why avisynth interlacing is so slow?

[marcorocchini]

Hi

please enter the cats only.

However

I would use

assumeFPS(50)
AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()

but interlacing is extremely slow. Is a specific plugin for avisynth that do interlace more fast? thanks

[poisondeathray]

Use multithreading

Prefetch(some value)

[Alwyn]

Add Prefetch to the very end of your script in this format:

Prefetch(x)

where X is the number of logical processors in your CPU -1 or 2.

“The Prefetch number should be set to 1 or 2 less than the number of "logical processors" in your system. To check what that is, go to Task Manager in Windows, select the Performance tab, then click on the CPU item. You'll see down below in the window an entry for the number of logical processors in your system. For some people, it will be the same as the number of physical cores, for others it might be double that if your CPU supports Hyperthreading. In my case, I have 12 logical processors, so I set Prefetch to 10.”

Quoted from here (and it works from personal experience; Prefetch on my system increased encoding speed by 250%):

http://macilatthefront.blogspot.com/

Number of Logical Processors:


[Attachment 72864 - Click to enlarge]

[Alwyn]

Did some encoding time tests on an AVI capture using QTGMC. 9:40 length.

Prefetch (I have 24 logical processors) set to
0: 6:20
5: 2:00
10: 1:26
15 and above: 1:25

[jagabo]

Interlacing like that is very fast in AviSynth. Whether you use prefetch or not. Something else in or about your script is slow.

A 12 minute 40 second 1280x720, 50 fps video took about 11 seconds to process (VirtualDub2, Video -> Fast Recompress, File -> Run Video Analysis Pass, i7 9900K).

ColorBarsHD() #width=1280, height=720, pixel_type="YV24")
AssumeFPS(50)
Trim(0, 28999)
AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()

Adding prefetch(N) actually made it run a little slower. And in fact, removing the interlacing line actually slows the processing rate -- because twice as much data must be passed to the editor/encoder.

[marcorocchini]

[marcorocchini]

does exist an external plugin that do what the

AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()

do?

[jagabo]

does exist an external plugin that do what the

AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()

do?

I don't know of any. But why? It won't run significantly faster.

Why don't you post your entire script? Then someone will be able to tell you where the slowness is coming from.

[marcorocchini]

oh my ***

this is the source

https://www.swisstransfer.com/d/6bffde22-b94f-4666-a353-c9bcce3aa38f

LoadPlugin("V:\FFMS2_R1363_AGO2023\X64\ffms2.dll")

Import("V:\FFMS2_AVSI_2023\FFMS2.AVSI")

FFVideoSource("v:\concatHD.ts", fpsnum=50, fpsden=1).assumeFPS(50).converttoYV12(interlaced=false)

emask = mt_convolution(vertical="-1 2 -1").ColorYUV(off_y=-64).ColorYUV(gain_y=128).mt_expand(chroma="-128").Blur(0.8)
Overlay(last, Blur(0.5, 0.6), mask=emask)

ConvertToYUY2(interlaced=false)

assumeFPS(50)
AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()
Prefetch(8)

I would like to encode it in MXF 50i using "your" emask to reduce the artifacts dues by the progressive-->interlaced and overall: getting the exact duration of the source if possible


or for example consider to get the output file using virtualdub and encode to uncompressed YUY2 avi file: if I encode the script where the interlacing is due to avisynth, the duration of the outputted .avi file seems to me similar to the source.

But, if the encoding is due by virtuald, and removing the "AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()" lines in the script, and using the "interlace" filter in virtualdub, seems to me that the outputted .avi file have 1 problematic frame missing to the end.

So I I would like not to use avisynth's interlacing, but it seems that: not using it, results in a bad final file of 1 missing frame

[jagabo]

Interlacing requires pairs of frames from the input -- one field comes from each of the two input frames. Your source has an odd number of frames (923) so the very last output frame cannot be interlaced properly. 923 source frames would become 461.5 output frames. You can't have half a frame so AviSynth uses two fields from the last input frame to build the last output frame, resulting in 462 frames. That is 1/50 second longer than your source. It sounds like VirtualDub just discards the orphaned field at the end, giving a results that's 1/50 second shorter than the source.

[marcorocchini]

ahh

ok.

But using "AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()" in the script, the output seems alwais correct, even if the processing appear to slow.

Really there can be no solution to get interlacing outside the avisynth interlacing?

[jagabo]

Really there can be no solution to get interlacing outside the avisynth interlacing?

No program can interlace an odd number of input frames and produce the same exact running time as the source.

By the way, the major source of slowness in your script is the creation of the horizontal edge mask, emask.

[Cornucopia]

Yes, if X = # of 50p progressive frames, and is ODD #, and if Y = X / 2, then Y (interlaced frames) is never a Whole Number, requiring either loss of 1/2 frame, addition of dupe 1/2 frame, replacement/blend of 1 or 2 half frames, or error.

Scott

[marcorocchini]

ahh ok

however my question now is:

is there an alternative to the

emask = mt_convolution(vertical="-1 2 -1").ColorYUV(off_y=-64).ColorYUV(gain_y=128).mt_expand(chroma="-128").Blur(0.8)
Overlay(last, Blur(0.5, 0.6), mask=emask)

to speedup the encoding?

I know the the the emask works fine but in certain cases requires too much time, so is there a different way even if not as so good?

[jagabo]

Your source is oversharpened on both the vertical and horizontal. So you might as well blur all sharp edges. This runs much faster, about 5 seconds (no encoding) for the entire clip, on my computer:

FFVideoSource("concatHD.ts", fpsnum=50, fpsden=1)

emask = mt_edge(mode="hprewitt", thy1=200, thy2=200).mt_expand().Blur(1.0)
Overlay(last, Blur(0.7), mask=emask)

ConvertToYUY2(interlaced=false)

AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()

Note the absence of prefilter().

And if you don't mind blurring everything this takes about 2 seconds:

FFVideoSource("concatHD.ts", fpsnum=50, fpsden=1)

Blur(0.7)

ConvertToYUY2(interlaced=false)

AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()

[jagabo]

mt_edge() can be used to detect horizontal edges only and is much faster than mt_convolution(). So if you want something like my original horizontal edge detection use this:

FFVideoSource("concatHD.ts", fpsnum=50, fpsden=1)#.assumeFPS(50)#.converttoYV12(interlaced=false)

emask = mt_edge("1 2 1 0 0 0 -1 -2 -1 1", thy1=250, thy2=250).mt_expand().Blur(1.0) # horizontal edges only
Overlay(last, Blur(0.7), mask=emask)

ConvertToYUY2(interlaced=false)

AssumeTFF().SeparateFields().SelectEvery(4, 0, 3).Weave()