KNLMeansCL: OpenCL NLMeans de-noising algorithm [2018-01-29]

[real.finder]

Originally Posted by pinterf

Hint: with

info=true

you can display the actual driver config.

[pinterf]

Same for me, except that I have 750. And the driver is a bit older, 32.0.101.6078.
Anyway, I narrowed down the problem, which is in one or more of the opencl program kernel functions which do the actual horizontal and vertical processing. Seems that next week(s) I'm going to entertain myself with the millions of trial and error modifications.
You have to wait a bit further, I will only appear when I'm ready.

[real.finder]

Originally Posted by pinterf

Same for me, except that I have 750. And the driver is a bit older, 32.0.101.6078.
Anyway, I narrowed down the problem, which is in one or more of the opencl program kernel functions which do the actual horizontal and vertical processing. Seems that next week(s) I'm going to entertain myself with the millions of trial and error modifications.
You have to wait a bit further, I will only appear when I'm ready.

no news since months, forgotten?

[pinterf]

Originally Posted by real.finder

no news since months, forgotten?

It's just not finalized. Most possibly a driver error, on the 4th crazy day and after the fifth possible workaround, I got it. I have almost finished and assembled a meaningful bug report (the smallest possible code which still presents the bug), then the holidays came then a huge pile of Avisynth and other plugin related stuffs. And I had to arrange and nicely commit the modifications I've done to KNL since 2020. So please wait a little bit more I hope not that much.

[tormento]

I find BM3D to be superior to KNLMeansCL almost in every aspect.

Is there any niche application where it could shine?

[pinterf]

Bug reported to Intel. At least I consider it a bug:

https://community.intel.com/t5/GPU-C.../1673206#M1764
EDIT: Intel (or Intel's bot) almost instantly categorized my bug report as spam, causing it to disappear from the flow of posts. I mentioned this to the moderator... We'll see what happens. It's not easy to communicate with them.

There may be a possible workaround: manually specifying parameters ocl_x and ocl_y to 16.
Report back if it works.

[tormento]

Originally Posted by pinterf

Bug reported to Intel. At least I consider it a bug

The link you posted brings to:

The message you are trying to access is not available.

[real.finder]

Originally Posted by pinterf

Bug reported to Intel. At least I consider it a bug:

https://community.intel.com/t5/GPU-C.../1673206#M1764
EDIT: Intel (or Intel's bot) almost instantly categorized my bug report as spam, causing it to disappear from the flow of posts. I mentioned this to the moderator... We'll see what happens. It's not easy to communicate with them.

There may be a possible workaround: manually specifying parameters ocl_x and ocl_y to 16.
Report back if it works.

KNLMeansCL(info=true, ocl_x=16, ocl_y = 16, ocl_r=16)
seems work, dont know if it act like KNLMeansCL(info=true) with other non bugged gpu drivers

[pinterf]

Originally Posted by real.finder

KNLMeansCL(info=true, ocl_x=16, ocl_y = 16, ocl_r=16)
seems work, dont know if it act like KNLMeansCL(info=true) with other non bugged gpu drivers

Probably, the 32x8 configuration was chosen for performance reasons on Intel GPUs (I'm not sure if this was based on tests or manufacturer recommendations). The 16x16 configuration should work as well. Anyway, a lot of time has passed since these values were established in the code, and the optimal setting may have changed since then.

[Katie Boundary]

Has anyone compiled a version of this that doesn't rely on external DLL bullshit? Visual C++ 2013 isn't installing on one of my computers for some stupid bullshit reason or another and I think that's screwing up a bunch of my filters.

UPDATE: I was able to find slightly older versions (12.0.30501) of Visual C++ 2013, and they installed correctly, but KNLmeansCL and two other filters are still whining about a missing DLL file.

Dynamic linking is one of the worst ideas in computing history.

[Jamaika]

I have a statically linked library using gcc/mingw.
https://github.com/FFmpeg/FFmpeg/com...220c4595f6a33c
https://github.com/AviSynth/AviSynth...900b0997749207
https://github.com/HomeOfAviSynthPlu...7c9c971cd626d9
https://github.com/Khanattila/KNLMea...ba0d1e76ed7850
https://github.com/KhronosGroup/Open...ebfa4d7568c0de
https://github.com/KhronosGroup/Open...7391b3e23efa81

LibavSource2("input.avi",fpsnum=30000,fpsden=1001)
KNLMeansCL(device_type="gpu",device_id=0,info=true)

https://www.sendspace.com/file/k2xj7v

Edit:
I will only add that all add-ons should be for ffms2 and not ffmpeg. If we want to use ffmpeg, the library should be modified, e.g. for l-smash. There aren't modifications for the latest avisynth.c 3.7.4.
https://github.com/HomeOfAviSynthPlu...?author=qyot27

[Katie Boundary]

Originally Posted by Jamaika

I have a statically linked library using gcc/mingw.
https://github.com/FFmpeg/FFmpeg/com...220c4595f6a33c
https://github.com/AviSynth/AviSynth...900b0997749207
https://github.com/HomeOfAviSynthPlu...7c9c971cd626d9
https://github.com/Khanattila/KNLMea...ba0d1e76ed7850
https://github.com/KhronosGroup/Open...ebfa4d7568c0de
https://github.com/KhronosGroup/Open...7391b3e23efa81

LibavSource2("input.avi",fpsnum=30000,fpsden=1001)
KNLMeansCL(device_type="gpu",device_id=0,info=true)

https://www.sendspace.com/file/k2xj7v

Okay... which of those 20 or 30 links is it?

EDIT: after a little bit of digging, it turns out that version 0.6.11 was compiled with Visual C++ 2013 instead of 2015, and therefore does not require UCRT. It was kept around for compatibility with older versions of OpenCL, but XP compatibility was an accidental side effect. I can use this for now.

[Khanattila]

Originally Posted by pinterf

Bug reported to Intel. At least I consider it a bug:

https://community.intel.com/t5/GPU-C.../1673206#M1764
EDIT: Intel (or Intel's bot) almost instantly categorized my bug report as spam, causing it to disappear from the flow of posts. I mentioned this to the moderator... We'll see what happens. It's not easy to communicate with them.

There may be a possible workaround: manually specifying parameters ocl_x and ocl_y to 16.
Report back if it works.

Great job with the bug report.

Originally Posted by pinterf

Probably, the 32x8 configuration was chosen for performance reasons on Intel GPUs (I'm not sure if this was based on tests or manufacturer recommendations). The 16x16 configuration should work as well. Anyway, a lot of time has passed since these values were established in the code, and the optimal setting may have changed since then.

Testing, and a general understanding of how the Intel integrated GPU architecture was designed (memory alignment, cache size, compute units per group). Those values obviously depend on the architecture and are not optimal forever.

[GeoffreyA]

Not related, but I recently discovered a slight performance loss with KNLMeansCL on the Arc B580's recent drivers versus one from January.

https://github.com/IGCIT/Intel-GPU-C...IT/issues/1134

[takla]

FFMPEG has merged a rewrite of nlmeans vulkan, which offers increased processing speed.

Newest FFMPEG download

Usage example:

ffmpeg -y -benchmark -init_hw_device vulkan -i INPUT.mkv -vf "format=nv12,hwupload,nlmeans_vulkan,hwdownload,format=nv12" -c:v libx264 OUTPUT.mkv

On my system (Ryzen 7600X, RTX 5070 Ti) I get 39.2 FPS with 1920x1080 BluRay footage.
Edit: And 44.4 FPS with +300/+3000 core/memory overclock on the GPU.

[GeoffreyA]

Originally Posted by takla

FFMPEG has merged a rewrite of nlmeans vulkan, which offers increased processing speed.

Newest FFMPEG download

Usage example:

ffmpeg -y -benchmark -init_hw_device vulkan -i INPUT.mkv -vf "format=nv12,hwupload,nlmeans_vulkan,hwdownload,format=nv12" -c:v libx264 OUTPUT.mkv

On my system (Ryzen 7600X, RTX 5070 Ti) I get 39.2 FPS with 1920x1080 BluRay footage.
Edit: And 44.4 FPS with +300/+3000 core/memory overclock on the GPU.

On the 4600G and B580, I get 12 fps on an old build, and 14 fps on the new one. Notably, though, is that the pervasive artefacts are gone.

[tormento]

Now we just need to wait (and hope) that someone will port that to AVS+.

[takla]

Originally Posted by GeoffreyA

On the 4600G and B580, I get 12 fps on an old build, and 14 fps on the new one. Notably, though, is that the pervasive artefacts are gone.

Thats a 16.6% speedup, nice.