Home
last modified time | relevance | path

Searched refs:vo1_lo (Results 1 – 25 of 45) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3p1-minmax-neon-2x4-acc2.c204 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
206 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
210 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
213 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
D3x3p1-minmax-neonfma-2x4-acc2.c204 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
206 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
210 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
213 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
D3x3p1-minmax-neon-2x4.c200 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4() local
202 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
206 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
209 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
D3x3p1-minmax-neonfma-2x4.c200 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4() local
202 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
206 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
209 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
D3x3s2p1-minmax-neonfma-2x4-acc2.c210 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
212 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
216 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
219 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
D3x3s2p1-minmax-neon-2x4.c206 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
208 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
212 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
215 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
D3x3s2p1-minmax-neonfma-2x4.c206 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
208 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
212 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
215 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
D3x3p1-minmax-neonfma-3x4.c242 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4() local
246 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
250 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
255 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
D3x3p1-minmax-neon-3x4.c242 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4() local
246 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
250 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
255 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
D3x3s2p1-minmax-neonfma-3x4.c258 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
262 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
266 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
271 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
D3x3s2p1-minmax-neon-3x4.c258 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
262 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
266 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
271 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
D3x3p1-minmax-neonfma-4x4.c284 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4() local
290 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4()
294 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4()
301 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4()
D3x3p1-minmax-neon-4x4.c284 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4() local
290 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4()
294 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4()
301 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4()
D3x3p1-minmax-neon-5x4.c326 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
334 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
338 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
347 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
D3x3s2p1-minmax-neonfma-4x4.c310 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
316 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
320 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
327 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
D3x3s2p1-minmax-neon-4x4.c310 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
316 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
320 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
327 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
D3x3p1-minmax-neonfma-5x4.c326 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
334 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
338 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
347 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
D5x5s2p2-minmax-neon-2x4-acc2.c390 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
392 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
396 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
399 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
D5x5s2p2-minmax-neonfma-2x4-acc2.c390 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
392 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
396 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
399 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
D5x5s2p2-minmax-neon-2x4.c386 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
388 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
392 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
395 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
D5x5s2p2-minmax-neon-2x4-acc3.c394 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
396 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
400 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
403 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
D3x3p1-minmax-neon-6x4.c368 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
378 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
382 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
393 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
D5x5s2p2-minmax-neonfma-2x4-acc3.c394 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
396 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
400 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
403 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4.c386 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
388 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
392 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
395 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
D3x3p1-minmax-neonfma-6x4.c368 float32x2_t vo1_lo = vget_low_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
378 vst1_f32(o1, vo1_lo); o1 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
382 vo1_lo = vget_high_f32(vo1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
393 vst1_lane_f32(o1, vo1_lo, 0); o1 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()

12