82         const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); i0 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
83         const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); i1 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
84         const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
85         const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
86         const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); i4 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
87         const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); i5 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
88         const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
89         const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
90         const __m128i vi8 = _mm_loadu_si128((const __m128i*) i8); i8 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
105         const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
106         const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
107         const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
108         const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
109         const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
110         const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
111         const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
112         const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
113         const __m128i vi8 = _mm_loadu_si128((const __m128i*) i8);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
189         const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); i0 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
190         const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); i1 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
191         const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
192         const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
193         const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); i4 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
194         const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); i5 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
195         const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
196         const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16;  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
197         const __m128i vo = _mm_loadu_si128((const __m128i*) o);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
213         const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
214         const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
215         const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
216         const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
217         const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
218         const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
219         const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
220         const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
221         const __m128i vo = _mm_loadu_si128((const __m128i*) o);  in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()