OpenCL Local memory and Xcode - xcode6

Question

I'm trying to learn OpenCL on a Mac, which appears to have some differences in implementation from the OpenCL book I'm reading. I want to be able to dynamically allocate local memory on the GPU. What I'm reading is I need to use the clSetKernelArg function, but that doesn't work within Xcode 6.4. Here's the code as it stands (never mind it's a pointless program, just trying to learn the syntax for shared memory). In Xcode, the kernel is written as a stand-alone .cl file similar to CUDA, so that's a separate file.
add.cl:
kernel void add(int a, int b, global int* c, local int* d)
{
d[0] = a;
d[1] = b;
*c = d[0] + d[1];
}
main.c:
#include <stdio.h>
#include <OpenCL/opencl.h>
#include "add.cl.h"
int main(int argc, const char * argv[]) {
int a = 3;
int b = 5;
int c;
int* cptr = &c;
dispatch_queue_t queue = gcl_create_dispatch_queue(CL_DEVICE_TYPE_GPU, NULL);
void* dev_c = gcl_malloc(sizeof(cl_int), NULL, CL_MEM_WRITE_ONLY);
// attempt to create local memory buffer
void* dev_d = gcl_malloc(2*sizeof(cl_int), NULL, CL_MEM_READ_WRITE);
// clSetKernelArg(add_kernel, 3, 2*sizeof(cl_int), NULL);
dispatch_sync(queue, ^{
cl_ndrange range = { 1, {0, 0, 0}, {1, 0, 0}, {1, 0, 0} };
// This gives a warning:
// Warning: Incompatible pointer to integer conversion passing 'cl_int *'
// (aka 'int *') to parameter of type 'size_t' (aka 'unsigned long')
add_kernel(&range, a, b, (cl_int*)dev_c, (cl_int*)dev_d);
gcl_memcpy((void*)cptr, dev_c, sizeof(cl_int));
});
printf("%d + %d = %d\n", a, b, c);
gcl_free(dev_c);
dispatch_release(queue);
return 0;
}
I've tried putting clSetKernelArg where indicated and it doesn't like the first argument:
Error: Passing 'void (^)(const cl_ndrange *, cl_int, cl_int, cl_int *, size_t)' to parameter of incompatible type 'cl_kernel' (aka 'struct _cl_kernel *')
I've looked and looked but can't find any examples illustrating this point within the Xcode environment. Can you point me in the right direction?

Answer 1

Managed to solve this by ditching Apple's extensions and using standard OpenCL 1.2 calls. That means replacing gcl_malloc with clCreateBuffer, replacing dispatch_sync with clEnqueueNDRangeKernel, and most importantly, using clSetKernelArg with NULL in the last argument for local variables. Works like a charm.
Here's the new version:
char kernel_add[1024] =
"kernel void add(int a, int b, global int* c, local int* d) \
{\
d[0] = a;\
d[1] = b;\
*c = d[0] + d[1];\
}";
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <OpenCL/opencl.h>
int main(int argc, const char * argv[]) {
int a = 3;
int b = 5;
int c;
cl_device_id device_id;
int err = clGetDeviceIDs(NULL, CL_DEVICE_TYPE_GPU, 1, &device_id, NULL);
cl_context context = clCreateContext(0, 1, &device_id, NULL, NULL, &err);
cl_command_queue queue = clCreateCommandQueue(context, device_id, 0, &err);
const char* srccode = kernel;
cl_program program = clCreateProgramWithSource(context, 1, &srccode, NULL, &err);
err = clBuildProgram(program, 0, NULL, NULL, NULL, NULL);
cl_kernel kernel = clCreateKernel(program, "kernel_add", &err);
cl_mem dev_c = clCreateBuffer(context, CL_MEM_WRITE_ONLY, sizeof(int), NULL, NULL);
err = clSetKernelArg(kernel, 0, sizeof(int), &a);
err |= clSetKernelArg(kernel, 1, sizeof(int), &b);
err |= clSetKernelArg(kernel, 2, sizeof(cl_mem), &dev_c);
err |= clSetKernelArg(kernel, 3, sizeof(int), NULL);
size_t one = 1;
err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &one, NULL, 0, NULL, NULL);
clFinish(queue);
err = clEnqueueReadBuffer(queue, dev_c, true, 0, sizeof(int), &c, 0, NULL, NULL);
clReleaseMemObject(dev_c);
clReleaseKernel(kernel);
clReleaseProgram(program);
clReleaseCommandQueue(queue);
clReleaseContext(context);
return 0;
}

Answer 2

In regular OpenCL, for a kernel parameter declared as a local pointer, you don't allocate a host buffer and pass it in (like you're doing with dev_d). Instead you do a clSetKernelArg with the size of the desired local storage but a NULL pointer (like this: clSetKernelArg(kernel, 2, sizeof(cl_int) * local_work_size[0], NULL)). You'll have to translate that into the Xcode way if you insist on being platform-specific.