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3943:   } else {
3944:     PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(Y, PETSC_FALSE));
3945:     PetscCall(MatAXPY_SeqAIJ(Y, a, X, str));
3946:   }
3947:   PetscFunctionReturn(PETSC_SUCCESS);
3948: }

3950: static PetscErrorCode MatScale_SeqAIJCUSPARSE(Mat Y, PetscScalar a)
3951: {
3952:   Mat_SeqAIJ    *y = (Mat_SeqAIJ *)Y->data;
3953:   PetscScalar   *ay;
3954:   cublasHandle_t cublasv2handle;
3955:   PetscBLASInt   one = 1, bnz = 1;

3957:   PetscFunctionBegin;
3958:   PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3959:   PetscCall(PetscCUBLASGetHandle(&cublasv2handle));
3960:   PetscCall(PetscBLASIntCast(y->nz, &bnz));
3961:   PetscCall(PetscLogGpuTimeBegin());
3962:   PetscCallCUBLAS(cublasXscal(cublasv2handle, bnz, &a, ay, one));
3963:   PetscCall(PetscLogGpuFlops(bnz));
3964:   PetscCall(PetscLogGpuTimeEnd());
3965:   PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3966:   PetscFunctionReturn(PETSC_SUCCESS);
3967: }

3969: static PetscErrorCode MatZeroEntries_SeqAIJCUSPARSE(Mat A)
3970: {
3971:   PetscBool   gpu = PETSC_FALSE;
3972:   Mat_SeqAIJ *a   = (Mat_SeqAIJ *)A->data;

3974:   PetscFunctionBegin;
3975:   if (A->factortype == MAT_FACTOR_NONE) {
3976:     Mat_SeqAIJCUSPARSE *spptr = (Mat_SeqAIJCUSPARSE *)A->spptr;
3977:     if (spptr->mat) {
3978:       CsrMatrix *matrix = (CsrMatrix *)spptr->mat->mat;
3979:       if (matrix->values) {
3980:         gpu = PETSC_TRUE;
3981:         thrust::fill(thrust::device, matrix->values->begin(), matrix->values->end(), 0.);
3982:       }
3983:     }
3984:     if (spptr->matTranspose) {
3985:       CsrMatrix *matrix = (CsrMatrix *)spptr->matTranspose->mat;
3986:       if (matrix->values) thrust::fill(thrust::device, matrix->values->begin(), matrix->values->end(), 0.);
3987:     }
3988:   }
3989:   if (gpu) A->offloadmask = PETSC_OFFLOAD_GPU;
3990:   else {
3991:     PetscCall(PetscArrayzero(a->a, a->i[A->rmap->n]));
3992:     A->offloadmask = PETSC_OFFLOAD_CPU;
3993:   }
3994:   PetscFunctionReturn(PETSC_SUCCESS);
3995: }

3997: static PetscErrorCode MatGetCurrentMemType_SeqAIJCUSPARSE(PETSC_UNUSED Mat A, PetscMemType *m)
3998: {
3999:   PetscFunctionBegin;
4000:   *m = PETSC_MEMTYPE_CUDA;
4001:   PetscFunctionReturn(PETSC_SUCCESS);
4002: }

4004: static PetscErrorCode MatBindToCPU_SeqAIJCUSPARSE(Mat A, PetscBool flg)
4005: {
4006:   Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;

4008:   PetscFunctionBegin;
4009:   if (A->factortype != MAT_FACTOR_NONE) {
4010:     A->boundtocpu = flg;
4011:     PetscFunctionReturn(PETSC_SUCCESS);
4012:   }
4013:   if (flg) {
4014:     PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));

4016:     A->ops->scale                     = MatScale_SeqAIJ;
4017:     A->ops->getdiagonal               = MatGetDiagonal_SeqAIJ;
4018:     A->ops->axpy                      = MatAXPY_SeqAIJ;
4019:     A->ops->zeroentries               = MatZeroEntries_SeqAIJ;
4020:     A->ops->mult                      = MatMult_SeqAIJ;
4021:     A->ops->multadd                   = MatMultAdd_SeqAIJ;
4022:     A->ops->multtranspose             = MatMultTranspose_SeqAIJ;
4023:     A->ops->multtransposeadd          = MatMultTransposeAdd_SeqAIJ;
4024:     A->ops->multhermitiantranspose    = NULL;
4025:     A->ops->multhermitiantransposeadd = NULL;
4026:     A->ops->productsetfromoptions     = MatProductSetFromOptions_SeqAIJ;
4027:     A->ops->getcurrentmemtype         = NULL;
4028:     PetscCall(PetscMemzero(a->ops, sizeof(Mat_SeqAIJOps)));
4029:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", NULL));
4030:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", NULL));
4031:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", NULL));
4032:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
4033:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
4034:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", NULL));
4035:   } else {
4036:     A->ops->scale                     = MatScale_SeqAIJCUSPARSE;
4037:     A->ops->getdiagonal               = MatGetDiagonal_SeqAIJCUSPARSE;
4038:     A->ops->axpy                      = MatAXPY_SeqAIJCUSPARSE;
4039:     A->ops->zeroentries               = MatZeroEntries_SeqAIJCUSPARSE;
4040:     A->ops->mult                      = MatMult_SeqAIJCUSPARSE;
4041:     A->ops->multadd                   = MatMultAdd_SeqAIJCUSPARSE;
4042:     A->ops->multtranspose             = MatMultTranspose_SeqAIJCUSPARSE;
4043:     A->ops->multtransposeadd          = MatMultTransposeAdd_SeqAIJCUSPARSE;
4044:     A->ops->multhermitiantranspose    = MatMultHermitianTranspose_SeqAIJCUSPARSE;
4045:     A->ops->multhermitiantransposeadd = MatMultHermitianTransposeAdd_SeqAIJCUSPARSE;
4046:     A->ops->productsetfromoptions     = MatProductSetFromOptions_SeqAIJCUSPARSE;
4047:     A->ops->getcurrentmemtype         = MatGetCurrentMemType_SeqAIJCUSPARSE;
4048:     a->ops->getarray                  = MatSeqAIJGetArray_SeqAIJCUSPARSE;
4049:     a->ops->restorearray              = MatSeqAIJRestoreArray_SeqAIJCUSPARSE;
4050:     a->ops->getarrayread              = MatSeqAIJGetArrayRead_SeqAIJCUSPARSE;
4051:     a->ops->restorearrayread          = MatSeqAIJRestoreArrayRead_SeqAIJCUSPARSE;
4052:     a->ops->getarraywrite             = MatSeqAIJGetArrayWrite_SeqAIJCUSPARSE;
4053:     a->ops->restorearraywrite         = MatSeqAIJRestoreArrayWrite_SeqAIJCUSPARSE;
4054:     a->ops->getcsrandmemtype          = MatSeqAIJGetCSRAndMemType_SeqAIJCUSPARSE;

4056:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", MatSeqAIJCopySubArray_SeqAIJCUSPARSE));
4057:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4058:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4059:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJCUSPARSE));
4060:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJCUSPARSE));
4061:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4062:   }
4063:   A->boundtocpu = flg;
4064:   if (flg && a->inode.size_csr) {
4065:     a->inode.use = PETSC_TRUE;
4066:   } else {
4067:     a->inode.use = PETSC_FALSE;
4068:   }
4069:   PetscFunctionReturn(PETSC_SUCCESS);
4070: }

4072: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat A, MatType, MatReuse reuse, Mat *newmat)
4073: {
4074:   Mat B;

4076:   PetscFunctionBegin;
4077:   PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUDA)); /* first use of CUSPARSE may be via MatConvert */
4078:   if (reuse == MAT_INITIAL_MATRIX) {
4079:     PetscCall(MatDuplicate(A, MAT_COPY_VALUES, newmat));
4080:   } else if (reuse == MAT_REUSE_MATRIX) {
4081:     PetscCall(MatCopy(A, *newmat, SAME_NONZERO_PATTERN));
4082:   }
4083:   B = *newmat;

4085:   PetscCall(PetscFree(B->defaultvectype));
4086:   PetscCall(PetscStrallocpy(VECCUDA, &B->defaultvectype));

4088:   if (reuse != MAT_REUSE_MATRIX && !B->spptr) {
4089:     if (B->factortype == MAT_FACTOR_NONE) {
4090:       Mat_SeqAIJCUSPARSE *spptr;
4091:       PetscCall(PetscNew(&spptr));
4092:       PetscCallCUSPARSE(cusparseCreate(&spptr->handle));
4093:       PetscCallCUSPARSE(cusparseSetStream(spptr->handle, PetscDefaultCudaStream));
4094:       spptr->format = MAT_CUSPARSE_CSR;
4095: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4096:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
4097:       spptr->spmvAlg = CUSPARSE_SPMV_CSR_ALG1; /* default, since we only support csr */
4098:   #else
4099:       spptr->spmvAlg = CUSPARSE_CSRMV_ALG1; /* default, since we only support csr */
4100:   #endif
4101:       spptr->spmmAlg    = CUSPARSE_SPMM_CSR_ALG1; /* default, only support column-major dense matrix B */
4102:       spptr->csr2cscAlg = CUSPARSE_CSR2CSC_ALG1;
4103: #endif
4104:       B->spptr = spptr;
4105:     } else {
4106:       Mat_SeqAIJCUSPARSETriFactors *spptr;

4108:       PetscCall(PetscNew(&spptr));
4109:       PetscCallCUSPARSE(cusparseCreate(&spptr->handle));
4110:       PetscCallCUSPARSE(cusparseSetStream(spptr->handle, PetscDefaultCudaStream));
4111:       B->spptr = spptr;
4112:     }
4113:     B->offloadmask = PETSC_OFFLOAD_UNALLOCATED;
4114:   }
4115:   B->ops->assemblyend       = MatAssemblyEnd_SeqAIJCUSPARSE;
4116:   B->ops->destroy           = MatDestroy_SeqAIJCUSPARSE;
4117:   B->ops->setoption         = MatSetOption_SeqAIJCUSPARSE;
4118:   B->ops->setfromoptions    = MatSetFromOptions_SeqAIJCUSPARSE;
4119:   B->ops->bindtocpu         = MatBindToCPU_SeqAIJCUSPARSE;
4120:   B->ops->duplicate         = MatDuplicate_SeqAIJCUSPARSE;
4121:   B->ops->getcurrentmemtype = MatGetCurrentMemType_SeqAIJCUSPARSE;

4123:   PetscCall(MatBindToCPU_SeqAIJCUSPARSE(B, PETSC_FALSE));
4124:   PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJCUSPARSE));
4125:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatCUSPARSESetFormat_C", MatCUSPARSESetFormat_SeqAIJCUSPARSE));
4126: #if defined(PETSC_HAVE_HYPRE)
4127:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaijcusparse_hypre_C", MatConvert_AIJ_HYPRE));
4128: #endif
4129:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatCUSPARSESetUseCPUSolve_C", MatCUSPARSESetUseCPUSolve_SeqAIJCUSPARSE));
4130:   PetscFunctionReturn(PETSC_SUCCESS);
4131: }

4133: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJCUSPARSE(Mat B)
4134: {
4135:   PetscFunctionBegin;
4136:   PetscCall(MatCreate_SeqAIJ(B));
4137:   PetscCall(MatConvert_SeqAIJ_SeqAIJCUSPARSE(B, MATSEQAIJCUSPARSE, MAT_INPLACE_MATRIX, &B));
4138:   PetscFunctionReturn(PETSC_SUCCESS);
4139: }

4141: /*MC
4142:    MATSEQAIJCUSPARSE - MATAIJCUSPARSE = "(seq)aijcusparse" - A matrix type to be used for sparse matrices on NVIDIA GPUs.

4144:    Options Database Keys:
4145: +  -mat_type aijcusparse                 - Sets the matrix type to "seqaijcusparse" during a call to `MatSetFromOptions()`
4146: .  -mat_cusparse_storage_format csr      - Sets the storage format of matrices (for `MatMult()` and factors in `MatSolve()`).
4147:                                            Other options include ell (ellpack) or hyb (hybrid).
4148: .  -mat_cusparse_mult_storage_format csr - Sets the storage format of matrices (for `MatMult()`). Other options include ell (ellpack) or hyb (hybrid).
4149: -  -mat_cusparse_use_cpu_solve           - Performs the `MatSolve()` on the CPU

4151:   Level: beginner

4153:   Notes:
4154:   These matrices can be in either CSR, ELL, or HYB format.

4156:   All matrix calculations are performed on NVIDIA GPUs using the cuSPARSE library.

4158:   Uses 32-bit integers internally. If PETSc is configured `--with-64-bit-indices`, the integer row and column indices are stored on the GPU with `int`. It is unclear what happens
4159:   if some integer values passed in do not fit in `int`.

4161: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJCUSPARSE()`, `MatCUSPARSESetUseCPUSolve()`, `MATAIJCUSPARSE`, `MatCreateAIJCUSPARSE()`, `MatCUSPARSESetFormat()`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
4162: M*/

4164: PETSC_INTERN PetscErrorCode MatSolverTypeRegister_CUSPARSE(void)
4165: {
4166:   PetscFunctionBegin;
4167:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_LU, MatGetFactor_seqaijcusparse_cusparse));
4168:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_CHOLESKY, MatGetFactor_seqaijcusparse_cusparse));
4169:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_ILU, MatGetFactor_seqaijcusparse_cusparse));
4170:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_ICC, MatGetFactor_seqaijcusparse_cusparse));
4171:   PetscFunctionReturn(PETSC_SUCCESS);
4172: }

4174: static PetscErrorCode MatSeqAIJCUSPARSE_Destroy(Mat mat)
4175: {
4176:   Mat_SeqAIJCUSPARSE *cusp = static_cast<Mat_SeqAIJCUSPARSE *>(mat->spptr);

4178:   PetscFunctionBegin;
4179:   if (cusp) {
4180:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->mat, cusp->format));
4181:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->matTranspose, cusp->format));
4182:     delete cusp->workVector;
4183:     delete cusp->rowoffsets_gpu;
4184:     delete cusp->csr2csc_i;
4185:     delete cusp->coords;
4186:     if (cusp->handle) PetscCallCUSPARSE(cusparseDestroy(cusp->handle));
4187:     PetscCall(PetscFree(mat->spptr));
4188:   }
4189:   PetscFunctionReturn(PETSC_SUCCESS);
4190: }

4192: static PetscErrorCode CsrMatrix_Destroy(CsrMatrix **mat)
4193: {
4194:   PetscFunctionBegin;
4195:   if (*mat) {
4196:     delete (*mat)->values;
4197:     delete (*mat)->column_indices;
4198:     delete (*mat)->row_offsets;
4199:     delete *mat;
4200:     *mat = 0;
4201:   }
4202:   PetscFunctionReturn(PETSC_SUCCESS);
4203: }

4205: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
4206: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSETriFactorStruct **trifactor)
4207: {
4208:   PetscFunctionBegin;
4209:   if (*trifactor) {
4210:     if ((*trifactor)->descr) PetscCallCUSPARSE(cusparseDestroyMatDescr((*trifactor)->descr));
4211:     if ((*trifactor)->solveInfo) PetscCallCUSPARSE(cusparseDestroyCsrsvInfo((*trifactor)->solveInfo));
4212:     PetscCall(CsrMatrix_Destroy(&(*trifactor)->csrMat));
4213:     if ((*trifactor)->solveBuffer) PetscCallCUDA(cudaFree((*trifactor)->solveBuffer));
4214:     if ((*trifactor)->AA_h) PetscCallCUDA(cudaFreeHost((*trifactor)->AA_h));
4215:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4216:     if ((*trifactor)->csr2cscBuffer) PetscCallCUDA(cudaFree((*trifactor)->csr2cscBuffer));
4217:   #endif
4218:     PetscCall(PetscFree(*trifactor));
4219:   }
4220:   PetscFunctionReturn(PETSC_SUCCESS);
4221: }
4222: #endif

4224: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSEMultStruct **matstruct, MatCUSPARSEStorageFormat format)
4225: {
4226:   CsrMatrix *mat;

4228:   PetscFunctionBegin;
4229:   if (*matstruct) {
4230:     if ((*matstruct)->mat) {
4231:       if (format == MAT_CUSPARSE_ELL || format == MAT_CUSPARSE_HYB) {
4232: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4233:         SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
4234: #else
4235:         cusparseHybMat_t hybMat = (cusparseHybMat_t)(*matstruct)->mat;
4236:         PetscCallCUSPARSE(cusparseDestroyHybMat(hybMat));
4237: #endif
4238:       } else {
4239:         mat = (CsrMatrix *)(*matstruct)->mat;
4240:         PetscCall(CsrMatrix_Destroy(&mat));
4241:       }
4242:     }
4243:     if ((*matstruct)->descr) PetscCallCUSPARSE(cusparseDestroyMatDescr((*matstruct)->descr));
4244:     delete (*matstruct)->cprowIndices;
4245:     if ((*matstruct)->alpha_one) PetscCallCUDA(cudaFree((*matstruct)->alpha_one));
4246:     if ((*matstruct)->beta_zero) PetscCallCUDA(cudaFree((*matstruct)->beta_zero));
4247:     if ((*matstruct)->beta_one) PetscCallCUDA(cudaFree((*matstruct)->beta_one));

4249: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4250:     Mat_SeqAIJCUSPARSEMultStruct *mdata = *matstruct;
4251:     if (mdata->matDescr) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr));

4253:     for (int i = 0; i < 3; i++) {
4254:       if (mdata->cuSpMV[i].initialized) {
4255:         PetscCallCUDA(cudaFree(mdata->cuSpMV[i].spmvBuffer));
4256:         PetscCallCUSPARSE(cusparseDestroyDnVec(mdata->cuSpMV[i].vecXDescr));
4257:         PetscCallCUSPARSE(cusparseDestroyDnVec(mdata->cuSpMV[i].vecYDescr));
4258:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
4259:         if (mdata->matDescr_SpMV[i]) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr_SpMV[i]));
4260:         if (mdata->matDescr_SpMM[i]) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr_SpMM[i]));
4261:   #endif
4262:       }
4263:     }
4264: #endif
4265:     delete *matstruct;
4266:     *matstruct = NULL;
4267:   }
4268:   PetscFunctionReturn(PETSC_SUCCESS);
4269: }

4271: PetscErrorCode MatSeqAIJCUSPARSETriFactors_Reset(Mat_SeqAIJCUSPARSETriFactors_p *trifactors)
4272: {
4273:   Mat_SeqAIJCUSPARSETriFactors *fs = *trifactors;

4275:   PetscFunctionBegin;
4276:   if (fs) {
4277: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
4278:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->loTriFactorPtr));
4279:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->upTriFactorPtr));
4280:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->loTriFactorPtrTranspose));
4281:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->upTriFactorPtrTranspose));
4282:     delete fs->workVector;
4283:     fs->workVector = NULL;
4284: #endif
4285:     delete fs->rpermIndices;
4286:     delete fs->cpermIndices;
4287:     fs->rpermIndices  = NULL;
4288:     fs->cpermIndices  = NULL;
4289:     fs->init_dev_prop = PETSC_FALSE;
4290: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
4291:     PetscCallCUDA(cudaFree(fs->csrRowPtr));
4292:     PetscCallCUDA(cudaFree(fs->csrColIdx));
4293:     PetscCallCUDA(cudaFree(fs->csrRowPtr32));
4294:     PetscCallCUDA(cudaFree(fs->csrColIdx32));
4295:     PetscCallCUDA(cudaFree(fs->csrVal));
4296:     PetscCallCUDA(cudaFree(fs->diag));
4297:     PetscCallCUDA(cudaFree(fs->X));
4298:     PetscCallCUDA(cudaFree(fs->Y));
4299:     // PetscCallCUDA(cudaFree(fs->factBuffer_M)); /* No needed since factBuffer_M shares with one of spsvBuffer_L/U */
4300:     PetscCallCUDA(cudaFree(fs->spsvBuffer_L));
4301:     PetscCallCUDA(cudaFree(fs->spsvBuffer_U));
4302:     PetscCallCUDA(cudaFree(fs->spsvBuffer_Lt));
4303:     PetscCallCUDA(cudaFree(fs->spsvBuffer_Ut));
4304:     PetscCallCUSPARSE(cusparseDestroyMatDescr(fs->matDescr_M));
4305:     PetscCallCUSPARSE(cusparseDestroySpMat(fs->spMatDescr_L));
4306:     PetscCallCUSPARSE(cusparseDestroySpMat(fs->spMatDescr_U));
4307:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_L));
4308:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_Lt));
4309:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_U));
4310:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_Ut));
4311:     PetscCallCUSPARSE(cusparseDestroyDnVec(fs->dnVecDescr_X));
4312:     PetscCallCUSPARSE(cusparseDestroyDnVec(fs->dnVecDescr_Y));
4313:     PetscCallCUSPARSE(cusparseDestroyCsrilu02Info(fs->ilu0Info_M));
4314:     PetscCallCUSPARSE(cusparseDestroyCsric02Info(fs->ic0Info_M));
4315:     PetscCall(PetscFree(fs->csrRowPtr_h));
4316:     PetscCall(PetscFree(fs->csrVal_h));
4317:     PetscCall(PetscFree(fs->diag_h));
4318:     fs->createdTransposeSpSVDescr    = PETSC_FALSE;
4319:     fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;
4320: #endif
4321:   }
4322:   PetscFunctionReturn(PETSC_SUCCESS);
4323: }

4325: static PetscErrorCode MatSeqAIJCUSPARSETriFactors_Destroy(Mat_SeqAIJCUSPARSETriFactors **trifactors)
4326: {
4327:   PetscFunctionBegin;
4328:   if (*trifactors) {
4329:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(trifactors));
4330:     PetscCallCUSPARSE(cusparseDestroy((*trifactors)->handle));
4331:     PetscCall(PetscFree(*trifactors));
4332:   }
4333:   PetscFunctionReturn(PETSC_SUCCESS);
4334: }

4336: struct IJCompare {
4337:   __host__ __device__ inline bool operator()(const thrust::tuple<PetscInt, PetscInt> &t1, const thrust::tuple<PetscInt, PetscInt> &t2)
4338:   {
4339:     if (thrust::get<0>(t1) < thrust::get<0>(t2)) return true;
4340:     if (thrust::get<0>(t1) == thrust::get<0>(t2)) return thrust::get<1>(t1) < thrust::get<1>(t2);
4341:     return false;
4342:   }
4343: };

4345: static PetscErrorCode MatSeqAIJCUSPARSEInvalidateTranspose(Mat A, PetscBool destroy)
4346: {
4347:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;

4349:   PetscFunctionBegin;
4350:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4351:   if (!cusp) PetscFunctionReturn(PETSC_SUCCESS);
4352:   if (destroy) {
4353:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->matTranspose, cusp->format));
4354:     delete cusp->csr2csc_i;
4355:     cusp->csr2csc_i = NULL;
4356:   }
4357:   A->transupdated = PETSC_FALSE;
4358:   PetscFunctionReturn(PETSC_SUCCESS);
4359: }

4361: static PetscErrorCode MatCOOStructDestroy_SeqAIJCUSPARSE(void **data)
4362: {
4363:   MatCOOStruct_SeqAIJ *coo = (MatCOOStruct_SeqAIJ *)*data;

4365:   PetscFunctionBegin;
4366:   PetscCallCUDA(cudaFree(coo->perm));
4367:   PetscCallCUDA(cudaFree(coo->jmap));
4368:   PetscCall(PetscFree(coo));
4369:   PetscFunctionReturn(PETSC_SUCCESS);
4370: }

4372: static PetscErrorCode MatSetPreallocationCOO_SeqAIJCUSPARSE(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
4373: {
4374:   PetscBool            dev_ij = PETSC_FALSE;
4375:   PetscMemType         mtype  = PETSC_MEMTYPE_HOST;
4376:   PetscInt            *i, *j;
4377:   PetscContainer       container_h;
4378:   MatCOOStruct_SeqAIJ *coo_h, *coo_d;

4380:   PetscFunctionBegin;
4381:   PetscCall(PetscGetMemType(coo_i, &mtype));
4382:   if (PetscMemTypeDevice(mtype)) {
4383:     dev_ij = PETSC_TRUE;
4384:     PetscCall(PetscMalloc2(coo_n, &i, coo_n, &j));
4385:     PetscCallCUDA(cudaMemcpy(i, coo_i, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4386:     PetscCallCUDA(cudaMemcpy(j, coo_j, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4387:   } else {
4388:     i = coo_i;
4389:     j = coo_j;
4390:   }

4392:   PetscCall(MatSetPreallocationCOO_SeqAIJ(mat, coo_n, i, j));
4393:   if (dev_ij) PetscCall(PetscFree2(i, j));
4394:   mat->offloadmask = PETSC_OFFLOAD_CPU;
4395:   // Create the GPU memory
4396:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(mat));

4398:   // Copy the COO struct to device
4399:   PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container_h));
4400:   PetscCall(PetscContainerGetPointer(container_h, (void **)&coo_h));
4401:   PetscCall(PetscMalloc1(1, &coo_d));
4402:   *coo_d = *coo_h; // do a shallow copy and then amend some fields that need to be different
4403:   PetscCallCUDA(cudaMalloc((void **)&coo_d->jmap, (coo_h->nz + 1) * sizeof(PetscCount)));
4404:   PetscCallCUDA(cudaMemcpy(coo_d->jmap, coo_h->jmap, (coo_h->nz + 1) * sizeof(PetscCount), cudaMemcpyHostToDevice));
4405:   PetscCallCUDA(cudaMalloc((void **)&coo_d->perm, coo_h->Atot * sizeof(PetscCount)));
4406:   PetscCallCUDA(cudaMemcpy(coo_d->perm, coo_h->perm, coo_h->Atot * sizeof(PetscCount), cudaMemcpyHostToDevice));

4408:   // Put the COO struct in a container and then attach that to the matrix
4409:   PetscCall(PetscObjectContainerCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Device", coo_d, MatCOOStructDestroy_SeqAIJCUSPARSE));
4410:   PetscFunctionReturn(PETSC_SUCCESS);
4411: }

4413: __global__ static void MatAddCOOValues(const PetscScalar kv[], PetscCount nnz, const PetscCount jmap[], const PetscCount perm[], InsertMode imode, PetscScalar a[])
4414: {
4415:   PetscCount       i         = blockIdx.x * blockDim.x + threadIdx.x;
4416:   const PetscCount grid_size = gridDim.x * blockDim.x;
4417:   for (; i < nnz; i += grid_size) {
4418:     PetscScalar sum = 0.0;
4419:     for (PetscCount k = jmap[i]; k < jmap[i + 1]; k++) sum += kv[perm[k]];
4420:     a[i] = (imode == INSERT_VALUES ? 0.0 : a[i]) + sum;
4421:   }
4422: }

4424: static PetscErrorCode MatSetValuesCOO_SeqAIJCUSPARSE(Mat A, const PetscScalar v[], InsertMode imode)
4425: {
4426:   Mat_SeqAIJ          *seq  = (Mat_SeqAIJ *)A->data;
4427:   Mat_SeqAIJCUSPARSE  *dev  = (Mat_SeqAIJCUSPARSE *)A->spptr;
4428:   PetscCount           Annz = seq->nz;
4429:   PetscMemType         memtype;
4430:   const PetscScalar   *v1 = v;
4431:   PetscScalar         *Aa;
4432:   PetscContainer       container;
4433:   MatCOOStruct_SeqAIJ *coo;

4435:   PetscFunctionBegin;
4436:   if (!dev->mat) PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));

4438:   PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Device", (PetscObject *)&container));
4439:   PetscCall(PetscContainerGetPointer(container, (void **)&coo));

4441:   PetscCall(PetscGetMemType(v, &memtype));
4442:   if (PetscMemTypeHost(memtype)) { /* If user gave v[] in host, we might need to copy it to device if any */
4443:     PetscCallCUDA(cudaMalloc((void **)&v1, coo->n * sizeof(PetscScalar)));
4444:     PetscCallCUDA(cudaMemcpy((void *)v1, v, coo->n * sizeof(PetscScalar), cudaMemcpyHostToDevice));
4445:   }

4447:   if (imode == INSERT_VALUES) PetscCall(MatSeqAIJCUSPARSEGetArrayWrite(A, &Aa));
4448:   else PetscCall(MatSeqAIJCUSPARSEGetArray(A, &Aa));

4450:   PetscCall(PetscLogGpuTimeBegin());
4451:   if (Annz) {
4452:     MatAddCOOValues<<<((int)(Annz + 255) / 256), 256>>>(v1, Annz, coo->jmap, coo->perm, imode, Aa);
4453:     PetscCallCUDA(cudaPeekAtLastError());
4454:   }
4455:   PetscCall(PetscLogGpuTimeEnd());

4457:   if (imode == INSERT_VALUES) PetscCall(MatSeqAIJCUSPARSERestoreArrayWrite(A, &Aa));
4458:   else PetscCall(MatSeqAIJCUSPARSERestoreArray(A, &Aa));

4460:   if (PetscMemTypeHost(memtype)) PetscCallCUDA(cudaFree((void *)v1));
4461:   PetscFunctionReturn(PETSC_SUCCESS);
4462: }

4464: /*@C
4465:   MatSeqAIJCUSPARSEGetIJ - returns the device row storage `i` and `j` indices for `MATSEQAIJCUSPARSE` matrices.

4467:   Not Collective

4469:   Input Parameters:
4470: + A          - the matrix
4471: - compressed - `PETSC_TRUE` or `PETSC_FALSE` indicating the matrix data structure should be always returned in compressed form

4473:   Output Parameters:
4474: + i - the CSR row pointers, these are always `int` even when PETSc is configured with `--with-64-bit-indices`
4475: - j - the CSR column indices, these are always `int` even when PETSc is configured with `--with-64-bit-indices`

4477:   Level: developer

4479:   Note:
4480:   When compressed is true, the CSR structure does not contain empty rows

4482: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSERestoreIJ()`, `MatSeqAIJCUSPARSEGetArrayRead()`
4483: @*/
4484: PetscErrorCode MatSeqAIJCUSPARSEGetIJ(Mat A, PetscBool compressed, const int **i, const int **j)
4485: {
4486:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4487:   CsrMatrix          *csr;
4488:   Mat_SeqAIJ         *a = (Mat_SeqAIJ *)A->data;

4490:   PetscFunctionBegin;
4492:   if (!i || !j) PetscFunctionReturn(PETSC_SUCCESS);
4493:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4494:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4495:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4496:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4497:   csr = (CsrMatrix *)cusp->mat->mat;
4498:   if (i) {
4499:     if (!compressed && a->compressedrow.use) { /* need full row offset */
4500:       if (!cusp->rowoffsets_gpu) {
4501:         cusp->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
4502:         cusp->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);
4503:         PetscCall(PetscLogCpuToGpu((A->rmap->n + 1) * sizeof(PetscInt)));
4504:       }
4505:       *i = cusp->rowoffsets_gpu->data().get();
4506:     } else *i = csr->row_offsets->data().get();
4507:   }
4508:   if (j) *j = csr->column_indices->data().get();
4509:   PetscFunctionReturn(PETSC_SUCCESS);
4510: }

4512: /*@C
4513:   MatSeqAIJCUSPARSERestoreIJ - restore the device row storage `i` and `j` indices obtained with `MatSeqAIJCUSPARSEGetIJ()`

4515:   Not Collective

4517:   Input Parameters:
4518: + A          - the matrix
4519: . compressed - `PETSC_TRUE` or `PETSC_FALSE` indicating the matrix data structure should be always returned in compressed form
4520: . i          - the CSR row pointers
4521: - j          - the CSR column indices

4523:   Level: developer

4525: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetIJ()`
4526: @*/
4527: PetscErrorCode MatSeqAIJCUSPARSERestoreIJ(Mat A, PetscBool compressed, const int **i, const int **j)
4528: {
4529:   PetscFunctionBegin;
4531:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4532:   if (i) *i = NULL;
4533:   if (j) *j = NULL;
4534:   (void)compressed;
4535:   PetscFunctionReturn(PETSC_SUCCESS);
4536: }

4538: /*@C
4539:   MatSeqAIJCUSPARSEGetArrayRead - gives read-only access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix nonzero entries are stored

4541:   Not Collective

4543:   Input Parameter:
4544: . A - a `MATSEQAIJCUSPARSE` matrix

4546:   Output Parameter:
4547: . a - pointer to the device data

4549:   Level: developer

4551:   Note:
4552:   Will trigger host-to-device copies if the most up-to-date matrix data is on the host

4554: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`, `MatSeqAIJCUSPARSEGetArrayWrite()`, `MatSeqAIJCUSPARSERestoreArrayRead()`
4555: @*/
4556: PetscErrorCode MatSeqAIJCUSPARSEGetArrayRead(Mat A, const PetscScalar **a)
4557: {
4558:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4559:   CsrMatrix          *csr;

4561:   PetscFunctionBegin;
4563:   PetscAssertPointer(a, 2);
4564:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4565:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4566:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4567:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4568:   csr = (CsrMatrix *)cusp->mat->mat;
4569:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4570:   *a = csr->values->data().get();
4571:   PetscFunctionReturn(PETSC_SUCCESS);
4572: }

4574: /*@C
4575:   MatSeqAIJCUSPARSERestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJCUSPARSEGetArrayRead()`

4577:   Not Collective

4579:   Input Parameters:
4580: + A - a `MATSEQAIJCUSPARSE` matrix
4581: - a - pointer to the device data

4583:   Level: developer

4585: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayRead()`
4586: @*/
4587: PetscErrorCode MatSeqAIJCUSPARSERestoreArrayRead(Mat A, const PetscScalar **a)
4588: {
4589:   PetscFunctionBegin;
4591:   PetscAssertPointer(a, 2);
4592:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4593:   *a = NULL;
4594:   PetscFunctionReturn(PETSC_SUCCESS);
4595: }

4597: /*@C
4598:   MatSeqAIJCUSPARSEGetArray - gives read-write access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix is stored

4600:   Not Collective

4602:   Input Parameter:
4603: . A - a `MATSEQAIJCUSPARSE` matrix

4605:   Output Parameter:
4606: . a - pointer to the device data

4608:   Level: developer

4610:   Note:
4611:   Will trigger host-to-device copies if the most up-to-date matrix data is on the host

4613: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayRead()`, `MatSeqAIJCUSPARSEGetArrayWrite()`, `MatSeqAIJCUSPARSERestoreArray()`
4614: @*/
4615: PetscErrorCode MatSeqAIJCUSPARSEGetArray(Mat A, PetscScalar **a)
4616: {
4617:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4618:   CsrMatrix          *csr;

4620:   PetscFunctionBegin;
4622:   PetscAssertPointer(a, 2);
4623:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4624:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4625:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4626:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4627:   csr = (CsrMatrix *)cusp->mat->mat;
4628:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4629:   *a             = csr->values->data().get();
4630:   A->offloadmask = PETSC_OFFLOAD_GPU;
4631:   PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
4632:   PetscFunctionReturn(PETSC_SUCCESS);
4633: }
4634: /*@C
4635:   MatSeqAIJCUSPARSERestoreArray - restore the read-write access array obtained from `MatSeqAIJCUSPARSEGetArray()`

4637:   Not Collective

4639:   Input Parameters:
4640: + A - a `MATSEQAIJCUSPARSE` matrix
4641: - a - pointer to the device data

4643:   Level: developer

4645: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`
4646: @*/
4647: PetscErrorCode MatSeqAIJCUSPARSERestoreArray(Mat A, PetscScalar **a)
4648: {
4649:   PetscFunctionBegin;
4651:   PetscAssertPointer(a, 2);
4652:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4653:   PetscCall(PetscObjectStateIncrease((PetscObject)A));
4654:   *a = NULL;
4655:   PetscFunctionReturn(PETSC_SUCCESS);
4656: }

4658: /*@C
4659:   MatSeqAIJCUSPARSEGetArrayWrite - gives write access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix is stored

4661:   Not Collective

4663:   Input Parameter:
4664: . A - a `MATSEQAIJCUSPARSE` matrix

4666:   Output Parameter:
4667: . a - pointer to the device data

4669:   Level: developer

4671:   Note:
4672:   Does not trigger any host to device copies.

4674:   It marks the data GPU valid so users must set all the values in `a` to ensure out-of-date data is not considered current

4676: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`, `MatSeqAIJCUSPARSEGetArrayRead()`, `MatSeqAIJCUSPARSERestoreArrayWrite()`
4677: @*/
4678: PetscErrorCode MatSeqAIJCUSPARSEGetArrayWrite(Mat A, PetscScalar **a)
4679: {
4680:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4681:   CsrMatrix          *csr;

4683:   PetscFunctionBegin;
4685:   PetscAssertPointer(a, 2);
4686:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4687:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4688:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4689:   csr = (CsrMatrix *)cusp->mat->mat;
4690:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4691:   *a             = csr->values->data().get();
4692:   A->offloadmask = PETSC_OFFLOAD_GPU;
4693:   PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
4694:   PetscFunctionReturn(PETSC_SUCCESS);
4695: }

4697: /*@C
4698:   MatSeqAIJCUSPARSERestoreArrayWrite - restore the write-only access array obtained from `MatSeqAIJCUSPARSEGetArrayWrite()`

4700:   Not Collective

4702:   Input Parameters:
4703: + A - a `MATSEQAIJCUSPARSE` matrix
4704: - a - pointer to the device data

4706:   Level: developer

4708: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayWrite()`
4709: @*/
4710: PetscErrorCode MatSeqAIJCUSPARSERestoreArrayWrite(Mat A, PetscScalar **a)
4711: {
4712:   PetscFunctionBegin;
4714:   PetscAssertPointer(a, 2);
4715:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4716:   PetscCall(PetscObjectStateIncrease((PetscObject)A));
4717:   *a = NULL;
4718:   PetscFunctionReturn(PETSC_SUCCESS);
4719: }

4721: struct IJCompare4 {
4722:   __host__ __device__ inline bool operator()(const thrust::tuple<int, int, PetscScalar, int> &t1, const thrust::tuple<int, int, PetscScalar, int> &t2)
4723:   {
4724:     if (thrust::get<0>(t1) < thrust::get<0>(t2)) return true;
4725:     if (thrust::get<0>(t1) == thrust::get<0>(t2)) return thrust::get<1>(t1) < thrust::get<1>(t2);
4726:     return false;
4727:   }
4728: };

4730: struct Shift {
4731:   int _shift;

4733:   Shift(int shift) : _shift(shift) { }
4734:   __host__ __device__ inline int operator()(const int &c) { return c + _shift; }
4735: };

4737: /* merges two SeqAIJCUSPARSE matrices A, B by concatenating their rows. [A';B']' operation in MATLAB notation */
4738: PetscErrorCode MatSeqAIJCUSPARSEMergeMats(Mat A, Mat B, MatReuse reuse, Mat *C)
4739: {
4740:   Mat_SeqAIJ                   *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data, *c;
4741:   Mat_SeqAIJCUSPARSE           *Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr, *Bcusp = (Mat_SeqAIJCUSPARSE *)B->spptr, *Ccusp;
4742:   Mat_SeqAIJCUSPARSEMultStruct *Cmat;
4743:   CsrMatrix                    *Acsr, *Bcsr, *Ccsr;
4744:   PetscInt                      Annz, Bnnz;
4745:   cusparseStatus_t              stat;
4746:   PetscInt                      i, m, n, zero = 0;

4748:   PetscFunctionBegin;
4751:   PetscAssertPointer(C, 4);
4752:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4753:   PetscCheckTypeName(B, MATSEQAIJCUSPARSE);
4754:   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);
4755:   PetscCheck(reuse != MAT_INPLACE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_INPLACE_MATRIX not supported");
4756:   PetscCheck(Acusp->format != MAT_CUSPARSE_ELL && Acusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4757:   PetscCheck(Bcusp->format != MAT_CUSPARSE_ELL && Bcusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4758:   if (reuse == MAT_INITIAL_MATRIX) {
4759:     m = A->rmap->n;
4760:     n = A->cmap->n + B->cmap->n;
4761:     PetscCall(MatCreate(PETSC_COMM_SELF, C));
4762:     PetscCall(MatSetSizes(*C, m, n, m, n));
4763:     PetscCall(MatSetType(*C, MATSEQAIJCUSPARSE));
4764:     c                       = (Mat_SeqAIJ *)(*C)->data;
4765:     Ccusp                   = (Mat_SeqAIJCUSPARSE *)(*C)->spptr;
4766:     Cmat                    = new Mat_SeqAIJCUSPARSEMultStruct;
4767:     Ccsr                    = new CsrMatrix;
4768:     Cmat->cprowIndices      = NULL;
4769:     c->compressedrow.use    = PETSC_FALSE;
4770:     c->compressedrow.nrows  = 0;
4771:     c->compressedrow.i      = NULL;
4772:     c->compressedrow.rindex = NULL;
4773:     Ccusp->workVector       = NULL;
4774:     Ccusp->nrows            = m;
4775:     Ccusp->mat              = Cmat;
4776:     Ccusp->mat->mat         = Ccsr;
4777:     Ccsr->num_rows          = m;
4778:     Ccsr->num_cols          = n;
4779:     PetscCallCUSPARSE(cusparseCreateMatDescr(&Cmat->descr));
4780:     PetscCallCUSPARSE(cusparseSetMatIndexBase(Cmat->descr, CUSPARSE_INDEX_BASE_ZERO));
4781:     PetscCallCUSPARSE(cusparseSetMatType(Cmat->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
4782:     PetscCallCUDA(cudaMalloc((void **)&Cmat->alpha_one, sizeof(PetscScalar)));
4783:     PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_zero, sizeof(PetscScalar)));
4784:     PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_one, sizeof(PetscScalar)));
4785:     PetscCallCUDA(cudaMemcpy(Cmat->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4786:     PetscCallCUDA(cudaMemcpy(Cmat->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4787:     PetscCallCUDA(cudaMemcpy(Cmat->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4788:     PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4789:     PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));
4790:     PetscCheck(Acusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4791:     PetscCheck(Bcusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");

4793:     Acsr                 = (CsrMatrix *)Acusp->mat->mat;
4794:     Bcsr                 = (CsrMatrix *)Bcusp->mat->mat;
4795:     Annz                 = (PetscInt)Acsr->column_indices->size();
4796:     Bnnz                 = (PetscInt)Bcsr->column_indices->size();
4797:     c->nz                = Annz + Bnnz;
4798:     Ccsr->row_offsets    = new THRUSTINTARRAY32(m + 1);
4799:     Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
4800:     Ccsr->values         = new THRUSTARRAY(c->nz);
4801:     Ccsr->num_entries    = c->nz;
4802:     Ccusp->coords        = new THRUSTINTARRAY(c->nz);
4803:     if (c->nz) {
4804:       auto              Acoo = new THRUSTINTARRAY32(Annz);
4805:       auto              Bcoo = new THRUSTINTARRAY32(Bnnz);
4806:       auto              Ccoo = new THRUSTINTARRAY32(c->nz);
4807:       THRUSTINTARRAY32 *Aroff, *Broff;

4809:       if (a->compressedrow.use) { /* need full row offset */
4810:         if (!Acusp->rowoffsets_gpu) {
4811:           Acusp->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
4812:           Acusp->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);
4813:           PetscCall(PetscLogCpuToGpu((A->rmap->n + 1) * sizeof(PetscInt)));
4814:         }
4815:         Aroff = Acusp->rowoffsets_gpu;
4816:       } else Aroff = Acsr->row_offsets;
4817:       if (b->compressedrow.use) { /* need full row offset */
4818:         if (!Bcusp->rowoffsets_gpu) {
4819:           Bcusp->rowoffsets_gpu = new THRUSTINTARRAY32(B->rmap->n + 1);
4820:           Bcusp->rowoffsets_gpu->assign(b->i, b->i + B->rmap->n + 1);
4821:           PetscCall(PetscLogCpuToGpu((B->rmap->n + 1) * sizeof(PetscInt)));
4822:         }
4823:         Broff = Bcusp->rowoffsets_gpu;
4824:       } else Broff = Bcsr->row_offsets;
4825:       PetscCall(PetscLogGpuTimeBegin());
4826:       stat = cusparseXcsr2coo(Acusp->handle, Aroff->data().get(), Annz, m, Acoo->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4827:       PetscCallCUSPARSE(stat);
4828:       stat = cusparseXcsr2coo(Bcusp->handle, Broff->data().get(), Bnnz, m, Bcoo->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4829:       PetscCallCUSPARSE(stat);
4830:       /* Issues when using bool with large matrices on SUMMIT 10.2.89 */
4831:       auto Aperm = thrust::make_constant_iterator(1);
4832:       auto Bperm = thrust::make_constant_iterator(0);
4833: #if PETSC_PKG_CUDA_VERSION_GE(10, 0, 0)
4834:       auto Bcib = thrust::make_transform_iterator(Bcsr->column_indices->begin(), Shift(A->cmap->n));
4835:       auto Bcie = thrust::make_transform_iterator(Bcsr->column_indices->end(), Shift(A->cmap->n));
4836: #else
4837:       /* there are issues instantiating the merge operation using a transform iterator for the columns of B */
4838:       auto Bcib = Bcsr->column_indices->begin();
4839:       auto Bcie = Bcsr->column_indices->end();
4840:       thrust::transform(Bcib, Bcie, Bcib, Shift(A->cmap->n));
4841: #endif
4842:       auto wPerm = new THRUSTINTARRAY32(Annz + Bnnz);
4843:       auto Azb   = thrust::make_zip_iterator(thrust::make_tuple(Acoo->begin(), Acsr->column_indices->begin(), Acsr->values->begin(), Aperm));
4844:       auto Aze   = thrust::make_zip_iterator(thrust::make_tuple(Acoo->end(), Acsr->column_indices->end(), Acsr->values->end(), Aperm));
4845:       auto Bzb   = thrust::make_zip_iterator(thrust::make_tuple(Bcoo->begin(), Bcib, Bcsr->values->begin(), Bperm));
4846:       auto Bze   = thrust::make_zip_iterator(thrust::make_tuple(Bcoo->end(), Bcie, Bcsr->values->end(), Bperm));
4847:       auto Czb   = thrust::make_zip_iterator(thrust::make_tuple(Ccoo->begin(), Ccsr->column_indices->begin(), Ccsr->values->begin(), wPerm->begin()));
4848:       auto p1    = Ccusp->coords->begin();
4849:       auto p2    = Ccusp->coords->begin();
4850: #if CCCL_VERSION >= 3001000
4851:       cuda::std::advance(p2, Annz);
4852: #else
4853:       thrust::advance(p2, Annz);
4854: #endif
4855:       PetscCallThrust(thrust::merge(thrust::device, Azb, Aze, Bzb, Bze, Czb, IJCompare4()));
4856: #if PETSC_PKG_CUDA_VERSION_LT(10, 0, 0)
4857:       thrust::transform(Bcib, Bcie, Bcib, Shift(-A->cmap->n));
4858: #endif
4859:       auto cci = thrust::make_counting_iterator(zero);
4860:       auto cce = thrust::make_counting_iterator(c->nz);
4861: #if 0 //Errors on SUMMIT cuda 11.1.0
4862:       PetscCallThrust(thrust::partition_copy(thrust::device,cci,cce,wPerm->begin(),p1,p2,thrust::identity<int>()));
4863: #else
4864:   #if PETSC_PKG_CUDA_VERSION_LT(12, 9, 0) || PetscDefined(HAVE_THRUST)
4865:       auto pred = thrust::identity<int>();
4866:   #else
4867:       auto pred = cuda::std::identity();
4868:   #endif
4869:       PetscCallThrust(thrust::copy_if(thrust::device, cci, cce, wPerm->begin(), p1, pred));
4870:       PetscCallThrust(thrust::remove_copy_if(thrust::device, cci, cce, wPerm->begin(), p2, pred));
4871: #endif
4872:       stat = cusparseXcoo2csr(Ccusp->handle, Ccoo->data().get(), c->nz, m, Ccsr->row_offsets->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4873:       PetscCallCUSPARSE(stat);
4874:       PetscCall(PetscLogGpuTimeEnd());
4875:       delete wPerm;
4876:       delete Acoo;
4877:       delete Bcoo;
4878:       delete Ccoo;
4879: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4880:       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);
4881:       PetscCallCUSPARSE(stat);
4882: #endif
4883:       if (A->form_explicit_transpose && B->form_explicit_transpose) { /* if A and B have the transpose, generate C transpose too */
4884:         PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
4885:         PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(B));
4886:         PetscBool                     AT = Acusp->matTranspose ? PETSC_TRUE : PETSC_FALSE, BT = Bcusp->matTranspose ? PETSC_TRUE : PETSC_FALSE;
4887:         Mat_SeqAIJCUSPARSEMultStruct *CmatT = new Mat_SeqAIJCUSPARSEMultStruct;
4888:         CsrMatrix                    *CcsrT = new CsrMatrix;
4889:         CsrMatrix                    *AcsrT = AT ? (CsrMatrix *)Acusp->matTranspose->mat : NULL;
4890:         CsrMatrix                    *BcsrT = BT ? (CsrMatrix *)Bcusp->matTranspose->mat : NULL;

4892:         (*C)->form_explicit_transpose = PETSC_TRUE;
4893:         (*C)->transupdated            = PETSC_TRUE;
4894:         Ccusp->rowoffsets_gpu         = NULL;
4895:         CmatT->cprowIndices           = NULL;
4896:         CmatT->mat                    = CcsrT;
4897:         CcsrT->num_rows               = n;
4898:         CcsrT->num_cols               = m;
4899:         CcsrT->num_entries            = c->nz;

4901:         CcsrT->row_offsets    = new THRUSTINTARRAY32(n + 1);
4902:         CcsrT->column_indices = new THRUSTINTARRAY32(c->nz);
4903:         CcsrT->values         = new THRUSTARRAY(c->nz);

4905:         PetscCall(PetscLogGpuTimeBegin());
4906:         auto rT = CcsrT->row_offsets->begin();
4907:         if (AT) {
4908:           rT = thrust::copy(AcsrT->row_offsets->begin(), AcsrT->row_offsets->end(), rT);
4909: #if CCCL_VERSION >= 3001000
4910:           cuda::std::advance(rT, -1);
4911: #else
4912:           thrust::advance(rT, -1);
4913: #endif
4914:         }
4915:         if (BT) {
4916:           auto titb = thrust::make_transform_iterator(BcsrT->row_offsets->begin(), Shift(a->nz));
4917:           auto tite = thrust::make_transform_iterator(BcsrT->row_offsets->end(), Shift(a->nz));
4918:           thrust::copy(titb, tite, rT);
4919:         }
4920:         auto cT = CcsrT->column_indices->begin();
4921:         if (AT) cT = thrust::copy(AcsrT->column_indices->begin(), AcsrT->column_indices->end(), cT);
4922:         if (BT) thrust::copy(BcsrT->column_indices->begin(), BcsrT->column_indices->end(), cT);
4923:         auto vT = CcsrT->values->begin();
4924:         if (AT) vT = thrust::copy(AcsrT->values->begin(), AcsrT->values->end(), vT);
4925:         if (BT) thrust::copy(BcsrT->values->begin(), BcsrT->values->end(), vT);
4926:         PetscCall(PetscLogGpuTimeEnd());

4928:         PetscCallCUSPARSE(cusparseCreateMatDescr(&CmatT->descr));
4929:         PetscCallCUSPARSE(cusparseSetMatIndexBase(CmatT->descr, CUSPARSE_INDEX_BASE_ZERO));
4930:         PetscCallCUSPARSE(cusparseSetMatType(CmatT->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
4931:         PetscCallCUDA(cudaMalloc((void **)&CmatT->alpha_one, sizeof(PetscScalar)));
4932:         PetscCallCUDA(cudaMalloc((void **)&CmatT->beta_zero, sizeof(PetscScalar)));
4933:         PetscCallCUDA(cudaMalloc((void **)&CmatT->beta_one, sizeof(PetscScalar)));
4934:         PetscCallCUDA(cudaMemcpy(CmatT->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4935:         PetscCallCUDA(cudaMemcpy(CmatT->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4936:         PetscCallCUDA(cudaMemcpy(CmatT->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4937: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4938:         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);
4939:         PetscCallCUSPARSE(stat);
4940: #endif
4941:         Ccusp->matTranspose = CmatT;
4942:       }
4943:     }

4945:     c->free_a = PETSC_TRUE;
4946:     PetscCall(PetscShmgetAllocateArray(c->nz, sizeof(PetscInt), (void **)&c->j));
4947:     PetscCall(PetscShmgetAllocateArray(m + 1, sizeof(PetscInt), (void **)&c->i));
4948:     c->free_ij = PETSC_TRUE;
4949:     if (PetscDefined(USE_64BIT_INDICES)) { /* 32 to 64-bit conversion on the GPU and then copy to host (lazy) */
4950:       THRUSTINTARRAY ii(Ccsr->row_offsets->size());
4951:       THRUSTINTARRAY jj(Ccsr->column_indices->size());
4952:       ii = *Ccsr->row_offsets;
4953:       jj = *Ccsr->column_indices;
4954:       PetscCallCUDA(cudaMemcpy(c->i, ii.data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4955:       PetscCallCUDA(cudaMemcpy(c->j, jj.data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4956:     } else {
4957:       PetscCallCUDA(cudaMemcpy(c->i, Ccsr->row_offsets->data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4958:       PetscCallCUDA(cudaMemcpy(c->j, Ccsr->column_indices->data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4959:     }
4960:     PetscCall(PetscLogGpuToCpu((Ccsr->column_indices->size() + Ccsr->row_offsets->size()) * sizeof(PetscInt)));
4961:     PetscCall(PetscMalloc1(m, &c->ilen));
4962:     PetscCall(PetscMalloc1(m, &c->imax));
4963:     c->maxnz         = c->nz;
4964:     c->nonzerorowcnt = 0;
4965:     c->rmax          = 0;
4966:     for (i = 0; i < m; i++) {
4967:       const PetscInt nn = c->i[i + 1] - c->i[i];
4968:       c->ilen[i] = c->imax[i] = nn;
4969:       c->nonzerorowcnt += (PetscInt)!!nn;
4970:       c->rmax = PetscMax(c->rmax, nn);
4971:     }
4972:     PetscCall(PetscMalloc1(c->nz, &c->a));
4973:     (*C)->nonzerostate++;
4974:     PetscCall(PetscLayoutSetUp((*C)->rmap));
4975:     PetscCall(PetscLayoutSetUp((*C)->cmap));
4976:     Ccusp->nonzerostate = (*C)->nonzerostate;
4977:     (*C)->preallocated  = PETSC_TRUE;
4978:   } else {
4979:     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);
4980:     c = (Mat_SeqAIJ *)(*C)->data;
4981:     if (c->nz) {
4982:       Ccusp = (Mat_SeqAIJCUSPARSE *)(*C)->spptr;
4983:       PetscCheck(Ccusp->coords, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing coords");
4984:       PetscCheck(Ccusp->format != MAT_CUSPARSE_ELL && Ccusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4985:       PetscCheck(Ccusp->nonzerostate == (*C)->nonzerostate, PETSC_COMM_SELF, PETSC_ERR_COR, "Wrong nonzerostate");
4986:       PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4987:       PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));
4988:       PetscCheck(Acusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4989:       PetscCheck(Bcusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4990:       Acsr = (CsrMatrix *)Acusp->mat->mat;
4991:       Bcsr = (CsrMatrix *)Bcusp->mat->mat;
4992:       Ccsr = (CsrMatrix *)Ccusp->mat->mat;
4993:       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());
4994:       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());
4995:       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());
4996:       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);
4997:       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());
4998:       auto pmid = Ccusp->coords->begin();
4999: #if CCCL_VERSION >= 3001000
5000:       cuda::std::advance(pmid, Acsr->num_entries);
5001: #else
5002:       thrust::advance(pmid, Acsr->num_entries);
5003: #endif
5004:       PetscCall(PetscLogGpuTimeBegin());
5005:       auto zibait = thrust::make_zip_iterator(thrust::make_tuple(Acsr->values->begin(), thrust::make_permutation_iterator(Ccsr->values->begin(), Ccusp->coords->begin())));
5006:       auto zieait = thrust::make_zip_iterator(thrust::make_tuple(Acsr->values->end(), thrust::make_permutation_iterator(Ccsr->values->begin(), pmid)));
5007:       thrust::for_each(zibait, zieait, VecCUDAEquals());
5008:       auto zibbit = thrust::make_zip_iterator(thrust::make_tuple(Bcsr->values->begin(), thrust::make_permutation_iterator(Ccsr->values->begin(), pmid)));
5009:       auto ziebit = thrust::make_zip_iterator(thrust::make_tuple(Bcsr->values->end(), thrust::make_permutation_iterator(Ccsr->values->begin(), Ccusp->coords->end())));
5010:       thrust::for_each(zibbit, ziebit, VecCUDAEquals());
5011:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(*C, PETSC_FALSE));
5012:       if (A->form_explicit_transpose && B->form_explicit_transpose && (*C)->form_explicit_transpose) {
5013:         PetscCheck(Ccusp->matTranspose, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing transpose Mat_SeqAIJCUSPARSEMultStruct");
5014:         PetscBool  AT = Acusp->matTranspose ? PETSC_TRUE : PETSC_FALSE, BT = Bcusp->matTranspose ? PETSC_TRUE : PETSC_FALSE;
5015:         CsrMatrix *AcsrT = AT ? (CsrMatrix *)Acusp->matTranspose->mat : NULL;
5016:         CsrMatrix *BcsrT = BT ? (CsrMatrix *)Bcusp->matTranspose->mat : NULL;
5017:         CsrMatrix *CcsrT = (CsrMatrix *)Ccusp->matTranspose->mat;
5018:         auto       vT    = CcsrT->values->begin();
5019:         if (AT) vT = thrust::copy(AcsrT->values->begin(), AcsrT->values->end(), vT);
5020:         if (BT) thrust::copy(BcsrT->values->begin(), BcsrT->values->end(), vT);
5021:         (*C)->transupdated = PETSC_TRUE;
5022:       }
5023:       PetscCall(PetscLogGpuTimeEnd());
5024:     }
5025:   }
5026:   PetscCall(PetscObjectStateIncrease((PetscObject)*C));
5027:   (*C)->assembled     = PETSC_TRUE;
5028:   (*C)->was_assembled = PETSC_FALSE;
5029:   (*C)->offloadmask   = PETSC_OFFLOAD_GPU;
5030:   PetscFunctionReturn(PETSC_SUCCESS);
5031: }

5033: static PetscErrorCode MatSeqAIJCopySubArray_SeqAIJCUSPARSE(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
5034: {
5035:   bool               dmem;
5036:   const PetscScalar *av;

5038:   PetscFunctionBegin;
5039:   dmem = isCudaMem(v);
5040:   PetscCall(MatSeqAIJCUSPARSEGetArrayRead(A, &av));
5041:   if (n && idx) {
5042:     THRUSTINTARRAY widx(n);
5043:     widx.assign(idx, idx + n);
5044:     PetscCall(PetscLogCpuToGpu(n * sizeof(PetscInt)));

5046:     THRUSTARRAY                    *w = NULL;
5047:     thrust::device_ptr<PetscScalar> dv;
5048:     if (dmem) {
5049:       dv = thrust::device_pointer_cast(v);
5050:     } else {
5051:       w  = new THRUSTARRAY(n);
5052:       dv = w->data();
5053:     }
5054:     thrust::device_ptr<const PetscScalar> dav = thrust::device_pointer_cast(av);

5056:     auto zibit = thrust::make_zip_iterator(thrust::make_tuple(thrust::make_permutation_iterator(dav, widx.begin()), dv));
5057:     auto zieit = thrust::make_zip_iterator(thrust::make_tuple(thrust::make_permutation_iterator(dav, widx.end()), dv + n));
5058:     thrust::for_each(zibit, zieit, VecCUDAEquals());
5059:     if (w) PetscCallCUDA(cudaMemcpy(v, w->data().get(), n * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
5060:     delete w;
5061:   } else {
5062:     PetscCallCUDA(cudaMemcpy(v, av, n * sizeof(PetscScalar), dmem ? cudaMemcpyDeviceToDevice : cudaMemcpyDeviceToHost));
5063:   }
5064:   if (!dmem) PetscCall(PetscLogCpuToGpu(n * sizeof(PetscScalar)));
5065:   PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(A, &av));
5066:   PetscFunctionReturn(PETSC_SUCCESS);
5067: }