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

[Khanattila]

Originally Posted by Mangix

Latest Nvidia drivers support OpenCL 1.2. Any performance benefits?

Nvidia seems to don't like OpenCL, maybe because of CUDA.
OpenCL 1.2 / 1.1 API are different. The plugin should be rewritten to exploit all.

EDIT. OpenCL 1.2 was presented on November 15, 2011.

[Atak_Snajpera]

What are your thoughts about opencl 2.0? AMD has just added support in latest drivers.

[Khanattila]

Originally Posted by Atak_Snajpera

What are your thoughts about opencl 2.0? AMD has just added support in latest drivers.

Is a compatibility issue. I'd like to use OpenCL 2.0, but I can't.

To achieve the maximum diffusion I have to use OpenCL 1.1.
NVIDIA CUDA Toolkit v7.0 can't stand even OpenCL 1.2. (http://developer.download.nvidia.com/compute/cuda/7_0/Prod/doc/CUDA_Toolkit_Release_Notes.pdf)

I don't want to build two different versions. I hope so.

[Khanattila]

Originally Posted by Stereodude

Here you go. http://www.ozone3d.net/gpudb/gpu.php?which=49162&v=2

Sorry for the late response.
Probably a big "A" use too much private/local memory for yor GPU.

(GPU have 3 three different memories: private memory, local memory, global memory. Where private is the faster and global is the bigger.)

Please reduce the value of A.

[Khanattila]

Probably this week I will release a new version of NLMeansCL2.
Since it will be a major change, for now I'll call it NLMeansCL2b().
So you can keep both versions.

[Khanattila]

NLMeansCL2b (
    clip src, 
    int D (0),            // Temporal windows, disabled in Beta 1
    int A (4),            // Search window
    int S (2),            // Similarity neighborhood window
    int B (0),            // Base window
    int wmode (1),        // Weighting function
    float h (1.8),        // Strength of the filtering
    string device_type ("default"), 
    int y (3), 
    int u (2), 
    int v (2), 
    bool lsb_inout (false), 
    bool info (false)
)

Little explanation:

[wmode = 0] Cauchy weighting function has a very slow decay.
It assign larger weights to dissimilar blocks than the Leclerc
robust function, which will eventually lead to oversmoothing.

[wmode = 1] Leclerc weighting function has a faster decay,
but still assigns positive weights to dissimilar blocks. Original
NLMeans weighting function.

[wmode = 2] Bisquare weighting function use a soft threshold.

Download: removed.

################

New:

• Temporal windows search (disabled in Beta1).
• Cauchy weighting function.
• Bisquare weighting function.

Removed:

• Ay, Sy and By arguments.
• aa argument (simple Euclidean distance).
• sse argument (replaced by wmode).
• legacy kernel.

Changed:

• Patchwise Implementation
• Euclidean distance

Updated:

• NVIDIA CUDA Toolkit v7.0.
• AVS 2.6.0 RC 2 [150331].

[Reel.Deel]

Hi Khanattila, thanks for the update! I tried using NLMeansCL2b with AviSynth+ r1576 on my work computer but I got this error message: "Plugin was designed for a later version of Avisynth (6)"
I updated to r1779 and it works with that version. Was that intentional or is there something else going on? If so will this also be true for anyone using an older version before AviSynth 2.6 R2?

[Khanattila]

Originally Posted by Reel.Deel

Hi Khanattila, thanks for the update! I tried using NLMeansCL2b with AviSynth+ r1576 on my work computer but I got this error message: "Plugin was designed for a later version of Avisynth (6)"
I updated to r1779 and it works with that version. Was that intentional or is there something else going on? If so will this also be true for anyone using an older version before AviSynth 2.6 R2?

Hi! Avisynth 2.6.0 has many bug fixes and improvements. It is a good idea to upgrade. For this I forced users to use an updated version. However, I do not know how Avisynth+ works.

Version 6 is 2.6.0.
Version 5 is 2.6.0a1-a5.
Version 4 is reserved.
Version 3 is 2.5.6.

[StainlessS]

I think there were multiple versions of Avisynth Header VERSION 5, a change was made to the header at some point which made plugins compiled with it
require Avisynth versionv2.6 Alpha 4+. The previous to current version of ClipClop plugin crashed on Avisynth v2.6a3 and previous
(as with other plugins, immediately at startup).

[Groucho2004]

Originally Posted by StainlessS

I think there were multiple versions of Avisynth Header VERSION 5, a change was made to the header at some point which made plugins compiled with it
require Avisynth versionv2.6 Alpha 4+.

I think you're getting this mixed up with the introduction of "AVS_linkage" in 2.6 Alpha4.

As for the "AVISYNTH_INTERFACE_VERSION":
"3" : 2.5.x
"5" : 2.6.0 < RC1
"6" : 2.6.0 >= RC1

As for AVS+, the lastest builds (r17xx) have the header updated to v6.

[StainlessS]

Guilty as charged M'lud

[Khanattila]

KNLMeansCL (
    clip src, 
    int D (0),                        // Temporal window
    int A (4),                        // Search window
    int S (2),                        // Similarity neighborhood window
    int wmode (1),                    // Weighting function
    float h (1.8),                    // Strength of the filtering
    string device_type ("default"), 
    bool lsb_inout (false), 
    bool info (false)
)

Changelog

• v0.5.0 Beta2 (2015-04-27)
  - New: Temporal windows search.
  - New: Cauchy weighting function.
  - New: Bisquare weighting function.
  - Changed: plugin name!
  - Changed: now process always luminace (y).
  - Changed: simple Euclidean distance.
  - Changed: sse argument, replaced by wmode.
  - Removed: u and v arguments.
  - Removed: Ay and Sy.
  - Removed: patchwise Implementation.
  - Removed: aa argument.
  - Removed: legacy kernel.
  - Updated: NVIDIA CUDA Toolkit v7.0.
  - Updated: AVS 2.6.0 RC 2 [150331].

[Groucho2004]

Originally Posted by Khanattila

KNLMeansCL (
    clip src, 
    int D (0),                        // Temporal window
    int A (4),                        // Search window
    int S (2),                        // Similarity neighborhood window
    int wmode (1),                    // Weighting function
    float h (1.8),                    // Strength of the filtering
    string device_type ("default"), 
    bool lsb_inout (false), 
    bool info (false)
)

Changelog

• v0.5.0 Beta2 (2015-04-27)
  - New: Temporal windows search.
  - New: Cauchy weighting function.
  - New: Bisquare weighting function.
  - Changed: plugin name!
  - Changed: now process always luminace (y).
  - Changed: simple Euclidean distance.
  - Changed: sse argument, replaced by wmode.
  - Removed: u and v arguments.
  - Removed: Ay and Sy.
  - Removed: patchwise Implementation.
  - Removed: aa argument.
  - Removed: legacy kernel.
  - Updated: NVIDIA CUDA Toolkit v7.0.
  - Updated: AVS 2.6.0 RC 2 [150331].

Download: KNLMeansCL_v0.5_Beta2.

Thanks for this new version. I suppose the default for "D" (0) disables temporal operation? I get a huge performance hit when I set it to "1" which I guess is normal.
I really like this filter and I think I have to consider upgrading from my old GT240.
So, how about running a simple benchmark to see how various video cards fare? I'm thinking something like this:

colorbars(width = 1280, height = 720, pixel_type = "yv12").killaudio().assumefps(24000, 1001)
trim(0,99)
fadeio(48)
trim(0,99)
KNLMeansCL(D = 1)

Running this script through AVSMeter (-log -gpu), I get these results:

[Runtime info]
Frames processed:               100 (0 - 99)
FPS (min | max | average):      1.421 | 1.452 | 1.448
Memory usage (phys | virt):     44 | 43 MB
Thread count:                   8
CPU usage (average):            25%
GPU usage (average):            99%
Video engine load (average):    0%
GPU memory usage:               67 MB
Time (elapsed):                 00:01:09.051

[Graphics card info]
Card name:               NVIDIA GeForce GT 240
GPU name:                GT215
Memory size:             512
OpenCL version:          OpenCL 1.0 CUDA
Driver version:          6.14.13.4052 (ForceWare 340.52) / XP

[Khanattila]

D is the number of past and future frame that the filter uses for denoising the current frame.

D = 0, only current frame (n).
D = 1, use n - 1, n, n + 1.
D = 2, use n - 2, n - 1, n, n + 1, n + 2.
etc.

D = 0 also exploit the symmetry property of the weights, i.e. w(p, p + q) = w(p + q, p). But this requires an accumulation buffer, probably in an old GPU makes performance down.

(Symmetry is also possible in temporal but requires too many check).

KNLMeansCL(D = 0)

[Runtime info]
Frames processed: 100 (0 - 99)
FPS (min | max | average): 27.89 | 45.95 | 43.44
CPU usage (average): 24%
GPU usage (average): 73%
Thread count: 8
Memory usage (phys | virt): 60 | 75 MB
Time (elapsed): 00:00:02.302

KNLMeansCL(D = 1)

[Runtime info]
Frames processed: 100 (0 - 99)
FPS (min | max | average): 6.984 | 7.730 | 7.598
CPU usage (average): 25%
GPU usage (average): 94%
Thread count: 8
Memory usage (phys | virt): 63 | 77 MB
Time (elapsed): 00:00:13.161

KNLMeansCL(D = 2)

[Runtime info]
Frames processed: 100 (0 - 99)
FPS (min | max | average): 4.368 | 4.653 | 4.587
CPU usage (average): 25%
GPU usage (average): 95%
Thread count: 8
Memory usage (phys | virt): 64 | 79 MB
Time (elapsed): 00:00:21.799

EDIT.
Computational complexity: ((2 * A + 1) * (2 * A +1) * (2 * D + 1) - 1) / (D ? 1 : 2)

[Groucho2004]

I suppose you used the GTX760 for these measurements?

[Khanattila]

Originally Posted by Groucho2004

I suppose you used the GTX760 for these measurements?

[Graphics card info]
Card name:               NVIDIA GeForce GTX 760
GPU name:                GK104
Memory size:             2048
OpenCL version:          OpenCL 1.2 CUDA
Driver version:          9.18.13.5012 WHQL (ForceWare 350.12) / Win8.1 64

EDIT.
Would you try this?
- 720x480. KNLMeansCL(0, 2, 1)
- 720x480. KNLMeansCL(0, 3, 1)
- 720x480. KNLMeansCL(0, 5, 1)

Original 9600 GT take: 100.00 FPS / 52.46 FPS / 18.46 FPS.

(B. Goossens, H.Q. Luong, J. Aelterman, A. Pizurica, and W. Philips,
"A GPU-Accelerated Real-Time NLMeans Algorithm for Denoising Color Video Sequences",
in Proc. ACIVS (2), 2010, pp.46-57. )

[Groucho2004]

Originally Posted by Khanattila

Would you try this?
- 720x480. KNLMeansCL(0, 2, 1)
- 720x480. KNLMeansCL(0, 3, 1)
- 720x480. KNLMeansCL(0, 5, 1)

Original 9600 GT take: 100.00 FPS / 52.46 FPS / 18.46 FPS.

colorbars(width = 720, height = 480, pixel_type = "yv12").killaudio().assumefps(24000, 1001)
trim(0,99)
fadeio(48)
trim(0,99)

KNLMeansCL(0, x, 1) #x = 2, 3, 5

GT 240: 77 FPS / 40 FPS / 16.5 FPS.

How can this be? The 9600 is ancient.

[Khanattila]

Originally Posted by Groucho2004

colorbars(width = 720, height = 480, pixel_type = "yv12").killaudio().assumefps(24000, 1001)
trim(0,99)
fadeio(48)
trim(0,99)

KNLMeansCL(0, x, 1) #x = 2, 3, 5

GT 240: 77 FPS / 40 FPS / 16.5 FPS.

How can this be? The 9600 is ancient.

Code is not fully optimized... but about is the same card: http://www.tomshardware.com/reviews/...40,2475-5.html

[Khanattila]

          Beta2       Beta3
S=0    59.05 FPS    52.44 FPS  -11%
S=1    47.35 FPS    51.82 FPS  + 9%
S=2    43.99 FPS    51.43 FPS  +17%
S=3    38.88 FPS    50.39 FPS  +30%
S=4    36.89 FPS    49.20 FPS  +33%

Ready for the final release.

[Groucho2004]

Originally Posted by Khanattila

KNLMeansCL_v.0.5_Beta3.dll

I'm getting an error with this version:

[build_programm (CL_BUILD_PROGRAM_FAILURE)]

Also, a text file ("KNLMeansCL.txt") is created with this content:

"error: macro 'V_BLOCK_Y' contains embedded newline, text after the newline is ignored."