ILGPU

I believe these are SIMD. Tensor cores require MMA family of instructions. Ask me how I know. :) https://github.com/m4rs-mt/ILGPU/compare/master...lostmsu:IL... Good article: https://alexarmbr.github.io/2024/08/10/How-To-Write-A-Fast-M...

I actually also used it on .NET codebase, specifically https://github.com/m4rs-mt/ILGPU It is just poor at designing a generic solution despite repeated requests to follow the design of existing alternatives (present in the same repro). It tended to plug holes in a broken architecture it came up with on its own instead of redesigning or trying to simplify its code to be able to keep it in its own head. TBH I suspect this might be limited purely by context length. It produced fine(-ish) initial bits so a few tests would pass, but it dug itself a hole of introducing provenance and could not keep track of it properly. You can see it: https://github.com/lostmsu/ILGPU/tree/Vulkan-GPT-5-Stuck TBH2: this was a huge request. But also there are already other backends it could just mirror.

> https://github.com/m4rs-mt/ILGPU/releases/tag/v1.5.1 > Sept 2023. you guys just don't get it - there's a reason why CUDA is a dialect of C/C++ and not C# and it's not because the engineers at NVIDIA have just never heard of C#.

> RISC-V The work is underway: https://github.com/dotnet/runtime/pulls?q=label%3Aarch-riscv > Nvidia's PTX, AMD https://ilgpu.net/ and even https://github.com/m4rs-mt/ILGPU/blob/c3af8f368445d8e6443f36... While not PTX, there's also this project: https://github.com/Sergio0694/ComputeSharp which partially overlaps with what ILGPU offers Arguably, even C++ itself - you are not using "full" C++ but a special subset that works on top of specific abstraction to compile to GPUs, and I was told that CUDA C++ is considered legacy. The original context of discussion is performance and perceived issue of "having runtime", which is what my reply is targeted at. In that context, C# provides you the tools and a solution other languages in the class of Java, Go, TS and anything else interpreted just don't have. So you could reasonably replace a project

Given IL itself is an abstract stack-based bytecode, it can be compiled to the corresponding IR, which can then target corresponding back-end (CUDA, OpenCL, CPU, etc.) - this is what ILGPU does. Because all code is in the single repository in the post and is fairly easy to read, you can skim through it to draw your own conclusions if this interests you. Also, very easy to start using: just `dotnet add package ILGPU` on most configurations (as ADHD puts higher mental strain on activities involving complex configuration, I try to keep to the tools that have minimal ceremony) C# (and F# by extension) generally allow to write system-ish code, with references to locals and same C primitives, which means that you're likely not sacrificing in performance in this particular scenario by having the language be higher-level. After all, you're using

Take a look at ILGPU. It's very nice and has been around for a long time! (just no one knows about it, sadly) Short example: https://github.com/m4rs-mt/ILGPU/blob/master/Samples/SimpleM... Supports even advanced bits like inline PTX assembly: https://github.com/m4rs-mt/ILGPU/blob/master/Samples/InlineP...