Unable to get proper output from NRD

[StudenteChamp2]

I am integration NRD in my rendering software and i am close to have it running. But still I am "close". Here what i get when i use Optix denoiser. It works properly https://youtu.be/oGUy7VgZD0o

With NRD when denoising start i get a fullscreen pink, which is the color i use to init the output DirectX 12 textures https://youtu.be/UgYthvk9pPk

To debug here the first frame inputs https://www.dropbox.com/scl/fi/jfdwt470rnn6cjmcqy5yc/NRD-AOVs.zip?rlkey=7q1i2hvqddo01en0g52e3n5vh&st=44rr8jz5&dl=0

There is a lot to say so here the source code: https://www.dropbox.com/scl/fi/r07shlham0tlg94rm0xbv/Source-code.zip?rlkey=goh61m43lakhgnkal02p5wqw9&st=w486xmv8&dl=0

Denoising happens here:

	//=======================================================================================================
	// PERFORM DENOISING
	//=======================================================================================================

	NrdUserPool userPool = {};
	{
		NrdIntegration_SetResource(userPool, nrd::ResourceType::IN_MV, *In_NRD_MotionVectorsTexture);
		NrdIntegration_SetResource(userPool, nrd::ResourceType::IN_VIEWZ, *In_NRD_ViewZTexture);
		NrdIntegration_SetResource(userPool, nrd::ResourceType::IN_NORMAL_ROUGHNESS, *In_NRD_NormalRoughnessTexture);
		NrdIntegration_SetResource(userPool, nrd::ResourceType::IN_BASECOLOR_METALNESS, *In_NRD_BaseColorMetalnessTexture);

		NrdIntegration_SetResource(userPool, nrd::ResourceType::IN_DIFF_RADIANCE_HITDIST, *In_NRD_DiffuseRadianceHitTexture);
		NrdIntegration_SetResource(userPool, nrd::ResourceType::IN_SPEC_RADIANCE_HITDIST, *In_NRD_SpecularRadianceHitTexture);

		NrdIntegration_SetResource(userPool, nrd::ResourceType::OUT_DIFF_RADIANCE_HITDIST, *Out_NRD_DiffuseRadianceHitTexture);
		NrdIntegration_SetResource(userPool, nrd::ResourceType::OUT_SPEC_RADIANCE_HITDIST, *Out_NRD_SpecularRadianceHitTexture);

		NrdIntegration_SetResource(userPool, nrd::ResourceType::OUT_VALIDATION, *Out_NRD_ValidationTexture);
	};

	const nrd::Identifier denoiserId = NRD_ID(REBLUR_DIFFUSE_SPECULAR);
	m_NRD->Denoise(&denoiserId, 1, *ConstantNRDContextSingleton::instance()->getNRICommandBuffer(), userPool);

So when I run denoising I get the color from the OutDebugColor image. This means denoising is not performed at all. I mean NRD doesn't write anything to the output textures(Out_NRD_DiffuseRadianceHitTexture and Out_NRD_SpecularRadianceHitTexture).

By instance commenting this line gives the same result: // m_NRD->Denoise(&denoiserId, 1, *ConstantNRDContextSingleton::instance()->getNRICommandBuffer(), userPool);

So I am doing something wrong but what? Debug layers doesnt claim anything. And unfortunately my pipeline is not linked to my swapchain so take a PIX capture wont give more info. The only suspicious things i can think is that i am using a RGBA32FImage for any NRD inputs. I should use RGBA8 for the out validation texture but the issue is not there. I will try to implement the specific formats tomorow and see if its helps.

[dzhdanNV]

All provided info is useless, because it's just a prerequisite to get NRD running. Please, enable NRD validation:

• set CommonSettings::enableValidation = true
• pass OUT_VALIDATION (already here) and manually overlay it on top of your final image
• grab and attached a screenshot

https://github.com/NVIDIAGameWorks/RayTracingDenoiser?tab=readme-ov-file#validation-layer

[UPDATE]

i am using a RGBA32F image for any NRD inputs

If NormalEncoding and RoughnessEncoding are respected, it probably should work... except materialID. So, I hope R10_G10_B10_A2_UNORM has not been selected for normal encoding, i.e. NRD is compiled with another encoding, because this one is the default one)

[StudenteChamp2]

1. Here the validation output

2. Here the CPU input fed to NRD https://www.dropbox.com/scl/fi/jfdwt470rnn6cjmcqy5yc/NRD-AOVs.zip?rlkey=7q1i2hvqddo01en0g52e3n5vh&st=rgp6p87h&dl=0

Thanks for helping :)

[dzhdanNV]

I would assume that after getting this output you could say something like "Something is wrong. I will get back to you after fixing major problems on my side"...

Obvious observations:

• "Roughness" looks more like normals. Mismatched encoding?
• "Z" looks like a solid color and most likely wrong
• "Units" shows black, but should consist of "1 unit" blocks
• Diff and spec "hitT" are 0 (it's fine for the initial integration)

Knowledge:

• README
• NRD sample

[StudenteChamp2]

Yes something is clearly wrong. I am using a similar pipeline with DLSS and it works https://computergraphics.stackexchange.com/questions/14226/what-is-the-fastest-way-to-readback-a-directx12-texture-on-the-cpu

Will double check my current implementation.

[dzhdanNV]

Additional thought - to work NRD requires a valid sequence of frames. NRD has history, which is tried to be preserved in a good state. It means that NRD expects at least 10 FPS (bare minimum) if the camera moves, or at least a valid sequence of frames with any performance if the final image is absolutely static (but why not use REFERENCE denoiser here?). Does it meet your expectations?

[StudenteChamp2]

Camera is static for now but i do plan to support animation. I also need interactive framerates.

AMD have their own denoiser but it seems to only support reflections aka it doesn't denoise primary rays. https://gpuopen.com/fidelityfx-denoiser/

So NRD seems to be the best solution for me. But I am struggling to have it run fine. The fact i am CPU based doesnt help. I will give it another shot this weekend :)

For your first question I solve the monte carlo integration by using a cumulative moving average. Every time a new frame is rendered it is merged with the previous one then I feed the merge to NRD/Optix

[dzhdanNV]

The discussion is in stagnation. Closing, since it's not an NRD issue. Feel free to reopen, if there is a real need.