code example
OpenCL Open Computing Language
GPU computing

#include <cl.h>
#include <malloc.h>


// kernel
#define BLOCK (512)
const char *source =
"__kernel void sub1(__global float* fx,\
    __global const float* fy,\
    __local float* se, __global float* fe) {\
  const unsigned int t = get_global_id(0);\
  const unsigned int b = get_group_id(0);\
  const unsigned block = 512;\
  const unsigned int i = block*b+t;\
  float e;\
  /* do computation */\
  fx[t] += ( fy[t+2] + fy[t] )*.5;\
  e = fy[t+1] * fy[t+1];\
  /* reduction */\
  se[t] = e;\
  barrier(CLK_LOCAL_MEM_FENCE);\
  if (t<256) {\
   se[t] += se[t+256];\
   barrier(CLK_LOCAL_MEM_FENCE);\
  }\
  if (t<128) {\
   se[t] += se[t+128];\
   barrier(CLK_LOCAL_MEM_FENCE);\
  }\
  if (t<64) {\
   se[t] += se[t+64];\
   barrier(CLK_LOCAL_MEM_FENCE);\
  }\
  if (t<32) {\
   se[t] += se[t+32];\
   se[t] += se[t+16];\
   se[t] += se[t+8];\
   se[t] += se[t+4];\
   se[t] += se[t+2];\
   se[t] += se[t+1];\
  }\
  if (t==0)\
   fe[b] = se[0];\
}";

int main(int argc, char *argv[]) {
  int n = ...;
  float *x, *y;
  x = new float[n+1];
  y = new float[n+1];

  ... // fill x, y

  // allocate GPU
  cl_context ct = clCreateContextFromType(0, CL_DEVICE_TYPE_GPU, 0, 0, 0);
  size_t ctsize;
  clGetContextInfo(ct, CL_CONTEXT_DEVICES, 0, 0, &ctsize);
  cl_device_id *aDevices = (cl_device_id*)malloc(ctsize);
  clGetContextInfo(ct, CL_CONTEXT_DEVICES, ctsize, aDevices, 0);
  // compile kernel
  cl_program prog = clCreateProgramWithSource(ct, 1, &source, 0, 0);
  clBuildProgram(prog, 0, 0, 0, 0, 0);
  cl_kernel kern = clCreateKernel(prog, "sub1", 0);
  float e = 0;
  cl_command_queue queue = clCreateCommandQueue(ct, aDevices[0], 0, 0);
  // allocate GPU memory
  cl_mem fx = clCreateBuffer(ct, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
  (n-1)*sizeof(cl_float), &x[1], 0);
  cl_mem fy = clCreateBuffer(ct, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
  (n-1+2)*sizeof(cl_float), &y[1-1], 0);
  cl_mem se = clCreateBuffer(ct, CL_MEM_READ_WRITE,
  BLOCK*sizeof(cl_float), 0, 0);
  cl_mem fe = clCreateBuffer(ct, CL_MEM_WRITE_ONLY,
  (n-1)/BLOCK*sizeof(cl_float), 0, 0);
  clSetKernelArg(kern, 0, sizeof(cl_mem), (void *)&fx);
  clSetKernelArg(kern, 1, sizeof(cl_mem), (void *)&fx);
  clSetKernelArg(kern, 2, sizeof(cl_mem), (void *)&se);
  clSetKernelArg(kern, 3, sizeof(cl_mem), (void *)&fe);
  float *d = new float[(n-1)/BLOCK];
  // call GPU
  const unsigned int size = BLOCK;
  const unsigned int dim = n-1+2;
  clEnqueueNDRangeKernel(queue, kern, 1, 0, &dim, &size, 0, 0, 0);
  // copy to host memory
  clEnqueueReadBuffer(queue, fx, CL_TRUE, 0,
  (n-1) * sizeof(cl_float), &x[1], 0, 0, 0);
  clEnqueueReadBuffer(queue, fe, CL_TRUE, 0,
  (n-1) * sizeof(cl_float), d, 0, 0, 0);
  float ee = 0;
  for (int i=0; i<(n-1)/BLOCK; ++i)
   ee += d[i];
  e += ee;
  delete[] d;
  // release GPU memory
  clReleaseMemObject(fx);
  clReleaseMemObject(fy);
  clReleaseMemObject(se);
  clReleaseMemObject(fe);

  ... // output x, e

  delete[] x, y;
  return 0;
}

[start] [references] [download] [install]