Build warnings in <cuda/std/barrier>

[sleeepyjack]

Having build warnings when CMAKE_CUDA_ARCHITECTURES=75:

[ 98%] Linking CXX executable DYNAMIC_MAP_TEST
[ 98%] Built target DYNAMIC_MAP_TEST
/home/yunsongw/miniconda3/include/rapids/libcudacxx/cuda/std/barrier: In function ‘void cuda::__4::init(cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>*, ptrdiff_t, cuda::std::__4::__empty_completion)’:
/home/yunsongw/miniconda3/include/rapids/libcudacxx/cuda/std/barrier:158:155: warning: unused parameter ‘__completion’ [-Wunused-parameter]
  158 |     friend void init(barrier * __b, std::ptrdiff_t __expected, std::__empty_completion __completion = std::__empty_completion()) {
      |                                                                                                                        ~~~~~~~~~~~                        ^                                                  
/home/yunsongw/miniconda3/include/rapids/libcudacxx/cuda/std/barrier: In function ‘void cuda::__4::init(cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>*, ptrdiff_t, cuda::std::__4::__empty_completion)’:
/home/yunsongw/miniconda3/include/rapids/libcudacxx/cuda/std/barrier:158:155: warning: unused parameter ‘__completion’ [-Wunused-parameter]
  158 |     friend void init(barrier * __b, std::ptrdiff_t __expected, std::__empty_completion __completion = std::__empty_completion()) {
      |                                                                                                                        ~~~~~~~~~~~                        ^                                                  
/home/yunsongw/miniconda3/include/rapids/libcudacxx/cuda/std/barrier: In function ‘void cuda::__4::init(cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>*, ptrdiff_t, cuda::std::__4::__empty_completion)’:
/home/yunsongw/miniconda3/include/rapids/libcudacxx/cuda/std/barrier:158:155: warning: unused parameter ‘__completion’ [-Wunused-parameter]
  158 |     friend void init(barrier * __b, std::ptrdiff_t __expected, std::__empty_completion __completion = std::__empty_completion()) {
      |                                                                                                                        ~~~~~~~~~~~                        ^                                                  

Worthing reporting to libcudacxx.

Otherwise, LGTM.

Originally posted by @PointKernel in NVIDIA/cuCollections#201 (review)

[PointKernel]

Any updates on this issue?

I can put up a PR removing the argument __completion since it seems never to be used but not sure if it's the expected solution.
https://github.com/NVIDIA/libcudacxx/blob/31eadca674f62dca7535734c6aae8a1f082b2259/include/cuda/std/barrier#L158-L171

[jrhemstad]

@davedsth can you look at this?

[wmaxey]

Unsure if we have internal changes pending that might fix this. David might know more.

[davedsth]

Internal changes are coming but we haven't addressed the completion arguments
Removing it until we have API + test sounds reasonable

[PointKernel]

Updates: unused parameters are growing

/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier: In function ‘void cuda::__4::init(cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>*, ptrdiff_t, cuda::std::__4::__empty_completion)’:
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier:143:155: warning: unused parameter ‘__completion’ [-Wunused-parameter]
  143 |     friend void init(barrier * __b, std::ptrdiff_t __expected, std::__empty_completion __completion = std::__empty_completion()) {
      |                                                                                                                        ~~~~~~~~~~~                        ^                                                  
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier: In member function ‘bool cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>::__test_wait_sm_80(cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>::arrival_token) const’:
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier:247:47: warning: unused parameter ‘__token’ [-Wunused-parameter]
  247 |     inline bool __test_wait_sm_80(arrival_token __token) const {
      |                                 ~~~~~~~~~~~~~~^~~~~~~
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier: In member function ‘bool cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>::__test_wait_parity_sm_80(bool) const’:
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier:357:36: warning: unused parameter ‘__phase_parity’ [-Wunused-parameter]
  357 |     inline bool __test_wait_parity_sm_80(bool __phase_parity) const {

It was only __completion before, now I can see __token and __phase_parity as well.

[wmaxey]

@PointKernel Which version of libcu++ is this? Can you provide info on the compiler, flags, and how barrier is instatiated?

[PointKernel]

libcu++ version: branch/1.9.1
CXX compiler: GNU 10.3.0
CUDA compiler: NVCC 11.8.89

compiler flag and build log:

cd /home/yunsongw/Work/cuCollections/build/examples && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_CUDA -DTHRUST_HOST_SYSTEM=THRUST_HOST_SYSTEM_CPP -I/home/yunsongw/Work/cuCollections/examples -I/home/yunsongw/Work/cuCollections/include -I/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include -I/home/yunsongw/Work/cuCollections/build/_deps/thrust-src -I/home/yunsongw/Work/cuCollections/build/_deps/thrust-src/dependencies/cub -isystem=/usr/local/cuda/include -O3 -DNDEBUG --generate-code=arch=compute_75,code=[compute_75,sm_75] --compiler-options=-Wall --compiler-options=-Wextra --expt-extended-lambda --expt-relaxed-constexpr -Xcompiler -Wno-subobject-linkage -std=c++17 -MD -MT examples/CMakeFiles/STATIC_SET_HOST_BULK_EXAMPLE.dir/static_set/host_bulk_example.cu.o -MF CMakeFiles/STATIC_SET_HOST_BULK_EXAMPLE.dir/static_set/host_bulk_example.cu.o.d -x cu -c /home/yunsongw/Work/cuCollections/examples/static_set/host_bulk_example.cu -o CMakeFiles/STATIC_SET_HOST_BULK_EXAMPLE.dir/static_set/host_bulk_example.cu.o
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier: In function ‘void cuda::__4::init(cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>*, ptrdiff_t, cuda::std::__4::__empty_completion)’:
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier:143:155: warning: unused parameter ‘__completion’ [-Wunused-parameter]
  143 |     friend void init(barrier * __b, std::ptrdiff_t __expected, std::__empty_completion __completion = std::__empty_completion()) {
      |                                                                                                                        ~~~~~~~~~~~                        ^                                                  
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier: In member function ‘bool cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>::__test_wait_sm_80(cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>::arrival_token) const’:
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier:247:47: warning: unused parameter ‘__token’ [-Wunused-parameter]
  247 |     inline bool __test_wait_sm_80(arrival_token __token) const {
      |                                 ~~~~~~~~~~~~~~^~~~~~~
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier: In member function ‘bool cuda::__4::barrier<cuda::std::__4::__detail::thread_scope_block>::__test_wait_parity_sm_80(bool) const’:
/home/yunsongw/Work/cuCollections/build/_deps/libcudacxx-src/lib/cmake/libcudacxx/../../../include/cuda/std/barrier:357:36: warning: unused parameter ‘__phase_parity’ [-Wunused-parameter]
  357 |     inline bool __test_wait_parity_sm_80(bool __phase_parity) const {
      |                               ~~~~~^~~~~~~~~~~~~~

The warnings come out even we just include the header but not use it, here is the godbolt repro: https://godbolt.org/z/dzq6orsW3

[wmaxey]

Perfect! Thank you for the repro!

[miscco]

@PointKernel could you please verify that the issue is gone with ToT?

[PointKernel]

I don't see these warnings anymore. Thanks! @miscco @wmaxey

[PointKernel]

Somehow this issue came out again with recent CTK, e.g. https://godbolt.org/z/fGEocq58T.

[miscco]

That looks like a old library version. There is no line 143 in <cuda/std/barrier>

@wmaxey those changes should be in 12.1 shouldnt they?

[PointKernel]

To provide more information on this issue. If you check our CI run details,

• Building for sm60 is fine: https://github.com/NVIDIA/cuCollections/actions/runs/5941588210/job/16112766038?pr=356
• Building for sm70 will emit the warnings I mentioned above: https://github.com/NVIDIA/cuCollections/actions/runs/5941588210/job/16112767788?pr=356

And we are fetching the same version of libcudacxx in both cases:

-- CPM: adding package [email protected] (branch/1.9.1)
-- Found libcudacxx: /home/coder/cuCollections/build/local/_deps/libcudacxx-src/lib/cmake/libcudacxx/libcudacxx-config.cmake (found version "1.9.1.0")