#include <cstdlib>
#include <cstdio>
#include <CL/CL.h>

//#define BUILD_PROGRAM
//#define CREATE_KERNEL

void main()
{
	const char* kernelSource = "kernel void K(global read_only float* buffIn, global write_only float* buffOut){int index = get_global_id( 0 ); buffOut[ index ] = buffIn[ index ];}";
	const char* kernelName = "K";
	cl_int error;

	cl_uint platformCount;
	clGetPlatformIDs( 0, 0, &platformCount );
	cl_platform_id *platformList = new cl_platform_id[ platformCount ];
	clGetPlatformIDs( platformCount, platformList, 0 );

	size_t paramSize;
	clGetPlatformInfo( platformList[0], CL_PLATFORM_NAME, 0, 0, &paramSize );
	char *platformName = new char[paramSize];
	clGetPlatformInfo( platformList[0], CL_PLATFORM_NAME, paramSize, platformName, 0 );
	
	cl_uint deviceCount;
	clGetDeviceIDs( platformList[0], CL_DEVICE_TYPE_DEFAULT, 0, 0, &deviceCount );
	cl_device_id *deviceList = new cl_device_id[ deviceCount ];
	clGetDeviceIDs( platformList[0], CL_DEVICE_TYPE_DEFAULT, deviceCount, deviceList, 0 );

	cl_context_properties contextPropertyList[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)platformList[0], 0 };
	cl_context context = clCreateContext( contextPropertyList, deviceCount, deviceList, 0, 0, 0 );

	size_t bufferSize = 100;
	float *arrayIn = new float[bufferSize];	
	for( int i = 0; i < bufferSize; i++ ) arrayIn[i] = i;
	float *arrayOut = new float[bufferSize];
	
	cl_mem bufferIn = clCreateBuffer( context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, bufferSize * sizeof(float), arrayIn, 0 );
	cl_mem bufferOut = clCreateBuffer( context, CL_MEM_WRITE_ONLY, bufferSize * sizeof(float), 0, 0 );

	cl_command_queue queue = clCreateCommandQueue( context, deviceList[0], 0, 0 );

	cl_program program = clCreateProgramWithSource( context, 1, &kernelSource, 0, 0 );

#ifdef BUILD_PROGRAM
	while( true )
#endif
	clBuildProgram( program, 0, 0, 0, 0, 0 ); // leaks

#ifdef CREATE_KERNEL
	while( true ){
#endif
	cl_kernel kernel = clCreateKernel( program, kernelName, 0 ); // leaks
#ifdef CREATE_KERNEL
	clReleaseKernel( kernel );}
#endif
	
	clSetKernelArg( kernel, 0, sizeof( cl_mem ), &bufferIn );
	clSetKernelArg( kernel, 1, sizeof( cl_mem ), &bufferOut );

	size_t globalWorkSize[] = { bufferSize };
	
	while( true )
	{
		clEnqueueNDRangeKernel( queue, kernel, 1, 0, globalWorkSize, 0, 0, 0, 0 );
		clEnqueueReadBuffer( queue, bufferOut, (cl_bool)1, 0, bufferSize * sizeof(float), arrayOut, 0, 0, 0 );
	}

#ifndef CREATE_KERNEL
	clReleaseKernel( kernel );
#endif
	clReleaseProgram( program );
	clReleaseCommandQueue( queue );
	clReleaseMemObject( bufferOut );
	clReleaseMemObject( bufferIn );
	clReleaseContext( context );
	delete[] arrayOut;
	delete[] arrayIn;
	delete[] deviceList;
	delete[] platformName;
	delete[] platformList;
}