92         const __m128i vmax018 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vi8);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
93         const __m128i vmax23 = _mm_max_epu8(vi2, vi3);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
94         const __m128i vmax45 = _mm_max_epu8(vi4, vi5);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
95         const __m128i vmax67 = _mm_max_epu8(vi6, vi7);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
97         const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
98         const __m128i vmax01678 = _mm_max_epu8(vmax018, vmax67);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
99         const __m128i vmax = _mm_max_epu8(vmax2345, vmax01678);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
100         const __m128i vout = _mm_max_epu8(_mm_min_epu8(vmax, voutput_max), voutput_min);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
115         const __m128i vmax018 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vi8);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
116         const __m128i vmax23 = _mm_max_epu8(vi2, vi3);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
117         const __m128i vmax45 = _mm_max_epu8(vi4, vi5);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
118         const __m128i vmax67 = _mm_max_epu8(vi6, vi7);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
120         const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
121         const __m128i vmax01678 = _mm_max_epu8(vmax018, vmax67);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
122         const __m128i vmax = _mm_max_epu8(vmax2345, vmax01678);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
123         __m128i vout = _mm_max_epu8(_mm_min_epu8(vmax, voutput_max), voutput_min);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
199         const __m128i vmax01 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vo);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
200         const __m128i vmax23 = _mm_max_epu8(vi2, vi3);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
201         const __m128i vmax45 = _mm_max_epu8(vi4, vi5);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
202         const __m128i vmax67 = _mm_max_epu8(vi6, vi7);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
204         const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
205         const __m128i vmax0167 = _mm_max_epu8(vmax01, vmax67);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
206         const __m128i vmax = _mm_max_epu8(vmax2345, vmax0167);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
207         const __m128i vout = _mm_max_epu8(_mm_min_epu8(vmax, voutput_max), voutput_min);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
223         const __m128i vmax01 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vo);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
224         const __m128i vmax23 = _mm_max_epu8(vi2, vi3);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
225         const __m128i vmax45 = _mm_max_epu8(vi4, vi5);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
226         const __m128i vmax67 = _mm_max_epu8(vi6, vi7);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
228         const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
229         const __m128i vmax0167 = _mm_max_epu8(vmax01, vmax67);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
230         const __m128i vmax = _mm_max_epu8(vmax2345, vmax0167);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
231         __m128i vout = _mm_max_epu8(_mm_min_epu8(vmax, voutput_max), voutput_min);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()