/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
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-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-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.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-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 | 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-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 | 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-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 | neonfma-p5-x12.c | 116 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() 117 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() 118 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() 174 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() 231 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
|
D | neon-p5-x12-acc2.c | 118 …vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() 119 …vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() 120 …vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcltq_f32(vx89AB, vdenorm_… in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() 178 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() 235 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff))); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
|
/external/XNNPACK/src/f32-sigmoid/gen/ |
D | neonfma-rr1-p5-div-x24.c | 143 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() 144 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() 145 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() 146 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() 147 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() 148 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() 185 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() 211 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
|
D | neonfma-rr1-p5-div-x20.c | 128 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() 129 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() 130 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() 131 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() 132 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() 167 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() 193 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
|
D | neonfma-rr1-p5-div-x16.c | 113 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16() 114 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16() 115 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16() 116 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16() 149 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16() 175 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
|
D | neonfma-rr1-p5-nr2recps-x24.c | 164 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 165 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 166 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 167 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 168 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 169 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 210 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 240 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
|
D | neonfma-rr1-p5-nr1recps1fma-x24.c | 164 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 165 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 166 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 167 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 168 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 169 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 210 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 240 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
|
D | neonfma-rr1-p5-nr1recps1fma-x20.c | 146 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 147 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 148 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 149 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 150 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 189 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 219 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
D | neonfma-rr1-p5-nr2recps-x20.c | 146 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 147 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 148 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 149 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 150 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 189 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 219 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
|
/external/XNNPACK/src/f32-velu/gen/ |
D | velu-neon-rr2-p6-x24.c | 156 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() 158 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() 160 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() 162 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() 164 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() 166 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() 206 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x24() 235 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x24()
|
D | velu-neonfma-rr1-p6-x24.c | 148 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() 150 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() 152 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() 154 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() 156 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() 158 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() 197 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24() 225 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x24()
|
D | velu-neonfma-rr1-p6-x20.c | 133 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() 135 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() 137 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() 139 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() 141 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() 178 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20() 206 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neonfma_rr1_p6_x20()
|
D | velu-neon-rr2-p6-x20.c | 140 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() 142 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() 144 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() 146 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() 148 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() 186 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20() 215 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f)); in xnn_f32_velu_ukernel__neon_rr2_p6_x20()
|