benchmarks/mm/mm_main.c

   1 #include "common.h"
   2 #include <assert.h>
   3 #include <stdlib.h>
   4 #include <stdio.h>
   5 #include "util.h"
   6
   7 void thread_entry(int cid, int nc)
   8 {
   9   const int R = 8;
  10   int m, n, p;
  11
  12   if (have_vec) {
  13     m = HCBM;
  14     n = HCBN;
  15     p = HCBK;
  16   } else {
  17     m = CBM;
  18     n = CBN;
  19     p = CBK;
  20   }
  21
  22   t a[m*p];
  23   t b[p*n];
  24   t c[m*n];
  25
  26   for (size_t i = 0; i < m; i++)
  27     for (size_t j = 0; j < p; j++)
  28       a[i*p+j] = i+j;
  29   for (size_t i = 0; i < p; i++)
  30     for (size_t j = 0; j < n; j++)
  31       b[i*n+j] = i-j;
  32   memset(c, 0, m*n*sizeof(c[0]));
  33
  34   size_t instret, cycles;
  35   if (have_vec) {
  36     for (int i = 0; i < R; i++)
  37     {
  38       instret = -rdinstret();
  39       cycles = -rdcycle();
  40       mm_rb_hwacha(m, n, p, a, p, b, n, c, n);
  41       instret += rdinstret();
  42       cycles += rdcycle();
  43     }
  44   } else {
  45     for (int i = 0; i < R; i++)
  46     {
  47       instret = -rdinstret();
  48       cycles = -rdcycle();
  49       mm(m, n, p, a, p, b, n, c, n);
  50       instret += rdinstret();
  51       cycles += rdcycle();
  52     }
  53   }
  54
  55   printf("C%d: reg block %dx%dx%d, cache block %dx%dx%d\n",
  56          cid, RBM, RBN, RBK, CBM, CBN, CBK);
  57   printf("C%d: %d instructions\n", cid, (int)(instret));
  58   printf("C%d: %d cycles\n", cid, (int)(cycles));
  59   printf("C%d: %d flops\n", cid, 2*m*n*p);
  60   printf("C%d: %d Mflops @ 1 GHz\n", cid, 2000*m*n*p/(cycles));
  61
  62 #if 1
  63   for (size_t i = 0; i < m; i++)
  64   {
  65     for (size_t j = 0; j < n; j++)
  66     {
  67       t s = 0;
  68       for (size_t aik = i, bkj = -j; aik < i+p; aik++, bkj++)
  69         s += (t)aik*(t)bkj;
  70       if (fabs(c[i*n+j]-s*R) > 1e-6*s)
  71       {
  72         printf("C%d: c[%lu][%lu] %f != %f\n", cid, i, j, c[i*n+j], s);
  73         exit(1);
  74       }
  75     }
  76   }
  77 #endif
  78
  79   barrier(nc);
  80   exit(0);
  81 }