SoC 2009/ARM NEON Optimization
This project is part of Google Summer of Code 2009.
|
Abstract
Formerly used primarily in PDAs and cellphones, ARM is starting to target netbooks and laptops with its new Cortex-A series of processors. One key feature of these new processors is the NEON SIMD unit, the use of which will massively boost the performance of many multimedia applications. This project will consist of writing NEON/ARMv6 SIMD assembly for all of the major DSP functions in x264, ideally speeding x264 up by over 4-5 times.
Goals
| No. | Task Description | Status | Comment |
|---|---|---|---|
| 1 | ARM framework | Mostly Done | Waiting on more info on CPUID emulation |
| 2 | SAD/SATD | Done | Qualifying task |
| 3 | SSD | Done | |
| 4 | Luma MC | In progress | get_ref_neon complete |
| 5 | Chroma MC | In progress | Ported ffmpeg's and the general case works, but not the special cases |
| 6 | (I)DCT | In Progress | 4x4 (I)DCT done, 8x8 not started |
| 7 | (De)quant | In Progress | 4x4 quant done, rest not started |
| 8 | SSIM | In Progress | core done, end in progress |
| 9 | Zigzag | In Progress | 4x4 scan/sub frame done, rest in progress |
| 10 | Hadamard AC | Not Started | |
| 11 | Intra Prediction | Not Started | |
| 12 | sa8d | Not Started | |
| 13 | pixel variance | Not Started | |
| 14 | lowres_init | Not Started | |
| 15 | Loop Filter | Not Started | |
| 16 | intra sad/satd/sa8d | Not Started | |
| 17 | Integral init | Not Started | |
| 18 | Denoise | Not Started | |
| 19 | Decimate score | Not Started | |
| 20 | Coeff last/level_run | Not Started | |
| 21 | cabac | Not Started | |
| 22 | ARMv6 optimizations | In progress | For smaller functions, write code using ARMv6's simd extensions. For instance, ARMv6 SAD width 4 is faster on Cortex-A8. |
| 23 | Cortex-A9 optimizations | In progress | Primarily consists of running checkasm --bench on an A9, noting whether ARMv6 or NEON is faster, figuring out why, and ensuring that the faster function is selected. |
| 24 | gcc optimization | Not Started | Investigate what gcc produces for important C functions and attempt to optimize it for ARM |
| 25 | Performance Counter Optimization | Not Started | Utilize the ARMv7 performance counters to optimize stalls, cache usage, NEON->ARM data hazards, and more |
Timeline
This is an estimation.
March-April Qualifying task, work on the more important (>5% in profiles) functions
May Final projects and exams
May 23 Program Starts
May 23 - July 13 Work on writing NEON asm for
June 13 - July 13 Code
July 13 Mid Term deadline. Specific goals pending, but I'd like to have none of the functions without asm equivalents to be more than about 2% of common profiles. In other words, all the major functions would have been written in NEON ASM
July 13 - August 17 Code
August 17 Program end
