Received from Hugh Owens on Sun, Feb 12, 2012 at 05:41:40PM EST:
> Hello,
>
> I recently began a project using PyCUDA and I'm encountering some
> strange behavior that I cannot figure out the source of. Before I
> continue, I will preface this with a big disclaimer: I'm not a
> particularly good C/C++ programmer, and it's possible the heart of
> matter is just that. That said, my problem is as follows:
>
> Floats returned from the GPU are just... not the values they should
> be. This problem arose in the process of writing a much more
> complicated routine, but I've been able to duplicate it in the minimal
> example below.
>
>
> import numpy as np
> import pycuda.driver as drv
> import pycuda.autoinit
> from pycuda.compiler import SourceModule
>
>
> mod = SourceModule(
> """
> __global__ void min_example(float *a, float b)
> {
> const int i = threadIdx.x;
> a[i] = b;
> }
> """
> )
>
> example = mod.get_function("min_example")
> aa = np.ones(512)
> example(drv.InOut(aa),np.float32(1.0),block = (512,1,1))
> print aa[0] #Prints 0.0078125. What?
>
> In the case shown, rather than 1.0, the array contains "0.0078125"
> after running the kernel. Oddly enough, for b = 2.0, the output is
> actually correct, or correct enough, "2.00000047684". For 3.0, I get
> something like 32, for 4.0, "512.000123024", and for 5.0, 2048ish.
> Now, I do see a pattern here, but what's causing it is really beyond
> me. Any help would be tremendously appreciated.
>
> Thanks,
>
> Hugh Owens
In your code above, aa contains double precision floats while your
kernel operates on single precision floats. Try making the following
change:
aa = np.ones(512, dtype=np.float32)
L.G.
_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda
