On 2014-06-20 17:48, Lev Givon wrote:
When I tried running the above code against a K20Xm GPU with Ubuntu 14.04, CUDA
5.5, pycuda 2013.1.1 (stock Ubuntu package), and the latest code from
scikits.cuda master, I observed a different exception:
Traceback (most recent call last):
File "cuda.py", line 12, in <module>
X = gpuarray.to_gpu(np.random.randn(n, 2483).astype(np.float32))
File "/usr/lib/python2.7/dist-packages/pycuda/gpuarray.py", line 913, in
to_gpu
result.set(ary)
File "/usr/lib/python2.7/dist-packages/pycuda/gpuarray.py", line 228, in
set
drv.memcpy_htod(self.gpudata, ary)
pycuda._driver.LaunchError: cuMemcpyHtoD failed: launch failed
The exception occurs on the second iteration of the loop.
No errors observed when run against a Tesla S2050.
That reminds me that I've not given exact version numbers on my first mail:
The version of pycuda used was different on each machine: whatever was
the head of the master branch in the github repo at the time of the
intial machine installation (One of the machines was just installed a
few days ago, the others a few months back). The same goes for
scikit.cuda. The Tesla-K40 machines were all running CUDA 6, the GTX 780
machine was running CUDA 5.5.