76       __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()  local
87       vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c0, vacc2x01234567);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
91       vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc0, vacc2x89ABCDEF);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
96       va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
104       vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c1, vacc2x01234567);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
108       vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc1, vacc2x89ABCDEF);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
113       va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
121       vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c2, vacc2x01234567);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
125       vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc2, vacc2x89ABCDEF);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
130       va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
138       vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c3, vacc2x01234567);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
142       vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc3, vacc2x89ABCDEF);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
155         const __m256 va2 = _mm256_broadcast_ss(a2);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()  local
166         vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()
170         vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF);  in xnn_f32_gemm_ukernel_4x16s4__fma3_broadcast()