David Mertens


ExtUtils::nvcc - CUDA compiler and linker wrapper for Perl's toolchain


This documentation explains the use of ExtUtils::nvcc version 0.03.


I have included a fully working example for Inline::C, as well as other partial examples for ExtUtils::MakeMaker and Module::Build.


 use strict;
 use warnings;
 use ExtUtils::nvcc;
 # Here's the magic sauce
 use Inline C => DATA => ExtUtils::nvcc::Inline;
 # The rest of this is just a working example
 # Generate a series of 100 sequential values and pack them
 # as an array of floats:
 my $data = pack('f*', 1..100); 
 # Call the Perl-callable wrapper to the CUDA kernel:
 # Print the results
 print "Got ", join (', ', unpack('f*', $data)), "\n";
 END {
     # I am having trouble with memory leaks. This messgae
     # indicates that the segmentation fault occurrs after
     # the end of the script's execution.
     print "Really done!\n";
 // This is a very simple CUDA kernel that triples the value of the
 // global data associated with the location at threadIdx.x. NOTE: this
 // is a particularly good example of BAD programming - it should be
 // more defensive. It is just a proof of concept, to show that you can
 // indeed write CUDA kernels using Inline::C.
 __global__ void triple(float * data_g) {
     data_g[threadIdx.x] *= 3;
 // NOTE: Do not make such a kernel a regular habit. Generally, copying
 // data to and from the device is very, very slow (compared with all
 // other CUDA operations). This is just a proof of concept.
 void cuda_test(char * input) {
     // Inline::C knows how to massage a Perl scalar into a char
     // array (pointer), which I can easily cast as a float pointer:
     float * data = (float * ) input;
     // Allocate the memory of the device:
     float * data_d;
     unsigned int data_bytes = sizeof(float) * 100;
     cudaMalloc(&data_d, data_bytes);
     // Copy the host memory to the device:
     cudaMemcpy(data_d, data, data_bytes, cudaMemcpyHostToDevice);
     // Print a status indicator and execuate the kernel
     printf("Trippling values via CUDA\n");
     // Execute the kernel:
     triple <<<1, 100>>>(data_d);
     // Copy the contents back to the Perl scalar:
     cudaMemcpy(data, data_d, data_bytes, cudaMemcpyDeviceToHost);
     // Free the device memory


 # In your Makefile.PL:
 use ExtUtils::MakeMaker;
 use ExtUtils::nvcc;
     # ... other options ...


 # In your Build.PL file:
 use Module::Build;
 use ExtUtils::nvcc;
 my $build = Module::Build->new(
     # ... other options ...
     config => {ExtUtils::nvcc::MB},


This module serves as the configuration front-end to a Perl module that knows how to translate arbitrary command-line arguments into nvcc-digestable command-line arguments. This means you can use nvcc to compile CUDA code for Perl. I discuss that functionality in ExtUtils::nvcc::Backend.

This module functions for your day-to-day use of ExtUtils::nvcc (if there is such a thing as day-to-day use of a toolchain). It provides a few functions that generate the configuration keys necessary to get Inline::C, ExtUtils::MakeMaker, and Module::Build to use ExtUtils::nvcc::Backend to compile your CUDA code. The functions you would use are, respectively, ExtUtils::nvcc::Inline, ExtUtils::nvcc::EUMM, and ExtUtils::nvcc::MB.


If you want to use CUDA in your Inline::C scripts, you simply need to add the proper configuration options. Those options are generated for you (in an alternating key => value list) by the function ExtUtils::nvcc::Inline:

 use ExtUtils::nvcc;
 use Inline C => DATA => ExtUtils::nvcc::Inline;

This is equivalent to the following:

 use Inline C => DATA =>
   CC => "$^X -MExtUtils::nvcc::Backend -eExtUtils::nvcc::Backend::compiler --",
   LD => "$^X -MExtUtils::nvcc::Backend -eExtUtils::nvcc::Backend::linker --";

And now you understand why you would use the function that generates these options for you. If you find that you are having trouble, you can get more verbose output by calling Inline with the verbose key, as discussed under "Optional Arguments":

 use Inline C => DATA => ExtUtils::nvcc::Inline('verbose');


If you want to use CUDA in your XS files and you use ExtUtils::MakeMaker to handle your distribution, ExtUtils::nvcc provides a function to specify your compiler and linker so that they use ExtUtils::nvcc to properly handle the compiler and linker arguments for you. The function is called ExtUtils::nvcc::EUMM. You would place it directly into your options for WriteMakefile like so:

 use ExtUtils::nvcc;
     # ... other options ...

This is equivalent to the following:

     # ... other options ...
     CC => "$^X -MExtUtils::nvcc -eExtUtils::nvcc::compiler --",
     LD => "$^X -MExtUtils::nvcc -eExtUtils::nvcc::linker --",

If you want more verbose output, you can call ExtUtils::nvcc::EUMM with the verbose argument, as discussed under "Optional Arguments":

 use ExtUtils::nvcc;
     # ... other options ...


If you want to use CUDA in your XS files and you use Module::Build to manage your distribution, ExtUtils::nvcc provides a simple function that will help set up your build configuration. As with the others, it returns a useful key => value collection, this time applicable to the config anonymous hash that goes into Module::Build's constructor. Here's how you should use it:

 use ExtUtils::nvcc;
 my $build = Module::Build->new(
     # ... other options ...
     config => {ExtUtils::nvcc::MB},

If you run into compile trouble and you suspect it has to do with how ExtUtils::nvcc is processing your compiler's command-line arguments, you can request a verbose output by calling MB with the argument verbose, as discussed under "Optional Arguments":

 use ExtUtils::nvcc;
 my $build = Module::Build->new(
     # ... other options ...
     config => {ExtUtils::nvcc::MB('verbose')},

Optional Arguments

I would hope that if there's a problem during compilation, it's due to a mistake in your code and not this module. However, this toolchain is still quite new and rather untested, so errors may sometimes arise due to ExtUtils::nvcc::Backend mis-handling an argument. Verbosity, and possibly other arguments in the future, can be sent to ExtUtils::nvcc::Backend to help you sort it all out. All you need to do is supply the word 'verbose' as an argument to "Inline", "EUMM", or "MB".


This is not really a user-level function, but I feel obligated to document it. It is used by the user-level toolchain configuration functions Inline, EUMM, and MB to convert the mode (compiler or linker) along with the options into the command-line needed to invoke ExtUtils::nvcc::Backend as a compiler or linker. If you're not working on ExtUtils::nvcc itself, don't worry about this function.


Windows usage presents a couple of difficulties, as described in this section.

Visual Studio Only

Unfortunately, nVidia's compiler wrapper (nvcc) only supports the use of cl.exe on Windows. This means that Cygwin and Strawberry Perl users are out of luck for using ExtUtils::nvcc. I attempted to install Visual Studio alongside Strawberry Perl, but the Perl toolchain passes along the gcc flags, which cl.exe does not like. There may be a way to fiddle with the configuration a bit, but I wouldn't hold my breath.

An alternative may be to create a drop-in cl.exe replacement which parses the arguments for cl.exe and invokes gcc. If that's not reverse-engineering reverse-engineered, I don't know what is.

Visual Studio Command Prompt

When you install Visual Studio (as of Visual Studio 2010), you will get a Start Menu entry for Visual Studio Command Prompt. You should run your build processes (i.e. cpan) from one of these command prompts. Among other things, this command prompt sets all of the necessary environment variables to ensure that the compiler can be found, and that the compiler can find all the necessary libraries. This may or may not be necessary for using nvcc directly, but it is certainly is necessary for the rest of the Perl toolchain to find cl.exe and friends.


ExtUtils::nvcc could croak for a number of reasons. To keep things concise, I list both the front-end and the back-end diagnostic messages here. I begin with the front-end errors, errors that ExtUtils::nvcc will throw at you:

Bad mode; must be either 'compiler' or 'linker'

This is an internal error that gets thrown when build_args is called with an invalid mode. If you see this, either you are build_args yourself, and not supplying the string 'compiler' or 'linker', or there is an internal error. In the latter case, please report the error to the bug-tracker listed below.

Bad arguments: <bad-arg-1>, <bad-arg-2>, ...

This message means that you supplied an invalid argument to one of the user-level functions Inline, EUMM, or MB. Check your spelling and capitalization against "Optional Arguments" discussed above.

These are the back-end errors, errors that ExtUtils::nvcc::Backend will have:

Last argument [[<arg>]] left me expecting a value, but I didn't find one

Apparently you (or the build system) supplied a list of arguments to ExtUtils::nvcc::Backend ending with an option that expects an argument. For example, the -o option is a very common option that indicates the output filename from the compilation or linking process. If you supply a -o option to ExtUtils::nvcc::Backend, it expects the following argument to be the output filename. If this is your last argument and you don't supply a filename, this error will be thrown.

If the last argument is complete, to the best of your knowledge, it could be that ExtUtils::nvcc::Backend mis-parsed your command-line arguments in other ways. You should enable verbose output and study that for more details.

Nothing to do! You didn't give me any arguments, not even a file!

This means that you somehow invoked ExtUtils::nvcc::Backend without a single argument. Double-check your command-line invocation and try again.

You must provide at least one source file

Somehow you invoked ExtUtils::nvcc::Backend without a source file listed. Double-check your command-line invocation and try again.

Unable to run nvcc. Is it in your path?

This error means that nvcc cannot be found (or more precisely, that nvcc -V does not give a meaningful result). As the error suggests, be sure to check that nVidia's nvcc is in your path. You will also get this error if you do not have nvcc installed. In that case, install nVidia's CUDA toolkit and you should be ready to go.

nvcc encountered a problem

In this case, nvcc attempted to compile or link your code and failed. Look over the compiler/linker output for clues as to where your code went wrong. My guess is that there is a compiler error in your code.

However, it is possible that the Backend is out-of-touch with your version of nvcc, and it (for example) passed an nvcc argument through to your compiler. Your compiler won't like that and it will likely complain. In that case, file a bug report, as discussed in "BUGS AND LIMITATIONS".


This toolchain requires that you have the following pieces:

nVidia's CUDA toolkit

You must have nVidia's CUDA toolkit in order to compile CUDA code. This module ultimately calls nvcc to perform the compilation; it cannot compile your CUDA code itself. Furthermore, nvcc requires a C++ compiler, so you'll need to be sure you have one of those. The CUDA toolkit is only available for a handful of systems, and this module does not support building CUDA-capable modules for other systems. For example, the latest version of Ubuntu or Fedora may not be supported, and as of this time of writing Gentoo and Arch Linux (and many others) have no support at all.

A Perl development environment

You will need to have access to the Perl development toolchain, either ExtUtils::MakeMaker or Module::Build. (Note Inline uses EU::MM on the backend.) If you are in an environment in which you do not have these tools, you will be able to use ExtUtils::nvcc::Backend, but you'll have a hard time tying anything into Perl.

gcc (Linux) or Visual Studio (Windows)

The nvcc compiler only supports gcc on Linux, and cl.exe on Windows. You cannot specify an alternative compiler.


The code for ExtUtils::nvcc is hosted at github, but please file bugs at https://rt.cpan.org/Public/Bug/Report.html?Queue=ExtUtils-nvcc.

A major maintainability problem is that the Backend has a very ad-hoc parsing scheme that is not systematically tested at the moment. It would be better, I think, to query nvcc at runtime for arguments that it accepts so that there could never be a version skew for the arguments that the Backend parses and the arguments that nvcc accepts. However, nvcc does not have an easily parsed representation of its arguments, so this is probably equally troublesome.

For Windows users, a major issue is that nvcc only works with Microsoft's compiler, cl.exe, on Windows machines. As such, ExtUtils::nvcc will not operate correctly under Cygwin or Strawberry Perl. I would like to remedy this situation. Please let me know if you find a work-around for Strawberry Perl or Cygwin.

Furthermore, ExtUtils::nvcc doesn't even work with Windows at the moment. This module was developed on Ubuntu and I have only dabbled with the Windows build system. It is giving trouble, and any help would be much appreciated.


In the few months that ExtUtils::nvcc has been quietly sitting on CPAN, it has received zero test reports. That's because there isn't a single automated tester that has nvcc installed. If you try to install this module, please report your success or failure at cpantesters. The process is a bit involved, but your test reports will help to turn this piddly little thing into a useful tool! You can read more here: http://wiki.cpantesters.org/wiki/QuickStart.


The only major missing piece at this time is a true test of the build process using ExtUtils::MakeMake and/or Module::Build. I would like to add tests of these to the test suite, but it'll take some thought to figure out how to invoke the build system from inside the test suite.


I have obfuscated my email address. Simply remove the portion that would not be sensible for a Perl developer.

David Mertens <dcmertens.perl.csharp@gmail.com>


The source code for this project is on github at "run4flat/perl_nvcc" in github.com.

This is intended to be part of the toolchain to enable CUDA. A minimalistic Perl module for CUDA is in the works and can be found at "run4flat/perl-CUDA-Minimal" in github.com.

You can read more about CUDA at nVidia's website: http://www.nvidia.com/object/cuda_home_new.html

An alternative to embedding CUDA in C or XS is KappaCUDA.

Other important and related toolchain modules include Inline::C, Module::Build, ExtUtils::MakeMaker


Copyright (c) 2010-2011 David Mertens. All rights reserved.

This module is free software; you can redistribute it and/or modify it under the same terms as Perl itself. See perlartistic.

This software is distributed in the hope that it will be useful, but without any warranty; without even the implied warranty of merchantability or fitness for a particular purpose.

Hosting generously
sponsored by Bytemark