76       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
78       const __m128i vxk0 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
85       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
87       const __m128i vxk1 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
94       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
96       const __m128i vxk2 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
103       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
105       const __m128i vxk3 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
112       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
114       const __m128i vxk4 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
121       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
123       const __m128i vxk5 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
130       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
132       const __m128i vxk6 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
139       const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
141       const __m128i vxk7 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
148       const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
150       const __m128i vxk8 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
228       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
230       const __m128i vxk0 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
237       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
239       const __m128i vxk1 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
246       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
248       const __m128i vxk2 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
255       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
257       const __m128i vxk3 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
264       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
266       const __m128i vxk4 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
273       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
275       const __m128i vxk5 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
282       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
284       const __m128i vxk6 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
291       const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
293       const __m128i vxk7 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
300       const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
302       const __m128i vxk8 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()