2598     const v_float32 v_alpha = vx_setall_f32((float)alpha);  in accW_simd_()  local
2622             v_dst00 = v_fma(v_dst00, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src00)) * v_alpha);  in accW_simd_()
2623             v_dst01 = v_fma(v_dst01, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src01)) * v_alpha);  in accW_simd_()
2624             v_dst10 = v_fma(v_dst10, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src10)) * v_alpha);  in accW_simd_()
2625             v_dst11 = v_fma(v_dst11, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src11)) * v_alpha);  in accW_simd_()
2671 …elect(v_mf00, v_fma(v_dst00, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src00)) * v_alpha), v_dst00);  in accW_simd_()
2672 …elect(v_mf01, v_fma(v_dst01, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src01)) * v_alpha), v_dst01);  in accW_simd_()
2673 …elect(v_mf10, v_fma(v_dst10, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src10)) * v_alpha), v_dst10);  in accW_simd_()
2674 …elect(v_mf11, v_fma(v_dst11, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src11)) * v_alpha), v_dst11);  in accW_simd_()
2727 …lect(v_mf00, v_fma(v_dst00, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src000)) * v_alpha), v_dst00);  in accW_simd_()
2728 …lect(v_mf01, v_fma(v_dst01, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src001)) * v_alpha), v_dst01);  in accW_simd_()
2729 …lect(v_mf10, v_fma(v_dst02, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src010)) * v_alpha), v_dst02);  in accW_simd_()
2730 …lect(v_mf11, v_fma(v_dst03, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src011)) * v_alpha), v_dst03);  in accW_simd_()
2732 …lect(v_mf00, v_fma(v_dst10, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src100)) * v_alpha), v_dst10);  in accW_simd_()
2733 …lect(v_mf01, v_fma(v_dst11, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src101)) * v_alpha), v_dst11);  in accW_simd_()
2734 …lect(v_mf10, v_fma(v_dst12, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src110)) * v_alpha), v_dst12);  in accW_simd_()
2735 …lect(v_mf11, v_fma(v_dst13, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src111)) * v_alpha), v_dst13);  in accW_simd_()
2737 …lect(v_mf00, v_fma(v_dst20, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src200)) * v_alpha), v_dst20);  in accW_simd_()
2738 …lect(v_mf01, v_fma(v_dst21, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src201)) * v_alpha), v_dst21);  in accW_simd_()
2739 …lect(v_mf10, v_fma(v_dst22, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src210)) * v_alpha), v_dst22);  in accW_simd_()
2740 …lect(v_mf11, v_fma(v_dst23, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src211)) * v_alpha), v_dst23);  in accW_simd_()
2757     const v_float32 v_alpha = vx_setall_f32((float)alpha);  in accW_simd_()  local
2773             v_dst0 = v_fma(v_dst0, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_int0)) * v_alpha);  in accW_simd_()
2774             v_dst1 = v_fma(v_dst1, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_int1)) * v_alpha);  in accW_simd_()
2805 … v_select(v_mf0, v_fma(v_dst0, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src0)) * v_alpha), v_dst0);  in accW_simd_()
2806 … v_select(v_mf1, v_fma(v_dst1, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src1)) * v_alpha), v_dst1);  in accW_simd_()
2839 …select(v_mf0, v_fma(v_dst00, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src00)) * v_alpha), v_dst00);  in accW_simd_()
2840 …select(v_mf0, v_fma(v_dst10, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src10)) * v_alpha), v_dst10);  in accW_simd_()
2841 …select(v_mf0, v_fma(v_dst20, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src20)) * v_alpha), v_dst20);  in accW_simd_()
2843 …select(v_mf1, v_fma(v_dst01, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src01)) * v_alpha), v_dst01);  in accW_simd_()
2844 …select(v_mf1, v_fma(v_dst11, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src11)) * v_alpha), v_dst11);  in accW_simd_()
2845 …select(v_mf1, v_fma(v_dst21, v_beta, v_cvt_f32(v_reinterpret_as_s32(v_src21)) * v_alpha), v_dst21);  in accW_simd_()
2860     const __m256 v_alpha = _mm256_set1_ps((float)alpha);  in accW_simd_()  local
2869 …mm256_mul_ps(_mm256_loadu_ps(dst + x), v_beta), _mm256_mul_ps(_mm256_loadu_ps(src + x), v_alpha)));  in accW_simd_()
2870 …l_ps(_mm256_loadu_ps(dst + x + 8), v_beta), _mm256_mul_ps(_mm256_loadu_ps(src + x + 8), v_alpha)));  in accW_simd_()
2874     const v_float32 v_alpha = vx_setall_f32((float)alpha);  in accW_simd_()  local
2887             v_dst0 = v_fma(v_dst0, v_beta, vx_load(src + x) * v_alpha);  in accW_simd_()
2888             v_dst1 = v_fma(v_dst1, v_beta, vx_load(src + x + step) * v_alpha);  in accW_simd_()
2902     const v_float64 v_alpha = vx_setall_f64(alpha);  in accW_simd_()  local
2930             v_dst0 = v_fma(v_dst0, v_beta, v_src0 * v_alpha);  in accW_simd_()
2931             v_dst1 = v_fma(v_dst1, v_beta, v_src1 * v_alpha);  in accW_simd_()
2932             v_dst2 = v_fma(v_dst2, v_beta, v_src2 * v_alpha);  in accW_simd_()
2933             v_dst3 = v_fma(v_dst3, v_beta, v_src3 * v_alpha);  in accW_simd_()
2949     const v_float64 v_alpha = vx_setall_f64(alpha);  in accW_simd_()  local
2976             v_dst00 = v_fma(v_dst00, v_beta, v_src00 * v_alpha);  in accW_simd_()
2977             v_dst01 = v_fma(v_dst01, v_beta, v_src01 * v_alpha);  in accW_simd_()
2978             v_dst10 = v_fma(v_dst10, v_beta, v_src10 * v_alpha);  in accW_simd_()
2979             v_dst11 = v_fma(v_dst11, v_beta, v_src11 * v_alpha);  in accW_simd_()
2995     const __m256d v_alpha = _mm256_set1_pd(alpha);  in accW_simd_()  local
3011 …, _mm256_add_pd(_mm256_mul_pd(_mm256_loadu_pd(dst + x), v_beta), _mm256_mul_pd(v_src00, v_alpha)));  in accW_simd_()
3012 …m256_add_pd(_mm256_mul_pd(_mm256_loadu_pd(dst + x + 4), v_beta), _mm256_mul_pd(v_src01, v_alpha)));  in accW_simd_()
3013 …m256_add_pd(_mm256_mul_pd(_mm256_loadu_pd(dst + x + 8), v_beta), _mm256_mul_pd(v_src10, v_alpha)));  in accW_simd_()
3014 …256_add_pd(_mm256_mul_pd(_mm256_loadu_pd(dst + x + 12), v_beta), _mm256_mul_pd(v_src11, v_alpha)));  in accW_simd_()
3018     const v_float64 v_alpha = vx_setall_f64(alpha);  in accW_simd_()  local
3040             v_dst00 = v_fma(v_dst00, v_beta, v_src00 * v_alpha);  in accW_simd_()
3041             v_dst01 = v_fma(v_dst01, v_beta, v_src01 * v_alpha);  in accW_simd_()
3042             v_dst10 = v_fma(v_dst10, v_beta, v_src10 * v_alpha);  in accW_simd_()
3043             v_dst11 = v_fma(v_dst11, v_beta, v_src11 * v_alpha);  in accW_simd_()
3059     const __m256d v_alpha = _mm256_set1_pd(alpha);  in accW_simd_()  local
3071 …x, _mm256_add_pd(_mm256_mul_pd(_mm256_loadu_pd(dst + x), v_beta), _mm256_mul_pd(v_src0, v_alpha)));  in accW_simd_()
3072 …mm256_add_pd(_mm256_mul_pd(_mm256_loadu_pd(dst + x + 4), v_beta), _mm256_mul_pd(v_src1, v_alpha)));  in accW_simd_()
3076     const v_float64 v_alpha = vx_setall_f64(alpha);  in accW_simd_()  local
3092             v_dst0 = v_fma(v_dst0, v_beta, v_src0 * v_alpha);  in accW_simd_()
3093             v_dst1 = v_fma(v_dst1, v_beta, v_src1 * v_alpha);  in accW_simd_()