a premake5 extension for cuda

premake5-CUDA

Compiles CUDA code using the Visual Studio CUDA Toolkit extension on Windows and nvcc on Linux. Enabled macros (listed in src/cuda-exported-variables.lua):

• cudaFiles (Table) -> list of files to be compiled by NVCC to binary (relative path from solution root).
• cudaPTXFiles (Table) -> list of files to be compiled by NVCC to PTX (relative path from solution root) - Windows only.
• cudaRelocatableCode (Bool) -> triggers -rdc=true.
• cudaExtensibleWholeProgram (Bool) -> triggers extensible whole program compilation.
• cudaCompilerOptions (Table) -> passed to nvcc.
• cudaLinkerOptions (Table) -> passed to nvlink.
• cudaFastMath (Bool) -> triggers fast math optimizations.
• cudaVerbosePTXAS (Bool) -> triggers code gen verbosity.
• cudaMaxRegCount (String) -> number to determine the max used registers.
• cudaKeep (Bool) -> keeps preprocessed output.
• cudaPath (String) -> custom CUDA install path.
• cudaGenLineInfo (Bool) -> generates line info.
• cudaIntDir (String) -> Intermediary directory for CUDA files - Windows only.
• cudaKeepDir (String) -> Directory to place cudaKeep files (on Linux, requires an existing directory).

The following functions are provided:

• detectNvccVersion() -> try to detect the default version of nvcc on the system.
• detectNvccVersion(cudaPath) -> try to detect the version of nvcc from a provided path.

---

Notes for Windows:

Files specified by files are compiled by cl and not nvcc.

An example is provided in the test folder where a CUDA executable project containing C++, PTX and CUDA files is linked against a CUDA shared library project. If you clone this repo recursively (i.e. with -recursive), it will also pull the premake5 repo, which can be used via the makefile to build premake and then the tests (e.g. via the nmake command). You do not need the premake5 repo unless you want to build the tests.

To use:

• Install the CUDA toolkit, along with its Visual Studio integration.
• Copy the premake5-cuda folder to your project.
• Include it in your premake5.lua file as shown in the example.

Tested with Visual Studio 2022 (toolkit v143) with CUDA toolkit 12.1 VS integration.

Note: If PTX is requested, it will be found in the output object folder, with the .obj extension, though, it can be opened with a text editor for inspection.

---

Notes for Linux:

This extension was primarily made for VS on Windows, with the CUDA toolkit extension. Linux is a work in progress and does not have feature parity with Windows yet. Differences are:

• toolset must be set as "nvcc".
• rules must be set to 'cu'.
• cudaPTXFiles and cudaIntDir not supported.
• the list of cuda files must be provided in files instead of cudaFiles.
• unlike Windows, the whole project is compiled by nvcc and not just the .cu files.

See test premake5 file for how linux config differs. The differences are summarised here:

  if os.target() == "windows" then
    cudaFiles { "exe/cu/**.cu" } -- files to be compiled into binaries by VS CUDA.
    cudaPTXFiles { "exe/ptx/**.cu" } -- files to be compiled into ptx, Windows only.
  else
    toolset "nvcc"
    files { "exe/cu/**.cu" }
    rules {"cu"}
  end

Admittedly, Linux support is a bit clunky but it should get the job done.