InqWell's At it today: Radeon 9700 Specs

[Dave Baumann]

http://www.theinquirer.net/?article=4431

Not much we've not already heard / guessed other than this snippet:

The Radeon 9700 will have new feature called VideoShader. This will use the pixel shader unit to accelerate MPEG video, but it won't have support for MPEG-4 decoding or encoding in hardware.

I find that believable since this is what P10 does as well - this could gve us an indication that R300's pipelines may be somewhat programmable similar to P10 - no doubt this level of programmability will be where everyone is going over time.

 

[MuFu]

http://www.theinquirer.net/?article=4431

Not much we've not already heard / guessed other than this snippet:

The Radeon 9700 will have new feature called VideoShader. This will use the pixel shader unit to accelerate MPEG video..

Aren't shading engines optimised for basic arithmetic? You'd think that use of the a Pixel Shader makes no sense for accelerating video decoding as this involves mostly cosine transform ops. Maybe they are talking about overlay/filter acceleration. I dunnoooo....

:-\

MuFu.

 

[pcchen]

You mean DCT? Although DCT has cosine in its name, all "cosines" are constants. DCT and iDCT are essentially just matrix multiplications. Of course, there are several fast algorithms that can reduce the number of multiplications. The only problem with DCT/iDCT is that they requires higher precision and range to be useful.

Another time consuming operation for decoding is motion compensation. It is just a simple addition.

 

[MuFu]

You mean DCT? Although DCT has cosine in its name, all "cosines" are constants. DCT and iDCT are essentially just matrix multiplications. Of course, there are several fast algorithms that can reduce the number of multiplications. The only problem with DCT/iDCT is that they requires higher precision and range to be useful.

Another time consuming operation for decoding is motion compensation. It is just a simple addition.

Ah, ok thanks. So basically iDCT is an multiplication operation on a matrix of cosine coefficients?

*Fires up MATLAB*

MuFu.

 

[MfA]

Has ATI not always had an iDCT engine in their chips? I doubt they would bother taking it out and trying to do it with the pixel shader hardware ... with the kind of gate counts of modern chips its more bother than its worth.

 

[MuFu]

Yes they have... but when you think how much they are packing into the R300 as it is, I'm sure they'll integrate and cut redundant cells where possible. I wouldn't be suprised if this ASIC doesn't have a fixed-funcion TnL unit as they can also cover the functionality through emulation in the shader pipeline.

MuFu.

 

[Basic]

DCT (and a lot of other transforms) can be seen as a weighted sum of base functions. DCT in mpeg uses base functions that are 8x8 pixel "images" (consisiting of cosines). Put the base images in textures and do multitexturing to sum them. Weights should also be put in textures. This would trade the beauty of DCT tricks to reduce calculations, for nice regularity that fits well into the brute force of the rendering pipeline (PS2.0).

Motion compensation:
Put last frame in a texture and do the motion compensation by either rendering lots of 8x8 pixel quads, or by dependent texture reads on a full frame quad.

All of it could be done with as little as two quads for a monochrome video frame, or four for a colour video. But it would be with some pretty complex PS2.0 though. And I don't know if it actually would be good.