/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c16-minmax-neon-mlal-padal.c | 219 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 278 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 2x16c8-minmax-neon-mull-padal.c | 187 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 246 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 242 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 313 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 326 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 385 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 290 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 361 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 361 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 444 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 297 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 380 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 431 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 502 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 536 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 619 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c8-minmax-neon-mull-padal.c | 203 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 262 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 235 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 294 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 342 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 401 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 260 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 331 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 308 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 379 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 317 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 400 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 381 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 464 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 449 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 520 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 556 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 639 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|