20     const int8_t* input_x,  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()  argument
36     const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
38     const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
40     const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
42     const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
44     const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
46     const __m128i vxKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 20));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
48     const __m128i vxOPQR = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 24));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
50     const __m128i vxSTUV = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 28));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
52     input_x += 32;  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
115       const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
117       const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
119       input_x += 8;  in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()