On Thu, 16 Jun 2011 18:17:07 +0200, Tomasz Rybak <[email protected]> wrote: > Dnia 2011-06-11, sob o godzinie 21:45 -0400, Andreas Kloeckner pisze: > > > Sorry to be a pain--but can you add some (at least light-duty) tests of > > all this new stuff? :) > > > > Sorry for the delay - was busy with work. > > Here is patch adding CURAND4 with tests and documentation. > I also added ability to call skip_sequence on XORWOW, > and fixed problem with skipahead: XORWOW and Sobol64 use > 64-bit argument, Sobol32 accepts only 32-bit argument. > For CURAND 3.2 this was not a problem, but CURAND 4.0 > started using templates and trying to use 32-bit > skip-ahead for XORWOW resulted in kernel compilation > error. > > Now everything is OK, and all tests work.
Failed for me on GTX 590.
Traceback (most recent call last):
File "/home/kloeckner/src/pycuda/pycuda/tools.py", line 549, in f
inner_f(*args, **kwargs)
File "/home/kloeckner/src/pycuda/test/test_gpuarray.py", line 286, in
test_curand_wrappers
gen = gen_type()
File "/home/kloeckner/src/pycuda/pycuda/curandom.py", line 836, in __init__
'curandStateScrambledSobol64',
sobol64_random_source+random_skip_ahead64_source)
File "/home/kloeckner/src/pycuda/pycuda/curandom.py", line 725, in __init__
raise ValueError("Initialisation failed. Decrease number of threads.")
ValueError: Initialisation failed. Decrease number of threads.
I don't understand this error message--I added a facility that queries
max_threads_per_block on each kernel it knows about. (This is in the
generators_per_block property getter). That should make manual
intervention of this sort entirely unnecessary. Any idea why that's not
working?
Also, I encountered these warnings, which I'd like fixed.
/home/kloeckner/src/pycuda/pycuda/curandom.py:636: DeprecationWarning: setting
the block size in Function.prepare is deprecated
p.prepare("PiPi", block=(self.generators_per_block, 1, 1))
/home/kloeckner/src/pycuda/pycuda/curandom.py:652: DeprecationWarning: Not
passing the block size to prepared_call is deprecated as of version 2011.1.
self.block_count * self.generators_per_block, dir_vector.gpudata, offset)
/home/kloeckner/src/pycuda/pycuda/curandom.py:702: DeprecationWarning: setting
the block size in Function.prepare is deprecated
p.prepare("PiPPi", block=(self.generators_per_block, 1, 1))
/home/kloeckner/src/pycuda/pycuda/curandom.py:719: DeprecationWarning: Not
passing the block size to prepared_call is deprecated as of version 2011.1.
dir_vector.gpudata, scramble_vector.gpudata, offset)
Andreas
pgpF6xXj5oFy4.pgp
Description: PGP signature
_______________________________________________ PyCUDA mailing list [email protected] http://lists.tiker.net/listinfo/pycuda
