Skip to content

[libcu++] Dynamically load CUDA library instead of using the runtime #6899

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
miscco merged 4 commits into from
Dec 15, 2025
+34 −13
Merged

[libcu++] Dynamically load CUDA library instead of using the runtime #6899

Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
816b857
Dynamically load CUDA library instead of using the runtime
pciolkosz Dec 5, 2025
8fda00c
Switch back to windows.h
pciolkosz Dec 9, 2025
1c52941
Review feedback
pciolkosz Dec 11, 2025
935a975
More review feedback
pciolkosz Dec 11, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
46 changes: 33 additions & 13 deletions libcudacxx/include/cuda/__driver/driver_api.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,11 @@
# include <cuda/std/__internal/namespaces.h>
# include <cuda/std/__type_traits/always_false.h>
# include <cuda/std/__type_traits/is_same.h>
# if _CCCL_OS(WINDOWS)
# include <windows.h>
# else
# include <dlfcn.h>
# endif

# include <cuda.h>

Expand All @@ -46,21 +51,36 @@ _CCCL_BEGIN_NAMESPACE_CUDA_DRIVER
_CCCL_SUPPRESS_DEPRECATED_PUSH

//! @brief Gets the cuGetProcAddress function pointer.
[[nodiscard]] _CCCL_HOST_API inline auto __getProcAddressFn() -> decltype(cuGetProcAddress)*
[[nodiscard]] _CCCL_PUBLIC_HOST_API inline auto __getProcAddressFn() -> decltype(cuGetProcAddress)*
{
// TODO switch to dlopen of libcuda.so instead of the below
void* __fn;
::cudaDriverEntryPointQueryResult __result;
# if _CCCL_CTK_AT_LEAST(13, 0)
::cudaError_t __status =
::cudaGetDriverEntryPointByVersion("cuGetProcAddress", &__fn, 13000, ::cudaEnableDefault, &__result);
# else
::cudaError_t __status = ::cudaGetDriverEntryPoint("cuGetProcAddress", &__fn, ::cudaEnableDefault, &__result);
# endif
if (__status != ::cudaSuccess || __result != ::cudaDriverEntryPointSuccess)
# if _CCCL_OS(WINDOWS)
static auto __driver_library = ::LoadLibraryExA("nvcuda.dll", nullptr, LOAD_LIBRARY_SEARCH_SYSTEM32);
if (__driver_library == nullptr)
{
::cuda::__throw_cuda_error(::cudaErrorUnknown, "Failed to load nvcuda.dll");
}
static void* __fn = ::GetProcAddress(__driver_library, "cuGetProcAddress_v2");
Copy link
Contributor

@davebayer davebayer Dec 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we have a separate variable for the symbol name? It's used twice and I can see us changing it only in one place in the future..

if (__fn == nullptr)
{
::cuda::__throw_cuda_error(::cudaErrorUnknown, "Failed to get cuGetProcAddress_v2 from nvcuda.dll");
}
# else // ^^^ _CCCL_OS(WINDOWS) ^^^ / vvv !_CCCL_OS(WINDOWS) vvv
# if _CCCL_OS(ANDROID)
const char* __driver_library_name = "libcuda.so";
# else // ^^^ _CCCL_OS(ANDROID) ^^^ / vvv !_CCCL_OS(ANDROID) vvv
const char* __driver_library_name = "libcuda.so.1";
# endif // ^^^ !_CCCL_OS(ANDROID) ^^^
static void* __driver_library = ::dlopen(__driver_library_name, RTLD_NOW);
if (__driver_library == nullptr)
{
::cuda::__throw_cuda_error(::cudaErrorUnknown, "Failed to load libcuda.so.1");
Copy link
Contributor

@davebayer davebayer Dec 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Shouldn't we throw cudaErrorInitializationError?

}
static void* __fn = ::dlsym(__driver_library, "cuGetProcAddress_v2");
if (__fn == nullptr)
{
::cuda::__throw_cuda_error(::cudaErrorUnknown, "Failed to get cuGetProcAddress");
::cuda::__throw_cuda_error(::cudaErrorUnknown, "Failed to get cuGetProcAddress_v2 from libcuda.so.1");
}
# endif // ^^^ !_CCCL_OS(WINDOWS) ^^^
return reinterpret_cast<decltype(cuGetProcAddress)*>(__fn);
}

Expand Down Expand Up @@ -151,7 +171,7 @@ _CCCL_HOST_API inline void __call_driver_fn(Fn __fn, const char* __err_msg, Args
//! @return The address of the symbol.
//!
//! @throws @c cuda::cuda_error if the symbol cannot be obtained or the CUDA driver failed to initialize.
[[nodiscard]] _CCCL_HOST_API inline void*
[[nodiscard]] _CCCL_PUBLIC_HOST_API inline void*
__get_driver_entry_point(const char* __name, [[maybe_unused]] int __major = 12, [[maybe_unused]] int __minor = 0)
{
// Get cuGetProcAddress function and call cuInit(0) only on the first call
Expand Down
Loading