Support forum for DG Tools

Earlier I was a bit concerned about whether an external Avisynth+ filter could change the video format, e.g., from YUV420P16 to YV12. Then I realized that is exactly what z_ConvertFormat() does. So I tried it and it is actually quite easy. So I now have a filter that just converts YUV420P16 to YV12 by truncating the lower 8 bits and it works fine. Of course without proper gamut mapping and tone mapping it is washed out as expected. But this forms the basis for implementing the process I gave a few posts up. So, well on the way to nirvana!

admin wrote: ↑

Tue Apr 24, 2018 7:37 am

Earlier I was a bit concerned about whether an external Avisynth+ filter could change the video format, e.g., from YUV420P16 to YV12. Then I realized that is exactly what z_ConvertFormat() does. So I tried it and it is actually quite easy. So I now have a filter that just converts YUV420P16 to YV12 by truncating the lower 8 bits and it works fine. Of course without proper gamut mapping and tone mapping it is washed out as expected. But this forms the basis for implementing the process I gave a few posts up. So, well on the way to nirvana!

Good Avatar.

I have coded the proposed algorithm in a C-code filter, eliminating the calls to z_ConvertFormat() in the script. It works quite well. Surprisingly, the speed is almost identical to the existing DGTonemap(). It's surprising because z_ConvertFormat() is assembler optimized. But apparently that's canceled out by the fact that the script uses two z_ConvertFormat() calls. The image quality is comparable to what we have from the existing DGTonemap script. I expect to see a large speedup when implementing it in CUDA. Keep your fingers crossed.

I have tossed out Hable/Reinhard and rolled my own locally adaptive tonemapping operator. Below is an example. Note the lack of blowout in the clouds and retention of detail there. Also, detail is retained on the dark areas of the left-hand rider. Detail in the hills is brought out nicely. To me, it nicely conveys the 'wow' factor of the original HDR. It's computationally more intensive but I think it's worth it and CUDA should give acceptable frame rates.

Here is the process (all done in C code for now):

* Convert full-depth YUV to 32-bit float RGB.
* Normalize R, G, and B to [0.0,1.0].
* Linearize with ST.2084 EOTF.
* Map the gamut from 2020 to 709 using the inverse of REC.2087.
* Tonemap with locally adaptive algorithm.
* Compress to 709 gamma.
* Scale to 0-255 and convert to YV12.

Amazing...

When do we get to try this ourselves

Thanks. I have to port it to CUDA to make it practical. Patience!

Looking very impressive

Does porting to CUDA imply that it'll be integrated with DGSource and the indexing? Or still a standalone plugin?
Wondering as I can't currently use the indexing for HEVC

Thanks for the hard work

Narkyy wrote: ↑

Tue May 08, 2018 8:14 pm

Does porting to CUDA imply that it'll be integrated with DGSource and the indexing? Or still a standalone plugin?
Wondering as I can't currently use the indexing for HEVC

Thanks for the hard work

You're welcome.

It'll probably be a standalone filter requiring YUV420P16. The utility filters currently in DGDecodeNV.dll can be used independently of DGSource() and indexing, BTW, so even if I integrated it you could still use it. Why can't you currently use indexing for HEVC? No suitable nVidia card?

Yep, and from past support emails from a while ago I couldn't get the 970 to work with DXVA + Win10 for HEVC.
That sounds good then

I have the initial port to CUDA completed with Reinhard tonemapping. It's running at 80 fps on HDR 3840x2160. Pretty good. Software mode runs at about 1 fps, so that's an 80x speedup.

I'll add the option for my local tonemapping and then give y'all a test version. I expect the frame rate to drop by about 25% but we'll see. I'll allow selection of the desired tonemapping operator.

Can hardly wait, you are a busy one

admin wrote: ↑

Thu May 10, 2018 5:05 pm

I have the initial port to CUDA completed with Reinhard tonemapping. It's running at 80 fps on HDR 3840x2160. Pretty good. Software mode runs at about 1 fps, so that's an 80x speedup.

I'll add the option for my local tonemapping and then give y'all a test version. I expect the frame rate to drop by about 25% but we'll see. I'll allow selection of the desired tonemapping operator.

80x, wow.
local tonemapping ... copyrighted ?

Thank you ... again

You're welcome. Actually, local tonemapping is a well-known thing, with several approaches having been tried. The challenge is to make it computationally feasible for video. I am using a histogram equalization approach.

I have a beta for y'all to try:

http://rationalqm.us/misc/DGHDRtoSDR_1.0.rar

Read the included text file DGHDRtoSDR_1.0.txt for usage instructions. Feedback will be gratefully received.

Some parts are not yet ported to CUDA. I am working on that and also improvements to the local tonemapping. I am getting ~80 fps (measured with AVSMeter64) with this version.

Really brings out the background details atm.

DGHDRtoSDR(impl="255", light=50, equal=0)


DGHDRtoSDR(impl="255", light=50, equal=10)

Thanks for your results. Try equal=15 or 20 to really bring out the details. I'm working on not amplifying noise in flat areas at those high levels of equalization.

I'm also curious to know your GPU, CPU, and attained frame rate.

https://i.imgur.com/GJMK9wr.jpg
The flower looks distorted.

My script:

LwLibAvVideoSource("G:\SOURCES\MARI_TO_MAJO_NO_HANA_UHD\BDMV\STREAM\00001.m2ts")
ConvertFromDoubleWidth(bits=10)
ConvertBits(16)
DGHDRtoSDR(impl="255",light=90,equal=2)
DeBicubicResizeMT(1920,1080)

I've got near 14 fps with software decoder.

Upd: imgur used jpeg so it's not that visible on my first screenshot. Hope that one is better: https://i.imgur.com/hcihLxO.png

Had a quick go at it, here's some comparisons

SDR
Hable
Reinhard
DGHDRtoSDR(impl="255",light=50,equal=10)
DGHDRtoSDR(impl="255",light=50,equal=5 )

Some other than default settings
DGHDRtoSDR(impl="255",light=100,equal=5)
DGHable(exposure=1.5, w=1.8)

Around 29 fps average for just the tonemapping at 3840x2160 with prefetch(2)
Ryzen 2700X, GTX 970
Script for speed test:

FFVideoSource("source.mkv", colorspace="YUV420P16")

DGHDRtoSDR(impl="255",light=50,equal=5)
trim(30000,35000)
prefetch(2)

I think it does a great job at keeping detail, a bit messy in darker areas though

Thanks for your testing.

The flower looks distorted.

Can you please give me a source sample that produces this? I can PM you an FTP if you need it.

Maybe I should put a CUDA resizer in there too as your scripted resize is losing you a lot of time.

Thanks for your testing.

Narkyy wrote: ↑

Sat May 12, 2018 2:48 pm

a bit messy in darker areas though

Can you please tell me what you mean by messy?

admin wrote: ↑

Sat May 12, 2018 4:35 pm

Thanks for your testing.

Narkyy wrote: ↑

Sat May 12, 2018 2:48 pm

a bit messy in darker areas though

Can you please tell me what you mean by messy?

Reinhard (equal=0)
DGHDRtoSDR(impl="255",light=50,equal=10)

From equal=0 to equal=10, it seems to affect everything in the frame. If you look at the bottom left, there's a lot of noise introduced because 'detail' is recovered from the dark. Maybe there could be a luma threshold.

Here's another example:

Reinhard (equal=0)
DGHDRtoSDR(impl="255",light=50,equal=10)

Thank you so much for your work.
From my initial tests, I've experienced some heavy banding with the "equal" value being anything other than 0.
Screenshots:

Narkyy wrote: ↑

Sat May 12, 2018 4:54 pm

there's a lot of noise introduced because 'detail' is recovered from the dark.

That's not noise. It is real detail.

dmcs wrote: ↑

Sat May 12, 2018 5:23 pm

From my initial tests, I've experienced some heavy banding with the "equal" value being anything other than 0.

Yes, as I mentioned I have work to do for flat areas. Can you please provide me with your unprocessed source samples? Let me know if you need my FTP.

Software mode 8.5 FPS
GPU mode 112 FPS
I'll post images later, I am getting a HTTP error when trying to attach