syoyo / cuew

CUDA Extension Wrangler

syoyo's modification to CUEW.

Currently tested on Ubuntu 20.04 x86-64 and Windows 10 64bit.

Changes compared to original version

• Support CUDA 11.5
  • cuGetProcAddress is not supported.
• Support CUDNN 8.0.3

Supported API

• cuda.h
• cudnn.h
• nvrtc.h
• cudaGL.h
  • Use cuda_gl_interop.h instead

Buld tests

Linux

$ mkdir build
$ cd build
$ cmake ..
$ make

Windows

> vcsetup-2019.bat

Then open .sln file in build folder.

Add cuew to your project

Simply copy include/cuew.h and src/cuew.c to your project.

Generate cuew(for developer)

You need python3. Install pycparser.

$ python -m pip install pycparser

$ cd auto
# Edit header path in cuew_ge.py if required
$ ./cuew_gen.sh

GL header

If you encounter the parse error on GL header, there is a work around.

Copy cudaGL.h to mycudaGL.h and remove <GL/gl.h> include line. and use mycudaGL.h as an input .h file. Content of <gl.h> is not used when generating cuew header/source.

Known issues

• Combining with Address Sanitizer(-fsanitizer=address) won't work
  • calling CUDA API results in undefined behavior or seg faults
  • google/sanitizers#629
• CUEW does not report warning when using deprecated CUDA API
• CUDA_API_PER_THREAD_DEFAULT_STREAM is not supported.

TODO

• Test on MSVC.
  • clang-cl works
• Test CUDA-GL interop API
• Test cuDNN API call.
• Find a way to co-exist with Address Sanitizer
• CUDA_API_PER_THREAD_DEFAULT_STREAM support

=================

The CUDA Extension Wrangler Library (CUEW) is a cross-platform open-source C/C++ extension loading library. CUEW provides efficient run-time mechanisms for determining which CUDA functions and extensions extensions are supported on the target platform.

CUDA core and extension functionality is exposed in a single header file. CUEW has been tested on a variety of operating systems, including Windows, Linux, Mac OS X.

LICENSE

CUEW library is released under the Apache 2.0 license.