Compatibility with CUDA-Toolkit 13.0.0

[lujobi]

It seems that this project is not (yet) compatible with the newest version of the CUDA Toolkit (Version 13.0).

Please find attached our investigation.
Any help would be greatly appreciated!

Thank you in advance.

---

Investigation

Simply building it with the documented command yields the following error:

$ cmake -B build -S . -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=bin
...
CMake Error at /usr/share/cmake-3.28/Modules/FindPackageHandleStandardArgs.cmake:230 (message):
  Could NOT find thrust (missing: THRUST_INCLUDE_DIR) (Required is at least
  version "1.13.1")
Call Stack (most recent call first):
  /usr/share/cmake-3.28/Modules/FindPackageHandleStandardArgs.cmake:600 (_FPHSA_FAILURE_MESSAGE)
  cmake/Findthrust.cmake:50 (find_package_handle_standard_args)
  src/stdgpu/CMakeLists.txt:7 (find_package)

After some investigating, it seems that the include directory $CUDAToolkit_INCLUDE_DIRS has been restructured. As such thrust has been moved into a cccl subdir.

After manually setting CUB_INCLUDE_DIR, LIBCUDACXX_INCLUDE_DIR and THRUST_INCLUDE_DIR the configure step succeeds. However the build command fails with an error that originates from within the c++ std-lib.

Configure Command:

cmake -B build -S . \
        -DCMAKE_BUILD_TYPE=Release \
        -DCMAKE_INSTALL_PREFIX=bin \
        -DCUB_INCLUDE_DIR="/usr/local/cuda/targets/sbsa-linux/include/cccl" \
        -DLIBCUDACXX_INCLUDE_DIR="/usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/detail/libcxx/include" \
        -DTHRUST_INCLUDE_DIR="/usr/local/cuda/targets/sbsa-linux/include/cccl"

Build Output:

cmake --build build --config Release --parallel 8

[ 66%] Building CXX object src/stdgpu/CMakeFiles/stdgpu.dir/cuda/impl/memory.cpp.o
In file included from /usr/include/c++/13/iterator:65,
                 from /usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/__type_traits/is_primary_template.h:33,
                 from /usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/__iterator/incrementable_traits.h:32,
                 from /usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/__iterator/concepts.h:40,
                 from /usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/__algorithm/ranges_iterator_concept.h:23,
                 from /usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/__algorithm/iterator_operations.h:24,
                 from /usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/__algorithm/lower_bound.h:25,
                 from /usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/__algorithm/binary_search.h:25,
                 from /usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/detail/libcxx/include/algorithm:651,
                 from /home/user/Open3D/build/stdgpu/src/ext_stdgpu/src/stdgpu/cuda/impl/device.cpp:18:
/usr/include/c++/13/bits/stream_iterator.h:55:31: error: ‘char_traits’ does not name a type
   55 |            typename _Traits = char_traits<_CharT>, typename _Dist = ptrdiff_t>
      |                                           ^~~~~~~~~~~
/usr/include/c++/13/bits/stream_iterator.h:55:42: error: expected ‘>’ before ‘<’ token
   55 |            typename _Traits = char_traits<_CharT>, typename _Dist = ptrdiff_t>
      |                                          ^
/usr/include/c++/13/bits/stream_iterator.h:57:48: error: ‘_Dist’ was not declared in this scope
   57 |     : public iterator<input_iterator_tag, _Tp, _Dist, const _Tp*, const _Tp&>
      |                                                ^~~~~
/usr/include/c++/13/bits/stream_iterator.h:57:77: error: template argument 3 is invalid
   57 |     : public iterator<input_iterator_tag, _Tp, _Dist, const _Tp*, const _Tp&>
      |                                                                             ^
/usr/include/c++/13/bits/stream_iterator.h:62:15: error: ‘basic_istream’ does not name a type
   62 |       typedef basic_istream<_CharT, _Traits> istream_type;
      |               ^~~~~~~~~~~~~
/usr/include/c++/13/bits/stream_iterator.h:65:7: error: ‘istream_type’ does not name a type
   65 |       istream_type*     _M_stream;
      |       ^~~~~~~~~~~~
/usr/include/c++/13/bits/stream_iterator.h:79:36: error: expected ‘)’ before ‘&’ token
   79 |       istream_iterator(istream_type& __s)
      |                       ~            ^
      |                                    )
/usr/include/c++/13/bits/stream_iterator.h: In constructor ‘constexpr std::istream_iterator<_Tp, _CharT, _Traits>::istream_iterator()’:
/usr/include/c++/13/bits/stream_iterator.h:76:9: error: class ‘std::istream_iterator<_Tp, _CharT, _Traits>’ does not have any field named ‘_M_stream’
   76 |       : _M_stream(0), _M_value(), _M_ok(false) {}
      |        ^~~~~~~~~

...
===================== OUTPUT CROPPED BY ~18000 Lines =====================
...

/home/user/Open3D/build/stdgpu/src/ext_stdgpu/src/stdgpu/../stdgpu/impl/limits_detail.h: In static member function ‘static constexpr long double stdgpu::numeric_limits<long double>::infinity()’:
/home/user/Open3D/build/stdgpu/src/ext_stdgpu/src/stdgpu/../stdgpu/impl/limits_detail.h:720:12: error: ‘HUGE_VALL’ was not declared in this scope
  720 |     return HUGE_VALL;
      |            ^~~~~~~~~
make[5]: *** [src/stdgpu/CMakeFiles/stdgpu.dir/build.make:90: src/stdgpu/CMakeFiles/stdgpu.dir/impl/iterator.cpp.o] Error 1
make[4]: *** [CMakeFiles/Makefile2:157: src/stdgpu/CMakeFiles/stdgpu.dir/all] Error 2
make[3]: *** [Makefile:136: all] Error 2
make[2]: *** [CMakeFiles/ext_stdgpu.dir/build.make:86: stdgpu/src/ext_stdgpu-stamp/ext_stdgpu-build] Error 2
make[1]: *** [CMakeFiles/Makefile2:1827: CMakeFiles/ext_stdgpu.dir/all] Error 2

I can provide a full log if this seems helpful. However I fear that the problem lies somewhere else within our config.

System:

• OS: Ubuntu 24.04
• Target Machine: Nvida DGX Spark (ARMv9)
• Compiler: GCC 13
• Backend: CUDA (CUDA Toolkit 13.0)
• Library version: master

[stotko]

Thanks for the report. CUDA 13 indeed includes several breaking changes that likely affect stdgpu. I will have a closer look at their migration guide when I get a chance. We might need to fix the support at several locations in the code.

1. According to your investigation, we definitvely need to update the detection of thrust in CMake. Regarding your workaround, this line

   -DLIBCUDACXX_INCLUDE_DIR="/usr/local/cuda/targets/sbsa-linux/include/cccl/cuda/std/detail/libcxx/include" \

   looks odd to me. Does the CMake configure step still work if you use the same path as for the other variables, i.e. using -DLIBCUDACXX_INCLUDE_DIR="/usr/local/cuda/targets/sbsa-linux/include/cccl" \ instead?

   Alternatively, could you try whether the following change in the file cmake/cuda/determine_thrust_paths.cmake (without specifying any of the *_INCLUDE_DIR variables) would fix the detection of thrust as well?

   -    set(${STDGPU_OUTPUT_THRUST_PATHS} "${CUDAToolkit_INCLUDE_DIRS}")
   +    set(${STDGPU_OUTPUT_THRUST_PATHS} "${CUDAToolkit_INCLUDE_DIRS}/cccl")

2. The build errors indicate that there is an issue with libcudacxx. This could potentially be caused by the above point. After having tried out the suggested changes, could you provide me with the full log of both the CMake configuring and building steps?

[lujobi]

Thank you for the quick reply.

Indeed the suggested change (set(${STDGPU_OUTPUT_THRUST_PATHS} "${CUDAToolkit_INCLUDE_DIRS}/cccl")) showed some promising effects.

The next error is easily fixed: cudaDeviceProp does not have a field named clockRate anymore. As such we removed line 79f in src/stdgpu/cuda/impl/device.cpp.

The examples do not build yet (see log below), but can be disabled using -DSTDGPU_BUILD_EXAMPLES=OFF.

Build commands are apart from this the very same as on the website:

cmake -B build -S . -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=bin
cmake --build build --config Release --parallel 8

buildlog.txt

[stotko]

I was able to get CUDA 13 running with a few minor adaptions, see #466. Could you check if this fixes the errors for you as well? Note that you need a newer CMake version for this to work, i.e. version 3.31.9 or higher, as CMake also had to adapt to the changed include layout (see https://gitlab.kitware.com/cmake/cmake/-/merge_requests/11079).

[lujobi]

Thank you very much for the quick fix, will take a look at it tomorrow in the morning and keep you posted.

[cwlm]

Still meeting this problem compiling, on the cuda_13 branch

CMake Error at /usr/share/cmake-3.28/Modules/FindPackageHandleStandardArgs.cmake:230 (message):

  Could NOT find thrust (missing: THRUST_INCLUDE_DIR) (Required is at least

  version "1.13.1")

Call Stack (most recent call first):

  /usr/share/cmake-3.28/Modules/FindPackageHandleStandardArgs.cmake:600 (_FPHSA_FAILURE_MESSAGE)

  cmake/Findthrust.cmake:50 (find_package_handle_standard_args)

  src/stdgpu/CMakeLists.txt:7 (find_package)

However, this seems to be resolved by

-    set(${STDGPU_OUTPUT_THRUST_PATHS} "${CUDAToolkit_INCLUDE_DIRS}")
+    set(${STDGPU_OUTPUT_THRUST_PATHS} "${CUDAToolkit_INCLUDE_DIRS}/cccl")

System:

• OS: Ubuntu 24.04
• Compiler: GCC 13
• Backend: CUDA (CUDA Toolkit 13.0)
• Library version: cuda_13

[lujobi]

The patched version works fine if the CMake version is recent enough.

Thank you very much @stotko 👍

[stotko]

Thanks for testing. Given that older CMake versions are quite likely to be in use, I extended the detection logic with a cleaner solution than the single-line workaround. If you find any further issues with the patch, please let me know. I will give this a few more days and then merge it next week.