Lines Matching refs:PetscCallCUDA

229       PetscCallCUDA(cudaMalloc(&fs->csrRowPtr, sizeof(*fs->csrRowPtr) * (m + 1)));  in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
230 PetscCallCUDA(cudaMalloc(&fs->csrColIdx, sizeof(*fs->csrColIdx) * Mnz)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
231 PetscCallCUDA(cudaMalloc(&fs->csrVal, sizeof(*fs->csrVal) * Mnz)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
232PetscCallCUDA(cudaMemcpy(fs->csrRowPtr, Mi, sizeof(*fs->csrRowPtr) * (m + 1), cudaMemcpyHostToDevi… in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
233PetscCallCUDA(cudaMemcpy(fs->csrColIdx, Mj, sizeof(*fs->csrColIdx) * Mnz, cudaMemcpyHostToDevice)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
255 PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(*fs->X) * m)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
256 PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(*fs->Y) * m)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
266 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
267 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
285 PetscCallCUDA(cudaMemcpy(fs->csrVal, Ma, sizeof(*Ma) * Mnz, cudaMemcpyHostToDevice)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_LU()
326 PetscCallCUDA(cudaMallocHost((void **)&AALo, nzLower * sizeof(PetscScalar))); in MatSeqAIJCUSPARSEBuildILULowerTriMatrix()
329 PetscCallCUDA(cudaMallocHost((void **)&AiLo, (n + 1) * sizeof(PetscInt))); in MatSeqAIJCUSPARSEBuildILULowerTriMatrix()
330 PetscCallCUDA(cudaMallocHost((void **)&AjLo, nzLower * sizeof(PetscInt))); in MatSeqAIJCUSPARSEBuildILULowerTriMatrix()
397 PetscCallCUDA(cudaMalloc(&loTriFactor->solveBuffer, loTriFactor->solveBufferSize)); in MatSeqAIJCUSPARSEBuildILULowerTriMatrix()
403 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSEBuildILULowerTriMatrix()
409 PetscCallCUDA(cudaFreeHost(AiLo)); in MatSeqAIJCUSPARSEBuildILULowerTriMatrix()
410 PetscCallCUDA(cudaFreeHost(AjLo)); in MatSeqAIJCUSPARSEBuildILULowerTriMatrix()
413 …if (!loTriFactor->AA_h) PetscCallCUDA(cudaMallocHost((void **)&loTriFactor->AA_h, nzLower * sizeof… in MatSeqAIJCUSPARSEBuildILULowerTriMatrix()
458 PetscCallCUDA(cudaMallocHost((void **)&AAUp, nzUpper * sizeof(PetscScalar))); in MatSeqAIJCUSPARSEBuildILUUpperTriMatrix()
461 PetscCallCUDA(cudaMallocHost((void **)&AiUp, (n + 1) * sizeof(PetscInt))); in MatSeqAIJCUSPARSEBuildILUUpperTriMatrix()
462 PetscCallCUDA(cudaMallocHost((void **)&AjUp, nzUpper * sizeof(PetscInt))); in MatSeqAIJCUSPARSEBuildILUUpperTriMatrix()
526 PetscCallCUDA(cudaMalloc(&upTriFactor->solveBuffer, upTriFactor->solveBufferSize)); in MatSeqAIJCUSPARSEBuildILUUpperTriMatrix()
533 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSEBuildILUUpperTriMatrix()
539 PetscCallCUDA(cudaFreeHost(AiUp)); in MatSeqAIJCUSPARSEBuildILUUpperTriMatrix()
540 PetscCallCUDA(cudaFreeHost(AjUp)); in MatSeqAIJCUSPARSEBuildILUUpperTriMatrix()
543 …if (!upTriFactor->AA_h) PetscCallCUDA(cudaMallocHost((void **)&upTriFactor->AA_h, nzUpper * sizeof… in MatSeqAIJCUSPARSEBuildILUUpperTriMatrix()
644 PetscCallCUDA(cudaMalloc(&fs->csrRowPtr, sizeof(*fs->csrRowPtr) * (m + 1))); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
645 PetscCallCUDA(cudaMalloc(&fs->csrColIdx, sizeof(*fs->csrColIdx) * Mnz)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
646 PetscCallCUDA(cudaMalloc(&fs->csrVal, sizeof(*fs->csrVal) * Mnz)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
647 PetscCallCUDA(cudaMalloc(&fs->diag, sizeof(*fs->diag) * m)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
648 PetscCallCUDA(cudaMemcpy(fs->csrRowPtr, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyHostToDevice)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
649 PetscCallCUDA(cudaMemcpy(fs->csrColIdx, Mj, sizeof(*Mj) * Mnz, cudaMemcpyHostToDevice)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
665 PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(*fs->X) * m)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
666 PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(*fs->Y) * m)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
674 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
678 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Ut, fs->spsvBufferSize_Ut)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
694 PetscCallCUDA(cudaMemcpy(fs->csrVal, Ma, sizeof(*Ma) * Mnz, cudaMemcpyHostToDevice)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
695 PetscCallCUDA(cudaMemcpy(fs->diag, D, sizeof(*D) * m, cudaMemcpyHostToDevice)); in MatSeqAIJCUSPARSEBuildFactoredMatrix_Cholesky()
791 PetscCallCUDA(cudaMallocHost((void **)&AAUp, nzUpper * sizeof(PetscScalar))); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
792 PetscCallCUDA(cudaMallocHost((void **)&AALo, nzUpper * sizeof(PetscScalar))); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
795 PetscCallCUDA(cudaMallocHost((void **)&AiUp, (n + 1) * sizeof(PetscInt))); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
796 PetscCallCUDA(cudaMallocHost((void **)&AjUp, nzUpper * sizeof(PetscInt))); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
865 PetscCallCUDA(cudaMalloc(&upTriFactor->solveBuffer, upTriFactor->solveBufferSize)); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
872 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
917 PetscCallCUDA(cudaMalloc(&loTriFactor->solveBuffer, loTriFactor->solveBufferSize)); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
924 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
931 PetscCallCUDA(cudaFreeHost(AiUp)); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
932 PetscCallCUDA(cudaFreeHost(AjUp)); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
961 PetscCallCUDA(cudaFreeHost(AAUp)); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
962 PetscCallCUDA(cudaFreeHost(AALo)); in MatSeqAIJCUSPARSEBuildICCTriMatrices()
1094 PetscCallCUDA(cudaMalloc(&loTriFactor->csr2cscBuffer, loTriFactor->csr2cscBufferSize)); in MatSeqAIJCUSPARSEAnalyzeTransposeForSolve()
1110 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSEAnalyzeTransposeForSolve()
1119 PetscCallCUDA(cudaMalloc(&loTriFactorT->solveBuffer, loTriFactorT->solveBufferSize)); in MatSeqAIJCUSPARSEAnalyzeTransposeForSolve()
1126 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSEAnalyzeTransposeForSolve()
1170 PetscCallCUDA(cudaMalloc(&upTriFactor->csr2cscBuffer, upTriFactor->csr2cscBufferSize)); in MatSeqAIJCUSPARSEAnalyzeTransposeForSolve()
1186 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSEAnalyzeTransposeForSolve()
1195 PetscCallCUDA(cudaMalloc(&upTriFactorT->solveBuffer, upTriFactorT->solveBufferSize)); in MatSeqAIJCUSPARSEAnalyzeTransposeForSolve()
1203 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSEAnalyzeTransposeForSolve()
1242 PetscCallCUDA(cudaMalloc((void **)&matstructT->alpha_one, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEFormExplicitTranspose()
1243 PetscCallCUDA(cudaMalloc((void **)&matstructT->beta_zero, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEFormExplicitTranspose()
1244 PetscCallCUDA(cudaMalloc((void **)&matstructT->beta_one, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEFormExplicitTranspose()
1245PetscCallCUDA(cudaMemcpy(matstructT->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemc… in MatSeqAIJCUSPARSEFormExplicitTranspose()
1246PetscCallCUDA(cudaMemcpy(matstructT->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMem… in MatSeqAIJCUSPARSEFormExplicitTranspose()
1247PetscCallCUDA(cudaMemcpy(matstructT->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcp… in MatSeqAIJCUSPARSEFormExplicitTranspose()
1367 PetscCallCUDA(cudaMalloc(&csr2cscBuffer, csr2cscBufferSize)); in MatSeqAIJCUSPARSEFormExplicitTranspose()
1393 PetscCallCUDA(cudaFree(csr2cscBuffer)); in MatSeqAIJCUSPARSEFormExplicitTranspose()
1483 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Lt, fs->spsvBufferSize_Lt)); in MatSolveTranspose_SeqAIJCUSPARSE_LU()
1484 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Ut, fs->spsvBufferSize_Ut)); in MatSolveTranspose_SeqAIJCUSPARSE_LU()
1716PetscCallCUDA(cudaMemcpyAsync(fs->csrVal, Acsr->values->data().get(), sizeof(PetscScalar) * nz, cu… in MatILUFactorNumeric_SeqAIJCUSPARSE_ILU0()
1802 PetscCallCUDA(cudaMalloc((void **)&fs->csrRowPtr32, sizeof(*fs->csrRowPtr32) * (m + 1))); in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1803 PetscCallCUDA(cudaMalloc((void **)&fs->csrColIdx32, sizeof(*fs->csrColIdx32) * nz)); in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1804 PetscCallCUDA(cudaMalloc((void **)&fs->csrVal, sizeof(*fs->csrVal) * nz)); in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1806PetscCallCUDA(cudaMemcpyAsync(fs->csrRowPtr32, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyDeviceToDevice… in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1807PetscCallCUDA(cudaMemcpyAsync(fs->csrColIdx32, Aj, sizeof(*Aj) * nz, cudaMemcpyDeviceToDevice, Pet… in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1845 PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(PetscScalar) * m)); in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1846 PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(PetscScalar) * m)); in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1863PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_L, (size_t)fs->fa… in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1865 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_U, fs->spsvBufferSize_U)); in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1867PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_U, (size_t)fs->fa… in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1869 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L)); in MatILUFactorSymbolic_SeqAIJCUSPARSE_ILU0()
1966PetscCallCUDA(cudaMemcpyAsync(fs->csrVal, Acsr->values->data().get(), sizeof(PetscScalar) * nz, cu… in MatICCFactorNumeric_SeqAIJCUSPARSE_ICC0()
2051 PetscCallCUDA(cudaMalloc((void **)&fs->csrRowPtr32, sizeof(*fs->csrRowPtr32) * (m + 1))); in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2052 PetscCallCUDA(cudaMalloc((void **)&fs->csrColIdx32, sizeof(*fs->csrColIdx32) * nz)); in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2053 PetscCallCUDA(cudaMalloc((void **)&fs->csrVal, sizeof(PetscScalar) * nz)); in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2055PetscCallCUDA(cudaMemcpyAsync(fs->csrRowPtr32, Ai, sizeof(*Ai) * (m + 1), cudaMemcpyDeviceToDevice… in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2056PetscCallCUDA(cudaMemcpyAsync(fs->csrColIdx32, Aj, sizeof(*Aj) * nz, cudaMemcpyDeviceToDevice, Pet… in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2086 PetscCallCUDA(cudaMalloc((void **)&fs->X, sizeof(PetscScalar) * m)); in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2087 PetscCallCUDA(cudaMalloc((void **)&fs->Y, sizeof(PetscScalar) * m)); in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2102PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_L, (size_t)fs->fa… in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2104 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_Lt, fs->spsvBufferSize_Lt)); in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2106PetscCallCUDA(cudaMalloc((void **)&fs->factBuffer_M, PetscMax(fs->spsvBufferSize_Lt, (size_t)fs->f… in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2108 PetscCallCUDA(cudaMalloc((void **)&fs->spsvBuffer_L, fs->spsvBufferSize_L)); in MatICCFactorSymbolic_SeqAIJCUSPARSE_ICC0()
2329PetscCallCUDA(cudaMemcpy(a->a, matrix->values->data().get(), a->nz * sizeof(PetscScalar), cudaMemc… in MatSeqAIJCUSPARSECopyFromGPU()
2334PetscCallCUDA(cudaMemcpy(a->a, fs->csrVal, a->nz * sizeof(PetscScalar), cudaMemcpyDeviceToHost)); in MatSeqAIJCUSPARSECopyFromGPU()
2442 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSECopyToGPU()
2479 PetscCallCUDA(cudaMalloc((void **)&matstruct->alpha_one, sizeof(PetscScalar))); in MatSeqAIJCUSPARSECopyToGPU()
2480 PetscCallCUDA(cudaMalloc((void **)&matstruct->beta_zero, sizeof(PetscScalar))); in MatSeqAIJCUSPARSECopyToGPU()
2481 PetscCallCUDA(cudaMalloc((void **)&matstruct->beta_one, sizeof(PetscScalar))); in MatSeqAIJCUSPARSECopyToGPU()
2482PetscCallCUDA(cudaMemcpy(matstruct->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcp… in MatSeqAIJCUSPARSECopyToGPU()
2483PetscCallCUDA(cudaMemcpy(matstruct->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemc… in MatSeqAIJCUSPARSECopyToGPU()
2484PetscCallCUDA(cudaMemcpy(matstruct->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpy… in MatSeqAIJCUSPARSECopyToGPU()
2563 PetscCallCUDA(WaitForCUDA()); in MatSeqAIJCUSPARSECopyToGPU()
2626 PetscCallCUDA(cudaFree(mmdata->Bt)); in MatProductCtxDestroy_MatMatCusparse()
2634 if (mmdata->dBuffer4) PetscCallCUDA(cudaFree(mmdata->dBuffer4)); in MatProductCtxDestroy_MatMatCusparse()
2635 if (mmdata->dBuffer5) PetscCallCUDA(cudaFree(mmdata->dBuffer5)); in MatProductCtxDestroy_MatMatCusparse()
2637 if (mmdata->mmBuffer) PetscCallCUDA(cudaFree(mmdata->mmBuffer)); in MatProductCtxDestroy_MatMatCusparse()
2638 if (mmdata->mmBuffer2) PetscCallCUDA(cudaFree(mmdata->mmBuffer2)); in MatProductCtxDestroy_MatMatCusparse()
2767 PetscCallCUDA(cudaFree(mmdata->mmBuffer)); in MatProductNumeric_SeqAIJCUSPARSE_SeqDENSECUDA()
2768 PetscCallCUDA(cudaMalloc(&mmdata->mmBuffer, mmBufferSize)); in MatProductNumeric_SeqAIJCUSPARSE_SeqDENSECUDA()
2884 …if (product->type == MATPRODUCT_ABt || product->type == MATPRODUCT_RARt) PetscCallCUDA(cudaMalloc(… in MatProductSymbolic_SeqAIJCUSPARSE_SeqDENSECUDA()
3008 PetscCallCUDA(WaitForCUDA()); in MatProductNumeric_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3145 PetscCallCUDA(cudaMalloc((void **)&Cmat->alpha_one, sizeof(PetscScalar))); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3146 PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_zero, sizeof(PetscScalar))); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3147 PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_one, sizeof(PetscScalar))); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3148PetscCallCUDA(cudaMemcpy(Cmat->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHost… in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3149PetscCallCUDA(cudaMemcpy(Cmat->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHos… in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3150PetscCallCUDA(cudaMemcpy(Cmat->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostT… in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3239 PetscCallCUDA(cudaMalloc((void **)&dBuffer1, bufferSize1)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3246 PetscCallCUDA(cudaMalloc((void **)&dBuffer2, bufferSize2)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3247 PetscCallCUDA(cudaMalloc((void **)&dBuffer3, bufferSize3)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3248 PetscCallCUDA(cudaMalloc((void **)&mmdata->dBuffer4, bufferSize4)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3251 PetscCallCUDA(cudaFree(dBuffer1)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3252 PetscCallCUDA(cudaFree(dBuffer2)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3259 PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */ in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3261 PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */ in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3268 PetscCallCUDA(cudaMalloc((void **)&mmdata->dBuffer5, bufferSize5)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3271 PetscCallCUDA(cudaFree(dBuffer3)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3281 PetscCallCUDA(cudaMalloc((void **)&mmdata->mmBuffer2, bufSize2)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3293 PetscCallCUDA(cudaMalloc((void **)&mmdata->mmBuffer, mmdata->mmBufferSize)); in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3303 PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */ in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3305 PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */ in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3318 PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */ in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3320 PetscCallCUDA(cudaPeekAtLastError()); /* catch out of memory errors */ in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3344PetscCallCUDA(cudaMemcpy(d_i, ii.data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cudaM… in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3345PetscCallCUDA(cudaMemcpy(c->j, jj.data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), c… in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3349PetscCallCUDA(cudaMemcpy(d_i, Ccsr->row_offsets->data().get(), Ccsr->row_offsets->size() * sizeof(… in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3350PetscCallCUDA(cudaMemcpy(c->j, Ccsr->column_indices->data().get(), Ccsr->column_indices->size() * … in MatProductSymbolic_SeqAIJCUSPARSE_SeqAIJCUSPARSE()
3664PetscCallCUDA(cudaMalloc(&matstruct->cuSpMV[opA].spmvBuffer, matstruct->cuSpMV[opA].spmvBufferSize… in MatMultAddKernel_SeqAIJCUSPARSE()
3770 PetscCallCUDA(cudaPeekAtLastError()); in MatGetDiagonal_SeqAIJCUSPARSE()
3912 PetscCallCUDA(cudaMalloc(&buffer, bufferSize)); in MatAXPY_SeqAIJCUSPARSE()
3918 PetscCallCUDA(cudaFree(buffer)); in MatAXPY_SeqAIJCUSPARSE()
4213 if ((*trifactor)->solveBuffer) PetscCallCUDA(cudaFree((*trifactor)->solveBuffer)); in MatSeqAIJCUSPARSEMultStruct_Destroy()
4214 if ((*trifactor)->AA_h) PetscCallCUDA(cudaFreeHost((*trifactor)->AA_h)); in MatSeqAIJCUSPARSEMultStruct_Destroy()
4216 if ((*trifactor)->csr2cscBuffer) PetscCallCUDA(cudaFree((*trifactor)->csr2cscBuffer)); in MatSeqAIJCUSPARSEMultStruct_Destroy()
4245 if ((*matstruct)->alpha_one) PetscCallCUDA(cudaFree((*matstruct)->alpha_one)); in MatSeqAIJCUSPARSEMultStruct_Destroy()
4246 if ((*matstruct)->beta_zero) PetscCallCUDA(cudaFree((*matstruct)->beta_zero)); in MatSeqAIJCUSPARSEMultStruct_Destroy()
4247 if ((*matstruct)->beta_one) PetscCallCUDA(cudaFree((*matstruct)->beta_one)); in MatSeqAIJCUSPARSEMultStruct_Destroy()
4255 PetscCallCUDA(cudaFree(mdata->cuSpMV[i].spmvBuffer)); in MatSeqAIJCUSPARSEMultStruct_Destroy()
4291 PetscCallCUDA(cudaFree(fs->csrRowPtr)); in MatSeqAIJCUSPARSETriFactors_Reset()
4292 PetscCallCUDA(cudaFree(fs->csrColIdx)); in MatSeqAIJCUSPARSETriFactors_Reset()
4293 PetscCallCUDA(cudaFree(fs->csrRowPtr32)); in MatSeqAIJCUSPARSETriFactors_Reset()
4294 PetscCallCUDA(cudaFree(fs->csrColIdx32)); in MatSeqAIJCUSPARSETriFactors_Reset()
4295 PetscCallCUDA(cudaFree(fs->csrVal)); in MatSeqAIJCUSPARSETriFactors_Reset()
4296 PetscCallCUDA(cudaFree(fs->diag)); in MatSeqAIJCUSPARSETriFactors_Reset()
4297 PetscCallCUDA(cudaFree(fs->X)); in MatSeqAIJCUSPARSETriFactors_Reset()
4298 PetscCallCUDA(cudaFree(fs->Y)); in MatSeqAIJCUSPARSETriFactors_Reset()
4300 PetscCallCUDA(cudaFree(fs->spsvBuffer_L)); in MatSeqAIJCUSPARSETriFactors_Reset()
4301 PetscCallCUDA(cudaFree(fs->spsvBuffer_U)); in MatSeqAIJCUSPARSETriFactors_Reset()
4302 PetscCallCUDA(cudaFree(fs->spsvBuffer_Lt)); in MatSeqAIJCUSPARSETriFactors_Reset()
4303 PetscCallCUDA(cudaFree(fs->spsvBuffer_Ut)); in MatSeqAIJCUSPARSETriFactors_Reset()
4366 PetscCallCUDA(cudaFree(coo->perm)); in MatCOOStructDestroy_SeqAIJCUSPARSE()
4367 PetscCallCUDA(cudaFree(coo->jmap)); in MatCOOStructDestroy_SeqAIJCUSPARSE()
4385 PetscCallCUDA(cudaMemcpy(i, coo_i, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost)); in MatSetPreallocationCOO_SeqAIJCUSPARSE()
4386 PetscCallCUDA(cudaMemcpy(j, coo_j, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost)); in MatSetPreallocationCOO_SeqAIJCUSPARSE()
4403 PetscCallCUDA(cudaMalloc((void **)&coo_d->jmap, (coo_h->nz + 1) * sizeof(PetscCount))); in MatSetPreallocationCOO_SeqAIJCUSPARSE()
4404PetscCallCUDA(cudaMemcpy(coo_d->jmap, coo_h->jmap, (coo_h->nz + 1) * sizeof(PetscCount), cudaMemcp… in MatSetPreallocationCOO_SeqAIJCUSPARSE()
4405 PetscCallCUDA(cudaMalloc((void **)&coo_d->perm, coo_h->Atot * sizeof(PetscCount))); in MatSetPreallocationCOO_SeqAIJCUSPARSE()
4406PetscCallCUDA(cudaMemcpy(coo_d->perm, coo_h->perm, coo_h->Atot * sizeof(PetscCount), cudaMemcpyHos… in MatSetPreallocationCOO_SeqAIJCUSPARSE()
4443 PetscCallCUDA(cudaMalloc((void **)&v1, coo->n * sizeof(PetscScalar))); in MatSetValuesCOO_SeqAIJCUSPARSE()
4444 PetscCallCUDA(cudaMemcpy((void *)v1, v, coo->n * sizeof(PetscScalar), cudaMemcpyHostToDevice)); in MatSetValuesCOO_SeqAIJCUSPARSE()
4453 PetscCallCUDA(cudaPeekAtLastError()); in MatSetValuesCOO_SeqAIJCUSPARSE()
4460 if (PetscMemTypeHost(memtype)) PetscCallCUDA(cudaFree((void *)v1)); in MatSetValuesCOO_SeqAIJCUSPARSE()
4782 PetscCallCUDA(cudaMalloc((void **)&Cmat->alpha_one, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEMergeMats()
4783 PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_zero, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEMergeMats()
4784 PetscCallCUDA(cudaMalloc((void **)&Cmat->beta_one, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEMergeMats()
4785PetscCallCUDA(cudaMemcpy(Cmat->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHost… in MatSeqAIJCUSPARSEMergeMats()
4786PetscCallCUDA(cudaMemcpy(Cmat->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHos… in MatSeqAIJCUSPARSEMergeMats()
4787PetscCallCUDA(cudaMemcpy(Cmat->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHostT… in MatSeqAIJCUSPARSEMergeMats()
4931 PetscCallCUDA(cudaMalloc((void **)&CmatT->alpha_one, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEMergeMats()
4932 PetscCallCUDA(cudaMalloc((void **)&CmatT->beta_zero, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEMergeMats()
4933 PetscCallCUDA(cudaMalloc((void **)&CmatT->beta_one, sizeof(PetscScalar))); in MatSeqAIJCUSPARSEMergeMats()
4934PetscCallCUDA(cudaMemcpy(CmatT->alpha_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHos… in MatSeqAIJCUSPARSEMergeMats()
4935PetscCallCUDA(cudaMemcpy(CmatT->beta_zero, &PETSC_CUSPARSE_ZERO, sizeof(PetscScalar), cudaMemcpyHo… in MatSeqAIJCUSPARSEMergeMats()
4936PetscCallCUDA(cudaMemcpy(CmatT->beta_one, &PETSC_CUSPARSE_ONE, sizeof(PetscScalar), cudaMemcpyHost… in MatSeqAIJCUSPARSEMergeMats()
4954PetscCallCUDA(cudaMemcpy(c->i, ii.data().get(), Ccsr->row_offsets->size() * sizeof(PetscInt), cuda… in MatSeqAIJCUSPARSEMergeMats()
4955PetscCallCUDA(cudaMemcpy(c->j, jj.data().get(), Ccsr->column_indices->size() * sizeof(PetscInt), c… in MatSeqAIJCUSPARSEMergeMats()
4957PetscCallCUDA(cudaMemcpy(c->i, Ccsr->row_offsets->data().get(), Ccsr->row_offsets->size() * sizeof… in MatSeqAIJCUSPARSEMergeMats()
4958PetscCallCUDA(cudaMemcpy(c->j, Ccsr->column_indices->data().get(), Ccsr->column_indices->size() * … in MatSeqAIJCUSPARSEMergeMats()
5059 …if (w) PetscCallCUDA(cudaMemcpy(v, w->data().get(), n * sizeof(PetscScalar), cudaMemcpyDeviceToHos… in MatSeqAIJCopySubArray_SeqAIJCUSPARSE()
5062PetscCallCUDA(cudaMemcpy(v, av, n * sizeof(PetscScalar), dmem ? cudaMemcpyDeviceToDevice : cudaMem… in MatSeqAIJCopySubArray_SeqAIJCUSPARSE()