Results from profiling decoder

[dattrax]

All developers,

I have attached the results from my profiling experiments using rational quantify. You may find them suprising....

to sumerise

23.74% of the decode was spent in a routine yv12_to_yuv_c, which there is no mmx version of this yet, I don't know if this is a big issue, but I know a lot of programs prefer raster 4:2:0 out over other modes. It may be worth writing an mmx version of this.

12.56% in decoder_mbinter

9.49% in idct_mmx

6.61% in decoder_mbintra

Jim

[Koepi]

Thx Jim,

at least I am surprised, you're right.

But I start to dislike MMXed colour space conversion, YUV2RGB is a little unprecise in XviD, and that can be noticed if you know what you have to look for ;)

Ok, trying to enable the attachment...

Best regards,
Koepi

[-h]

Wow! I didn't expect such a big colorspace conversion hit.

I'm writing colorspace conversion code for another codec I'm working on (sigh), I might give it a shot.

Well actually the conversion you're talking about is just a couple memcpy() calls, which would be very easy to do. But it's waiting on that damn DSL modem to get here..

-h

[dattrax]

That colourspace routine is only memcpy(), so its not the fact it needs mmxing, but moreso prefetched. If we can remove the latency to the memory system, then this should come down vastly

Jim