94       __m128 vmax = vi0;  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()  local
97       const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
98       vmax = _mm_max_ps(vi1, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
101       const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
102       vmax = _mm_max_ps(vi2, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
105       const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
106       vmax = _mm_max_ps(vi3, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
109       const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
110       vmax = _mm_max_ps(vi4, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
113       const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
114       vmax = _mm_max_ps(vi5, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
117       const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
118       vmax = _mm_max_ps(vi6, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
121       const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
122       vmax = _mm_max_ps(vi7, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
125       const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
126       vmax = _mm_max_ps(vi8, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
129       _mm_storeu_ps(output, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
145       __m128 vmax = vi0;  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()  local
148       const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
149       vmax = _mm_max_ps(vi1, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
152       const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
153       vmax = _mm_max_ps(vi2, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
156       const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
157       vmax = _mm_max_ps(vi3, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
160       const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
161       vmax = _mm_max_ps(vi4, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
164       const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
165       vmax = _mm_max_ps(vi5, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
168       const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
169       vmax = _mm_max_ps(vi6, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
172       const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
173       vmax = _mm_max_ps(vi7, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
176       const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax));  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
177       vmax = _mm_max_ps(vi8, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
181         _mm_storel_pi((__m64*) output, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
183         vmax = _mm_movehl_ps(vmax, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
189         _mm_store_ss(output, vmax);  in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()