Futhark 0.9.1 released – now with CUDA backend | Hacker News Reader

Futhark 0.9.1 released – now with CUDA backend

futhark-lang.org

72 pointsAthas14 comments

Interesting. I wonder if there's a curated list of languages that have GPU backends? I think it's high time for most mainstream languages to add GPUs to their concurrency / parallelism story.

Out of my head, Haskell, Java, .NET, Rust, C, C++, Fortran, D, Julia.

So LLVM?

Actually SPIR and PTX, eventually OpenCL as well.

But is it a SPIR/PTX backend for LLVM (and any language that targets it), or is it a direct backend for Haskell/Rust/Java/C# and all the other implementations that you listed?

It a mix of SPIR/PTX backends for LLVM and proprietary toolchains.

Java,

https://www.ibm.com/support/knowledgecenter/en/SSYKE2_8.0.0/...

http://www.jcuda.de/

https://clojurecuda.uncomplicate.org/

Haskell

Accelerate, http://www.acceleratehs.org/

Obsidian, https://github.com/svenssonjoel/Obsidian

.NET

Hybridizer , http://www.altimesh.com/hybridizer-essentials/,

Alea GPU, http://www.quantalea.com/

NMmath, https://www.centerspace.net/nmath-premium/

DCompute, https://github.com/libmir/dcompute

Julia,

JuliaGPU, https://github.com/JuliaGPU

What's the closest thing to a C implementation currently? Some languages implement at least subsets of it.

Had the wonderful opportunity to not only play around with Futhark last year but also have a lecture from Troels himself when I took a course in parallel functional programming last year. Really like the language and are really interested in what it could do in the future. When I used it last there were still a few kinks to work out to use it but from what I've seen from the releases lately more and more of them are being worked out. Love the direction it is going. Best of luck for the future of the project.

Anyone use Futhark for genetic algorithms or other evolutionary computing uses yet?

I've done a lot of work with GAs, and in my experience the effort is better spent parallelizing the model rather than the evaluations. (You can use a fast model for lots of other stuff besides optimization.) However, if you already have a compact model that's fast to evaluate yet difficult to optimize, then you might have a good candidate for this approach.

The problems I'm interested in have over 12k constraints and need to solve pretty fast. They're among the hardest MIP problems out there and I'd love to see how well a GA compares on a real model, but it will take me a long time to create a comparable GA model.

12k constraints? Without a major change in formulation, you're almost certainly way better off with convex optimization. As long as the problem is convex, anyway.

Points for helping against monopolies. I have been following the project for some years and am very happy to see it grow and on top of a functional front end at that!

Now where is Futhark's GEMM implementation & the vs. CuDNN benchmarks?

cuBLAS does GEMM.

Reply on news.ycombinator.com