89         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
91         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
93         _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
95         _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
101         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
103         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
105         _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
107         _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
113         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
115         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
117         _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
119         _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
126         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
128         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
130         _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
132         _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
155         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
157         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
159         _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
161         _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
169           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
171           _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
173           _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
175           _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
183             _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
185             _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
187             _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()
189             _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));  in xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2()