680   __m256d w0, w1, w2, w3;  in MatMult_SeqBAIJ_9_AVX2()  local
751       w3 = _mm256_set1_pd(work[j * 9 + 3]);  in MatMult_SeqBAIJ_9_AVX2()
753       z0 = _mm256_fmadd_pd(a0, w3, z0);  in MatMult_SeqBAIJ_9_AVX2()
755       z1 = _mm256_fmadd_pd(a1, w3, z1);  in MatMult_SeqBAIJ_9_AVX2()
757       z2 = _mm256_fmadd_pd(a2, w3, z2);  in MatMult_SeqBAIJ_9_AVX2()
787       w3 = _mm256_set1_pd(work[j * 9 + 7]);  in MatMult_SeqBAIJ_9_AVX2()
789       z0 = _mm256_fmadd_pd(a3, w3, z0);  in MatMult_SeqBAIJ_9_AVX2()
791       z1 = _mm256_fmadd_pd(a4, w3, z1);  in MatMult_SeqBAIJ_9_AVX2()
793       z2 = _mm256_fmadd_pd(a5, w3, z2);  in MatMult_SeqBAIJ_9_AVX2()
1334   __m256d w0, w1, w2, w3;  in MatMult_SeqBAIJ_12_AVX2()  local
1390       w3 = _mm256_set1_pd(work[3]);  in MatMult_SeqBAIJ_12_AVX2()
1392       z0 = _mm256_fmadd_pd(a3, w3, z0);  in MatMult_SeqBAIJ_12_AVX2()
1394       z1 = _mm256_fmadd_pd(a4, w3, z1);  in MatMult_SeqBAIJ_12_AVX2()
1396       z2 = _mm256_fmadd_pd(a5, w3, z2);  in MatMult_SeqBAIJ_12_AVX2()
1426       w3 = _mm256_set1_pd(work[7]);  in MatMult_SeqBAIJ_12_AVX2()
1428       z0 = _mm256_fmadd_pd(a3, w3, z0);  in MatMult_SeqBAIJ_12_AVX2()
1430       z1 = _mm256_fmadd_pd(a4, w3, z1);  in MatMult_SeqBAIJ_12_AVX2()
1432       z2 = _mm256_fmadd_pd(a5, w3, z2);  in MatMult_SeqBAIJ_12_AVX2()
1462       w3 = _mm256_set1_pd(work[11]);  in MatMult_SeqBAIJ_12_AVX2()
1464       z0 = _mm256_fmadd_pd(a3, w3, z0);  in MatMult_SeqBAIJ_12_AVX2()
1466       z1 = _mm256_fmadd_pd(a4, w3, z1);  in MatMult_SeqBAIJ_12_AVX2()
1468       z2 = _mm256_fmadd_pd(a5, w3, z2);  in MatMult_SeqBAIJ_12_AVX2()
2516   __m256d w0, w1, w2, w3;  in MatMultAdd_SeqBAIJ_9_AVX2()  local
2587       w3 = _mm256_set1_pd(work[j * 9 + 3]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2589       z0 = _mm256_fmadd_pd(a0, w3, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2591       z1 = _mm256_fmadd_pd(a1, w3, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2593       z2 = _mm256_fmadd_pd(a2, w3, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()
2623       w3 = _mm256_set1_pd(work[j * 9 + 7]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2625       z0 = _mm256_fmadd_pd(a3, w3, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2627       z1 = _mm256_fmadd_pd(a4, w3, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2629       z2 = _mm256_fmadd_pd(a5, w3, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()