grCUDA: Polyglot GPU Access in GraalVM

This Truffle language exposes GPUs to the polyglot GraalVM. The goal is to

1. make data exchange between the host language and the GPU efficient without burdening the programmer.

2. allow programmers to invoke existing GPU kernels from their host language.

Supported and tested GraalVM languages:

• Python

• JavaScript/NodeJS

• Ruby

• R

• Java

• C and Rust through the Graal Sulong Component

For details and features of grCUDA language see the grCUDA documentation.

How to bind precompiled kernels to callables, compile and launch kernels is described in the polyglot kernel launch documentation.

Using grCUDA in the GraalVM

grCUDA can be used in the binaries of the GraalVM languages (lli, graalpython, js, R, and ruby). The JAR file containing grCUDA must be appended to the classpath or copied into jre/languages/grcuda of the Graal installation. Note that --jvm and --polyglot must be specified in both cases as well.

The following example shows how create a GPU kernel and two device arrays in JavaScript (NodeJS) and invoke the kernel:

// build kernel from CUDA C/C++ source code const kernelSource = ` __global__ void increment(int *arr, int n) {   auto idx = blockIdx.x * blockDim.x + threadIdx.x;   if (idx < n) {     arr[idx] += 1;   } }` const buildkernel = Polyglot.eval('grcuda', 'buildkernel') const incKernel = buildkernel(   kernelSource, // CUDA kernel source code string   'increment', // kernel name   'pointer, sint32') // kernel signature // allocate device array const n = 100 const deviceArray = Polyglot.eval('grcuda', 'int[100]') for (let i = 0; i < n; i++) {   deviceArray[i] = i // ... and initialize on the host } // launch kernel in grid of 1 block with 128 threads incKernel(1, 128)(deviceArray, n) // print elements from updated array for (let i = 0; i < n; ++i) {   console.log(deviceArray[i]) }

$GRAALVM_DIR/bin/node --polyglot --jvm example.js 1 2 ... 100

The next example shows how to launch an existing compiled GPU kernel from Python. The CUDA kernel

extern "C" __global__ void increment(int *arr, int n) {   auto idx = blockIdx.x * blockDim.x + threadIdx.x;   if (idx < n) {     arr[idx] += 1;   } }

is compiled using nvcc --cubin into a cubin file. The  kernel function can be loaded from the cubin and bound to a callable  object in the host language, here Python.

import polyglot n = 100 device_array = polyglot.eval(language='grcuda', string='int[100]') for i in range(n):   device_array[i] = i inc_kernel = polyglot.eval(   # bind kernel from binary   language='cuda',   string='bindkernel("kernel.cubin", "increment", "pointer, sint32")') # launch kernel as 1 block with 128 threads inc_kernel(1, 128)(device_array, n) for i in range(n):   print(device_array[i])

nvcc --cubin  --generate-code arch=compute_70,code=sm_70 kernel.cu $GRAALVM_DIR/bin/graalpython --polyglot --jvm example.py 1 2 ... 100

For more details on how to invoke existing GPU kernels, see the Documentation on polyglot kernel launches.

Installation

grCUDA can be downloaded as a binary JAR from grcuda/releases and manually copied into a GraalVM installation.

1. Download GraalVM CE 19.2.0.1 for Linux graalvm-ce-linux-amd64-19.2.0.1.tar.gz from GitHub and untar it in your installation directory.

   cd <your installation directory> tar xfz graalvm-ce-linux-amd64-19.2.0.1.tar.gz export GRAALVM_DIR=`pwd`/graalvm-ce-19.2.0.1

2. Download the grCUDA JAR from grcuda/releases

   cd $GRAALVM_DIR/jre/languages mkdir grcuda cp <download folder>/grcuda-0.1.0.jar grcuda

3. Test grCUDA in Node.JS from GraalVM.

   cd $GRAALVM_DIR/bin ./node --jvm --polyglot > arr = Polyglot.eval('grcuda', 'int[5]') [Array: null prototype] [ 0, 0, 0, 0, 0 ]

4. Download other GraalVM languages.

   cd $GRAAL_VM/bin ./gu available ./gu install python ./gu install R ./gu install ruby

Instructions to build grCUDA from Sources

grCUDA requires the mx build tool. Clone the mx reposistory and add the directory into $PATH, such that the mx can be invoked from the command line.

Build grCUDA and the unit tests:

cd <directory containing this REAMDE> mx build

Note that this will also checkout the graal repository.

To run unit tests:

mx unittest com.nvidia

Using grCUDA in a JDK

Make sure that you use the OpenJDK+JVMCI-0.55.

To use the CUDA language from Python:

mx --dynamicimports graalpython --cp-sfx `pwd`/mxbuild/dists/jdk1.8/grcuda.jar     python --polyglot ... >>> import polyglot >>> da = polyglot.eval(language='grcuda', string='double[1000]') >>> da[0] 0.0 >>> da[0] = 1.2 >>> da[0:10] [1.2, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]