Shaders
Ok, so, someone a million years ago decided that GPUs get their own programming language. This is fairly understandable; GPUs execute code in a fundamentally different way than CPUs. It would probably still have been easier for developers if whoever made the first shader language just made a GCC backend or something.
There are probably reasons as to why they didn’t. My guess? probably because GPU architectures change often (think Ampere → Ada Lovelace → Blackwell), unlike x86 which has only been updated once to support 64-bit computing. Still doesn’t mean they couldn’t have made our lives easier.
Architecture
Shaders are generally compiled at runtime. The one exception is SPIR-V shaders which are compiled into an intermediate bytecode beforehand. OpenGL, Windows DirectX, and Apple’s Metal all have their own, separate shader languages.
When compiled, the shaders resemble assembly.1 Then, each GPU will execute a wavefront2 in groups of 64 threads. Each thread still runs one instruction at a time, but every thread runs the same one instruction in parallel.3 Computations themselves are done on ALUs.4
Why the different languages
So whenever something sucks to use, I’ve been trained to assume it’s just Microsoft’s fault.5 So when OpenGL came out with the first mainstream, cross-platform, and open-source shader language known as GLSL, Microsoft came along and said “Hey, you know what would be cool? Our own language” and came out with their own proprietary, Windows-only HLSL soon after.
The Vulkan Era
In the mid-2010s, the brand-new replacement for OpenGL known as Vulkan brought with it SPIR-V, an IR for shaders that aimed to unify this mess. And it was good on paper… until Microsoft decided to be themselves again and introduce their own, proprietary DXIL.
Oh wait I forgot about Apple
No one cares about Apple, we can forget about them.
SPIR-V Now
I’ve been writing shaders for my fluid simulation using a blend of the Rust-like WGSL, but I’m going to port it over to the up-and-coming rust-gpu project which aims to compile honest-to-goodness Rust code to SPIR-V shaders which can then, hopefully, be run by WGPU and not suck.
Footnotes
-
Not like
.sfiles, just the instructions are kind of similar ↩ -
Or a “warp” for Nvidia chips ↩
-
Conditionals and jumps create inefficiencies because certain threads will have to execute extra or fewer instructions ↩
-
Arithmetic Logic Units, aka CUDA for Nvidia ↩
-
IE existing, Visual Studio, DirectX/D3D, UWP, old Edge, the dotnet rewrite, DevOps and VSTS, UEFI and secure boot, WSL existing, Github sucks since purchase, Minecraft sucks since purchase, Windows deleting my
bootctlentry every few months. And those are just the ones I could come up with off the top of my head. ↩