680   __m256d w0, w1, w2, w3;  in MatMult_SeqBAIJ_9_AVX2()  local
742       w2 = _mm256_set1_pd(work[j * 9 + 2]);  in MatMult_SeqBAIJ_9_AVX2()
744       z0 = _mm256_fmadd_pd(a3, w2, z0);  in MatMult_SeqBAIJ_9_AVX2()
746       z1 = _mm256_fmadd_pd(a4, w2, z1);  in MatMult_SeqBAIJ_9_AVX2()
748       z2 = _mm256_fmadd_pd(a5, w2, z2);  in MatMult_SeqBAIJ_9_AVX2()
778       w2 = _mm256_set1_pd(work[j * 9 + 6]);  in MatMult_SeqBAIJ_9_AVX2()
780       z0 = _mm256_fmadd_pd(a0, w2, z0);  in MatMult_SeqBAIJ_9_AVX2()
782       z1 = _mm256_fmadd_pd(a1, w2, z1);  in MatMult_SeqBAIJ_9_AVX2()
784       z2 = _mm256_fmadd_pd(a2, w2, z2);  in MatMult_SeqBAIJ_9_AVX2()
1334   __m256d w0, w1, w2, w3;  in MatMult_SeqBAIJ_12_AVX2()  local
1381       w2 = _mm256_set1_pd(work[2]);  in MatMult_SeqBAIJ_12_AVX2()
1383       z0 = _mm256_fmadd_pd(a0, w2, z0);  in MatMult_SeqBAIJ_12_AVX2()
1385       z1 = _mm256_fmadd_pd(a1, w2, z1);  in MatMult_SeqBAIJ_12_AVX2()
1387       z2 = _mm256_fmadd_pd(a2, w2, z2);  in MatMult_SeqBAIJ_12_AVX2()
1417       w2 = _mm256_set1_pd(work[6]);  in MatMult_SeqBAIJ_12_AVX2()
1419       z0 = _mm256_fmadd_pd(a0, w2, z0);  in MatMult_SeqBAIJ_12_AVX2()
1421       z1 = _mm256_fmadd_pd(a1, w2, z1);  in MatMult_SeqBAIJ_12_AVX2()
1423       z2 = _mm256_fmadd_pd(a2, w2, z2);  in MatMult_SeqBAIJ_12_AVX2()
1453       w2 = _mm256_set1_pd(work[10]);  in MatMult_SeqBAIJ_12_AVX2()
1455       z0 = _mm256_fmadd_pd(a0, w2, z0);  in MatMult_SeqBAIJ_12_AVX2()
1457       z1 = _mm256_fmadd_pd(a1, w2, z1);  in MatMult_SeqBAIJ_12_AVX2()
1459       z2 = _mm256_fmadd_pd(a2, w2, z2);  in MatMult_SeqBAIJ_12_AVX2()
2516   __m256d w0, w1, w2, w3;  in MatMultAdd_SeqBAIJ_9_AVX2()  local
2578       w2 = _mm256_set1_pd(work[j * 9 + 2]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2580       z0 = _mm256_fmadd_pd(a3, w2, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2582       z1 = _mm256_fmadd_pd(a4, w2, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2584       z2 = _mm256_fmadd_pd(a5, w2, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()
2614       w2 = _mm256_set1_pd(work[j * 9 + 6]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2616       z0 = _mm256_fmadd_pd(a0, w2, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2618       z1 = _mm256_fmadd_pd(a1, w2, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2620       z2 = _mm256_fmadd_pd(a2, w2, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()