ContextQMD
Back to Bear

Generate compile_commands.json for a CUDA project

site/src/guides/recipes/cuda.md

4.2.03.9 KB
Original Source
<!-- Diataxis type: how-to -->

Generate compile_commands.json for a CUDA project

Run the build under Bear exactly as you already run it:

sh
bear -- make

Bear recognizes nvcc, NVIDIA's CUDA compiler driver, under the cuda family (as: cuda), including a cross-compilation target prefix and a version suffix (aarch64-linux-gnu-nvcc, nvcc-12.0). The recorded entry looks like any other compiler entry, with nvcc-specific flags (--gpu-architecture, -gencode, -Xcompiler, and similar) parsed and kept in arguments verbatim:

json
[
  {
    "file": "main.cu",
    "arguments": ["nvcc", "-c", "main.cu", "-o", "main.o"],
    "directory": "/path/to/project",
    "output": "main.o"
  }
]

See Supported compilers for the full recognized-name table and how prefix/suffix recognition works in general.

clangd and clang-tidy do not speak the same nvcc dialect

nvcc's command-line flags are its own, not Clang's CUDA mode: options such as -gencode, --generate-code, -Xcompiler, -Xptxas, and --extended-lambda have no equivalent that Clang-based tooling accepts, so an entry recorded straight from an nvcc invocation can make clangd or clang-tidy reject the command line outright when they try to parse a .cu file with it. Bear itself has no option to filter or rewrite individual flags in a recorded entry; format.arguments in the bear(1) man page only covers @file expansion and a couple of environment-variable foldings, and sources only decides which files get an entry, not what their recorded arguments contain.

The fix is on the consumer side. A .clangd file can strip the unsupported flags before clangd parses the entry:

yaml
CompileFlags:
  Remove: [-gencode*, --generate-code*, -Xcompiler*, -Xptxas*, --extended-lambda]

If you would rather not have clangd touch .cu files at all, suppress its diagnostics on them the same way the bear(1) man page recommends for the valac entries of a mixed C/Vala project:

yaml
If:
  PathMatch: .*\.cu
Diagnostics:
  Suppress: '*'

Or drop CUDA sources from the database entirely with a sources rule in bear.yml (see CONFIGURATION in the man page):

yaml
sources:
  files:
    - pattern: "*.cu"
      action: exclude

nvcc's host-compiler sub-invocations

nvcc splits a .cu file into a device-code path and a host-code path, and for the host side it generates an intermediate file and compiles that with the host compiler (gcc, clang, or MSVC's cl, depending on platform) as a separate process. Bear intercepts that process the same way it intercepts nvcc itself, and because the host compiler's intermediate file has a different name than the original .cu source, this sub-invocation does not collide with nvcc's own entry under Bear's default duplicate matching (directory and file): both land in compile_commands.json as separate entries, one for the .cu file and one for the generated intermediate file the host compiler actually saw. The intermediate file's entry is rarely useful to an editor, since the file it names does not exist once the build finishes; exclude it the same way as .cu files above, with a pattern matching the intermediate filename your local nvcc generates (for example "*.cudafe1.cpp"), or scope a sources rule to directories instead if the host-compiler pass runs from its own temporary directory.