Please make sure to keep the list cc'd. Joe Haywood <[email protected]> writes: > The code uses a rng based on Marsaglia's MWC which I implemented, the Cuda > and Pyopencl codes are the same. There are no complex numbers. I did use some > print statements in both the cuda code and opencl code and the numbers > printf'd are the same, within reason, throughout. One thing I notice was even > though I had initialized the output array to all zeros when it comes back to > python it has nans and infs in places. I verified that there are no division > by zero but I'm not sure if I'm overflowing the float type. I will check this > more tomorrow. > > I also noticed something in another opencl list that said a race condition > could still exist even when doing global atomic addition. That is, the first > batch of calculations know the value of the global variable as one thing but > the next batches know the global variable as its initial value not the > updated value. I will investigate this more tomorrow as well. > > I don't have the code available online but I will post the relevant parts > soon. > > I have only ever been able to get the Pyopencl from the Ubuntu repo to > work. Is building a different version as easy as "pip install > pyopencl=version"?
If you've got the CL headers and the Python headers ('python-dev'
package), it should be.
Andreas
signature.asc
Description: PGP signature
_______________________________________________ PyOpenCL mailing list [email protected] http://lists.tiker.net/listinfo/pyopencl
