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 = a->diag;
206:   const MatScalar              *Aa = a->a;
207:   PetscInt                     *Mi, *Mj, Mnz;
208:   PetscScalar                  *Ma;

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

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

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

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

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

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

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

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

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

297:       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));
298:       fs->updatedSpSVAnalysis          = PETSC_TRUE;
299:       fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;
300:     }
301:   }
302:   PetscFunctionReturn(PETSC_SUCCESS);
303: }
304: #else
305: static PetscErrorCode MatSeqAIJCUSPARSEBuildILULowerTriMatrix(Mat A)
306: {
307:   Mat_SeqAIJ                        *a                  = (Mat_SeqAIJ *)A->data;
308:   PetscInt                           n                  = A->rmap->n;
309:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
310:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
311:   const PetscInt                    *ai = a->i, *aj = a->j, *vi;
312:   const MatScalar                   *aa = a->a, *v;
313:   PetscInt                          *AiLo, *AjLo;
314:   PetscInt                           i, nz, nzLower, offset, rowOffset;

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

473:           /* decrement the offset */
474:           offset -= (nz + 1);

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

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

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

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

500:         /* set the operation */
501:         upTriFactor->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

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

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

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

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

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

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

531:         PetscCallCUDA(WaitForCUDA());
532:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

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

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

550:           /* decrement the offset */
551:           offset -= (nz + 1);

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

568: static PetscErrorCode MatSeqAIJCUSPARSEILUAnalysisAndCopyToGPU(Mat A)
569: {
570:   Mat_SeqAIJ                   *a                  = (Mat_SeqAIJ *)A->data;
571:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
572:   IS                            isrow = a->row, isicol = a->icol;
573:   PetscBool                     row_identity, col_identity;
574:   PetscInt                      n = A->rmap->n;

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

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

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

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

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

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

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

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

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

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

661:       // Allocate work vectors in SpSv
662:       PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(*fs->X) * m));
663:       PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(*fs->Y) * m));

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

668:       // Query buffer sizes for SpSV and then allocate buffers
669:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_U));
670:       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));
671:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U));

673:       PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Ut)); // Ut solve uses the same matrix (spMatDescr_U), but different descr and buffer
674:       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));
675:       PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Ut, fs->spsvBufferSize_Ut));

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

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

710: // Solve Ut D U x = b
711: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_Cholesky(Mat A, Vec b, Vec x)
712: {
713:   Mat_SeqAIJCUSPARSETriFactors         *fs  = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
714:   Mat_SeqAIJ                           *aij = static_cast<Mat_SeqAIJ *>(A->data);
715:   const PetscScalar                    *barray;
716:   PetscScalar                          *xarray;
717:   thrust::device_ptr<const PetscScalar> bGPU;
718:   thrust::device_ptr<PetscScalar>       xGPU;
719:   const cusparseSpSVAlg_t               alg = CUSPARSE_SPSV_ALG_DEFAULT;
720:   PetscInt                              m   = A->rmap->n;

722:   PetscFunctionBegin;
723:   PetscCall(PetscLogGpuTimeBegin());
724:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
725:   PetscCall(VecCUDAGetArrayRead(b, &barray));
726:   xGPU = thrust::device_pointer_cast(xarray);
727:   bGPU = thrust::device_pointer_cast(barray);

729:   // Reorder b with the row permutation if needed, and wrap the result in fs->X
730:   if (fs->rpermIndices) {
731:     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)));
732:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
733:   } else {
734:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
735:   }

737:   // Solve Ut Y = X
738:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
739:   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));

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

749:   // Solve U X = Y
750:   if (fs->cpermIndices) { // if need to permute, we need to use the intermediate buffer X
751:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
752:   } else {
753:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
754:   }
755:   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));

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

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

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

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

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

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

823:         /* allocate space for the triangular factor information */
824:         PetscCall(PetscNew(&upTriFactor));
825:         upTriFactor->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

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

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

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

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

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

853:         /* set the operation */
854:         upTriFactor->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

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

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

869:         PetscCallCUDA(WaitForCUDA());
870:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

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

875:         /* allocate space for the triangular factor information */
876:         PetscCall(PetscNew(&loTriFactor));
877:         loTriFactor->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

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

890:         /* set the operation */
891:         loTriFactor->solveOp = CUSPARSE_OPERATION_TRANSPOSE;

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

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

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

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

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

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

921:         PetscCallCUDA(WaitForCUDA());
922:         PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

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

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

938:           /* first, set the diagonal elements */
939:           AAUp[offset] = 1.0 / v[nz];
940:           AALo[offset] = 1.0 / v[nz];

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

968: static PetscErrorCode MatSeqAIJCUSPARSEICCAnalysisAndCopyToGPU(Mat A)
969: {
970:   Mat_SeqAIJ                   *a                  = (Mat_SeqAIJ *)A->data;
971:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
972:   IS                            ip                 = a->row;
973:   PetscBool                     perm_identity;
974:   PetscInt                      n = A->rmap->n;

976:   PetscFunctionBegin;
977:   PetscCheck(cusparseTriFactors, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing cusparseTriFactors");

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

987:   A->offloadmask = PETSC_OFFLOAD_BOTH;

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

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

1010: static PetscErrorCode MatCholeskyFactorNumeric_SeqAIJCUSPARSE(Mat B, Mat A, const MatFactorInfo *info)
1011: {
1012:   PetscFunctionBegin;
1013:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
1014:   PetscCall(MatCholeskyFactorNumeric_SeqAIJ(B, A, info));
1015:   B->offloadmask = PETSC_OFFLOAD_CPU;

1017: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
1018:   B->ops->solve          = MatSolve_SeqAIJCUSPARSE_Cholesky;
1019:   B->ops->solvetranspose = MatSolve_SeqAIJCUSPARSE_Cholesky;
1020: #else
1021:   /* determine which version of MatSolve needs to be used. */
1022:   Mat_SeqAIJ *b  = (Mat_SeqAIJ *)B->data;
1023:   IS          ip = b->row;
1024:   PetscBool   perm_identity;

1026:   PetscCall(ISIdentity(ip, &perm_identity));
1027:   if (perm_identity) {
1028:     B->ops->solve          = MatSolve_SeqAIJCUSPARSE_NaturalOrdering;
1029:     B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering;
1030:   } else {
1031:     B->ops->solve          = MatSolve_SeqAIJCUSPARSE;
1032:     B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE;
1033:   }
1034: #endif
1035:   B->ops->matsolve          = NULL;
1036:   B->ops->matsolvetranspose = NULL;

1038:   /* get the triangular factors */
1039:   PetscCall(MatSeqAIJCUSPARSEICCAnalysisAndCopyToGPU(B));
1040:   PetscFunctionReturn(PETSC_SUCCESS);
1041: }

1043: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
1044: static PetscErrorCode MatSeqAIJCUSPARSEAnalyzeTransposeForSolve(Mat A)
1045: {
1046:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1047:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
1048:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
1049:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorT;
1050:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorT;
1051:   cusparseIndexBase_t                indexBase;
1052:   cusparseMatrixType_t               matrixType;
1053:   cusparseFillMode_t                 fillMode;
1054:   cusparseDiagType_t                 diagType;

1056:   PetscFunctionBegin;
1057:   /* allocate space for the transpose of the lower triangular factor */
1058:   PetscCall(PetscNew(&loTriFactorT));
1059:   loTriFactorT->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

1061:   /* set the matrix descriptors of the lower triangular factor */
1062:   matrixType = cusparseGetMatType(loTriFactor->descr);
1063:   indexBase  = cusparseGetMatIndexBase(loTriFactor->descr);
1064:   fillMode   = cusparseGetMatFillMode(loTriFactor->descr) == CUSPARSE_FILL_MODE_UPPER ? CUSPARSE_FILL_MODE_LOWER : CUSPARSE_FILL_MODE_UPPER;
1065:   diagType   = cusparseGetMatDiagType(loTriFactor->descr);

1067:   /* Create the matrix description */
1068:   PetscCallCUSPARSE(cusparseCreateMatDescr(&loTriFactorT->descr));
1069:   PetscCallCUSPARSE(cusparseSetMatIndexBase(loTriFactorT->descr, indexBase));
1070:   PetscCallCUSPARSE(cusparseSetMatType(loTriFactorT->descr, matrixType));
1071:   PetscCallCUSPARSE(cusparseSetMatFillMode(loTriFactorT->descr, fillMode));
1072:   PetscCallCUSPARSE(cusparseSetMatDiagType(loTriFactorT->descr, diagType));

1074:   /* set the operation */
1075:   loTriFactorT->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

1077:   /* allocate GPU space for the CSC of the lower triangular factor*/
1078:   loTriFactorT->csrMat                 = new CsrMatrix;
1079:   loTriFactorT->csrMat->num_rows       = loTriFactor->csrMat->num_cols;
1080:   loTriFactorT->csrMat->num_cols       = loTriFactor->csrMat->num_rows;
1081:   loTriFactorT->csrMat->num_entries    = loTriFactor->csrMat->num_entries;
1082:   loTriFactorT->csrMat->row_offsets    = new THRUSTINTARRAY32(loTriFactorT->csrMat->num_rows + 1);
1083:   loTriFactorT->csrMat->column_indices = new THRUSTINTARRAY32(loTriFactorT->csrMat->num_entries);
1084:   loTriFactorT->csrMat->values         = new THRUSTARRAY(loTriFactorT->csrMat->num_entries);

1086:   /* compute the transpose of the lower triangular factor, i.e. the CSC */
1087:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1088:   PetscCallCUSPARSE(cusparseCsr2cscEx2_bufferSize(cusparseTriFactors->handle, loTriFactor->csrMat->num_rows, loTriFactor->csrMat->num_cols, loTriFactor->csrMat->num_entries, loTriFactor->csrMat->values->data().get(),
1089:                                                   loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactorT->csrMat->values->data().get(), loTriFactorT->csrMat->row_offsets->data().get(),
1090:                                                   loTriFactorT->csrMat->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, CUSPARSE_CSR2CSC_ALG1, &loTriFactor->csr2cscBufferSize));
1091:   PetscCallCUDA(cudaMalloc(&loTriFactor->csr2cscBuffer, loTriFactor->csr2cscBufferSize));
1092:   #endif

1094:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));
1095:   {
1096:     // there is no clean way to have PetscCallCUSPARSE wrapping this function...
1097:     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(),
1098:                                  loTriFactor->csrMat->column_indices->data().get(), loTriFactorT->csrMat->values->data().get(),
1099:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1100:                                  loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, CUSPARSE_CSR2CSC_ALG1, loTriFactor->csr2cscBuffer);
1101:   #else
1102:                                  loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->csrMat->row_offsets->data().get(), CUSPARSE_ACTION_NUMERIC, indexBase);
1103:   #endif
1104:     PetscCallCUSPARSE(stat);
1105:   }

1107:   PetscCallCUDA(WaitForCUDA());
1108:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));

1110:   /* Create the solve analysis information */
1111:   PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
1112:   PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&loTriFactorT->solveInfo));
1113:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
1114:   PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, loTriFactorT->solveOp, loTriFactorT->csrMat->num_rows, loTriFactorT->csrMat->num_entries, loTriFactorT->descr, loTriFactorT->csrMat->values->data().get(),
1115:                                             loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->solveInfo, &loTriFactorT->solveBufferSize));
1116:   PetscCallCUDA(cudaMalloc(&loTriFactorT->solveBuffer, loTriFactorT->solveBufferSize));
1117:   #endif

1119:   /* perform the solve analysis */
1120:   PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, loTriFactorT->solveOp, loTriFactorT->csrMat->num_rows, loTriFactorT->csrMat->num_entries, loTriFactorT->descr, loTriFactorT->csrMat->values->data().get(),
1121:                                             loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->solveInfo, loTriFactorT->solvePolicy, loTriFactorT->solveBuffer));

1123:   PetscCallCUDA(WaitForCUDA());
1124:   PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

1126:   /* assign the pointer */
1127:   ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->loTriFactorPtrTranspose = loTriFactorT;

1129:   /*********************************************/
1130:   /* Now the Transpose of the Upper Tri Factor */
1131:   /*********************************************/

1133:   /* allocate space for the transpose of the upper triangular factor */
1134:   PetscCall(PetscNew(&upTriFactorT));
1135:   upTriFactorT->solvePolicy = CUSPARSE_SOLVE_POLICY_USE_LEVEL;

1137:   /* set the matrix descriptors of the upper triangular factor */
1138:   matrixType = cusparseGetMatType(upTriFactor->descr);
1139:   indexBase  = cusparseGetMatIndexBase(upTriFactor->descr);
1140:   fillMode   = cusparseGetMatFillMode(upTriFactor->descr) == CUSPARSE_FILL_MODE_UPPER ? CUSPARSE_FILL_MODE_LOWER : CUSPARSE_FILL_MODE_UPPER;
1141:   diagType   = cusparseGetMatDiagType(upTriFactor->descr);

1143:   /* Create the matrix description */
1144:   PetscCallCUSPARSE(cusparseCreateMatDescr(&upTriFactorT->descr));
1145:   PetscCallCUSPARSE(cusparseSetMatIndexBase(upTriFactorT->descr, indexBase));
1146:   PetscCallCUSPARSE(cusparseSetMatType(upTriFactorT->descr, matrixType));
1147:   PetscCallCUSPARSE(cusparseSetMatFillMode(upTriFactorT->descr, fillMode));
1148:   PetscCallCUSPARSE(cusparseSetMatDiagType(upTriFactorT->descr, diagType));

1150:   /* set the operation */
1151:   upTriFactorT->solveOp = CUSPARSE_OPERATION_NON_TRANSPOSE;

1153:   /* allocate GPU space for the CSC of the upper triangular factor*/
1154:   upTriFactorT->csrMat                 = new CsrMatrix;
1155:   upTriFactorT->csrMat->num_rows       = upTriFactor->csrMat->num_cols;
1156:   upTriFactorT->csrMat->num_cols       = upTriFactor->csrMat->num_rows;
1157:   upTriFactorT->csrMat->num_entries    = upTriFactor->csrMat->num_entries;
1158:   upTriFactorT->csrMat->row_offsets    = new THRUSTINTARRAY32(upTriFactorT->csrMat->num_rows + 1);
1159:   upTriFactorT->csrMat->column_indices = new THRUSTINTARRAY32(upTriFactorT->csrMat->num_entries);
1160:   upTriFactorT->csrMat->values         = new THRUSTARRAY(upTriFactorT->csrMat->num_entries);

1162:   /* compute the transpose of the upper triangular factor, i.e. the CSC */
1163:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1164:   PetscCallCUSPARSE(cusparseCsr2cscEx2_bufferSize(cusparseTriFactors->handle, upTriFactor->csrMat->num_rows, upTriFactor->csrMat->num_cols, upTriFactor->csrMat->num_entries, upTriFactor->csrMat->values->data().get(),
1165:                                                   upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactorT->csrMat->values->data().get(), upTriFactorT->csrMat->row_offsets->data().get(),
1166:                                                   upTriFactorT->csrMat->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, CUSPARSE_CSR2CSC_ALG1, &upTriFactor->csr2cscBufferSize));
1167:   PetscCallCUDA(cudaMalloc(&upTriFactor->csr2cscBuffer, upTriFactor->csr2cscBufferSize));
1168:   #endif

1170:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));
1171:   {
1172:     // there is no clean way to have PetscCallCUSPARSE wrapping this function...
1173:     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(),
1174:                                  upTriFactor->csrMat->column_indices->data().get(), upTriFactorT->csrMat->values->data().get(),
1175:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1176:                                  upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, CUSPARSE_CSR2CSC_ALG1, upTriFactor->csr2cscBuffer);
1177:   #else
1178:                                  upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->csrMat->row_offsets->data().get(), CUSPARSE_ACTION_NUMERIC, indexBase);
1179:   #endif
1180:     PetscCallCUSPARSE(stat);
1181:   }

1183:   PetscCallCUDA(WaitForCUDA());
1184:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));

1186:   /* Create the solve analysis information */
1187:   PetscCall(PetscLogEventBegin(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));
1188:   PetscCallCUSPARSE(cusparseCreateCsrsvInfo(&upTriFactorT->solveInfo));
1189:   #if PETSC_PKG_CUDA_VERSION_GE(9, 0, 0)
1190:   PetscCallCUSPARSE(cusparseXcsrsv_buffsize(cusparseTriFactors->handle, upTriFactorT->solveOp, upTriFactorT->csrMat->num_rows, upTriFactorT->csrMat->num_entries, upTriFactorT->descr, upTriFactorT->csrMat->values->data().get(),
1191:                                             upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->solveInfo, &upTriFactorT->solveBufferSize));
1192:   PetscCallCUDA(cudaMalloc(&upTriFactorT->solveBuffer, upTriFactorT->solveBufferSize));
1193:   #endif

1195:   /* perform the solve analysis */
1196:   /* christ, would it have killed you to put this stuff in a function????????? */
1197:   PetscCallCUSPARSE(cusparseXcsrsv_analysis(cusparseTriFactors->handle, upTriFactorT->solveOp, upTriFactorT->csrMat->num_rows, upTriFactorT->csrMat->num_entries, upTriFactorT->descr, upTriFactorT->csrMat->values->data().get(),
1198:                                             upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->solveInfo, upTriFactorT->solvePolicy, upTriFactorT->solveBuffer));

1200:   PetscCallCUDA(WaitForCUDA());
1201:   PetscCall(PetscLogEventEnd(MAT_CUSPARSESolveAnalysis, A, 0, 0, 0));

1203:   /* assign the pointer */
1204:   ((Mat_SeqAIJCUSPARSETriFactors *)A->spptr)->upTriFactorPtrTranspose = upTriFactorT;
1205:   PetscFunctionReturn(PETSC_SUCCESS);
1206: }
1207: #endif

1209: struct PetscScalarToPetscInt {
1210:   __host__ __device__ PetscInt operator()(PetscScalar s) { return (PetscInt)PetscRealPart(s); }
1211: };

1213: static PetscErrorCode MatSeqAIJCUSPARSEFormExplicitTranspose(Mat A)
1214: {
1215:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
1216:   Mat_SeqAIJCUSPARSEMultStruct *matstruct, *matstructT;
1217:   Mat_SeqAIJ                   *a = (Mat_SeqAIJ *)A->data;
1218:   cusparseStatus_t              stat;
1219:   cusparseIndexBase_t           indexBase;

1221:   PetscFunctionBegin;
1222:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
1223:   matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
1224:   PetscCheck(matstruct, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing mat struct");
1225:   matstructT = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->matTranspose;
1226:   PetscCheck(!A->transupdated || matstructT, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing matTranspose struct");
1227:   if (A->transupdated) PetscFunctionReturn(PETSC_SUCCESS);
1228:   PetscCall(PetscLogEventBegin(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));
1229:   PetscCall(PetscLogGpuTimeBegin());
1230:   if (cusparsestruct->format != MAT_CUSPARSE_CSR) PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_TRUE));
1231:   if (!cusparsestruct->matTranspose) { /* create cusparse matrix */
1232:     matstructT = new Mat_SeqAIJCUSPARSEMultStruct;
1233:     PetscCallCUSPARSE(cusparseCreateMatDescr(&matstructT->descr));
1234:     indexBase = cusparseGetMatIndexBase(matstruct->descr);
1235:     PetscCallCUSPARSE(cusparseSetMatIndexBase(matstructT->descr, indexBase));
1236:     PetscCallCUSPARSE(cusparseSetMatType(matstructT->descr, CUSPARSE_MATRIX_TYPE_GENERAL));

1238:     /* set alpha and beta */
1239:     PetscCallCUDA(cudaMalloc((void **)&matstructT->alpha_one, sizeof(PetscScalar)));
1240:     PetscCallCUDA(cudaMalloc((void **)&matstructT->beta_zero, sizeof(PetscScalar)));
1241:     PetscCallCUDA(cudaMalloc((void **)&matstructT->beta_one, sizeof(PetscScalar)));
1242:     PetscCallCUDA(cudaMemcpy(matstructT->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
1243:     PetscCallCUDA(cudaMemcpy(matstructT->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
1244:     PetscCallCUDA(cudaMemcpy(matstructT->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));

1246:     if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
1247:       CsrMatrix *matrixT      = new CsrMatrix;
1248:       matstructT->mat         = matrixT;
1249:       matrixT->num_rows       = A->cmap->n;
1250:       matrixT->num_cols       = A->rmap->n;
1251:       matrixT->num_entries    = a->nz;
1252:       matrixT->row_offsets    = new THRUSTINTARRAY32(matrixT->num_rows + 1);
1253:       matrixT->column_indices = new THRUSTINTARRAY32(a->nz);
1254:       matrixT->values         = new THRUSTARRAY(a->nz);

1256:       if (!cusparsestruct->rowoffsets_gpu) cusparsestruct->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
1257:       cusparsestruct->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);

1259: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1260:   #if PETSC_PKG_CUDA_VERSION_GE(11, 2, 1)
1261:       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 */
1262:                                indexBase, cusparse_scalartype);
1263:       PetscCallCUSPARSE(stat);
1264:   #else
1265:       /* cusparse-11.x returns errors with zero-sized matrices until 11.2.1,
1266:            see https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#cusparse-11.2.1

1268:            I don't know what a proper value should be for matstructT->matDescr with empty matrices, so I just set
1269:            it to NULL to blow it up if one relies on it. Per https://docs.nvidia.com/cuda/cusparse/index.html#csr2cscEx2,
1270:            when nnz = 0, matrixT->row_offsets[] should be filled with indexBase. So I also set it accordingly.
1271:         */
1272:       if (matrixT->num_entries) {
1273:         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);
1274:         PetscCallCUSPARSE(stat);

1276:       } else {
1277:         matstructT->matDescr = NULL;
1278:         matrixT->row_offsets->assign(matrixT->row_offsets->size(), indexBase);
1279:       }
1280:   #endif
1281: #endif
1282:     } else if (cusparsestruct->format == MAT_CUSPARSE_ELL || cusparsestruct->format == MAT_CUSPARSE_HYB) {
1283: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1284:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
1285: #else
1286:       CsrMatrix *temp  = new CsrMatrix;
1287:       CsrMatrix *tempT = new CsrMatrix;
1288:       /* First convert HYB to CSR */
1289:       temp->num_rows       = A->rmap->n;
1290:       temp->num_cols       = A->cmap->n;
1291:       temp->num_entries    = a->nz;
1292:       temp->row_offsets    = new THRUSTINTARRAY32(A->rmap->n + 1);
1293:       temp->column_indices = new THRUSTINTARRAY32(a->nz);
1294:       temp->values         = new THRUSTARRAY(a->nz);

1296:       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());
1297:       PetscCallCUSPARSE(stat);

1299:       /* Next, convert CSR to CSC (i.e. the matrix transpose) */
1300:       tempT->num_rows       = A->rmap->n;
1301:       tempT->num_cols       = A->cmap->n;
1302:       tempT->num_entries    = a->nz;
1303:       tempT->row_offsets    = new THRUSTINTARRAY32(A->rmap->n + 1);
1304:       tempT->column_indices = new THRUSTINTARRAY32(a->nz);
1305:       tempT->values         = new THRUSTARRAY(a->nz);

1307:       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(),
1308:                               tempT->column_indices->data().get(), tempT->row_offsets->data().get(), CUSPARSE_ACTION_NUMERIC, indexBase);
1309:       PetscCallCUSPARSE(stat);

1311:       /* Last, convert CSC to HYB */
1312:       cusparseHybMat_t hybMat;
1313:       PetscCallCUSPARSE(cusparseCreateHybMat(&hybMat));
1314:       cusparseHybPartition_t partition = cusparsestruct->format == MAT_CUSPARSE_ELL ? CUSPARSE_HYB_PARTITION_MAX : CUSPARSE_HYB_PARTITION_AUTO;
1315:       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);
1316:       PetscCallCUSPARSE(stat);

1318:       /* assign the pointer */
1319:       matstructT->mat = hybMat;
1320:       A->transupdated = PETSC_TRUE;
1321:       /* delete temporaries */
1322:       if (tempT) {
1323:         if (tempT->values) delete (THRUSTARRAY *)tempT->values;
1324:         if (tempT->column_indices) delete (THRUSTINTARRAY32 *)tempT->column_indices;
1325:         if (tempT->row_offsets) delete (THRUSTINTARRAY32 *)tempT->row_offsets;
1326:         delete (CsrMatrix *)tempT;
1327:       }
1328:       if (temp) {
1329:         if (temp->values) delete (THRUSTARRAY *)temp->values;
1330:         if (temp->column_indices) delete (THRUSTINTARRAY32 *)temp->column_indices;
1331:         if (temp->row_offsets) delete (THRUSTINTARRAY32 *)temp->row_offsets;
1332:         delete (CsrMatrix *)temp;
1333:       }
1334: #endif
1335:     }
1336:   }
1337:   if (cusparsestruct->format == MAT_CUSPARSE_CSR) { /* transpose mat struct may be already present, update data */
1338:     CsrMatrix *matrix  = (CsrMatrix *)matstruct->mat;
1339:     CsrMatrix *matrixT = (CsrMatrix *)matstructT->mat;
1340:     PetscCheck(matrix, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrix");
1341:     PetscCheck(matrix->row_offsets, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrix rows");
1342:     PetscCheck(matrix->column_indices, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrix cols");
1343:     PetscCheck(matrix->values, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrix values");
1344:     PetscCheck(matrixT, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrixT");
1345:     PetscCheck(matrixT->row_offsets, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrixT rows");
1346:     PetscCheck(matrixT->column_indices, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrixT cols");
1347:     PetscCheck(matrixT->values, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CsrMatrixT values");
1348:     if (!cusparsestruct->rowoffsets_gpu) { /* this may be absent when we did not construct the transpose with csr2csc */
1349:       cusparsestruct->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
1350:       cusparsestruct->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);
1351:       PetscCall(PetscLogCpuToGpu((A->rmap->n + 1) * sizeof(PetscInt)));
1352:     }
1353:     if (!cusparsestruct->csr2csc_i) {
1354:       THRUSTARRAY csr2csc_a(matrix->num_entries);
1355:       PetscCallThrust(thrust::sequence(thrust::device, csr2csc_a.begin(), csr2csc_a.end(), 0.0));

1357:       indexBase = cusparseGetMatIndexBase(matstruct->descr);
1358: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1359:       void  *csr2cscBuffer;
1360:       size_t csr2cscBufferSize;
1361:       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(),
1362:                                            matrixT->row_offsets->data().get(), matrixT->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, cusparsestruct->csr2cscAlg, &csr2cscBufferSize);
1363:       PetscCallCUSPARSE(stat);
1364:       PetscCallCUDA(cudaMalloc(&csr2cscBuffer, csr2cscBufferSize));
1365: #endif

1367:       if (matrix->num_entries) {
1368:         /* When there are no nonzeros, this routine mistakenly returns CUSPARSE_STATUS_INVALID_VALUE in
1369:            mat_tests-ex62_15_mpiaijcusparse on ranks 0 and 2 with CUDA-11. But CUDA-10 is OK.
1370:            I checked every parameters and they were just fine. I have no clue why cusparse complains.

1372:            Per https://docs.nvidia.com/cuda/cusparse/index.html#csr2cscEx2, when nnz = 0, matrixT->row_offsets[]
1373:            should be filled with indexBase. So I just take a shortcut here.
1374:         */
1375:         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(),
1376: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1377:                                 matrixT->row_offsets->data().get(), matrixT->column_indices->data().get(), cusparse_scalartype, CUSPARSE_ACTION_NUMERIC, indexBase, cusparsestruct->csr2cscAlg, csr2cscBuffer);
1378:         PetscCallCUSPARSE(stat);
1379: #else
1380:                                 matrixT->column_indices->data().get(), matrixT->row_offsets->data().get(), CUSPARSE_ACTION_NUMERIC, indexBase);
1381:         PetscCallCUSPARSE(stat);
1382: #endif
1383:       } else {
1384:         matrixT->row_offsets->assign(matrixT->row_offsets->size(), indexBase);
1385:       }

1387:       cusparsestruct->csr2csc_i = new THRUSTINTARRAY(matrix->num_entries);
1388:       PetscCallThrust(thrust::transform(thrust::device, matrixT->values->begin(), matrixT->values->end(), cusparsestruct->csr2csc_i->begin(), PetscScalarToPetscInt()));
1389: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
1390:       PetscCallCUDA(cudaFree(csr2cscBuffer));
1391: #endif
1392:     }
1393:     PetscCallThrust(
1394:       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()));
1395:   }
1396:   PetscCall(PetscLogGpuTimeEnd());
1397:   PetscCall(PetscLogEventEnd(MAT_CUSPARSEGenerateTranspose, A, 0, 0, 0));
1398:   /* the compressed row indices is not used for matTranspose */
1399:   matstructT->cprowIndices = NULL;
1400:   /* assign the pointer */
1401:   ((Mat_SeqAIJCUSPARSE *)A->spptr)->matTranspose = matstructT;
1402:   A->transupdated                                = PETSC_TRUE;
1403:   PetscFunctionReturn(PETSC_SUCCESS);
1404: }

1406: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
1407: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_LU(Mat A, Vec b, Vec x)
1408: {
1409:   const PetscScalar                    *barray;
1410:   PetscScalar                          *xarray;
1411:   thrust::device_ptr<const PetscScalar> bGPU;
1412:   thrust::device_ptr<PetscScalar>       xGPU;
1413:   Mat_SeqAIJCUSPARSETriFactors         *fs  = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
1414:   const Mat_SeqAIJ                     *aij = static_cast<Mat_SeqAIJ *>(A->data);
1415:   const cusparseOperation_t             op  = CUSPARSE_OPERATION_NON_TRANSPOSE;
1416:   const cusparseSpSVAlg_t               alg = CUSPARSE_SPSV_ALG_DEFAULT;
1417:   PetscInt                              m   = A->rmap->n;

1419:   PetscFunctionBegin;
1420:   PetscCall(PetscLogGpuTimeBegin());
1421:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
1422:   PetscCall(VecCUDAGetArrayRead(b, &barray));
1423:   xGPU = thrust::device_pointer_cast(xarray);
1424:   bGPU = thrust::device_pointer_cast(barray);

1426:   // Reorder b with the row permutation if needed, and wrap the result in fs->X
1427:   if (fs->rpermIndices) {
1428:     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)));
1429:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
1430:   } else {
1431:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
1432:   }

1434:   // Solve L Y = X
1435:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
1436:   // Note that cusparseSpSV_solve() secretly uses the external buffer used in cusparseSpSV_analysis()!
1437:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, op, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_L));

1439:   // Solve U X = Y
1440:   if (fs->cpermIndices) {
1441:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
1442:   } else {
1443:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
1444:   }
1445:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, op, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_Y, fs->dnVecDescr_X, cusparse_scalartype, alg, fs->spsvDescr_U));

1447:   // Reorder X with the column permutation if needed, and put the result back to x
1448:   if (fs->cpermIndices) {
1449:     PetscCallThrust(thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X), fs->cpermIndices->begin()),
1450:                                  thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X + m), fs->cpermIndices->end()), xGPU));
1451:   }
1452:   PetscCall(VecCUDARestoreArrayRead(b, &barray));
1453:   PetscCall(VecCUDARestoreArrayWrite(x, &xarray));
1454:   PetscCall(PetscLogGpuTimeEnd());
1455:   PetscCall(PetscLogGpuFlops(2.0 * aij->nz - m));
1456:   PetscFunctionReturn(PETSC_SUCCESS);
1457: }

1459: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE_LU(Mat A, Vec b, Vec x)
1460: {
1461:   Mat_SeqAIJCUSPARSETriFactors         *fs  = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(A->spptr);
1462:   Mat_SeqAIJ                           *aij = static_cast<Mat_SeqAIJ *>(A->data);
1463:   const PetscScalar                    *barray;
1464:   PetscScalar                          *xarray;
1465:   thrust::device_ptr<const PetscScalar> bGPU;
1466:   thrust::device_ptr<PetscScalar>       xGPU;
1467:   const cusparseOperation_t             opA = CUSPARSE_OPERATION_TRANSPOSE;
1468:   const cusparseSpSVAlg_t               alg = CUSPARSE_SPSV_ALG_DEFAULT;
1469:   PetscInt                              m   = A->rmap->n;

1471:   PetscFunctionBegin;
1472:   PetscCall(PetscLogGpuTimeBegin());
1473:   if (!fs->createdTransposeSpSVDescr) { // Call MatSolveTranspose() for the first time
1474:     PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Lt));
1475:     PetscCallCUSPARSE(cusparseSpSV_bufferSize(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, /* The matrix is still L. We only do transpose solve with it */
1476:                                               fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_Lt, &fs->spsvBufferSize_Lt));

1478:     PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Ut));
1479:     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));
1480:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Lt, fs->spsvBufferSize_Lt));
1481:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Ut, fs->spsvBufferSize_Ut));
1482:     fs->createdTransposeSpSVDescr = PETSC_TRUE;
1483:   }

1485:   if (!fs->updatedTransposeSpSVAnalysis) {
1486:     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));

1488:     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));
1489:     fs->updatedTransposeSpSVAnalysis = PETSC_TRUE;
1490:   }

1492:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
1493:   PetscCall(VecCUDAGetArrayRead(b, &barray));
1494:   xGPU = thrust::device_pointer_cast(xarray);
1495:   bGPU = thrust::device_pointer_cast(barray);

1497:   // Reorder b with the row permutation if needed, and wrap the result in fs->X
1498:   if (fs->rpermIndices) {
1499:     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)));
1500:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
1501:   } else {
1502:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
1503:   }

1505:   // Solve Ut Y = X
1506:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
1507:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_U, fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, alg, fs->spsvDescr_Ut));

1509:   // Solve Lt X = Y
1510:   if (fs->cpermIndices) { // if need to permute, we need to use the intermediate buffer X
1511:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, fs->X));
1512:   } else {
1513:     PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
1514:   }
1515:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, opA, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, fs->dnVecDescr_Y, fs->dnVecDescr_X, cusparse_scalartype, alg, fs->spsvDescr_Lt));

1517:   // Reorder X with the column permutation if needed, and put the result back to x
1518:   if (fs->cpermIndices) {
1519:     PetscCallThrust(thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X), fs->cpermIndices->begin()),
1520:                                  thrust::make_permutation_iterator(thrust::device_pointer_cast(fs->X + m), fs->cpermIndices->end()), xGPU));
1521:   }

1523:   PetscCall(VecCUDARestoreArrayRead(b, &barray));
1524:   PetscCall(VecCUDARestoreArrayWrite(x, &xarray));
1525:   PetscCall(PetscLogGpuTimeEnd());
1526:   PetscCall(PetscLogGpuFlops(2.0 * aij->nz - A->rmap->n));
1527:   PetscFunctionReturn(PETSC_SUCCESS);
1528: }
1529: #else
1530: /* Why do we need to analyze the transposed matrix again? Can't we just use op(A) = CUSPARSE_OPERATION_TRANSPOSE in MatSolve_SeqAIJCUSPARSE? */
1531: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE(Mat A, Vec bb, Vec xx)
1532: {
1533:   PetscInt                              n = xx->map->n;
1534:   const PetscScalar                    *barray;
1535:   PetscScalar                          *xarray;
1536:   thrust::device_ptr<const PetscScalar> bGPU;
1537:   thrust::device_ptr<PetscScalar>       xGPU;
1538:   Mat_SeqAIJCUSPARSETriFactors         *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1539:   Mat_SeqAIJCUSPARSETriFactorStruct    *loTriFactorT       = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtrTranspose;
1540:   Mat_SeqAIJCUSPARSETriFactorStruct    *upTriFactorT       = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtrTranspose;
1541:   THRUSTARRAY                          *tempGPU            = (THRUSTARRAY *)cusparseTriFactors->workVector;

1543:   PetscFunctionBegin;
1544:   /* Analyze the matrix and create the transpose ... on the fly */
1545:   if (!loTriFactorT && !upTriFactorT) {
1546:     PetscCall(MatSeqAIJCUSPARSEAnalyzeTransposeForSolve(A));
1547:     loTriFactorT = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtrTranspose;
1548:     upTriFactorT = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtrTranspose;
1549:   }

1551:   /* Get the GPU pointers */
1552:   PetscCall(VecCUDAGetArrayWrite(xx, &xarray));
1553:   PetscCall(VecCUDAGetArrayRead(bb, &barray));
1554:   xGPU = thrust::device_pointer_cast(xarray);
1555:   bGPU = thrust::device_pointer_cast(barray);

1557:   PetscCall(PetscLogGpuTimeBegin());
1558:   /* First, reorder with the row permutation */
1559:   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);

1561:   /* First, solve U */
1562:   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(),
1563:                                          upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->solveInfo, xarray, tempGPU->data().get(), upTriFactorT->solvePolicy, upTriFactorT->solveBuffer));

1565:   /* Then, solve L */
1566:   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(),
1567:                                          loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->solveInfo, tempGPU->data().get(), xarray, loTriFactorT->solvePolicy, loTriFactorT->solveBuffer));

1569:   /* Last, copy the solution, xGPU, into a temporary with the column permutation ... can't be done in place. */
1570:   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());

1572:   /* Copy the temporary to the full solution. */
1573:   thrust::copy(thrust::cuda::par.on(PetscDefaultCudaStream), tempGPU->begin(), tempGPU->end(), xGPU);

1575:   /* restore */
1576:   PetscCall(VecCUDARestoreArrayRead(bb, &barray));
1577:   PetscCall(VecCUDARestoreArrayWrite(xx, &xarray));
1578:   PetscCall(PetscLogGpuTimeEnd());
1579:   PetscCall(PetscLogGpuFlops(2.0 * cusparseTriFactors->nnz - A->cmap->n));
1580:   PetscFunctionReturn(PETSC_SUCCESS);
1581: }

1583: static PetscErrorCode MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering(Mat A, Vec bb, Vec xx)
1584: {
1585:   const PetscScalar                 *barray;
1586:   PetscScalar                       *xarray;
1587:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1588:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactorT       = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtrTranspose;
1589:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactorT       = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtrTranspose;
1590:   THRUSTARRAY                       *tempGPU            = (THRUSTARRAY *)cusparseTriFactors->workVector;

1592:   PetscFunctionBegin;
1593:   /* Analyze the matrix and create the transpose ... on the fly */
1594:   if (!loTriFactorT && !upTriFactorT) {
1595:     PetscCall(MatSeqAIJCUSPARSEAnalyzeTransposeForSolve(A));
1596:     loTriFactorT = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtrTranspose;
1597:     upTriFactorT = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtrTranspose;
1598:   }

1600:   /* Get the GPU pointers */
1601:   PetscCall(VecCUDAGetArrayWrite(xx, &xarray));
1602:   PetscCall(VecCUDAGetArrayRead(bb, &barray));

1604:   PetscCall(PetscLogGpuTimeBegin());
1605:   /* First, solve U */
1606:   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(),
1607:                                          upTriFactorT->csrMat->row_offsets->data().get(), upTriFactorT->csrMat->column_indices->data().get(), upTriFactorT->solveInfo, barray, tempGPU->data().get(), upTriFactorT->solvePolicy, upTriFactorT->solveBuffer));

1609:   /* Then, solve L */
1610:   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(),
1611:                                          loTriFactorT->csrMat->row_offsets->data().get(), loTriFactorT->csrMat->column_indices->data().get(), loTriFactorT->solveInfo, tempGPU->data().get(), xarray, loTriFactorT->solvePolicy, loTriFactorT->solveBuffer));

1613:   /* restore */
1614:   PetscCall(VecCUDARestoreArrayRead(bb, &barray));
1615:   PetscCall(VecCUDARestoreArrayWrite(xx, &xarray));
1616:   PetscCall(PetscLogGpuTimeEnd());
1617:   PetscCall(PetscLogGpuFlops(2.0 * cusparseTriFactors->nnz - A->cmap->n));
1618:   PetscFunctionReturn(PETSC_SUCCESS);
1619: }

1621: static PetscErrorCode MatSolve_SeqAIJCUSPARSE(Mat A, Vec bb, Vec xx)
1622: {
1623:   const PetscScalar                    *barray;
1624:   PetscScalar                          *xarray;
1625:   thrust::device_ptr<const PetscScalar> bGPU;
1626:   thrust::device_ptr<PetscScalar>       xGPU;
1627:   Mat_SeqAIJCUSPARSETriFactors         *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1628:   Mat_SeqAIJCUSPARSETriFactorStruct    *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
1629:   Mat_SeqAIJCUSPARSETriFactorStruct    *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
1630:   THRUSTARRAY                          *tempGPU            = (THRUSTARRAY *)cusparseTriFactors->workVector;

1632:   PetscFunctionBegin;
1633:   /* Get the GPU pointers */
1634:   PetscCall(VecCUDAGetArrayWrite(xx, &xarray));
1635:   PetscCall(VecCUDAGetArrayRead(bb, &barray));
1636:   xGPU = thrust::device_pointer_cast(xarray);
1637:   bGPU = thrust::device_pointer_cast(barray);

1639:   PetscCall(PetscLogGpuTimeBegin());
1640:   /* First, reorder with the row permutation */
1641:   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());

1643:   /* Next, solve L */
1644:   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(),
1645:                                          loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, tempGPU->data().get(), xarray, loTriFactor->solvePolicy, loTriFactor->solveBuffer));

1647:   /* Then, solve U */
1648:   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(),
1649:                                          upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, xarray, tempGPU->data().get(), upTriFactor->solvePolicy, upTriFactor->solveBuffer));

1651:   /* Last, reorder with the column permutation */
1652:   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);

1654:   PetscCall(VecCUDARestoreArrayRead(bb, &barray));
1655:   PetscCall(VecCUDARestoreArrayWrite(xx, &xarray));
1656:   PetscCall(PetscLogGpuTimeEnd());
1657:   PetscCall(PetscLogGpuFlops(2.0 * cusparseTriFactors->nnz - A->cmap->n));
1658:   PetscFunctionReturn(PETSC_SUCCESS);
1659: }

1661: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_NaturalOrdering(Mat A, Vec bb, Vec xx)
1662: {
1663:   const PetscScalar                 *barray;
1664:   PetscScalar                       *xarray;
1665:   Mat_SeqAIJCUSPARSETriFactors      *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
1666:   Mat_SeqAIJCUSPARSETriFactorStruct *loTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->loTriFactorPtr;
1667:   Mat_SeqAIJCUSPARSETriFactorStruct *upTriFactor        = (Mat_SeqAIJCUSPARSETriFactorStruct *)cusparseTriFactors->upTriFactorPtr;
1668:   THRUSTARRAY                       *tempGPU            = (THRUSTARRAY *)cusparseTriFactors->workVector;

1670:   PetscFunctionBegin;
1671:   /* Get the GPU pointers */
1672:   PetscCall(VecCUDAGetArrayWrite(xx, &xarray));
1673:   PetscCall(VecCUDAGetArrayRead(bb, &barray));

1675:   PetscCall(PetscLogGpuTimeBegin());
1676:   /* First, solve L */
1677:   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(),
1678:                                          loTriFactor->csrMat->row_offsets->data().get(), loTriFactor->csrMat->column_indices->data().get(), loTriFactor->solveInfo, barray, tempGPU->data().get(), loTriFactor->solvePolicy, loTriFactor->solveBuffer));

1680:   /* Next, solve U */
1681:   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(),
1682:                                          upTriFactor->csrMat->row_offsets->data().get(), upTriFactor->csrMat->column_indices->data().get(), upTriFactor->solveInfo, tempGPU->data().get(), xarray, upTriFactor->solvePolicy, upTriFactor->solveBuffer));

1684:   PetscCall(VecCUDARestoreArrayRead(bb, &barray));
1685:   PetscCall(VecCUDARestoreArrayWrite(xx, &xarray));
1686:   PetscCall(PetscLogGpuTimeEnd());
1687:   PetscCall(PetscLogGpuFlops(2.0 * cusparseTriFactors->nnz - A->cmap->n));
1688:   PetscFunctionReturn(PETSC_SUCCESS);
1689: }
1690: #endif

1692: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
1693: static PetscErrorCode MatILUFactorNumeric_SeqAIJCUSPARSE_ILU0(Mat fact, Mat A, const MatFactorInfo *)
1694: {
1695:   Mat_SeqAIJCUSPARSETriFactors *fs    = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1696:   Mat_SeqAIJ                   *aij   = (Mat_SeqAIJ *)fact->data;
1697:   Mat_SeqAIJCUSPARSE           *Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
1698:   CsrMatrix                    *Acsr;
1699:   PetscInt                      m, nz;
1700:   PetscBool                     flg;

1702:   PetscFunctionBegin;
1703:   if (PetscDefined(USE_DEBUG)) {
1704:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
1705:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
1706:   }

1708:   /* Copy A's value to fact */
1709:   m  = fact->rmap->n;
1710:   nz = aij->nz;
1711:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
1712:   Acsr = (CsrMatrix *)Acusp->mat->mat;
1713:   PetscCallCUDA(cudaMemcpyAsync(fs->csrVal, Acsr->values->data().get(), sizeof(PetscScalar) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

1715:   PetscCall(PetscLogGpuTimeBegin());
1716:   /* Factorize fact inplace */
1717:   if (m)
1718:     PetscCallCUSPARSE(cusparseXcsrilu02(fs->handle, m, nz, /* cusparseXcsrilu02 errors out with empty matrices (m=0) */
1719:                                         fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ilu0Info_M, fs->policy_M, fs->factBuffer_M));
1720:   if (PetscDefined(USE_DEBUG)) {
1721:     int              numerical_zero;
1722:     cusparseStatus_t status;
1723:     status = cusparseXcsrilu02_zeroPivot(fs->handle, fs->ilu0Info_M, &numerical_zero);
1724:     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);
1725:   }

1727:   #if PETSC_PKG_CUDA_VERSION_GE(12, 1, 1)
1728:   if (fs->updatedSpSVAnalysis) {
1729:     if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_L, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
1730:     if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_U, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
1731:   } else
1732:   #endif
1733:   {
1734:     /* cusparseSpSV_analysis() is numeric, i.e., it requires valid matrix values, therefore, we do it after cusparseXcsrilu02()
1735:      See discussion at https://github.com/NVIDIA/CUDALibrarySamples/issues/78
1736:     */
1737:     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));

1739:     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));

1741:     fs->updatedSpSVAnalysis = PETSC_TRUE;
1742:     /* L, U values have changed, reset the flag to indicate we need to redo cusparseSpSV_analysis() for transpose solve */
1743:     fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;
1744:   }

1746:   fact->offloadmask            = PETSC_OFFLOAD_GPU;
1747:   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.
1748:   fact->ops->solvetranspose    = MatSolveTranspose_SeqAIJCUSPARSE_LU;
1749:   fact->ops->matsolve          = NULL;
1750:   fact->ops->matsolvetranspose = NULL;
1751:   PetscCall(PetscLogGpuTimeEnd());
1752:   PetscCall(PetscLogGpuFlops(fs->numericFactFlops));
1753:   PetscFunctionReturn(PETSC_SUCCESS);
1754: }

1756: static PetscErrorCode MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0(Mat fact, Mat A, IS, IS, const MatFactorInfo *info)
1757: {
1758:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1759:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
1760:   PetscInt                      m, nz;

1762:   PetscFunctionBegin;
1763:   if (PetscDefined(USE_DEBUG)) {
1764:     PetscInt  i;
1765:     PetscBool flg, missing;

1767:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
1768:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
1769:     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);
1770:     PetscCall(MatMissingDiagonal(A, &missing, &i));
1771:     PetscCheck(!missing, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry %" PetscInt_FMT, i);
1772:   }

1774:   /* Free the old stale stuff */
1775:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&fs));

1777:   /* Copy over A's meta data to fact. Note that we also allocated fact's i,j,a on host,
1778:      but they will not be used. Allocate them just for easy debugging.
1779:    */
1780:   PetscCall(MatDuplicateNoCreate_SeqAIJ(fact, A, MAT_DO_NOT_COPY_VALUES, PETSC_TRUE /*malloc*/));

1782:   fact->offloadmask            = PETSC_OFFLOAD_BOTH;
1783:   fact->factortype             = MAT_FACTOR_ILU;
1784:   fact->info.factor_mallocs    = 0;
1785:   fact->info.fill_ratio_given  = info->fill;
1786:   fact->info.fill_ratio_needed = 1.0;

1788:   aij->row = NULL;
1789:   aij->col = NULL;

1791:   /* ====================================================================== */
1792:   /* Copy A's i, j to fact and also allocate the value array of fact.       */
1793:   /* We'll do in-place factorization on fact                                */
1794:   /* ====================================================================== */
1795:   const int *Ai, *Aj;

1797:   m  = fact->rmap->n;
1798:   nz = aij->nz;

1800:   PetscCallCUDA(cudaMalloc((void **)&fs->csrRowPtr32, sizeof(*fs->csrRowPtr32) * (m + 1)));
1801:   PetscCallCUDA(cudaMalloc((void **)&fs->csrColIdx32, sizeof(*fs->csrColIdx32) * nz));
1802:   PetscCallCUDA(cudaMalloc((void **)&fs->csrVal, sizeof(*fs->csrVal) * nz));
1803:   PetscCall(MatSeqAIJCUSPARSEGetIJ(A, PETSC_FALSE, &Ai, &Aj)); /* Do not use compressed Ai.  The returned Ai, Aj are 32-bit */
1804:   PetscCallCUDA(cudaMemcpyAsync(fs->csrRowPtr32, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));
1805:   PetscCallCUDA(cudaMemcpyAsync(fs->csrColIdx32, Aj, sizeof(*Aj) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

1807:   /* ====================================================================== */
1808:   /* Create descriptors for M, L, U                                         */
1809:   /* ====================================================================== */
1810:   cusparseFillMode_t fillMode;
1811:   cusparseDiagType_t diagType;

1813:   PetscCallCUSPARSE(cusparseCreateMatDescr(&fs->matDescr_M));
1814:   PetscCallCUSPARSE(cusparseSetMatIndexBase(fs->matDescr_M, CUSPARSE_INDEX_BASE_ZERO));
1815:   PetscCallCUSPARSE(cusparseSetMatType(fs->matDescr_M, CUSPARSE_MATRIX_TYPE_GENERAL));

1817:   /* https://docs.nvidia.com/cuda/cusparse/index.html#cusparseDiagType_t
1818:     cusparseDiagType_t: This type indicates if the matrix diagonal entries are unity. The diagonal elements are always
1819:     assumed to be present, but if CUSPARSE_DIAG_TYPE_UNIT is passed to an API routine, then the routine assumes that
1820:     all diagonal entries are unity and will not read or modify those entries. Note that in this case the routine
1821:     assumes the diagonal entries are equal to one, regardless of what those entries are actually set to in memory.
1822:   */
1823:   fillMode = CUSPARSE_FILL_MODE_LOWER;
1824:   diagType = CUSPARSE_DIAG_TYPE_UNIT;
1825:   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));
1826:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
1827:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

1829:   fillMode = CUSPARSE_FILL_MODE_UPPER;
1830:   diagType = CUSPARSE_DIAG_TYPE_NON_UNIT;
1831:   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));
1832:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
1833:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_U, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

1835:   /* ========================================================================= */
1836:   /* Query buffer sizes for csrilu0, SpSV and allocate buffers                 */
1837:   /* ========================================================================= */
1838:   PetscCallCUSPARSE(cusparseCreateCsrilu02Info(&fs->ilu0Info_M));
1839:   if (m)
1840:     PetscCallCUSPARSE(cusparseXcsrilu02_bufferSize(fs->handle, m, nz, /* cusparseXcsrilu02 errors out with empty matrices (m=0) */
1841:                                                    fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ilu0Info_M, &fs->factBufferSize_M));

1843:   PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(PetscScalar) * m));
1844:   PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(PetscScalar) * m));

1846:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_X, m, fs->X, cusparse_scalartype));
1847:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_Y, m, fs->Y, cusparse_scalartype));

1849:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_L));
1850:   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));

1852:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_U));
1853:   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));

1855:   /* From my experiment with the example at https://github.com/NVIDIA/CUDALibrarySamples/tree/master/cuSPARSE/bicgstab,
1856:      and discussion at https://github.com/NVIDIA/CUDALibrarySamples/issues/77,
1857:      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.
1858:      To save memory, we make factBuffer_M share with the bigger of spsvBuffer_L/U.
1859:    */
1860:   if (fs->spsvBufferSize_L > fs->spsvBufferSize_U) {
1861:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_L, (size_t)fs->factBufferSize_M)));
1862:     fs->spsvBuffer_L = fs->factBuffer_M;
1863:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U));
1864:   } else {
1865:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_U, (size_t)fs->factBufferSize_M)));
1866:     fs->spsvBuffer_U = fs->factBuffer_M;
1867:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L));
1868:   }

1870:   /* ========================================================================== */
1871:   /* Perform analysis of ilu0 on M, SpSv on L and U                             */
1872:   /* The lower(upper) triangular part of M has the same sparsity pattern as L(U)*/
1873:   /* ========================================================================== */
1874:   int              structural_zero;
1875:   cusparseStatus_t status;

1877:   fs->policy_M = CUSPARSE_SOLVE_POLICY_USE_LEVEL;
1878:   if (m)
1879:     PetscCallCUSPARSE(cusparseXcsrilu02_analysis(fs->handle, m, nz, /* cusparseXcsrilu02 errors out with empty matrices (m=0) */
1880:                                                  fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ilu0Info_M, fs->policy_M, fs->factBuffer_M));
1881:   if (PetscDefined(USE_DEBUG)) {
1882:     /* cusparseXcsrilu02_zeroPivot() is a blocking call. It calls cudaDeviceSynchronize() to make sure all previous kernels are done. */
1883:     status = cusparseXcsrilu02_zeroPivot(fs->handle, fs->ilu0Info_M, &structural_zero);
1884:     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);
1885:   }

1887:   /* Estimate FLOPs of the numeric factorization */
1888:   {
1889:     Mat_SeqAIJ    *Aseq = (Mat_SeqAIJ *)A->data;
1890:     PetscInt      *Ai, *Adiag, nzRow, nzLeft;
1891:     PetscLogDouble flops = 0.0;

1893:     PetscCall(MatMarkDiagonal_SeqAIJ(A));
1894:     Ai    = Aseq->i;
1895:     Adiag = Aseq->diag;
1896:     for (PetscInt i = 0; i < m; i++) {
1897:       if (Ai[i] < Adiag[i] && Adiag[i] < Ai[i + 1]) { /* There are nonzeros left to the diagonal of row i */
1898:         nzRow  = Ai[i + 1] - Ai[i];
1899:         nzLeft = Adiag[i] - Ai[i];
1900:         /* We want to eliminate nonzeros left to the diagonal one by one. Assume each time, nonzeros right
1901:           and include the eliminated one will be updated, which incurs a multiplication and an addition.
1902:         */
1903:         nzLeft = (nzRow - 1) / 2;
1904:         flops += nzLeft * (2.0 * nzRow - nzLeft + 1);
1905:       }
1906:     }
1907:     fs->numericFactFlops = flops;
1908:   }
1909:   fact->ops->lufactornumeric = MatILUFactorNumeric_SeqAIJCUSPARSE_ILU0;
1910:   PetscFunctionReturn(PETSC_SUCCESS);
1911: }

1913: static PetscErrorCode MatSolve_SeqAIJCUSPARSE_ICC0(Mat fact, Vec b, Vec x)
1914: {
1915:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1916:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
1917:   const PetscScalar            *barray;
1918:   PetscScalar                  *xarray;

1920:   PetscFunctionBegin;
1921:   PetscCall(VecCUDAGetArrayWrite(x, &xarray));
1922:   PetscCall(VecCUDAGetArrayRead(b, &barray));
1923:   PetscCall(PetscLogGpuTimeBegin());

1925:   /* Solve L*y = b */
1926:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, (void *)barray));
1927:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_Y, fs->Y));
1928:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, /* L Y = X */
1929:                                        fs->dnVecDescr_X, fs->dnVecDescr_Y, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_L));

1931:   /* Solve Lt*x = y */
1932:   PetscCallCUSPARSE(cusparseDnVecSetValues(fs->dnVecDescr_X, xarray));
1933:   PetscCallCUSPARSE(cusparseSpSV_solve(fs->handle, CUSPARSE_OPERATION_TRANSPOSE, &PETSC_CUSPARSE_ONE, fs->spMatDescr_L, /* Lt X = Y */
1934:                                        fs->dnVecDescr_Y, fs->dnVecDescr_X, cusparse_scalartype, CUSPARSE_SPSV_ALG_DEFAULT, fs->spsvDescr_Lt));

1936:   PetscCall(VecCUDARestoreArrayRead(b, &barray));
1937:   PetscCall(VecCUDARestoreArrayWrite(x, &xarray));

1939:   PetscCall(PetscLogGpuTimeEnd());
1940:   PetscCall(PetscLogGpuFlops(2.0 * aij->nz - fact->rmap->n));
1941:   PetscFunctionReturn(PETSC_SUCCESS);
1942: }

1944: static PetscErrorCode MatICCFactorNumeric_SeqAIJCUSPARSE_ICC0(Mat fact, Mat A, const MatFactorInfo *)
1945: {
1946:   Mat_SeqAIJCUSPARSETriFactors *fs    = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
1947:   Mat_SeqAIJ                   *aij   = (Mat_SeqAIJ *)fact->data;
1948:   Mat_SeqAIJCUSPARSE           *Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
1949:   CsrMatrix                    *Acsr;
1950:   PetscInt                      m, nz;
1951:   PetscBool                     flg;

1953:   PetscFunctionBegin;
1954:   if (PetscDefined(USE_DEBUG)) {
1955:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
1956:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
1957:   }

1959:   /* Copy A's value to fact */
1960:   m  = fact->rmap->n;
1961:   nz = aij->nz;
1962:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
1963:   Acsr = (CsrMatrix *)Acusp->mat->mat;
1964:   PetscCallCUDA(cudaMemcpyAsync(fs->csrVal, Acsr->values->data().get(), sizeof(PetscScalar) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

1966:   /* Factorize fact inplace */
1967:   /* https://docs.nvidia.com/cuda/cusparse/index.html#csric02_solve
1968:      csric02() only takes the lower triangular part of matrix A to perform factorization.
1969:      The matrix type must be CUSPARSE_MATRIX_TYPE_GENERAL, the fill mode and diagonal type are ignored,
1970:      and the strictly upper triangular part is ignored and never touched. It does not matter if A is Hermitian or not.
1971:      In other words, from the point of view of csric02() A is Hermitian and only the lower triangular part is provided.
1972:    */
1973:   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));
1974:   if (PetscDefined(USE_DEBUG)) {
1975:     int              numerical_zero;
1976:     cusparseStatus_t status;
1977:     status = cusparseXcsric02_zeroPivot(fs->handle, fs->ic0Info_M, &numerical_zero);
1978:     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);
1979:   }

1981:   #if PETSC_PKG_CUDA_VERSION_GE(12, 1, 1)
1982:   if (fs->updatedSpSVAnalysis) {
1983:     if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_L, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
1984:     if (fs->csrVal) PetscCallCUSPARSE(cusparseSpSV_updateMatrix(fs->handle, fs->spsvDescr_Lt, fs->csrVal, CUSPARSE_SPSV_UPDATE_GENERAL));
1985:   } else
1986:   #endif
1987:   {
1988:     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));

1990:     /* Note that cusparse reports this error if we use double and CUSPARSE_OPERATION_CONJUGATE_TRANSPOSE
1991:     ** On entry to cusparseSpSV_analysis(): conjugate transpose (opA) is not supported for matA data type, current -> CUDA_R_64F
1992:   */
1993:     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));
1994:     fs->updatedSpSVAnalysis = PETSC_TRUE;
1995:   }

1997:   fact->offloadmask            = PETSC_OFFLOAD_GPU;
1998:   fact->ops->solve             = MatSolve_SeqAIJCUSPARSE_ICC0;
1999:   fact->ops->solvetranspose    = MatSolve_SeqAIJCUSPARSE_ICC0;
2000:   fact->ops->matsolve          = NULL;
2001:   fact->ops->matsolvetranspose = NULL;
2002:   PetscCall(PetscLogGpuFlops(fs->numericFactFlops));
2003:   PetscFunctionReturn(PETSC_SUCCESS);
2004: }

2006: static PetscErrorCode MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0(Mat fact, Mat A, IS, const MatFactorInfo *info)
2007: {
2008:   Mat_SeqAIJCUSPARSETriFactors *fs  = (Mat_SeqAIJCUSPARSETriFactors *)fact->spptr;
2009:   Mat_SeqAIJ                   *aij = (Mat_SeqAIJ *)fact->data;
2010:   PetscInt                      m, nz;

2012:   PetscFunctionBegin;
2013:   if (PetscDefined(USE_DEBUG)) {
2014:     PetscInt  i;
2015:     PetscBool flg, missing;

2017:     PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQAIJCUSPARSE, &flg));
2018:     PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Expected MATSEQAIJCUSPARSE, but input is %s", ((PetscObject)A)->type_name);
2019:     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);
2020:     PetscCall(MatMissingDiagonal(A, &missing, &i));
2021:     PetscCheck(!missing, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry %" PetscInt_FMT, i);
2022:   }

2024:   /* Free the old stale stuff */
2025:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&fs));

2027:   /* Copy over A's meta data to fact. Note that we also allocated fact's i,j,a on host,
2028:      but they will not be used. Allocate them just for easy debugging.
2029:    */
2030:   PetscCall(MatDuplicateNoCreate_SeqAIJ(fact, A, MAT_DO_NOT_COPY_VALUES, PETSC_TRUE /*malloc*/));

2032:   fact->offloadmask            = PETSC_OFFLOAD_BOTH;
2033:   fact->factortype             = MAT_FACTOR_ICC;
2034:   fact->info.factor_mallocs    = 0;
2035:   fact->info.fill_ratio_given  = info->fill;
2036:   fact->info.fill_ratio_needed = 1.0;

2038:   aij->row = NULL;
2039:   aij->col = NULL;

2041:   /* ====================================================================== */
2042:   /* Copy A's i, j to fact and also allocate the value array of fact.       */
2043:   /* We'll do in-place factorization on fact                                */
2044:   /* ====================================================================== */
2045:   const int *Ai, *Aj;

2047:   m  = fact->rmap->n;
2048:   nz = aij->nz;

2050:   PetscCallCUDA(cudaMalloc((void **)&fs->csrRowPtr32, sizeof(*fs->csrRowPtr32) * (m + 1)));
2051:   PetscCallCUDA(cudaMalloc((void **)&fs->csrColIdx32, sizeof(*fs->csrColIdx32) * nz));
2052:   PetscCallCUDA(cudaMalloc((void **)&fs->csrVal, sizeof(PetscScalar) * nz));
2053:   PetscCall(MatSeqAIJCUSPARSEGetIJ(A, PETSC_FALSE, &Ai, &Aj)); /* Do not use compressed Ai */
2054:   PetscCallCUDA(cudaMemcpyAsync(fs->csrRowPtr32, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));
2055:   PetscCallCUDA(cudaMemcpyAsync(fs->csrColIdx32, Aj, sizeof(*Aj) * nz, cudaMemcpyDeviceToDevice, PetscDefaultCudaStream));

2057:   /* ====================================================================== */
2058:   /* Create mat descriptors for M, L                                        */
2059:   /* ====================================================================== */
2060:   cusparseFillMode_t fillMode;
2061:   cusparseDiagType_t diagType;

2063:   PetscCallCUSPARSE(cusparseCreateMatDescr(&fs->matDescr_M));
2064:   PetscCallCUSPARSE(cusparseSetMatIndexBase(fs->matDescr_M, CUSPARSE_INDEX_BASE_ZERO));
2065:   PetscCallCUSPARSE(cusparseSetMatType(fs->matDescr_M, CUSPARSE_MATRIX_TYPE_GENERAL));

2067:   /* https://docs.nvidia.com/cuda/cusparse/index.html#cusparseDiagType_t
2068:     cusparseDiagType_t: This type indicates if the matrix diagonal entries are unity. The diagonal elements are always
2069:     assumed to be present, but if CUSPARSE_DIAG_TYPE_UNIT is passed to an API routine, then the routine assumes that
2070:     all diagonal entries are unity and will not read or modify those entries. Note that in this case the routine
2071:     assumes the diagonal entries are equal to one, regardless of what those entries are actually set to in memory.
2072:   */
2073:   fillMode = CUSPARSE_FILL_MODE_LOWER;
2074:   diagType = CUSPARSE_DIAG_TYPE_NON_UNIT;
2075:   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));
2076:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_FILL_MODE, &fillMode, sizeof(fillMode)));
2077:   PetscCallCUSPARSE(cusparseSpMatSetAttribute(fs->spMatDescr_L, CUSPARSE_SPMAT_DIAG_TYPE, &diagType, sizeof(diagType)));

2079:   /* ========================================================================= */
2080:   /* Query buffer sizes for csric0, SpSV of L and Lt, and allocate buffers     */
2081:   /* ========================================================================= */
2082:   PetscCallCUSPARSE(cusparseCreateCsric02Info(&fs->ic0Info_M));
2083:   if (m) PetscCallCUSPARSE(cusparseXcsric02_bufferSize(fs->handle, m, nz, fs->matDescr_M, fs->csrVal, fs->csrRowPtr32, fs->csrColIdx32, fs->ic0Info_M, &fs->factBufferSize_M));

2085:   PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(PetscScalar) * m));
2086:   PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(PetscScalar) * m));

2088:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_X, m, fs->X, cusparse_scalartype));
2089:   PetscCallCUSPARSE(cusparseCreateDnVec(&fs->dnVecDescr_Y, m, fs->Y, cusparse_scalartype));

2091:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_L));
2092:   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));

2094:   PetscCallCUSPARSE(cusparseSpSV_createDescr(&fs->spsvDescr_Lt));
2095:   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));

2097:   /* To save device memory, we make the factorization buffer share with one of the solver buffer.
2098:      See also comments in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0().
2099:    */
2100:   if (fs->spsvBufferSize_L > fs->spsvBufferSize_Lt) {
2101:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_L, (size_t)fs->factBufferSize_M)));
2102:     fs->spsvBuffer_L = fs->factBuffer_M;
2103:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Lt, fs->spsvBufferSize_Lt));
2104:   } else {
2105:     PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_Lt, (size_t)fs->factBufferSize_M)));
2106:     fs->spsvBuffer_Lt = fs->factBuffer_M;
2107:     PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L));
2108:   }

2110:   /* ========================================================================== */
2111:   /* Perform analysis of ic0 on M                                               */
2112:   /* The lower triangular part of M has the same sparsity pattern as L          */
2113:   /* ========================================================================== */
2114:   int              structural_zero;
2115:   cusparseStatus_t status;

2117:   fs->policy_M = CUSPARSE_SOLVE_POLICY_USE_LEVEL;
2118:   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));
2119:   if (PetscDefined(USE_DEBUG)) {
2120:     /* cusparseXcsric02_zeroPivot() is a blocking call. It calls cudaDeviceSynchronize() to make sure all previous kernels are done. */
2121:     status = cusparseXcsric02_zeroPivot(fs->handle, fs->ic0Info_M, &structural_zero);
2122:     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);
2123:   }

2125:   /* Estimate FLOPs of the numeric factorization */
2126:   {
2127:     Mat_SeqAIJ    *Aseq = (Mat_SeqAIJ *)A->data;
2128:     PetscInt      *Ai, nzRow, nzLeft;
2129:     PetscLogDouble flops = 0.0;

2131:     Ai = Aseq->i;
2132:     for (PetscInt i = 0; i < m; i++) {
2133:       nzRow = Ai[i + 1] - Ai[i];
2134:       if (nzRow > 1) {
2135:         /* We want to eliminate nonzeros left to the diagonal one by one. Assume each time, nonzeros right
2136:           and include the eliminated one will be updated, which incurs a multiplication and an addition.
2137:         */
2138:         nzLeft = (nzRow - 1) / 2;
2139:         flops += nzLeft * (2.0 * nzRow - nzLeft + 1);
2140:       }
2141:     }
2142:     fs->numericFactFlops = flops;
2143:   }
2144:   fact->ops->choleskyfactornumeric = MatICCFactorNumeric_SeqAIJCUSPARSE_ICC0;
2145:   PetscFunctionReturn(PETSC_SUCCESS);
2146: }
2147: #endif

2149: static PetscErrorCode MatLUFactorNumeric_SeqAIJCUSPARSE(Mat B, Mat A, const MatFactorInfo *info)
2150: {
2151:   // use_cpu_solve is a field in Mat_SeqAIJCUSPARSE. B, a factored matrix, uses Mat_SeqAIJCUSPARSETriFactors.
2152:   Mat_SeqAIJCUSPARSE *cusparsestruct = static_cast<Mat_SeqAIJCUSPARSE *>(A->spptr);

2154:   PetscFunctionBegin;
2155:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2156:   PetscCall(MatLUFactorNumeric_SeqAIJ(B, A, info));
2157:   B->offloadmask = PETSC_OFFLOAD_CPU;

2159:   if (!cusparsestruct->use_cpu_solve) {
2160: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2161:     B->ops->solve          = MatSolve_SeqAIJCUSPARSE_LU;
2162:     B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE_LU;
2163: #else
2164:     /* determine which version of MatSolve needs to be used. */
2165:     Mat_SeqAIJ *b     = (Mat_SeqAIJ *)B->data;
2166:     IS          isrow = b->row, iscol = b->col;
2167:     PetscBool   row_identity, col_identity;

2169:     PetscCall(ISIdentity(isrow, &row_identity));
2170:     PetscCall(ISIdentity(iscol, &col_identity));
2171:     if (row_identity && col_identity) {
2172:       B->ops->solve          = MatSolve_SeqAIJCUSPARSE_NaturalOrdering;
2173:       B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE_NaturalOrdering;
2174:     } else {
2175:       B->ops->solve          = MatSolve_SeqAIJCUSPARSE;
2176:       B->ops->solvetranspose = MatSolveTranspose_SeqAIJCUSPARSE;
2177:     }
2178: #endif
2179:   }
2180:   B->ops->matsolve          = NULL;
2181:   B->ops->matsolvetranspose = NULL;

2183:   /* get the triangular factors */
2184:   if (!cusparsestruct->use_cpu_solve) PetscCall(MatSeqAIJCUSPARSEILUAnalysisAndCopyToGPU(B));
2185:   PetscFunctionReturn(PETSC_SUCCESS);
2186: }

2188: static PetscErrorCode MatLUFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS isrow, IS iscol, const MatFactorInfo *info)
2189: {
2190:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = static_cast<Mat_SeqAIJCUSPARSETriFactors *>(B->spptr);

2192:   PetscFunctionBegin;
2193:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2194:   PetscCall(MatLUFactorSymbolic_SeqAIJ(B, A, isrow, iscol, info));
2195:   B->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJCUSPARSE;
2196:   PetscFunctionReturn(PETSC_SUCCESS);
2197: }

2199: static PetscErrorCode MatILUFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS isrow, IS iscol, const MatFactorInfo *info)
2200: {
2201:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2203:   PetscFunctionBegin;
2204: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2205:   PetscBool row_identity = PETSC_FALSE, col_identity = PETSC_FALSE;
2206:   if (!info->factoronhost) {
2207:     PetscCall(ISIdentity(isrow, &row_identity));
2208:     PetscCall(ISIdentity(iscol, &col_identity));
2209:   }
2210:   if (!info->levels && row_identity && col_identity) {
2211:     PetscCall(MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0(B, A, isrow, iscol, info));
2212:   } else
2213: #endif
2214:   {
2215:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2216:     PetscCall(MatILUFactorSymbolic_SeqAIJ(B, A, isrow, iscol, info));
2217:     B->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJCUSPARSE;
2218:   }
2219:   PetscFunctionReturn(PETSC_SUCCESS);
2220: }

2222: static PetscErrorCode MatICCFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS perm, const MatFactorInfo *info)
2223: {
2224:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2226:   PetscFunctionBegin;
2227: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2228:   PetscBool perm_identity = PETSC_FALSE;
2229:   if (!info->factoronhost) PetscCall(ISIdentity(perm, &perm_identity));
2230:   if (!info->levels && perm_identity) {
2231:     PetscCall(MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0(B, A, perm, info));
2232:   } else
2233: #endif
2234:   {
2235:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2236:     PetscCall(MatICCFactorSymbolic_SeqAIJ(B, A, perm, info));
2237:     B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJCUSPARSE;
2238:   }
2239:   PetscFunctionReturn(PETSC_SUCCESS);
2240: }

2242: static PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJCUSPARSE(Mat B, Mat A, IS perm, const MatFactorInfo *info)
2243: {
2244:   Mat_SeqAIJCUSPARSETriFactors *cusparseTriFactors = (Mat_SeqAIJCUSPARSETriFactors *)B->spptr;

2246:   PetscFunctionBegin;
2247:   PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(&cusparseTriFactors));
2248:   PetscCall(MatCholeskyFactorSymbolic_SeqAIJ(B, A, perm, info));
2249:   B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJCUSPARSE;
2250:   PetscFunctionReturn(PETSC_SUCCESS);
2251: }

2253: static PetscErrorCode MatFactorGetSolverType_seqaij_cusparse(Mat, MatSolverType *type)
2254: {
2255:   PetscFunctionBegin;
2256:   *type = MATSOLVERCUSPARSE;
2257:   PetscFunctionReturn(PETSC_SUCCESS);
2258: }

2260: /*MC
2261:   MATSOLVERCUSPARSE = "cusparse" - A matrix type providing triangular solvers for seq matrices
2262:   on a single GPU of type, `MATSEQAIJCUSPARSE`. Currently supported
2263:   algorithms are ILU(k) and ICC(k). Typically, deeper factorizations (larger k) results in poorer
2264:   performance in the triangular solves. Full LU, and Cholesky decompositions can be solved through the
2265:   CuSPARSE triangular solve algorithm. However, the performance can be quite poor and thus these
2266:   algorithms are not recommended. This class does NOT support direct solver operations.

2268:   Level: beginner

2270: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJCUSPARSE`, `PCFactorSetMatSolverType()`, `MatSolverType`, `MatCreateSeqAIJCUSPARSE()`,
2271:           `MATAIJCUSPARSE`, `MatCreateAIJCUSPARSE()`, `MatCUSPARSESetFormat()`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
2272: M*/

2274: PETSC_EXTERN PetscErrorCode MatGetFactor_seqaijcusparse_cusparse(Mat A, MatFactorType ftype, Mat *B)
2275: {
2276:   PetscInt n = A->rmap->n;

2278:   PetscFunctionBegin;
2279:   PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
2280:   PetscCall(MatSetSizes(*B, n, n, n, n));
2281:   (*B)->factortype = ftype; // factortype makes MatSetType() allocate spptr of type Mat_SeqAIJCUSPARSETriFactors
2282:   PetscCall(MatSetType(*B, MATSEQAIJCUSPARSE));

2284:   if (A->boundtocpu && A->bindingpropagates) PetscCall(MatBindToCPU(*B, PETSC_TRUE));
2285:   if (ftype == MAT_FACTOR_LU || ftype == MAT_FACTOR_ILU || ftype == MAT_FACTOR_ILUDT) {
2286:     PetscCall(MatSetBlockSizesFromMats(*B, A, A));
2287:     if (!A->boundtocpu) {
2288:       (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJCUSPARSE;
2289:       (*B)->ops->lufactorsymbolic  = MatLUFactorSymbolic_SeqAIJCUSPARSE;
2290:     } else {
2291:       (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJ;
2292:       (*B)->ops->lufactorsymbolic  = MatLUFactorSymbolic_SeqAIJ;
2293:     }
2294:     PetscCall(PetscStrallocpy(MATORDERINGND, (char **)&(*B)->preferredordering[MAT_FACTOR_LU]));
2295:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ILU]));
2296:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ILUDT]));
2297:   } else if (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) {
2298:     if (!A->boundtocpu) {
2299:       (*B)->ops->iccfactorsymbolic      = MatICCFactorSymbolic_SeqAIJCUSPARSE;
2300:       (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJCUSPARSE;
2301:     } else {
2302:       (*B)->ops->iccfactorsymbolic      = MatICCFactorSymbolic_SeqAIJ;
2303:       (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJ;
2304:     }
2305:     PetscCall(PetscStrallocpy(MATORDERINGND, (char **)&(*B)->preferredordering[MAT_FACTOR_CHOLESKY]));
2306:     PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ICC]));
2307:   } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Factor type not supported for CUSPARSE Matrix Types");

2309:   PetscCall(MatSeqAIJSetPreallocation(*B, MAT_SKIP_ALLOCATION, NULL));
2310:   (*B)->canuseordering = PETSC_TRUE;
2311:   PetscCall(PetscObjectComposeFunction((PetscObject)*B, "MatFactorGetSolverType_C", MatFactorGetSolverType_seqaij_cusparse));
2312:   PetscFunctionReturn(PETSC_SUCCESS);
2313: }

2315: static PetscErrorCode MatSeqAIJCUSPARSECopyFromGPU(Mat A)
2316: {
2317:   Mat_SeqAIJ         *a    = (Mat_SeqAIJ *)A->data;
2318:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
2319: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2320:   Mat_SeqAIJCUSPARSETriFactors *fs = (Mat_SeqAIJCUSPARSETriFactors *)A->spptr;
2321: #endif

2323:   PetscFunctionBegin;
2324:   if (A->offloadmask == PETSC_OFFLOAD_GPU) {
2325:     PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyFromGPU, A, 0, 0, 0));
2326:     if (A->factortype == MAT_FACTOR_NONE) {
2327:       CsrMatrix *matrix = (CsrMatrix *)cusp->mat->mat;
2328:       PetscCallCUDA(cudaMemcpy(a->a, matrix->values->data().get(), a->nz * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
2329:     }
2330: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
2331:     else if (fs->csrVal) {
2332:       /* We have a factorized matrix on device and are able to copy it to host */
2333:       PetscCallCUDA(cudaMemcpy(a->a, fs->csrVal, a->nz * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
2334:     }
2335: #endif
2336:     else
2337:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for copying this type of factorized matrix from device to host");
2338:     PetscCall(PetscLogGpuToCpu(a->nz * sizeof(PetscScalar)));
2339:     PetscCall(PetscLogEventEnd(MAT_CUSPARSECopyFromGPU, A, 0, 0, 0));
2340:     A->offloadmask = PETSC_OFFLOAD_BOTH;
2341:   }
2342:   PetscFunctionReturn(PETSC_SUCCESS);
2343: }

2345: static PetscErrorCode MatSeqAIJGetArray_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2346: {
2347:   PetscFunctionBegin;
2348:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2349:   *array = ((Mat_SeqAIJ *)A->data)->a;
2350:   PetscFunctionReturn(PETSC_SUCCESS);
2351: }

2353: static PetscErrorCode MatSeqAIJRestoreArray_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2354: {
2355:   PetscFunctionBegin;
2356:   A->offloadmask = PETSC_OFFLOAD_CPU;
2357:   *array         = NULL;
2358:   PetscFunctionReturn(PETSC_SUCCESS);
2359: }

2361: static PetscErrorCode MatSeqAIJGetArrayRead_SeqAIJCUSPARSE(Mat A, const PetscScalar *array[])
2362: {
2363:   PetscFunctionBegin;
2364:   PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));
2365:   *array = ((Mat_SeqAIJ *)A->data)->a;
2366:   PetscFunctionReturn(PETSC_SUCCESS);
2367: }

2369: static PetscErrorCode MatSeqAIJRestoreArrayRead_SeqAIJCUSPARSE(Mat, const PetscScalar *array[])
2370: {
2371:   PetscFunctionBegin;
2372:   *array = NULL;
2373:   PetscFunctionReturn(PETSC_SUCCESS);
2374: }

2376: static PetscErrorCode MatSeqAIJGetArrayWrite_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2377: {
2378:   PetscFunctionBegin;
2379:   *array = ((Mat_SeqAIJ *)A->data)->a;
2380:   PetscFunctionReturn(PETSC_SUCCESS);
2381: }

2383: static PetscErrorCode MatSeqAIJRestoreArrayWrite_SeqAIJCUSPARSE(Mat A, PetscScalar *array[])
2384: {
2385:   PetscFunctionBegin;
2386:   A->offloadmask = PETSC_OFFLOAD_CPU;
2387:   *array         = NULL;
2388:   PetscFunctionReturn(PETSC_SUCCESS);
2389: }

2391: static PetscErrorCode MatSeqAIJGetCSRAndMemType_SeqAIJCUSPARSE(Mat A, const PetscInt **i, const PetscInt **j, PetscScalar **a, PetscMemType *mtype)
2392: {
2393:   Mat_SeqAIJCUSPARSE *cusp;
2394:   CsrMatrix          *matrix;

2396:   PetscFunctionBegin;
2397:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
2398:   PetscCheck(A->factortype == MAT_FACTOR_NONE, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2399:   cusp = static_cast<Mat_SeqAIJCUSPARSE *>(A->spptr);
2400:   PetscCheck(cusp != NULL, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "cusp is NULL");
2401:   matrix = (CsrMatrix *)cusp->mat->mat;

2403:   if (i) {
2404: #if !defined(PETSC_USE_64BIT_INDICES)
2405:     *i = matrix->row_offsets->data().get();
2406: #else
2407:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSparse does not supported 64-bit indices");
2408: #endif
2409:   }
2410:   if (j) {
2411: #if !defined(PETSC_USE_64BIT_INDICES)
2412:     *j = matrix->column_indices->data().get();
2413: #else
2414:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "cuSparse does not supported 64-bit indices");
2415: #endif
2416:   }
2417:   if (a) *a = matrix->values->data().get();
2418:   if (mtype) *mtype = PETSC_MEMTYPE_CUDA;
2419:   PetscFunctionReturn(PETSC_SUCCESS);
2420: }

2422: PETSC_INTERN PetscErrorCode MatSeqAIJCUSPARSECopyToGPU(Mat A)
2423: {
2424:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
2425:   Mat_SeqAIJCUSPARSEMultStruct *matstruct      = cusparsestruct->mat;
2426:   Mat_SeqAIJ                   *a              = (Mat_SeqAIJ *)A->data;
2427:   PetscInt                      m              = A->rmap->n, *ii, *ridx, tmp;
2428:   cusparseStatus_t              stat;
2429:   PetscBool                     both = PETSC_TRUE;

2431:   PetscFunctionBegin;
2432:   PetscCheck(!A->boundtocpu, PETSC_COMM_SELF, PETSC_ERR_GPU, "Cannot copy to GPU");
2433:   if (A->offloadmask == PETSC_OFFLOAD_UNALLOCATED || A->offloadmask == PETSC_OFFLOAD_CPU) {
2434:     if (A->nonzerostate == cusparsestruct->nonzerostate && cusparsestruct->format == MAT_CUSPARSE_CSR) { /* Copy values only */
2435:       CsrMatrix *matrix;
2436:       matrix = (CsrMatrix *)cusparsestruct->mat->mat;

2438:       PetscCheck(!a->nz || a->a, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR values");
2439:       PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2440:       matrix->values->assign(a->a, a->a + a->nz);
2441:       PetscCallCUDA(WaitForCUDA());
2442:       PetscCall(PetscLogCpuToGpu(a->nz * sizeof(PetscScalar)));
2443:       PetscCall(PetscLogEventEnd(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2444:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
2445:     } else {
2446:       PetscInt nnz;
2447:       PetscCall(PetscLogEventBegin(MAT_CUSPARSECopyToGPU, A, 0, 0, 0));
2448:       PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusparsestruct->mat, cusparsestruct->format));
2449:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_TRUE));
2450:       delete cusparsestruct->workVector;
2451:       delete cusparsestruct->rowoffsets_gpu;
2452:       cusparsestruct->workVector     = NULL;
2453:       cusparsestruct->rowoffsets_gpu = NULL;
2454:       try {
2455:         if (a->compressedrow.use) {
2456:           m    = a->compressedrow.nrows;
2457:           ii   = a->compressedrow.i;
2458:           ridx = a->compressedrow.rindex;
2459:         } else {
2460:           m    = A->rmap->n;
2461:           ii   = a->i;
2462:           ridx = NULL;
2463:         }
2464:         PetscCheck(ii, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR row data");
2465:         if (!a->a) {
2466:           nnz  = ii[m];
2467:           both = PETSC_FALSE;
2468:         } else nnz = a->nz;
2469:         PetscCheck(!nnz || a->j, PETSC_COMM_SELF, PETSC_ERR_GPU, "Missing CSR column data");

2471:         /* create cusparse matrix */
2472:         cusparsestruct->nrows = m;
2473:         matstruct             = new Mat_SeqAIJCUSPARSEMultStruct;
2474:         PetscCallCUSPARSE(cusparseCreateMatDescr(&matstruct->descr));
2475:         PetscCallCUSPARSE(cusparseSetMatIndexBase(matstruct->descr, CUSPARSE_INDEX_BASE_ZERO));
2476:         PetscCallCUSPARSE(cusparseSetMatType(matstruct->descr, CUSPARSE_MATRIX_TYPE_GENERAL));

2478:         PetscCallCUDA(cudaMalloc((void **)&matstruct->alpha_one, sizeof(PetscScalar)));
2479:         PetscCallCUDA(cudaMalloc((void **)&matstruct->beta_zero, sizeof(PetscScalar)));
2480:         PetscCallCUDA(cudaMalloc((void **)&matstruct->beta_one, sizeof(PetscScalar)));
2481:         PetscCallCUDA(cudaMemcpy(matstruct->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2482:         PetscCallCUDA(cudaMemcpy(matstruct->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2483:         PetscCallCUDA(cudaMemcpy(matstruct->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
2484:         PetscCallCUSPARSE(cusparseSetPointerMode(cusparsestruct->handle, CUSPARSE_POINTER_MODE_DEVICE));

2486:         /* Build a hybrid/ellpack matrix if this option is chosen for the storage */
2487:         if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
2488:           /* set the matrix */
2489:           CsrMatrix *mat   = new CsrMatrix;
2490:           mat->num_rows    = m;
2491:           mat->num_cols    = A->cmap->n;
2492:           mat->num_entries = nnz;
2493:           PetscCallCXX(mat->row_offsets = new THRUSTINTARRAY32(m + 1));
2494:           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 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 MatDestroy_MatMatCusparse(void *data)
2622: {
2623:   MatMatCusparse *mmdata = (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:   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 = (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:   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 = MatDestroy_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:   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 = (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:   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 = MatDestroy_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(MatMarkDiagonal_SeqAIJ(C));
3375:   PetscCall(PetscMalloc1(c->nz, &c->a));
3376:   Ccsr->num_entries = c->nz;

3378:   C->nonzerostate++;
3379:   PetscCall(PetscLayoutSetUp(C->rmap));
3380:   PetscCall(PetscLayoutSetUp(C->cmap));
3381:   Ccusp->nonzerostate = C->nonzerostate;
3382:   C->offloadmask      = PETSC_OFFLOAD_UNALLOCATED;
3383:   C->preallocated     = PETSC_TRUE;
3384:   C->assembled        = PETSC_FALSE;
3385:   C->was_assembled    = PETSC_FALSE;
3386:   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 */
3387:     mmdata->reusesym = PETSC_TRUE;
3388:     C->offloadmask   = PETSC_OFFLOAD_GPU;
3389:   }
3390:   C->ops->productnumeric = MatProductNumeric_SeqAIJCUSPARSE_SeqAIJCUSPARSE;
3391:   PetscFunctionReturn(PETSC_SUCCESS);
3392: }

3394: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ_SeqDense(Mat);

3396: /* handles sparse or dense B */
3397: static PetscErrorCode MatProductSetFromOptions_SeqAIJCUSPARSE(Mat mat)
3398: {
3399:   Mat_Product *product = mat->product;
3400:   PetscBool    isdense = PETSC_FALSE, Biscusp = PETSC_FALSE, Ciscusp = PETSC_TRUE;

3402:   PetscFunctionBegin;
3403:   MatCheckProduct(mat, 1);
3404:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)product->B, MATSEQDENSE, &isdense));
3405:   if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, MATSEQAIJCUSPARSE, &Biscusp));
3406:   if (product->type == MATPRODUCT_ABC) {
3407:     Ciscusp = PETSC_FALSE;
3408:     if (!product->C->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->C, MATSEQAIJCUSPARSE, &Ciscusp));
3409:   }
3410:   if (Biscusp && Ciscusp) { /* we can always select the CPU backend */
3411:     PetscBool usecpu = PETSC_FALSE;
3412:     switch (product->type) {
3413:     case MATPRODUCT_AB:
3414:       if (product->api_user) {
3415:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
3416:         PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
3417:         PetscOptionsEnd();
3418:       } else {
3419:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
3420:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
3421:         PetscOptionsEnd();
3422:       }
3423:       break;
3424:     case MATPRODUCT_AtB:
3425:       if (product->api_user) {
3426:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
3427:         PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
3428:         PetscOptionsEnd();
3429:       } else {
3430:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
3431:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
3432:         PetscOptionsEnd();
3433:       }
3434:       break;
3435:     case MATPRODUCT_PtAP:
3436:       if (product->api_user) {
3437:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
3438:         PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
3439:         PetscOptionsEnd();
3440:       } else {
3441:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
3442:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
3443:         PetscOptionsEnd();
3444:       }
3445:       break;
3446:     case MATPRODUCT_RARt:
3447:       if (product->api_user) {
3448:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatRARt", "Mat");
3449:         PetscCall(PetscOptionsBool("-matrart_backend_cpu", "Use CPU code", "MatRARt", usecpu, &usecpu, NULL));
3450:         PetscOptionsEnd();
3451:       } else {
3452:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_RARt", "Mat");
3453:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatRARt", usecpu, &usecpu, NULL));
3454:         PetscOptionsEnd();
3455:       }
3456:       break;
3457:     case MATPRODUCT_ABC:
3458:       if (product->api_user) {
3459:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMatMult", "Mat");
3460:         PetscCall(PetscOptionsBool("-matmatmatmult_backend_cpu", "Use CPU code", "MatMatMatMult", usecpu, &usecpu, NULL));
3461:         PetscOptionsEnd();
3462:       } else {
3463:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_ABC", "Mat");
3464:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMatMult", usecpu, &usecpu, NULL));
3465:         PetscOptionsEnd();
3466:       }
3467:       break;
3468:     default:
3469:       break;
3470:     }
3471:     if (usecpu) Biscusp = Ciscusp = PETSC_FALSE;
3472:   }
3473:   /* dispatch */
3474:   if (isdense) {
3475:     switch (product->type) {
3476:     case MATPRODUCT_AB:
3477:     case MATPRODUCT_AtB:
3478:     case MATPRODUCT_ABt:
3479:     case MATPRODUCT_PtAP:
3480:     case MATPRODUCT_RARt:
3481:       if (product->A->boundtocpu) {
3482:         PetscCall(MatProductSetFromOptions_SeqAIJ_SeqDense(mat));
3483:       } else {
3484:         mat->ops->productsymbolic = MatProductSymbolic_SeqAIJCUSPARSE_SeqDENSECUDA;
3485:       }
3486:       break;
3487:     case MATPRODUCT_ABC:
3488:       mat->ops->productsymbolic = MatProductSymbolic_ABC_Basic;
3489:       break;
3490:     default:
3491:       break;
3492:     }
3493:   } else if (Biscusp && Ciscusp) {
3494:     switch (product->type) {
3495:     case MATPRODUCT_AB:
3496:     case MATPRODUCT_AtB:
3497:     case MATPRODUCT_ABt:
3498:       mat->ops->productsymbolic = MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE;
3499:       break;
3500:     case MATPRODUCT_PtAP:
3501:     case MATPRODUCT_RARt:
3502:     case MATPRODUCT_ABC:
3503:       mat->ops->productsymbolic = MatProductSymbolic_ABC_Basic;
3504:       break;
3505:     default:
3506:       break;
3507:     }
3508:   } else { /* fallback for AIJ */
3509:     PetscCall(MatProductSetFromOptions_SeqAIJ(mat));
3510:   }
3511:   PetscFunctionReturn(PETSC_SUCCESS);
3512: }

3514: static PetscErrorCode MatMult_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3515: {
3516:   PetscFunctionBegin;
3517:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_FALSE, PETSC_FALSE));
3518:   PetscFunctionReturn(PETSC_SUCCESS);
3519: }

3521: static PetscErrorCode MatMultAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3522: {
3523:   PetscFunctionBegin;
3524:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_FALSE, PETSC_FALSE));
3525:   PetscFunctionReturn(PETSC_SUCCESS);
3526: }

3528: static PetscErrorCode MatMultHermitianTranspose_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3529: {
3530:   PetscFunctionBegin;
3531:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_TRUE, PETSC_TRUE));
3532:   PetscFunctionReturn(PETSC_SUCCESS);
3533: }

3535: static PetscErrorCode MatMultHermitianTransposeAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3536: {
3537:   PetscFunctionBegin;
3538:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_TRUE, PETSC_TRUE));
3539:   PetscFunctionReturn(PETSC_SUCCESS);
3540: }

3542: static PetscErrorCode MatMultTranspose_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy)
3543: {
3544:   PetscFunctionBegin;
3545:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, NULL, yy, PETSC_TRUE, PETSC_FALSE));
3546:   PetscFunctionReturn(PETSC_SUCCESS);
3547: }

3549: __global__ static void ScatterAdd(PetscInt n, PetscInt *idx, const PetscScalar *x, PetscScalar *y)
3550: {
3551:   int i = blockIdx.x * blockDim.x + threadIdx.x;
3552:   if (i < n) y[idx[i]] += x[i];
3553: }

3555: /* z = op(A) x + y. If trans & !herm, op = ^T; if trans & herm, op = ^H; if !trans, op = no-op */
3556: static PetscErrorCode MatMultAddKernel_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz, PetscBool trans, PetscBool herm)
3557: {
3558:   Mat_SeqAIJ                   *a              = (Mat_SeqAIJ *)A->data;
3559:   Mat_SeqAIJCUSPARSE           *cusparsestruct = (Mat_SeqAIJCUSPARSE *)A->spptr;
3560:   Mat_SeqAIJCUSPARSEMultStruct *matstruct;
3561:   PetscScalar                  *xarray, *zarray, *dptr, *beta, *xptr;
3562:   cusparseOperation_t           opA = CUSPARSE_OPERATION_NON_TRANSPOSE;
3563:   PetscBool                     compressed;
3564: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3565:   PetscInt nx, ny;
3566: #endif

3568:   PetscFunctionBegin;
3569:   PetscCheck(!herm || trans, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "Hermitian and not transpose not supported");
3570:   if (!a->nz) {
3571:     if (yy) PetscCall(VecSeq_CUDA::Copy(yy, zz));
3572:     else PetscCall(VecSeq_CUDA::Set(zz, 0));
3573:     PetscFunctionReturn(PETSC_SUCCESS);
3574:   }
3575:   /* The line below is necessary due to the operations that modify the matrix on the CPU (axpy, scale, etc) */
3576:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
3577:   if (!trans) {
3578:     matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
3579:     PetscCheck(matstruct, PetscObjectComm((PetscObject)A), PETSC_ERR_GPU, "SeqAIJCUSPARSE does not have a 'mat' (need to fix)");
3580:   } else {
3581:     if (herm || !A->form_explicit_transpose) {
3582:       opA       = herm ? CUSPARSE_OPERATION_CONJUGATE_TRANSPOSE : CUSPARSE_OPERATION_TRANSPOSE;
3583:       matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->mat;
3584:     } else {
3585:       if (!cusparsestruct->matTranspose) PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
3586:       matstruct = (Mat_SeqAIJCUSPARSEMultStruct *)cusparsestruct->matTranspose;
3587:     }
3588:   }
3589:   /* Does the matrix use compressed rows (i.e., drop zero rows)? */
3590:   compressed = matstruct->cprowIndices ? PETSC_TRUE : PETSC_FALSE;

3592:   try {
3593:     PetscCall(VecCUDAGetArrayRead(xx, (const PetscScalar **)&xarray));
3594:     if (yy == zz) PetscCall(VecCUDAGetArray(zz, &zarray)); /* read & write zz, so need to get up-to-date zarray on GPU */
3595:     else PetscCall(VecCUDAGetArrayWrite(zz, &zarray));     /* write zz, so no need to init zarray on GPU */

3597:     PetscCall(PetscLogGpuTimeBegin());
3598:     if (opA == CUSPARSE_OPERATION_NON_TRANSPOSE) {
3599:       /* z = A x + beta y.
3600:          If A is compressed (with less rows), then Ax is shorter than the full z, so we need a work vector to store Ax.
3601:          When A is non-compressed, and z = y, we can set beta=1 to compute y = Ax + y in one call.
3602:       */
3603:       xptr = xarray;
3604:       dptr = compressed ? cusparsestruct->workVector->data().get() : zarray;
3605:       beta = (yy == zz && !compressed) ? matstruct->beta_one : matstruct->beta_zero;
3606: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3607:       /* Get length of x, y for y=Ax. ny might be shorter than the work vector's allocated length, since the work vector is
3608:           allocated to accommodate different uses. So we get the length info directly from mat.
3609:        */
3610:       if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3611:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3612:         nx             = mat->num_cols; // since y = Ax
3613:         ny             = mat->num_rows;
3614:       }
3615: #endif
3616:     } else {
3617:       /* z = A^T x + beta y
3618:          If A is compressed, then we need a work vector as the shorter version of x to compute A^T x.
3619:          Note A^Tx is of full length, so we set beta to 1.0 if y exists.
3620:        */
3621:       xptr = compressed ? cusparsestruct->workVector->data().get() : xarray;
3622:       dptr = zarray;
3623:       beta = yy ? matstruct->beta_one : matstruct->beta_zero;
3624:       if (compressed) { /* Scatter x to work vector */
3625:         thrust::device_ptr<PetscScalar> xarr = thrust::device_pointer_cast(xarray);

3627:         thrust::for_each(
3628: #if PetscDefined(HAVE_THRUST_ASYNC)
3629:           thrust::cuda::par.on(PetscDefaultCudaStream),
3630: #endif
3631:           thrust::make_zip_iterator(thrust::make_tuple(cusparsestruct->workVector->begin(), thrust::make_permutation_iterator(xarr, matstruct->cprowIndices->begin()))),
3632:           thrust::make_zip_iterator(thrust::make_tuple(cusparsestruct->workVector->begin(), thrust::make_permutation_iterator(xarr, matstruct->cprowIndices->begin()))) + matstruct->cprowIndices->size(), VecCUDAEqualsReverse());
3633:       }
3634: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3635:       if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3636:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3637:         nx             = mat->num_rows; // since y = A^T x
3638:         ny             = mat->num_cols;
3639:       }
3640: #endif
3641:     }

3643:     /* csr_spmv does y = alpha op(A) x + beta y */
3644:     if (cusparsestruct->format == MAT_CUSPARSE_CSR) {
3645: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3646:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
3647:       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.
3648:   #else
3649:       cusparseSpMatDescr_t &matDescr = matstruct->matDescr;
3650:   #endif

3652:       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");
3653:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
3654:       if (!matDescr) {
3655:         CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3656:         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));
3657:       }
3658:   #endif

3660:       if (!matstruct->cuSpMV[opA].initialized) { /* built on demand */
3661:         PetscCallCUSPARSE(cusparseCreateDnVec(&matstruct->cuSpMV[opA].vecXDescr, nx, xptr, cusparse_scalartype));
3662:         PetscCallCUSPARSE(cusparseCreateDnVec(&matstruct->cuSpMV[opA].vecYDescr, ny, dptr, cusparse_scalartype));
3663:         PetscCallCUSPARSE(
3664:           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));
3665:         PetscCallCUDA(cudaMalloc(&matstruct->cuSpMV[opA].spmvBuffer, matstruct->cuSpMV[opA].spmvBufferSize));
3666:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0) // cusparseSpMV_preprocess is added in 12.4
3667:         PetscCallCUSPARSE(
3668:           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));
3669:   #endif
3670:         matstruct->cuSpMV[opA].initialized = PETSC_TRUE;
3671:       } else {
3672:         /* x, y's value pointers might change between calls, but their shape is kept, so we just update pointers */
3673:         PetscCallCUSPARSE(cusparseDnVecSetValues(matstruct->cuSpMV[opA].vecXDescr, xptr));
3674:         PetscCallCUSPARSE(cusparseDnVecSetValues(matstruct->cuSpMV[opA].vecYDescr, dptr));
3675:       }

3677:       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));
3678: #else
3679:       CsrMatrix *mat = (CsrMatrix *)matstruct->mat;
3680:       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));
3681: #endif
3682:     } else {
3683:       if (cusparsestruct->nrows) {
3684: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3685:         SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
3686: #else
3687:         cusparseHybMat_t hybMat = (cusparseHybMat_t)matstruct->mat;
3688:         PetscCallCUSPARSE(cusparse_hyb_spmv(cusparsestruct->handle, opA, matstruct->alpha_one, matstruct->descr, hybMat, xptr, beta, dptr));
3689: #endif
3690:       }
3691:     }
3692:     PetscCall(PetscLogGpuTimeEnd());

3694:     if (opA == CUSPARSE_OPERATION_NON_TRANSPOSE) {
3695:       if (yy) {                                      /* MatMultAdd: zz = A*xx + yy */
3696:         if (compressed) {                            /* A is compressed. We first copy yy to zz, then ScatterAdd the work vector to zz */
3697:           PetscCall(VecSeq_CUDA::Copy(yy, zz));      /* zz = yy */
3698:         } else if (zz != yy) {                       /* A is not compressed. zz already contains A*xx, and we just need to add yy */
3699:           PetscCall(VecSeq_CUDA::AXPY(zz, 1.0, yy)); /* zz += yy */
3700:         }
3701:       } else if (compressed) { /* MatMult: zz = A*xx. A is compressed, so we zero zz first, then ScatterAdd the work vector to zz */
3702:         PetscCall(VecSeq_CUDA::Set(zz, 0));
3703:       }

3705:       /* ScatterAdd the result from work vector into the full vector when A is compressed */
3706:       if (compressed) {
3707:         PetscCall(PetscLogGpuTimeBegin());
3708:         PetscInt n = (PetscInt)matstruct->cprowIndices->size();
3709:         ScatterAdd<<<(int)((n + 255) / 256), 256, 0, PetscDefaultCudaStream>>>(n, matstruct->cprowIndices->data().get(), cusparsestruct->workVector->data().get(), zarray);
3710:         PetscCall(PetscLogGpuTimeEnd());
3711:       }
3712:     } else {
3713:       if (yy && yy != zz) PetscCall(VecSeq_CUDA::AXPY(zz, 1.0, yy)); /* zz += yy */
3714:     }
3715:     PetscCall(VecCUDARestoreArrayRead(xx, (const PetscScalar **)&xarray));
3716:     if (yy == zz) PetscCall(VecCUDARestoreArray(zz, &zarray));
3717:     else PetscCall(VecCUDARestoreArrayWrite(zz, &zarray));
3718:   } catch (char *ex) {
3719:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "CUSPARSE error: %s", ex);
3720:   }
3721:   if (yy) {
3722:     PetscCall(PetscLogGpuFlops(2.0 * a->nz));
3723:   } else {
3724:     PetscCall(PetscLogGpuFlops(2.0 * a->nz - a->nonzerorowcnt));
3725:   }
3726:   PetscFunctionReturn(PETSC_SUCCESS);
3727: }

3729: static PetscErrorCode MatMultTransposeAdd_SeqAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
3730: {
3731:   PetscFunctionBegin;
3732:   PetscCall(MatMultAddKernel_SeqAIJCUSPARSE(A, xx, yy, zz, PETSC_TRUE, PETSC_FALSE));
3733:   PetscFunctionReturn(PETSC_SUCCESS);
3734: }

3736: static PetscErrorCode MatAssemblyEnd_SeqAIJCUSPARSE(Mat A, MatAssemblyType mode)
3737: {
3738:   PetscFunctionBegin;
3739:   PetscCall(MatAssemblyEnd_SeqAIJ(A, mode));
3740:   PetscFunctionReturn(PETSC_SUCCESS);
3741: }

3743: /*@
3744:   MatCreateSeqAIJCUSPARSE - Creates a sparse matrix in `MATAIJCUSPARSE` (compressed row) format for use on NVIDIA GPUs

3746:   Collective

3748:   Input Parameters:
3749: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3750: . m    - number of rows
3751: . n    - number of columns
3752: . nz   - number of nonzeros per row (same for all rows), ignored if `nnz` is provide
3753: - nnz  - array containing the number of nonzeros in the various rows (possibly different for each row) or `NULL`

3755:   Output Parameter:
3756: . A - the matrix

3758:   Level: intermediate

3760:   Notes:
3761:   This matrix will ultimately pushed down to NVIDIA GPUs and use the CuSPARSE library for
3762:   calculations. For good matrix assembly performance the user should preallocate the matrix
3763:   storage by setting the parameter `nz` (or the array `nnz`).

3765:   It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
3766:   MatXXXXSetPreallocation() paradgm instead of this routine directly.
3767:   [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]

3769:   The AIJ format, also called
3770:   compressed row storage, is fully compatible with standard Fortran
3771:   storage.  That is, the stored row and column indices can begin at
3772:   either one (as in Fortran) or zero.

3774:   Specify the preallocated storage with either nz or nnz (not both).
3775:   Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3776:   allocation.

3778:   When working with matrices for GPUs, it is often better to use the `MatSetPreallocationCOO()` and `MatSetValuesCOO()` paradigm rather than using this routine and `MatSetValues()`

3780: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJCUSPARSE`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MATAIJCUSPARSE`,
3781:           `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
3782: @*/
3783: PetscErrorCode MatCreateSeqAIJCUSPARSE(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3784: {
3785:   PetscFunctionBegin;
3786:   PetscCall(MatCreate(comm, A));
3787:   PetscCall(MatSetSizes(*A, m, n, m, n));
3788:   PetscCall(MatSetType(*A, MATSEQAIJCUSPARSE));
3789:   PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, (PetscInt *)nnz));
3790:   PetscFunctionReturn(PETSC_SUCCESS);
3791: }

3793: static PetscErrorCode MatDestroy_SeqAIJCUSPARSE(Mat A)
3794: {
3795:   PetscFunctionBegin;
3796:   if (A->factortype == MAT_FACTOR_NONE) {
3797:     PetscCall(MatSeqAIJCUSPARSE_Destroy(A));
3798:   } else {
3799:     PetscCall(MatSeqAIJCUSPARSETriFactors_Destroy((Mat_SeqAIJCUSPARSETriFactors **)&A->spptr));
3800:   }
3801:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", NULL));
3802:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatCUSPARSESetFormat_C", NULL));
3803:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatCUSPARSESetUseCPUSolve_C", NULL));
3804:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", NULL));
3805:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", NULL));
3806:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", NULL));
3807:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
3808:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
3809:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
3810:   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijcusparse_hypre_C", NULL));
3811:   PetscCall(MatDestroy_SeqAIJ(A));
3812:   PetscFunctionReturn(PETSC_SUCCESS);
3813: }

3815: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
3816: static PetscErrorCode       MatBindToCPU_SeqAIJCUSPARSE(Mat, PetscBool);
3817: static PetscErrorCode       MatDuplicate_SeqAIJCUSPARSE(Mat A, MatDuplicateOption cpvalues, Mat *B)
3818: {
3819:   PetscFunctionBegin;
3820:   PetscCall(MatDuplicate_SeqAIJ(A, cpvalues, B));
3821:   PetscCall(MatConvert_SeqAIJ_SeqAIJCUSPARSE(*B, MATSEQAIJCUSPARSE, MAT_INPLACE_MATRIX, B));
3822:   PetscFunctionReturn(PETSC_SUCCESS);
3823: }

3825: static PetscErrorCode MatAXPY_SeqAIJCUSPARSE(Mat Y, PetscScalar a, Mat X, MatStructure str)
3826: {
3827:   Mat_SeqAIJ         *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data;
3828:   Mat_SeqAIJCUSPARSE *cy;
3829:   Mat_SeqAIJCUSPARSE *cx;
3830:   PetscScalar        *ay;
3831:   const PetscScalar  *ax;
3832:   CsrMatrix          *csry, *csrx;

3834:   PetscFunctionBegin;
3835:   cy = (Mat_SeqAIJCUSPARSE *)Y->spptr;
3836:   cx = (Mat_SeqAIJCUSPARSE *)X->spptr;
3837:   if (X->ops->axpy != Y->ops->axpy) {
3838:     PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(Y, PETSC_FALSE));
3839:     PetscCall(MatAXPY_SeqAIJ(Y, a, X, str));
3840:     PetscFunctionReturn(PETSC_SUCCESS);
3841:   }
3842:   /* if we are here, it means both matrices are bound to GPU */
3843:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(Y));
3844:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(X));
3845:   PetscCheck(cy->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)Y), PETSC_ERR_GPU, "only MAT_CUSPARSE_CSR supported");
3846:   PetscCheck(cx->format == MAT_CUSPARSE_CSR, PetscObjectComm((PetscObject)X), PETSC_ERR_GPU, "only MAT_CUSPARSE_CSR supported");
3847:   csry = (CsrMatrix *)cy->mat->mat;
3848:   csrx = (CsrMatrix *)cx->mat->mat;
3849:   /* see if we can turn this into a cublas axpy */
3850:   if (str != SAME_NONZERO_PATTERN && x->nz == y->nz && !x->compressedrow.use && !y->compressedrow.use) {
3851:     bool eq = thrust::equal(thrust::device, csry->row_offsets->begin(), csry->row_offsets->end(), csrx->row_offsets->begin());
3852:     if (eq) eq = thrust::equal(thrust::device, csry->column_indices->begin(), csry->column_indices->end(), csrx->column_indices->begin());
3853:     if (eq) str = SAME_NONZERO_PATTERN;
3854:   }
3855:   /* spgeam is buggy with one column */
3856:   if (Y->cmap->n == 1 && str != SAME_NONZERO_PATTERN) str = DIFFERENT_NONZERO_PATTERN;

3858:   if (str == SUBSET_NONZERO_PATTERN) {
3859:     PetscScalar b = 1.0;
3860: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3861:     size_t bufferSize;
3862:     void  *buffer;
3863: #endif

3865:     PetscCall(MatSeqAIJCUSPARSEGetArrayRead(X, &ax));
3866:     PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3867:     PetscCallCUSPARSE(cusparseSetPointerMode(cy->handle, CUSPARSE_POINTER_MODE_HOST));
3868: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
3869:     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(),
3870:                                                      csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get(), &bufferSize));
3871:     PetscCallCUDA(cudaMalloc(&buffer, bufferSize));
3872:     PetscCall(PetscLogGpuTimeBegin());
3873:     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(),
3874:                                           csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get(), buffer));
3875:     PetscCall(PetscLogGpuFlops(x->nz + y->nz));
3876:     PetscCall(PetscLogGpuTimeEnd());
3877:     PetscCallCUDA(cudaFree(buffer));
3878: #else
3879:     PetscCall(PetscLogGpuTimeBegin());
3880:     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(),
3881:                                           csry->column_indices->data().get(), cy->mat->descr, ay, csry->row_offsets->data().get(), csry->column_indices->data().get()));
3882:     PetscCall(PetscLogGpuFlops(x->nz + y->nz));
3883:     PetscCall(PetscLogGpuTimeEnd());
3884: #endif
3885:     PetscCallCUSPARSE(cusparseSetPointerMode(cy->handle, CUSPARSE_POINTER_MODE_DEVICE));
3886:     PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(X, &ax));
3887:     PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3888:     PetscCall(MatSeqAIJInvalidateDiagonal(Y));
3889:   } else if (str == SAME_NONZERO_PATTERN) {
3890:     cublasHandle_t cublasv2handle;
3891:     PetscBLASInt   one = 1, bnz = 1;

3893:     PetscCall(MatSeqAIJCUSPARSEGetArrayRead(X, &ax));
3894:     PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3895:     PetscCall(PetscCUBLASGetHandle(&cublasv2handle));
3896:     PetscCall(PetscBLASIntCast(x->nz, &bnz));
3897:     PetscCall(PetscLogGpuTimeBegin());
3898:     PetscCallCUBLAS(cublasXaxpy(cublasv2handle, bnz, &a, ax, one, ay, one));
3899:     PetscCall(PetscLogGpuFlops(2.0 * bnz));
3900:     PetscCall(PetscLogGpuTimeEnd());
3901:     PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(X, &ax));
3902:     PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3903:     PetscCall(MatSeqAIJInvalidateDiagonal(Y));
3904:   } else {
3905:     PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(Y, PETSC_FALSE));
3906:     PetscCall(MatAXPY_SeqAIJ(Y, a, X, str));
3907:   }
3908:   PetscFunctionReturn(PETSC_SUCCESS);
3909: }

3911: static PetscErrorCode MatScale_SeqAIJCUSPARSE(Mat Y, PetscScalar a)
3912: {
3913:   Mat_SeqAIJ    *y = (Mat_SeqAIJ *)Y->data;
3914:   PetscScalar   *ay;
3915:   cublasHandle_t cublasv2handle;
3916:   PetscBLASInt   one = 1, bnz = 1;

3918:   PetscFunctionBegin;
3919:   PetscCall(MatSeqAIJCUSPARSEGetArray(Y, &ay));
3920:   PetscCall(PetscCUBLASGetHandle(&cublasv2handle));
3921:   PetscCall(PetscBLASIntCast(y->nz, &bnz));
3922:   PetscCall(PetscLogGpuTimeBegin());
3923:   PetscCallCUBLAS(cublasXscal(cublasv2handle, bnz, &a, ay, one));
3924:   PetscCall(PetscLogGpuFlops(bnz));
3925:   PetscCall(PetscLogGpuTimeEnd());
3926:   PetscCall(MatSeqAIJCUSPARSERestoreArray(Y, &ay));
3927:   PetscCall(MatSeqAIJInvalidateDiagonal(Y));
3928:   PetscFunctionReturn(PETSC_SUCCESS);
3929: }

3931: static PetscErrorCode MatZeroEntries_SeqAIJCUSPARSE(Mat A)
3932: {
3933:   PetscBool   both = PETSC_FALSE;
3934:   Mat_SeqAIJ *a    = (Mat_SeqAIJ *)A->data;

3936:   PetscFunctionBegin;
3937:   if (A->factortype == MAT_FACTOR_NONE) {
3938:     Mat_SeqAIJCUSPARSE *spptr = (Mat_SeqAIJCUSPARSE *)A->spptr;
3939:     if (spptr->mat) {
3940:       CsrMatrix *matrix = (CsrMatrix *)spptr->mat->mat;
3941:       if (matrix->values) {
3942:         both = PETSC_TRUE;
3943:         thrust::fill(thrust::device, matrix->values->begin(), matrix->values->end(), 0.);
3944:       }
3945:     }
3946:     if (spptr->matTranspose) {
3947:       CsrMatrix *matrix = (CsrMatrix *)spptr->matTranspose->mat;
3948:       if (matrix->values) thrust::fill(thrust::device, matrix->values->begin(), matrix->values->end(), 0.);
3949:     }
3950:   }
3951:   PetscCall(PetscArrayzero(a->a, a->i[A->rmap->n]));
3952:   PetscCall(MatSeqAIJInvalidateDiagonal(A));
3953:   if (both) A->offloadmask = PETSC_OFFLOAD_BOTH;
3954:   else A->offloadmask = PETSC_OFFLOAD_CPU;
3955:   PetscFunctionReturn(PETSC_SUCCESS);
3956: }

3958: static PetscErrorCode MatGetCurrentMemType_SeqAIJCUSPARSE(PETSC_UNUSED Mat A, PetscMemType *m)
3959: {
3960:   PetscFunctionBegin;
3961:   *m = PETSC_MEMTYPE_CUDA;
3962:   PetscFunctionReturn(PETSC_SUCCESS);
3963: }

3965: static PetscErrorCode MatBindToCPU_SeqAIJCUSPARSE(Mat A, PetscBool flg)
3966: {
3967:   Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;

3969:   PetscFunctionBegin;
3970:   if (A->factortype != MAT_FACTOR_NONE) {
3971:     A->boundtocpu = flg;
3972:     PetscFunctionReturn(PETSC_SUCCESS);
3973:   }
3974:   if (flg) {
3975:     PetscCall(MatSeqAIJCUSPARSECopyFromGPU(A));

3977:     A->ops->scale                     = MatScale_SeqAIJ;
3978:     A->ops->axpy                      = MatAXPY_SeqAIJ;
3979:     A->ops->zeroentries               = MatZeroEntries_SeqAIJ;
3980:     A->ops->mult                      = MatMult_SeqAIJ;
3981:     A->ops->multadd                   = MatMultAdd_SeqAIJ;
3982:     A->ops->multtranspose             = MatMultTranspose_SeqAIJ;
3983:     A->ops->multtransposeadd          = MatMultTransposeAdd_SeqAIJ;
3984:     A->ops->multhermitiantranspose    = NULL;
3985:     A->ops->multhermitiantransposeadd = NULL;
3986:     A->ops->productsetfromoptions     = MatProductSetFromOptions_SeqAIJ;
3987:     A->ops->getcurrentmemtype         = NULL;
3988:     PetscCall(PetscMemzero(a->ops, sizeof(Mat_SeqAIJOps)));
3989:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", NULL));
3990:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", NULL));
3991:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", NULL));
3992:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
3993:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
3994:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", NULL));
3995:   } else {
3996:     A->ops->scale                     = MatScale_SeqAIJCUSPARSE;
3997:     A->ops->axpy                      = MatAXPY_SeqAIJCUSPARSE;
3998:     A->ops->zeroentries               = MatZeroEntries_SeqAIJCUSPARSE;
3999:     A->ops->mult                      = MatMult_SeqAIJCUSPARSE;
4000:     A->ops->multadd                   = MatMultAdd_SeqAIJCUSPARSE;
4001:     A->ops->multtranspose             = MatMultTranspose_SeqAIJCUSPARSE;
4002:     A->ops->multtransposeadd          = MatMultTransposeAdd_SeqAIJCUSPARSE;
4003:     A->ops->multhermitiantranspose    = MatMultHermitianTranspose_SeqAIJCUSPARSE;
4004:     A->ops->multhermitiantransposeadd = MatMultHermitianTransposeAdd_SeqAIJCUSPARSE;
4005:     A->ops->productsetfromoptions     = MatProductSetFromOptions_SeqAIJCUSPARSE;
4006:     A->ops->getcurrentmemtype         = MatGetCurrentMemType_SeqAIJCUSPARSE;
4007:     a->ops->getarray                  = MatSeqAIJGetArray_SeqAIJCUSPARSE;
4008:     a->ops->restorearray              = MatSeqAIJRestoreArray_SeqAIJCUSPARSE;
4009:     a->ops->getarrayread              = MatSeqAIJGetArrayRead_SeqAIJCUSPARSE;
4010:     a->ops->restorearrayread          = MatSeqAIJRestoreArrayRead_SeqAIJCUSPARSE;
4011:     a->ops->getarraywrite             = MatSeqAIJGetArrayWrite_SeqAIJCUSPARSE;
4012:     a->ops->restorearraywrite         = MatSeqAIJRestoreArrayWrite_SeqAIJCUSPARSE;
4013:     a->ops->getcsrandmemtype          = MatSeqAIJGetCSRAndMemType_SeqAIJCUSPARSE;

4015:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", MatSeqAIJCopySubArray_SeqAIJCUSPARSE));
4016:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdensecuda_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4017:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqdense_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4018:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJCUSPARSE));
4019:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJCUSPARSE));
4020:     PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJCUSPARSE));
4021:   }
4022:   A->boundtocpu = flg;
4023:   if (flg && a->inode.size_csr) {
4024:     a->inode.use = PETSC_TRUE;
4025:   } else {
4026:     a->inode.use = PETSC_FALSE;
4027:   }
4028:   PetscFunctionReturn(PETSC_SUCCESS);
4029: }

4031: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat A, MatType, MatReuse reuse, Mat *newmat)
4032: {
4033:   Mat B;

4035:   PetscFunctionBegin;
4036:   PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUDA)); /* first use of CUSPARSE may be via MatConvert */
4037:   if (reuse == MAT_INITIAL_MATRIX) {
4038:     PetscCall(MatDuplicate(A, MAT_COPY_VALUES, newmat));
4039:   } else if (reuse == MAT_REUSE_MATRIX) {
4040:     PetscCall(MatCopy(A, *newmat, SAME_NONZERO_PATTERN));
4041:   }
4042:   B = *newmat;

4044:   PetscCall(PetscFree(B->defaultvectype));
4045:   PetscCall(PetscStrallocpy(VECCUDA, &B->defaultvectype));

4047:   if (reuse != MAT_REUSE_MATRIX && !B->spptr) {
4048:     if (B->factortype == MAT_FACTOR_NONE) {
4049:       Mat_SeqAIJCUSPARSE *spptr;
4050:       PetscCall(PetscNew(&spptr));
4051:       PetscCallCUSPARSE(cusparseCreate(&spptr->handle));
4052:       PetscCallCUSPARSE(cusparseSetStream(spptr->handle, PetscDefaultCudaStream));
4053:       spptr->format = MAT_CUSPARSE_CSR;
4054: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4055:   #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
4056:       spptr->spmvAlg = CUSPARSE_SPMV_CSR_ALG1; /* default, since we only support csr */
4057:   #else
4058:       spptr->spmvAlg = CUSPARSE_CSRMV_ALG1; /* default, since we only support csr */
4059:   #endif
4060:       spptr->spmmAlg    = CUSPARSE_SPMM_CSR_ALG1; /* default, only support column-major dense matrix B */
4061:       spptr->csr2cscAlg = CUSPARSE_CSR2CSC_ALG1;
4062: #endif
4063:       B->spptr = spptr;
4064:     } else {
4065:       Mat_SeqAIJCUSPARSETriFactors *spptr;

4067:       PetscCall(PetscNew(&spptr));
4068:       PetscCallCUSPARSE(cusparseCreate(&spptr->handle));
4069:       PetscCallCUSPARSE(cusparseSetStream(spptr->handle, PetscDefaultCudaStream));
4070:       B->spptr = spptr;
4071:     }
4072:     B->offloadmask = PETSC_OFFLOAD_UNALLOCATED;
4073:   }
4074:   B->ops->assemblyend       = MatAssemblyEnd_SeqAIJCUSPARSE;
4075:   B->ops->destroy           = MatDestroy_SeqAIJCUSPARSE;
4076:   B->ops->setoption         = MatSetOption_SeqAIJCUSPARSE;
4077:   B->ops->setfromoptions    = MatSetFromOptions_SeqAIJCUSPARSE;
4078:   B->ops->bindtocpu         = MatBindToCPU_SeqAIJCUSPARSE;
4079:   B->ops->duplicate         = MatDuplicate_SeqAIJCUSPARSE;
4080:   B->ops->getcurrentmemtype = MatGetCurrentMemType_SeqAIJCUSPARSE;

4082:   PetscCall(MatBindToCPU_SeqAIJCUSPARSE(B, PETSC_FALSE));
4083:   PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJCUSPARSE));
4084:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatCUSPARSESetFormat_C", MatCUSPARSESetFormat_SeqAIJCUSPARSE));
4085: #if defined(PETSC_HAVE_HYPRE)
4086:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaijcusparse_hypre_C", MatConvert_AIJ_HYPRE));
4087: #endif
4088:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatCUSPARSESetUseCPUSolve_C", MatCUSPARSESetUseCPUSolve_SeqAIJCUSPARSE));
4089:   PetscFunctionReturn(PETSC_SUCCESS);
4090: }

4092: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJCUSPARSE(Mat B)
4093: {
4094:   PetscFunctionBegin;
4095:   PetscCall(MatCreate_SeqAIJ(B));
4096:   PetscCall(MatConvert_SeqAIJ_SeqAIJCUSPARSE(B, MATSEQAIJCUSPARSE, MAT_INPLACE_MATRIX, &B));
4097:   PetscFunctionReturn(PETSC_SUCCESS);
4098: }

4100: /*MC
4101:    MATSEQAIJCUSPARSE - MATAIJCUSPARSE = "(seq)aijcusparse" - A matrix type to be used for sparse matrices on NVIDIA GPUs.

4103:    Options Database Keys:
4104: +  -mat_type aijcusparse                 - Sets the matrix type to "seqaijcusparse" during a call to `MatSetFromOptions()`
4105: .  -mat_cusparse_storage_format csr      - Sets the storage format of matrices (for `MatMult()` and factors in `MatSolve()`).
4106:                                            Other options include ell (ellpack) or hyb (hybrid).
4107: .  -mat_cusparse_mult_storage_format csr - Sets the storage format of matrices (for `MatMult()`). Other options include ell (ellpack) or hyb (hybrid).
4108: -  -mat_cusparse_use_cpu_solve           - Performs the `MatSolve()` on the CPU

4110:   Level: beginner

4112:   Notes:
4113:   These matrices can be in either CSR, ELL, or HYB format.

4115:   All matrix calculations are performed on NVIDIA GPUs using the cuSPARSE library.

4117:   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
4118:   if some integer values passed in do not fit in `int`.

4120: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJCUSPARSE()`, `MatCUSPARSESetUseCPUSolve()`, `MATAIJCUSPARSE`, `MatCreateAIJCUSPARSE()`, `MatCUSPARSESetFormat()`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
4121: M*/

4123: PETSC_INTERN PetscErrorCode MatSolverTypeRegister_CUSPARSE(void)
4124: {
4125:   PetscFunctionBegin;
4126:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_LU, MatGetFactor_seqaijcusparse_cusparse));
4127:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_CHOLESKY, MatGetFactor_seqaijcusparse_cusparse));
4128:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_ILU, MatGetFactor_seqaijcusparse_cusparse));
4129:   PetscCall(MatSolverTypeRegister(MATSOLVERCUSPARSE, MATSEQAIJCUSPARSE, MAT_FACTOR_ICC, MatGetFactor_seqaijcusparse_cusparse));
4130:   PetscFunctionReturn(PETSC_SUCCESS);
4131: }

4133: static PetscErrorCode MatSeqAIJCUSPARSE_Destroy(Mat mat)
4134: {
4135:   Mat_SeqAIJCUSPARSE *cusp = static_cast<Mat_SeqAIJCUSPARSE *>(mat->spptr);

4137:   PetscFunctionBegin;
4138:   if (cusp) {
4139:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->mat, cusp->format));
4140:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->matTranspose, cusp->format));
4141:     delete cusp->workVector;
4142:     delete cusp->rowoffsets_gpu;
4143:     delete cusp->csr2csc_i;
4144:     delete cusp->coords;
4145:     if (cusp->handle) PetscCallCUSPARSE(cusparseDestroy(cusp->handle));
4146:     PetscCall(PetscFree(mat->spptr));
4147:   }
4148:   PetscFunctionReturn(PETSC_SUCCESS);
4149: }

4151: static PetscErrorCode CsrMatrix_Destroy(CsrMatrix **mat)
4152: {
4153:   PetscFunctionBegin;
4154:   if (*mat) {
4155:     delete (*mat)->values;
4156:     delete (*mat)->column_indices;
4157:     delete (*mat)->row_offsets;
4158:     delete *mat;
4159:     *mat = 0;
4160:   }
4161:   PetscFunctionReturn(PETSC_SUCCESS);
4162: }

4164: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
4165: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSETriFactorStruct **trifactor)
4166: {
4167:   PetscFunctionBegin;
4168:   if (*trifactor) {
4169:     if ((*trifactor)->descr) PetscCallCUSPARSE(cusparseDestroyMatDescr((*trifactor)->descr));
4170:     if ((*trifactor)->solveInfo) PetscCallCUSPARSE(cusparseDestroyCsrsvInfo((*trifactor)->solveInfo));
4171:     PetscCall(CsrMatrix_Destroy(&(*trifactor)->csrMat));
4172:     if ((*trifactor)->solveBuffer) PetscCallCUDA(cudaFree((*trifactor)->solveBuffer));
4173:     if ((*trifactor)->AA_h) PetscCallCUDA(cudaFreeHost((*trifactor)->AA_h));
4174:   #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4175:     if ((*trifactor)->csr2cscBuffer) PetscCallCUDA(cudaFree((*trifactor)->csr2cscBuffer));
4176:   #endif
4177:     PetscCall(PetscFree(*trifactor));
4178:   }
4179:   PetscFunctionReturn(PETSC_SUCCESS);
4180: }
4181: #endif

4183: static PetscErrorCode MatSeqAIJCUSPARSEMultStruct_Destroy(Mat_SeqAIJCUSPARSEMultStruct **matstruct, MatCUSPARSEStorageFormat format)
4184: {
4185:   CsrMatrix *mat;

4187:   PetscFunctionBegin;
4188:   if (*matstruct) {
4189:     if ((*matstruct)->mat) {
4190:       if (format == MAT_CUSPARSE_ELL || format == MAT_CUSPARSE_HYB) {
4191: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4192:         SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_CUSPARSE_ELL and MAT_CUSPARSE_HYB are not supported since CUDA-11.0");
4193: #else
4194:         cusparseHybMat_t hybMat = (cusparseHybMat_t)(*matstruct)->mat;
4195:         PetscCallCUSPARSE(cusparseDestroyHybMat(hybMat));
4196: #endif
4197:       } else {
4198:         mat = (CsrMatrix *)(*matstruct)->mat;
4199:         PetscCall(CsrMatrix_Destroy(&mat));
4200:       }
4201:     }
4202:     if ((*matstruct)->descr) PetscCallCUSPARSE(cusparseDestroyMatDescr((*matstruct)->descr));
4203:     delete (*matstruct)->cprowIndices;
4204:     if ((*matstruct)->alpha_one) PetscCallCUDA(cudaFree((*matstruct)->alpha_one));
4205:     if ((*matstruct)->beta_zero) PetscCallCUDA(cudaFree((*matstruct)->beta_zero));
4206:     if ((*matstruct)->beta_one) PetscCallCUDA(cudaFree((*matstruct)->beta_one));

4208: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4209:     Mat_SeqAIJCUSPARSEMultStruct *mdata = *matstruct;
4210:     if (mdata->matDescr) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr));

4212:     for (int i = 0; i < 3; i++) {
4213:       if (mdata->cuSpMV[i].initialized) {
4214:         PetscCallCUDA(cudaFree(mdata->cuSpMV[i].spmvBuffer));
4215:         PetscCallCUSPARSE(cusparseDestroyDnVec(mdata->cuSpMV[i].vecXDescr));
4216:         PetscCallCUSPARSE(cusparseDestroyDnVec(mdata->cuSpMV[i].vecYDescr));
4217:   #if PETSC_PKG_CUDA_VERSION_GE(12, 4, 0)
4218:         if (mdata->matDescr_SpMV[i]) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr_SpMV[i]));
4219:         if (mdata->matDescr_SpMM[i]) PetscCallCUSPARSE(cusparseDestroySpMat(mdata->matDescr_SpMM[i]));
4220:   #endif
4221:       }
4222:     }
4223: #endif
4224:     delete *matstruct;
4225:     *matstruct = NULL;
4226:   }
4227:   PetscFunctionReturn(PETSC_SUCCESS);
4228: }

4230: PetscErrorCode MatSeqAIJCUSPARSETriFactors_Reset(Mat_SeqAIJCUSPARSETriFactors_p *trifactors)
4231: {
4232:   Mat_SeqAIJCUSPARSETriFactors *fs = *trifactors;

4234:   PetscFunctionBegin;
4235:   if (fs) {
4236: #if PETSC_PKG_CUDA_VERSION_LT(11, 4, 0)
4237:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->loTriFactorPtr));
4238:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->upTriFactorPtr));
4239:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->loTriFactorPtrTranspose));
4240:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&fs->upTriFactorPtrTranspose));
4241:     delete fs->workVector;
4242:     fs->workVector = NULL;
4243: #endif
4244:     delete fs->rpermIndices;
4245:     delete fs->cpermIndices;
4246:     fs->rpermIndices  = NULL;
4247:     fs->cpermIndices  = NULL;
4248:     fs->init_dev_prop = PETSC_FALSE;
4249: #if PETSC_PKG_CUDA_VERSION_GE(11, 4, 0)
4250:     PetscCallCUDA(cudaFree(fs->csrRowPtr));
4251:     PetscCallCUDA(cudaFree(fs->csrColIdx));
4252:     PetscCallCUDA(cudaFree(fs->csrRowPtr32));
4253:     PetscCallCUDA(cudaFree(fs->csrColIdx32));
4254:     PetscCallCUDA(cudaFree(fs->csrVal));
4255:     PetscCallCUDA(cudaFree(fs->diag));
4256:     PetscCallCUDA(cudaFree(fs->X));
4257:     PetscCallCUDA(cudaFree(fs->Y));
4258:     // PetscCallCUDA(cudaFree(fs->factBuffer_M)); /* No needed since factBuffer_M shares with one of spsvBuffer_L/U */
4259:     PetscCallCUDA(cudaFree(fs->spsvBuffer_L));
4260:     PetscCallCUDA(cudaFree(fs->spsvBuffer_U));
4261:     PetscCallCUDA(cudaFree(fs->spsvBuffer_Lt));
4262:     PetscCallCUDA(cudaFree(fs->spsvBuffer_Ut));
4263:     PetscCallCUSPARSE(cusparseDestroyMatDescr(fs->matDescr_M));
4264:     PetscCallCUSPARSE(cusparseDestroySpMat(fs->spMatDescr_L));
4265:     PetscCallCUSPARSE(cusparseDestroySpMat(fs->spMatDescr_U));
4266:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_L));
4267:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_Lt));
4268:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_U));
4269:     PetscCallCUSPARSE(cusparseSpSV_destroyDescr(fs->spsvDescr_Ut));
4270:     PetscCallCUSPARSE(cusparseDestroyDnVec(fs->dnVecDescr_X));
4271:     PetscCallCUSPARSE(cusparseDestroyDnVec(fs->dnVecDescr_Y));
4272:     PetscCallCUSPARSE(cusparseDestroyCsrilu02Info(fs->ilu0Info_M));
4273:     PetscCallCUSPARSE(cusparseDestroyCsric02Info(fs->ic0Info_M));
4274:     PetscCall(PetscFree(fs->csrRowPtr_h));
4275:     PetscCall(PetscFree(fs->csrVal_h));
4276:     PetscCall(PetscFree(fs->diag_h));
4277:     fs->createdTransposeSpSVDescr    = PETSC_FALSE;
4278:     fs->updatedTransposeSpSVAnalysis = PETSC_FALSE;
4279: #endif
4280:   }
4281:   PetscFunctionReturn(PETSC_SUCCESS);
4282: }

4284: static PetscErrorCode MatSeqAIJCUSPARSETriFactors_Destroy(Mat_SeqAIJCUSPARSETriFactors **trifactors)
4285: {
4286:   PetscFunctionBegin;
4287:   if (*trifactors) {
4288:     PetscCall(MatSeqAIJCUSPARSETriFactors_Reset(trifactors));
4289:     PetscCallCUSPARSE(cusparseDestroy((*trifactors)->handle));
4290:     PetscCall(PetscFree(*trifactors));
4291:   }
4292:   PetscFunctionReturn(PETSC_SUCCESS);
4293: }

4295: struct IJCompare {
4296:   __host__ __device__ inline bool operator()(const thrust::tuple<PetscInt, PetscInt> &t1, const thrust::tuple<PetscInt, PetscInt> &t2)
4297:   {
4298:     if (thrust::get<0>(t1) < thrust::get<0>(t2)) return true;
4299:     if (thrust::get<0>(t1) == thrust::get<0>(t2)) return thrust::get<1>(t1) < thrust::get<1>(t2);
4300:     return false;
4301:   }
4302: };

4304: static PetscErrorCode MatSeqAIJCUSPARSEInvalidateTranspose(Mat A, PetscBool destroy)
4305: {
4306:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;

4308:   PetscFunctionBegin;
4309:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4310:   if (!cusp) PetscFunctionReturn(PETSC_SUCCESS);
4311:   if (destroy) {
4312:     PetscCall(MatSeqAIJCUSPARSEMultStruct_Destroy(&cusp->matTranspose, cusp->format));
4313:     delete cusp->csr2csc_i;
4314:     cusp->csr2csc_i = NULL;
4315:   }
4316:   A->transupdated = PETSC_FALSE;
4317:   PetscFunctionReturn(PETSC_SUCCESS);
4318: }

4320: static PetscErrorCode MatCOOStructDestroy_SeqAIJCUSPARSE(void **data)
4321: {
4322:   MatCOOStruct_SeqAIJ *coo = (MatCOOStruct_SeqAIJ *)*data;

4324:   PetscFunctionBegin;
4325:   PetscCallCUDA(cudaFree(coo->perm));
4326:   PetscCallCUDA(cudaFree(coo->jmap));
4327:   PetscCall(PetscFree(coo));
4328:   PetscFunctionReturn(PETSC_SUCCESS);
4329: }

4331: static PetscErrorCode MatSetPreallocationCOO_SeqAIJCUSPARSE(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
4332: {
4333:   PetscBool            dev_ij = PETSC_FALSE;
4334:   PetscMemType         mtype  = PETSC_MEMTYPE_HOST;
4335:   PetscInt            *i, *j;
4336:   PetscContainer       container_h;
4337:   MatCOOStruct_SeqAIJ *coo_h, *coo_d;

4339:   PetscFunctionBegin;
4340:   PetscCall(PetscGetMemType(coo_i, &mtype));
4341:   if (PetscMemTypeDevice(mtype)) {
4342:     dev_ij = PETSC_TRUE;
4343:     PetscCall(PetscMalloc2(coo_n, &i, coo_n, &j));
4344:     PetscCallCUDA(cudaMemcpy(i, coo_i, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4345:     PetscCallCUDA(cudaMemcpy(j, coo_j, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4346:   } else {
4347:     i = coo_i;
4348:     j = coo_j;
4349:   }

4351:   PetscCall(MatSetPreallocationCOO_SeqAIJ(mat, coo_n, i, j));
4352:   if (dev_ij) PetscCall(PetscFree2(i, j));
4353:   mat->offloadmask = PETSC_OFFLOAD_CPU;
4354:   // Create the GPU memory
4355:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(mat));

4357:   // Copy the COO struct to device
4358:   PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container_h));
4359:   PetscCall(PetscContainerGetPointer(container_h, (void **)&coo_h));
4360:   PetscCall(PetscMalloc1(1, &coo_d));
4361:   *coo_d = *coo_h; // do a shallow copy and then amend some fields that need to be different
4362:   PetscCallCUDA(cudaMalloc((void **)&coo_d->jmap, (coo_h->nz + 1) * sizeof(PetscCount)));
4363:   PetscCallCUDA(cudaMemcpy(coo_d->jmap, coo_h->jmap, (coo_h->nz + 1) * sizeof(PetscCount), cudaMemcpyHostToDevice));
4364:   PetscCallCUDA(cudaMalloc((void **)&coo_d->perm, coo_h->Atot * sizeof(PetscCount)));
4365:   PetscCallCUDA(cudaMemcpy(coo_d->perm, coo_h->perm, coo_h->Atot * sizeof(PetscCount), cudaMemcpyHostToDevice));

4367:   // Put the COO struct in a container and then attach that to the matrix
4368:   PetscCall(PetscObjectContainerCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Device", coo_d, MatCOOStructDestroy_SeqAIJCUSPARSE));
4369:   PetscFunctionReturn(PETSC_SUCCESS);
4370: }

4372: __global__ static void MatAddCOOValues(const PetscScalar kv[], PetscCount nnz, const PetscCount jmap[], const PetscCount perm[], InsertMode imode, PetscScalar a[])
4373: {
4374:   PetscCount       i         = blockIdx.x * blockDim.x + threadIdx.x;
4375:   const PetscCount grid_size = gridDim.x * blockDim.x;
4376:   for (; i < nnz; i += grid_size) {
4377:     PetscScalar sum = 0.0;
4378:     for (PetscCount k = jmap[i]; k < jmap[i + 1]; k++) sum += kv[perm[k]];
4379:     a[i] = (imode == INSERT_VALUES ? 0.0 : a[i]) + sum;
4380:   }
4381: }

4383: static PetscErrorCode MatSetValuesCOO_SeqAIJCUSPARSE(Mat A, const PetscScalar v[], InsertMode imode)
4384: {
4385:   Mat_SeqAIJ          *seq  = (Mat_SeqAIJ *)A->data;
4386:   Mat_SeqAIJCUSPARSE  *dev  = (Mat_SeqAIJCUSPARSE *)A->spptr;
4387:   PetscCount           Annz = seq->nz;
4388:   PetscMemType         memtype;
4389:   const PetscScalar   *v1 = v;
4390:   PetscScalar         *Aa;
4391:   PetscContainer       container;
4392:   MatCOOStruct_SeqAIJ *coo;

4394:   PetscFunctionBegin;
4395:   if (!dev->mat) PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));

4397:   PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Device", (PetscObject *)&container));
4398:   PetscCall(PetscContainerGetPointer(container, (void **)&coo));

4400:   PetscCall(PetscGetMemType(v, &memtype));
4401:   if (PetscMemTypeHost(memtype)) { /* If user gave v[] in host, we might need to copy it to device if any */
4402:     PetscCallCUDA(cudaMalloc((void **)&v1, coo->n * sizeof(PetscScalar)));
4403:     PetscCallCUDA(cudaMemcpy((void *)v1, v, coo->n * sizeof(PetscScalar), cudaMemcpyHostToDevice));
4404:   }

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

4409:   PetscCall(PetscLogGpuTimeBegin());
4410:   if (Annz) {
4411:     MatAddCOOValues<<<((int)(Annz + 255) / 256), 256>>>(v1, Annz, coo->jmap, coo->perm, imode, Aa);
4412:     PetscCallCUDA(cudaPeekAtLastError());
4413:   }
4414:   PetscCall(PetscLogGpuTimeEnd());

4416:   if (imode == INSERT_VALUES) PetscCall(MatSeqAIJCUSPARSERestoreArrayWrite(A, &Aa));
4417:   else PetscCall(MatSeqAIJCUSPARSERestoreArray(A, &Aa));

4419:   if (PetscMemTypeHost(memtype)) PetscCallCUDA(cudaFree((void *)v1));
4420:   PetscFunctionReturn(PETSC_SUCCESS);
4421: }

4423: /*@C
4424:   MatSeqAIJCUSPARSEGetIJ - returns the device row storage `i` and `j` indices for `MATSEQAIJCUSPARSE` matrices.

4426:   Not Collective

4428:   Input Parameters:
4429: + A          - the matrix
4430: - compressed - `PETSC_TRUE` or `PETSC_FALSE` indicating the matrix data structure should be always returned in compressed form

4432:   Output Parameters:
4433: + i - the CSR row pointers, these are always `int` even when PETSc is configured with `--with-64-bit-indices`
4434: - j - the CSR column indices, these are always `int` even when PETSc is configured with `--with-64-bit-indices`

4436:   Level: developer

4438:   Note:
4439:   When compressed is true, the CSR structure does not contain empty rows

4441: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSERestoreIJ()`, `MatSeqAIJCUSPARSEGetArrayRead()`
4442: @*/
4443: PetscErrorCode MatSeqAIJCUSPARSEGetIJ(Mat A, PetscBool compressed, const int **i, const int **j)
4444: {
4445:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4446:   CsrMatrix          *csr;
4447:   Mat_SeqAIJ         *a = (Mat_SeqAIJ *)A->data;

4449:   PetscFunctionBegin;
4451:   if (!i || !j) PetscFunctionReturn(PETSC_SUCCESS);
4452:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4453:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4454:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4455:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4456:   csr = (CsrMatrix *)cusp->mat->mat;
4457:   if (i) {
4458:     if (!compressed && a->compressedrow.use) { /* need full row offset */
4459:       if (!cusp->rowoffsets_gpu) {
4460:         cusp->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
4461:         cusp->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);
4462:         PetscCall(PetscLogCpuToGpu((A->rmap->n + 1) * sizeof(PetscInt)));
4463:       }
4464:       *i = cusp->rowoffsets_gpu->data().get();
4465:     } else *i = csr->row_offsets->data().get();
4466:   }
4467:   if (j) *j = csr->column_indices->data().get();
4468:   PetscFunctionReturn(PETSC_SUCCESS);
4469: }

4471: /*@C
4472:   MatSeqAIJCUSPARSERestoreIJ - restore the device row storage `i` and `j` indices obtained with `MatSeqAIJCUSPARSEGetIJ()`

4474:   Not Collective

4476:   Input Parameters:
4477: + A          - the matrix
4478: . compressed - `PETSC_TRUE` or `PETSC_FALSE` indicating the matrix data structure should be always returned in compressed form
4479: . i          - the CSR row pointers
4480: - j          - the CSR column indices

4482:   Level: developer

4484: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetIJ()`
4485: @*/
4486: PetscErrorCode MatSeqAIJCUSPARSERestoreIJ(Mat A, PetscBool compressed, const int **i, const int **j)
4487: {
4488:   PetscFunctionBegin;
4490:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4491:   if (i) *i = NULL;
4492:   if (j) *j = NULL;
4493:   (void)compressed;
4494:   PetscFunctionReturn(PETSC_SUCCESS);
4495: }

4497: /*@C
4498:   MatSeqAIJCUSPARSEGetArrayRead - gives read-only access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix nonzero entries are stored

4500:   Not Collective

4502:   Input Parameter:
4503: . A - a `MATSEQAIJCUSPARSE` matrix

4505:   Output Parameter:
4506: . a - pointer to the device data

4508:   Level: developer

4510:   Note:
4511:   Will trigger host-to-device copies if the most up-to-date matrix data is on the host

4513: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`, `MatSeqAIJCUSPARSEGetArrayWrite()`, `MatSeqAIJCUSPARSERestoreArrayRead()`
4514: @*/
4515: PetscErrorCode MatSeqAIJCUSPARSEGetArrayRead(Mat A, const PetscScalar **a)
4516: {
4517:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4518:   CsrMatrix          *csr;

4520:   PetscFunctionBegin;
4522:   PetscAssertPointer(a, 2);
4523:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4524:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4525:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4526:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4527:   csr = (CsrMatrix *)cusp->mat->mat;
4528:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4529:   *a = csr->values->data().get();
4530:   PetscFunctionReturn(PETSC_SUCCESS);
4531: }

4533: /*@C
4534:   MatSeqAIJCUSPARSERestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJCUSPARSEGetArrayRead()`

4536:   Not Collective

4538:   Input Parameters:
4539: + A - a `MATSEQAIJCUSPARSE` matrix
4540: - a - pointer to the device data

4542:   Level: developer

4544: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayRead()`
4545: @*/
4546: PetscErrorCode MatSeqAIJCUSPARSERestoreArrayRead(Mat A, const PetscScalar **a)
4547: {
4548:   PetscFunctionBegin;
4550:   PetscAssertPointer(a, 2);
4551:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4552:   *a = NULL;
4553:   PetscFunctionReturn(PETSC_SUCCESS);
4554: }

4556: /*@C
4557:   MatSeqAIJCUSPARSEGetArray - gives read-write access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix is stored

4559:   Not Collective

4561:   Input Parameter:
4562: . A - a `MATSEQAIJCUSPARSE` matrix

4564:   Output Parameter:
4565: . a - pointer to the device data

4567:   Level: developer

4569:   Note:
4570:   Will trigger host-to-device copies if the most up-to-date matrix data is on the host

4572: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayRead()`, `MatSeqAIJCUSPARSEGetArrayWrite()`, `MatSeqAIJCUSPARSERestoreArray()`
4573: @*/
4574: PetscErrorCode MatSeqAIJCUSPARSEGetArray(Mat A, PetscScalar **a)
4575: {
4576:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4577:   CsrMatrix          *csr;

4579:   PetscFunctionBegin;
4581:   PetscAssertPointer(a, 2);
4582:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4583:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4584:   PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4585:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4586:   csr = (CsrMatrix *)cusp->mat->mat;
4587:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4588:   *a             = csr->values->data().get();
4589:   A->offloadmask = PETSC_OFFLOAD_GPU;
4590:   PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
4591:   PetscFunctionReturn(PETSC_SUCCESS);
4592: }
4593: /*@C
4594:   MatSeqAIJCUSPARSERestoreArray - restore the read-write access array obtained from `MatSeqAIJCUSPARSEGetArray()`

4596:   Not Collective

4598:   Input Parameters:
4599: + A - a `MATSEQAIJCUSPARSE` matrix
4600: - a - pointer to the device data

4602:   Level: developer

4604: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`
4605: @*/
4606: PetscErrorCode MatSeqAIJCUSPARSERestoreArray(Mat A, PetscScalar **a)
4607: {
4608:   PetscFunctionBegin;
4610:   PetscAssertPointer(a, 2);
4611:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4612:   PetscCall(MatSeqAIJInvalidateDiagonal(A));
4613:   PetscCall(PetscObjectStateIncrease((PetscObject)A));
4614:   *a = NULL;
4615:   PetscFunctionReturn(PETSC_SUCCESS);
4616: }

4618: /*@C
4619:   MatSeqAIJCUSPARSEGetArrayWrite - gives write access to the array where the device data for a `MATSEQAIJCUSPARSE` matrix is stored

4621:   Not Collective

4623:   Input Parameter:
4624: . A - a `MATSEQAIJCUSPARSE` matrix

4626:   Output Parameter:
4627: . a - pointer to the device data

4629:   Level: developer

4631:   Note:
4632:   Does not trigger any host to device copies.

4634:   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

4636: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArray()`, `MatSeqAIJCUSPARSEGetArrayRead()`, `MatSeqAIJCUSPARSERestoreArrayWrite()`
4637: @*/
4638: PetscErrorCode MatSeqAIJCUSPARSEGetArrayWrite(Mat A, PetscScalar **a)
4639: {
4640:   Mat_SeqAIJCUSPARSE *cusp = (Mat_SeqAIJCUSPARSE *)A->spptr;
4641:   CsrMatrix          *csr;

4643:   PetscFunctionBegin;
4645:   PetscAssertPointer(a, 2);
4646:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4647:   PetscCheck(cusp->format != MAT_CUSPARSE_ELL && cusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4648:   PetscCheck(cusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4649:   csr = (CsrMatrix *)cusp->mat->mat;
4650:   PetscCheck(csr->values, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing CUDA memory");
4651:   *a             = csr->values->data().get();
4652:   A->offloadmask = PETSC_OFFLOAD_GPU;
4653:   PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(A, PETSC_FALSE));
4654:   PetscFunctionReturn(PETSC_SUCCESS);
4655: }

4657: /*@C
4658:   MatSeqAIJCUSPARSERestoreArrayWrite - restore the write-only access array obtained from `MatSeqAIJCUSPARSEGetArrayWrite()`

4660:   Not Collective

4662:   Input Parameters:
4663: + A - a `MATSEQAIJCUSPARSE` matrix
4664: - a - pointer to the device data

4666:   Level: developer

4668: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJCUSPARSEGetArrayWrite()`
4669: @*/
4670: PetscErrorCode MatSeqAIJCUSPARSERestoreArrayWrite(Mat A, PetscScalar **a)
4671: {
4672:   PetscFunctionBegin;
4674:   PetscAssertPointer(a, 2);
4675:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4676:   PetscCall(MatSeqAIJInvalidateDiagonal(A));
4677:   PetscCall(PetscObjectStateIncrease((PetscObject)A));
4678:   *a = NULL;
4679:   PetscFunctionReturn(PETSC_SUCCESS);
4680: }

4682: struct IJCompare4 {
4683:   __host__ __device__ inline bool operator()(const thrust::tuple<int, int, PetscScalar, int> &t1, const thrust::tuple<int, int, PetscScalar, int> &t2)
4684:   {
4685:     if (thrust::get<0>(t1) < thrust::get<0>(t2)) return true;
4686:     if (thrust::get<0>(t1) == thrust::get<0>(t2)) return thrust::get<1>(t1) < thrust::get<1>(t2);
4687:     return false;
4688:   }
4689: };

4691: struct Shift {
4692:   int _shift;

4694:   Shift(int shift) : _shift(shift) { }
4695:   __host__ __device__ inline int operator()(const int &c) { return c + _shift; }
4696: };

4698: /* merges two SeqAIJCUSPARSE matrices A, B by concatenating their rows. [A';B']' operation in MATLAB notation */
4699: PetscErrorCode MatSeqAIJCUSPARSEMergeMats(Mat A, Mat B, MatReuse reuse, Mat *C)
4700: {
4701:   Mat_SeqAIJ                   *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data, *c;
4702:   Mat_SeqAIJCUSPARSE           *Acusp = (Mat_SeqAIJCUSPARSE *)A->spptr, *Bcusp = (Mat_SeqAIJCUSPARSE *)B->spptr, *Ccusp;
4703:   Mat_SeqAIJCUSPARSEMultStruct *Cmat;
4704:   CsrMatrix                    *Acsr, *Bcsr, *Ccsr;
4705:   PetscInt                      Annz, Bnnz;
4706:   cusparseStatus_t              stat;
4707:   PetscInt                      i, m, n, zero = 0;

4709:   PetscFunctionBegin;
4712:   PetscAssertPointer(C, 4);
4713:   PetscCheckTypeName(A, MATSEQAIJCUSPARSE);
4714:   PetscCheckTypeName(B, MATSEQAIJCUSPARSE);
4715:   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);
4716:   PetscCheck(reuse != MAT_INPLACE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MAT_INPLACE_MATRIX not supported");
4717:   PetscCheck(Acusp->format != MAT_CUSPARSE_ELL && Acusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4718:   PetscCheck(Bcusp->format != MAT_CUSPARSE_ELL && Bcusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4719:   if (reuse == MAT_INITIAL_MATRIX) {
4720:     m = A->rmap->n;
4721:     n = A->cmap->n + B->cmap->n;
4722:     PetscCall(MatCreate(PETSC_COMM_SELF, C));
4723:     PetscCall(MatSetSizes(*C, m, n, m, n));
4724:     PetscCall(MatSetType(*C, MATSEQAIJCUSPARSE));
4725:     c                       = (Mat_SeqAIJ *)(*C)->data;
4726:     Ccusp                   = (Mat_SeqAIJCUSPARSE *)(*C)->spptr;
4727:     Cmat                    = new Mat_SeqAIJCUSPARSEMultStruct;
4728:     Ccsr                    = new CsrMatrix;
4729:     Cmat->cprowIndices      = NULL;
4730:     c->compressedrow.use    = PETSC_FALSE;
4731:     c->compressedrow.nrows  = 0;
4732:     c->compressedrow.i      = NULL;
4733:     c->compressedrow.rindex = NULL;
4734:     Ccusp->workVector       = NULL;
4735:     Ccusp->nrows            = m;
4736:     Ccusp->mat              = Cmat;
4737:     Ccusp->mat->mat         = Ccsr;
4738:     Ccsr->num_rows          = m;
4739:     Ccsr->num_cols          = n;
4740:     PetscCallCUSPARSE(cusparseCreateMatDescr(&Cmat->descr));
4741:     PetscCallCUSPARSE(cusparseSetMatIndexBase(Cmat->descr, CUSPARSE_INDEX_BASE_ZERO));
4742:     PetscCallCUSPARSE(cusparseSetMatType(Cmat->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
4743:     PetscCallCUDA(cudaMalloc((void **)&Cmat->alpha_one, sizeof(PetscScalar)));
4744:     PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_zero, sizeof(PetscScalar)));
4745:     PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_one, sizeof(PetscScalar)));
4746:     PetscCallCUDA(cudaMemcpy(Cmat->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4747:     PetscCallCUDA(cudaMemcpy(Cmat->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4748:     PetscCallCUDA(cudaMemcpy(Cmat->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4749:     PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4750:     PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));
4751:     PetscCheck(Acusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4752:     PetscCheck(Bcusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");

4754:     Acsr                 = (CsrMatrix *)Acusp->mat->mat;
4755:     Bcsr                 = (CsrMatrix *)Bcusp->mat->mat;
4756:     Annz                 = (PetscInt)Acsr->column_indices->size();
4757:     Bnnz                 = (PetscInt)Bcsr->column_indices->size();
4758:     c->nz                = Annz + Bnnz;
4759:     Ccsr->row_offsets    = new THRUSTINTARRAY32(m + 1);
4760:     Ccsr->column_indices = new THRUSTINTARRAY32(c->nz);
4761:     Ccsr->values         = new THRUSTARRAY(c->nz);
4762:     Ccsr->num_entries    = c->nz;
4763:     Ccusp->coords        = new THRUSTINTARRAY(c->nz);
4764:     if (c->nz) {
4765:       auto              Acoo = new THRUSTINTARRAY32(Annz);
4766:       auto              Bcoo = new THRUSTINTARRAY32(Bnnz);
4767:       auto              Ccoo = new THRUSTINTARRAY32(c->nz);
4768:       THRUSTINTARRAY32 *Aroff, *Broff;

4770:       if (a->compressedrow.use) { /* need full row offset */
4771:         if (!Acusp->rowoffsets_gpu) {
4772:           Acusp->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
4773:           Acusp->rowoffsets_gpu->assign(a->i, a->i + A->rmap->n + 1);
4774:           PetscCall(PetscLogCpuToGpu((A->rmap->n + 1) * sizeof(PetscInt)));
4775:         }
4776:         Aroff = Acusp->rowoffsets_gpu;
4777:       } else Aroff = Acsr->row_offsets;
4778:       if (b->compressedrow.use) { /* need full row offset */
4779:         if (!Bcusp->rowoffsets_gpu) {
4780:           Bcusp->rowoffsets_gpu = new THRUSTINTARRAY32(B->rmap->n + 1);
4781:           Bcusp->rowoffsets_gpu->assign(b->i, b->i + B->rmap->n + 1);
4782:           PetscCall(PetscLogCpuToGpu((B->rmap->n + 1) * sizeof(PetscInt)));
4783:         }
4784:         Broff = Bcusp->rowoffsets_gpu;
4785:       } else Broff = Bcsr->row_offsets;
4786:       PetscCall(PetscLogGpuTimeBegin());
4787:       stat = cusparseXcsr2coo(Acusp->handle, Aroff->data().get(), Annz, m, Acoo->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4788:       PetscCallCUSPARSE(stat);
4789:       stat = cusparseXcsr2coo(Bcusp->handle, Broff->data().get(), Bnnz, m, Bcoo->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4790:       PetscCallCUSPARSE(stat);
4791:       /* Issues when using bool with large matrices on SUMMIT 10.2.89 */
4792:       auto Aperm = thrust::make_constant_iterator(1);
4793:       auto Bperm = thrust::make_constant_iterator(0);
4794: #if PETSC_PKG_CUDA_VERSION_GE(10, 0, 0)
4795:       auto Bcib = thrust::make_transform_iterator(Bcsr->column_indices->begin(), Shift(A->cmap->n));
4796:       auto Bcie = thrust::make_transform_iterator(Bcsr->column_indices->end(), Shift(A->cmap->n));
4797: #else
4798:       /* there are issues instantiating the merge operation using a transform iterator for the columns of B */
4799:       auto Bcib = Bcsr->column_indices->begin();
4800:       auto Bcie = Bcsr->column_indices->end();
4801:       thrust::transform(Bcib, Bcie, Bcib, Shift(A->cmap->n));
4802: #endif
4803:       auto wPerm = new THRUSTINTARRAY32(Annz + Bnnz);
4804:       auto Azb   = thrust::make_zip_iterator(thrust::make_tuple(Acoo->begin(), Acsr->column_indices->begin(), Acsr->values->begin(), Aperm));
4805:       auto Aze   = thrust::make_zip_iterator(thrust::make_tuple(Acoo->end(), Acsr->column_indices->end(), Acsr->values->end(), Aperm));
4806:       auto Bzb   = thrust::make_zip_iterator(thrust::make_tuple(Bcoo->begin(), Bcib, Bcsr->values->begin(), Bperm));
4807:       auto Bze   = thrust::make_zip_iterator(thrust::make_tuple(Bcoo->end(), Bcie, Bcsr->values->end(), Bperm));
4808:       auto Czb   = thrust::make_zip_iterator(thrust::make_tuple(Ccoo->begin(), Ccsr->column_indices->begin(), Ccsr->values->begin(), wPerm->begin()));
4809:       auto p1    = Ccusp->coords->begin();
4810:       auto p2    = Ccusp->coords->begin();
4811: #if CCCL_VERSION >= 3001000
4812:       cuda::std::advance(p2, Annz);
4813: #else
4814:       thrust::advance(p2, Annz);
4815: #endif
4816:       PetscCallThrust(thrust::merge(thrust::device, Azb, Aze, Bzb, Bze, Czb, IJCompare4()));
4817: #if PETSC_PKG_CUDA_VERSION_LT(10, 0, 0)
4818:       thrust::transform(Bcib, Bcie, Bcib, Shift(-A->cmap->n));
4819: #endif
4820:       auto cci = thrust::make_counting_iterator(zero);
4821:       auto cce = thrust::make_counting_iterator(c->nz);
4822: #if 0 //Errors on SUMMIT cuda 11.1.0
4823:       PetscCallThrust(thrust::partition_copy(thrust::device,cci,cce,wPerm->begin(),p1,p2,thrust::identity<int>()));
4824: #else
4825:   #if PETSC_PKG_CUDA_VERSION_LT(12, 9, 0) || PetscDefined(HAVE_THRUST)
4826:       auto pred = thrust::identity<int>();
4827:   #else
4828:       auto pred = cuda::std::identity();
4829:   #endif
4830:       PetscCallThrust(thrust::copy_if(thrust::device, cci, cce, wPerm->begin(), p1, pred));
4831:       PetscCallThrust(thrust::remove_copy_if(thrust::device, cci, cce, wPerm->begin(), p2, pred));
4832: #endif
4833:       stat = cusparseXcoo2csr(Ccusp->handle, Ccoo->data().get(), c->nz, m, Ccsr->row_offsets->data().get(), CUSPARSE_INDEX_BASE_ZERO);
4834:       PetscCallCUSPARSE(stat);
4835:       PetscCall(PetscLogGpuTimeEnd());
4836:       delete wPerm;
4837:       delete Acoo;
4838:       delete Bcoo;
4839:       delete Ccoo;
4840: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4841:       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);
4842:       PetscCallCUSPARSE(stat);
4843: #endif
4844:       if (A->form_explicit_transpose && B->form_explicit_transpose) { /* if A and B have the transpose, generate C transpose too */
4845:         PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(A));
4846:         PetscCall(MatSeqAIJCUSPARSEFormExplicitTranspose(B));
4847:         PetscBool                     AT = Acusp->matTranspose ? PETSC_TRUE : PETSC_FALSE, BT = Bcusp->matTranspose ? PETSC_TRUE : PETSC_FALSE;
4848:         Mat_SeqAIJCUSPARSEMultStruct *CmatT = new Mat_SeqAIJCUSPARSEMultStruct;
4849:         CsrMatrix                    *CcsrT = new CsrMatrix;
4850:         CsrMatrix                    *AcsrT = AT ? (CsrMatrix *)Acusp->matTranspose->mat : NULL;
4851:         CsrMatrix                    *BcsrT = BT ? (CsrMatrix *)Bcusp->matTranspose->mat : NULL;

4853:         (*C)->form_explicit_transpose = PETSC_TRUE;
4854:         (*C)->transupdated            = PETSC_TRUE;
4855:         Ccusp->rowoffsets_gpu         = NULL;
4856:         CmatT->cprowIndices           = NULL;
4857:         CmatT->mat                    = CcsrT;
4858:         CcsrT->num_rows               = n;
4859:         CcsrT->num_cols               = m;
4860:         CcsrT->num_entries            = c->nz;

4862:         CcsrT->row_offsets    = new THRUSTINTARRAY32(n + 1);
4863:         CcsrT->column_indices = new THRUSTINTARRAY32(c->nz);
4864:         CcsrT->values         = new THRUSTARRAY(c->nz);

4866:         PetscCall(PetscLogGpuTimeBegin());
4867:         auto rT = CcsrT->row_offsets->begin();
4868:         if (AT) {
4869:           rT = thrust::copy(AcsrT->row_offsets->begin(), AcsrT->row_offsets->end(), rT);
4870: #if CCCL_VERSION >= 3001000
4871:           cuda::std::advance(rT, -1);
4872: #else
4873:           thrust::advance(rT, -1);
4874: #endif
4875:         }
4876:         if (BT) {
4877:           auto titb = thrust::make_transform_iterator(BcsrT->row_offsets->begin(), Shift(a->nz));
4878:           auto tite = thrust::make_transform_iterator(BcsrT->row_offsets->end(), Shift(a->nz));
4879:           thrust::copy(titb, tite, rT);
4880:         }
4881:         auto cT = CcsrT->column_indices->begin();
4882:         if (AT) cT = thrust::copy(AcsrT->column_indices->begin(), AcsrT->column_indices->end(), cT);
4883:         if (BT) thrust::copy(BcsrT->column_indices->begin(), BcsrT->column_indices->end(), cT);
4884:         auto vT = CcsrT->values->begin();
4885:         if (AT) vT = thrust::copy(AcsrT->values->begin(), AcsrT->values->end(), vT);
4886:         if (BT) thrust::copy(BcsrT->values->begin(), BcsrT->values->end(), vT);
4887:         PetscCall(PetscLogGpuTimeEnd());

4889:         PetscCallCUSPARSE(cusparseCreateMatDescr(&CmatT->descr));
4890:         PetscCallCUSPARSE(cusparseSetMatIndexBase(CmatT->descr, CUSPARSE_INDEX_BASE_ZERO));
4891:         PetscCallCUSPARSE(cusparseSetMatType(CmatT->descr, CUSPARSE_MATRIX_TYPE_GENERAL));
4892:         PetscCallCUDA(cudaMalloc((void **)&CmatT->alpha_one, sizeof(PetscScalar)));
4893:         PetscCallCUDA(cudaMalloc((void **)&CmatT->beta_zero, sizeof(PetscScalar)));
4894:         PetscCallCUDA(cudaMalloc((void **)&CmatT->beta_one, sizeof(PetscScalar)));
4895:         PetscCallCUDA(cudaMemcpy(CmatT->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4896:         PetscCallCUDA(cudaMemcpy(CmatT->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4897:         PetscCallCUDA(cudaMemcpy(CmatT->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostToDevice));
4898: #if PETSC_PKG_CUDA_VERSION_GE(11, 0, 0)
4899:         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);
4900:         PetscCallCUSPARSE(stat);
4901: #endif
4902:         Ccusp->matTranspose = CmatT;
4903:       }
4904:     }

4906:     c->free_a = PETSC_TRUE;
4907:     PetscCall(PetscShmgetAllocateArray(c->nz, sizeof(PetscInt), (void **)&c->j));
4908:     PetscCall(PetscShmgetAllocateArray(m + 1, sizeof(PetscInt), (void **)&c->i));
4909:     c->free_ij = PETSC_TRUE;
4910:     if (PetscDefined(USE_64BIT_INDICES)) { /* 32 to 64-bit conversion on the GPU and then copy to host (lazy) */
4911:       THRUSTINTARRAY ii(Ccsr->row_offsets->size());
4912:       THRUSTINTARRAY jj(Ccsr->column_indices->size());
4913:       ii = *Ccsr->row_offsets;
4914:       jj = *Ccsr->column_indices;
4915:       PetscCallCUDA(cudaMemcpy(c->i, ii.data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4916:       PetscCallCUDA(cudaMemcpy(c->j, jj.data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4917:     } else {
4918:       PetscCallCUDA(cudaMemcpy(c->i, Ccsr->row_offsets->data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4919:       PetscCallCUDA(cudaMemcpy(c->j, Ccsr->column_indices->data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), cudaMemcpyDeviceToHost));
4920:     }
4921:     PetscCall(PetscLogGpuToCpu((Ccsr->column_indices->size() + Ccsr->row_offsets->size()) * sizeof(PetscInt)));
4922:     PetscCall(PetscMalloc1(m, &c->ilen));
4923:     PetscCall(PetscMalloc1(m, &c->imax));
4924:     c->maxnz         = c->nz;
4925:     c->nonzerorowcnt = 0;
4926:     c->rmax          = 0;
4927:     for (i = 0; i < m; i++) {
4928:       const PetscInt nn = c->i[i + 1] - c->i[i];
4929:       c->ilen[i] = c->imax[i] = nn;
4930:       c->nonzerorowcnt += (PetscInt)!!nn;
4931:       c->rmax = PetscMax(c->rmax, nn);
4932:     }
4933:     PetscCall(MatMarkDiagonal_SeqAIJ(*C));
4934:     PetscCall(PetscMalloc1(c->nz, &c->a));
4935:     (*C)->nonzerostate++;
4936:     PetscCall(PetscLayoutSetUp((*C)->rmap));
4937:     PetscCall(PetscLayoutSetUp((*C)->cmap));
4938:     Ccusp->nonzerostate = (*C)->nonzerostate;
4939:     (*C)->preallocated  = PETSC_TRUE;
4940:   } else {
4941:     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);
4942:     c = (Mat_SeqAIJ *)(*C)->data;
4943:     if (c->nz) {
4944:       Ccusp = (Mat_SeqAIJCUSPARSE *)(*C)->spptr;
4945:       PetscCheck(Ccusp->coords, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing coords");
4946:       PetscCheck(Ccusp->format != MAT_CUSPARSE_ELL && Ccusp->format != MAT_CUSPARSE_HYB, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
4947:       PetscCheck(Ccusp->nonzerostate == (*C)->nonzerostate, PETSC_COMM_SELF, PETSC_ERR_COR, "Wrong nonzerostate");
4948:       PetscCall(MatSeqAIJCUSPARSECopyToGPU(A));
4949:       PetscCall(MatSeqAIJCUSPARSECopyToGPU(B));
4950:       PetscCheck(Acusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4951:       PetscCheck(Bcusp->mat, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing Mat_SeqAIJCUSPARSEMultStruct");
4952:       Acsr = (CsrMatrix *)Acusp->mat->mat;
4953:       Bcsr = (CsrMatrix *)Bcusp->mat->mat;
4954:       Ccsr = (CsrMatrix *)Ccusp->mat->mat;
4955:       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());
4956:       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());
4957:       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());
4958:       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);
4959:       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());
4960:       auto pmid = Ccusp->coords->begin();
4961: #if CCCL_VERSION >= 3001000
4962:       cuda::std::advance(pmid, Acsr->num_entries);
4963: #else
4964:       thrust::advance(pmid, Acsr->num_entries);
4965: #endif
4966:       PetscCall(PetscLogGpuTimeBegin());
4967:       auto zibait = thrust::make_zip_iterator(thrust::make_tuple(Acsr->values->begin(), thrust::make_permutation_iterator(Ccsr->values->begin(), Ccusp->coords->begin())));
4968:       auto zieait = thrust::make_zip_iterator(thrust::make_tuple(Acsr->values->end(), thrust::make_permutation_iterator(Ccsr->values->begin(), pmid)));
4969:       thrust::for_each(zibait, zieait, VecCUDAEquals());
4970:       auto zibbit = thrust::make_zip_iterator(thrust::make_tuple(Bcsr->values->begin(), thrust::make_permutation_iterator(Ccsr->values->begin(), pmid)));
4971:       auto ziebit = thrust::make_zip_iterator(thrust::make_tuple(Bcsr->values->end(), thrust::make_permutation_iterator(Ccsr->values->begin(), Ccusp->coords->end())));
4972:       thrust::for_each(zibbit, ziebit, VecCUDAEquals());
4973:       PetscCall(MatSeqAIJCUSPARSEInvalidateTranspose(*C, PETSC_FALSE));
4974:       if (A->form_explicit_transpose && B->form_explicit_transpose && (*C)->form_explicit_transpose) {
4975:         PetscCheck(Ccusp->matTranspose, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing transpose Mat_SeqAIJCUSPARSEMultStruct");
4976:         PetscBool  AT = Acusp->matTranspose ? PETSC_TRUE : PETSC_FALSE, BT = Bcusp->matTranspose ? PETSC_TRUE : PETSC_FALSE;
4977:         CsrMatrix *AcsrT = AT ? (CsrMatrix *)Acusp->matTranspose->mat : NULL;
4978:         CsrMatrix *BcsrT = BT ? (CsrMatrix *)Bcusp->matTranspose->mat : NULL;
4979:         CsrMatrix *CcsrT = (CsrMatrix *)Ccusp->matTranspose->mat;
4980:         auto       vT    = CcsrT->values->begin();
4981:         if (AT) vT = thrust::copy(AcsrT->values->begin(), AcsrT->values->end(), vT);
4982:         if (BT) thrust::copy(BcsrT->values->begin(), BcsrT->values->end(), vT);
4983:         (*C)->transupdated = PETSC_TRUE;
4984:       }
4985:       PetscCall(PetscLogGpuTimeEnd());
4986:     }
4987:   }
4988:   PetscCall(PetscObjectStateIncrease((PetscObject)*C));
4989:   (*C)->assembled     = PETSC_TRUE;
4990:   (*C)->was_assembled = PETSC_FALSE;
4991:   (*C)->offloadmask   = PETSC_OFFLOAD_GPU;
4992:   PetscFunctionReturn(PETSC_SUCCESS);
4993: }

4995: static PetscErrorCode MatSeqAIJCopySubArray_SeqAIJCUSPARSE(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
4996: {
4997:   bool               dmem;
4998:   const PetscScalar *av;

5000:   PetscFunctionBegin;
5001:   dmem = isCudaMem(v);
5002:   PetscCall(MatSeqAIJCUSPARSEGetArrayRead(A, &av));
5003:   if (n && idx) {
5004:     THRUSTINTARRAY widx(n);
5005:     widx.assign(idx, idx + n);
5006:     PetscCall(PetscLogCpuToGpu(n * sizeof(PetscInt)));

5008:     THRUSTARRAY                    *w = NULL;
5009:     thrust::device_ptr<PetscScalar> dv;
5010:     if (dmem) {
5011:       dv = thrust::device_pointer_cast(v);
5012:     } else {
5013:       w  = new THRUSTARRAY(n);
5014:       dv = w->data();
5015:     }
5016:     thrust::device_ptr<const PetscScalar> dav = thrust::device_pointer_cast(av);

5018:     auto zibit = thrust::make_zip_iterator(thrust::make_tuple(thrust::make_permutation_iterator(dav, widx.begin()), dv));
5019:     auto zieit = thrust::make_zip_iterator(thrust::make_tuple(thrust::make_permutation_iterator(dav, widx.end()), dv + n));
5020:     thrust::for_each(zibit, zieit, VecCUDAEquals());
5021:     if (w) PetscCallCUDA(cudaMemcpy(v, w->data().get(), n * sizeof(PetscScalar), cudaMemcpyDeviceToHost));
5022:     delete w;
5023:   } else {
5024:     PetscCallCUDA(cudaMemcpy(v, av, n * sizeof(PetscScalar), dmem ? cudaMemcpyDeviceToDevice : cudaMemcpyDeviceToHost));
5025:   }
5026:   if (!dmem) PetscCall(PetscLogCpuToGpu(n * sizeof(PetscScalar)));
5027:   PetscCall(MatSeqAIJCUSPARSERestoreArrayRead(A, &av));
5028:   PetscFunctionReturn(PETSC_SUCCESS);
5029: }