GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#1

Post by markelliot » Tue, Nov 12, 2019, 2:03 PM UTC

GrCUDA: A Polyglot Language Binding for CUDA in GraalVM
devblogs.nvidia.com

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#2

Post by mister_hn » Tue, Nov 12, 2019, 3:45 PM UTC

Why should one use this vs. using C/C++ bindings for CUDA and load them in other languages (if required)?

I believe it underperforms in comparison with raw C or C++ implementation, just because of the overhead that goes through GraalVM

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#3

Post by jjoonathan » Tue, Nov 12, 2019, 3:59 PM UTC

mister_hn wrote: post #2

Why should one use this vs. using C/C++ bindings for CUDA and load them in other languages (if required)? I believe it underperforms in comparison with raw C or C++ implementation, just because of the overhead that goes through GraalVM

Can GrCUDA give you a good testing / CI story? IIRC if you want to test you code on a GPU-less server, the other option is GPUOcelot, which is (also) 3rd party and no longer maintained.

EDIT: or you can install CUDA from 2009 before they stripped out the software emulator and use no features from the last decade, but then you might as well just use OpenCL :)

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#4

Post by diggan » Tue, Nov 12, 2019, 4:04 PM UTC

mister_hn wrote: post #2

Why should one use this vs. using C/C++ bindings for CUDA and load them in other languages (if required)? I believe it underperforms in comparison with raw C or C++ implementation, just because of the overhead that goes through GraalVM

Guess it would be easier to build a library using Cuda in Java + JS + Ruby + Python + Rust at the same time, as that's kind of what GraalVM would help a lot with. Basically, anything polyglot + Cuda would be easier with this.

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#5

Post by pjmlp » Tue, Nov 12, 2019, 4:22 PM UTC

mister_hn wrote: post #2

Why should one use this vs. using C/C++ bindings for CUDA and load them in other languages (if required)? I believe it underperforms in comparison with raw C or C++ implementation, just because of the overhead that goes through GraalVM

Because it is more productive and safer, instead of forcing everyone to learn C and C++.

A lesson that Khronos learnt too late regarding OpenCL.

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#6

Post by chrisseaton » Tue, Nov 12, 2019, 4:29 PM UTC

mister_hn wrote: post #2

Why should one use this vs. using C/C++ bindings for CUDA and load them in other languages (if required)? I believe it underperforms in comparison with raw C or C++ implementation, just because of the overhead that goes through GraalVM

This is an O(1) solution (one integration works for all languages). You're proposing an O(n) solution (binding into individual languages.) That's why.

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#7

Post by hellofunk » Tue, Nov 12, 2019, 4:56 PM UTC

See also: the Futhark programming language.

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#8

Post by xiphias2 » Tue, Nov 12, 2019, 6:02 PM UTC

It's great to have GraalVM CUDA support, but LLVM PTX output of many languages is full of bugs and missing features. I wish NVIDIA would give more support to those integrations that already exist, but experimental for many years now (i.e. Julia, Rust).

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#9

Post by The_rationalist » Tue, Nov 12, 2019, 6:34 PM UTC

What if web browsers included graalvm as an alternative to WebAssembly?

Re: GrCUDA: A Polyglot Language Binding for CUDA in GraalVM

#10

Post by pjmlp » Tue, Nov 12, 2019, 7:08 PM UTC

The_rationalist wrote: post #9

What if web browsers included graalvm as an alternative to WebAssembly?

It isn't fashionable.