[Open-graphics] Nvidia has 100% decode of h.264

[Tim Schmidt]

On 4/30/07, James Richard Tyrer <tyrerj at acm.org> wrote:
> Perhaps you are correct.  When we can buy a motherboard with 2 quad core
> CPUs running at 3 GHz, then decoding H.264 should be no problem.
> Actually, you will be able to do so in a few months but the price will
> be rather high.
>
> But ... (tm)
>
> The question is how much of the system's power will be required to do
> this.  If it takes over 50%, this is still going to slow down whatever
> else you are doing on your system.
>
> So, no matter how powerful your PC (within reason) it is still going to
> be useful to have hardware that will do DCT/IDCT and YUV <> RGB
> conversion since such hardware will always be more efficient at doing
> these specialized tasks.  The same technology that makes the 2 X 4 CPU
> computer affordable will also allow 8x8 DCT/IDCT and YUV <> RGB with a
> latency of a few clocks (get it down to the theoretical minimum).  You
> should be able to get YUV <> RGB down to 3  clocks since it is a [3
> vector] * [3 x 3 matrix] + a [3 vector] -- 3 multipliers and 3 adders
> per color (you can do this today but it is too expensive).  Still
> haven't found my DSP text book so I don't know what the theoretical
> minimum would be for the DCT/IDCT
>
> If wavelets become the standard method of compression then hardware that
> is good at performing a 2D FIR filter will be the next thing.

All sounds good...  I agree that specialized hardware can be more
efficient - it's certainly more efficient than doing the exact same
thing with generalized hardware, but we need the generalized hardware
anyway, so specialized hardware needs to be a LOT more efficient to be
worthwhile in terms of efficiency.  Depending on the usage of the
system, it may be dead weight most of the time.

Implementing enough hardware to decode a specific type of video stream
(and ostensibly other types as well) seems a bit much.  It seems the
sort of problem best attacked by implementing hardware that
accelerates the tough part of the decoding as a general purpose
instruction, and allows us to support newer (and older!) codecs with a
little bit of soft/firm ware.  MMX / 3DNow! / SSE attempt to do this
on the CPU side, a GPU equivalent seems warranted.

:)

--tim