Making python bindings for a project with CUDA kernels.

[pmichels19]

Hi everyone! I am currently working on a project that requires me to build python bindings. For performance reasons I decided that nanobind would be a great option for this. However, I am running into a bit of an issue with compiling the CUDA kernels that are included in the project.

My CMakeLists.txt looks like this:

cmake_minimum_required(VERSION 3.15...3.27)
project(my_project LANGUAGES CXX CUDA)
enable_language(CUDA)

# Warn if the user invokes CMake directly
if (NOT SKBUILD)
  message(WARNING <standard warning is here>)
endif()

# Try to import all Python components potentially needed by nanobind
find_package(Python 3.11
    REQUIRED COMPONENTS Interpreter Development.Module
    OPTIONAL_COMPONENTS Development.SABIModule
)
find_package(nanobind CONFIG REQUIRED)
find_package(CUDAToolkit REQUIRED)

set(MY_PROJECT_SOURCES
    # Source files go here
    src/pch.cpp
    src/types.cpp
    src/logging.cpp
    src/display.cpp
    src/main.cpp
)

set(MY_PROJECT_INCLUDES
    # Include folders go here
    Dependencies/x64/argparse/include
    Dependencies/x64/CImg/include
    Dependencies/x64/cuFFTDx/include
    Dependencies/x64/fftw/include
    Dependencies/x64/glm/include
    Dependencies/x64/libpng/include
    Dependencies/x64/zlib/include
    ${CUDAToolkit_INCLUDE_DIRS}
)

set(MY_PROJECT_DEPENDENCIES
    "${CMAKE_SOURCE_DIR}/Dependencies/x64/fftw/lib/libfftw3-3.lib"
    "${CMAKE_SOURCE_DIR}/Dependencies/x64/fftw/lib/libfftw3f-3.lib"
    "${CMAKE_SOURCE_DIR}/Dependencies/x64/libpng/lib/libpng16.lib"
    "${CMAKE_SOURCE_DIR}/Dependencies/x64/libpng/lib/libpng16_static.lib"
    "${CMAKE_SOURCE_DIR}/Dependencies/x64/zlib/lib/zlib.lib"
    "${CMAKE_SOURCE_DIR}/Dependencies/x64/zlib/lib/zlibstatic.lib"
    CUDA::cufft
    CUDA::cudart
    #${CUDAToolkit_LIBRARY_DIR}
)

add_library(MY_PROJECT_CUDA_KERNELS STATIC my-project/src/kernels.cu)
set_target_properties(MY_PROJECT_CUDA_KERNELS
    PROPERTIES
    POSITION_INDEPENDENT_CODE ON
    CUDA_VISIBILITY_PRESET "hidden"
    CUDA_SEPARABLE_COMPILATION ON
)

target_link_libraries(MY_PROJECT_CUDA_KERNELS PRIVATE ${MY_PROJECT_DEPENDENCIES} ${Python_LIBRARIES})
target_include_directories(MY_PROJECT_CUDA_KERNELS PRIVATE ${MY_PROJECT_INCLUDES})

add_library(MY_PROJECT_LIBRARY STATIC ${MY_PROJECT_SOURCES})
set_target_properties(MY_PROJECT_LIBRARY
    PROPERTIES
    CXX_STANDARD 17
    POSITION_INDEPENDENT_CODE ON
    CUDA_VISIBILITY_PRESET "hidden"
    CUDA_SEPARABLE_COMPILATION ON
)

target_link_libraries(MY_PROJECT_LIBRARY PRIVATE ${MY_PROJECT_DEPENDENCIES} ${MY_PROJECT_CUDA_KERNELS})
target_include_directories(MY_PROJECT_LIBRARY PRIVATE ${MY_PROJECT_INCLUDES})

nanobind_add_module(my_project_ext
    STABLE_ABI
    NB_STATIC
    src/my_project_ext.cpp
    ${MY_PROJECT_SOURCES}
    src/kernels.cu
)

target_link_libraries(my_project_ext PRIVATE ${MY_PROJECT_DEPENDENCIES} ${MY_PROJECT_LIBRARY})
target_include_directories(my_project _ext PRIVATE ${MY_PROJECT_INCLUDES})

install(TARGETS my_project_ext LIBRARY DESTINATION my_project)

In my_project_ext.cpp I have a very generic test function as well as a constructor for the main class:

#include "pch.h"

#include <nanobind/nanobind.h>

namespace nb = nanobind;
using namespace nb::literals;

int add( int i1, int i2 ) {
    return i1 + i2;
}

NB_MODULE(my_project_ext, m) {
    m.def( "add", &add, "a"_a, "b"_a = 1,
           "This function adds two numbers and increments if only one is provided." );

    nb::class_<Main>(m, "Main").def(nb::init<int, int, int, const std::string&>());
}

Unfortunately, I am getting quite an extensive error that I am not sure how to fix:

"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3\bin\nvcc.exe"  --use-local-env -ccbin "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.38.33130\bin\HostX64\x64" -x cu   -I"C:\Users\<user>\Documents\my-project\Dependencies\x64\argparse\include" -I"C:\Users\<user>\Documents\my-project\Dependencies\x64\CImg\include" -I"C:\Users\<user>\Documents\my-project\Dependencies\x64\cuFFTDx\include" -I"C:\Users\<user>\Documents\my-project\Dependencies\x64\fftw\include" -I"C:\Users\<user>\Documents\my-project\Dependencies\x64\glm\include" -I"C:\Users\<user>\Documents\my-project\Dependencies\x64\libpng\include" -I"C:\Users\<user>\Documents\my-project\Dependencies\x64\zlib\include" -I"C:\Users\<user>\AppData\Local\Programs\Python\Python311\Lib\site-packages\nanobind\include" -I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3\include" -IC:\Users\<user>\AppData\Local\Programs\Python\Python311\Include -I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3\include"     --keep-dir x64\Release  -maxrregcount=0   --machine 64 --compile -cudart static -std=c++17 --generate-code=arch=compute_52,code=[compute_52,sm_52] /bigobj /MP /Os -Xcompiler="/EHsc -Ob2"   -D_WINDOWS -DNDEBUG -D"CMAKE_INTDIR=\"Release\"" -Dmy_project_ext_EXPORTS -D_WINDLL -D_MBCS -DWIN32 -D_WINDOWS -DNDEBUG -D"CMAKE_INTDIR=\"Release\"" -Dmy_project_ext_EXPORTS -Xcompiler "/EHsc /W1 /nologo /O2 /FS   /MD " -Xcompiler "/Fdmy_project_ext.dir\Release\vc143.pdb" -o my_project_ext.dir\Release\kernels.obj "C:\Users\<user>\Documents\my-project\src\kernels.cu"
        nvcc fatal   : A single input file is required for a non-link phase when an outputfile is specified

Is there anyone who has made python bindings with a CUDA kernel before who is capable of providing some insight or maybe sharing how they did it? I have been stuck on this problem for...longer than I would like to admit really 😅

[wjakob]

Could it be something like your projected containing spaces, and one of the include directives not having quotes?

[pmichels19]

Alright, I'll give that a go! As for the quotes: it doesn't actually seem to matter 😅 still thanks for the quick responses though!

[pmichels19]

After some trying I've been able to at least nail down the problem a little more: it seems to be the cufft library.
My my_project_ext.cpp looks like this at the moment:

#include <cufft.h>

#include <nanobind/nanobind.h>

namespace nb = nanobind;

using namespace nb::literals;

int add( int i1, int i2 ) {
    cufftPlan3d(nullptr, 0, 0, 0, CUFFT_C2C);
    return i1 + i2;
}

NB_MODULE(my_project_ext, m) {
    m.def( "add", &add, "a"_a, "b"_a = 1,
           "This function adds two numbers and increments if only one is provided." );
}

With a CMakeLists.txt like this:

cmake_minimum_required(VERSION 3.15...3.27)
project(my_project LANGUAGES CXX CUDA)
enable_language(CUDA)

# Warn if the user invokes CMake directly
if (NOT SKBUILD)
    message(WARNING <standard warning is here>)
endif()

# Try to import all Python components potentially needed by nanobind
find_package(Python 3.11 REQUIRED COMPONENTS Interpreter Development.Module OPTIONAL_COMPONENTS Development.SABIModule)
find_package(nanobind CONFIG REQUIRED)
find_package(CUDAToolkit REQUIRED)

set(MY_PROJECT_INCLUDES
    # Include folders go here
    ${CUDAToolkit_INCLUDE_DIRS}
)

set(MY_PROJECT_DEPENDENCIES
    "${CUDAToolkit_LIBRARY_DIR}/cufft.lib"
)

nanobind_add_module(my_project_ext
    STABLE_ABI
    NB_STATIC
    src/my_project_ext.cpp
)

target_link_libraries(my_project_ext PRIVATE ${MY_PROJECT_DEPENDENCIES})
target_include_directories(my_project_ext PRIVATE ${MY_PROJECT_INCLUDES})

install(TARGETS my_project_ext LIBRARY DESTINATION my_project)

This still gives the same DLL import error.

I don't know if this is still an issue that should be kept on this GitHub by the way. I don't know if it is maybe a good idea to post this issue on the NVIDIA rather than here.

[wjakob]

I've read at some point there is some way to get DLL loading status messages on windows that could help to further pinpoint the issue. But I don't remember the details, perhaps someone else here knows?

[wjakob]

You need Process Monitor: https://learn.microsoft.com/en-us/sysinternals/downloads/procmon

[pmichels19]

Ah that actually got me a lot further! Apparently it was looking for the msvcp140, cufft and cudart dlls and failing to find them in the folders where it was looking. I'll see if I can fix that somehow, thanks for the fast replies and the help!

[matthewdcong]

nvcc doesn't support the /bigobj, /MP/, and /Os compiler options and is interpreting those as paths. Are you on the latest ToT master nanobind? PR #383 was sufficient to fix the issue for me.

[pmichels19]

Ah that has at least fixed the compiling issue! I do get strange behavior now when trying to import the module 🤔 seems like the CUDA libraries might not be linked properly (I think? Looking at the dependencies the cudart has an incorrect checksum and cufft isn't even there...)