Home
last modified time | relevance | path

Searched refs:vo0_lo (Results 1 – 25 of 82) sorted by relevance

1234

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3p1-minmax-neonfma-1x4.c157 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4() local
159 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4()
161 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4()
164 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4()
D3x3p1-minmax-neonfma-1x4-acc2.c159 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2() local
161 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2()
163 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2()
166 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2()
D3x3p1-minmax-neonfma-1x4-acc3.c161 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3() local
163 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3()
165 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3()
168 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3()
D3x3p1-minmax-neon-1x4-acc2.c159 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2() local
161 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2()
163 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2()
166 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2()
D3x3p1-minmax-neon-1x4.c157 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4() local
159 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4()
161 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4()
164 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4()
D3x3p1-minmax-neon-1x4-acc3.c161 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3() local
163 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3()
165 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3()
168 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3()
D3x3p1-minmax-neon-1x4-acc4.c163 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4() local
165 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4()
167 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4()
170 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4()
D3x3p1-minmax-neonfma-1x4-acc4.c163 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4() local
165 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4()
167 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4()
170 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4()
D3x3s2p1-minmax-neon-1x4-acc3.c156 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() local
158 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3()
160 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3()
163 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3()
D3x3s2p1-minmax-neon-1x4-acc2.c154 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() local
156 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2()
158 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2()
161 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2()
D3x3s2p1-minmax-neonfma-1x4-acc2.c154 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() local
156 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2()
158 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2()
161 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2()
D3x3s2p1-minmax-neon-1x4.c152 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4() local
154 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4()
156 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4()
159 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4()
D3x3s2p1-minmax-neon-1x4-acc4.c158 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() local
160 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4()
162 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4()
165 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4()
D3x3s2p1-minmax-neonfma-1x4-acc3.c156 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() local
158 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3()
160 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3()
163 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3()
D3x3s2p1-minmax-neonfma-1x4.c152 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4() local
154 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4()
156 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4()
159 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4()
D3x3s2p1-minmax-neonfma-1x4-acc4.c158 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() local
160 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4()
162 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4()
165 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4()
D3x3p1-minmax-neonfma-2x4.c199 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4() local
203 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
205 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
210 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
D3x3p1-minmax-neonfma-2x4-acc2.c203 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
207 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
209 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
214 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
D3x3p1-minmax-neon-2x4-acc2.c203 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
207 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
209 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
214 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
D3x3p1-minmax-neon-2x4.c199 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4() local
203 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
205 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
210 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
D3x3s2p1-minmax-neonfma-2x4.c205 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
209 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
211 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
216 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
D3x3s2p1-minmax-neon-2x4-acc2.c209 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() local
213 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2()
215 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2()
220 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2()
D3x3s2p1-minmax-neon-2x4.c205 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
209 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
211 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
216 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
D3x3s2p1-minmax-neonfma-2x4-acc2.c209 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
213 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
215 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
220 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
D3x3p1-minmax-neonfma-3x4.c241 float32x2_t vo0_lo = vget_low_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4() local
247 vst1_f32(o0, vo0_lo); o0 += 2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
249 vo0_lo = vget_high_f32(vo0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
256 vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()

1234