Lines Matching refs:acolidx
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()
401 acolidx += 8; in MatMult_SeqSELL()
404 acolidx += 8; in MatMult_SeqSELL()
407 acolidx += 8; in MatMult_SeqSELL()
410 acolidx += 8; in MatMult_SeqSELL()
427 PetscPrefetchBlock(acolidx, a->sliidx[i + 1] - a->sliidx[i], 0, PETSC_PREFETCH_HINT_T0); in MatMult_SeqSELL()
437 for (j = 0; j < nnz_in_row; ++j) yval += aval[8 * j + r] * x[acolidx[8 * j + r]]; in MatMult_SeqSELL()
452 acolidx += 4; in MatMult_SeqSELL()
455 acolidx += 4; in MatMult_SeqSELL()
464 PetscPrefetchBlock(acolidx, a->sliidx[i + 1] - a->sliidx[i], 0, PETSC_PREFETCH_HINT_T0); in MatMult_SeqSELL()
477 for (j = 0; j < nnz_in_row; ++j) yval += aval[8 * j + r] * x[acolidx[8 * j + r]]; in MatMult_SeqSELL()
489 vec_x_tmp = _mm_loadl_pd(vec_x_tmp, x + *acolidx++); in MatMult_SeqSELL()
490 vec_x_tmp = _mm_loadh_pd(vec_x_tmp, x + *acolidx++); in MatMult_SeqSELL()
492 vec_x_tmp = _mm_loadl_pd(vec_x_tmp, x + *acolidx++); in MatMult_SeqSELL()
493 vec_x_tmp = _mm_loadh_pd(vec_x_tmp, x + *acolidx++); in MatMult_SeqSELL()
499 vec_x_tmp = _mm_loadl_pd(vec_x_tmp, x + *acolidx++); in MatMult_SeqSELL()
500 vec_x_tmp = _mm_loadh_pd(vec_x_tmp, x + *acolidx++); in MatMult_SeqSELL()
502 vec_x_tmp = _mm_loadl_pd(vec_x_tmp, x + *acolidx++); in MatMult_SeqSELL()
503 vec_x_tmp = _mm_loadh_pd(vec_x_tmp, x + *acolidx++); in MatMult_SeqSELL()
517 …r (k = a->sliidx[i] + j; k < a->sliidx[i + 1]; k += sliceheight) sum[j] += aval[k] * x[acolidx[k]]; in MatMult_SeqSELL()
542 const PetscInt *acolidx = a->colidx; in MatMultAdd_SeqSELL() local
574 PetscPrefetchBlock(acolidx, a->sliidx[i + 1] - a->sliidx[i], 0, PETSC_PREFETCH_HINT_T0); in MatMultAdd_SeqSELL()
591 acolidx += 8; in MatMultAdd_SeqSELL()
594 acolidx += 8; in MatMultAdd_SeqSELL()
597 acolidx += 8; in MatMultAdd_SeqSELL()
603 acolidx += 8; in MatMultAdd_SeqSELL()
606 acolidx += 8; in MatMultAdd_SeqSELL()
612 acolidx += 8; in MatMultAdd_SeqSELL()
620 acolidx += 8; in MatMultAdd_SeqSELL()
623 acolidx += 8; in MatMultAdd_SeqSELL()
626 acolidx += 8; in MatMultAdd_SeqSELL()
629 acolidx += 8; in MatMultAdd_SeqSELL()
645 PetscPrefetchBlock(acolidx, a->sliidx[i + 1] - a->sliidx[i], 0, PETSC_PREFETCH_HINT_T0); in MatMultAdd_SeqSELL()
654 for (j = 0; j < nnz_in_row; ++j) yval += aval[8 * j + r] * x[acolidx[8 * j + r]]; in MatMultAdd_SeqSELL()
667 vec_x_tmp = _mm_loadl_pd(vec_x_tmp, x + *acolidx++); in MatMultAdd_SeqSELL()
668 vec_x_tmp = _mm_loadh_pd(vec_x_tmp, x + *acolidx++); in MatMultAdd_SeqSELL()
671 vec_x_tmp = _mm_loadl_pd(vec_x_tmp, x + *acolidx++); in MatMultAdd_SeqSELL()
672 vec_x_tmp = _mm_loadh_pd(vec_x_tmp, x + *acolidx++); in MatMultAdd_SeqSELL()
678 vec_x_tmp = _mm_loadl_pd(vec_x_tmp, x + *acolidx++); in MatMultAdd_SeqSELL()
679 vec_x_tmp = _mm_loadh_pd(vec_x_tmp, x + *acolidx++); in MatMultAdd_SeqSELL()
681 vec_x_tmp = _mm_loadl_pd(vec_x_tmp, x + *acolidx++); in MatMultAdd_SeqSELL()
682 vec_x_tmp = _mm_loadh_pd(vec_x_tmp, x + *acolidx++); in MatMultAdd_SeqSELL()
696 …r (k = a->sliidx[i] + j; k < a->sliidx[i + 1]; k += sliceheight) sum[j] += aval[k] * x[acolidx[k]]; in MatMultAdd_SeqSELL()
719 const PetscInt *acolidx = a->colidx; in MatMultTransposeAdd_SeqSELL() local
741 …for (j = 0; j < nnz_in_row; ++j) y[acolidx[sliceheight * j + r]] += aval[sliceheight * j + r] * x[… in MatMultTransposeAdd_SeqSELL()
746 …for (j = a->sliidx[i] + r; j < a->sliidx[i + 1]; j += sliceheight) y[acolidx[j]] += aval[j] * x[sl… in MatMultTransposeAdd_SeqSELL()