5476     const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);  in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()  local
5481     const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5482     const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5525       const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);  in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()  local
5528       const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5529       const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5596     const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);  in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()  local
5601     const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5602     const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5642       const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);  in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()  local
5645       const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
5646       const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
8388     const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);  in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()  local
8393     const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
8394     const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
8437       const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567);  in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()  local
8440       const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
8441       const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
8508     const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);  in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()  local
8513     const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
8514     const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
8554       const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);  in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()  local
8557       const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()
8558       const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);  in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16()