87       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()  local
97       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
106       vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
115       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
124       vacc01234567p0 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
133       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
142       vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
151       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
160       vacc01234567p0 = _mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
169       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
175       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
185       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()  local
191       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
197       vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
203       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
209       vacc01234567p0 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
215       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
221       vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
227       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
233       vacc01234567p0 = _mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
239       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
244       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
255       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()  local
259       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
263       vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
267       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
271       vacc01234567p0 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
275       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
279       vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
283       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
287       vacc01234567p0 = _mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
291       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()
294       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3()