Raytracing in one weekend in c++

GitHub Repository

Part 10

Currently doing a research project & the theme I've chosen for myself is "Optimization".

I was kinda annoyed at the render time especially when going through the 2nd book the final render I did took 5 Hrs and was still quite noisy, so today we will be looking into optimization techniques :

- SIMD

- Multithreading

- Compute Shaders

Each of these should shorten render times by quite a bit and allow us to use more samples & bounces.

Performance Document

Optimization

This code is the finished code of book 2

Should look a little something like this :

img. Source

SIMD

Things to Remember :

- If you're using the correct memory layouts when implementing SIMD, the rest of the problems will solve themselves.

- Parallel performance largely depends on the extent to which your program is parallelized.

- Compiling code with the -march=native flag generates optimized code for your CPU.

With SIMD we managed to cut render time in 2.

case 10: FinalSceneB2(600, 200, 100, 20, 250) :

- Non-Optimized : 5 Hrs

- SIMD Optimized : 2 Hrs 18 Min

MultiThreading

Things to Remember :

After writing shared data, other threads must throw all the work, retrieve the latest value of the variable since the cached version is no longer valid, and then restart their tasks. This phenomenon is often referred to as "false sharing."

Here is an interesting video talking about Cpu Caches.

In the end we managed to boost render time by 2.5x for renders using Multithreading, but we don't gain as much as expected from Hyper-Threading, but honestly I'm not very experienced in hyper-threading outside of this project.

case 10: FinalSceneB2(600, 200, 100, 20, 250) :

- Non-Optimized : 5 Hrs

- Multithreading Optimized : 2 Hrs 9 Min

Compute Shaders

Base Raytracing code & Compute shader code aren't that far apart, so it's not too hard to transfer one to the other just annoying.

One of the issues with GPU porting is that GLSL doesn't map memory into zeros. You should be aware that it doesn't have a constructor that we use in C++ for filling struct memory with zeros

you should take alignment into count since one of the main causes for a black screen was alignment. When using UBO or SSBO.

With Compute Shaders it's almost 10x faster.

case 10: FinalSceneB2(600, 200, 100, 20, 250) :

- Non-Optimized : 5 Hrs

- Compute Shader Optimized : TODO record time

Next up is ...

Buh bye