NVIDIA CUDA加速的VC-2候选视频编解码器[Dirac Video Codec]

Edison

AbstractWith the introduction of high definition video formats (HDTV), it has become necessary to develop efficient methods and algorithms which achieve better compression rates and real-time decompression for streaming over low-bandwidth networks. In this work we focus on accelerating a modern, wavelet-based video codec, the Dirac Video Codec (DVC), by mapping most computations involved to graphics hardware. We show that these computations can be performed efficiently on modern, programmable GPUs, regarded as massively parallel co-processors through NVidia's CUDA compute paradigm. We implemented a number of widely useful CUDA computational kernels such as: overlapped block motion compensation, wavelet transforms based on the lifting scheme, and frame arithmetic. Our efficient motion compensation and frame arithmetic implementations can be used to accelerate other video codecs as well. Additionally, wavelet lifting has a wide range of applications ranging from image denoising to volume rendering. By implementing all these steps of the DVC decompressor, we achieved considerable speed-up compared to an optimized CPU implementation.

PaperA paper that explains the methods we came up with in detail has been submitted. As soon as it is accepted, the link will be put here.

Features

• Large speedup in Dirac video decoding: allows for smooth, high-definition video decoding on systems that were incapable of this before
• GPU-Accelerated operations: subband transfer, wavelet transforms, motion compensation and frame arithmethic
• Fully asynchronous and parallel processing: the GPU works in tandem with the CPU

Requirements

• NVidia CUDA Toolkit 1.1, and a matching driver installed on the system
• NVidia GeForce 8xxx series video card
• GStreamer multimedia framework (recommended)
• liboil library for vectorized inner loops

DownloadDownload the source tarball here.
Version: 4 (latest) (based on schrodinger SVN revision 512)

GITYou can always get the latest, cutting-edge version using the GIT version management system:
git clone git://diracvideo.schleef.org/git/schroedinger.gitgit checkout cudaLinks

• Dirac Video Codec Homepage
• Schroedinger Project -- The implementation on which I based this
• BBC R&D section on the Dirac Video Codec
• NVIDIA CUDA Homepage
• NVIDIA Forums > CUDA GPU Computing

AuthorsPaper:

• Wladimir J. van der Laan
• Andrei C. Jalba
• Jos B.T.M. Roerdink

Software:

• Wladimir J. van der Laan
• David Schleef (original Schroedinger library)

G81

只知道VC-1，不知道VC-2为何物？:funk:

Edison

http://dirac.sourceforge.net/specification.html

Dirac Specification


The Dirac Specification is now essentially complete. The bitstream is deemed to be frozen except for bug fixes and constraints due to levels and profiles (to be added over time). Two versions of the specification exist: Version 1 (Dirac Pro) and Version 2 (Full Dirac).

Dirac Pro is an intra-only version of Dirac and has been proposed to the Society of Motion Picture and Television Engineers (SMPTE) for standardisation as VC-2. It's suitable for professional applications e.g. in television studios or post-production. Version 1.0 is available here:

Dirac Specification Version 1.1.0 (Dirac Pro or VC-2)
pdf


(Note that this is converted from Word 2003 as this is the SMPTE format of choice, and may have some minor conversion errors. The Word version is available here.)

The full Dirac specification has reached version 2.2 and is the same as Dirac Pro version 1 except that it also includes motion compensation so that it is suitable for distribution applications (broadcasting and streaming) also. It is available here:

Dirac Specification Version 2.2.0
pdf


In both cases, the specification covers the stream syntax, stream parsing operations and decoding processes. It specifies both the core syntax for general operation and a "low delay" syntax for professional use.

bigpao007

原帖由 G81 于 2008-5-18 00:23 发表
只知道VC-1，不知道VC-2为何物？:funk:

VC1   既然是 高清的一个解码方案 ， 那估计 VC2 也是把 :loveliness:

Edison

原帖由 A骨头 于 2008-5-18 01:31 发表
shader解码X1900就干过，费力不讨好

shader解码？AMD或者是部分喜欢造谣的fans撒谎罢了，根本就是不具备硬件解码器，那些不具备的解码阶段实际是由CPU执行的。

littlemouse

cuda 的h264解码器和编码器如果能马上出来，就好了，最好还是开源的。
解码器的用户更多。


随着gpu发展，我觉得有必要对软解/硬解重新定义了。
以前通过cpu的叫软解，通过gpu bitstream的叫硬解
如果通过cuda的方式解码，我觉得应该属于软解，区别于bitstream的处理方式。
或者干脆加个名词叫G解好了（好像读起来有点……）:p

[ 本帖最后由 littlemouse 于 2008-6-8 14:24 编辑 ]

EveningRain

原帖由 A骨头 于 2008-5-18 01:31 发表
shader解码X1900就干过，费力不讨好

x1900没听说有这功能，到是r600没出来之前有说可以用shader硬接h.264 vc1

aibo

这样的话以后就不用了vp了

CUDA一下，什么码都能解:loveliness:

ayanamei

原帖由 helanmouse 于 2008-6-8 14:58 发表




坟之，E大坟头值钱啊~

CUDA加速 编码 解码
没有说是完全GPU硬编码 /解码

我也没看出这是什么坟头
r5xx 什么时候支持Shader硬解过？ 还有一度说的ATI 编码器 那个也是CPU实现的好不好

[ 本帖最后由 ayanamei 于 2008-6-8 17:30 编辑 ]

teddytao

学习了。。。。

bull

我关心的是什么时候
CUDA支持后处理，
不过么
这种没有利润的事情
NV是肯定不会去做的