SoC 2009/ARM NEON Optimization

From VideoLAN Wiki
Jump to navigation Jump to search
This project is part of Google Summer of Code 2009.
Student: David Conrad
Mentor: Holger Lubitz


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.


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 Done 4x4 (I)DCT and 8x8 IDCT done, 8x8 DCT not started
7 (De)quant In Progress 4x4 quant done, rest not started
8 SSIM Done
9 Zigzag In Progress 4x4 scan/sub frame done, rest in progress
10 Hadamard AC Done
11 Intra Prediction In Progress Simpler functions done
12 sa8d Done
13 pixel variance Done
14 lowres_init Done
15 hpel_filter Not Started
16 Loop Filter Not Started
17 intra sad/satd/sa8d Not Started
18 Integral init Not Started
19 Denoise Not Started
20 Decimate score Not Started
21 Coeff last/level_run Not Started
22 cabac Not Started
23 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.
24 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.
25 gcc optimization Not Started Investigate what gcc produces for important C functions and attempt to optimize it for ARM
26 Performance Counter Optimization Not Started Utilize the ARMv7 performance counters to optimize stalls, cache usage, NEON->ARM data hazards, and more


I'm currently using gitorious to publish my work [1]


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 - June 14 Code

June 15 - June 21 Vacation on Chincoteague

June 22 - July 12 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