gflops.br

用于GPU通用计算的编程语言BrookGPU 0.4

// sum.br
// The absolute minimal test. It tests
// using a kernel to sum both 1-dimensional
// and two-dimensional input streams.

#include <stdio.h>


void resetTimer(void);
float getTimer(void);

kernel void gflop( float a<>, out float c<> ) {
  c = a;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
}

kernel void gflop2( float2 a<>, out float2 c<> ) {
  c = a;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
}

kernel void gflop3( float3 a<>, out float3 c<> ) {
  c = a;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
}

kernel void gflop4( float4 a<>, out float4 c<> ) {
  c = a;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
  c += a*c;
}

float input_a[1000][1000];
float output_c[1000][1000];
float2 input_a2[1000][1000];
float2 output_c2[1000][1000];
float3 input_a3[1000][1000];
float3 output_c3[1000][1000];
float4 input_a4[1000][1000];
float4 output_c4[1000][1000];

main() {

  float msec;
  int i, j;

  {
    float  a< 1000, 1000 >;
    float  c< 1000, 1000 >;
  
    // initialize data
    for( i = 0; i < 1000; i++ )
      for( j = 0; j < 1000; j++ )
	input_a[i][j] = (float) i*1000+j;

    resetTimer();
    
    streamRead( a, input_a );
    
    fprintf (stderr, "float1: ");
    fflush(stderr);
    
    for (i=0; i<1000; i++) {
      gflop(a, c);
      if (!(i%100)) {
	fprintf (stderr, ".");
	fflush(stderr);
      }
    }
    
    streamWrite( c, output_c );

    msec = getTimer();

    fprintf (stderr, "  %f\n", msec);
  }

  {
    float2 a< 1000, 1000 >;
    float2 c< 1000, 1000 >;

    resetTimer();
    
    streamRead( a, input_a2 );
    
    fprintf (stderr, "float2: ");
    fflush(stderr);

    for (i=0; i<1000; i++) {
      gflop2(a, c);
      if (!(i%100)) {
	fprintf (stderr, ".");
	fflush(stderr);
      }
    }
    
    streamWrite( c, output_c2 );

    msec = getTimer();

    fprintf (stderr, "  %f\n", msec);
  }

  {
    float3 a< 1000, 1000 >;
    float3 c< 1000, 1000 >;

    resetTimer();
    
    streamRead( a, input_a3 );
    
    fprintf (stderr, "float3: ");
    fflush(stderr);

    for (i=0; i<1000; i++) {
      gflop3(a, c);
      if (!(i%100)) {
	fprintf (stderr, ".");
	fflush(stderr);
      }
    }
    
    streamWrite( c, output_c3 );

    msec = getTimer();

    fprintf (stderr, "  %f\n", msec);
  }
  
  {
    float4 a< 1000, 1000 >;
    float4 c< 1000, 1000 >;

    resetTimer();
    
    streamRead( a, input_a4 );
    
    fprintf (stderr, "float4: ");
    fflush(stderr);

    for (i=0; i<1000; i++) {
      gflop4(a, c);
      if (!(i%100)) {
	fprintf (stderr, ".");
	fflush(stderr);
      }
    }
    
    streamWrite( c, output_c4 );

    msec = getTimer();

    fprintf (stderr, "  %f\n", msec);
  }


  {
    float a1< 1000, 1000 >;
    float c1< 1000, 1000 >;
    float4 a4< 1000, 1000 >;
    float4 c4< 1000, 1000 >;

    resetTimer();
    
    streamRead( a1, input_a );
    streamRead( a4, input_a4 );
    
    fprintf (stderr, "float1 & float4: ");
    fflush(stderr);

    for (i=0; i<500; i++) {
      gflop(a1, c1);
      gflop4(a4, c4);
      if (!(i%50)) {
	fprintf (stderr, ".");
	fflush(stderr);
      }
    }
    
    streamWrite( c1, output_c );
    streamWrite( c4, output_c4 );

    msec = getTimer();

    fprintf (stderr, "  %f\n", msec);
  }
  


  return 0;
}