Question about SATD.

[brunogm]

hi,

Looking the GSoC 2009 GPU, made me a question where exatly SATD is used and where else it can be applied if a GPU version make it cost negligible. Also a list of functions that could benefit of GPU implementations would be nice to understand the potential of such approach.

Currently i'm trying to better understand the internal working of x264 using the callgrind profiler.

thks

[Dark Shikari]

Putting SATD on the GPU doesn't make sense because it's a small DSP function that's called a lot; the latency to the graphics card is a thousand times higher than the runtime of the function.

You have to port an entire algorithm to the GPU to be useful, not just a single tiny DSP function.

SATD is used for qpel motion search and non-RD mode decision and up to 15% of runtime is spent in it (potentially as high as 20% or more before Holger's speedup patch).

[brunogm]

thanks! , i'm trying to find some thing useful to make an hardware IP that could be later incorporated in some System-on-Chip for H.264 and because of the 15% maybe SATD would be a good start. Well, I will use x264 for validating the hardware implementation and the thing is to find something useful for both ways.