Give compute shader raytracing a significant speedup with this one trick.

Показать описание

#gamedev #gamedevelopment #programming

Рекомендации по теме

Комментарии

I remember making a CUDA pathtracer and I found a group size of 8x8 or 16x16 was really good and increased my performance just like you have discovered.

awiseseal

I think this translates well to CUDA, where you don't launch <<<1, 1>>>, but rather <<<64, 512>>> size kernels depending on what's happening. I assume the same thing is happening here. Each compute unit (CU) has a bunch of hardware-implemented threads (like AVX, for example), so 1, 1 will just use 1 thread on 1CU. Nvidia GPUs have 32 "threads" per CU, AMD have 64.

By making the launch 8, 8, that's 64 threads CU, so you might ask, aren't the next 32 gonna wait, since a CU can only fit 32? Yep! But the scheduler can interleave them, or when one batch of 32 threads is waiting on a memory request, the other 32 can sneak in and run. That's why you still see a gain.

Going from 1, 1 -> 2, 2 that's 4x improvement to "occupancy" (1 thread to 16 threads per CU), so we may expect to see a ~4x speedup. The FPS went from 50 -> 180 = 3.6. That's pretty close!
And then, going from 2, 2 -> 4, 4 that's another 4x expected, and we again see ~2.33. Hmm not as much, but hey still good! Remember though there are other parameters at play too! Memory read/write, opengl overhead etc.
Now going from 4, 4 -> 8, 8 doesn't give us all that much. Here, we're probably memory bound, rather than being limited by compute capacity of a CU.

But, the important takeaway is, make sure you fill your CU! Know your hardware! Software abstractions are fine, but at the end of the day, you're juggling electrons, and pushing them down the pathways purpose built for the task is the way to achieve the highest performance!

(I'm either speaking out of my ass here or the concept still applies haha)

dexterman

That's all working because for GPU it's easier to compute all pixels in the buffer, that work on thousand subgroups...

davnoa

It is wild to me that your workgroup size was 1x1. I have never seen anyone go below 8x8.

SilverXenolupus

This is an incredible trick!! The only downside Ive encountered is sometimes it doesnt render the pixels on the far right and bottom of the window.

christophercoronaios

Just wow!! in less than 5 some videos take decades to explain!! #ExponentialDelivery *whoa*!!!

t.e.k.profitstraders

Can someone explains the reason behind this speed up?

alexlau

Give compute shader raytracing a significant speedup with this one trick.

Give compute shader raytracing a significant speedup with this one trick.

WebGPU Realtime Ray Tracing: Compute Shaders 101

Modern OpenGL Tutorial - Compute Shaders

Vulkan Compute Shader Raytracing 1: A single sphere

Learn GODOT 4 Compute Shaders with RAYTRACING!!

Metal with SwiftUI: Realtime Ray Tracing, Compute Shaders

Realtime Raytrace in Unity with Compute Shaders

Getting Started with Compute Shaders in Unity

Realtime Raytracing in Python: Compute Shaders 101

Coding Adventure: Ray Tracing

Vulkan Compute Shader Raytracing 2: Look Mom, no Renderpass!

Ray Tracing in Compute Shaders vs Screen Quad: Performance and Quality Comparison

Vulkan Compute Shader Raytracing 3: Many Spheres

The POWER of COMPUTE SHADERS!!!

OpenGL with C++ 13: Compute Shaders

Planet Terrain Generation using Compute Shaders | Astrodition Indie Game Devlog #6

Real-Time Ray Tracing | 'RTX ON in Vulkan' | Vulkan Lecture Series Ep. 6, Real-Time Render...

Accelerated GPU Ray Tracing

Bringing Ray Tracing to Vulkan

OpenGL fragment shader raytracer

Ray Tracing: How NVIDIA Solved the Impossible!

I Tried Making A Fully Ray Traced Game

Diyu Hu --- Introduction to Vulkan Compute and Compute Shader --- 2022-03-08

Manual Ray Tracing