680   __m256d w0, w1, w2, w3;  in MatMult_SeqBAIJ_9_AVX2()  local
733       w1 = _mm256_set1_pd(work[j * 9 + 1]);  in MatMult_SeqBAIJ_9_AVX2()
735       z0 = _mm256_fmadd_pd(a0, w1, z0);  in MatMult_SeqBAIJ_9_AVX2()
737       z1 = _mm256_fmadd_pd(a1, w1, z1);  in MatMult_SeqBAIJ_9_AVX2()
739       z2 = _mm256_fmadd_pd(a2, w1, z2);  in MatMult_SeqBAIJ_9_AVX2()
769       w1 = _mm256_set1_pd(work[j * 9 + 5]);  in MatMult_SeqBAIJ_9_AVX2()
771       z0 = _mm256_fmadd_pd(a0, w1, z0);  in MatMult_SeqBAIJ_9_AVX2()
773       z1 = _mm256_fmadd_pd(a1, w1, z1);  in MatMult_SeqBAIJ_9_AVX2()
775       z2 = _mm256_fmadd_pd(a2, w1, z2);  in MatMult_SeqBAIJ_9_AVX2()
1334   __m256d w0, w1, w2, w3;  in MatMult_SeqBAIJ_12_AVX2()  local
1372       w1 = _mm256_set1_pd(work[1]);  in MatMult_SeqBAIJ_12_AVX2()
1374       z0 = _mm256_fmadd_pd(a3, w1, z0);  in MatMult_SeqBAIJ_12_AVX2()
1376       z1 = _mm256_fmadd_pd(a4, w1, z1);  in MatMult_SeqBAIJ_12_AVX2()
1378       z2 = _mm256_fmadd_pd(a5, w1, z2);  in MatMult_SeqBAIJ_12_AVX2()
1408       w1 = _mm256_set1_pd(work[5]);  in MatMult_SeqBAIJ_12_AVX2()
1410       z0 = _mm256_fmadd_pd(a3, w1, z0);  in MatMult_SeqBAIJ_12_AVX2()
1412       z1 = _mm256_fmadd_pd(a4, w1, z1);  in MatMult_SeqBAIJ_12_AVX2()
1414       z2 = _mm256_fmadd_pd(a5, w1, z2);  in MatMult_SeqBAIJ_12_AVX2()
1444       w1 = _mm256_set1_pd(work[9]);  in MatMult_SeqBAIJ_12_AVX2()
1446       z0 = _mm256_fmadd_pd(a3, w1, z0);  in MatMult_SeqBAIJ_12_AVX2()
1448       z1 = _mm256_fmadd_pd(a4, w1, z1);  in MatMult_SeqBAIJ_12_AVX2()
1450       z2 = _mm256_fmadd_pd(a5, w1, z2);  in MatMult_SeqBAIJ_12_AVX2()
2516   __m256d w0, w1, w2, w3;  in MatMultAdd_SeqBAIJ_9_AVX2()  local
2569       w1 = _mm256_set1_pd(work[j * 9 + 1]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2571       z0 = _mm256_fmadd_pd(a0, w1, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2573       z1 = _mm256_fmadd_pd(a1, w1, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2575       z2 = _mm256_fmadd_pd(a2, w1, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()
2605       w1 = _mm256_set1_pd(work[j * 9 + 5]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2607       z0 = _mm256_fmadd_pd(a0, w1, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2609       z1 = _mm256_fmadd_pd(a1, w1, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2611       z2 = _mm256_fmadd_pd(a2, w1, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()