Home
last modified time | relevance | path

Searched refs:vadd_f32 (Results 1 – 25 of 63) sorted by relevance

123

/external/XNNPACK/src/f32-gavgpool-cw/
Dneon-x4.c78 const float32x4_t vsum01 = vcombine_f32(vadd_f32(vget_low_f32(vsum0), vget_high_f32(vsum0)), in xnn_f32_gavgpool_cw_ukernel__neon_x4()
79 vadd_f32(vget_low_f32(vsum1), vget_high_f32(vsum1))); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
80 const float32x4_t vsum23 = vcombine_f32(vadd_f32(vget_low_f32(vsum2), vget_high_f32(vsum2)), in xnn_f32_gavgpool_cw_ukernel__neon_x4()
81 vadd_f32(vget_low_f32(vsum3), vget_high_f32(vsum3))); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
114 float32x2_t vsum = vadd_f32(vget_low_f32(vsum0), vget_high_f32(vsum0)); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Dneonfma-p5-x4.c96 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x4()
153 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x4()
166 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x4()
Dneon-p5-x4.c97 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x4()
154 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x4()
167 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x4()
Dneonfma-lut64-p2-x4.c115 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x4()
191 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x4()
204 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x4()
Dneon-lut64-p2-x4.c116 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x4()
192 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x4()
205 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x4()
Dneonfma-p5-x8-acc2.c172 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
229 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
242 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
Dneon-p5-x8.c170 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
227 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
240 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
Dneon-p5-x8-acc2.c173 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
230 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
243 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
Dneonfma-p5-x8.c169 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
226 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
239 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
Dneonfma-p5-x12.c185 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
242 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
255 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
Dneon-p5-x12-acc2.c189 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
246 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
259 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
Dneonfma-p5-x12-acc2.c188 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
245 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
258 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
Dneon-p5-x12.c186 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
243 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
256 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
Dneon-lut64-p2-x8.c214 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
290 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
303 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
Dneonfma-lut64-p2-x8-acc2.c216 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
292 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
305 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
Dneon-p5-x12-acc3.c191 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
248 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
261 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
Dneon-lut64-p2-x8-acc2.c217 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
293 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
306 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
Dneonfma-p5-x12-acc3.c190 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
247 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
260 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
Dneonfma-lut64-p2-x8.c213 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
289 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
302 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
Dneon-p5-x16.c202 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
259 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
272 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
Dneonfma-p5-x16.c201 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
258 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
271 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
Dneonfma-p5-x16-acc2.c204 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
261 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
274 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
Dneon-p5-x16-acc2.c205 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
262 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
275 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
Dneonfma-p5-x16-acc4.c208 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
265 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
278 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
Dneon-p5-x16-acc4.c209 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
266 vacc_lo = vadd_f32(vacc_lo, vf_lo); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
279 … vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()

123