Searched refs:vec_y (Results 1 – 3 of 3) sorted by relevance
| /petsc/src/mat/impls/aij/seq/aijperm/ |
| H A D | aijperm.c | 285 __m512d vec_x, vec_y, vec_vals; in MatMult_SeqAIJPERM() local 378 vec_y = _mm512_setzero_pd(); in MatMult_SeqAIJPERM() 384 vec_y = _mm512_fmadd_pd(vec_x, vec_vals, vec_y); in MatMult_SeqAIJPERM() 392 vec_y = _mm512_mask3_fmadd_pd(vec_x, vec_vals, vec_y, mask); in MatMult_SeqAIJPERM() 399 yp[i] += _mm512_reduce_add_pd(vec_y); in MatMult_SeqAIJPERM() 415 vec_y = _mm512_loadu_pd(&yp[i]); in MatMult_SeqAIJPERM() 421 vec_y = _mm512_fmadd_pd(vec_x, vec_vals, vec_y); in MatMult_SeqAIJPERM() 422 _mm512_storeu_pd(&yp[i], vec_y); in MatMult_SeqAIJPERM()
|
| /petsc/src/mat/impls/sell/seq/ |
| H A D | sell.c | 29 #define AVX512_Mult_Private(vec_idx, vec_x, vec_vals, vec_y) \ argument 34 vec_y = _mm512_fmadd_pd(vec_x, vec_vals, vec_y) 36 #define AVX2_Mult_Private(vec_idx, vec_x, vec_vals, vec_y) \ argument 40 vec_y = _mm256_fmadd_pd(vec_x, vec_vals, vec_y) 330 __m512d vec_x, vec_y, vec_vals; in MatMult_SeqSELL() local 337 __m256d vec_x, vec_y, vec_y2, vec_vals; in MatMult_SeqSELL() local 342 __m256d vec_x, vec_y, vec_y2, vec_vals; in MatMult_SeqSELL() local 363 vec_y = _mm512_setzero_pd(); in MatMult_SeqSELL() 371 AVX512_Mult_Private(vec_idx, vec_x, vec_vals, vec_y); in MatMult_SeqSELL() 383 AVX512_Mult_Private(vec_idx, vec_x, vec_vals, vec_y); in MatMult_SeqSELL() [all …]
|
| /petsc/src/mat/impls/aij/seq/ |
| H A D | aij.h | 626 __m512d vec_x, vec_y, vec_vals; in PetscSparseDensePlusDot_AVX512_Private() local 630 vec_y = _mm512_setzero_pd(); in PetscSparseDensePlusDot_AVX512_Private() 635 vec_y = _mm512_fmadd_pd(vec_x, vec_vals, vec_y); in PetscSparseDensePlusDot_AVX512_Private() 647 vec_y = _mm512_mask3_fmadd_pd(vec_x, vec_vals, vec_y, mask); in PetscSparseDensePlusDot_AVX512_Private() 649 *sum += _mm512_reduce_add_pd(vec_y); in PetscSparseDensePlusDot_AVX512_Private() 651 *sum += _mm512_reduce_add_pd(vec_y); in PetscSparseDensePlusDot_AVX512_Private()
|