Skip to content

Enhance RDC detection and add _CCCL_HAS_DEVICE_RUNTIME() macro #7049

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
davebayer merged 1 commit into from
Jan 5, 2026
Merged

Enhance RDC detection and add _CCCL_HAS_DEVICE_RUNTIME() macro #7049

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
9 changes: 2 additions & 7 deletions cub/cub/detail/detect_cuda_runtime.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,11 +20,6 @@
# pragma system_header
#endif // no system header

// CUDA headers might not be present when using NVRTC, see NVIDIA/cccl#2095 for detail
#if !_CCCL_COMPILER(NVRTC)
# include <cuda_runtime_api.h>
#endif // !_CCCL_COMPILER(NVRTC)

#ifdef _CCCL_DOXYGEN_INVOKED // Only parse this during doxygen passes:
//! Defined if RDC is enabled and CUB_DISABLE_CDP is not defined.
//! Deprecated [Since 3.2]
Expand All @@ -40,9 +35,9 @@
# define CUB_RUNTIME_FUNCTION
#else // Non-doxygen pass:

# if _CCCL_HAS_RDC()
# if _CCCL_HAS_CDP()
# define CUB_RDC_ENABLED
# endif // _CCCL_HAS_RDC()
# endif // _CCCL_HAS_CDP()

# ifndef CUB_RUNTIME_FUNCTION
# define CUB_RUNTIME_FUNCTION _CCCL_CDP_API
Expand Down
25 changes: 16 additions & 9 deletions libcudacxx/include/cuda/__nvtx/nvtx.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,19 +94,26 @@ _CCCL_END_NAMESPACE_CUDA
# define _CCCL_BEFORE_NVTX_RANGE_SCOPE(name)
# endif // !CCCL_DETAIL_BEFORE_NVTX_RANGE_SCOPE

# if _CCCL_HOST_COMPILATION()
// Conditionally inserts a NVTX range starting here until the end of the current function scope in host code. Does
// nothing in device code.
// The optional is needed to defer the construction of an NVTX range (host-only code) and message string registration
// into a dispatch region running only on the host, while preserving the semantic scope where the range is declared.
# define _CCCL_NVTX_RANGE_SCOPE_IF(condition, name) \
_CCCL_BEFORE_NVTX_RANGE_SCOPE(name) \
::cuda::std::optional<::nvtx3::v1::scoped_range_in<::cuda::detail::NVTXCCCLDomain>> __cuda_nvtx3_range; \
NV_IF_TARGET( \
NV_IS_HOST, \
static const ::nvtx3::v1::registered_string_in<::cuda::detail::NVTXCCCLDomain> __cuda_nvtx3_func_name{name}; \
static const ::nvtx3::v1::event_attributes __cuda_nvtx3_func_attr{__cuda_nvtx3_func_name}; \
if (condition) __cuda_nvtx3_range.emplace(__cuda_nvtx3_func_attr); \
(void) __cuda_nvtx3_range;)
# define _CCCL_NVTX_RANGE_SCOPE_IF(condition, name) \
_CCCL_BEFORE_NVTX_RANGE_SCOPE(name) \
::cuda::std::optional<::nvtx3::v1::scoped_range_in<::cuda::detail::NVTXCCCLDomain>> __cuda_nvtx3_range; \
NV_IF_TARGET( \
NV_IS_HOST, ({ \
static const ::nvtx3::v1::registered_string_in<::cuda::detail::NVTXCCCLDomain> __cuda_nvtx3_func_name{name}; \
static const ::nvtx3::v1::event_attributes __cuda_nvtx3_func_attr{__cuda_nvtx3_func_name}; \
if (condition) \
{ \
__cuda_nvtx3_range.emplace(__cuda_nvtx3_func_attr); \
} \
}))
# else // ^^^ _CCCL_HOST_COMPILATION() ^^^ / vvv !_CCCL_HOST_COMPILATION() vvv
# define _CCCL_NVTX_RANGE_SCOPE_IF(condition, name)
# endif // ^^^ !_CCCL_HOST_COMPILATION() ^^^
Comment on lines +97 to +116
Copy link
Contributor

@bernhardmgruber bernhardmgruber Jan 5, 2026

Choose a reason for hiding this comment

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

Remark: this looks a bit weird, but I think it's correct. For nvc++, _CCCL_HOST_COMPILATION is defined for the combined host/device path, so we still need the NV_IF_TARGET inside.


# define _CCCL_NVTX_RANGE_SCOPE(name) _CCCL_NVTX_RANGE_SCOPE_IF(true, name)

Expand Down
41 changes: 41 additions & 0 deletions libcudacxx/include/cuda/std/__cccl/cuda_capabilities.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,4 +48,45 @@
# define _CCCL_PDL_TRIGGER_NEXT_LAUNCH()
#endif // _CCCL_HAS_PDL()

// Check whether the relocatable device code (RDC) is being generated.
#if defined(__CUDACC_RDC__) || defined(__CLANG_RDC__) || defined(_NVHPC_RDC)
# define _CCCL_HAS_RDC() 1
#else // ^^^ has RDC ^^^ / vvv no RDC vvv
# define _CCCL_HAS_RDC() 0
#endif // ^^^ no RDC ^^^

// Check whether extensible whole program is being compiled.
#if defined(__CUDACC_EWP__)
# define _CCCL_HAS_EWP() 1
#else // ^^^ has EWP ^^^ / vvv no EWP vvv
# define _CCCL_HAS_EWP() 0
#endif // ^^^ no EWP ^^^

// Control whether device runtime APIs can be used, because they require libcudadevrt to be linked. Defaults to true
// when RDC or EWP are enabled. Can be disabled by defining CCCL_DISABLE_DEVICE_RUNTIME.
#if (_CCCL_HAS_RDC() || _CCCL_HAS_EWP()) && !defined(CCCL_DISABLE_DEVICE_RUNTIME)
# define _CCCL_HAS_DEVICE_RUNTIME() 1
Copy link
Contributor

@fbusato fbusato Jan 5, 2026

Choose a reason for hiding this comment

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

relocatable device code does not strictly require libcudadevrt, but only for CDP

Copy link
Contributor Author

@davebayer davebayer Jan 5, 2026

Choose a reason for hiding this comment

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

I don't understand what do you mean. Using CUDA device runtime APIs require either RDC or EWP + link libcudadevrt. This macro returns 1 if device runtime (stuff from <cuda_device_runtime_api.h>) can be used by CCCL, that means if RDC or EWP are enabled. The user has the option to explicitly disable device runtime (in CCCL) by defining CCCL_DISABLE_DEVICE_RUNTIME. CDP calls cudaDeviceLaunch which is one of the device runtime APIs, thus CDP depends on _CCCL_HAS_DEVICE_RUNTIME()

Copy link
Contributor

@fbusato fbusato Jan 5, 2026

Choose a reason for hiding this comment

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

maybe I'm reading it in the wrong way.

  • The user has the option to explicitly disable device runtime (in CCCL) by defining CCCL_DISABLE_DEVICE_RUNTIME.
  • CDP depends on _CCCL_HAS_DEVICE_RUNTIME()

these are perfectly fine.

My issue is about RDC, relocatable device code. If we have RDC, this doesn't automatically translate in having _CCCL_HAS_DEVICE_RUNTIME. Am I missing something?

Copy link
Contributor Author

@davebayer davebayer Jan 5, 2026

Choose a reason for hiding this comment

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

It does not, because the user needn't to link libcudadevrt

#else // ^^^ has device runtime ^^^ / vvv no device runtime vvv
# define _CCCL_HAS_DEVICE_RUNTIME() 0
#endif // ^^^ no device runtime ^^^

// Some functions can be called from host or device code and launch kernels inside. Thus, they use CUDA Dynamic
// Parallelism (CDP) and require compiling with Relocatable Device Code (RDC) or extensible whole program (EWP) and link
// with device runtime library. CDP is unsupported with clang-cuda below 22.
// TODO(bgruber): remove CUB_DISABLE_CDP in CCCL 4.0
#if _CCCL_HAS_DEVICE_RUNTIME() && !defined(CCCL_DISABLE_CDP) && !defined(CUB_DISABLE_CDP) \
&& !_CCCL_CUDA_COMPILER(CLANG, <, 22)
// We have CDP, so host and device APIs can call kernels
# define _CCCL_HAS_CDP() 1
#else // ^^^ has CDP ^^^ / vvv no CDP vvv
// We don't have CDP, only host APIs can call kernels
# define _CCCL_HAS_CDP() 0
#endif // ^^^ no CDP ^^^

#if _CCCL_HAS_CDP()
# ifdef CUDA_FORCE_CDP1_IF_SUPPORTED
# error "CUDA Dynamic Parallelism 1 is no longer supported. Please undefine CUDA_FORCE_CDP1_IF_SUPPORTED."
# endif // CUDA_FORCE_CDP1_IF_SUPPORTED
#endif // _CCCL_HAS_CDP()

#endif // __CCCL_CUDA_CAPABILITIES
24 changes: 7 additions & 17 deletions libcudacxx/include/cuda/std/__cccl/visibility.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,7 @@
#endif // no system header

#include <cuda/std/__cccl/attributes.h>
#include <cuda/std/__cccl/cuda_capabilities.h>
#include <cuda/std/__cccl/execution_space.h>

// For unknown reasons, nvc++ need to selectively disable this warning
Expand Down Expand Up @@ -143,24 +144,13 @@
# define CCCL_DISABLE_CDP
#endif // _CCCL_DOXYGEN_INVOKED

// Some functions can be called from host or device code and launch kernels inside. Thus, they use CUDA Dynamic
// Parallelism (CDP) and require compiling with Relocatable Device Code (RDC).
// TODO(bgruber): remove CUB_DISABLE_CDP in CCCL 4.0
#if defined(__CUDACC_RDC__) && !defined(CCCL_DISABLE_CDP) && !defined(CUB_DISABLE_CDP)
# define _CCCL_HAS_RDC() 1
// We have RDC, so host and device APIs can call kernels
#if _CCCL_HAS_CDP()
// We have CDP, so host and device APIs can call kernels
# define _CCCL_CDP_API _CCCL_API
#else // defined(__CUDACC_RDC__) && !defined(CCCL_DISABLE_CDP) && !defined(CUB_DISABLE_CDP)
# define _CCCL_HAS_RDC() 0
// We don't have RDC, only host APIs can call kernels
# define _CCCL_CDP_API _CCCL_HOST_API
#endif // defined(__CUDACC_RDC__) && !defined(CCCL_DISABLE_CDP) && !defined(CUB_DISABLE_CDP)

#if _CCCL_HAS_RDC()
# ifdef CUDA_FORCE_CDP1_IF_SUPPORTED
# error "CUDA Dynamic Parallelism 1 is no longer supported. Please undefine CUDA_FORCE_CDP1_IF_SUPPORTED."
# endif // CUDA_FORCE_CDP1_IF_SUPPORTED
#endif // _CCCL_HAS_RDC()
#else // ^^^ _CCCL_HAS_CDP() ^^^ / vvv !_CCCL_HAS_CDP() vvv
// We don't have CDP, only host APIs can call kernels
# define _CCCL_CDP_API _CCCL_HOST_API
#endif // ^^^ !_CCCL_HAS_CDP() ^^^

//! _LIBCUDACXX_HIDE_FROM_ABI is for backwards compatibility for external projects.
//! _CCCL_API and its variants are the preferred way to declare functions
Expand Down
Loading