streamcopy.br

用于GPU通用计算的编程语言BrookGPU 0.4

#include <stdio.h>
#include <string.h>
#include <assert.h>
#include <math.h>

kernel void copyKernel(float4 src<>, out float4 dst<>) {
  dst = src;
}


void doCopy(float* src, float* dst, int dimx, int dimy) {
  
  float4 srcStream<dimx, dimy>;
  float4 dstStream<dimx, dimy>;
  
  streamRead(srcStream, src);
  copyKernel(srcStream, dstStream);
  streamWrite(dstStream, dst);

}

int fequals(float a, float b) {
  
  float TOLERANCE = 0.05f;
  
  return (fabs(a-b) < TOLERANCE);

}


int main(int argc, char* argv[]) {

  int i,j,k;
  int verbose = 0;
  int success;

  int minx, maxx;
  int miny, maxy;
  int numElements;

  float* src;
  float* dst;

  i=1;
  if (i<argc && strcmp(argv[i], "-v")==0) {
    i++;
    verbose = 1;
  }
  
  if(i+4<argc) {
    fprintf(stderr, "Too few commandline arguments.\n");
    fprintf(stderr, "Usage: %s [-v] minx maxx miny maxy\n", argv[0]);
    exit(1);
  }
  
  minx = atoi(argv[i]);
  maxx = atoi(argv[i+1]);
  miny = atoi(argv[i+2]);
  maxy = atoi(argv[i+3]);

  assert(minx!=0 && maxx!=0 && 
	 miny!=0 && maxy!=0);

  for (j=miny;j<=maxy;j++) {
    for (i=minx;i<=maxx;i++) {

      success = 1;
      numElements = 4 * i * j;

      src = (float*)malloc( sizeof(float) * numElements);
      dst = (float*)malloc( sizeof(float) * numElements);
      
      for (k=0;k<numElements;k++) {
	/* keep elements small, prevent overflow */
	src[k] = (float)(k % 100);
	dst[k] = 0.0f;
      }

      doCopy(src, dst, i, j);

      for (k=0;k<numElements;k++) {

	if (!fequals(src[k], dst[k])) {
	  if (verbose) 
	    printf("[stream size %dx%d] pixel (%d,%d)[%d]: expected %f obtained %f\n",
		    i, j, (k/4) % i, (k/4) / i, k%4, src[k], dst[k]);
	  success = 0;
	}
      }
      
      if (!success)
	printf("stream size %dx%d failed.\n", i, j);
	
     
    }
  }

}