[gjaegy]

Hi,

I am currently wondering if premultiplied alpha can still be used when using a full HDR pipeline, where the tone mapping appends at the very end of the frame rendering operation...

My pipeline relies on premultiplied alpha for several stuffs:
- normal 3D objects
- impostors
- low-resolution off-screen particles rendering, kind of impostor, where the content is blit onto the back buffer before tone map

Now, in my opinion, for premultiplied-alpha to produce correct results, it should happen after the tone mapping, and not before.

But then the issue is that it is not possible anymore to tone map at the very end of the pipeline, but each pixel should be tone mapped in the pixel shader...

any thoughts ?

[Andrew Lauritzen]

Originally Posted by gjaegy

Now, in my opinion, for premultiplied-alpha to produce correct results, it should happen after the tone mapping, and not before.

Why do you say this? Shouldn't it happen in linear space? Tone mapping should operate on the final colors reaching the camera/eye.

[gjaegy]

Hi Andy,

I was wondering because of the way I used to tone map each pixel at the end of the pixel shader, in each shader. Now I use the (more correct and cheaper) way to render all pixels in HDR, and then tone map at the very end, by doing a custom-MSAA resolve once the scene has been fully rendered, and blitting the resolved and tone-mapped pixel onto the back buffer.

However, what I used to do in my main shader (for standard objects) was:

- render pixel in HDR
- tone map
- premultiply by alpha
- blend onto RT

Now, I am doing the following:

- render in HDR
- once the scene has been fully rendered, bind the HDR RT
- render a full-screen quad and a custom blit shader
- custom blit shader performs a custom MSAA resolve by tone-mapping each subsample and averaging (paper page 18)

So, basically we don't have the same math, and I was in fact wondering if the second method is still giving correct results...

Below the equation for a given pixel:

method 1:
alpha * (1.0 - exp(-exposure * colour))

method2:
1.0 - exp(-exposure * alpha * colour)

any thoughts ?

[Andrew Lauritzen]

My intuition tells me that all the alpha stuff should be done in linear space since the pre-multiplied RGB portion represents how much color the surface is *adding* (linearly) to the reflected light and the alpha portion represents how much it is occluding (also linearly) from the light behind it. All should be done in linear space (both in terms of gamma and pre-tone map). Thus my intuition is that your second pipeline is more correct.

[gjaegy]

Thanks for your answer Andy.

With a bit of hindsight, and seing the equation again, I think you are right.

The blending operation should happen in linear colour space, and therefore the alpha pre-multiplication should also logically take place in linear space.

I will keep an eye on the graphical result for a while, to make sure it also produces good results visually with different scenes.

Thanks for your feedback man !

Greg