• Home
  • Raw
  • Download

Lines Matching refs:vo0p0

102       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02));  in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()  local
103 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
104 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
105 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
106 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
131 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
132 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
133 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
134 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
135 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
159 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
160 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
161 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
162 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
163 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
171 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
172 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
173 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
174 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
175 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
183 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
184 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
185 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
186 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
187 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
190 __m128 vo0 = _mm_max_ps(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
199 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
200 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
201 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
202 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
203 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
228 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
229 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
230 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
231 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
232 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
256 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
257 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
258 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
259 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
260 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
268 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
269 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
270 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
271 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
272 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
280 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
281 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
282 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
283 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
284 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
287 __m128 vo0 = _mm_max_ps(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
305 __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
306 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk12)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
307 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk22)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
308 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
309 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x4567, vk42)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
323 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
324 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
325 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
326 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
327 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
342 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x2345, vk00)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
343 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
344 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x2345, vk20)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
345 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
346 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
354 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
355 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
356 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
357 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
358 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
366 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
367 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
368 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
369 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
370 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
373 __m128 vo0 = _mm_max_ps(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()