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_acc2() local
88 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
90 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
94 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
104 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
108 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
123 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
127 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
131 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
141 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
145 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
160 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
164 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
168 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
170 vo0p0 = vaddq_f32(vo0p0, vo0p1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
172 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
180 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
196 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
198 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
202 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
212 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
216 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
231 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
235 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
239 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
249 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
253 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
268 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
272 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
276 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
278 vo0p0 = vaddq_f32(vo0p0, vo0p1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
280 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
291 float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
301 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
303 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x4567, vget_low_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
307 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
317 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
321 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
331 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x2345, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
335 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x2345, vget_high_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
339 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
350 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
354 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
364 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
368 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
372 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
374 vo0p0 = vaddq_f32(vo0p0, vo0p1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
376 float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()