BM3D (VapourSynth) versus MVTools

[DJATOM]

CUDA based BM3D (V-BM3D) is fast enough on decent GPU and probably faster than mvtools2 with small temporal radius. You might want to consider using it as faster denoising backend.
Some numbers if you interested:
1080p, radius=3, fast=True - ~21.85 fps
1080p, radius=3, fast=False - ~14.17 fps
1080p, radius=2, fast=True - ~38.93 fps
1080p, radius=2, fast=False - ~19.69 fps
1080p, radius=1, fast=True - ~63.21 fps
1080p, radius=1, fast=False - ~31.49 fps
1080p, radius=0, fast=True - ~159.75 fps (that's a spatial mode)
1080p, radius=0, fast=False - ~79.24 fps (that's also a spatial mode)
"fast" option squeezes more fps with the same quality at expense of flooding PCIe lanes, probably my card can preform better, but due to SSD attached to CPU lanes it works at x8 mode (still enough for me).

[ChaosKing]

ahh finally some numbers for bm3dcuda. Could you also post some numbers for the bm3dCPU version? I wonder how big the speed gap is.

[DJATOM]

radius=0 - ~114.46 fps
radius=1 - ~65.87 fps
radius=2 - ~46.11 fps
radius=3 - ~36.03 fps
radius=4 - ~29.26 fps
I'm measuring with 12 cores CPU and overclocked RAM (yet no turbo boost enabled).

[ChaosKing]

Didn't expect that. It's even faster than your 2070. I guess a ~8 cores cpu could be about the same speed as the 2070 then.

[tormento]

CUDA based BM3D (V-BM3D) is fast enough on decent GPU and probably faster than mvtools2 with small temporal radius. You might want to consider using it as faster denoising backend.
Thanks for your suggestion, I did not know that BM3D had a CUDA version. Where can I find a Windows build, better with HBD support?

However, there are a few CUDA denoiser around already. What we need is a MVTool on CUDA or OpenCL.

[ChaosKing]

Thanks for your suggestion, I did not know that BM3D had a CUDA version. Where can I find a Windows build, better with HBD support?

However, there are a few CUDA denoiser around already. What we need is a MVTool on CUDA or OpenCL.

https://github.com/WolframRhodium/VapourSynth-BM3DCUDA

Scroll to version 2.1 for cuda version.

[tormento]

Scroll to version 2.1 for cuda version.
Thanks but it is for VS not AVS+.

[kedautinh12]

In the present, BM3D only in VS

[WolframRhodium]

The numbers from DJATOM looks interesting.

The GPU performance aligns well with the estimation that

fps * (2 * radius + 1) ≈ constant

, since the calculation is dominated by block matching.

The CPU performance is unexpected to me. Maybe the cache plays a nontrivial role.

Anyway, I have no experience in writing an AVS+ plugin. Will take a look if people are interested.

[kedautinh12]

Yeah, i'm waiting many years for avs+ ver

[real.finder]

Anyway, I have no experience in writing an AVS+ plugin. Will take a look if people are interested.

you can make the plugin work in both https://github.com/HomeOfAviSynthPlusEvolution/neo_DFTTest (this is backport of vs DFTTest and now it's work on both)

[WolframRhodium]

you can make the plugin work in both https://github.com/HomeOfAviSynthPlusEvolution/neo_DFTTest (this is backport of vs DFTTest and now it's work on both)

Thanks, I am aware of the DualSynth implementation. I need to read more about the parallelism mechanism of AVS+, thought.

[Reel.Deel]

Anyway, I have no experience in writing an AVS+ plugin. Will take a look if people are interested.

+1 ... definitely some interest here.

[Dogway]

Obviously super interested in this. I think it can easily double current scripts speed that make use of mvtools. Probably for implementation reference you can check source of AVS+ 3.7.1 CUDA.

[real.finder]

WolframRhodium, any news about avs+ port?

@Moderators, can you move BM3D posts to new thread?

[tebasuna51]

@Moderators, can you move BM3D posts to new thread?

Moved. It is ok now?

[real.finder]

Moved. It is ok now?

Yes. It's fine. Thanks

[WolframRhodium]

I'm sorry about being lazy on avs+ port... Will consider it this weekend.

[Dogway]

I think OP was replying to my post:
Currently mvtools is the bottleneck of most scripts, I calculated that for a 2h feature we need a performance of at least 5fps to finish an encode in 10h, good value for an overnight encode. I think this is still a bit far when using refinemotion (a must in my book) -1080p @ 16-bit-.

Thank you WolframRhodium, it would actually be an achievement and a revolution for AVS since mvtools is the major bottleneck in most scripts.

[lansing]

I'm confused, so the 3900x is faster than the RTX 2070??

[DJATOM]

Usually faster for me for certain job, but if I need to do all at once, gpu offloading helps.

[tormento]

Moved. It is ok now?
I think this post is not belong to this thread.

[tormento]

Usually faster for me for certain job, but if I need to do all at once, gpu offloading helps.
GPU offloading means more CPU resources to assign to encoder. So it helps a lot.

[WolframRhodium]

Please check the following test binaries:
nightly releases (https://github.com/WolframRhodium/VapourSynth-BM3DCUDA/issues/7#issuecomment-886173346)

The functions are named "BM3D_CUDA" and "BM3D_VAggregate". They only support planar float inputs (please do the conversion yourself). The usage is the same as the VapourSynth version.

Please feel free to report any issues. I rarely use AVS+ and I may make mistakes everywhere.

[DJATOM]

Script
DGSource(...)
ConvertBits(bits=32)
BM3D_CUDA(sigma=0.5, radius=2)
BM3D_VAggregate(radius=2)
ConvertBits(bits=16)
Prefetch(12) #more than 12 had worse performance on my machine

Results
radius=0: 165.55 fps
radius=1: 54.57 fps
radius=2: 36.33 fps

[kedautinh12]

Thanks very much, can you port cpu ver of BM3D???

[Dogway]

Thanks a lot, runs at 20fps with DJATOM's script on my 1070!

[Reel.Deel]

Very nice WolframRhodium, thank you very much.

Here are my results with the same script as DJATOM on a GTX 1050 Ti:
[Clip info]
Number of frames: 10001
Length (hh:mm:ss.ms): 00:06:57.125
Frame width: 1920
Frame height: 1080
Framerate: 23.976 (24000/1001)
Colorspace: YUV420P16


[Runtime info]
Frames processed: 10001 (0 - 10000)
FPS (min | max | average): 2.132 | 237298 | 27.10
Process memory usage (max): 3214 MiB
Thread count: 36
CPU usage (average): 54.9%

Time (elapsed): 00:06:09.091

[WolframRhodium]

Thanks very much, can you port cpu ver of BM3D???

Maybe next week.

[tormento]

Thanks a lot, runs at 20fps with DJATOM's script on my 1070!
Do you think it worths to be used as prefilter in SMDegrain or it's more than complete in its own? Could it be used instead of MVTools in SMDegrain too?

[tormento]

Results on 1060 3GB

Number of frames: 2904
Length (hh:mm:ss.ms): 00:02:01.121
Frame width: 1920
Frame height: 800
Framerate: 23.976 (24000/1001)
Colorspace: YUV420P16

Frames processed: 2904 (0 - 2903)
FPS (min | max | average): 2.520 | 153846 | 18.69
Process memory usage (max): 1491 MiB
Thread count: 23
CPU usage (average): 27.2%

GPU usage (average): 63%
VPU usage (average): 8%
GPU memory usage: 991 MiB
GPU Power Consumption (average): 70.1 W

Time (elapsed): 00:02:35.347

[Dogway]

I didn't give it a deep look but it seemed to me to soften the clip a bit. I would look into vapoursynth superscripts which mix mvtools and BM3D.

Currently I'm on the task of porting Adaptive Sharpen to avisynth (half done) to fully replace Contrasharpen, that's taking me the whole time.

[Quadratic]

I didn't give it a deep look but it seemed to me to soften the clip a bit. I would look into vapoursynth superscripts which mix mvtools and BM3D.
BM3D can use mvtools as a reference clip which greatly helps when the goal is to reduce grain but not entirely flatten the clip. Generally functions such as SMDegrain are going to be visually more faithful to the original clip, whereas BM3D is unmatched when it comes to complete removal of noise (which makes it an excellent prefilter, among other things).

[tormento]

BM3D is unmatched when it comes to complete removal of noise (which makes it an excellent prefilter, among other things).
Do you have any suggested values for parameters while using as prefilter?

[Quadratic]

Do you have any suggested values for parameters while using as prefilter?

Not specific to prefiltering, but watch out for blocking https://github.com/WolframRhodium/VapourSynth-BM3DCUDA/issues/5

[WolframRhodium]

The CPU (AVX2) version is uploaded.

[kedautinh12]

Wow, thanks, can you up for AVX same VS and share link download???

[WolframRhodium]

Wow, thanks, can you up for AVX same VS and share link download???

The link remains the same. AVX2 is required for the cpu version for efficient data movement in register, the same as in VS.

[tormento]

The CPU (AVX2) version is uploaded.
Does Test2 work for CUDA too or is it AVX2 bonded?

[kedautinh12]

CUDA for GPU ver, AVX2 for CPU ver

[WolframRhodium]

Does Test2 work for CUDA too or is it AVX2 bonded?

Both are included and are implemented in different dlls.

[real.finder]

ColorBars(width=640, height=480, pixel_type="yv12")
LoadPlugin("BM3DCPU_AVS.dll")
ConvertBits(bits=32)
BM3D_cpu(sigma=0.5, radius=2)
BM3D_VAggregate(radius=2)
ConvertBits(8)

22 fps with no mt (no Prefetch) on i7-8700

I have a question about BM3D_VAggregate, it's cpu only? and seems it's Duplicate in both cpu and cuda dlls and that seems not ok

[WolframRhodium]

ColorBars(width=640, height=480, pixel_type="yv12")
LoadPlugin("BM3DCPU_AVS.dll")
ConvertBits(bits=32)
BM3D_cpu(sigma=0.5, radius=2)
BM3D_VAggregate(radius=2)
ConvertBits(8)

22 fps with no mt (no Prefetch) on i7-8700

I have a question about BM3D_VAggregate, it's cpu only? and seems it's Duplicate in both cpu and cuda dlls and that seems not ok

It's cpu only since its performance is limited by memory bandwidth and has low arithmetic intensity. Test2 moved that function to the cpu dll so there should be no conflict.

[tormento]

Both are included and are implemented in different dlls.
Any changelog for test2 on CUDA?

[kedautinh12]

Read readme file in file downloaded

[WolframRhodium]

Any changelog for test2 on CUDA?

Nothing besides now it depends on the CPU version on VAggregate.

[Reel.Deel]

Nothing besides now it depends on the CPU version on VAggregate.

Any possibility of also supporting AVX? I see the VS version supports it, according to the documentation: The cpu version is implemented in AVX/AVX2 intrinsics...

[kedautinh12]

Any possibility of also supporting AVX? I see the VS version supports it, according to the documentation: The cpu version is implemented in AVX/AVX2 intrinsics...

But in Requirements, he wrote only AVX2

[WolframRhodium]

I mean AVX and AVX2 intrinsics. Sorry for the confusion.

[Reel.Deel]

I mean AVX and AVX2 intrinsics. Sorry for the confusion.

Ahh gotcha, I guess that explains this error in AVSMeter:

Exception: 0xC000001D [STATUS_ILLEGAL_INSTRUCTION]
Module: BM3DCPU_AVS.dll
Address: 0x000007FEC45204E6

Figured it was worth asking about AVX since I see just "AVX" in the requirements of the VS version, but in other parts it says AVX/AVX2.... well I guess that leaves me out of it till I build my next PC :D