Slashdot Mirror
Transcoding in 1/5 the Time with Help from the GPU
mikemuch writes "ExtremeTech's Jason Cross got a lead about a technology ATI is developing called Avivo Transcode that will use ATI graphics cards to cut down the time it takes to transcode video by a factor of five. It's part of the general-purpose computation on GPU movement. The Aviva Transcode software can only work with ATI's latest 1000-series GPUs, and the company is working on profiles that will allow, for example, transcoding DVDs for Sony's PSP."
My educated guess is, No, there won't be Linux support..
ATI was the leader in MPEG2 acceleration, enabling iDCT+MC offload to their video processor almost 10 years ago. How'd that go in terms of Linux support, you ask? Well, we're still waiting for that to be enabled in Linux.
Nvidia and S3/VIA/Unichrome have drivers that support XvMC, but ATI is notably absent from the game they created. So, I won't hold my breath on Linux support for this very cool feature.
I wonder if http://www.gpgpu.org/ could offload some of the Slashdot effect to their GPU?
Maybe others have had this idea. Maybe it's too expensive or just not practical. Imagine using PCI cards with a handful of FPGAs on board to provide reconfigurable heavy number crunching abilities to specific applications. Processes designed to use them will use one or more FPGAs if they are available, else they'll fall back to using the main CPU in "software mode."
This should be written in Shader Language (or whatever it's called these days) which is portable between cards. There's no reason NOT to release this on any platform. Since it only runs on the latest ATI cards it probably uses some feature that nVidia will have in it's next batch of cards as well. If ATI doesn't release it for Linux and the Mac hopefully it won't be that difficult to duplicate their efforts. After all, shader programs are uploaded to the video driver as plain text.... ;)
I'd like to see it but I wonder what the quality is going to be like as compared to the best current encoders. I mean you can already see a big difference between cinema craft and j. random mpeg2 encoder...
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
But will the outputs have to be certified by Hollywood or the media industry? You know, because the only reason for processing audio or video is to steal profits from Sony, BMG, Warner, ... and renegade hacker tactics like A/D conversion should be legislated back to the hell they came from
GPUs are massively parallel DSP engines. That makes them ideally suited for the task. They can do things like "let's multiply 8 different floats in parallel at once". Which is useful when doing transforms like the iDCT or DCT which are capable of taking advantge of the parallelism.
But don't take that out of context. Ask a GPU to compile the linux kernel [which is possible] and an AMD64 will spank it something nasty. *GENERAL* purpose processors are slower at these very dedicated tasks but at the same time capable of doing quite a bit with reasonable performance.
By the same token, a custom circuit can compute AES in 11 cycles [1 if pipelined] at 300Mhz which when you scale to 2.2Ghz [for your typical AMD64] amounts to ~80 cycles. AES on the AMD64 takes 260 cycles. But, ask that circuit to compute SHA-1 and it can't. Or ask it render a dialog box, etc...
Tom
Someday, I'll have a real sig.
It looks like they're using their own codec to produce MPEG-2 and MPEG-4 material. How would you get an existing, x86-only aware application to utilize the GPU, which is not x86 instruction compatible? It's a good bet that codecs will be rewritten to utilize the GPU once code becomes available from ATI, nVidia, etc.
I'd actually be willing to spend more than $50 on a video card if more multimedia apps took advantage of the GPU's capabilities.
wow, for once there's a slashdot article i have insight on! (whether it's modded that way remains to be seen.... ;) )
i would actually be shocked if there weren't linux support. the ability to do what they want only need to be in the drivers. i've been doing a gpgpu feasability study as an internship and did an mpi video compressor (based on ffmpeg) in school. using a gpu for compression/transcoding is a project i was thinking of starting once i finally had some free time since it seems built for it. something like 24 instances running at once at a ridiculous amount of flops (puts a lot of cpus to shame, actually). if you have a simd project with 4D or under vectors, this is the way to go.
like i said, it really depends on the drivers. as long as they support some of the latest opengl extensions, you're good to go. languages like Cg and BrookGPU, as well as other shader languages, are cross-platform. they can also be used with directx, but fuck that. i prefer Cg, but ymmv. actually, the project might not be that hard, just needs enought people porting the algorithms to something like Cg.
that said, don't expect this to be great unless your video card is pci-express. the agp bus is heavily asymmetric towards data going out to the gpu. as more people start getting the fatter, more symmetric pipes of pci-e, look for more gpgpu projects to take off.
The IRS is the one organization that you don't want to fuck with. Remember, these are the guys who took down Al Capone.
Though I am sure you wrote that as a pure joke, this has already been done long ago. During the fierce competion on the demo scene between the ATARI ST and the Amiga, crews were exploiting every speck of power they could from their machine. The ATARI ST being a general purpose machine compared to the Amiga (which had very advanced sound and graphical custom processors), the programmers who wanted to pull off the same graphical effects went as far as using the processor managing the keyboard (a 68xx 8bit motorola chip) for added computational power.
1. On another note, as polygon counts skyrocket they approach single pixel size
This is not happening. Not anywhere (except maybe production rendering). It is far too time-consuming, expensive, and labor-intensive to produce huge numbers of high-polygon-count models for games. Vertex pipes are currently under-utilized in most games and applications now. Efforts are underway to allow procedural geometry creation on the GPU to better fill the vertex pipe without requiring huge content creation efforts. See this paper for details.
2. A second core that most apps don't know how to take advantage of will make this all the more obvious.
This undercuts the argument you make in the next paragraph. Also, it's not true. Both the PS3 and XBOX 360 have multiple CPU cores. It's true that current-gen engines aren't optimized for this technology, but next-gen engines will be.
3. multicore CPUs are nearing the point where full screen, real time ray tracing will be possible. GPUs will not stand a chance.
This might be true, but so what? Ray tracing offers few advantages over the current-gen programmable pipeline. I can only think of 2 things that a ray-tracer can do that the programmable pipeline can't: multilevel reflections and refraction. BRDFs, soft shadows, self-shadowing, etc. can all be handled in the GPU these days. Now, you can get great results by coupling a ray-tracer with a global illumination system like photon mapping, but that technique is nowhere near real-time. Typical acceleration schemes for ray-tracing and photon mapping will not work well in dynamic environments, but the GPU could care less whether a polygon was somewhere else on the previous frame.
Hate to break it to you, but the GPU is here to stay. Why? GPUs are specialized for processing 4-vectors, not single floats (or doubles) like the CPU + FPU. True, there are CPU extensions for this, such as SSE and 3DNOW, but typical CPUs have a single SSE processor, compared to a current-gen GPU with 8 vertex pipes and 24 pixel pipes. Finally, do you really want to burden your extra CPU with rendering when it could be handling physics or AI?