/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x16c8-minmax-neon-mull-padal.c | 250 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 341 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 298 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 389 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 369 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 472 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 305 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 408 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 439 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 530 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 544 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 647 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x16c8-minmax-neon-mull-padal.c | 268 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 359 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 316 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 407 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 325 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 428 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 389 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 492 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 457 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 548 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 564 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 667 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|