40   const __m128i vzero = _mm_setzero_si128();  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()  local
52     const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
53     const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
54     const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
55     const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
56     const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
57     const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
58     const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
68     const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero));  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
69     const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero));  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
96       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
97       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
98       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
99       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
100       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
101       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
102       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
112       vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero));  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
113       vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero));  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
164     const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
165     const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
166     const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
167     const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
168     const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
169     const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
170     const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
180     vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero));  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
181     vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero));  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
235     const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
236     const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
237     const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
238     const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
239     const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
240     const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
241     const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
251     vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero));  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
252     vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero));  in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()