54         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
57         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
60         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
63         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
66         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1);  in xnn_x8_zip_xm_ukernel__sse2()
69         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1);  in xnn_x8_zip_xm_ukernel__sse2()
72         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1);  in xnn_x8_zip_xm_ukernel__sse2()
75         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1);  in xnn_x8_zip_xm_ukernel__sse2()
78         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw2);  in xnn_x8_zip_xm_ukernel__sse2()
81         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw2);  in xnn_x8_zip_xm_ukernel__sse2()
84         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw2);  in xnn_x8_zip_xm_ukernel__sse2()
87         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw2);  in xnn_x8_zip_xm_ukernel__sse2()
90         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw3);  in xnn_x8_zip_xm_ukernel__sse2()
93         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw3);  in xnn_x8_zip_xm_ukernel__sse2()
96         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw3);  in xnn_x8_zip_xm_ukernel__sse2()
99         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw3);  in xnn_x8_zip_xm_ukernel__sse2()
117         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
120         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
123         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
126         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
129         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1);  in xnn_x8_zip_xm_ukernel__sse2()
132         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1);  in xnn_x8_zip_xm_ukernel__sse2()
135         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1);  in xnn_x8_zip_xm_ukernel__sse2()
138         *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw1);  in xnn_x8_zip_xm_ukernel__sse2()
161           *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
164           *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
167           *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
170           *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
176           *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
179           *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()
184           *((uint32_t*) output) = _mm_cvtsi128_si32(vxyzw0);  in xnn_x8_zip_xm_ukernel__sse2()