TDeintMod

[HolyWu]

https://github.com/HomeOfVapourSynth...ynth-TDeintMod

It's a combination of TDeint + TMM port.

[HolyWu]

Update r3.

Add SIMD to motion mask related processing, speed up about 30~40%.
Add 'show' parameter to display the motion mask instead of the deinterlaced frame.

[HolyWu]

Update r5.

Add 9-16 bits support.
IsCombed's combed frame detection now works properly for different color spaces.

EDIT: I just ninja update the r5 archive to fix no frame returned at the end of processing bug, which only happen when mtql, mthl, mtqc, mthc are all set to -2 and show=True.

[jackoneill]

Is it wise to read from getReadPtr()-1? It happens in the SIMD version of the threshMask function. I see that the value doesn't make its way into the output (I think?), but are you sure it will never crash?

[Myrsloik]

It will probably overread the end if the width is mod32 too.

[HolyWu]

Thanks. I didn't think about that carefully when I vectorized the function. Should have been fixed now.

[HolyWu]

Update r6.

• TDeintMod now will use the field order specified in the source frames (_FieldBased frame property) and will only fall back to the specified order if not present.
• The framerate and the frame duration are properly adjusted and normalized for double rate output.
• Return an error if the returned clip is longer than INT_MAX.
• Fix potential crash in threshMask's SIMD code.

[HolyWu]

Update r7.

• Actually return an error when the value of the order, field, mode or metric parameter is out of range.
• The output frame's _FieldBased property is now set to 0.
• IsCombed: Actually return an error for unsupported chroma subsampling.
• The parameter value is now scaled in a saner way.
• Remove the upper bound of the length parameter.

[kolak]

Is this expected to compile on OSX?

[HolyWu]

Update r8.

• Code refactoring and minor performance improvement.
• Add AVX2 code path.
• Add opt and planes parameters.

[HolyWu]

Update r9.

• Change show parameter to display binary comb mask.
• Add athresh, metric, expand and link parameters.

Someone requested for TCombMask a while ago, but we really don't need another filter for duplicate job.

[Sparktank]

Hmm... Every day, I think about learning VS.

So many more updates.
How is this compared to the AVS/AVS+ branches?

The only thing holding me back is not enough examples for VS usage.
I get it's python. But I don't even know python. I might as well learn Wookie.

But, I love seeing any update. Wether it be AVS+ or VS.
All updates just mean a step forward.

[jackoneill]

Originally Posted by Sparktank

The only thing holding me back is not enough examples for VS usage.
I get it's python. But I don't even know python. I might as well learn Wookie.

Here is an example:

import vapoursynth as vs

core = vs.get_core()

clip = core.std.BlankClip(format=vs.YUV420P8)
clip = core.text.CoreInfo(clip)

clip.set_output()

vspipe blankclip.py - --y4m | x264.exe --demuxer y4m --crf 16 -o blankclip.mkv -

[DJATOM]

@Sparktank
VS is pretty good. I tried it this fall and now I'm using VS for almost all my releases. IVTC (tfm) stuff is done using AVS+, anything else with VS.

[kgrabs]

Originally Posted by DJATOM

@Sparktank
VS is pretty good. I tried it this fall and now I'm using VS for almost all my releases. IVTC (tfm) stuff is done using AVS+, anything else with VS.

If/when you try switching to VIVTC, it might benefit you to know it doesn't remove the field-based prop, so if you try to crop to mod 2 you have to remove it first: https://pastebin.com/raw/eujeQxvb

[jackoneill]

Originally Posted by kgrabs

If/when you try switching to VIVTC, it might benefit you to know it doesn't remove the field-based prop, so if you try to crop to mod 2 you have to remove it first: https://pastebin.com/raw/eujeQxvb

An easier way to remove a frame property:

src = c.std.SetFrameProp(src, "_FieldBased", delete=True)

http://www.vapoursynth.com/doc/funct...frameprop.html

[HolyWu]

Update r10.

• Fix link parameter not working in 9-16 bit depth.

[DJATOM]

Originally Posted by kgrabs

If/when you try switching to VIVTC, it might benefit you to know it doesn't remove the field-based prop, so if you try to crop to mod 2 you have to remove it first: https://pastebin.com/raw/eujeQxvb

VIVTC can't perform 2 pass VFR, that's why I stick with tfm/tdecimate for that task.

[l33tmeatwad]

Both r9 and r10 fail to compile on OSX.

TDeintMod/TDeintMod_SSE2.cpp:270:43: error: use of overloaded operator '==' is
      ambiguous (with operand types 'Vec16uc' and '__m128i' (vector of 2
      'long long' values))
            select(T2().load_a(srcp0 + x) == zero_128b() && T2().load_a(...
                   ~~~~~~~~~~~~~~~~~~~~~~ ^  ~~~~~~~~~~~
TDeintMod/TDeintMod_SSE2.cpp:281:15: note: in instantiation of function template
      specialization 'combineMasks_sse2<unsigned char, Vec16uc, 16>' requested
      here
template void combineMasks_sse2<uint8_t, Vec16uc, 16>(const VSFrameRef *...
              ^
TDeintMod/vectorclass/vectori128.h:679:23: note: candidate function
static inline Vec16cb operator == (Vec16c const & a, Vec16c const & b) {
                      ^
TDeintMod/TDeintMod_SSE2.cpp:270:43: note: built-in candidate
      operator==(__attribute__((__vector_size__(2 * sizeof(long long)))) long
      long, __attribute__((__vector_size__(2 * sizeof(long long)))) long long)
            select(T2().load_a(srcp0 + x) == zero_128b() && T2().load_a(...
                                          ^
TDeintMod/TDeintMod_SSE2.cpp:270:43: error: use of overloaded operator '==' is
      ambiguous (with operand types 'Vec8us' and '__m128i' (vector of 2
      'long long' values))
            select(T2().load_a(srcp0 + x) == zero_128b() && T2().load_a(...
                   ~~~~~~~~~~~~~~~~~~~~~~ ^  ~~~~~~~~~~~
TDeintMod/TDeintMod_SSE2.cpp:282:15: note: in instantiation of function template
      specialization 'combineMasks_sse2<unsigned short, Vec8us, 8>' requested
      here
template void combineMasks_sse2<uint16_t, Vec8us, 8>(const VSFrameRef *...
              ^
TDeintMod/vectorclass/vectori128.h:1504:22: note: candidate function
static inline Vec8sb operator == (Vec8s const & a, Vec8s const & b) {
                     ^
TDeintMod/TDeintMod_SSE2.cpp:270:43: note: built-in candidate
      operator==(__attribute__((__vector_size__(2 * sizeof(long long)))) long
      long, __attribute__((__vector_size__(2 * sizeof(long long)))) long long)
            select(T2().load_a(srcp0 + x) == zero_128b() && T2().load_a(...
                                          ^
2 errors generated.
make: *** [TDeintMod/TDeintMod_SSE2.lo] Error 1

[HolyWu]

Originally Posted by l33tmeatwad

Both r9 and r10 fail to compile on OSX.

Fixed in the latest commit.