Adaptive sharpen - Image sharpening shader

[Popwax]

I tried the GLSL port with MPV, but the result is looking weird, could anyone post a sample screenshot (before and after sharpening) of the HLSL shader, so that I can make a comparison?

[bacondither]

Some test images, used the 2015-11-05 shader with default settings.

Blurred test image

AVS HD 709 - Blu-ray & MP4 Calibration, Sharpness & Overscan

Ghost in the shell

[Popwax]

Originally Posted by bacondither

Some test images, used the 2015-11-05 shader with default settings.

Blurred test image

AVS HD 709 - Blu-ray & MP4 Calibration, Sharpness & Overscan

Ghost in the shell

The GLSL port doesn't produce the exact same sharpened image. It widen the edge sigificantly for the first test image, doesn't change anything visiblely for the second one. There are also slight differences for the third test image. Is this due to version different (GLSL port is based on 2015-09 I believe)I? What's the possible cause which can lead to widen edge in the first test image?

[ts1]

Here is mine comparison (this is without scaling):
http://screenshotcomparison.com/comparison/149655

[Popwax]

Originally Posted by ts1

Here is mine comparison (this is without scaling):
http://screenshotcomparison.com/comparison/149655

Actually what I want to compare are the sharpened images of GLSL and HLSL shader, since I notice distortion of the GLSL shader for noisy images.

bacondither: Could you post more test images for video with heavy film grains?

[ts1]

Yeah, output result is different. MPC-HC vs MPV:
http://screenshotcomparison.com/comparison/149670

[ts1]

This is only with opengl-hq preset (sigmoid-upscaling in particular). With opengl result is identical.

[vood007]

There is some sort of overshot in the glsl (one-pass) version. Upper is HLSL, lower GLSL.

[ts1]

vood007
Try updated glsl 1 pass shader.

[vood007]

Originally Posted by ts1

vood007
Try updated glsl 1 pass shader.

Yep looks fine now, thanks!

[Popwax]

Originally Posted by ts1

This is only with opengl-hq preset (sigmoid-upscaling in particular). With opengl result is identical.

Oh, that make sense. With "sigmoid-upscaling" option the shader is applied on linear RGB, but the original shader explicitly stated to require gamma light.

EDIT: I can confirm that with "sigmoid-upscaling" disabled, the GLSL shader can produce exactly same sharpened image.

Is there any chance the latest experimental AdaptiveSharpen (11/5/2015) will get an update soon to work properly with DirectX 11? I mainly use AdaptiveSharpen in games through ReShade and right now it says that this latest AdaptiveSharpen does not work as intended in DirectX 11. Does it create some odd artifacts in DirectX 11 or slow down performance too much?

[bacondither]

Originally Posted by XMonarchY

Is there any chance the latest experimental AdaptiveSharpen (11/5/2015) will get an update soon to work properly with DirectX 11? I mainly use AdaptiveSharpen in games through ReShade and right now it says that this latest AdaptiveSharpen does not work as intended in DirectX 11. Does it create some odd artifacts in DirectX 11 or slow down performance too much?

Hmm.... does it complain about "warning X3081: potentially unintended use of a comma expression in a variable initializer"?

If so, i have already fixed it. It did not result in any errors just the compiler complaining.

[CruNcher]

if i look @ the GPU usage even that small amount 10% hit is huge for games resources.

[JPulowski]

Originally Posted by bacondither

Hmm.... does it complain about "warning X3081: potentially unintended use of a comma expression in a variable initializer"?

If so, i have already fixed it. It did not result in any errors just the compiler complaining.

No, the compiler does not give any errors. In DX11, white dot artifacts occur around very thin or very bright surfaces (e.g. stars in the sky, in-game text etc.). Will upload a screenshot when I am available. Not sure what is the problem exactly, ReShade's transcompiler might be the culprit but I don't think that is the case unless there are very tricky shader model specific instructions. Lowering the strength of the shader removes some of the artifacts, but it doesn't make any sense since in DX9 the shader works as it should.

Edit: Here it is, notice the black dots around text and stars:
https://i.imgur.com/9lmTylj.png

[bacondither]

Originally Posted by JPulowski

No, the compiler does not give any errors. In DX11, white dot artifacts occur around very thin or very bright surfaces (e.g. stars in the sky, in-game text etc.). Will upload a screenshot when I am available. Not sure what is the problem exactly, ReShade's transcompiler might be the culprit but I don't think that is the case unless there are very tricky shader model specific instructions. Lowering the strength of the shader removes some of the artifacts, but it doesn't make any sense since in DX9 the shader works as it should.

Tested reshade with adaptive sharpen on some dx11 samples.
Found that the problem was with the tanh function that overflowed and resulted in a NaN in dx11 and a 0 in dx9.
It was not really visible in dx9 because it would return 0 when often it should be limited to a very small value near 0.

Modified the AdaptiveSharpen.h so it should work fine now.
Fixed AdaptiveSharpen.h

[bacondither]

New stable version 2015-11-17 and experimental version 2015-11-17-E.

[JPulowski]

Originally Posted by bacondither

Tested reshade with adaptive sharpen on some dx11 samples.
Found that the problem was with the tanh function that overflowed and resulted in a NaN in dx11 and a 0 in dx9.
It was not really visible in dx9 because it would return 0 when often it should be limited to a very small value near 0.

Modified the AdaptiveSharpen.h so it should work fine now.
Fixed AdaptiveSharpen.h

Thanks, I have also uploaded the fix to ReShade's repository.

Would you mind telling me which part of the code have you modified? Or did you completely fix it, so there is no need to apply a hotfix after the code has been ported?

[bacondither]

Originally Posted by JPulowski

Thanks, I have also uploaded the fix to ReShade's repository.

Would you mind telling me which part of the code have you modified? Or did you completely fix it, so there is no need to apply a hotfix after the code has been ported?

Updated everything to version 2015-11-17. The code that fixed the overflow can be found at line 61 and 243-244 in the new AdaptiveSharpen.h.

Diff for AdaptiveSharpen.h

[aufkrawall]

I should have looked at the image in unscaled size to judge brightness, sorry.
Don't want to sound like a broken record, but maybe you could still do something regarding brightness?
It still gets noticeably brighter with a strength of 0.5.

off:


0.5: