Matrix Computations on the GPU in Clojure | Hacker News Reader

Matrix Computations on the GPU in Clojure

neanderthal.uncomplicate.org

116 pointsdragandj17 comments

Finally! This looks awesome. But then, now i have to find an AMD graphics card. Any chance that this works on NVIDIA?

In principle, it works on Nvidia, but Nvidia still does not support 1.2... But, new pluggable engine for OpenCL 1.2 is in plan, and it will have optimizations for Nvidia, so, it is not something that is out of reach.

How about Intel/AMD on-die GPUs? (And any thoughts on if that's likely to help -- now, or with future generations -- as opposed to just sticking to the CPU side?)

So the library does not depend on OpenCL version or does?

The library does NOT depend on the version, since it supports pluggable engines. Currently the only available engine is focused on AMD GCN GPUs and uses OpenCL 2.0 features in its kernels. Other engines are possible and planned, notably one that would work with older versions of OpenCL and support tuning for Nvidia.

Going by the last paragraph in the article, the library currently supports "OpenCL 2.0" only, but "OpenCL 1.2" support is planned.

This is really exciting, thanks for sharing. OpenCL is probably the future of everything. AMD is going to be an important part of it for sure. Fantastic to see this explained in the Clojure context; it feels right at home!

If you are interested in programming your own GPU algorithms in clojure and opencl, also look at http://clojurecl.uncomplicate.org

This is great stuff, but it should really be designed as a core.matrix implementation rather than providing it's own API.

The last thing the Clojure community needs is fragmentation with more incompatible libraries.

What are a few good practical examples of the kinds of problems I'd solve with this?

I wrote my master thesis on General purpose GPU programming. The only two reasons you want to code for the GPU is that it can be much faster for some workloads (basically identical computation which can be trivially parralized) and (in some cases) because it can use much less power.

In the particular case of matrix multiplications, lots of problems can be described as matrices (say googles website ranking) and doing calculations on huge matrices is a thing the GPU is great at. That said it is unlikely to be something most people want.

Whenever you need to implement an algorithm that is described using matrices it is much easier and faster to implement it by calling linear algebra functions instead of coding loops yourself. Neanderthal comes into picture when you have lots of numbers and still need it to be fast (nano, micro, milli)seconds instead of minutes and hours (or perhaps days). So, actually any kind of numerical software written in Clojure (and Java) can hugely benefit from this.

So what is the most practical language currently for GPU processing?

It entirely depends on how you want to program your GPU. If you want fine grained control, you should really use raw OpenCL/CUDA. If however you need something higher level, it may be better to choose one of a number of DSLs that are available. E.g. Haskell's accellerate (for array programming), a DSL from the Delite/Forge project.

For general purpose high level programming however, there isn't really anything acceptable. Skeletal/pattern based programming is covered fairly well with frameworks such as SkelCL or SkePU, however (from what I've seen) there doesn't seem to be a nice functional, or even procedural general purpose language for GPU programming.

Python with Theano is surprisingly expressive and performant.

C++.

Or, hopefully, Clojure ;)

Reply on news.ycombinator.com