Implement ReSTIR sampling

[abalfoort]

Code from Sirtsu55:
Sirtsu55@33d91e3

I was reading through the issues and found this one and remembered I saw an implementation of ReSTIR of Sirtsu55.

So, all credits go to Sirtsu55. I don't know if this is even wanted, but it was a relative small change.

[res2k]

Generally, the changes seem to based off an older version and probably need more "merge work".
Eg I don't think get_spherical_triangle_pdfw should simply be removed.

[abalfoort]

I was afraid of that and this is also where my knowledge ends.
Do you want me to close this PR, or do you want to keep it?

[WabbitToonz]

sorry if this is not the propper place to show this
i tried abalfoort restir in actions (date 6/2/2025 )and notised shaows working where they were not and noise seems cleaner .
Here os a picture left is before and right is abalfoort restir
https://ibb.co/27L28PSd

[abalfoort]

Very nice.
Could you also post links of the mods you're using?
I usually keep it clean for development purposes, but that looks really good.

[WabbitToonz]

my mod is unreleased mix of q2rtx /quake 1/ sprawl models
other older videos are on my youtube
https://www.youtube.com/@erickarz9581

[res2k]

i tried abalfoort restir in actions (date 6/2/2025 )and notised shaows working where they were not and noise seems cleaner .
Here os a picture left is before and right is abalfoort restir
https://ibb.co/27L28PSd

Handy. Putting it inline for convenience:

There's certainly less noise and more detail visible (esp. in the scratches).
But there's also light in some places where there was only shadow previously - eg in the left corridor, or the space above it.
Not sure which side is more accurate, tho.

[Edit] Also interesting: the FPS counter is visible, indicating that ReSTIR's FPS are ~10% lower.

[WabbitToonz]

so i wanted to share 2 videos of that arera that in the picture (without and with restir).
the video with restir shows shadows where they were not before.
https://www.dropbox.com/scl/fi/r2mgezyjdmmobdw50yfkd/Quake-2-RTX-Remaster.zip?rlkey=dysm5vqojvnqd5troma5h0nfw&st=zs8pee69&dl=0

[abalfoort]

Nice, it's looking a lot better than expected.
I have been playing for some time now and, apart from the FPS loss, I haven't seen any issues.

@res2k What is needed, if any, to get this PR approved?

[res2k]

Nice, it's looking a lot better than expected. I have been playing for some time now and, apart from the FPS loss, I haven't seen any issues.

@res2k What is needed, if any, to get this PR approved?

Well, you should rather ask that @apanteleev, as he's the only one who can actually commit.

Personally, as stated earlier, I think the changes should be updated to not remove features.
From a quick glance, the following features seem to be steamrolled:

• Sampling of poly lights using spherical triangles. Added in Change sampling of polygonal light to use true spherical triangle sampling #266 to provide better sampling in some cases.
• Spot light emission profiles.

Additionally, I would try to check if using an UBO flag to choose ReSTIR vs "old" sampling, and having shaders carrying all the logic, would have a performance difference compared to eg specialization constant(s) or different variants decided via preprocessor.

[res2k]

Ok, so I'll try to tackle this (as in "bringing it into the shape I'd like it to be"). No idea how long I'll take, though.

[abalfoort]

Thanks. That would be great.
I tried to get into the code with the above mentioned commit #266, but was taken aback by the complexity. That'll take time.
So, if you're willing to upload changes in batches, that would be very helpful.

[abalfoort]

Today I restored #266. That was straight forward.

But now I'm looking at the spotlight emission profiles.
I suppose, in src/refresh/vkpt/main.c the "add_dlight_spot" function needs to be restored. It is called from "add_dlights" after the "case DLIGHT_SPOT" which is heavily changed with this PR.

So, what would be best to do with this part of the code:

				case DLIGHT_SPOT:
					// Old code
					light->type = DYNLIGHT_SPOT;
					add_dlight_spot(light, dynlight_data);

					// New Code - Copy spot data
					VectorCopy(dlight->spot.direction, light->positions + 6);
					light->positions[4] = dlight->spot.cos_total_width;
					light->positions[5] = dlight->spot.cos_falloff_start;
					hash.model = 0xFD;
					break;

[res2k]

But now I'm looking at the spotlight emission profiles.

I've actually already created a branch with the notes features preserved: https://github.com/res2k/Q2RTX/tree/restir-update

However, I'm still collecting some additional things for a PR write-up, hence I didn't publish that branch yet...

[abalfoort]

I see, you're splitting the change in separate commits. Thanks, that way I can better understand the code changes.

Now that you created a new branch, shall we close this PR?

[res2k]

Now that you created a new branch, shall we close this PR?

If you prefer. Though I'd probably wait until the "other" PR.

[res2k]

I see, you're splitting the change in separate commits. Thanks, that way I can better understand the code changes.

Now that you created a new branch, shall we close this PR?

Created #477. There are some more observations in there, take a look.

[apanteleev]

Is this PR superseded by #477? If so, please close this one.

[abalfoort]

Superseded by #477 - closing