benchmarks/mm/mm_main.c

   1 // See LICENSE for license details.
   2
   3 #include "common.h"
   4 #include <assert.h>
   5 #include <stdlib.h>
   6 #include <stdio.h>
   7 #include "util.h"
   8
   9 void thread_entry(int cid, int nc)
  10 {
  11   const int R = 8;
  12   int m, n, p;
  13   uint64_t s = 0xdeadbeefU;
  14
  15   m = CBM;
  16   n = CBN;
  17   p = CBK;
  18
  19   t a[m*p];
  20   t b[p*n];
  21   t c[m*n];
  22
  23   for (size_t i = 0; i < m; i++)
  24     for (size_t j = 0; j < p; j++)
  25       a[i*p+j] = (t)(s = lfsr(s));
  26   for (size_t i = 0; i < p; i++)
  27     for (size_t j = 0; j < n; j++)
  28       b[i*n+j] = (t)(s = lfsr(s));
  29   memset(c, 0, m*n*sizeof(c[0]));
  30
  31   size_t instret, cycles;
  32   for (int i = 0; i < R; i++)
  33   {
  34     instret = -read_csr(minstret);
  35     cycles = -read_csr(mcycle);
  36     mm(m, n, p, a, p, b, n, c, n);
  37     instret += read_csr(minstret);
  38     cycles += read_csr(mcycle);
  39   }
  40
  41   asm volatile("fence");
  42
  43   printf("C%d: reg block %dx%dx%d, cache block %dx%dx%d\n",
  44          cid, RBM, RBN, RBK, CBM, CBN, CBK);
  45   printf("C%d: %d instructions\n", cid, (int)(instret));
  46   printf("C%d: %d cycles\n", cid, (int)(cycles));
  47   printf("C%d: %d flops\n", cid, 2*m*n*p);
  48   printf("C%d: %d Mflops @ 1 GHz\n", cid, 2000*m*n*p/(cycles));
  49
  50 #if 1
  51   for (size_t i = 0; i < m; i++)
  52   {
  53     for (size_t j = 0; j < n; j++)
  54     {
  55       t s = 0;
  56       for (size_t k = 0; k < p; k++)
  57         s += a[i*p+k] * b[k*n+j];
  58       s *= R;
  59       if (fabs(c[i*n+j]-s) > fabs(1e-6*s))
  60       {
  61         printf("C%d: c[%lu][%lu] %f != %f\n", cid, i, j, c[i*n+j], s);
  62         exit(1);
  63       }
  64     }
  65   }
  66 #endif
  67
  68   barrier(nc);
  69   exit(0);
  70 }