mpeg2dec3 idct=5: 0% speed change

[JohnMK]

I used this idct on my P4 just now and experienced, in a benchmark, a 0% speed increase. I use mpeg2dec3 in my avs scripts created by GordianKnot, and encode to DivX in VDubMod, all YV12 colorspace of course.

Any theories?

[MaTTeR]

Actually I've been experimenting with iDCT=3 and iDCT=4 over the last fews days and suprisingly I noticed hardly any drop in performance. Running more benchmarks as we speak but from all accounts I see very little performance difference(if any) between the various iDCT settings. I can't comment on iDCT=5 though, running dual AMD chips here.

[JohnMK]

Maybe the kind of code in mpeg2dec3 simply isn't suited for SIMD, or perhaps, the right programmers weren't selected for the task?

I wonder if ICL7 could do any better . . . w/ SSE2 forced optimization switches. This wouldn't require anybody's work except the compiler's program, all . . . 10 seconds of the compile.

[trbarry]

The net result of the project to speed up DVD2AVI/MPEG2DEC2 with SSE2 last spring was just under 10% savings. But 3% of that was just letting MPEG2DEC2 do the cropping. You can read all about it on page ten of that long DVD2AVI Sourceforge Project (http://forum.doom9.org/showthread.php?s=&threadid=11442&perpage=20&highlight=sse2&pagenumber=10) thread.

But the bottom line is probably just the table from some of my own tests, copied below from that other thread:


Test Seconds Tot Pcnt Saved

mpeg2dec binary from Doom9 234

Doom9 mpeg2dec source 234 0.0 %

SSE2,IDCT=2. 225 3.8 %

SSE2,IDCT=5 218 6.8 %

SSE2, IDCT=5, mpeg2dec crop 211 9.8 %


Note that turning on IDCT=5 also (I think) turns on some other SSE2 code. I'm not sure whether MPEG2DEC3 implements all this. And since it is still in alpha I haven't tried to compare YV12 vs older YUY2 speeds yet.

- Tom

[JohnMK]

It's certainly not implemented yet.