Question

Is there any example of using clEnqueueNativeKernel in OpenCL? In this way one can write a kernel in a c or c++ language. Do other commands remain unchanged?

Answer 1

Native C++ “kernels” are essentially just functions that you want to execute within command queue to preserve order of commands.

AFAIK they are not supported on GPU. If you want to execute C++ functions across all devices, you should consider to use cl_event callbacks (when status == CL_COMPLETE).

Assume you have a buffer object you want to read from device and pass to your C++ function. Also you want to pass some integer value as well
(I use C++ OpenCL wrapper):

// First of all, we should define a struct which describes our arguments list.
struct Arguments
{
    int integer;
    void* buffer_host;
};

// Define C/C++ function you want to call.
void CFunction(void *args)
{
    Arguments args = reinterpret_cast<Arguments*>(args);
    // Do something with args->integer and args->buffer_host.
}

// ...
Arguments args = {.integer = 0, .buffer_host = NULL};

// First, we should define Buffer objects in arguments.
std::vector<cl::Memory> buffers_dev;
buffers_dev.push_back(a_buffer);

// Then we should define pointers to *pointer in args* which will be set
// when OpenCL read data from buffers_dev to the host memory.
std::vector<const void*> buffers_host;
buffers_host.push_back(&args.buffer_host);

// Finally, set integer
args.integer = 10;

queue.enqueueNativeKernel(CFunction,
                          std::make_pair(&args, siezof(Arguments)),
                          &buffers_dev,
                          &buffers_host);

// At this point args were copied by OpenCL and you may reuse or delete it.