ctypes_approach_to_accessing_cuda_runtime_api.py

import ctypes as C
from ctypes.util import find_library

cuda = C.cdll.LoadLibrary(find_library("cudart"))

dev_ct = C.c_int()
dev_ct_addr = C.addressof(dev_ct)
dev_ct_ptr = C.cast(dev_ct_addr, C.POINTER(C.c_int))
error_t = cuda.cudaGetDeviceCount(dev_ct_ptr)

print("Error code (no error if zero): %d" % error_t)
print("Device count: %d" % dev_ct.value)

Tested on Linux. Sample output:

Error code (no error if zero): 0
Device count: 2

Should probably work on the Mac (libcudart.dylib) and on Windows (libcudart.dll) as well (find_library looks up the shared lib regardless of platform-specific extension).

Neat -- so this is your preferred way of direct CUDA API interaction via Python?

No, not really. There's PyCUDA, but I wanted to find a simple demo that didn't rely on PyCUDA. For example, DL frameworks don't like to include PyCUDA for various reasons (or CuPy for that matter, which is another alternative). Also, you can always use SWIG, Ctypes are way too low level. But even though SWIG is prettier, SWIG builds are messy. Ctypes are nice only in the sense that if you just want to do something trivial in a few lines, it will be simpler than dealing with SWIG. But for large C/C++=Python bindings, I'd use SWIG, Pybind11, etc. Especially for C++ specifically, because PyBind11 has extensive support for C++11.