Vapoursynth

[Jukus]

Is there any way to use spline144 with Vapour?
Is it true that this is the best resizer for any situation?

[poisondeathray]

Is there any way to use spline144 with Vapour?


With fmtconv kernel="spline", taps=6


"spline"

Generic splines, number of sample points is twice the taps parameter, so you can use taps = 6 to get a Spline144Resize equivalent.





Is it true that this is the best resizer for any situation?

Probably not. Likely some ringing artifacts, oversharpen halos

[Myrsloik]

R48-RC3 (https://github.com/vapoursynth/vapoursynth/releases/tag/R48-RC3)

Fixes the aformentioned pink line at top and bottom (general expr bug) and another 32bit bug in expr as well. Keep testing it.

Keep testing it! So far the only discovered bug in RC3 is in Convolution with float input, everything else should work fine. I'll probably do daily builds from now on until every single regression is gone.

[poisondeathray]

If you have RGBS and use Expr to add 1 to channel R , is it suppose to round the value ? Or is this the same thing as the Convolution with float input issue ?

clip2 = core.std.Expr(clip, ["x 1 +", "", ""])

e.g. in vsedit, the color picker shows R=0.921468 for clip, R=1.92147 for clip2

I would have expected 1.921468 (G, B remain unchanged at 0.921468)

core.std.SetMaxCPU("none") does not affect the result, R48-RC3 x64

[Myrsloik]

If you have RGBS and use Expr to add 1 to channel R , is it suppose to round the value ? Or is this the same thing as the Convolution with float input issue ?

clip2 = core.std.Expr(clip, ["x 1 +", "", ""])

e.g. in vsedit, the color picker shows R=0.921468 for clip, R=1.92147 for clip2

I would have expected 1.921468 (G, B remain unchanged at 0.921468)

core.std.SetMaxCPU("none") does not affect the result, R48-RC3 x64

That's how floating point math works. It's not exact.

[StainlessS]

From VS2008 Float.h


#define FLT_DIG 6 /* # of decimal digits of precision */
#define FLT_EPSILON 1.192092896e-07F /* smallest such that 1.0+FLT_EPSILON != 1.0 */
#define FLT_GUARD 0
#define FLT_MANT_DIG 24 /* # of bits in mantissa */
#define FLT_MAX 3.402823466e+38F /* max value */
#define FLT_MAX_10_EXP 38 /* max decimal exponent */
#define FLT_MAX_EXP 128 /* max binary exponent */
#define FLT_MIN 1.175494351e-38F /* min positive value */
#define FLT_MIN_10_EXP (-37) /* min decimal exponent */
#define FLT_MIN_EXP (-125) /* min binary exponent */
#define FLT_NORMALIZE 0
#define FLT_RADIX 2 /* exponent radix */
#define FLT_ROUNDS 1 /* addition rounding: near */

[poisondeathray]

Thanks Myrsloik, StainlessS ; If it's using FLT_DIG 6 , shouldn't it round to 1.921468 ? Or does it count decimal digits differently ? Or was this covered in grade school math that I might have slept through? :D

[StainlessS]

Its really the binary digits that count, and 10 is not a power of 2, so it works (rounds) a bit weird.

EDIT: If it were viewed in binary, it would make sense.

[poisondeathray]

Its really the binary digits that count, and 10 is not a power of 2, so it works (rounds) a bit weird.

Thanks

[StainlessS]

PDR, a bit more.

There are a fractional number of binary digits per decimal digit, and although float only precise to 6 [EDIT: significant] digits, there will likely be one or more decimal digits which are not quite correct.

EDIT: Note, below 6 significant digits

0.921468
^^^^^^

1.921468 # Adding 1.0, last digit is not within first 6 significant digits : Note that the extra digit is printed because of the default print format, rather than it being within first 6 significant digits.
^ ^^^^^



a = 0.921468 # This was probably not exact to begin with

b = a + 1.0

DIGS=32 # Digits to the right of decimal point

BlankClip

RT_DebugF("a=%.*f\nb=%.*f",DIGS,a,DIGS,b)
RT_subtitle("a=%.*f\nb=%.*f",DIGS,a,DIGS,b)

return last



00000057 0.29118466 [1068] RT_DebugF: a=0.92146801948547363000000000000000
00000058 0.29125640 [1068] RT_DebugF: b=1.92146801948547360000000000000000


EDIT:

RT_subtitle("a=%.*f\nb=%.*f",DIGS,a,DIGS,b)

Equiv

RT_subtitle("a=%.32f\nb=%.32f",a,b)


EDIT: No reply necessary.

EDIT:
There are a fractional number of binary digits per decimal digit

Function Log2(float n) {
return Log(n) / Log(2.0)
}

blankclip
DIGS=32 # Digits to the right of decimal point

BitsPerDecDigit = Log2(10.0)
Test10=Pow(2.0,BitsPerDecDigit)
RT_debugF ("BitsPerDecDigit=%.*f\nTest10 =%.*f",DIGS,BitsPerDecDigit,DIGS,Test10)
RT_Subtitle("BitsPerDecDigit=%.*f\nTest10 =%.*f",DIGS,BitsPerDecDigit,DIGS,Test10)
return last



00000231 0.29308608 [4064] RT_DebugF: BitsPerDecDigit=3.32192802429199220000000000000000
00000232 0.29313752 [4064] RT_DebugF: Test10 =9.99999904632568360000000000000000

[ifb]

Speed comparisons with R46 also welcome.

Threadripper 1920X w/64GB ECC and Windows 10 x64
Local build of VS with the latest AVX2 fix (ac74e9b)
Three runs on a script that's all float and 16-bit (CinemaDNG sequence):
time vspipe -e 1000 script.vpy .

1 2 3 avg stddev relative
R48-RC3+1 avx2 432.83 434.95 435.22 434.33 1.309 1.0080x
sse2 431.05 433.80 436.28 433.71 2.616 1.0095x
none 467.84 465.92 463.87 465.88 1.985 0.9398x
R46 sse2 434.29 440.49 438.68 437.82 3.188 1.0000x

~1% faster vs R46 even though AVX2 doesn't help poor Zen1. :p

[Myrsloik]

Threadripper 1920X w/64GB ECC and Windows 10 x64
Local build of VS with the latest AVX2 fix (ac74e9b)
Three runs on a script that's all float and 16-bit (CinemaDNG sequence):
time vspipe -e 1000 script.vpy .

1 2 3 avg stddev relative
R48-RC3+1 avx2 432.83 434.95 435.22 434.33 1.309 1.0080x
sse2 431.05 433.80 436.28 433.71 2.616 1.0095x
none 467.84 465.92 463.87 465.88 1.985 0.9398x
R46 sse2 434.29 440.49 438.68 437.82 3.188 1.0000x

~1% faster vs R46 even though AVX2 doesn't help poor Zen1. :p

You've most likely hit the ram bandwidth limit. Run it with a single/few threads and the difference should be obvious.

[Myrsloik]

R48-RC4 (https://github.com/vapoursynth/vapoursynth/releases/tag/R48-RC4)

Fixes the floating point edge handling of RC3.

Test harder!

[Myrsloik]

It seems that the PATH is always added even though I have unchecked the option.

Congratulations! You've won a free RC5 build which you can redeem tomorrow! That was a really stupid typo. At least there are no known image corruption problems. (test the unprivileged installs too just to be sure)

[Myrsloik]

R48-RC5 (https://github.com/vapoursynth/vapoursynth/releases/tag/R48-RC5)

The only change from RC4 is that the installer now correctly respects the add to PATH options.

[l00t]

I've found another bug. When I try to use nyuszika7h's FFInfo the result is a pile of garbage with R48-RC4, while it was perfectly fine with R47.2. The picture is also okay without the script.

nyuszika7h's FFInfo script:
https://gist.github.com/nyuszika7h/340cd5bd529a70746f0b35a464ef9a91

with FFInfo:
https://thumbs2.imgbox.com/c5/75/lfIpTn7P_t.png (http://imgbox.com/lfIpTn7P)

without FFInfo:
https://thumbs2.imgbox.com/81/7d/ctRXj1Pu_t.png (http://imgbox.com/ctRXj1Pu)

import vapoursynth as vs
from vapoursynth import core
clip = core.ffms2.Source(source=r'some_hd_footage.mkv')
clip = vs_ffinfo.FFInfo(clip, text='some text', frame_num=True, frame_type=True, frame_time=True)
clip.set_output()

FFMS2 is the latest version from 10/07/2019 from StvG

Video parameters:
Resolution: 1920x1038
Format: YUV420P8
FPS: 24000/1001

Thanks in advance

[ChaosKing]

The problem is in: clip.sub.Subtitle("this is art")

[Myrsloik]

R48-RC6 (https://github.com/vapoursynth/vapoursynth/releases/tag/R48-RC6)

Fixes premultiplied maskedmerge and a few expr problems that most likely were never encountered by anyone.

[l00t]

Thanks, now it's perfect.

BTW: If someone's interested in FFInfo, I've pumped up a bit nyuszika's version: https://gist.github.com/l00tzOMG/404e81cdcae2e7415dd96b57b08f32ce

[_Al_]

I've pumped up a bit nyuszika's version: https://gist.github.com/l00tzOMG/404...d96b57b08f32ce
To get string equivalents for those numerical frame property values, you can use dictionaries/ tables and then just get value you need.
With if , elif approach you can end up having hundreds of them is database is large. Look in that example:
https://forum.doom9.org/showthread.php?p=1885759#post1885759

[ifb]

You've most likely hit the ram bandwidth limit. Run it with a single/few threads and the difference should be obvious.

core.num_threads = 1
R48 RC4 avx2 5975.40 1.0062
sse2 5938.56 1.0125
R46 sse2 6012.63 1.0000

I only did 1 run each, but still ~1% faster.

[_Al_]

@l00t
MATRIX = {
0:'rgb',
1:'709',
.
.
.
}
PRIMARIES = { .....}
TRANSFER = {....}

try:
lines.append(f'Matrix: {MATRIX[f.props["_Matrix"]]}')
except Exception as e:
lines.append(f'Matrix: {str(e)}')

try:
lines.append(f'Primaries: {PRIMARIES[f.props["_Primaries"]]}')
except Exception as e:
lines.append(f'Primaries: {str(e)}')

[l00t]

Thanks for the idea, modified the script accordingly :)

...

[MonoS]

Since version R46 i get the error "Failed to initialize VapourSynth environment" even execuiting a simple vspipe -v
I've tried R46, R47 and R48-RC6, all giving me the same error, this on two different machine, one Win 10 the other Win Server 2019 even after reboot

Installing R45 it works without problem.

What could be the issue? what can i do to help you diagnose the problem?

[Myrsloik]

Since version R46 i get the error "Failed to initialize VapourSynth environment" even execuiting a simple vspipe -v
I've tried R46, R47 and R48-RC6, all giving me the same error, this on two different machine, one Win 10 the other Win Server 2019 even after reboot

Installing R45 it works without problem.

What could be the issue? what can i do to help you diagnose the problem?

Which install type? Did you really install the vs2019 runtimes? You can simply run vspipe in a debugger if you still can't figure it out since I've even included pdbs for the main dlls.

[MonoS]

Which install type? Did you really install the vs2019 runtimes? You can simply run vspipe in a debugger if you still can't figure it out since I've even included pdbs for the main dlls.

64bit non portable version, installed vs2019 runtimes the first time, then in subsequent install skipped that.

For setting up the debugger what should i do? Download the source and put a breakpoint inside real_init (https://github.com/vapoursynth/vapoursynth/blob/7c488b5d33991115e148da60b7afe30040da9245/src/vsscript/vsscript.cpp#L46) and see what fail?

[Myrsloik]

64bit non portable version, installed vs2019 runtimes the first time, then in subsequent install skipped that.

For setting up the debugger what should i do? Download the source and put a breakpoint inside real_init (https://github.com/vapoursynth/vapoursynth/blob/7c488b5d33991115e148da60b7afe30040da9245/src/vsscript/vsscript.cpp#L46) and see what fail?

Yes, that should work and give you the best information.

[MonoS]

Yes, that should work and give you the best information.

Ok, i'll try to take a look, but it will probably take me about a week as i'm currently doing an encode.

[Myrsloik]

R48-RC6 (https://github.com/vapoursynth/vapoursynth/releases/tag/R48-RC6)

Fixes premultiplied maskedmerge and a few expr problems that most likely were never encountered by anyone.

4 days without any bugs found! Final release when bug free for a whole week. Keep testing!

[Patman]

Hi,

R48 final based on python 3.7.x module? Which version of Vapoursynth will be based on python 3.8?

Gesendet von meinem HMA-L09 mit Tapatalk

[Myrsloik]

Hi,

R48 final based on python 3.7.x module? Which version of Vapoursynth will be based on python 3.8?

Gesendet von meinem HMA-L09 mit Tapatalk

Yes, it will use python 3.7.x. I'll switch to 3.8.x when cython adds official support for it.

[Patman]

Yes, it will use python 3.7.x. I'll switch to 3.8.x when cython adds official support for it.Thanks for the info.

Gesendet von meinem HMA-L09 mit Tapatalk

[Myrsloik]

R48 is released!

[stax76]

Is there a particular reason why vsrepo downloaded files don't have original timestamps?

staxrip has packages with lost timestamps and I would like to recover them.

[Myrsloik]

Is there a particular reason why vsrepo downloaded files don't have original timestamps?

staxrip has packages with lost timestamps and I would like to recover them.

Because the way I'm handling things it's kinda annoying to set them. And it's not like you can ever trust them anyway...

[hydra3333]

Yes, it will use python 3.7.x. I'll switch to 3.8.x when cython adds official support for it.

just checking,

does that mean the v48 portable version
will run under portable python 3.7.5 ?
but not 3.8.0 ?

[l00t]

just checking,

does that mean the v48 portable version
will run under portable python 3.7.5 ?
but not 3.8.0 ?

exactly

[Jukus]

What needs to be done to make a video have different FPS?
For example, I want to do something like that

from vapoursynth import core
import havsfunc as haf

clip = core.d2v.Source('index.d2v')
clip1 = core.std.Trim(clip, 0, 5579)
clip1 = haf.QTGMC(clip1, Preset='Very Slow', Sharpness=0.5, FPSDivisor=1, SourceMatch=3, Lossless=2, MatchEnhance=0.75, TFF=True)
clip2 = core.std.Trim(clip, 5580, 24186)
clip2 = haf.QTGMC(clip2, Preset='Very Slow', Sharpness=0.5, SourceMatch=3, MatchEnhance=0.75, InputType=1)
clip = clip1+clip2

clip.set_output()

but this code will not work correctly.

[poisondeathray]

What needs to be done to make a video have different FPS?
For example, I want to do something like that

from vapoursynth import core
import havsfunc as haf

clip = core.d2v.Source('index.d2v')
clip1 = core.std.Trim(clip, 0, 5579)
clip1 = haf.QTGMC(clip1, Preset='Very Slow', Sharpness=0.5, FPSDivisor=1, SourceMatch=3, Lossless=2, MatchEnhance=0.75, TFF=True)
clip2 = core.std.Trim(clip, 5580, 24186)
clip2 = haf.QTGMC(clip2, Preset='Very Slow', Sharpness=0.5, SourceMatch=3, MatchEnhance=0.75, InputType=1)
clip = clip1+clip2

clip.set_output()

but this code will not work correctly.




Internally in vapoursynth, it has to be CFR (constant frame rate)

1) You can duplicate frames and framerate in the second section (but same content speed). eg. by using core.std.Interleave

Or ,

2) you can temporarily assign a 2xFPS to the 2nd section using core.std.AssumeFPS to append sections (so it plays double speed), then use external timecodes (timestamps) method to make it VFR

[DJATOM]

Internally in vapoursynth, it has to be CFR (constant frame rate)
No. You can splice mixed fps clips, or even mixed resolution clips, vapoursynth can handle that. But encoding app might fail to understand such clip, that depends on what you're using.
I made x264 input filter and successfully provided vapoursynth timecodes (but code is a bit dirty), you can pick it here and compile: https://pastebin.com/QhjQ26qG

[_Al_]

http://www.vapoursynth.com/doc/installation.html#installation-via-pip-pypi says that pip install should be done only after Vapoursynth is installed. So what is it for, or what is the purpose of that PIP installation?

[MonoS]

Yes, that should work and give you the best information.

I've reinstalled my whole system so i am now on a clean OS.
Installed Python 3.7.5 for all user (so it is in "C:\Program Files\Python37" ), installed Vapoursynth R48 and it gives me the same error.

I've then started debuging and the line that is failing is the PyImport_ImportModule in vapoursynth_api.h (https://github.com/vapoursynth/vapoursynth/blob/7c488b5d33991115e148da60b7afe30040da9245/include/cython/vapoursynth_api.h#L113), the module returned is NULL.
Folder "Lib\site-packages", where i would put my script, is empty, would expect to find file "vapoursynth.pth" and the folder "vapoursynth" (checked in another system with R45), probably is that that is making VSPipe fails?

[stax76]

Because the way I'm handling things it's kinda annoying to set them. And it's not like you can ever trust them anyway...

If both timestamps and module version info is missing (it's missing very often in my experience), what else do we have? The file size will be the only thing left and that isn't very much.

[ChaosKing]

We have the hash...
https://github.com/vapoursynth/vsrepo/blob/master/local/fft3dfilter.json

[MonoS]

I've reinstalled my whole system so i am now on a clean OS.
Installed Python 3.7.5 for all user (so it is in "C:\Program Files\Python37" ), installed Vapoursynth R48 and it gives me the same error.

I've then started debuging and the line that is failing is the PyImport_ImportModule in vapoursynth_api.h (https://github.com/vapoursynth/vapoursynth/blob/7c488b5d33991115e148da60b7afe30040da9245/include/cython/vapoursynth_api.h#L113), the module returned is NULL.
Folder "Lib\site-packages", where i would put my script, is empty, would expect to find file "vapoursynth.pth" and the folder "vapoursynth" (checked in another system with R45), probably is that that is making VSPipe fails?

I think i fixed the issue downloading the portable version of Vapoursynth and putting the file vapoursynth.cp37-win_amd64.pyd inside the site-packages folder under Python37. Hope this could help you fix the problem.

[Lypheo]

core.resize.Bicubic(core.std.BlankClip(format=vs.YUV444PS), format=vs.RGB24, matrix_in=1, primaries_in=1, primaries=1) #errors out (Resize error 3074: invalid colorspace definition (1/2/1 => 0/2/1). May need to specify additional colorspace parameters.)
core.resize.Bicubic(core.std.BlankClip(format=vs.YUV444PS), format=vs.RGB24, matrix_in=1, primaries_in=1, primaries=1, transfer_in=1, transfer=1) #works
core.resize.Bicubic(core.std.BlankClip(format=vs.YUV444PS), format=vs.RGB24, matrix_in=1) #works

Seems like a bug to me. vsresize/zimg seems to think it needs to do a primary conversion when in and out primaries are specified (even when equal), and thus demands that the transfer function be specified. This causes the error mentioned above when previewing clips with vsedit (which calls vsresize internally, I presume) that ffms2 attached _Primaries != 2 and _Transfer = 2 to (this is the case for files which have a primaries flag but no transfer flag).

[Richard1485]

http://www.vapoursynth.com/doc/installation.html#installation-via-pip-pypi says that pip install should be done only after Vapoursynth is installed. So what is it for, or what is the purpose of that PIP installation?

I was wondering the same thing.

[Myrsloik]

I was wondering the same thing.

It's useful if you want to install vapoursynth into multiple python environments. The normal installer only lets you choose one.

[Myrsloik]

core.resize.Bicubic(core.std.BlankClip(format=vs.YUV444PS), format=vs.RGB24, matrix_in=1, primaries_in=1, primaries=1) #errors out (Resize error 3074: invalid colorspace definition (1/2/1 => 0/2/1). May need to specify additional colorspace parameters.)
core.resize.Bicubic(core.std.BlankClip(format=vs.YUV444PS), format=vs.RGB24, matrix_in=1, primaries_in=1, primaries=1, transfer_in=1, transfer=1) #works
core.resize.Bicubic(core.std.BlankClip(format=vs.YUV444PS), format=vs.RGB24, matrix_in=1) #works

Seems like a bug to me. vsresize/zimg seems to think it needs to do a primary conversion when in and out primaries are specified (even when equal), and thus demands that the transfer function be specified. This causes the error mentioned above when previewing clips with vsedit (which calls vsresize internally, I presume) that ffms2 attached _Primaries != 2 and _Transfer = 2 to (this is the case for files which have a primaries flag but no transfer flag).

You always need to specify both primaries and transfer. You can't do just one. It's how it works.

[Lypheo]

You always need to specify both primaries and transfer. You can't do just one. It's how it works.
Yes, I know, but this is about cases where the output primaries are equal to the input primaries (so no conversion is performed). Requiring the transfer parameters to be specified as well in that situation doesn’t make a lot of sense because they’re not needed at all.
Again, this isn’t just a hypothetical case of passing these particular arguments to resize manually: When you have a YCbCR clip that has _Primaries other than 2 but _Transfer == 2 (=undefined) and try to convert to RGB (or to preview with VSEdit), vsresize will throw this error: Resize error 3074: invalid colorspace definition (1/2/1 => 0/2/1). May need to specify additional colorspace parameters.
This behaviour is likely very confusing to the unsuspecting user (as it was to me), because in theory, matrix is all the colorspace parameters that should be needed for a YUV->RGB conversion.