fuhailin / rules_cuda

CUDA rules for Bazel

CUDA rules for Bazel

The @rules_cuda repository primarily provides a cuda_library() macro which allows compiling a C++ bazel target containing CUDA device code using nvcc or clang.

For this, rules_cc's auto-configured toolchain is patched to support a cuda feature which is then enabled for cuda_library() targets.

A secondary @local_cuda repository contains bazel targets for the CUDA toolkit installed on the execution machine.

Requirements

Requires bazel 4.0 or later.

Setup

Add the following snippet to your WORKSPACE file:

load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
http_archive(
    name = "rules_cuda",
    sha256 = "f80438bee9906e9ecb1a8a4ae2365374ac1e8a283897281a2db2fb7fcf746333",
    strip_prefix = "runtime-b1c7cce21ba4661c17ac72421c6a0e2015e7bef3/third_party/rules_cuda",
    urls = ["https://github.com/tensorflow/runtime/archive/b1c7cce21ba4661c17ac72421c6a0e2015e7bef3.tar.gz"],
)
load("//cuda:dependencies.bzl", "rules_cuda_dependencies")
rules_cuda_dependencies()
load("@rules_cc//cc:repositories.bzl", "rules_cc_toolchains")
rules_cc_toolchains()

You can test your setup by running the example :

bazel run --@rules_cuda//cuda:enable_cuda @rules_cuda//examples:hello_cuda

Using cuda_library

Then, in the BUILD and/or *.bzl files in your own workspace, you can create C++ targets containing CUDA device code with the cuda_library macro:

load("@rules_cuda//cuda:defs.bzl", "cuda_library")

cuda_library(
    name = "kernel",
    srcs = ["kernel.cu"],
)

Configuration

The normal bazel build command allows configuring a few properties of the @rules_cuda repository.

• --@rules_cuda//cuda:enable_cuda: triggers the is_cuda_enabled config setting. This config setting is not used in the package itself, but is rather intended as a central switch to let users know whether building with CUDA support has been requested.
• --@rules_cuda//cuda:cuda_targets=sm_xy,...: configures the list of CUDA compute architectures to compile for as a comma-separated list. The default is "sm_52". For details, please consult the --cuda-gpu-arch clang flag.
• --@rules_cuda//cuda:copts=...: comma-separated list of arguments to add to cuda_library() compile commands.
• --@rules_cuda//cuda:cuda_runtime=<label>: configures the CUDA runtime target. The default is "@local_cuda//:cuda_runtime_static". This target is implicitly added as a dependency to cuda_library() targets. @rules_cuda//cuda:cuda_runtime should be used as the CUDA runtime target everywhere so that the actual runtime can be configured in a central place.
• --@rules_cuda//cuda:compiler=<compiler>: configures the compiler to use for cuda_library() sources. Supported values are nvcc (default) and clang. The latter uses the auto-configured host compiler with clang-specific CUDA flags, and should therefore be combined with --repo_env=CC=clang.
• --repo_env=CUDA_PATH=<path>: Specifies the path to the locally installed CUDA toolkit. The default is /usr/local/cuda.

Toolchains

This repository patches rules_cc's auto-config toolchain to add support for a 'cuda' feature which triggers compiling CUDA code for the chosen target architectures. The functionality for this is provided by cuda/toolchain.bzl and can also be applied to custom toolchains.