Lines Matching refs:vo0p0
78 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
86 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
88 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
90 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
92 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
94 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
102 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
104 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
106 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
108 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
110 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
123 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
125 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
127 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
129 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
131 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
139 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
141 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
143 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
145 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
147 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
160 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
162 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
164 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
166 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
168 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
171 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
179 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
193 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
195 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
197 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
199 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
201 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
209 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
211 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
213 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
215 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
217 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
230 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
232 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
234 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
236 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
238 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
246 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
248 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
250 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
252 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
254 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
267 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
269 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
271 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
273 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
275 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
278 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
289 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
297 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
299 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
301 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
303 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
305 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
313 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
315 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
317 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
319 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
321 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
329 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
331 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x2345, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
333 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
335 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x2345, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
337 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
346 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
348 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
350 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
352 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
354 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
362 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
364 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
366 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
368 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
370 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
373 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()