cudart.py

#!/usr/bin/env python

"""
Python interface to CUDA runtime functions.
"""

import atexit, ctypes, platform, re, sys, warnings
import numpy as np

# Load library:
_linux_version_list = [11.0, 10.2, 10.1, 10.0, 9.2, 9.1, 9.0, 8.0, 7.5, 7.0, 6.5, 6.0, 5.5, 5.0, 4.0]
_win32_version_list = [110, 102, 101, 100, 92, 91, 90, 80, 75, 70, 65, 60, 55, 50, 40]
if 'linux' in sys.platform:
    _libcudart_libname_list = ['libcudart.so'] + \
                              ['libcudart.so.%s' % v for v in _linux_version_list]
elif sys.platform == 'darwin':
    _libcudart_libname_list = ['libcudart.dylib']
elif sys.platform == 'win32':
    if sys.maxsize > 2**32:
        _libcudart_libname_list = ['cudart.dll'] + \
                                  ['cudart64_%s.dll' % v for v in _win32_version_list]
    else:
        _libcudart_libname_list = ['cudart.dll'] + \
                                  ['cudart32_%s.dll' % v for v in _win32_version_list]
else:
    raise RuntimeError('unsupported platform')

# Print understandable error message when library cannot be found:
_libcudart = None
for _libcudart_libname in _libcudart_libname_list:
    try:
        if sys.platform == 'win32':
            _libcudart = ctypes.windll.LoadLibrary(_libcudart_libname)
        else:
            _libcudart = ctypes.cdll.LoadLibrary(_libcudart_libname)
    except OSError:
        pass
    else:
        break
if _libcudart == None:
    raise OSError('CUDA runtime library not found')

# Code adapted from PARRET:
def POINTER(obj):
    """
    Create ctypes pointer to object.

    Notes
    -----
    This function converts None to a real NULL pointer because of bug
    in how ctypes handles None on 64-bit platforms.

    """

    p = ctypes.POINTER(obj)
    if not isinstance(p.from_param, classmethod):
        def from_param(cls, x):
            if x is None:
                return cls()
            else:
                return x
        p.from_param = classmethod(from_param)

    return p

# Classes corresponding to CUDA vector structures:
class float2(ctypes.Structure):
    _fields_ = [
        ('x', ctypes.c_float),
        ('y', ctypes.c_float)
        ]

class cuFloatComplex(float2):
    @property
    def value(self):
        return complex(self.x, self.y)

class double2(ctypes.Structure):
    _fields_ = [
        ('x', ctypes.c_double),
        ('y', ctypes.c_double)
        ]

class cuDoubleComplex(double2):
    @property
    def value(self):
        return complex(self.x, self.y)

def gpuarray_ptr(g):
    """
    Return ctypes pointer to data in GPUAarray object.

    """

    addr = int(g.gpudata)
    if g.dtype == np.int8:
        return ctypes.cast(addr, POINTER(ctypes.c_byte))
    if g.dtype == np.uint8:
        return ctypes.cast(addr, POINTER(ctypes.c_ubyte))
    if g.dtype == np.int16:
        return ctypes.cast(addr, POINTER(ctypes.c_short))
    if g.dtype == np.uint16:
        return ctypes.cast(addr, POINTER(ctypes.c_ushort))
    if g.dtype == np.int32:
        return ctypes.cast(addr, POINTER(ctypes.c_int))
    if g.dtype == np.uint32:
        return ctypes.cast(addr, POINTER(ctypes.c_uint))
    if g.dtype == np.int64:
        return ctypes.cast(addr, POINTER(ctypes.c_long))
    if g.dtype == np.uint64:
        return ctypes.cast(addr, POINTER(ctypes.c_ulong))
    if g.dtype == np.float32:
        return ctypes.cast(addr, POINTER(ctypes.c_float))
    elif g.dtype == np.float64:
        return ctypes.cast(addr, POINTER(ctypes.c_double))
    elif g.dtype == np.complex64:
        return ctypes.cast(addr, POINTER(cuFloatComplex))
    elif g.dtype == np.complex128:
        return ctypes.cast(addr, POINTER(cuDoubleComplex))
    else:
        raise ValueError('unrecognized type')

_libcudart.cudaGetErrorString.restype = ctypes.c_char_p
_libcudart.cudaGetErrorString.argtypes = [ctypes.c_int]
def cudaGetErrorString(e):
    """
    Retrieve CUDA error string.

    Return the string associated with the specified CUDA error status
    code.

    Parameters
    ----------
    e : int
        Error number.

    Returns
    -------
    s : str
        Error string.

    """

    return _libcudart.cudaGetErrorString(e)

# Generic CUDA error:
class cudaError(Exception):
    """CUDA error."""
    pass

# Exceptions corresponding to various CUDA runtime errors:
class cudaErrorMissingConfiguration(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(1)
    pass

class cudaErrorMemoryAllocation(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(2)
    pass

class cudaErrorInitializationError(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(3)
    pass

class cudaErrorLaunchFailure(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(4)
    pass

class cudaErrorPriorLaunchFailure(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(5)
    pass

class cudaErrorLaunchTimeout(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(6)
    pass

class cudaErrorLaunchOutOfResources(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(7)
    pass

class cudaErrorInvalidDeviceFunction(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(8)
    pass

class cudaErrorInvalidConfiguration(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(9)
    pass

class cudaErrorInvalidDevice(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(10)
    pass

class cudaErrorInvalidValue(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(11)
    pass

class cudaErrorInvalidPitchValue(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(12)
    pass

class cudaErrorInvalidSymbol(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(13)
    pass

class cudaErrorMapBufferObjectFailed(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(14)
    pass

class cudaErrorUnmapBufferObjectFailed(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(15)
    pass

class cudaErrorInvalidHostPointer(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(16)
    pass

class cudaErrorInvalidDevicePointer(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(17)
    pass

class cudaErrorInvalidTexture(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(18)
    pass

class cudaErrorInvalidTextureBinding(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(19)
    pass

class cudaErrorInvalidChannelDescriptor(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(20)
    pass

class cudaErrorInvalidMemcpyDirection(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(21)
    pass

class cudaErrorTextureFetchFailed(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(23)
    pass

class cudaErrorTextureNotBound(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(24)
    pass

class cudaErrorSynchronizationError(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(25)
    pass

class cudaErrorInvalidFilterSetting(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(26)
    pass

class cudaErrorInvalidNormSetting(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(27)
    pass

class cudaErrorMixedDeviceExecution(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(28)
    pass

class cudaErrorCudartUnloading(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(29)
    pass

class cudaErrorUnknown(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(30)
    pass

class cudaErrorNotYetImplemented(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(31)
    pass

class cudaErrorMemoryValueTooLarge(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(32)
    pass

class cudaErrorInvalidResourceHandle(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(33)
    pass

class cudaErrorNotReady(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(34)
    pass

class cudaErrorInsufficientDriver(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(35)
    pass

class cudaErrorSetOnActiveProcess(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(36)
    pass

class cudaErrorInvalidSurface(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(37)
    pass

class cudaErrorNoDevice(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(38)
    pass

class cudaErrorECCUncorrectable(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(39)
    pass

class cudaErrorSharedObjectSymbolNotFound(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(40)
    pass

class cudaErrorSharedObjectInitFailed(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(41)
    pass

class cudaErrorUnsupportedLimit(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(42)
    pass

class cudaErrorDuplicateVariableName(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(43)
    pass

class cudaErrorDuplicateTextureName(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(44)
    pass

class cudaErrorDuplicateSurfaceName(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(45)
    pass

class cudaErrorDevicesUnavailable(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(46)
    pass

class cudaErrorInvalidKernelImage(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(47)
    pass

class cudaErrorNoKernelImageForDevice(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(48)
    pass

class cudaErrorIncompatibleDriverContext(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(49)
    pass

class cudaErrorPeerAccessAlreadyEnabled(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(50)
    pass

class cudaErrorPeerAccessNotEnabled(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(51)
    pass

class cudaErrorDeviceAlreadyInUse(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(54)
    pass

class cudaErrorProfilerDisabled(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(55)
    pass

class cudaErrorProfilerNotInitialized(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(56)
    pass

class cudaErrorProfilerAlreadyStarted(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(57)
    pass

class cudaErrorProfilerAlreadyStopped(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(58)
    pass

class cudaErrorAssert(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(59)
    pass

class cudaErrorTooManyPeers(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(60)
    pass

class cudaErrorHostMemoryAlreadyRegistered(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(61)
    pass

class cudaErrorHostMemoryNotRegistered(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(62)
    pass

class cudaErrorOperatingSystem(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(63)
    pass

class cudaErrorPeerAccessUnsupported(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(64)
    pass

class cudaErrorLaunchMaxDepthExceeded(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(65)
    pass

class cudaErrorLaunchFileScopedTex(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(66)
    pass

class cudaErrorLaunchFileScopedSurf(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(67)
    pass

class cudaErrorSyncDepthExceeded(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(68)
    pass

class cudaErrorLaunchPendingCountExceeded(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(69)
    pass

class cudaErrorNotPermitted(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(70)
    pass

class cudaErrorNotSupported(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(71)
    pass

class cudaErrorHardwareStackError(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(72)
    pass

class cudaErrorIllegalInstruction(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(73)
    pass

class cudaErrorMisalignedAddress(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(74)
    pass

class cudaErrorInvalidAddressSpace(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(75)
    pass

class cudaErrorInvalidPc(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(76)
    pass

class cudaErrorIllegalAddress(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(77)
    pass

class cudaErrorInvalidPtx(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(78)
    pass

class cudaErrorInvalidGraphicsContext(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(79)

class cudaErrorStartupFailure(cudaError):
    __doc__ = _libcudart.cudaGetErrorString(127)
    pass

cudaExceptions = {
    1: cudaErrorMissingConfiguration,
    2: cudaErrorMemoryAllocation,
    3: cudaErrorInitializationError,
    4: cudaErrorLaunchFailure,
    5: cudaErrorPriorLaunchFailure,
    6: cudaErrorLaunchTimeout,
    7: cudaErrorLaunchOutOfResources,
    8: cudaErrorInvalidDeviceFunction,
    9: cudaErrorInvalidConfiguration,
    10: cudaErrorInvalidDevice,
    11: cudaErrorInvalidValue,
    12: cudaErrorInvalidPitchValue,
    13: cudaErrorInvalidSymbol,
    14: cudaErrorMapBufferObjectFailed,
    15: cudaErrorUnmapBufferObjectFailed,
    16: cudaErrorInvalidHostPointer,
    17: cudaErrorInvalidDevicePointer,
    18: cudaErrorInvalidTexture,
    19: cudaErrorInvalidTextureBinding,
    20: cudaErrorInvalidChannelDescriptor,
    21: cudaErrorInvalidMemcpyDirection,
    22: cudaError,
    23: cudaErrorTextureFetchFailed,
    24: cudaErrorTextureNotBound,
    25: cudaErrorSynchronizationError,
    26: cudaErrorInvalidFilterSetting,
    27: cudaErrorInvalidNormSetting,
    28: cudaErrorMixedDeviceExecution,
    29: cudaErrorCudartUnloading,
    30: cudaErrorUnknown,
    31: cudaErrorNotYetImplemented,
    32: cudaErrorMemoryValueTooLarge,
    33: cudaErrorInvalidResourceHandle,
    34: cudaErrorNotReady,
    35: cudaErrorInsufficientDriver,
    36: cudaErrorSetOnActiveProcess,
    37: cudaErrorInvalidSurface,
    38: cudaErrorNoDevice,
    39: cudaErrorECCUncorrectable,
    40: cudaErrorSharedObjectSymbolNotFound,
    41: cudaErrorSharedObjectInitFailed,
    42: cudaErrorUnsupportedLimit,
    43: cudaErrorDuplicateVariableName,
    44: cudaErrorDuplicateTextureName,
    45: cudaErrorDuplicateSurfaceName,
    46: cudaErrorDevicesUnavailable,
    47: cudaErrorInvalidKernelImage,
    48: cudaErrorNoKernelImageForDevice,
    49: cudaErrorIncompatibleDriverContext,
    50: cudaErrorPeerAccessAlreadyEnabled,
    51: cudaErrorPeerAccessNotEnabled,
    52: cudaError,
    53: cudaError,
    54: cudaErrorDeviceAlreadyInUse,
    55: cudaErrorProfilerDisabled,
    56: cudaErrorProfilerNotInitialized,
    57: cudaErrorProfilerAlreadyStarted,
    58: cudaErrorProfilerAlreadyStopped,
    59: cudaErrorAssert,
    60: cudaErrorTooManyPeers,
    61: cudaErrorHostMemoryAlreadyRegistered,
    62: cudaErrorHostMemoryNotRegistered,
    63: cudaErrorOperatingSystem,
    64: cudaErrorPeerAccessUnsupported,
    65: cudaErrorLaunchMaxDepthExceeded,
    66: cudaErrorLaunchFileScopedTex,
    67: cudaErrorLaunchFileScopedSurf,
    68: cudaErrorSyncDepthExceeded,
    69: cudaErrorLaunchPendingCountExceeded,
    70: cudaErrorNotPermitted,
    71: cudaErrorNotSupported,
    72: cudaErrorHardwareStackError,
    73: cudaErrorIllegalInstruction,
    74: cudaErrorMisalignedAddress,
    75: cudaErrorInvalidAddressSpace,
    76: cudaErrorInvalidPc,
    77: cudaErrorIllegalAddress,
    78: cudaErrorInvalidPtx,
    79: cudaErrorInvalidGraphicsContext,
    127: cudaErrorStartupFailure
    }

def cudaCheckStatus(status):
    """
    Raise CUDA exception.

    Raise an exception corresponding to the specified CUDA runtime error
    code.

    Parameters
    ----------
    status : int
        CUDA runtime error code.

    See Also
    --------
    cudaExceptions
    """

    if status != 0:
        try:
            e = cudaExceptions[status]
        except KeyError:
            raise cudaError('unknown CUDA error %s' % status)
        else:
            raise e

# Memory allocation functions (adapted from pystream):
_libcudart.cudaMalloc.restype = int
_libcudart.cudaMalloc.argtypes = [ctypes.POINTER(ctypes.c_void_p),
                                  ctypes.c_size_t]
def cudaMalloc(count, ctype=None):
    """
    Allocate device memory.

    Allocate memory on the device associated with the current active
    context.

    Parameters
    ----------
    count : int
        Number of bytes of memory to allocate
    ctype : _ctypes.SimpleType, optional
        ctypes type to cast returned pointer.

    Returns
    -------
    ptr : ctypes pointer
        Pointer to allocated device memory.

    """

    ptr = ctypes.c_void_p()
    status = _libcudart.cudaMalloc(ctypes.byref(ptr), count)
    cudaCheckStatus(status)
    if ctype != None:
        ptr = ctypes.cast(ptr, ctypes.POINTER(ctype))
    return ptr

_libcudart.cudaFree.restype = int
_libcudart.cudaFree.argtypes = [ctypes.c_void_p]
def cudaFree(ptr):
    """
    Free device memory.

    Free allocated memory on the device associated with the current active
    context.

    Parameters
    ----------
    ptr : ctypes pointer
        Pointer to allocated device memory.

    """

    status = _libcudart.cudaFree(ptr)
    cudaCheckStatus(status)

_libcudart.cudaMallocPitch.restype = int
_libcudart.cudaMallocPitch.argtypes = [ctypes.POINTER(ctypes.c_void_p),
                                       ctypes.POINTER(ctypes.c_size_t),
                                       ctypes.c_size_t, ctypes.c_size_t]
def cudaMallocPitch(pitch, rows, cols, elesize):
    """
    Allocate pitched device memory.

    Allocate pitched memory on the device associated with the current active
    context.

    Parameters
    ----------
    pitch : int
        Pitch for allocation.
    rows : int
        Requested pitched allocation height.
    cols : int
        Requested pitched allocation width.
    elesize : int
        Size of memory element.

    Returns
    -------
    ptr : ctypes pointer
        Pointer to allocated device memory.

    """

    ptr = ctypes.c_void_p()
    status = _libcudart.cudaMallocPitch(ctypes.byref(ptr),
                                        ctypes.c_size_t(pitch), cols*elesize,
                                        rows)
    cudaCheckStatus(status)
    return ptr, pitch

# Memory copy modes:
cudaMemcpyHostToHost = 0
cudaMemcpyHostToDevice = 1
cudaMemcpyDeviceToHost = 2
cudaMemcpyDeviceToDevice = 3
cudaMemcpyDefault = 4

_libcudart.cudaMemcpy.restype = int
_libcudart.cudaMemcpy.argtypes = [ctypes.c_void_p, ctypes.c_void_p,
                                  ctypes.c_size_t, ctypes.c_int]
def cudaMemcpy_htod(dst, src, count):
    """
    Copy memory from host to device.

    Copy data from host memory to device memory.

    Parameters
    ----------
    dst : ctypes pointer
        Device memory pointer.
    src : ctypes pointer
        Host memory pointer.
    count : int
        Number of bytes to copy.

    """

    status = _libcudart.cudaMemcpy(dst, src,
                                   ctypes.c_size_t(count),
                                   cudaMemcpyHostToDevice)
    cudaCheckStatus(status)

def cudaMemcpy_dtoh(dst, src, count):
    """
    Copy memory from device to host.

    Copy data from device memory to host memory.

    Parameters
    ----------
    dst : ctypes pointer
        Host memory pointer.
    src : ctypes pointer
        Device memory pointer.
    count : int
        Number of bytes to copy.

    """

    status = _libcudart.cudaMemcpy(dst, src,
                                   ctypes.c_size_t(count),
                                   cudaMemcpyDeviceToHost)
    cudaCheckStatus(status)

_libcudart.cudaMemGetInfo.restype = int
_libcudart.cudaMemGetInfo.argtypes = [ctypes.c_void_p,
                                      ctypes.c_void_p]
def cudaMemGetInfo():
    """
    Return the amount of free and total device memory.

    Returns
    -------
    free : long
        Free memory in bytes.
    total : long
        Total memory in bytes.

    """

    free = ctypes.c_size_t()
    total = ctypes.c_size_t()
    status = _libcudart.cudaMemGetInfo(ctypes.byref(free),
                                       ctypes.byref(total))
    cudaCheckStatus(status)
    return free.value, total.value

_libcudart.cudaSetDevice.restype = int
_libcudart.cudaSetDevice.argtypes = [ctypes.c_int]
def cudaSetDevice(dev):
    """
    Set current CUDA device.

    Select a device to use for subsequent CUDA operations.

    Parameters
    ----------
    dev : int
        Device number.

    """

    status = _libcudart.cudaSetDevice(dev)
    cudaCheckStatus(status)

_libcudart.cudaGetDevice.restype = int
_libcudart.cudaGetDevice.argtypes = [ctypes.POINTER(ctypes.c_int)]
def cudaGetDevice():
    """
    Get current CUDA device.

    Return the identifying number of the device currently used to
    process CUDA operations.

    Returns
    -------
    dev : int
        Device number.

    """

    dev = ctypes.c_int()
    status = _libcudart.cudaGetDevice(ctypes.byref(dev))
    cudaCheckStatus(status)
    return dev.value

_libcudart.cudaDriverGetVersion.restype = int
_libcudart.cudaDriverGetVersion.argtypes = [ctypes.POINTER(ctypes.c_int)]
def cudaDriverGetVersion():
    """
    Get installed CUDA driver version.

    Return the version of the installed CUDA driver as an integer. If
    no driver is detected, 0 is returned.

    Returns
    -------
    version : int
        Driver version.
    """

    version = ctypes.c_int()
    status = _libcudart.cudaDriverGetVersion(ctypes.byref(version))
    cudaCheckStatus(status)
    return version.value

_libcudart.cudaRuntimeGetVersion.restype = int
_libcudart.cudaRuntimeGetVersion.argtypes = [ctypes.POINTER(ctypes.c_int)]
def cudaRuntimeGetVersion():
    """
    Get installed CUDA runtime version.

    Return the version of the installed CUDA runtime as an integer. If
    no driver is detected, 0 is returned.

    Returns
    -------
    version : int
        Runtime version.
    """

    version = ctypes.c_int()
    status = _libcudart.cudaRuntimeGetVersion(ctypes.byref(version))
    cudaCheckStatus(status)
    return version.value

try:
    _cudart_version = cudaRuntimeGetVersion()
except:
    _cudart_version = 99999

class _cudart_version_req(object):
    """
    Decorator to replace function with a placeholder that raises an exception
    if the installed CUDA Runtime version is not greater than `v`.
    """

    def __init__(self, v):
        self.vs = str(v)
        if isinstance(v, int):
            major = str(v)
            minor = '0'
        else:
            major, minor = re.search(r'(\d+)\.(\d+)', self.vs).groups()
        self.vi = int(major.ljust(len(major)+1, '0')+minor.ljust(2, '0'))

    def __call__(self,f):
        def f_new(*args,**kwargs):
            raise NotImplementedError('CUDART '+self.vs+' required')
        f_new.__doc__ = f.__doc__

        if _cudart_version >= self.vi:
            return f
        else:
            return f_new

# Memory types:
cudaMemoryTypeHost = 1
cudaMemoryTypeDevice = 2

class cudaPointerAttributes(ctypes.Structure):
    _fields_ = [
        ('memoryType', ctypes.c_int),
        ('device', ctypes.c_int),
        ('devicePointer', ctypes.c_void_p),
        ('hostPointer', ctypes.c_void_p)
        ]

_libcudart.cudaPointerGetAttributes.restype = int
_libcudart.cudaPointerGetAttributes.argtypes = [ctypes.c_void_p,
                                                ctypes.c_void_p]
def cudaPointerGetAttributes(ptr):
    """
    Get memory pointer attributes.

    Returns attributes of the specified pointer.

    Parameters
    ----------
    ptr : ctypes pointer
        Memory pointer to examine.

    Returns
    -------
    memory_type : int
        Memory type; 1 indicates host memory, 2 indicates device
        memory.
    device : int
        Number of device associated with pointer.

    Notes
    -----
    This function only works with CUDA 4.0 and later.

    """

    attributes = cudaPointerAttributes()
    status = \
        _libcudart.cudaPointerGetAttributes(ctypes.byref(attributes), ptr)
    cudaCheckStatus(status)
    return attributes.memoryType, attributes.device