'module' object has no attribute 'cublasCgemmBatched' about scikit-cuda

[lebedov]

What version of CUDA do you have installed? Also, can you try running the CUBLAS unit tests in tests/test_cublas.py and let me know what happens?

[samhumeau]

Hi, thanks for the quick answer,

I have currenty CUDA 5.5 installed (Ubuntu 14.04).
When I run the test_cublas, it shows 10 errors (on 54 tests) that follow the same pattern:

ERROR: test_cublasSgeam (main.test_cublas)
Traceback (most recent call last):
File "test_cublas.py", line 429, in test_cublasSgeam
cublas.cublasSgeam(self.cublas_handle, 'n', 'n', 2, 3,
AttributeError: 'module' object has no attribute 'cublasSgeam'

and the same thing for:
ERROR: test_cublasSgeam (main.test_cublas)
ERROR: test_cublasCgeam (main.test_cublas)
ERROR: test_cublasSgemmBatched (main.test_cublas)
ERROR: test_cublasStrsmBatched (main.test_cublas)
ERROR: test_cublasSgetrfBatched (main.test_cublas)
ERROR: test_cublasDgeam (main.test_cublas)
ERROR: test_cublasZgeam (main.test_cublas)
ERROR: test_cublasDgemmBatched (main.test_cublas)
ERROR: test_cublasDtrsmBatched (main.test_cublas)
ERROR: test_cublasDgetrfBatched (main.test_cublas)

Sam

[lebedov]

If you run

import scikits.cuda.cublas
print scikits.cuda.cublas._cublas_version

in an interactive Python session, what do you observe?

[samhumeau]

I observe 5000 (???)

Sam

[lebedov]

(The CUBLAS version number is slightly off because the mechanism for finding it looks at the soname major number of actual library file rather than calling cublasGetVersion(). I'll update this method to look at both the major and minor number. I don't believe this is related to your problem, though, because I also am running Ubuntu 14.04 with CUDA 5.5 and have not observed any problems.)

[lebedov]

Can you try installing scikits.cuda in a virtualenv and running the CUBLAS tests with it? Here is how to do so:

sudo apt-get install python-virtualenv
cd ~/
unset PYTHONPATH
/usr/bin/virtualenv MYENV
./MYENV/bin/pip install -e git+https://github.com/lebedov/scikits.cuda#egg=scikits.cuda
./MYENV/bin/pip install mako
./MYENV/bin/pip install scipy
cd ./MYENV/src/scikits.cuda
~/MYENV/bin/python tests/test_cublas.py

[benanne]

Hey Samuel,

please note that you should no longer use the implementation I have on my GitHub account. This code has been integrated into Theano in the meantime, and it has been improved as well. So if you'd like to use this, just get the git version of Theano. You can enable it easily, have a look at the 'conv2d_fft' docs here: http://deeplearning.net/software/theano/library/tensor/nnet/conv.html

That said, with the Theano version you will probably run into the same issue (it also uses cublasCgemmBatched).

[samhumeau]

Hey
@ Sander Dieleman I actually directly the theano version of your code. I said "the code of Sander Dieleman" because I have known it via your blog.

@lebedov. OK, so in the virtual env I do not have the error for cublasCgemmBatched (even if there are two errors):

...........................E...........................E
ERROR: test_cublasSgetrfBatched (main.test_cublas)

Traceback (most recent call last):
File "tests/test_cublas.py", line 632, in test_cublasSgetrfBatched
from scipy.linalg import lu_factor
ImportError: No module named scipy.linalg

ERROR: test_cublasDgetrfBatched (main.test_cublas)
Traceback (most recent call last):
File "tests/test_cublas.py", line 652, in test_cublasDgetrfBatched
from scipy.linalg import lu_factor
ImportError: No module named scipy.linalg

---

Ran 56 tests in 9.809s

I guess this simply has something to do with some things that I would have installed. I will try very soon to make my code that uses theano in a virtual env, it would be an acceptable solution.

Sam

[lebedov]

The two errors you observed are caused by scipy not being found. Can you successfully import scipy using the python instance in the virtualenv?

[benanne]

Ohh I see, alright. Hope you can figure out what's going on!

[samhumeau]

ok, so now everything is working fine on the virtual environment (scipy re-installed), the tests return 0 errors and I have a way to enjoy the amazing faster convolutions offered by the code of benanne.

On the main installation of python, however, I still have this problem, which I can't explain. But since I have a way to make it work (and that it is not really fun to dig into this kind of problems), I think I will let it go like this.

So thanks you guys for your advices and your awesome libraries.

Sam

[lebedov]

Glad you got it to work!

[francisquintallauzon]

I had the same problem today. What made the trick was simply to install the development version (0.5.0 as of today) after installing the pypi version. For those unfamiliar with pip simply run: sudo pip install --upgrade --no-deps git+https://github.com/lebedov/scikits.cuda.git