Lines Matching refs:z0
681 __m256d z0, z1, z2; in MatMult_SeqBAIJ_9_AVX2() local
718 z0 = _mm256_setzero_pd(); in MatMult_SeqBAIJ_9_AVX2()
726 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_9_AVX2()
735 z0 = _mm256_fmadd_pd(a0, w1, z0); in MatMult_SeqBAIJ_9_AVX2()
744 z0 = _mm256_fmadd_pd(a3, w2, z0); in MatMult_SeqBAIJ_9_AVX2()
753 z0 = _mm256_fmadd_pd(a0, w3, z0); in MatMult_SeqBAIJ_9_AVX2()
762 z0 = _mm256_fmadd_pd(a3, w0, z0); in MatMult_SeqBAIJ_9_AVX2()
771 z0 = _mm256_fmadd_pd(a0, w1, z0); in MatMult_SeqBAIJ_9_AVX2()
780 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMult_SeqBAIJ_9_AVX2()
789 z0 = _mm256_fmadd_pd(a3, w3, z0); in MatMult_SeqBAIJ_9_AVX2()
798 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_9_AVX2()
805 _mm256_storeu_pd(&z[0], z0); in MatMult_SeqBAIJ_9_AVX2()
1335 __m256d z0, z1, z2; in MatMult_SeqBAIJ_12_AVX2() local
1353 z0 = _mm256_setzero_pd(); in MatMult_SeqBAIJ_12_AVX2()
1365 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_12_AVX2()
1374 z0 = _mm256_fmadd_pd(a3, w1, z0); in MatMult_SeqBAIJ_12_AVX2()
1383 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMult_SeqBAIJ_12_AVX2()
1392 z0 = _mm256_fmadd_pd(a3, w3, z0); in MatMult_SeqBAIJ_12_AVX2()
1401 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_12_AVX2()
1410 z0 = _mm256_fmadd_pd(a3, w1, z0); in MatMult_SeqBAIJ_12_AVX2()
1419 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMult_SeqBAIJ_12_AVX2()
1428 z0 = _mm256_fmadd_pd(a3, w3, z0); in MatMult_SeqBAIJ_12_AVX2()
1437 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMult_SeqBAIJ_12_AVX2()
1446 z0 = _mm256_fmadd_pd(a3, w1, z0); in MatMult_SeqBAIJ_12_AVX2()
1455 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMult_SeqBAIJ_12_AVX2()
1464 z0 = _mm256_fmadd_pd(a3, w3, z0); in MatMult_SeqBAIJ_12_AVX2()
1473 _mm256_storeu_pd(&z[0], z0); in MatMult_SeqBAIJ_12_AVX2()
2517 __m256d z0, z1, z2; in MatMultAdd_SeqBAIJ_9_AVX2() local
2554 z0 = _mm256_loadu_pd(&z[0]); in MatMultAdd_SeqBAIJ_9_AVX2()
2562 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2571 z0 = _mm256_fmadd_pd(a0, w1, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2580 z0 = _mm256_fmadd_pd(a3, w2, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2589 z0 = _mm256_fmadd_pd(a0, w3, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2598 z0 = _mm256_fmadd_pd(a3, w0, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2607 z0 = _mm256_fmadd_pd(a0, w1, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2616 z0 = _mm256_fmadd_pd(a0, w2, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2625 z0 = _mm256_fmadd_pd(a3, w3, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2634 z0 = _mm256_fmadd_pd(a0, w0, z0); in MatMultAdd_SeqBAIJ_9_AVX2()
2641 _mm256_storeu_pd(&z[0], z0); in MatMultAdd_SeqBAIJ_9_AVX2()