dcompute ~master

Native Heterogeneous Computing for D


To use this package, run the following command in your project's root directory:

Manual usage
Put the following dependency into your project's dependences section:

dcompute

Latest version Latest version License Gitter

About

This project is a set of libraries designed to work with LDC to enable native execution of D on GPUs (and other more exotic targets of OpenCL such as FPGAs DSPs, hereafter just 'GPUs') on the OpenCL and CUDA runtimes. As DCompute depends on developments in LDC for the code generation, a relatively recent LDC is required, use 1.8.0 or newer.

There are four main parts:

  • std: A library containing standard functionality for targetting GPUs and abstractions over the intrinsics of OpenCL and CUDA.
  • driver: For handling all the compute API interactions and provide a friendly, easy-to-use, consistent interface. Of course you can always get down to a lower level of interaction if you need to. You can also use this to execute non-D kernels (e.g. OpenCL or CUDA).
  • kernels: A set of standard kernels and primitives to cover a large number of use cases and serve as documentation on how (and how not) to use this library.
  • tests: A framework for testing kernels. The suite is runnable with dub test (see dub.json for the configuration used).

Examples

Kernel:

@kernel void saxpy(GlobalPointer!(float) res,
                   float alpha,
                   GlobalPointer!(float) x,
                   GlobalPointer!(float) y, 
                   size_t N)
{
    auto i = GlobalIndex.x;
    if (i >= N) return;
    res[i] = alpha*x[i] + y[i];
}

Invoke with (CUDA):

q.enqueue!(saxpy)
    ([N,1,1],[1,1,1]) // Grid & block & optional shared memory
    (b_res,alpha,b_x,b_y, N); // kernel arguments

equivalent to the CUDA code

saxpy<<<1,N,0,q>>>(b_res,alpha,b_x,b_y, N);

For more examples and the full code see source/dcompute/tests.

Build Instructions

To build DCompute you will need:

  • ldc as the D dcompiler.
  • a SPIRV capable LLVM (available here to build ldc to to support SPIRV (required for OpenCL)).
  • or LDC built with any LLVM 3.9.1 or greater that has the NVPTX backend enabled, to support CUDA.
  • dub and then just run $dub build or add "dcompute": "~>0.1.0" to your dub.json or dependency "dcompute" version="~>0.1.0" to your dub.sdl.

If you get an error like Error: unrecognized switch '-mdcompute-targets=cuda-210, make sure you are using LDC and not DMD: passing --compiler=/path/to/ldc2 to dub will force it to use /path/to/ldc2 as the D compiler.

A dmd compatible d compiler,dmd, ldmd or gdmd (available as part of ldc and gdc respectively), and cmake for building ldc is also required if you need to build ldc yourself.

Getting Started

Please see the documentation.

TODO

Generate OpenCL builtins from here

Our sponsors

<img src="https://raw.githubusercontent.com/libmir/mir-algorithm/master/images/symmetry.png" height="80" />         <img src="https://raw.githubusercontent.com/libmir/mir-algorithm/master/images/kaleidic.jpeg" height="80" />

Authors:
  • Nicholas Wilson
Dependencies:
derelict-cl, derelict-cuda, taggedalgebraic
Versions:
0.1.0 2017-Oct-15
0.0.0-alpha0 2016-Sep-20
~master 2022-Mar-01
~thewilsonator-patch-3 2018-Apr-09
~thewilsonator-patch-2 2018-Feb-13
Show all 6 versions
Download Stats:
  • 0 downloads today

  • 0 downloads this week

  • 1 downloads this month

  • 320 downloads total

Score:
2.3
Short URL:
dcompute.dub.pm