/external/XNNPACK/src/qs8-igemm/gen/ |
D | 8x16c4-minmax-neondot.c | 438 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 439 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 440 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 441 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 442 …int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 443 …int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 444 …int8x16_t vout6x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc6x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 445 …int8x16_t vout7x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc7x01234567), vqmovn_s16(vacc7x89ABCD… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 491 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 492 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() [all …]
|
D | 6x16c4-minmax-neondot.c | 352 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 353 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 354 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 355 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 356 …int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 357 …int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCD… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 395 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 396 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 397 …int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 405 …vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0… in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() [all …]
|
D | 4x16c4-minmax-neondot.c | 266 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 267 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 268 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 269 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 299 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 300 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 306 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 307 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
|
D | 4x16c2-minmax-neon-mull-padal-dup.c | 460 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 461 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 462 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 463 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 493 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 494 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 500 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 501 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
|
D | 4x16c8-minmax-neon-mull-padal.c | 546 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 547 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 548 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 549 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 579 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 580 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 586 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() 587 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 2x16c2-minmax-neon-mull-padal-dup.c | 286 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() 287 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() 309 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() 313 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
|
D | 8x8c4-minmax-neondot.c | 300 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x012… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 301 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x012… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 302 …int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x012… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot() 303 …int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x012… in xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot()
|
D | 4x16-minmax-neon-mull-addw-dup.c | 625 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup() 626 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup() 627 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup() 628 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup() 658 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup() 659 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup() 665 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup() 666 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
|
D | 4x16-minmax-neon-mlal-lane.c | 544 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane() 545 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane() 546 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane() 547 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane() 577 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane() 578 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane() 584 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane() 585 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 8x16c4-minmax-neondot.c | 414 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 415 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 416 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 417 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 418 …int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 419 …int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 420 …int8x16_t vout6x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc6x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 421 …int8x16_t vout7x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc7x01234567), vqmovn_s16(vacc7x89ABCD… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 477 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 478 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() [all …]
|
D | 6x16c4-minmax-neondot.c | 332 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 333 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 334 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 335 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 336 …int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 337 …int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCD… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 383 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 384 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 385 …int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 393 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() [all …]
|
D | 4x16c4-minmax-neondot.c | 250 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 251 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 252 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 253 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 289 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 290 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 296 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 297 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
|
D | 4x16c2-minmax-neon-mull-padal-dup.c | 440 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 441 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 442 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 443 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 476 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 477 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 483 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 484 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
|
D | 4x16c8-minmax-neon-mull-padal.c | 526 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 527 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 528 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 529 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 562 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 563 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 569 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() 570 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 8x8c4-minmax-neondot.c | 276 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x012… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 277 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x012… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 278 …int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x012… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot() 279 …int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x012… in xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot()
|
D | 2x16c2-minmax-neon-mull-padal-dup.c | 270 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() 271 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() 294 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup() 298 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
|
D | 4x16-minmax-neon-mlal-lane.c | 525 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane() 526 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane() 527 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane() 528 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane() 561 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane() 562 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane() 568 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane() 569 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
|
D | 4x16-minmax-neon-mull-addw-dup.c | 605 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup() 606 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup() 607 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup() 608 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup() 641 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup() 642 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup() 648 …vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup() 649 …vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0… in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
|
/external/ruy/ruy/ |
D | pack_arm.cc | 2430 vst1q_s8(dst_ptr, vcombine_s8(val0, val8)); in Pack8bitRowMajorForNeon() 2431 vst1q_s8(dst_ptr + 16, vcombine_s8(val1, val9)); in Pack8bitRowMajorForNeon() 2433 vst1q_s8(dst_ptr, vcombine_s8(val2, val10)); in Pack8bitRowMajorForNeon() 2434 vst1q_s8(dst_ptr + 16, vcombine_s8(val3, val11)); in Pack8bitRowMajorForNeon() 2437 vst1q_s8(dst_ptr, vcombine_s8(val4, val12)); in Pack8bitRowMajorForNeon() 2438 vst1q_s8(dst_ptr + 16, vcombine_s8(val5, val13)); in Pack8bitRowMajorForNeon() 2440 vst1q_s8(dst_ptr, vcombine_s8(val6, val14)); in Pack8bitRowMajorForNeon() 2441 vst1q_s8(dst_ptr + 16, vcombine_s8(val7, val15)); in Pack8bitRowMajorForNeon()
|
/external/libgav1/libgav1/src/dsp/arm/ |
D | motion_field_projection_neon.cc | 40 const int8x16_t kOneQ = vcombine_s8(kOne, kOne); in LoadDivision() 43 const int8x16_t t1 = vcombine_s8(tt.val[0], tt.val[1]); in LoadDivision() 117 const int8x16_t position_xy = vcombine_s8(*position_x8, *position_y8); in GetPosition() 124 const int8x16_t floor_xy = vcombine_s8(x8_floor8, y8_floor8); in GetPosition() 125 const int8x16_t ceiling_xy = vcombine_s8(x8_ceiling8, y8_ceiling8); in GetPosition()
|
/external/XNNPACK/src/qs8-vaddc/gen/ |
D | minmax-neon-ld64-x32.c | 82 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32() 83 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32()
|
/external/llvm-project/clang/test/CodeGen/ |
D | aarch64-neon-vcombine.c | 11 return vcombine_s8(low, high); in test_vcombine_s8()
|
/external/clang/test/CodeGen/ |
D | aarch64-neon-vcombine.c | 11 return vcombine_s8(low, high); in test_vcombine_s8()
|
/external/XNNPACK/src/qs8-vadd/gen/ |
D | minmax-neon-ld64-x32.c | 94 … int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF)); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() 95 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
|
/external/libjpeg-turbo/simd/arm/ |
D | jidctfst-neon.c | 412 int8x16_t cols_01_s8 = vcombine_s8(vqshrn_n_s16(col0, PASS1_BITS + 3), in jsimd_idct_ifast_neon() 414 int8x16_t cols_45_s8 = vcombine_s8(vqshrn_n_s16(col4, PASS1_BITS + 3), in jsimd_idct_ifast_neon() 416 int8x16_t cols_23_s8 = vcombine_s8(vqshrn_n_s16(col2, PASS1_BITS + 3), in jsimd_idct_ifast_neon() 418 int8x16_t cols_67_s8 = vcombine_s8(vqshrn_n_s16(col6, PASS1_BITS + 3), in jsimd_idct_ifast_neon()
|