Pycuda - memcpy_dtoh, not giving what appears to have been set

Accelerated Computing CUDA CUDA Programming and Performance
,
1

I have a very simple function where I’m passing in a char array and doing a simple character match. I want to return an array of 1/0 values depending on which characters are matched.

Problem: although I can see the value has been set in the data structure (as I print it in the function after it’s assigned) when the int array is copied back from the device the values aren’t as expected.

I’m sure it’s something silly.

import pycuda.driver as cuda
import pycuda.autoinit
from pycuda.compiler import SourceModule
import numpy as np

mod = SourceModule("""
__global__ void test(const char *q, const int chrSize, int *d, const int intSize) {
    int v = 0;
    if( q[threadIdx.x * chrSize] == 'a' ||  q[threadIdx.x * chrSize] == 'c' ) {
        v = 1;
    }
    d[threadIdx.x * intSize] = v;
    printf("x=%d, y=%d, val=%c ret=%d\\n", threadIdx.x, threadIdx.y, q[threadIdx.x * chrSize], d[threadIdx.x * intSize]);
}
""")
func = mod.get_function("test")

# input data
a = np.asarray(['a','b','c','d'], dtype=np.str_)
# allocate/copy to device
a_gpu = cuda.mem_alloc(a.nbytes)
cuda.memcpy_htod(a_gpu, a)

# destination array
d = np.zeros((4), dtype=np.int16)
# allocate/copy to device
d_gpu = cuda.mem_alloc(d.nbytes)
cuda.memcpy_htod(d_gpu, d)

# run the function
func(a_gpu, np.int8(a.dtype.itemsize), d_gpu, np.int8(d.dtype.itemsize), block=(4,1,1))

# copy data back and print
cuda.memcpy_dtoh(d, d_gpu)
print(d)

Output:

x=0, y=0, val=a ret=1
x=1, y=0, val=b ret=0
x=2, y=0, val=c ret=1
x=3, y=0, val=d ret=0
[1 0 0 0]

Expected output:

x=0, y=0, val=a ret=1
x=1, y=0, val=b ret=0
x=2, y=0, val=c ret=1
x=3, y=0, val=d ret=0
[1 0 1 0]

Thanks in advance.

This has also been asked on stack overflow.

2

My error was identified in the stack overflow post:

You have two main problems, neither of which have anything to do with memcpy_dtoh :

  1. You have declared d and d_gpu as dtype np.int16 , but the kernel is expecting C++ int , leading to a type mistmatch. You should use the np.int32 type to define the arrays.
  2. The indexing of d within the kernel is incorrect. If you have declared the array to the compiler as a 32 bit type, indexing the array as d[threadIdx.x] will automatically include the correct alignment for the type. Passing and using intSize to the kernel for indexing d is not required and it is incorrect to do so.

If you fix those two issues, I suspect the code will work as intended.

The indicated modifications solved the problem.