680   __m256d w0, w1, w2, w3;  in MatMult_SeqBAIJ_9_AVX2()  local
724       w0 = _mm256_set1_pd(work[j * 9]);  in MatMult_SeqBAIJ_9_AVX2()
726       z0 = _mm256_fmadd_pd(a0, w0, z0);  in MatMult_SeqBAIJ_9_AVX2()
728       z1 = _mm256_fmadd_pd(a1, w0, z1);  in MatMult_SeqBAIJ_9_AVX2()
730       z2 = _mm256_fmadd_pd(a2, w0, z2);  in MatMult_SeqBAIJ_9_AVX2()
760       w0 = _mm256_set1_pd(work[j * 9 + 4]);  in MatMult_SeqBAIJ_9_AVX2()
762       z0 = _mm256_fmadd_pd(a3, w0, z0);  in MatMult_SeqBAIJ_9_AVX2()
764       z1 = _mm256_fmadd_pd(a4, w0, z1);  in MatMult_SeqBAIJ_9_AVX2()
766       z2 = _mm256_fmadd_pd(a5, w0, z2);  in MatMult_SeqBAIJ_9_AVX2()
796       w0 = _mm256_set1_pd(work[j * 9 + 8]);  in MatMult_SeqBAIJ_9_AVX2()
798       z0 = _mm256_fmadd_pd(a0, w0, z0);  in MatMult_SeqBAIJ_9_AVX2()
800       z1 = _mm256_fmadd_pd(a1, w0, z1);  in MatMult_SeqBAIJ_9_AVX2()
802       z2 = _mm256_fmadd_pd(a2, w0, z2);  in MatMult_SeqBAIJ_9_AVX2()
1334   __m256d w0, w1, w2, w3;  in MatMult_SeqBAIJ_12_AVX2()  local
1363       w0 = _mm256_set1_pd(work[0]);  in MatMult_SeqBAIJ_12_AVX2()
1365       z0 = _mm256_fmadd_pd(a0, w0, z0);  in MatMult_SeqBAIJ_12_AVX2()
1367       z1 = _mm256_fmadd_pd(a1, w0, z1);  in MatMult_SeqBAIJ_12_AVX2()
1369       z2 = _mm256_fmadd_pd(a2, w0, z2);  in MatMult_SeqBAIJ_12_AVX2()
1399       w0 = _mm256_set1_pd(work[4]);  in MatMult_SeqBAIJ_12_AVX2()
1401       z0 = _mm256_fmadd_pd(a0, w0, z0);  in MatMult_SeqBAIJ_12_AVX2()
1403       z1 = _mm256_fmadd_pd(a1, w0, z1);  in MatMult_SeqBAIJ_12_AVX2()
1405       z2 = _mm256_fmadd_pd(a2, w0, z2);  in MatMult_SeqBAIJ_12_AVX2()
1435       w0 = _mm256_set1_pd(work[8]);  in MatMult_SeqBAIJ_12_AVX2()
1437       z0 = _mm256_fmadd_pd(a0, w0, z0);  in MatMult_SeqBAIJ_12_AVX2()
1439       z1 = _mm256_fmadd_pd(a1, w0, z1);  in MatMult_SeqBAIJ_12_AVX2()
1441       z2 = _mm256_fmadd_pd(a2, w0, z2);  in MatMult_SeqBAIJ_12_AVX2()
2516   __m256d w0, w1, w2, w3;  in MatMultAdd_SeqBAIJ_9_AVX2()  local
2560       w0 = _mm256_set1_pd(work[j * 9]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2562       z0 = _mm256_fmadd_pd(a0, w0, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2564       z1 = _mm256_fmadd_pd(a1, w0, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2566       z2 = _mm256_fmadd_pd(a2, w0, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()
2596       w0 = _mm256_set1_pd(work[j * 9 + 4]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2598       z0 = _mm256_fmadd_pd(a3, w0, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2600       z1 = _mm256_fmadd_pd(a4, w0, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2602       z2 = _mm256_fmadd_pd(a5, w0, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()
2632       w0 = _mm256_set1_pd(work[j * 9 + 8]);  in MatMultAdd_SeqBAIJ_9_AVX2()
2634       z0 = _mm256_fmadd_pd(a0, w0, z0);  in MatMultAdd_SeqBAIJ_9_AVX2()
2636       z1 = _mm256_fmadd_pd(a1, w0, z1);  in MatMultAdd_SeqBAIJ_9_AVX2()
2638       z2 = _mm256_fmadd_pd(a2, w0, z2);  in MatMultAdd_SeqBAIJ_9_AVX2()