87       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()  local
97       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
115       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
133       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
151       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
169       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
175       vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
178       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
188       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()  local
194       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
206       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
218       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
230       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
242       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
247       vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
249       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
260       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()  local
264       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
272       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
280       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
288       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
296       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
299       vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()
301       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2()