/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c8-minmax-neon-mull-padal.c | 268 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 270 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 300 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 302 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 337 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 339 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 407 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 409 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 385 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 387 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 406 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 408 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 470 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 472 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 526 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 528 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 645 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 647 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c8-minmax-neon-mull-padal.c | 252 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 254 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 284 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 286 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 319 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 321 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 391 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 393 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 367 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 369 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 386 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 388 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 508 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 510 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 450 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 452 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 625 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 627 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|