benchmarks/mt-matmul/mt-matmul.c

   1 //**************************************************************************
   2 // Multi-threaded Matrix Multiply benchmark
   3 //--------------------------------------------------------------------------
   4 // TA     : Christopher Celio
   5 // Student:
   6 //
   7 //
   8 // This benchmark multiplies two 2-D arrays together and writes the results to
   9 // a third vector. The input data (and reference data) should be generated
  10 // using the matmul_gendata.pl perl script and dumped to a file named
  11 // dataset.h.
  12
  13
  14 // print out arrays, etc.
  15 //#define DEBUG
  16
  17 //--------------------------------------------------------------------------
  18 // Includes
  19
  20 #include <string.h>
  21 #include <stdlib.h>
  22 #include <stdio.h>
  23
  24
  25 //--------------------------------------------------------------------------
  26 // Input/Reference Data
  27
  28 typedef double data_t;
  29 #include "dataset.h"
  30
  31
  32 //--------------------------------------------------------------------------
  33 // Basic Utilities and Multi-thread Support
  34
  35 __thread unsigned long coreid;
  36
  37 #include "util.h"
  38
  39 #define stringify_1(s) #s
  40 #define stringify(s) stringify_1(s)
  41 #define stats(code) do { \
  42     unsigned long _c = -rdcycle(), _i = -rdinstret(); \
  43     code; \
  44     _c += rdcycle(), _i += rdinstret(); \
  45     if (coreid == 0) \
  46       printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
  47              stringify(code), _c, _c/DIM_SIZE/DIM_SIZE/DIM_SIZE, 10*_c/DIM_SIZE/DIM_SIZE/DIM_SIZE%10, _c/_i, 10*_c/_i%10); \
  48   } while(0)
  49
  50 //--------------------------------------------------------------------------
  51 // matmul function
  52
  53 // single-thread, naive version
  54 void __attribute__((noinline)) matmul_naive(const int lda,  const data_t A[], const data_t B[], data_t C[] )
  55 {
  56    int i, j, k;
  57
  58    if (coreid > 0)
  59       return;
  60
  61    for ( i = 0; i < lda; i++ )
  62       for ( j = 0; j < lda; j++ )
  63       {
  64          for ( k = 0; k < lda; k++ )
  65          {
  66             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
  67          }
  68       }
  69
  70 }
  71
  72
  73
  74 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
  75 {
  76
  77    // ***************************** //
  78    // **** ADD YOUR CODE HERE ***** //
  79    // ***************************** //
  80    //
  81    // feel free to make a separate function for MI and MSI versions.
  82
  83 }
  84
  85 //--------------------------------------------------------------------------
  86 // Main
  87 //
  88 // all threads start executing thread_entry(). Use their "coreid" to
  89 // differentiate between threads (each thread is running on a separate core).
  90
  91 void thread_entry(int cid, int nc)
  92 {
  93    coreid = cid;
  94
  95    // static allocates data in the binary, which is visible to both threads
  96    static data_t results_data[ARRAY_SIZE];
  97
  98
  99    // Execute the provided, naive matmul
 100    barrier(nc);
 101    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 102
 103
 104    // verify
 105    int res = verifyDouble(ARRAY_SIZE, results_data, verify_data);
 106    if (res)
 107       exit(res);
 108
 109 #if 0
 110    // clear results from the first trial
 111    size_t i;
 112    if (coreid == 0)
 113       for (i=0; i < ARRAY_SIZE; i++)
 114          results_data[i] = 0;
 115    barrier(nc);
 116
 117
 118    // Execute your faster matmul
 119    barrier(nc);
 120    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 121
 122 #ifdef DEBUG
 123    printArray("results:", ARRAY_SIZE, results_data);
 124    printArray("verify :", ARRAY_SIZE, verify_data);
 125 #endif
 126
 127    // verify
 128    res = verify(ARRAY_SIZE, results_data, verify_data);
 129    if (res)
 130       exit(res);
 131    barrier(nc);
 132 #endif
 133
 134    exit(0);
 135 }