5

Should be simple enough; I literally want to send an int to the a SourceModule kernel declaration, where the C function

__global__......(int value,.....)

with the value being declared and called...

value = 256
...
...
func(value,...)

But I'm getting a variety of errors from pycuda that I'm using the wrong type.

Framester
  • 33,341
  • 51
  • 130
  • 192
Bolster
  • 7,460
  • 13
  • 61
  • 96

1 Answers1

9

The standard PyCUDA function interface requires argument have numpy dtypes, because it internally does mapping to C types under the hood. So for scalar arguments which are passed by value, you need to "cast" to a suitable numpy dtype first. Something like

value = 256
va = numpy.int32(value)

func(va)

should work. If you are passing single precision floating point values or arrays, make sure to explicitly use a dtype of np.float32, because numpy uses double precision by default and you will wind up with similar errors.

talonmies
  • 70,661
  • 34
  • 192
  • 269
  • Thank you again. The next (mini) part of the question is the pointer-ing; if my function is expecting an int, will this still behave as 'expected' or is it a case of doing some pointer related fun in the kernel code? I ask because it seems a little wasteful cudamallocing for the sake of a single int. – Bolster Apr 19 '11 at 15:18
  • 2
    pycuda will map scalars to arguments passed by values and gpuarrays to pointers without your needing to do anything in the standard function interface. There is also an option to set up your own static interface list, which can be useful if you have to call a function many times because it reduces the python interpreter overhead in doing the same interface list over and over. That is a "last 5%" type optimization, though... – talonmies Apr 19 '11 at 15:23