Compute Shaders: Optimize your engine using compute / Lou Kramer, AMD

This talk will give an in-depth introduction to compute shaders - how do they differ from vertex and pixel shaders from a programming and from a hardware perspective? What are the typical use cases for compute shaders, what are their advantages and disadvantages - how can you make use of them for your game engine to improve the overall performance? Learn how to write optimal compute shader code which is aware of the structure of GPUs and what functionality the hardware exposes. After the explanation of the basic concepts, we will walk through a practical example which uses a compute shader to perform downsampling, comparing it with an equivalent pixel shader solution.

[@LahceneBelbachir]

Great talk, helped me a lot writing my first compute sahder, thanks.

[@TheHardSteppa]

Great talk! Jam packed with important concepts.

[@WindWolf28]

Thank you, this was a very informative talk!

[@JoeBain]

I'd love to see the source code for the mip map generators.

[@DigitalDuo2211]

Fed goops FTW!

[@DukeAbbaddon]

Lou the copy back is quite important for games developers using the CPU to perform for example user impute like hacking consoles ... "If you anyway transfer the texture from the CPU to the GPU, and you already have the mips or have enough CPU spare time to compute them on the CPU, it makes probably sense to do so. But if you only have the texture on the GPU, and are only using it on the GPU, there is no reason to generate the mips on the CPU. Because this would add at least 2 synchronization points plus the PCI transfer costs for GPU to CPU and CPU to GPU. This overhead can be completely avoided by just computing the mips on the GPU. Thanks, Lou" science.n-helix.com/2019/06/vulkan-stack.html

[@yonjuunininjin]

I think the beginning (what are compute shaders in the graphics pipeline) is much better explained in this video kzbin.info/www/bejne/eHqZkHd_mq17iK8

[@PixelThorn]

That one pixel shader nerd got triggered

[@guywithknife]

hahaha yeah, and I feel he misunderstood a key point. You can see on the second performance chart that at the end, there IS a loss of parallelism, but its a slow ramp down. You still can run all the different thread groups in parallel, you can still run (as Lou said) the lower mip levels in parallel with some of the still running higher mip levels. So, besides the removal of the synchronisation, there's also much better parallel pipelining between mip levels.

[@thegabrielcho]

When will I understand all this stuff? :(

[@amicloud_yt]

Once you have practiced it enough.

[@DukeAbbaddon]

compute shader stage attack for quality enhancement & game mechanics: Physics renderer & gravity I think that we need compute shaders as our angle of attack on the quality of the final output on the console.. The reason i believe this is because you strictly stated that resources can be saved like mipmaps & tessellation & audio... But we need immersive and we need the resultant usable for : Shader cache.. Processing of optimal vector numbers combined with localised & stored vertices arrays... There are many reasons to invoke the pipeline at the compute rather than the graphics shader pipeline.. Execution pipelining, Static resource upgrade & image/Texture & audio Processing of data & storage optimisation or drive spin management... Encryption & High performance computing nodes. There are many reasons we need to aggressively work the shader market: HPC ML Science Gaming Processing is not always about temporary results; Good ones are preferred. RS science.n-helix.com

[@TryNotToHate1]

I’m glad the smart people are still around

[@YourCRTube]

Good talk, but no code, c'mon.

[@Alkis05]

Nice talk, but that accent... Thicc. Couldn't understand a few parts.