62       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()  local
72       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
81       vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
90       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
99       vacc01234567p0 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
105       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
115       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()  local
121       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
127       vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
133       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
139       vacc01234567p0 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
144       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
155       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()  local
159       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
163       vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
167       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
171       vacc01234567p0 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()
174       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3()