99         const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
100         const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
101         const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
102         const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
103         const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
104         const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
105         const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
106         const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
107         const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
183         const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
184         const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
185         const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
186         const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
187         const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
188         const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
189         const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
190         const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
289         const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
290         const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
291         const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
292         const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
293         const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
294         const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
295         const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
296         const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
364         const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
365         const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
366         const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
367         const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
368         const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
369         const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
370         const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
371         const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()