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

[WolframRhodium]

Originally Posted by anton_foy

Too bad, it is really fast and has so much better temporal stability compared to knlmeansCL.

The output of nlm-cuda should be closed to KNLMeansCL and should not result in an increase in image quality.

[anton_foy]

Originally Posted by WolframRhodium

The output of nlm-cuda should be closed to KNLMeansCL and should not result in an increase in image quality.

My mistake, yes they seem to give identical results yet only the first 4 frames seems to differ from eachother even when both using the same settings. But that doesn't bother me. Great plugin

@Tormento
Is the author here on Doom9 still? I cannot access his own forum.

[tormento]

Originally Posted by anton_foy

Is the author here on Doom9 still? I cannot access his own forum.

I think he is lurking but not officially here.

You can’t access or you can’t register?

In the second case, perhaps some users here and on his forum can ask him to create you an account.

He is very capable but also a bit weird.

[JamesJohnston]

I reported this issue on GitHub, but I'm not sure if anyone's monitoring that, so posting it here as well. More details in the GitHub issue: https://github.com/pinterf/KNLMeansCL/issues/6

In short, if the image width is a multiple of 64, KNLMeansCL outputs garbage - as seen in this test case:

plugin_dir = "C:\VideoProject\Software\AviSynth-plugin\"
LoadPlugin(plugin_dir + "KNLMeansCL.dll")

function test(int width) {
    input = ColorBars(width=width, height=480, pixel_type="YV24").Trim(0, 50).Subtitle(String(width), size=36).FadeOut(50).ConvertToY8()
    output = input.KNLMeansCL(ocl_x=8, ocl_y=16, ocl_r=1)
    StackVertical(input, output)
}

StackHorizontal( \
    test(8*1),  test(8*2),  test(8*3),  test(8*4),  test(8*5),  test(8*6),  test(8*7),  test(8*8),  \
    test(8*9),  test(8*10), test(8*11), test(8*12), test(8*13), test(8*14), test(8*15), test(8*16), \
    test(8*17), test(8*18), test(8*19), test(8*20), test(8*21), test(8*22), test(8*23), test(8*24), \
    test(8*25), test(8*26), test(8*27), test(8*28), test(8*29), test(8*30), test(8*31), test(8*32), \
    test(8*33), test(8*34), test(8*35), test(8*36), test(8*37), test(8*38), test(8*39), test(8*40), \
    test(8*41), test(8*42), test(8*43), test(8*44), test(8*45), test(8*46), test(8*47), test(8*48) \
)

I posted an example of an output frame in the GitHub issue. But what happens is the bottom row (filter output) is missing outputs (they are all black) whenever width is a multiple of 64.

In real-world videos, I've seen where the filter output gets stuck at a specific frame, and doesn't change when I move the AvsPmod slider bar to other frames. Other times, it will show partial or total garbage pixels, as if it was displaying uninitialized memory.

In practice, I'm trying to crop a standard definition 720x480 video down by a multiple of 16, which is coming out with a width of 704. That seems like it would be a fairly common scenario...? but that's a multiple of 64, so the rendered output comes out as garbage.

[real.finder]

with intel igpu in 13th gen

ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL()

it's intel problem or KNLMeansCL?

[tormento]

Originally Posted by real.finder

it's intel problem or KNLMeansCL?

OpenCL should be GPU agnostic.

You tried with other GPUs as well?

[real.finder]

Originally Posted by tormento

OpenCL should be GPU agnostic.

You tried with other GPUs as well?

there are problems with KNLMeansCL and amd gpu already, only nvidia seems work fine with KNLMeansCL

I think CPU is the only standard (Especially X86-X64 since in apple m1-m4 is kind of different from another arm from what I've heard, I know someone use x86 port of his program to make it work in m1-m4), so maybe it's good idea to make cpu port of KNLMeansCL (opencl in cpu is dropped in modren cpus) edit: there are https://forum.doom9.org/showthread.php?t=172966 already but not as KNLMeansCL, xNLMeans maybe got things that not in KNLMeansCL but there are also things in KNLMeansCL that not in xNLMeans, so maybe someone update xNLMeans to make it ultimate NLMeans

[qyot27]

Originally Posted by real.finder

with intel igpu in 13th gen

ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL()

it's intel problem or KNLMeansCL?

I can't reproduce that here with an Arc A770, Ubuntu 24.10.

[tormento]

Originally Posted by real.finder

it's good idea to make cpu port of KNLMeansCL

Or not.

There are plugins that works really well under cpu, others really good on GPU. I think KNLMeansCL could be a real performance hit on a CPU.

Why don’t you simply use the CPU version of BM3D? I find it in many ways superior.

[Z2697]

CPU KNLM Choices:
1) CPU "port": https://github.com/AmusementClub/vs-nlm-ispc
2) CPU OpenCL runtime: https://www.intel.com/content/www/us...l-support.html

[Jamaika]

Originally Posted by real.finder

there are problems with KNLMeansCL and amd gpu already, only nvidia seems work fine with KNLMeansCL

I think CPU is the only standard (Especially X86-X64 since in apple m1-m4 is kind of different from another arm from what I've heard, I know someone use x86 port of his program to make it work in m1-m4), so maybe it's good idea to make cpu port of KNLMeansCL (opencl in cpu is dropped in modren cpus) edit: there are https://forum.doom9.org/showthread.php?t=172966 already but not as KNLMeansCL, xNLMeans maybe got things that not in KNLMeansCL but there are also things in KNLMeansCL that not in xNLMeans, so maybe someone update xNLMeans to make it ultimate NLMeans

KNLMeansCL is old __AVISYNTH_8_H__ and openCL 1.2.

plugins/Khanattila/NLMAvisynth.cpp: In constructor 'NLMAvisynth::NLMAvisynth(PClip, int, int, int, double, const char*, int, double, PClip, const char*, int, int, int, int, bool, bool, int, IScriptEnvironment*)':
plugins/Khanattila/NLMAvisynth.cpp:215:38: warning: converting to non-pointer type 'cl_channel_type' {aka 'unsigned int'} from NULL [-Wconversion-null]
  215 |     cl_channel_type channel_type_u = NULL, channel_type_p = NULL;
      |                                      ^~~~
plugins/Khanattila/NLMAvisynth.cpp:215:61: warning: converting to non-pointer type 'cl_channel_type' {aka 'unsigned int'} from NULL [-Wconversion-null]
  215 |     cl_channel_type channel_type_u = NULL, channel_type_p = NULL;
      |                                                             ^~~~
plugins/Khanattila/NLMAvisynth.cpp:216:42: warning: converting to non-pointer type 'cl_channel_type' {aka 'unsigned int'} from NULL [-Wconversion-null]
  216 |     cl_channel_type channel_type_p_out = NULL;
      |                                          ^~~~

ffplay_avx2.exe KNLMeansCL.avs =
ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL()

https://forum.doom9.org/showthread.p...21#post2010721

[real.finder]

Originally Posted by qyot27

I can't reproduce that here with an Arc A770, Ubuntu 24.10.

maybe windows driver bug? like this https://forum.doom9.org/showthread.p...73#post1963373

Originally Posted by tormento

Or not.

There are plugins that works really well under cpu, others really good on GPU. I think KNLMeansCL could be a real performance hit on a CPU.

Why don’t you simply use the CPU version of BM3D? I find it in many ways superior.

the cpu plugins of NLMeans in avs can't be used as drop-in replacement for KNLMeansCL
BM3D is another story

Originally Posted by Z2697

CPU KNLM Choices:
1) CPU "port": https://github.com/AmusementClub/vs-nlm-ispc
2) CPU OpenCL runtime: https://www.intel.com/content/www/us...l-support.html

that for vs only?
anyway thanks for CPU OpenCL runtime link, I thought Intel no longer supported that

Originally Posted by Jamaika

KNLMeansCL is old __AVISYNTH_8_H__ and openCL 1.2.

plugins/Khanattila/NLMAvisynth.cpp: In constructor 'NLMAvisynth::NLMAvisynth(PClip, int, int, int, double, const char*, int, double, PClip, const char*, int, int, int, int, bool, bool, int, IScriptEnvironment*)':
plugins/Khanattila/NLMAvisynth.cpp:215:38: warning: converting to non-pointer type 'cl_channel_type' {aka 'unsigned int'} from NULL [-Wconversion-null]
  215 |     cl_channel_type channel_type_u = NULL, channel_type_p = NULL;
      |                                      ^~~~
plugins/Khanattila/NLMAvisynth.cpp:215:61: warning: converting to non-pointer type 'cl_channel_type' {aka 'unsigned int'} from NULL [-Wconversion-null]
  215 |     cl_channel_type channel_type_u = NULL, channel_type_p = NULL;
      |                                                             ^~~~
plugins/Khanattila/NLMAvisynth.cpp:216:42: warning: converting to non-pointer type 'cl_channel_type' {aka 'unsigned int'} from NULL [-Wconversion-null]
  216 |     cl_channel_type channel_type_p_out = NULL;
      |                                          ^~~~

ffplay_avx2.exe KNLMeansCL.avs =
ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL()

don't know, but avs should be backward compatible unless there are bugs in KNLMeansCL, anyway it still work with nvidia for me (in my home pc, the igpu bug using remote server)

[Jamaika]

You tried my version. Is it the same or not?

[real.finder]

Originally Posted by Jamaika

You tried my version. Is it the same or not?

where? your link is for ffmpeg

[Jamaika]

I added it above. Yes, you run it via ffplay_avx2
https://www.sendspace.com/file/mou7uc

[real.finder]

Originally Posted by Jamaika

I added it above. Yes, you run it via ffplay_avx2
https://www.sendspace.com/file/mou7uc

no need to do it in this case

anyway, it worked with

ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL(device_type="cpu")

after install w_opencl_runtime_p_2024.2.0.980.exe so it's gpu problem (maybe the driver)

[Jamaika]

I will just add that I have installed latest opencl.
https://github.com/KhronosGroup/Open...04b5522fdffcd6

[real.finder]

Originally Posted by Jamaika

I will just add that I have installed latest opencl.
https://github.com/KhronosGroup/Open...04b5522fdffcd6

it's same as https://i.postimg.cc/02HPRr4f/Untitled.png

[pinterf]

Originally Posted by real.finder

with intel igpu in 13th gen

ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL()

it's intel problem or KNLMeansCL?

Reproduced.

device_id = 0 (Nvidia driver)
no problems, this setting is using NVidia 1660 on my machine.

device_id = 1: (Intel driver)
artifacts.
This id is using my Intel UHD 750 driver. It looks like if there were internal format or endian-ness problems (e.g. how a 32 bit data is ordered YUVX or XVUY)

Hint: with

info=true

you can display the actual driver config.

ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL(channels="auto", device_id=1, info=true)

My processor is 11th Gen Intel(R) Core(TM) i7-11700, but obviously, there must be a hidden driver setting which has to be checked to figure out that a pixel format which is good for one driver why is not good for the other.

[Jamaika]

Originally Posted by pinterf

Hint: with

info=true

you can display the actual driver config.

ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL(channels="auto", device_id=1, info=true)

My processor is 11th Gen Intel(R) Core(TM) i7-11700, but obviously, there must be a hidden driver setting which has to be checked to figure out that a pixel format which is good for one driver why is not good for the other.

ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL(channels="auto", device_id=1, info=true)
[avisynth @ 0000023e209d9f60] KNLMeansCL: no compatible opencl platforms available!
(AudioBoost.avs, line 2)
AudioBoost.avs: Unknown error occurred

ColorBars(width=640, height=480, pixel_type="yv12")
KNLMeansCL(channels="auto", device_id=0, info=true)
KNLMeansCL
Version 1.1.1e
Copyright(C) Khanattila
Bits per sample: 8
Search window: 5x5x3
Similarity neightborhood: 9x9
Num of ref pixels: 74
Local work size: 16x16 - 16x8x3
Platform info
Name: NVIDIA CUDA
Version: OpenCL 3.0 CUDA 12.6.65
Device info
Name: NVIDIA Geforce RTX 3050
Vendor: NVIDIA Corporation
Version: OpenCL 3.0 CUDA 560.94

It works.