/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c8-minmax-neon-mull-padal.c | 265 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 269 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 297 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 301 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 334 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 338 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 404 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 408 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 382 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 386 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 403 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 407 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 467 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 471 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 523 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 527 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 642 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 646 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c8-minmax-neon-mull-padal.c | 249 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 253 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 281 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 285 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 316 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 320 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 388 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 392 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 364 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 368 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 383 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 387 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 505 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 509 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 447 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 451 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 622 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 626 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|