hannes-brt / cudnn-python-wrappers Goto Github PK

I've tried to use your branch for CUDNN RC2 API, but had a strange thing: when I tried to use it with double, the convolution gave me some strange values: about 10 ^ -312. When I switched to float, everithing was okay again.

Here I made a small test for this error:
https://github.com/Felix-neko/cudnn-python-bugreport

Cannot you look at it?

P.S. Thank your for your patience.
^__^

Hi! Cannot you somehow add cudaSetDevice function to your library?

For example,

_libcudart.cudaSetDevice.restype = int
_libcudart.cudaSetDevice.argtypes = [ctypes.c_int]
# Added to select device to execute CUDA streams
def cudaSetDevice(devid):
    """
    Args:
        devid (int): device from 0 to N_DEV - 1
    Returns:
        int
    """
    status = _libcudart.cudaSetDevice(ctypes.c_int(devid))
    return status

Alas, there is no cudaSetDevice in pycuda, so we, the users, have to call it manually.

i have edited the filename of the cudnn64_x.dll,so it can load cudnnv5.1 now.
if sys.platform in ('linux2', 'linux'): _libcudnn_libname_list = ['libcudnn.so', 'libcudnn.so.6.5', 'libcudnn.so.6.5.18'] elif sys.platform == 'win32': _libcudnn_libname_list = ['cudnn64_5.dll'] else: raise RuntimeError('unsupported platform')

But in line311,it got another error. cudnnv5.1 dosen't have "cudnnCreateTensor4dDescriptor" function.But it has a function called**"cudnnCreateTensorDescriptor"** instead.
>>> import libcudnn, ctypes Traceback (most recent call last): File "<stdin>", line 1, in <module> File "C:\Anaconda2\Lib\site-packages\libcudnn.py", line 311, in <module> _libcudnn.cudnnCreateTensor4dDescriptor.restype = int File "C:\Anaconda2\lib\ctypes\__init__.py", line 375, in __getattr__ func = self.__getitem__(name) File "C:\Anaconda2\lib\ctypes\__init__.py", line 380, in __getitem__ func = self._FuncPtr((name_or_ordinal, self)) AttributeError: function 'cudnnCreateTensor4dDescriptor' not found

Guten morgen!
There is a new CUDNN release (R2). Aren't you going to support it?

Just now it gives the following error:

import libcudnn as cudnn
File "/usr/local/lib/python3.4/dist-packages/libcudnn.py", line 311, in <module>
_libcudnn.cudnnCreateTensor4dDescriptor.restype = int
File "/usr/lib/python3.4/ctypes/__init__.py", line 364, in __getattr__
func = self.__getitem__(name)
File "/usr/lib/python3.4/ctypes/__init__.py", line 369, in __getitem__
func = self._FuncPtr((name_or_ordinal, self))
AttributeError: /usr/local/cuda-6.5/lib64/libcudnn.so: undefined symbol: cudnnCreateTensor4dDescriptor

UPD: As I can see, the API has changed a bit.

Hi, I am trying to run a sample code to test my cudnn installation.
I can import libcudnn without any error. But when I run:
cudnn_context = libcudnn.cudnnCreate()
to initialize cuDNN and return a handle to the cuDNN context

I get the following error:

Traceback (most recent call last):
File "", line 1, in
File "C:\Anaconda\lib\site-packages\libcudnn.py", line 277, in cudnnCreate
cudnnCheckStatus(status)
File "C:\Anaconda\lib\site-packages\libcudnn.py", line 246, in cudnnCheckStatu
s
raise cudnnExceptions[status]
libcudnn.cudnnStatusArchMismatch

I have a nvidia geforce 635m with cuda 6.5 installed on my computer, and VS2013 professional.
is there any solution please ?

When I try to do import libcudnn I get the following error:

ImportError: dynamic module does not define init function (initlibcudnn)

I'm running Ubuntu 14.10 and using cudnn-6.5-linux-x64-v2-rc2.

On Ubuntu 14.04 and Python 3.4 there is sys.platform == "linux", not "linux2".

Cannot you fix it?

Placeholder for cuDNN v4 support development effort.

I am only set up to test on Linux, so it would be helpful if somebody else could test on Windows and Mac.