Lines Matching refs:vo0p0
86 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
96 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
99 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
102 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
105 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
108 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
118 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
121 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
124 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
127 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
130 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
146 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
149 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
152 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
155 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
158 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
168 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
171 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
174 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
180 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
196 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
199 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
202 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
205 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
208 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
212 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
223 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
240 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
243 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
246 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
249 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
252 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
262 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
265 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
268 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
271 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
274 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
290 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
293 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
296 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
299 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
302 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
312 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
315 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
318 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
321 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
324 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
340 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
343 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
346 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
349 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
352 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
356 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
370 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
380 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
383 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
386 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
389 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
392 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
402 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
405 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
408 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
411 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
414 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
424 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
427 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
430 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
433 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
436 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
447 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
450 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
453 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
456 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
459 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
469 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
472 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
475 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
478 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
481 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
485 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()