Lines Matching refs:a0

679   __m256d a0, a1, a2, a3, a4, a5;  in MatMult_SeqBAIJ_9_AVX2()  local
725 a0 = _mm256_loadu_pd(&v[j * 81]); in MatMult_SeqBAIJ_9_AVX2()
726 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_9_AVX2()
734 a0 = _mm256_loadu_pd(&v[j * 81 + 9]); in MatMult_SeqBAIJ_9_AVX2()
735 z0 = _mm256_fmadd_pd(a0, w1, z0); in MatMult_SeqBAIJ_9_AVX2()
752 a0 = _mm256_loadu_pd(&v[j * 81 + 27]); in MatMult_SeqBAIJ_9_AVX2()
753 z0 = _mm256_fmadd_pd(a0, w3, z0); in MatMult_SeqBAIJ_9_AVX2()
770 a0 = _mm256_loadu_pd(&v[j * 81 + 45]); in MatMult_SeqBAIJ_9_AVX2()
771 z0 = _mm256_fmadd_pd(a0, w1, z0); in MatMult_SeqBAIJ_9_AVX2()
779 a0 = _mm256_loadu_pd(&v[j * 81 + 54]); in MatMult_SeqBAIJ_9_AVX2()
780 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMult_SeqBAIJ_9_AVX2()
797 a0 = _mm256_loadu_pd(&v[j * 81 + 72]); in MatMult_SeqBAIJ_9_AVX2()
798 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_9_AVX2()
1333 __m256d a0, a1, a2, a3, a4, a5; in MatMult_SeqBAIJ_12_AVX2() local
1364 a0 = _mm256_loadu_pd(v + 0); in MatMult_SeqBAIJ_12_AVX2()
1365 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_12_AVX2()
1382 a0 = _mm256_loadu_pd(v + 24); in MatMult_SeqBAIJ_12_AVX2()
1383 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMult_SeqBAIJ_12_AVX2()
1400 a0 = _mm256_loadu_pd(v + 48); in MatMult_SeqBAIJ_12_AVX2()
1401 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_12_AVX2()
1418 a0 = _mm256_loadu_pd(v + 72); in MatMult_SeqBAIJ_12_AVX2()
1419 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMult_SeqBAIJ_12_AVX2()
1436 a0 = _mm256_loadu_pd(v + 96); in MatMult_SeqBAIJ_12_AVX2()
1437 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_12_AVX2()
1454 a0 = _mm256_loadu_pd(v + 120); in MatMult_SeqBAIJ_12_AVX2()
1455 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMult_SeqBAIJ_12_AVX2()
2515 __m256d a0, a1, a2, a3, a4, a5; in MatMultAdd_SeqBAIJ_9_AVX2() local
2561 a0 = _mm256_loadu_pd(&v[j * 81]); in MatMultAdd_SeqBAIJ_9_AVX2()
2562 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2570 a0 = _mm256_loadu_pd(&v[j * 81 + 9]); in MatMultAdd_SeqBAIJ_9_AVX2()
2571 z0 = _mm256_fmadd_pd(a0, w1, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2588 a0 = _mm256_loadu_pd(&v[j * 81 + 27]); in MatMultAdd_SeqBAIJ_9_AVX2()
2589 z0 = _mm256_fmadd_pd(a0, w3, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2606 a0 = _mm256_loadu_pd(&v[j * 81 + 45]); in MatMultAdd_SeqBAIJ_9_AVX2()
2607 z0 = _mm256_fmadd_pd(a0, w1, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2615 a0 = _mm256_loadu_pd(&v[j * 81 + 54]); in MatMultAdd_SeqBAIJ_9_AVX2()
2616 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2633 a0 = _mm256_loadu_pd(&v[j * 81 + 72]); in MatMultAdd_SeqBAIJ_9_AVX2()
2634 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMultAdd_SeqBAIJ_9_AVX2()