YV12 and interlaced .d2v source

[trbarry]

Well, there should not be any SSE2 currently in TomsMoComp, but something might have sneaked in somewhere and I'm testing on a P4.

But it is probably a mixed version problem. I'll down fresh copies of the current everything, including my own stuff, and test again.

- Tom

[trbarry]

I spent today in the worst of all worlds. The release version of TomsMoComp would crash but the moment I'd recompile it for debug it would start working so I couldn't find the problem.

I think that is related to the somewhat automatic inlining that VS 6 helpfully does. Anyway I turned that (back) off and TomsMoComp seems to work better now. There are too darn many compiler options and I think the default ones may have changed in Avisynth 2.5. And I don't understand most of them.

But please try www.trbarry.com/TomsMoComp.zip and let me know.

- Tom

[CruNcher]

No access violations anymore on my P4 SSE2 optimisation would be
indeed a great idear it powered mpeg2dec well

[lighty]

Yup- same here! It works on my AthlonXP without any errors (so far). As for speed testing I still have to do those but I will report ASAP.

Thx for the effort trbarry!

OK- I tested speed and I experienced speed drop from 26-28fps to 15-16fps. It seems also that when TomsMoComp deinterlace algorithm kicks in fps suddenly drops down to 3-5fps and after it's done processing those frames it gets back to 15-16fps.

LoadPlugin("C:\PROGRAM FILES\AVISYNTH2\PLUGINS\MPEG2DEC3.DLL")
LoadPlugin("C:\PROGRAM FILES\AVISYNTH2\PLUGINS\TOMSMOCOMP.DLL")
MPEG2Source("D:\Red Dwarf\video_ts\1\11.d2v")
SetMemoryMax(256)

TomsMoComp(1,5,0)

crop(18,0,688,574)
LanczosResize(608,464)

[trbarry]

No access violations anymore on my P4 SSE2 optimisation would be indeed a great idear it powered mpeg2dec well

Glad to hear that.

Actually I previously wrote at least 3 different versions of TomsMoComp using SSE2. None of them showed any significant performance improvement so I ripped it all back out just to keep things simpler. SSE2 gains an advantage when you can access a lot of 16 byte aligned fields in linear fashion but TomsMoComp is looking at surrounding pixels at offsets of +1,-1, +2, etc so it doesn't seem to help as much.

Of all my Avisynth filters, only SimpleResize has any SSE2 code that really seems to give an advantage. And even there I wrote some code that I eventually removed.

But DVD2AVI and MPEG2DEC gained about 10-20% IIRC.

I think many MMX assembler video filters are limited by the memory bandwidth of accessing all the data, not the calculations that are done on it.

- Tom

[hakko504]

@Bach

You have a valid point, written the way I have done it in that script will produce some bobbing, and previously I've used this script with good results.

function GreedyBob(clip clip, int "Topfirst")
{
Top=clip.GreedyHMA(default(Topfirst,1),0,0,0,0,0,0,0)
Bottom=clip.GreedyHMA(default(1-Topfirst,0),0,0,0,0,0,0,0)
return Interleave(Top,Bottom)
}
video=SegmentedAVISource("mjpeg_capture.avi").ConvertToYuY2() #352x576 PAL captures
GreedyBob(video)
Convolution3D(preset="MovieLQ")
VericalReduceBy2()

But that's the point I'm trying to make: by adding this function into a deinterlacer, a frame/field dependancy can be added to get rid of the bobbing effects, as it is possible to determine where edges and other moving parts are.

EDIT: Forgot C3D

[wing1]

@trbarry

This version does NOT yet support the SearchEffort=-1 parm that does a vertical screen doubling. Sorry, that needs more debugging and is turned off.

I just switched over to 2.5 and everything are working great, but I missed that vertical screen doubling feature in your filter; I used it on all my scripts. For now I do a convertoyuy2() and converttoyv12(), which defeat the whole purpose of 2.5 . Why do i do that? because of undot.dll and Marc FD's asharp.dll

Thank you guys for such great filters!

[trbarry]

Wing1 -

Sorry, I meant to get that part done right away but wrote a completely different filter instead. I'll fix that.

- Tom

[wing1]

@trbarry

thank you for responding....there is nothing to be sorry about. You are the developer ( free at that! ), and users like me can only request and eagerly await for the releases

I tried the -1 option on the 2.5 version, and it works for a few frames and then the pictures split up vertically with scramble video in between.

[trbarry]

It's funny it gives wierd results. I thought I had turned it off completely for YV12 for now.


- Tom

[JohnMK]

@ all

Thank you so much for this filter. You've undoubtedly put in a lot of effort to perfect it and I for one am appreciative.

@ trbarry

Don't lose hope with the P4 optimizations; they're certainly appreciated where and when you volunteer to implement them. More and more people are getting P4's. I just bought one in fact, after a 3-year hiatus from all Intel products.

Do you know if it would make any sense whatsoever to implement hyperthreading awareness, or is this something best done at the avisynth level? I don't know if avisynth uses multiple threads or not.

[trbarry]

Do you know if it would make any sense whatsoever to implement hyperthreading awareness, or is this something best done at the avisynth level? I don't know if avisynth uses multiple threads or not.

For up to 2 processors it's probably best at the Virtualdub level where it is now. I think Vdub uses a separate task for decoding and encoding, which already splits it rather well.

I have a single processor P4 running Windows/Me and know little about this part anyway. So I likely won't be doing it.

- Tom