single precision MVTools plugin (stable)

[feisty2]

@sl1pkn07
thanks to Are_'s help, the compiling files are available now

[sl1pkn07]

zankius!

[feisty2]

https://github.com/IFeelBloated/vapo...ses/tag/test13
test13
precision boost
1. SAD (float -> double)
2. SATD (int16_t -> double)
3. DCT (uint8_t -> even more precise than float)
and also features some cosmetic changes from @ jackoneill

[feisty2]

https://github.com/IFeelBloated/vapo...ses/tag/test14
test14
A. Full Precision Boost
1. DCT (float -> double)
2. Motion Analysis (float -> double)
3. Super (float -> double)
4. Overlap (float -> double)
5. Variance (float -> double)
6. Degrain (float -> double)
and more...
basically everything works at double precision, rounded to single precision only at the final output stage
B. Bug Fixes
fixed a bug inherited from the avisynth plugin (bit shift operation on negative values, reported by @Are_ via runtime debugging)

[Myrsloik]

Originally Posted by feisty2

https://github.com/IFeelBloated/vapo...ses/tag/test14
test14
A. Full Precision Boost
1. DCT (float -> double)
2. Motion Analysis (float -> double)
3. Super (float -> double)
4. Overlap (float -> double)
5. Variance (float -> double)
6. Degrain (float -> double)
and more...
basically everything works at double precision, rounded to single precision only at the final output stage
B. Bug Fixes
fixed a bug inherited from the avisynth plugin (bit shift operation on negative values, reported by @Are_ via runtime debugging)

I'm curious now. Did you compare the speed between your single and double precision versions? I suspect you've made it even slower now.

[feisty2]

Originally Posted by Myrsloik

I'm curious now. Did you compare the speed between your single and double precision versions? I suspect you've made it even slower now.

the speed is only a little bit slower than single precision version long as FFTW stays out of business, but drops heavily (~2x slower on my i7-4790k) if FFTW gets involved (dct=1-4)

[MonoS]

Probably you'll see some speed difference if you enable auto vectorization [on GCC -O3 and -march=corei7-avx]. Right now it isn't any slower because latency between float and double is the same [5 cycles on the latest intel generation], the little speed difference is probably due to memory access, fftw is 2x slower due to loss in vectorization [in a 128bit vector register can fit 4 floats but only 2 doubles].

I'll be more interested to know if this precision boost is really useful

[feisty2]

Originally Posted by MonoS

I'll be more interested to know if this precision boost is really useful

floats lose precision as the calculation goes on, and giant plugin like mvtools comes with hell lot of calculations, double precision is required to do the intermediate stuff if you want the final output features a real single precision, otherwise it's just low precision data in a single precision format..

[MonoS]

Originally Posted by feisty2

floats lose precision as the calculation goes on, and giant plugin like mvtools comes with hell lot of calculations, double precision is required to do the intermediate stuff if you want the final output features a real single precision, otherwise it's just low precision data in a single precision format..

Sorry if i ask, but i'd like to see some real world example.

While i understand the using of single precision for this kind of calculation, i think that double is rather overkill.
But let make clear that this my statement is not supported by actual facts, i'm willing to change my mind [but please, try to make some better example than the one you did on fmtc thread]

[feisty2]

https://github.com/IFeelBloated/vapo...ses/tag/test15
test15
A. Colorspace
all floating point colorspaces are supported now, GrayS, RGBS and YUV4xxPS (note that dct 1-4 on YUV clips might be kind of buggy, as chroma features a different range from luma, will be fixed in the next release)
B. Degrain
the stupid "thsadc" and "limitc" parameters got their asses canceled, "thsad" and "limit" have been made arrays
"plane" parameter won't do nothing on RGB and GRAY clips, all planes will be processed
C. stuff, here and there..
bug fixes shamelessly copied from @jackoneill

[feisty2]

so I been testing, RGB colorspace has been supported but,
I found motion estimating under RGB is not some kind of nice thing to do, I mean, it works, but the quality sucks comparing to just do it under YUV 4:4:4
convert your RGB input to YUV before applying mvtools on it, or, anyways, mvtools will happily take the RGB input and feed you back with some shitty motion estimation if you give it like some RGB input directly

[feisty2]

https://github.com/IFeelBloated/vapo...eleases/tag/r1
r1
first stable release!!!
A. sanity check.
will raise an error if the input is not single precision fp or features varying dimensions
B. DCT
fixed dct stuff on YUV input

[Mystery Keeper]

Works awesomely. Thank you!

[feisty2]

r2
merged bug fixes from jackoneill's branch since his last release
currently no plan to add depanning stuff

[feisty2]

r3
A. BlockFPS
1. added support to overlap (merged from Fizick's master branch)
2. new modes, mode 6-8, occlusion mask weighted on SAD (merged from Fizick's master branch)
B. Compensate
1. new parameter "time", use it to do partial time compensation (merged from Fizick's master branch)
C. Bug Fixes
1. vector length was clamped to 127/pel on motion flow functions, now it's 2147483647/pel, practically unlimited (Fizick relaxed it to 32767/pel, I decided to do it more thoroughly)
D. Precision Boost
1. internal masking for motion flow (uint8_t -> double)
2. simpleresize for masks (uint8_t -> double)
floating point precision MMask should be super easy to implement now (all internal stuff are double already), but I didn't do it anyways like, yeah, I'm all fucked up lazy

[Mystery Keeper]

You say it is possible to write QTGMC analog for floating point. Can you please give a code to replace TemporalSoften?

[feisty2]

r4
New Filter
binary: added mvsf.Flow
mvmulti: added mvmulti.Flow

[~SimpleX~]

VS Editor crashes on simple

import vapoursynth as vs
core = vs.get_core()

src = core.d2v.Source(...)
last = core.fmtc.bitdepth(src, bits=32, fulls=False, fulld=True)
super = core.mvsf.Super(last)
last.set_output()

Latest test version of VS r33, mvsf r4.

[feisty2]

Originally Posted by ~SimpleX~

VS Editor crashes on simple

import vapoursynth as vs
core = vs.get_core()

src = core.d2v.Source(...)
last = core.fmtc.bitdepth(src, bits=32, fulls=False, fulld=True)
super = core.mvsf.Super(last)
last.set_output()

Latest test version of VS r33, mvsf r4.

does your CPU feature the AVX2 extension?

[~SimpleX~]

Originally Posted by feisty2

does your CPU feature the AVX2 extension?

Oh, well, it doesn't. I have i7 2600k, which is Sandy Bridge.