Enhance RDC detection and add _CCCL_HAS_DEVICE_RUNTIME() macro
#7049
Conversation
|
/ok to test 5c0cc34 |
davebayer
force-pushed
the
allow_rdc_with_clang
branch
from
5c0cc34 to
39fd084
Compare
davebayer
changed the title
davebayer
force-pushed
the
allow_rdc_with_clang
branch
from
39fd084 to
cbeaaf1
Compare
davebayer
changed the title
davebayer
changed the title
_CCCL_HAS_DEVICE_RUNTIME() macro
davebayer
force-pushed
the
allow_rdc_with_clang
branch
from
cbeaaf1 to
e863934
Compare
This comment has been minimized.
This comment has been minimized.
davebayer
force-pushed
the
allow_rdc_with_clang
branch
3 times, most recently
from
d7696fc to
c8019bf
Compare
davebayer
force-pushed
the
allow_rdc_with_clang
branch
from
c8019bf to
3780d10
Compare
|
/ok to test 3780d10 |
miscco
left a comment
miscco
left a comment
There was a problem hiding this comment.
Looks good to me but would like some thoughts from @bernhardmgruber and @elstehle
This comment has been minimized.
This comment has been minimized.
| // 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 |
There was a problem hiding this comment.
relocatable device code does not strictly require libcudadevrt, but only for CDP
There was a problem hiding this comment.
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()
There was a problem hiding this comment.
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?
There was a problem hiding this comment.
It does not, because the user needn't to link libcudadevrt
| # 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() ^^^ |
There was a problem hiding this comment.
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.
🥳 CI Workflow Results🟩 Finished in 8h 56m: Pass: 100%/136 | Total: 6d 08h | Max: 5h 24m | Hits: 68%/271961See results here. |
4 participants
Currently, we are mixing several things together. This PR is trying to put things in order and extend RDC support.
Changes:
_CCCL_HAS_RDC()macro just reflects whether RDC is being generated. This macro will no longer be disabled by definingCCCL_DISABLE_CDPorCUB_DISABLE_CDP._CCCL_HAS_EWP()macro expands to1if the-ewp(extensive whole program) option was passed to the compiler._CCCL_HAS_DEVICE_RUNTIME()macro that expands to1if device runtime (cuda_device_runtime_api.h) stuff can be used by CCCL. This option requires RDC or EWP and can be disabled by the user by definingCCCL_DISABLE_DEVICE_RUNTIME._CCCL_HAS_CDP()macro that expands to1if CUDA dynamic parallelism is enabled. Requires_CCCL_HAS_DEVICE_RUNTIME()to be1