Have you tried initializing and releasing the CUDA context directly instead
of using pycuda.autoinit ?
- bryan
On Thu, Sep 8, 2011 at 4:20 PM, Matthew Graham <m...@cacr.caltech.edu> wrote:
> Hi,
>
> I've made some progress in this by moving the GPU initialization and source
> module compilation within the active function:
>
> class GPU():
>
> def __init__(self):
> self.kernel_code = '''...'''
>
> def culsp(self, data, params):
> # Initialize GPU
> import pycuda.autoinit
> mod = compiler.SourceModule(self.kernel_code)
> cuker = mod.get_function('culsp_kernel')
> # Call function
> cuker(data, params)
>
> When CherryPy receives an HTTP GET, it calls culsp with the data and
> params. This now works the first time it is called when CherryPy is running
> but subsequent calls give:
>
> mod = compiler.SourceModule(self.kernel_code)
> File
> "/Library/Frameworks/Python.framework/Versions/6.0.0/lib/python2.6/site-packages/pycuda-2011.1-py2.6-macosx-10.5-i386.egg/pycuda/compiler.py",
> line 271, in __init__
> arch, code, cache_dir, include_dirs)
> File
> "/Library/Frameworks/Python.framework/Versions/6.0.0/lib/python2.6/site-packages/pycuda-2011.1-py2.6-macosx-10.5-i386.egg/pycuda/compiler.py",
> line 218, in compile
> arch = "sm_%d%d" % Context.get_device().compute_capability()
> LogicError: cuCtxGetDevice failed: invalid context
>
> So it looks like with subsequent calls to the function (new thread), no
> context can be established. I'm sure that there must a simple way to do
> this?
>
> Cheers,
>
> Matthew
>
>
>
> On Sep 1, 2011, at 11:15 AM, Josh Bleecher Snyder wrote:
>
> > Perhaps CherryPy is spinning up new Python threads to handle the
> > inbound requests, and the handle to the CUDA context doesn't transfer
> > nicely across threads.
> >
> > I encountered a similar situation (in a different context), and ended
> > up with a solution in which I have standalone single-threaded worker
> > processes that do the GPU processing, and my request handler
> > communicates with them to get actual processing done. Or perhaps
> > CherryPy has a single-threaded mode...
> >
> > -josh
> >
> >
> >
> > On Thu, Sep 1, 2011 at 8:21 AM, Matthew Graham <m...@cacr.caltech.edu>
> wrote:
> >> Hi Andreas,
> >>
> >> In both cases, I am the user running the code.
> >>
> >> Cheers,
> >>
> >> Matthew
> >>
> >>
> >> On Sep 1, 2011, at 5:06 AM, Andreas Kloeckner wrote:
> >>
> >>> Hi Matthew,
> >>> Pu
> >>> On Wed, 31 Aug 2011 22:10:33 -0700, Matthew Graham <
> m...@cacr.caltech.edu> wrote:
> >>>> I have some code that runs fine as a standalone routine; however, I
> want to incorporate it as part of a web service using CherryPy as the
> hosting framework. When I try and call it from a client, I get:
> >>>>
> >>>> [31/Aug/2011:21:56:38] HTTP Traceback (most recent call last):
> >>>> File
> "/Library/Frameworks/Python.framework/Versions/6.0.0/lib/python2.6/site-packages/cherrypy/_cprequest.py",
> line 606, in respond
> >>>> cherrypy.response.body = self.handler()
> >>>> File
> "/Library/Frameworks/Python.framework/Versions/6.0.0/lib/python2.6/site-packages/cherrypy/_cpdispatch.py",
> line 25, in __call__
> >>>> return self.callable(*self.args, **self.kwargs)
> >>>> File "tscs.py", line 124, in POST
> >>>> response = self._analyze(data, kwargs)
> >>>> File "tscs.py", line 83, in _analyze
> >>>> row.append(self.stats[statid]['method'](data[id]))
> >>>> File "/Users/mjg/Projects/python/kiss/statistics.py", line 268, in
> gpuls
> >>>> lsp = Culsp()
> >>>> File "/Users/mjg/Projects/python/kiss/culsp.py", line 146, in
> __init__
> >>>> mod = compiler.SourceModule(kernel_code)
> >>>> File
> "/Library/Frameworks/Python.framework/Versions/6.0.0/lib/python2.6/site-packages/pycuda-2011.1-py2.6-macosx-10.5-i386.egg/pycuda/compiler.py",
> line 271, in __init__
> >>>> arch, code, cache_dir, include_dirs)
> >>>> File
> "/Library/Frameworks/Python.framework/Versions/6.0.0/lib/python2.6/site-packages/pycuda-2011.1-py2.6-macosx-10.5-i386.egg/pycuda/compiler.py",
> line 218, in compile
> >>>> arch = "sm_%d%d" % Context.get_device().compute_capability()
> >>>> LogicError: cuCtxGetDevice failed: invalid context
> >>>>
> >>>> Any ideas? I'm running CUDA 4.0 and PyCUDA 2011.1.2.
> >>>
> >>> Have you run your code as the web user? (Often www-data, but maybe
> >>> something different?) Does the same thing happen then? In any case, I
> >>> suspect that this is some permissions/security issue. It might be
> >>> helpful trying to replicate all the restrictions that web
> >>> server-executed code has on it...
> >>>
> >>> HTH,
> >>> Andreas
> >>
> >>
> >> _______________________________________________
> >> PyCUDA mailing list
> >> PyCUDA@tiker.net
> >> http://lists.tiker.net/listinfo/pycuda
> >>
> >
> > _______________________________________________
> > PyCUDA mailing list
> > PyCUDA@tiker.net
> > http://lists.tiker.net/listinfo/pycuda
> >
>
>
> _______________________________________________
> PyCUDA mailing list
> PyCUDA@tiker.net
> http://lists.tiker.net/listinfo/pycuda
>
_______________________________________________
PyCUDA mailing list
PyCUDA@tiker.net
http://lists.tiker.net/listinfo/pycuda
