Searched refs:acolidx (Results 1 – 4 of 4) sorted by relevance
| /petsc/src/mat/impls/sell/seq/seqcuda/ |
| H A D | sellcuda.cu | 75 …qsell_basic_kernel(PetscInt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar … in matmult_seqsell_basic_kernel() argument
85 …ice_id] + row_in_slice; i < sliidx[slice_id + 1]; i += sliceheight) sum += aval[i] * x[acolidx[i]]; in matmult_seqsell_basic_kernel()
90 …qsell_basic_kernel(PetscInt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar … in matmultadd_seqsell_basic_kernel() argument
100 …ice_id] + row_in_slice; i < sliidx[slice_id + 1]; i += sliceheight) sum += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_basic_kernel()
108 …sell_tiled_kernel9(PetscInt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar … in matmult_seqsell_tiled_kernel9() argument
120 …Idx.x + 32 * threadIdx.y; i < sliidx[slice_id + 1]; i += 32 * BLOCKY) t += aval[i] * x[acolidx[i]]; in matmult_seqsell_tiled_kernel9()
137 …sell_tiled_kernel9(PetscInt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar … in matmultadd_seqsell_tiled_kernel9() argument
149 …Idx.x + 32 * threadIdx.y; i < sliidx[slice_id + 1]; i += 32 * BLOCKY) t += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_tiled_kernel9()
185 …, PetscInt totalchunks, const PetscInt *chunk_slice_map, const PetscInt *acolidx, const MatScalar … in matmult_seqsell_tiled_kernel8() argument
209 if (row < nrows && gid < totalentries) t = aval[gid] * x[acolidx[gid]]; in matmult_seqsell_tiled_kernel8()
[all …]
|
| /petsc/src/mat/impls/sell/seq/seqhip/ |
| H A D | sellhip.hip.cxx | 78 …qsell_basic_kernel(PetscInt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar … in matmult_seqsell_basic_kernel() argument
88 …ice_id] + row_in_slice; i < sliidx[slice_id + 1]; i += sliceheight) sum += aval[i] * x[acolidx[i]]; in matmult_seqsell_basic_kernel()
93 …qsell_basic_kernel(PetscInt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar … in matmultadd_seqsell_basic_kernel() argument
103 …ice_id] + row_in_slice; i < sliidx[slice_id + 1]; i += sliceheight) sum += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_basic_kernel()
112 …sell_tiled_kernel9(PetscInt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar … in matmult_seqsell_tiled_kernel9() argument
124 …IZE * threadIdx.y; i < sliidx[slice_id + 1]; i += WARP_SIZE * BLOCKY) t += aval[i] * x[acolidx[i]]; in matmult_seqsell_tiled_kernel9()
141 …sell_tiled_kernel9(PetscInt nrows, PetscInt sliceheight, const PetscInt *acolidx, const MatScalar … in matmultadd_seqsell_tiled_kernel9() argument
153 …IZE * threadIdx.y; i < sliidx[slice_id + 1]; i += WARP_SIZE * BLOCKY) t += aval[i] * x[acolidx[i]]; in matmultadd_seqsell_tiled_kernel9()
189 …, PetscInt totalchunks, const PetscInt *chunk_slice_map, const PetscInt *acolidx, const MatScalar … in matmult_seqsell_tiled_kernel8() argument
213 if (row < nrows && gid < totalentries) t = aval[gid] * x[acolidx[gid]]; in matmult_seqsell_tiled_kernel8()
[all …]
|
| /petsc/src/mat/impls/sell/seq/ |
| H A D | sell.c | 31 vec_idx = _mm256_loadu_si256((__m256i const *)acolidx); \
38 vec_idx = _mm_loadu_si128((__m128i const *)acolidx); /* SSE2 */ \
327 const PetscInt *acolidx = a->colidx; in MatMult_SeqSELL() local
360 PetscPrefetchBlock(acolidx, a->sliidx[i + 1] - a->sliidx[i], 0, PETSC_PREFETCH_HINT_T0); in MatMult_SeqSELL()
372 acolidx += 8; in MatMult_SeqSELL()
375 acolidx += 8; in MatMult_SeqSELL()
378 acolidx += 8; in MatMult_SeqSELL()
384 acolidx += 8; in MatMult_SeqSELL()
387 acolidx += 8; in MatMult_SeqSELL()
393 acolidx += 8; in MatMult_SeqSELL()
[all …]
|
| /petsc/src/mat/impls/sell/mpi/ |
| H A D | mpisell.c | 603 PetscInt M = mat->rmap->N, N = mat->cmap->N, *acolidx, row, col, i, j; in MatView_MPISELL_ASCIIorDraworSocket() local
620 acolidx = Aloc->colidx; in MatView_MPISELL_ASCIIorDraworSocket()
627 col = *acolidx + mat->rmap->rstart; in MatView_MPISELL_ASCIIorDraworSocket()
631 acolidx++; in MatView_MPISELL_ASCIIorDraworSocket()
637 acolidx = Aloc->colidx; in MatView_MPISELL_ASCIIorDraworSocket()
644 col = sell->garray[*acolidx]; in MatView_MPISELL_ASCIIorDraworSocket()
648 acolidx++; in MatView_MPISELL_ASCIIorDraworSocket()
|