LightmapGI: Pack L1 SH coefficients before denoising

[BlueCube3310]

Modifies the code that handles denoising to always pack the L1 coefficients first, then denoise the directional lightmaps. This slightly improves quality and eliminates errors that occur due to dealing with negative color values.

TODO:

• Verify that this approach doesn't alter the directional data,
• Test with OIDN

[jcostello]

Seems that directional data is preserved and works just fine with OIDN. It fixed the issue of losing directional data when denoising with OIDN

[Calinou]

Tested locally, it works as expected. Code looks good to me.

However, JNLM now has a tendency to darken bright static lights, as seen below. This doesn't happen with OIDN.

Testing project: test_pr_109972.zip

No denoiser	JNLM before	JNLM after

OIDN before	OIDN after	"Ground truth"1

Footnotes

1. Ultra quality, supersampling 2.0, no denoising ↩

[BlueCube3310]

However, JNLM now has a tendency to darken bright static lights, as seen below. This doesn't happen with OIDN.

This can probably be fixed by adjusting the denoiser strength, since the directional data is now packed that value can be a constant. The L0 coefficient has 1/4 the brightness of a non-directional lightmap, so an adjustment may be necessary there as well.

[clayjohn]

Makes sense to me

[clayjohn]

@BlueCube3310 Should we go ahead and merge this, or did you want to poke at the JNLM issue first?

[BlueCube3310]

@BlueCube3310 Should we go ahead and merge this, or did you want to poke at the JNLM issue first?

I'll attempt to fix the issue first

[BlueCube3310]

The L0 light weights are now multiplied by 4 to ensure the directional lightmap has the same influence as a regular one. Additionally, the L1 coefficients now use the weights from the L0 layer.

I've also fixed a bug that caused shadowmasks to be unnecessarily denoised 4 times when baking with both them and directional enabled.

CC @Calinou could you confirm whether this fixes the JNLM denoising? (I'm no longer able to download the linked MRP)