Why renderer is too slow?

[tigrazone]

On my rtx 3060 I have 5-10 fps in reference mode. Other renderers do it faster.

I worked on my fork of rtxpt and seen unoptimal way to recreate env map for every frame. Can it be precalculated, saved somewhere and just use, not calculate every frame? One more bottleneck - by default b6 compression is on. If I switch to off rendering start to work little bit faster.

Maybe not good threading or so renderer is slow and it is boring.

Is ray_query used by default or rtx pipeline?
Ray_query works faster.

Also when I read code it is completely mess - function repeated many times dont know why.
For example many variants of octant convert of normal vector(float3 to uint) and f32tof16.

[fstrugar-nv]

envmap is not re-created every frame, it is created only once after load and then just reused, the code for skipping is here:

RTXPT/Rtxpt/Lighting/Distant/EnvMapBaker.cpp

Line 363 in 764e426

if (!contentsChanged)

The only case where re-creation happens every frame is with procedural sky, in which case using a lot lower resolution is recommended. For static environment we default to 2k x 2k - in case there's re-creation happening every frame, then the

W.r.t. to general performance, the older 30xx architecture has no Shader Execution Reordering which reduces performance significantly. Unfortunately, Vulkan port also doesn't support SER at the moment (known issue) and is in general slower than d3d12 because a lot of functionality has not yet been implemented - these are known issues but I don't know if Vulkan port will be prioritized.

Thanks for the comments on the code - we're working on improving it.

[tigrazone]

older 30xx architecture has no Shader Execution Reordering which reduces performance significantly

I mean performance gain and faster renderers is vk_gltf_renderer for example. Same scenes rendered up to 3-5 times faster w/o SER on same GPU

[fstrugar-nv]

RTX-PT is currently mainly optimized to work with SER and produce realtime output compatible with DLSS-RR denoiser in D3D12. Reference renderer is there for testing & validation.

On a laptop RTX 3060 and with DLSS-RR Performance 1920x1080, the Kitchen scene runs at around 30fps.

I do not know the details of vk_gltf_renderer path tracer and how it handles various PT scenarios, but if it is sufficient for your use case and faster then it's probably a better option for you!

[fstrugar-nv]

This is roughly what's expected on a (laptop) 3060:

[tigrazone]

All ok with realtime mode but reference rendering is slow and boring. Maybe code have many bottlenecks. vk_gltf_renderer use only reference mode rendering and almost same with rtxpt bsdf, pt calculation code

[tigrazone]

I'm try to add env map code from rtxpt to rtxgi and have slowdown up to 5-10 fps slower. Because of it I think env map sample code not too performant and can make slowdowns. And maybe other pieces of code also have bottlenecks in rtxpt

[tigrazone]

This piece of code works without any checks and called every frame

RTXPT/Rtxpt/Lighting/Distant/EnvMapBaker.cpp

Line 533 in 764e426

m_importanceSamplingBaker->Update(commandList, m_cubemap);

Why?