Hardware accelerated decoders

[Sagekilla]

Hello all, been a while since I've last been here.

I've been out of the loop for a while now, and I was wondering what was recommended for decoding Blu-ray disks.

Right now my bottleneck is the source, since I'm only seeing about ~20% CPU load on my machine (Core i7, I'm not expecting to see 100% but more like 60 - 80%).

I'm using FFVideoSource to open up my m2ts files, since I can't figure out how to get it to work otherwise.

On that note, what's the recommended way of actually opening .m2ts files so I can process them? DSS times out, and I have ffdshow tryouts installed.

Anyhow, any help is appreciated!

[Groucho2004]

Originally Posted by Sagekilla

Right now my bottleneck is the source, since I'm only seeing about ~20% CPU load on my machine

That's most likely because you're not using the multithreaded version of ffms2.

I have been using DGDevNV for quite a while and I'm quite happy with it so far. I get about 60-120 fps depending on the source which is fast enough for my C2Q at 3 GHz encoding to x264.

[Sagekilla]

I may try just going with the MT version of Avisynth. I don't really have that complex of a script that I use for processing my videos.

In essence it's just:

Source(...)
Crop(...)
Resize(...)
MDegrain2(...)

Hopefully the MT version of Avisynth etc works fine with that.

[Groucho2004]

You don't need the MT version of Avisynth. The multithreaded ffms2 will be sufficient, especially with a simple script.

Edit: OK, I don't know how slow MDegrain2 is so it might be worth testing various scenarios.

[Sagekilla]

I double checked, and I have the multithreaded version of ffms2.

Still getting only ~20% CPU usage during encoding though. This is my script:

src = FFVideoSource("video.m2ts", cachefile="video.m2ts.ffindex").AssumeFPS("ntsc_film")
src = src.Spline36Resize(1280, 720, 0, 128, 0, -136)

super = src.MSuper(pel=4)
bvec2 = super.MANalyse(isb=true, delta=2, blksize=8, overlap=4)
bvec1 = super.MANalyse(isb=true, delta=1, blksize=8, overlap=4)
fvec1 = super.MANalyse(isb=false, delta=1, blksize=8, overlap=4)
fvec2 = super.MANalyse(isb=false, delta=2, blksize=8, overlap=4)
src.MDegrain2(super, bvec1, fvec1, bvec2, fvec2)

Hardly what you'd call complicated.

[Groucho2004]

Originally Posted by Sagekilla

src = FFVideoSource("video.m2ts", cachefile="video.m2ts.ffindex").AssumeFPS("ntsc_film")
src = src.Spline36Resize(1280, 720, 0, 128, 0, -136)

super = src.MSuper(pel=4)
bvec2 = super.MANalyse(isb=true, delta=2, blksize=8, overlap=4)
bvec1 = super.MANalyse(isb=true, delta=1, blksize=8, overlap=4)
fvec1 = super.MANalyse(isb=false, delta=1, blksize=8, overlap=4)
fvec2 = super.MANalyse(isb=false, delta=2, blksize=8, overlap=4)
src.MDegrain2(super, bvec1, fvec1, bvec2, fvec2)

Hardly what you'd call complicated.

No, but it's what makes the script your bottleneck. I suppose you have little choice but to use Avisynth MT.

[Didée]

You want to try Avisynth-MT.

After that, you want to drop pel from 4 to 2. (pel=4 is so much slower, and so much of a waste.)

After that, you want to up blksize from 8 to 16 (and overlap from 4 to 8).

At that point, you'll start seeing "reasonable" speeds.

[Sagekilla]

How would you use SetMTMode with MVtools? I tried using the way suggested in the documentation, but I get
garbage for the video output.

[Didée]

SetMTmode(5,6)      # mode=5 at first (to not break [most] source filters)
setmemorymax(1000)

LoadPlugin("MVTools2.dll")
LoadPlugin("ffms2.dll")
Import("ffms2.avsi")

ffvideosource("source.mkv")
changefps(last,last,true) # cache a few frames of input - dont ask, just do. ;-)

setmtmode(2)  #  after source loading, switch to mode=2

o  = last

blks = 16  # sufficient for 720p / a-must for 1080p
pels =  2  #  or even 1 when in a hurry ...

sup1  = o.MSuper(pel=pels,sharp=2)
bvec1 = sup1.MAnalyse( isb=true, delta=1, blksize=blks, overlap=blks/2 )
fvec1 = sup1.MAnalyse( isb=false,delta=1, blksize=blks, overlap=blks/2 )
bvec2 = sup1.MAnalyse( isb=true, delta=2, blksize=blks, overlap=blks/2 )
fvec2 = sup1.MAnalyse( isb=false,delta=2, blksize=blks, overlap=blks/2 )

o.mdegrain2(sup1,fvec1,bvec1,fvec2,bvec2)

return(last)

[Sagekilla]

Thanks for the help Didee worked like a charm.

[onesloth]

Originally Posted by Didée

changefps(last,last,true) # cache a few frames of input - dont ask, just do. ;-)

Asking. Why do you use this? Is it discussed in another thread somewhere? Is the idea to get around the single-threaded source filter bottleneck with the cache?

[Didée]

I've encountered a few instances where a script like

setmtmode(5)
ffvideosource - or - DGsource/DGMultiSource
setmtmode(2)
filters()

actually would run *very* slow. Inserting the changefps line solved the problem. Hence, this seems like a "does never hurt, but may help tremendously" tweak.

[Dogway]

Can I omit setmtmode(5)?

[Didée]

Of course you can. Nobody forces you to use multithreading. (If you omit setmtmode(5), the script will run single-threaded.)

[Dogway]

sh** I thought setmtmode(5) was to force single threading (?) I lived in ignorance!

[Didée]

Yes.

1) To make multithreading via SetMTmode functionning, SetMTmode must be used before any video clip is loaded.

2) Most source filters are not compatible with mutlithreading.


1) + 2) =>implies=> mode(5,threads) -> sourcefilter -> mode(2) -> filterchain

[onesloth]

Originally Posted by Didée

2) Most source filters are not compatible with mutlithreading.

I haven't found this to be true. E.g. I'm pretty sure all built-in avisynth source filters are compatible.

[TheRyuu]

Originally Posted by onesloth

I haven't found this to be true. E.g. I'm pretty sure all built-in avisynth source filters are compatible.

They may be compatible (i.e. SetMTMode(2) appears to work fine) but in order to be safe any format with b frames should be using SetMTMode(5) (at least I think thats the reason if I remember IanB's post correctly).

If you're thinking about using MT avisynth though it's likely that your source filter will not be your bottleneck and with ffms2 which with some common formats is already threaded if you use the ffmpeg-mt compiled version.

[Sagekilla]

Originally Posted by Didée

I've encountered a few instances where a script like

setmtmode(5)
ffvideosource - or - DGsource/DGMultiSource
setmtmode(2)
filters()

actually would run *very* slow. Inserting the changefps line solved the problem. Hence, this seems like a "does never hurt, but may help tremendously" tweak.

So in essence all you're after is just to force Avisynth to do frame caching, since it has the implicit cache after each filter.

And what better way than through a filter that does essentially a very fast nop


Edit: On an unrelated note, it's been about 3 years since I've done any big encoding. I just re-encoded the Blu-ray version of Serenity to 720p. It came out 400 MB smaller and looks better than my older DVD rip

[Gavino]

Originally Posted by Sagekilla

So in essence all you're after is just to force Avisynth to do frame caching, since it has the implicit cache after each filter.

And what better way than through a filter that does essentially a very fast nop

The [linear=]true in the call to ChangeFPS is to ensure that frame requests to the preceding filter arrive in linear order, since seeking in a source filter can often be slow (depending on format).