/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x16c8-minmax-neon-mull-padal.c | 252 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 348 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 300 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 396 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 371 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 479 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 307 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 415 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 441 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 537 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 546 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 654 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x16c8-minmax-neon-mull-padal.c | 270 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 366 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 318 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 414 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 327 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 435 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 391 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 499 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 459 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 555 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 566 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 674 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|