27   const __m128i vy_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.y_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()  local
50     __m128i vyprod01234567hi = _mm_mulhi_epu16(vy01234567, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
52     const __m128i vyprod01234567lo = _mm_mullo_epi16(vy01234567, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
54     __m128i vyprod89ABCDEFhi = _mm_mulhi_epu16(vy89ABCDEF, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
56     const __m128i vyprod89ABCDEFlo = _mm_mullo_epi16(vy89ABCDEF, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
58     __m128i vyprodGHIJKLMNhi = _mm_mulhi_epu16(vyGHIJKLMN, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
60     const __m128i vyprodGHIJKLMNlo = _mm_mullo_epi16(vyGHIJKLMN, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
62     __m128i vyprodOPQRSTUVhi = _mm_mulhi_epu16(vyOPQRSTUV, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
64     const __m128i vyprodOPQRSTUVlo = _mm_mullo_epi16(vyOPQRSTUV, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
76 …= _mm_sub_epi16(vyprod01234567hi, _mm_and_si128(_mm_srai_epi16(vy01234567, 15), vy_multiplier_lo));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
78 …= _mm_sub_epi16(vyprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vy89ABCDEF, 15), vy_multiplier_lo));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
80 …= _mm_sub_epi16(vyprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vyGHIJKLMN, 15), vy_multiplier_lo));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
82 …= _mm_sub_epi16(vyprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vyOPQRSTUV, 15), vy_multiplier_lo));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
151       __m128i vyprod01234567hi = _mm_mulhi_epu16(vy01234567, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
153       const __m128i vyprod01234567lo = _mm_mullo_epi16(vy01234567, vy_multiplier_lo);  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()
159 …= _mm_sub_epi16(vyprod01234567hi, _mm_and_si128(_mm_srai_epi16(vy01234567, 15), vy_multiplier_lo));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x32()