95       __m128i vidx = _mm_setzero_si128();  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()  local
99       vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
103       vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
107       vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
111       vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
115       vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
119       vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
123       vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
127       vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
131       _mm_storeu_si128((__m128i*) index, vidx);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
146       __m128i vidx = _mm_setzero_si128();  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()  local
150       vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
154       vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
158       vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
162       vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
166       vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
170       vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
174       vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
178       vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8)));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
182         _mm_storel_epi64((__m128i*) index, vidx);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
184         vidx = _mm_unpackhi_epi64(vidx, vidx);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
190         *index = (uint32_t) _mm_cvtsi128_si32(vidx);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()