Dnia 2011-12-09, pią o godzinie 11:06 -0500, Thomas Wiecki pisze:
> That does seem to work. Its actually what I initially thought of doing
> (but didn't know that .state can be passed as a curandState. Thanks!

No problem. .state is just a pointer to device memory storing states.

> 
> Is the .fill(array) required? I tried it without and at least didn't crash.

No. All initialisation is done in constructor.
.fill() just calls curand() (or similar) using those states.

> 
> Just so I know what it's doing, upon instantiation,
> XORWOWRandomNumberGenerator() creates the maximum number of
> curandStates? What happens when I call my function with idx being
> greater than this number?

undefined behaviour^TM ;-)
It will try to go outside allocated memory and behaviour will
depend on the device, CUDA version, etc.
On some it will just read unitilialised data, on some hardware
might caught access to unallocated memory, ...
IMO it will be the same as IndexOutOfRange in C/C++ - we
do now know what will happen.

> 
> On Fri, Dec 9, 2011 at 8:12 AM, Tomasz Rybak <bogom...@post.pl> wrote:
> > Dnia 2011-12-09, pią o godzinie 08:01 -0500, Thomas Wiecki pisze:
> >> The problem is that I don't want an array of random numbers in the end
> >> (as your code does), but want to pass generators to a cuda function
> >> that then simulates a stochastic process. So I need access to the
> >> curandState.
> >
> > The memory storing array of generators' states in in property
> > XORWOWRandomNumberGenerator.states. Array has as many elements
> > as there is cores on the GPU (e.g. 512 for GTX 590).
> >
> > Regards.
> >
> > --
> > Tomasz Rybak <bogom...@post.pl> GPG/PGP key ID: 2AD5 9860
> > Fingerprint A481 824E 7DD3 9C0E C40A  488E C654 FB33 2AD5 9860
> > http://member.acm.org/~tomaszrybak


-- 
Tomasz Rybak <bogom...@post.pl> GPG/PGP key ID: 2AD5 9860
Fingerprint A481 824E 7DD3 9C0E C40A  488E C654 FB33 2AD5 9860
http://member.acm.org/~tomaszrybak

Attachment: signature.asc
Description: This is a digitally signed message part

_______________________________________________
PyCUDA mailing list
PyCUDA@tiker.net
http://lists.tiker.net/listinfo/pycuda

Reply via email to