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   uint64_t s = 0xdeadbeefU;
  12
  13   if (have_vec) {
  14     m = HCBM;
  15     n = HCBN;
  16     p = HCBK;
  17   } else {
  18     m = CBM;
  19     n = CBN;
  20     p = CBK;
  21   }
  22
  23   t a[m*p];
  24   t b[p*n];
  25   t c[m*n];
  26
  27   for (size_t i = 0; i < m; i++)
  28     for (size_t j = 0; j < p; j++)
  29       a[i*p+j] = (t)(s = lfsr(s));
  30   for (size_t i = 0; i < p; i++)
  31     for (size_t j = 0; j < n; j++)
  32       b[i*n+j] = (t)(s = lfsr(s));
  33   memset(c, 0, m*n*sizeof(c[0]));
  34
  35   size_t instret, cycles;
  36   if (have_vec) {
  37     for (int i = 0; i < R; i++)
  38     {
  39       instret = -rdinstret();
  40       cycles = -rdcycle();
  41       mm_rb_hwacha(m, n, p, a, p, b, n, c, n);
  42       instret += rdinstret();
  43       cycles += rdcycle();
  44     }
  45   } else {
  46     for (int i = 0; i < R; i++)
  47     {
  48       instret = -rdinstret();
  49       cycles = -rdcycle();
  50       mm(m, n, p, a, p, b, n, c, n);
  51       instret += rdinstret();
  52       cycles += rdcycle();
  53     }
  54   }
  55
  56   printf("C%d: reg block %dx%dx%d, cache block %dx%dx%d\n",
  57          cid, RBM, RBN, RBK, CBM, CBN, CBK);
  58   printf("C%d: %d instructions\n", cid, (int)(instret));
  59   printf("C%d: %d cycles\n", cid, (int)(cycles));
  60   printf("C%d: %d flops\n", cid, 2*m*n*p);
  61   printf("C%d: %d Mflops @ 1 GHz\n", cid, 2000*m*n*p/(cycles));
  62
  63 #if 1
  64   for (size_t i = 0; i < m; i++)
  65   {
  66     for (size_t j = 0; j < n; j++)
  67     {
  68       t s = 0;
  69       for (size_t k = 0; k < p; k++)
  70         s += a[i*p+k] * b[k*n+j];
  71       s *= R;
  72       if (fabs(c[i*n+j]-s) > fabs(1e-6*s))
  73       {
  74         printf("C%d: c[%lu][%lu] %f != %f\n", cid, i, j, c[i*n+j], s);
  75         exit(1);
  76       }
  77     }
  78   }
  79 #endif
  80
  81   barrier(nc);
  82   exit(0);
  83 }