Is NRD limited to 16 bits float textures?

[StudenteChamp2]

From NRD.hlsli:

	// X => IN_DIFF_RADIANCE_HITDIST
	// X => IN_SPEC_RADIANCE_HITDIST
	float4 RELAX_FrontEnd_PackRadianceAndHitDist( float3 radiance, float hitDist, bool sanitize )
	{
	  if( sanitize )
		{
  		  radiance = _NRD_IsInvalid( radiance ) ? float3( 0, 0, 0 ) : clamp( radiance, 0, NRD_FP16_MAX );
  		  hitDist = _NRD_IsInvalid( hitDist ) ? 0 : clamp( hitDist, 0, NRD_FP16_MAX );
		}

	 return float4( radiance, hitDist );
	}

What will happen if i use 32 bits float textures?

[dzhdanNV]

What will happen if i use 32 bits float textures?

You can use FP32 if your radiance passed to NRD <= NRD_FP16_MAX.

Is NRD limited to 16 bits float textures?

Internal processing for radiance is FP16. But when creating NRD resources, including internal ones, you can replace RGBA16f to RGBA32f without problems. It can be done by modifying g_NRD_NrdToNriFormat, if you use the NRD integration layer.

[StudenteChamp2]

Exellent thank you <3

[dzhdanNV]

NRD integration layer now supports FP16/FP32 promotion/demotion:

    // Demote FP32 to FP16 (slightly improves performance in exchange of precision loss)
    // (FP32 is used only for viewZ under the hood, all denoisers are FP16 compatible)
    bool demoteFloat32to16 = false;

    // Promote FP16 to FP32 (overkill, kills performance)
    bool promoteFloat16to32 = false;

Rarely needed, but still...