Lines Matching refs:vmax
75 __m128 vmax = vi0; in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
78 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
79 vmax = _mm_max_ps(vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
82 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
83 vmax = _mm_max_ps(vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
86 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
87 vmax = _mm_max_ps(vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
90 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
91 vmax = _mm_max_ps(vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
94 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
95 vmax = _mm_max_ps(vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
98 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
99 vmax = _mm_max_ps(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
102 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
103 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
106 const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
107 vmax = _mm_max_ps(vi8, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
110 _mm_store_ps(ab, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
160 __m128 vmax = _mm_load_ps(ab); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
163 const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
164 vmax = _mm_max_ps(vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
167 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
169 vmax = _mm_max_ps(vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
172 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
174 vmax = _mm_max_ps(vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
177 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
179 vmax = _mm_max_ps(vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
182 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
184 vmax = _mm_max_ps(vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
187 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
189 vmax = _mm_max_ps(vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
192 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
194 vmax = _mm_max_ps(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
197 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
199 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
202 _mm_store_ps(ab, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
273 __m128 vmax = _mm_load_ps(ab); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
278 const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
279 vmax = _mm_max_ps(vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
282 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
284 vmax = _mm_max_ps(vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
287 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
289 vmax = _mm_max_ps(vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
292 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
294 vmax = _mm_max_ps(vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
297 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
299 vmax = _mm_max_ps(vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
302 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
304 vmax = _mm_max_ps(vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
307 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
309 vmax = _mm_max_ps(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
312 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
314 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
317 _mm_storeu_ps(o, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
332 __m128 vmax = _mm_load_ps(ab); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
335 const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
336 vmax = _mm_max_ps(vi0, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
339 const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
341 vmax = _mm_max_ps(vi1, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
344 const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
346 vmax = _mm_max_ps(vi2, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
349 const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
351 vmax = _mm_max_ps(vi3, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
354 const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
356 vmax = _mm_max_ps(vi4, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
359 const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
361 vmax = _mm_max_ps(vi5, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
364 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
366 vmax = _mm_max_ps(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
369 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
371 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
375 _mm_storel_pi((__m64*) o, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
377 vmax = _mm_movehl_ps(vmax, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
383 _mm_store_ss(o, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()