59       __m128 vmax = vi0;  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()  local
62       const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
63       vmax = _mm_max_ps(vi1, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
66       const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
67       vmax = _mm_max_ps(vi2, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
70       const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
71       vmax = _mm_max_ps(vi3, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
74       _mm_storeu_ps(output, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
85       __m128 vmax = vi0;  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()  local
88       const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
89       vmax = _mm_max_ps(vi1, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
92       const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
93       vmax = _mm_max_ps(vi2, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
96       const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
97       vmax = _mm_max_ps(vi3, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
101         _mm_storel_pi((__m64*) output, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
103         vmax = _mm_movehl_ps(vmax, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()
109         _mm_store_ss(output, vmax);  in xnn_f32_argmaxpool_ukernel_4x__sse2_c4()