Demos failing with CUDA 5

[alexbw]

Run any demo with CUDA 5 installed (Linux, 64-bit, with newest CULA as well as of 2/3/2013), and all demos fail, either because np.allclose() fails, meaning computations are incorrect, or there is an ambiguous cublasInternalError() that occurs.

Any thoughts, or should I just go ahead and try downgrading to CUDA 4?

[lebedov]

I haven't had occasion to test scikits.cuda with CUDA 5 yet, so I can't comment at this point. Try downgrading to CUDA 4. I'm leaving this issue open until I have a chance to investigate.

[alexbw]

Will do tonight, and let you know.

On Feb 3, 2013, at 10:24 PM, Lev Givon notifications@github.com wrote:

I haven't had occasion to test scikits.cuda with CUDA 5 yet, so I can't comment at this point. Try downgrading to CUDA 4. I'm leaving this issue open until I have a chance to investigate.

—
Reply to this email directly or view it on GitHub.

[alexbw]

Have you tested it against 4.0, or some newer 4.x version?

On Feb 3, 2013, at 10:24 PM, Lev Givon notifications@github.com wrote:

I haven't had occasion to test scikits.cuda with CUDA 5 yet, so I can't comment at this point. Try downgrading to CUDA 4. I'm leaving this issue open until I have a chance to investigate.

—
Reply to this email directly or view it on GitHub.

[alexbw]

I downgraded to 4.2, and I still am receiving the same errors. Running dot_demo.py:

Testing matrix multiplication for type float32
Success status: False

Testing vector multiplication for type float32

cublasExecutionFailed Traceback (most recent call last)
/usr/local/lib/python2.7/dist-packages/IPython/utils/py3compat.pyc in execfile(fname, *where)
181 else:
182 filename = fname
--> 183 builtin.execfile(filename, *where)

/home/dattalab/Downloads/scikits.cuda/demos/dot_demo.py in ()
53 e_gpu = gpuarray.to_gpu(e)
54
---> 55 temp = culinalg.dot(d_gpu, e_gpu)
56 print 'Success status: ', np.allclose(np.dot(d, e), temp)

/usr/local/lib/python2.7/dist-packages/scikits.cuda-0.042-py2.7.egg/scikits/cuda/linalg.pyc in dot(x_gpu, y_gpu, transa, transb)
263
264 return cublas_func(x_gpu.size, x_gpu.gpudata, 1,
--> 265 y_gpu.gpudata, 1)
266 else:
267

/usr/local/lib/python2.7/dist-packages/scikits.cuda-0.042-py2.7.egg/scikits/cuda/cublas.pyc in cublasSdot(n, x, incx, y, incy)
1287 int(x), incx, int(y), incy,
1288 ctypes.byref(result))
-> 1289 cublasCheckStatus(status)
1290 return np.float32(result.value)
1291

/usr/local/lib/python2.7/dist-packages/scikits.cuda-0.042-py2.7.egg/scikits/cuda/cublas.pyc in cublasCheckStatus(status)
149 if status != 0:
150 try:
--> 151 raise cublasExceptions[status]
152 except KeyError:
153 raise cublasError

cublasExecutionFailed:
terminate called after throwing an instance of 'cudalib::CublasException'
what():
[1] 4153 abort (core dumped) ipython dot_demo.py

On Feb 3, 2013, at 10:24 PM, Lev Givon notifications@github.com wrote:

I haven't had occasion to test scikits.cuda with CUDA 5 yet, so I can't comment at this point. Try downgrading to CUDA 4. I'm leaving this issue open until I have a chance to investigate.

—
Reply to this email directly or view it on GitHub.

[alexbw]

Also, many tests failing.

---

Ran 134 tests in 14.409s
FAILED (errors=121, failures=5)

Here's a subset: https://gist.github.com/4705170

On Feb 3, 2013, at 10:24 PM, Lev Givon notifications@github.com wrote:

I haven't had occasion to test scikits.cuda with CUDA 5 yet, so I can't comment at this point. Try downgrading to CUDA 4. I'm leaving this issue open until I have a chance to investigate.

—
Reply to this email directly or view it on GitHub.

[lebedov]

Are you using the latest revision from Github? The tests seem to run without any issues when I tried running them against the latest revision on a system running CUDA 4.2.9 and CULA R15.

[alexbw]

I'll double-check that I'm running with exactly the same versions as you list.
Also, I'm running Ubuntu 12.04, with a GeForce GTX 660.

[lebedov]

I ran the tests on Ubuntu 12.10 with a GeForce GTX 460 using the stock CUDA 4.2 packages available for that version of the distro.

[lebedov]

Tested latest revision in Github against CUDA 5.0.35 on Ubuntu without any problems.