/aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/ |
H A D | 4x16c8-minmax-rndnu-neon-mull.c | 336 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local 465 const int32x2_t vsum3xEF = vpadd_s32(vpsum3xE, vpsum3xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
|
H A D | 4x16c16-minmax-rndnu-neon-mlal.c | 400 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local 528 const int32x2_t vsum3xEF = vpadd_s32(vpsum3xE, vpsum3xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
|
H A D | 4x16c8-minmax-rndnu-neon-mlal.c | 575 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local 704 const int32x2_t vsum3xEF = vpadd_s32(vpsum3xE, vpsum3xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
|
H A D | 4x16c4s2-minmax-rndnu-neon-mull.c | 339 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull() local
|
H A D | 4x16c4-minmax-rndnu-neon-mull-ld1r.c | 435 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
|
H A D | 4x16c4-minmax-rndnu-neon-mull-dup.c | 431 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
|
H A D | 4x16c4-minmax-rndnu-neon-mull-ld2r.c | 431 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
|
H A D | 4x16c4s2-minmax-rndnu-neon-mlal.c | 584 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
|
H A D | 4x16c4-minmax-rndnu-neon-mlal-dup.c | 687 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
|
H A D | 4x16c4-minmax-rndnu-neon-mlal-ld1r.c | 699 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
|
H A D | 4x16c4-minmax-rndnu-neon-mlal-ld2r.c | 687 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
|
/aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/ |
H A D | 4x16c8-minmax-rndnu-neon-mull.c | 316 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local 445 const int32x2_t vsum3xEF = vpadd_s32(vpsum3xE, vpsum3xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
|
H A D | 4x16c16-minmax-rndnu-neon-mlal.c | 380 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local 508 const int32x2_t vsum3xEF = vpadd_s32(vpsum3xE, vpsum3xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
|
H A D | 4x16c8-minmax-rndnu-neon-mlal.c | 555 const int32x4_t vsum3xEF = vpaddq_s32(vacc3x14, vacc3x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local 684 const int32x2_t vsum3xEF = vpadd_s32(vpsum3xE, vpsum3xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
|
H A D | 4x16c4s2-minmax-rndnu-neon-mull.c | 319 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull() local
|
H A D | 4x16c4-minmax-rndnu-neon-mull-dup.c | 412 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
|
H A D | 4x16c4-minmax-rndnu-neon-mull-ld1r.c | 416 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
|
H A D | 4x16c4-minmax-rndnu-neon-mull-ld2r.c | 412 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
|
H A D | 4x16c4s2-minmax-rndnu-neon-mlal.c | 564 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
|
H A D | 4x16c4-minmax-rndnu-neon-mlal-ld1r.c | 680 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
|
H A D | 4x16c4-minmax-rndnu-neon-mlal-dup.c | 668 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
|
H A D | 4x16c4-minmax-rndnu-neon-mlal-ld2r.c | 668 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
|