120     __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()  local
121     vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) params->sse2.output_zero_point));  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
122     vout = _mm_packus_epi16(vout, vout);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
123     vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_max));  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
124     vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_min));  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
126     _mm_storel_epi64((__m128i*) output, vout); output += 8;  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
189     __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()  local
190     vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) params->sse2.output_zero_point));  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
191     vout = _mm_packus_epi16(vout, vout);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
192     vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_max));  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
193     vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_min));  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
196       *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
198       vout = _mm_srli_epi64(vout, 32);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
201       *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout, 0);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
203       vout = _mm_srli_epi32(vout, 16);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()
206       *((uint8_t*) output) = (uint8_t) _mm_cvtsi128_si32(vout);  in xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8()