I want to have a global counter in OpenCL that can be increased by every Work Item in every Work Group.
In my Kernel I do:
#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
void increase(volatile __global int* counter)
{
atomic_inc(counter);
}
__kernel void test()
{
volatile __global int* counter = 0;
increase(counter);
printf("Counter: %i",&counter);
}
My Hostcode is minimum pyopencl to enqueue the Kernel:
import pyopencl as cl
platform = cl.get_platforms()[0]
devs = platform.get_devices()
device = devs[0]
ctx = cl.Context([device])
queue = cl.CommandQueue(ctx)
mf = cl.mem_flags
f = open('Minimal.cl', 'r')
fstr = "".join(f.readlines())
prg = cl.Program(ctx, fstr).build()
test_knl = prg.test
def f():
cl.enqueue_nd_range_kernel(queue,test_knl,(1,1,2),None)
f()
I expect an output to show a (possibly random ordered) Appearance of "Counter: i"
, where i
is the number of total work items (in my case 2).
Instead I don't get a printout. When I rerun the programm, it fails with
pyopencl.cffi_cl.LogicError: clcreatecontext failed: <unknown error -9999>
killing my IDE (Spyder) completely.
volatile __global int* counter = 0;
Creating a pointer to global memory is not enough. There must to be some global memory storage behind it.
There are two options at least:
1) You can create a program scope variable if you use OpenCL 2.0 implementation:
void increase(volatile __global int* counter)
{
atomic_inc(counter);
}
__global int counter = 0;
__kernel void test()
{
volatile __global int* counterPtr = &counter;
increase(counterPtr); // or increase(&counter);
printf("Counter: %i",*counterPtr);
}
2) Create an OpenCL buffer and pass it by a kernel argument:
void increase(volatile __global int* counter)
{
atomic_inc(counter);
}
__kernel void test(__global int *counterArg)
{
volatile __global int* counterPtr = counterArg;
increase(counterPtr); // or increase(counterArg);
printf("Counter: %i",*counterPtr);
}