benchmarks/mm/mm_main.c

   1 #include "common.h"
   2 #include <assert.h>
   3 #include <stdlib.h>
   4 #include "util.h"
   5
   6 void thread_entry(int cid, int nc)
   7 {
   8   const int R = 8;
   9   int m, n, p;
  10
  11   if (have_vec) {
  12     m = HCBM;
  13     n = HCBN;
  14     p = HCBK;
  15   } else {
  16     m = CBM;
  17     n = CBN;
  18     p = CBK;
  19   }
  20
  21   t a[m*p];
  22   t b[p*n];
  23   t c[m*n];
  24
  25   for (size_t i = 0; i < m; i++)
  26     for (size_t j = 0; j < p; j++)
  27       a[i*p+j] = i+j;
  28   for (size_t i = 0; i < p; i++)
  29     for (size_t j = 0; j < n; j++)
  30       b[i*n+j] = i-j;
  31   memset(c, 0, m*n*sizeof(c[0]));
  32
  33   size_t instret, cycles;
  34   if (have_vec) {
  35     for (int i = 0; i < R; i++)
  36     {
  37       instret = -rdinstret();
  38       cycles = -rdcycle();
  39       mm_rb_hwacha(m, n, p, a, p, b, n, c, n);
  40       instret += rdinstret();
  41       cycles += rdcycle();
  42     }
  43   } else {
  44     for (int i = 0; i < R; i++)
  45     {
  46       instret = -rdinstret();
  47       cycles = -rdcycle();
  48       mm(m, n, p, a, p, b, n, c, n);
  49       instret += rdinstret();
  50       cycles += rdcycle();
  51     }
  52   }
  53
  54   printf("C%d: reg block %dx%dx%d, cache block %dx%dx%d\n",
  55          cid, RBM, RBN, RBK, CBM, CBN, CBK);
  56   printf("C%d: %d instructions\n", cid, (int)(instret));
  57   printf("C%d: %d cycles\n", cid, (int)(cycles));
  58   printf("C%d: %d flops\n", cid, 2*m*n*p);
  59   printf("C%d: %d Mflops @ 1 GHz\n", cid, 2000*m*n*p/(cycles));
  60
  61 #if 1
  62   for (size_t i = 0; i < m; i++)
  63   {
  64     for (size_t j = 0; j < n; j++)
  65     {
  66       t s = 0;
  67       for (size_t aik = i, bkj = -j; aik < i+p; aik++, bkj++)
  68         s += (t)aik*(t)bkj;
  69       if (fabs(c[i*n+j]-s*R) > 1e-6*s)
  70       {
  71         printf("C%d: c[%lu][%lu] %f != %f\n", cid, i, j, c[i*n+j], s);
  72         exit(1);
  73       }
  74     }
  75   }
  76 #endif
  77
  78   barrier(nc);
  79   exit(0);
  80 }