61         __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()  local
68         vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
69         vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
71         va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
76         vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
77         vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
79         va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
84         vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
85         vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
87         va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
92         vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
93         vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
101           const __m256 va0 = _mm256_broadcast_ss(a0);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()  local
108           vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()
109           vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);  in xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast()