50       __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()  local
57       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
58       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
60       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
65       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
66       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
68       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
73       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
74       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
76       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
81       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
82       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
90         const __m256 va0 = _mm256_broadcast_ss(a0);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()  local
97         vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()
98         vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);  in xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast()