Support forum for DG Tools

Any news on DGDenoise?

I'm busy trying to implement deinterlacing in DGDecIM. I'll come back to the denoising.

thanks for replying

You're welcome, gonca my friend.

Yeah, but you are talented and determined.

You make me blush.

I got bored today and worked on DGDenoise.

I found a way to eliminate the artifacting that had stopped me before. Now the filter is comparable to KNLMeansCL in speed and quality. It's integrated into DGSource() but can also be invoked stand-alone (stand-alone use requires a DG license). So you can do:

LoadPlugin("DGDecodeNV.dll")
DGSource("file.dgi", strength=0.6)

or

LoadPlugin("DGDecodeNV.dll")
AVISource("file.avi") # or other source filter
DGDenoise(strength=0.6)

I'm pretty happy with it but have to do a bit more testing and write documentation. Also I want to look into CUDA kernel optimization. Then I will release a test version for y'all to bang on.

You da Man

+1

Thanks, gents. I'm working on a 35% speed improvement, but ReadModeNormalizedFloat is not fully cooperating. Don't worry, I will beat it into submission.

Persistence is the key, right gonca?

Yep!
And your good programming skills help as well

Ah, cuTexRefSetFlags(cu_texref, CU_TRSF_READ_AS_INTEGER) during my texture initialization suppresses the action of ReadModeNormalizedFloat! Delete that sucker. So now I have it working and achieved my big performance gain. I'll explain a bit in case anyone else is interested in CUDA kernel optimization.

I had a texture of unsigned chars (the YV12 input data). Textures are good to use because texture reads are optimized for 2D spatial locality, just what we need for NLM. I do a texture read and get an unsigned char out, of course. But I need it to be a float in the range [0,1] for use in the NLM algorithm. So naively I was just dividing it by 255.0f. Now in the innermost loop, I need two such texture reads, so I was getting a big overhead of these divisions. Then I discovered ReadModeNormalizedFloat. It maps to [0,1] automatically during the texture read and does it very fast. I implemented it and it gave the speed-up (because I deleted the divisions) but it messed up my colors. Then reading CUDA docs I found that CU_TRSF_READ_AS_INTEGER suppresses the action of ReadModeNormalizedFloat. I had inherited that flag setting from an nVidia texture sample I started from. So not setting that flag was the key.

Where are we now? Using my 1080p sample rat.264 I ran with both DGDenoise and KNLMeansCL default settings:

DGDenoise: 35 seconds
KNLMeansCL: 53 seconds

The quality is equal as the resulting frames are essentially indistinguishable.

But I'm not happy yet. I have a vector length calculation in the inner loop that I might be able to improve.

Getting close to a beta for y'all.

Wow! This speed is impressive. Can't wait to test it.......

Just finished the documentation. Assembling a beta...

Here you go:

http://rationalqm.us/misc/NLM.rar

I may uprev to 2053 when I include this in the formal distribution.

Don't I need a DGDenoise.dll for the standalone filter (avisynth plugin)?
Where to put the dgdenoise.ptx file?

I want to denoise an .avi using the sourcefilter AviSource() followed by DGDenoise(). I can't index an .avi with DGIndexNV .....

Put the PTX file with DGDecodeNV.dll.

Just load the DGDecodeNV.dll and then put DGDenoise() in your script. You don't need to invoke DGSource().

Is the documentation not clear?

Now it works, thanks. I had still an old DGDenoise.dll in my avisynth plugins folder.
Speed is very good. Congrats! With which settings of KNLMeansCL did you compare (d parameter)?

Give me an hour or so and I will try to run some tests on it
Hope I don't have the same brain fart as last time

Sharc wrote:With which settings of KNLMeansCL did you compare (d parameter)?

Default, i.e., 0.

Here some first results, tested on the fields of a noisy 720x576i video:

KNLMeansCL(0,3,4,4.0) 15.8fps
KNLMeansCL(1,3,4,4.0) 4.9fps
DGDenoise(strength=0.13) 13.2fps (strength>0.2 loosing details)

GPU: GT730

Can you please give me the video and your script so I can look into the detail loss you mentioned, and also compare performance? I have a 1050 Ti.

What did I mess up this time on the script

LoadPlugin("C:\Program Files (Portable)\dgdecnv\DGDecodeNV.dll")
DGSource("W:\HD\WORKFILES\VID_00000.DGI", fieldop=0)
DGDenoise(strength=0.5)
ConvertToYV12().AssumeFPS(24000,1001)

You don't need the ConverttoYV12() and probably also the AssumeFPS(). Also, you can invoke the denoising in DGSource() and omit the DGDenoise() if you like.

Are you having a problem? Make sure you don't have an old DGDenoise.dll lying around to be autoloaded.