87         __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()  local
96         vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c0, vacc2x01234567);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
99         vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc0, vacc2x89ABCDEF);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
103         va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
110         vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c1, vacc2x01234567);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
113         vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc1, vacc2x89ABCDEF);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
117         va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
124         vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c2, vacc2x01234567);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
127         vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc2, vacc2x89ABCDEF);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
131         va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
138         vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c3, vacc2x01234567);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
141         vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc3, vacc2x89ABCDEF);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
153           const __m256 va2 = _mm256_broadcast_ss(a2);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()  local
162           vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()
165           vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF);  in xnn_f32_igemm_ukernel_3x16s4__fma3_broadcast()