85       __m128 vacc0123p0 = _mm_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()  local
95       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
104       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
113       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
122       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
131       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
140       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
149       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
158       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
167       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
173       __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
183       __m128 vacc0123p0 = _mm_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()  local
189       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
195       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
201       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
207       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
213       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
219       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
225       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
231       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
237       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
242       __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
249       __m128 vacc0123p0 = _mm_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()  local
253       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
257       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
261       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
265       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
269       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
273       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
277       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
281       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
285       vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123));  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
288       __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()