So as confirmed in the answer and comment thread with @NAmorim, there are no accessible Python bindings to OpenCV's various CUDA modules.
I was able to get around this restriction by using Cython to gain access to the CUDA functions I needed and implementing the necessary logic to convert my Python objects (mainly NumPy arrays) to OpenCV C/C++ objects and back.
Working Code
I first wrote a Cython definition file, GpuWrapper.pxd. The purpose of this file is to reference external C/C++ classes and methods, such as the CUDA methods I am interested in.
from libcpp cimport bool
from cpython.ref cimport PyObject
# References PyObject to OpenCV object conversion code borrowed from OpenCV's own conversion file, cv2.cpp
cdef extern from 'pyopencv_converter.cpp':
cdef PyObject* pyopencv_from(const Mat& m)
cdef bool pyopencv_to(PyObject* o, Mat& m)
cdef extern from 'opencv2/imgproc.hpp' namespace 'cv':
cdef enum InterpolationFlags:
INTER_NEAREST = 0
cdef enum ColorConversionCodes:
COLOR_BGR2GRAY
cdef extern from 'opencv2/core/core.hpp':
cdef int CV_8UC1
cdef int CV_32FC1
cdef extern from 'opencv2/core/core.hpp' namespace 'cv':
cdef cppclass Size_[T]:
Size_() except +
Size_(T width, T height) except +
T width
T height
ctypedef Size_[int] Size2i
ctypedef Size2i Size
cdef cppclass Scalar[T]:
Scalar() except +
Scalar(T v0) except +
cdef extern from 'opencv2/core/core.hpp' namespace 'cv':
cdef cppclass Mat:
Mat() except +
void create(int, int, int) except +
void* data
int rows
int cols
cdef extern from 'opencv2/core/cuda.hpp' namespace 'cv::cuda':
cdef cppclass GpuMat:
GpuMat() except +
void upload(Mat arr) except +
void download(Mat dst) const
cdef cppclass Stream:
Stream() except +
cdef extern from 'opencv2/cudawarping.hpp' namespace 'cv::cuda':
cdef void warpPerspective(GpuMat src, GpuMat dst, Mat M, Size dsize, int flags, int borderMode, Scalar borderValue, Stream& stream)
# Function using default values
cdef void warpPerspective(GpuMat src, GpuMat dst, Mat M, Size dsize, int flags)
We also need the ability to convert Python objects to OpenCV objects. I copied the first couple hundred lines from OpenCV's modules/python/src2/cv2.cpp. You can find that code below in the appendix.
We can finally write our Cython wrapper methods to call OpenCV's CUDA functions! This is part of the Cython implementation file, GpuWrapper.pyx.
import numpy as np # Import Python functions, attributes, submodules of numpy
cimport numpy as np # Import numpy C/C++ API
def cudaWarpPerspectiveWrapper(np.ndarray[np.uint8_t, ndim=2] _src,
np.ndarray[np.float32_t, ndim=2] _M,
_size_tuple,
int _flags=INTER_NEAREST):
# Create GPU/device InputArray for src
cdef Mat src_mat
cdef GpuMat src_gpu
pyopencv_to(<PyObject*> _src, src_mat)
src_gpu.upload(src_mat)
# Create CPU/host InputArray for M
cdef Mat M_mat = Mat()
pyopencv_to(<PyObject*> _M, M_mat)
# Create Size object from size tuple
# Note that size/shape in Python is handled in row-major-order -- therefore, width is [1] and height is [0]
cdef Size size = Size(<int> _size_tuple[1], <int> _size_tuple[0])
# Create empty GPU/device OutputArray for dst
cdef GpuMat dst_gpu = GpuMat()
warpPerspective(src_gpu, dst_gpu, M_mat, size, INTER_NEAREST)
# Get result of dst
cdef Mat dst_host
dst_gpu.download(dst_host)
cdef np.ndarray out = <np.ndarray> pyopencv_from(dst_host)
return out
After running a setup script to cythonize and compile this code (see apendix), we can import GpuWrapper as a Python module and run cudaWarpPerspectiveWrapper. This allowed me to run the code through CUDA with only a mismatch of 0.34722222222222854% -- quite exciting!
References (can only post max of 2)
Appendix
pyopencv_converter.cpp
#include <Python.h>
#include "numpy/ndarrayobject.h"
#include "opencv2/core/core.hpp"
static PyObject* opencv_error = 0;
// === FAIL MESSAGE ====================================================================================================
static int failmsg(const char *fmt, ...)
{
char str[1000];
va_list ap;
va_start(ap, fmt);
vsnprintf(str, sizeof(str), fmt, ap);
va_end(ap);
PyErr_SetString(PyExc_TypeError, str);
return 0;
}
struct ArgInfo
{
const char * name;
bool outputarg;
// more fields may be added if necessary
ArgInfo(const char * name_, bool outputarg_)
: name(name_)
, outputarg(outputarg_) {}
// to match with older pyopencv_to function signature
operator const char *() const { return name; }
};
// === THREADING =======================================================================================================
class PyAllowThreads
{
public:
PyAllowThreads() : _state(PyEval_SaveThread()) {}
~PyAllowThreads()
{
PyEval_RestoreThread(_state);
}
private:
PyThreadState* _state;
};
class PyEnsureGIL
{
public:
PyEnsureGIL() : _state(PyGILState_Ensure()) {}
~PyEnsureGIL()
{
PyGILState_Release(_state);
}
private:
PyGILState_STATE _state;
};
// === ERROR HANDLING ==================================================================================================
#define ERRWRAP2(expr) \
try \
{ \
PyAllowThreads allowThreads; \
expr; \
} \
catch (const cv::Exception &e) \
{ \
PyErr_SetString(opencv_error, e.what()); \
return 0; \
}
// === USING NAMESPACE CV ==============================================================================================
using namespace cv;
// === NUMPY ALLOCATOR =================================================================================================
class NumpyAllocator : public MatAllocator
{
public:
NumpyAllocator() { stdAllocator = Mat::getStdAllocator(); }
~NumpyAllocator() {}
UMatData* allocate(PyObject* o, int dims, const int* sizes, int type, size_t* step) const
{
UMatData* u = new UMatData(this);
u->data = u->origdata = (uchar*)PyArray_DATA((PyArrayObject*) o);
npy_intp* _strides = PyArray_STRIDES((PyArrayObject*) o);
for( int i = 0; i < dims - 1; i++ )
step[i] = (size_t)_strides[i];
step[dims-1] = CV_ELEM_SIZE(type);
u->size = sizes[0]*step[0];
u->userdata = o;
return u;
}
UMatData* allocate(int dims0, const int* sizes, int type, void* data, size_t* step, int flags, UMatUsageFlags usageFlags) const
{
if( data != 0 )
{
CV_Error(Error::StsAssert, "The data should normally be NULL!");
// probably this is safe to do in such extreme case
return stdAllocator->allocate(dims0, sizes, type, data, step, flags, usageFlags);
}
PyEnsureGIL gil;
int depth = CV_MAT_DEPTH(type);
int cn = CV_MAT_CN(type);
const int f = (int)(sizeof(size_t)/8);
int typenum = depth == CV_8U ? NPY_UBYTE : depth == CV_8S ? NPY_BYTE :
depth == CV_16U ? NPY_USHORT : depth == CV_16S ? NPY_SHORT :
depth == CV_32S ? NPY_INT : depth == CV_32F ? NPY_FLOAT :
depth == CV_64F ? NPY_DOUBLE : f*NPY_ULONGLONG + (f^1)*NPY_UINT;
int i, dims = dims0;
cv::AutoBuffer<npy_intp> _sizes(dims + 1);
for( i = 0; i < dims; i++ )
_sizes[i] = sizes[i];
if( cn > 1 )
_sizes[dims++] = cn;
PyObject* o = PyArray_SimpleNew(dims, _sizes, typenum);
if(!o)
CV_Error_(Error::StsError, ("The numpy array of typenum=%d, ndims=%d can not be created", typenum, dims));
return allocate(o, dims0, sizes, type, step);
}
bool allocate(UMatData* u, int accessFlags, UMatUsageFlags usageFlags) const
{
return stdAllocator->allocate(u, accessFlags, usageFlags);
}
void deallocate(UMatData* u) const
{
if(!u)
return;
PyEnsureGIL gil;
CV_Assert(u->urefcount >= 0);
CV_Assert(u->refcount >= 0);
if(u->refcount == 0)
{
PyObject* o = (PyObject*)u->userdata;
Py_XDECREF(o);
delete u;
}
}
const MatAllocator* stdAllocator;
};
// === ALLOCATOR INITIALIZATION ========================================================================================
NumpyAllocator g_numpyAllocator;
// === CONVERTOR FUNCTIONS =============================================================================================
template<typename T> static
bool pyopencv_to(PyObject* obj, T& p, const char* name = "<unknown>");
template<typename T> static
PyObject* pyopencv_from(const T& src);
enum { ARG_NONE = 0, ARG_MAT = 1, ARG_SCALAR = 2 };
// special case, when the convertor needs full ArgInfo structure
static bool pyopencv_to(PyObject* o, Mat& m, const ArgInfo info)
{
bool allowND = true;
if(!o || o == Py_None)
{
if( !m.data )
m.allocator = &g_numpyAllocator;
return true;
}
if( PyInt_Check(o) )
{
double v[] = {static_cast<double>(PyInt_AsLong((PyObject*)o)), 0., 0., 0.};
m = Mat(4, 1, CV_64F, v).clone();
return true;
}
if( PyFloat_Check(o) )
{
double v[] = {PyFloat_AsDouble((PyObject*)o), 0., 0., 0.};
m = Mat(4, 1, CV_64F, v).clone();
return true;
}
if( PyTuple_Check(o) )
{
int i, sz = (int)PyTuple_Size((PyObject*)o);
m = Mat(sz, 1, CV_64F);
for( i = 0; i < sz; i++ )
{
PyObject* oi = PyTuple_GET_ITEM(o, i);
if( PyInt_Check(oi) )
m.at<double>(i) = (double)PyInt_AsLong(oi);
else if( PyFloat_Check(oi) )
m.at<double>(i) = (double)PyFloat_AsDouble(oi);
else
{
failmsg("%s is not a numerical tuple", info.name);
m.release();
return false;
}
}
return true;
}
if( !PyArray_Check(o) )
{
failmsg("%s is not a numpy array, neither a scalar", info.name);
return false;
}
PyArrayObject* oarr = (PyArrayObject*) o;
bool needcopy = false, needcast = false;
int typenum = PyArray_TYPE(oarr), new_typenum = typenum;
int type = typenum == NPY_UBYTE ? CV_8U :
typenum == NPY_BYTE ? CV_8S :
typenum == NPY_USHORT ? CV_16U :
typenum == NPY_SHORT ? CV_16S :
typenum == NPY_INT ? CV_32S :
typenum == NPY_INT32 ? CV_32S :
typenum == NPY_FLOAT ? CV_32F :
typenum == NPY_DOUBLE ? CV_64F : -1;
if( type < 0 )
{
if( typenum == NPY_INT64 || typenum == NPY_UINT64 || typenum == NPY_LONG )
{
needcopy = needcast = true;
new_typenum = NPY_INT;
type = CV_32S;
}
else
{
failmsg("%s data type = %d is not supported", info.name, typenum);
return false;
}
}
#ifndef CV_MAX_DIM
const int CV_MAX_DIM = 32;
#endif
int ndims = PyArray_NDIM(oarr);
if(ndims >= CV_MAX_DIM)
{
failmsg("%s dimensionality (=%d) is too high", info.name, ndims);
return false;
}
int size[CV_MAX_DIM+1];
size_t step[CV_MAX_DIM+1];
size_t elemsize = CV_ELEM_SIZE1(type);
const npy_intp* _sizes = PyArray_DIMS(oarr);
const npy_intp* _strides = PyArray_STRIDES(oarr);
bool ismultichannel = ndims == 3 && _sizes[2] <= CV_CN_MAX;
for( int i = ndims-1; i >= 0 && !needcopy; i-- )
{
// these checks handle cases of
// a) multi-dimensional (ndims > 2) arrays, as well as simpler 1- and 2-dimensional cases
// b) transposed arrays, where _strides[] elements go in non-descending order
// c) flipped arrays, where some of _strides[] elements are negative
// the _sizes[i] > 1 is needed to avoid spurious copies when NPY_RELAXED_STRIDES is set
if( (i == ndims-1 && _sizes[i] > 1 && (size_t)_strides[i] != elemsize) ||
(i < ndims-1 && _sizes[i] > 1 && _strides[i] < _strides[i+1]) )
needcopy = true;
}
if( ismultichannel && _strides[1] != (npy_intp)elemsize*_sizes[2] )
needcopy = true;
if (needcopy)
{
if (info.outputarg)
{
failmsg("Layout of the output array %s is incompatible with cv::Mat (step[ndims-1] != elemsize or step[1] != elemsize*nchannels)", info.name);
return false;
}
if( needcast ) {
o = PyArray_Cast(oarr, new_typenum);
oarr = (PyArrayObject*) o;
}
else {
oarr = PyArray_GETCONTIGUOUS(oarr);
o = (PyObject*) oarr;
}
_strides = PyArray_STRIDES(oarr);
}
// Normalize strides in case NPY_RELAXED_STRIDES is set
size_t default_step = elemsize;
for ( int i = ndims - 1; i >= 0; --i )
{
size[i] = (int)_sizes[i];
if ( size[i] > 1 )
{
step[i] = (size_t)_strides[i];
default_step = step[i] * size[i];
}
else
{
step[i] = default_step;
default_step *= size[i];
}
}
// handle degenerate case
if( ndims == 0) {
size[ndims] = 1;
step[ndims] = elemsize;
ndims++;
}
if( ismultichannel )
{
ndims--;
type |= CV_MAKETYPE(0, size[2]);
}
if( ndims > 2 && !allowND )
{
failmsg("%s has more than 2 dimensions", info.name);
return false;
}
m = Mat(ndims, size, type, PyArray_DATA(oarr), step);
m.u = g_numpyAllocator.allocate(o, ndims, size, type, step);
m.addref();
if( !needcopy )
{
Py_INCREF(o);
}
m.allocator = &g_numpyAllocator;
return true;
}
template<>
bool pyopencv_to(PyObject* o, Mat& m, const char* name)
{
return pyopencv_to(o, m, ArgInfo(name, 0));
}
template<>
PyObject* pyopencv_from(const Mat& m)
{
if( !m.data )
Py_RETURN_NONE;
Mat temp, *p = (Mat*)&m;
if(!p->u || p->allocator != &g_numpyAllocator)
{
temp.allocator = &g_numpyAllocator;
ERRWRAP2(m.copyTo(temp));
p = &temp;
}
PyObject* o = (PyObject*)p->u->userdata;
Py_INCREF(o);
return o;
}
setupGpuWrapper.py
import subprocess
import os
import numpy as np
from distutils.core import setup, Extension
from Cython.Build import cythonize
from Cython.Distutils import build_ext
"""
Run setup with the following command:
```
python setupGpuWrapper.py build_ext --inplace
```
"""
# Determine current directory of this setup file to find our module
CUR_DIR = os.path.dirname(__file__)
# Use pkg-config to determine library locations and include locations
opencv_libs_str = subprocess.check_output("pkg-config --libs opencv".split()).decode()
opencv_incs_str = subprocess.check_output("pkg-config --cflags opencv".split()).decode()
# Parse into usable format for Extension call
opencv_libs = [str(lib) for lib in opencv_libs_str.strip().split()]
opencv_incs = [str(inc) for inc in opencv_incs_str.strip().split()]
extensions = [
Extension('GpuWrapper',
sources=[os.path.join(CUR_DIR, 'GpuWrapper.pyx')],
language='c++',
include_dirs=[np.get_include()] + opencv_incs,
extra_link_args=opencv_libs)
]
setup(
cmdclass={'build_ext': build_ext},
name="GpuWrapper",
ext_modules=cythonize(extensions)
)
