83         vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1)));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
87         vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2)));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
91         vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3)));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
95         vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4)));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
99         vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5)));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
103         vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6)));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
107         vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7)));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
111         vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8)));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
168         vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
173         vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
178         vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
183         vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
188         vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
193         vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
198         vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
203         vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
283         vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
288         vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
293         vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
298         vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
303         vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
308         vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
313         vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
318         vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
342         vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
347         vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
352         vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
357         vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
362         vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
367         vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
372         vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
377         vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7));  in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()