OpenCL Interoperability with Host OpenMPΒΆ

The OpenCL APIs are defined to be thread safe. It is therefore safe for multiple threads created in an OpenMP parallel region to enqueue to a shared OpenCL queue. The following example C++ code will create one OpenCL command queue and will enqueue a kernel from multiple threads in an OpenMP parallel region. The number of times the kernel is enqueued is dependent on the default number of threads the OpenMP runtime creates. This will usually be equal to the number of CPU cores in the system. This example uses the C++ functor construct to make a kernel enqueue command appear to be a function call. This example also uses the one Q.finish() command outside the parallel region to effectively wait_all on the asynchronous hello kernel enqueues.:

#define __CL_ENABLE_EXCEPTIONS
#include <CL/cl.hpp>
#include <iostream>
#include <cstdlib>
#include <cstdio>

using namespace cl;
using namespace std;

const char * kernelStr = "kernel void Hello() { printf(\"Hello\\n\"); }";

int main(int argc, char *argv[])
{
   try
   {
     Context             context (CL_DEVICE_TYPE_ACCELERATOR);
     std::vector<Device> devices = context.getInfo<CL_CONTEXT_DEVICES>();
     Program::Sources    source(1, std::make_pair(kernelStr,strlen(kernelStr)));
     Program             program = Program(context, source);
     program.build(devices);
     CommandQueue        Q     (context, devices[0], CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE);
     Kernel              kernel(program, "Hello");
     KernelFunctor hello = kernel.bind(Q, NDRange(1), NDRange(1));

     #pragma omp parallel
         hello();

     Q.finish();
   }
   catch (Error err)
   { cerr << "ERROR: " << err.what() << "(" << err.err() << ")" << endl; }
}

The previous example contained one OpenCL Q that was shared across multiple threads. It is also valid to define multiple OpenCL command queues tied to the same device, where each command Queue is private to the thread. The following example is a slightly modified version of the above illustrating the use of private command queues. Note the definition of the command queue and the functor in the parallel region. Also note that since there will not be a command queue defined in the main thread, one finish() API call to wait on all threads is not possible, so a wait is attached to each kernel enqueue command.:

#define __CL_ENABLE_EXCEPTIONS
#include <CL/cl.hpp>
#include <iostream>
#include <cstdlib>
#include <cstdio>

using namespace cl;
using namespace std;

const char * kernelStr = "kernel void Hello() { printf(\"Hello\\n\"); }";

int main(int argc, char *argv[])
{
   try
   {
     Context             context (CL_DEVICE_TYPE_ACCELERATOR);
     std::vector<Device> devices = context.getInfo<CL_CONTEXT_DEVICES>();
     Program::Sources    source(1, std::make_pair(kernelStr,strlen(kernelStr)));
     Program             program = Program(context, source);
     program.build(devices);
     Kernel              kernel(program, "Hello");

     #pragma omp parallel
     {
         CommandQueue  Q (context, devices[0], CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE);
         KernelFunctor hello = kernel.bind(Q, NDRange(1), NDRange(1));
         hello().wait();
     }
   }
   catch (Error err)
   { cerr << "ERROR: " << err.what() << "(" << err.err() << ")" << endl; }
}