Thursday 9 March 2017

OpenCL Program for Vector / Array Addition

          To learn Parallel Computing with OpenCL, you should start with example of Array Addition as it illustrates the proper use of multi-threading paradigm.
           In this post, we will see OpenCL program for Array / Vector Addition.                        

           In OpenCL, we have to follow the following steps:

1. Get a list of platforms available on your system. (i.e. OpenCL installations on your system).
2. Get a list of devices(cpu or gpu) supported for available OpenCL platforms.
3. Get context properties list.
4. Create a context (collection of GPU and CPU).
5. Create command queue for the context and device.
6. Create a program object for all kernels.
7. Build/Compile the program.
8. Create kernel object for the specific kernel which you want to execute.
9. Load data into the input buffer.
10. Set the arguments for the kernel.
11. Enqueue the kernel for execution.
12. Copy the results from Output buffer to Host(CPU) variable.
13. Print the results.
14. Free OpenCL resources.

            Following program is run on a system which has NVIDIA gpu and CUDA installed on it.
            
            Program code and its output is as follows:
        
 Program: (openclprogram.c)
#include <stdio.h>
#include <stdlib.h>
#include <CL/cl.h>
  
/* Kernel (function to be run on device) */   
const char *code =
 "__kernel void arrayadd(__global int *x, __global int *y, __global int *z)\n"
    "{\n"
    "  size_t id = get_global_id(0);\n"
    "  z[id] = x[id] + y[id];\n"
    "}\n";

int main()
{
  cl_context context;
  cl_context_properties properties[3];
  cl_kernel kernel;
  cl_command_queue commandqueue;
  cl_program program;
  cl_int err;
  cl_uint num_of_platforms;
  cl_platform_id platform_id;
  cl_device_id device_id;
  cl_uint num_of_devices;
  cl_mem buffer1, buffer2, outputbuffer;

  size_t global;
  int arraysize;
  int a[20];
  int b[20];
  int results[20];
  int i;

  printf("Enter the size of arrays\n");
  scanf("%d",&arraysize);

  printf("Enter %d elements of First Array\n",arraysize);
  for(i=0;i<arraysize;i++)
        scanf("%d",&a[i]);

  printf("Enter %d elements of Second Array\n",arraysize);
  for(i=0;i<arraysize;i++)
        scanf("%d",&b[i]);

  /* Get a list of platforms available on your system. (i.e. OpenCL installations on your system).*/
  if (clGetPlatformIDs(1, &platform_id, &num_of_platforms)!= CL_SUCCESS)
   {
    printf("Not getting Platform id\n");
    return 1;
   }
 
  /* Get a list of devices(cpu or gpu) supported for available OpenCL platforms.*/
  if (clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_ALL, 1, &device_id, &num_of_devices) != CL_SUCCESS)
   {
    printf("Not getting Device id\n");
    return 1;
   }
  
   /* Get context properties list.*/
  properties[0]= CL_CONTEXT_PLATFORM;
  properties[1]= (cl_context_properties) platform_id;
  properties[2]= 0;

  /* Create a context (i.e. collection of GPU and CPU).*/
  context = clCreateContext(properties,1,&device_id,NULL,NULL,&err);
 
  /* Create command queue for the context and device.*/
  commandqueue = clCreateCommandQueue(context, device_id, 0, &err);
  
   /* Create a program object for all kernels.*/
   program = clCreateProgramWithSource(context,1,(const char **) &code, NULL, &err);
   
  /* Build/Compile the program.*/
  if (clBuildProgram(program, 0, NULL, NULL, NULL, NULL) != CL_SUCCESS)
   {
    printf("Error during building program\n");
    return 1;
   }
    
  /* Create kernel object for the specific kernel which you want to execute.*/
  kernel = clCreateKernel(program, "arrayadd", &err);


  buffer1 = clCreateBuffer(context, CL_MEM_READ_ONLY, sizeof(int) * arraysize, NULL, NULL);
  buffer2 = clCreateBuffer(context, CL_MEM_READ_ONLY, sizeof(int) * arraysize, NULL, NULL);
  outputbuffer = clCreateBuffer(context, CL_MEM_WRITE_ONLY, sizeof(int) * arraysize, NULL, NULL);  

  /* Load data into the input buffer.*/
  clEnqueueWriteBuffer(commandqueue, buffer1, CL_TRUE, 0, sizeof(int) * arraysize, a, 0, NULL, NULL);
  clEnqueueWriteBuffer(commandqueue, buffer2, CL_TRUE, 0, sizeof(int) * arraysize, b, 0, NULL, NULL);
 
  /* Set the arguments for the kernel.*/
  clSetKernelArg(kernel, 0, sizeof(cl_mem), &buffer1);
  clSetKernelArg(kernel, 1, sizeof(cl_mem), &buffer2);
  clSetKernelArg(kernel, 2, sizeof(cl_mem), &outputbuffer);
  
  global=arraysize;
   
  /* Enqueue the kernel for execution. */
  clEnqueueNDRangeKernel(commandqueue, kernel, 1, NULL, &global, NULL, 0, NULL, NULL);
  clFinish(commandqueue);

  /* Copy the results from Output buffer to Host(CPU) variable. */
  clEnqueueReadBuffer(commandqueue, outputbuffer, CL_TRUE, 0, sizeof(int) *arraysize, results, 0, NULL, NULL);

  /* Print the results. */
  printf("Addition of Two Arrays is as follows: \n");
 
  for(i=0;i<arraysize; i++)
   {
    printf("%d\n",results[i]);
   }
  
 /* Free OpenCL resources. */
  clReleaseMemObject(buffer1);
  clReleaseMemObject(buffer2);
  clReleaseMemObject(outputbuffer);
  clReleaseProgram(program);
  clReleaseKernel(kernel);
  clReleaseCommandQueue(commandqueue);
  clReleaseContext(context);
   
  return 0;
}

Output:
>>> gcc openclprogram.c -I /usr/local/cuda/include/ -L /usr/local/cuda/lib64/ -lOpenCL
>>> ./a.out
Enter the size of arrays
5
Enter 5 elements of First Array
4 2 7 3 8
Enter 5 elements of Second Array
9 10 22 3 6
Addition of Two Arrays is as follows:
13
12
29
6
14








No comments:

Post a Comment