Search code examples
pythongpupycuda

PyCUDA - ElementWise fails for equality checking

I'm trying to build an equality checker for two arrays that can I can run on my GPU using PyCUDA.

Following the example given on the PyCUDA GPU Arrays documentation page, I attempted to write my own implementation. But whilst the below code works as expected for arithmetic, e.g. "z[i] = x[i] + y[i]", it returns erroneous output for the equality checker operand "z[i] = x[i] == y[i]". 

import pycuda.gpuarray as gpuarray
import pycuda.driver as cuda
import pycuda.autoinit
import numpy as np
from pycuda.elementwise import ElementwiseKernel

matrix_size = (5,)
a = np.random.randint(2, size=matrix_size)
b = np.random.randint(2, size=matrix_size)

print a
print b

a_gpu = gpuarray.to_gpu(a) 
b_gpu = gpuarray.to_gpu(b)

eq_checker = ElementwiseKernel(
        "int *x, int *y, int *z",
        "z[i] = x[i] == y[i]",
        "equality_checker")

c_gpu = gpuarray.empty_like(a_gpu)
eq_checker(a_gpu, b_gpu, c_gpu)

print c_gpu

Which prints out something like:

[0 1 0 0 0]
[0 1 1 1 0]
[4294967297 4294967297          0          1          1]

Does anyone understand why this error is occurring, or at least have an alternative PyCUDA method to achieve the desired function?

Solution

  • Solved! The problem was that numpy automatically returns 64-bit integers, whereas PyCUDA only standardly accepts 32-bit integers.

    This is therefore fixed by specifying the type of ints numpy generates, such as:

    a = np.random.randint(2, size=matrix_size, dtype=np.int32)
b = np.random.randint(2, size=matrix_size, dtype=np.int32)

    after which it works as expected.