/aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/ |
H A D | 3x16c16-minmax-rndnu-neon-mlal.c | 301 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local 397 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
|
H A D | 3x16c8-minmax-rndnu-neon-mull.c | 253 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local 350 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
|
H A D | 4x16c8-minmax-rndnu-neon-mull.c | 308 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local 417 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
|
H A D | 3x16c8-minmax-rndnu-neon-mlal.c | 442 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local 539 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
|
H A D | 4x16c16-minmax-rndnu-neon-mlal.c | 372 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local 480 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
|
H A D | 4x16c8-minmax-rndnu-neon-mlal.c | 547 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local 656 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
|
H A D | 3x16c4s2-minmax-rndnu-neon-mull.c | 255 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull() local
|
H A D | 3x16c4-minmax-rndnu-neon-mull-ld2r.c | 329 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
|
H A D | 3x16c4-minmax-rndnu-neon-mull-ld1r.c | 332 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
|
H A D | 3x16c4-minmax-rndnu-neon-mull-dup.c | 329 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
|
H A D | 4x16c4s2-minmax-rndnu-neon-mull.c | 307 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull() local
|
H A D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 448 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
/aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/ |
H A D | 3x16c8-minmax-rndnu-neon-mull.c | 271 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local 368 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
|
H A D | 3x16c16-minmax-rndnu-neon-mlal.c | 319 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local 415 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
|
H A D | 4x16c8-minmax-rndnu-neon-mull.c | 328 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local 437 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
|
H A D | 4x16c16-minmax-rndnu-neon-mlal.c | 392 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local 500 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
|
H A D | 3x16c8-minmax-rndnu-neon-mlal.c | 460 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local 557 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
|
H A D | 4x16c8-minmax-rndnu-neon-mlal.c | 567 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local 676 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
|
H A D | 3x16c4s2-minmax-rndnu-neon-mull.c | 273 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull() local
|
H A D | 3x16c4-minmax-rndnu-neon-mull-ld2r.c | 346 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
|
H A D | 3x16c4-minmax-rndnu-neon-mull-dup.c | 346 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
|
H A D | 3x16c4-minmax-rndnu-neon-mull-ld1r.c | 349 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
|
H A D | 4x16c4s2-minmax-rndnu-neon-mull.c | 327 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull() local
|
H A D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 466 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
H A D | 4x16c4-minmax-rndnu-neon-mull-ld1r.c | 423 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
|