| /external/XNNPACK/src/f32-gavgpool-spchw/ |
| D | neon-x4.c | 60 vi0 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0))); in xnn_f32_gavgpool_spchw_ukernel__neon_x4()
61 vi1 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1))); in xnn_f32_gavgpool_spchw_ukernel__neon_x4()
62 vi2 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2))); in xnn_f32_gavgpool_spchw_ukernel__neon_x4()
63 vi3 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3))); in xnn_f32_gavgpool_spchw_ukernel__neon_x4()
110 vi0 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0))); in xnn_f32_gavgpool_spchw_ukernel__neon_x4()
|
| /external/XNNPACK/src/f32-dwconv-spchw/ |
| D | 3x3s2p1-neonfma.c | 115 …const float32x4_t vi0x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vuz… in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma()
116 …const float32x4_t vi0x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vuz… in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma()
117 …const float32x4_t vi1x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vuz… in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma()
118 …const float32x4_t vi1x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vuz… in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma()
119 …const float32x4_t vi2x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vuz… in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma()
120 …const float32x4_t vi2x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vuz… in xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma()
|
| D | 5x5s2p2-neonfma.c | 256 vi0x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x468A))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
257 vi1x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x468A))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
258 vi2x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x468A))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
259 vi3x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x468A))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
260 vi4x468A = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x468A))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
262 vi0x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x579B))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
263 vi1x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x579B))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
264 vi2x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x579B))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
265 vi3x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x579B))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
266 vi4x579B = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x579B))); in xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma()
|
| D | 5x5p2-neonfma.c | 265 vi0x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x89AB))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
266 vi1x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x89AB))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
267 vi2x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x89AB))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
268 vi3x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x89AB))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
269 vi4x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x89AB))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
270 vi5x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x89AB))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
271 vi6x89AB = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi6x89AB))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
449 vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
450 vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
451 vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567))); in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
[all …]
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | neonfma-rr1-p5-div-x24.c | 165 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
166 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
167 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
168 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
169 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
170 …vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
245 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
304 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
|
| D | neonfma-rr1-p5-div-x20.c | 151 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
152 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
153 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
154 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
155 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
227 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
286 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
|
| D | neonfma-rr1-p5-div-x16.c | 137 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
138 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
139 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
140 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
209 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
268 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
|
| D | neonfma-rr1-p5-nr1recps1fma-x24.c | 189 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
190 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
191 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
192 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
193 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
194 …vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
278 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
346 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
|
| D | neonfma-rr1-p5-nr2recps-x24.c | 189 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
190 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
191 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
192 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
193 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
194 …vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
278 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
346 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
|
| D | neonfma-rr1-p5-nr2fma-x24.c | 189 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
190 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
191 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
192 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
193 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
194 …vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
278 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
346 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
|
| D | neonfma-rr1-p5-nr1recps1fma-x20.c | 172 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
173 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
174 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
175 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
176 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
257 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
325 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
| D | neonfma-rr1-p5-nr2recps-x20.c | 172 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
173 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
174 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
175 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
176 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm… in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
257 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
325 … vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff))); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | neonfma-p5-x20-acc2.c | 143 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
144 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
145 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
146 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
147 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcltq_f32(vxGHIJ, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
209 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
266 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2()
|
| D | neonfma-p5-x20-acc5.c | 146 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
147 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
148 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
149 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
150 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcltq_f32(vxGHIJ, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
215 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
272 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5()
|
| D | neon-p5-x20-acc5.c | 147 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
148 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
149 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
150 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
151 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcltq_f32(vxGHIJ, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
216 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
273 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5()
|
| D | neonfma-p5-x20.c | 142 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
143 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
144 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
145 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
146 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcltq_f32(vxGHIJ, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
206 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
263 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20()
|
| D | neon-p5-x20.c | 143 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
144 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
145 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
146 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
147 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcltq_f32(vxGHIJ, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
207 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
264 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20()
|
| D | neon-p5-x20-acc2.c | 144 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
145 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
146 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
147 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
148 …vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcltq_f32(vxGHIJ, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
210 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
267 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2()
|
| D | neonfma-p5-x16-acc4.c | 132 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
133 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
134 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
135 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
197 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
254 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
|
| D | neon-p5-x16-acc2.c | 131 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
132 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
133 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
134 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
194 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
251 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
|
| D | neon-p5-x16.c | 130 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
131 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
132 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
133 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
191 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
248 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
|
| D | neonfma-p5-x16.c | 129 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
130 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
131 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
132 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
190 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
247 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
|
| D | neonfma-p5-x16-acc2.c | 130 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
131 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
132 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
133 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
193 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
250 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
|
| D | neon-p5-x16-acc4.c | 133 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
134 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
135 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
136 …vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcltq_f32(vxCDEF, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
198 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
255 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
|
| D | neon-p5-x12.c | 117 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
118 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
119 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
175 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
232 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|