/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c16-minmax-neon-mlal-padal.c | 222 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 2x16c8-minmax-neon-mull-padal.c | 190 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 245 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 329 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 293 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 364 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 300 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 434 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 539 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c8-minmax-neon-mull-padal.c | 206 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 238 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 345 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 263 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 311 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 320 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 384 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 452 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 559 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, 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() local
|