mt/as_matmul/matmul_mi.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 float data_t;
  29 #include "dataset.h"
  30
  31
  32 //--------------------------------------------------------------------------
  33 // Basic Utilities and Multi-thread Support
  34
  35 __thread unsigned long coreid;
  36 unsigned long ncores;
  37
  38 #include "util.h"
  39
  40 #define stringify_1(s) #s
  41 #define stringify(s) stringify_1(s)
  42 #define stats(code) do { \
  43     unsigned long _c = -rdcycle(), _i = -rdinstret(); \
  44     code; \
  45     _c += rdcycle(), _i += rdinstret(); \
  46     if (coreid == 0) \
  47       printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
  48              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); \
  49   } while(0)
  50
  51
  52 //--------------------------------------------------------------------------
  53 // Helper functions
  54
  55 void printArray( char name[], int n, data_t arr[] )
  56 {
  57    int i;
  58    if (coreid != 0)
  59       return;
  60
  61    printf( " %10s :", name );
  62    for ( i = 0; i < n; i++ )
  63       printf( " %3ld ", (long) arr[i] );
  64    printf( "\n" );
  65 }
  66
  67 void __attribute__((noinline)) verify(size_t n, const data_t* test, const data_t* correct)
  68 {
  69    if (coreid != 0)
  70       return;
  71
  72    size_t i;
  73    for (i = 0; i < n; i++)
  74    {
  75       if (test[i] != correct[i])
  76       {
  77          printf("FAILED test[%d]= %3ld, correct[%d]= %3ld\n",
  78             i, (long)test[i], i, (long)correct[i]);
  79          exit(-1);
  80       }
  81    }
  82
  83    return;
  84 }
  85
  86 //--------------------------------------------------------------------------
  87 // matmul function
  88
  89 // single-thread, naive version
  90 void __attribute__((noinline)) matmul_naive(const int lda,  const data_t A[], const data_t B[], data_t C[] )
  91 {
  92    int i, j, k;
  93
  94    if (coreid > 0)
  95       return;
  96
  97    for ( i = 0; i < lda; i++ )
  98       for ( j = 0; j < lda; j++ )
  99       {
 100          for ( k = 0; k < lda; k++ )
 101          {
 102             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 103          }
 104       }
 105
 106 }
 107
 108
 109
 110 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 111 {
 112
 113    // ***************************** //
 114    // **** ADD YOUR CODE HERE ***** //
 115    // ***************************** //
 116    //
 117    // feel free to make a separate function for MI and MSI versions.
 118
 119    int i, j, k, n, m, c1, c2;
 120
 121    //matmul_naive(32, input1_data, input2_data, results_data); barrier(): 952596 cycles, 29.0 cycles/iter, 3.6 CPI
 122    //matmul(32, input1_data, input2_data, results_data); barrier(): 570135 cycles, 17.3 cycles/iter, 3.4 CPI
 123
 124    for ( j = coreid; j < lda; j += 2*ncores ) {
 125       for ( i = 0; i < lda; i += 1 ){
 126          c1 = 0;     //global vars c1, c2
 127          c2 = 0;
 128          for ( k = 0; k < lda; k += 1 ) {
 129             c1 += A[j * lda + k] * B[k*lda + i];
 130             c2 += A[(j+2) * lda + k] * B[k*lda + i];
 131
 132             //barrier();
 133          }
 134
 135          C[i + j * lda] = c1;
 136          C[i + (j+2) * lda] = c2;
 137          barrier();
 138       }
 139    //barrier();
 140    }
 141
 142 }
 143
 144 //--------------------------------------------------------------------------
 145 // Main
 146 //
 147 // all threads start executing thread_entry(). Use their "coreid" to
 148 // differentiate between threads (each thread is running on a separate core).
 149
 150 void thread_entry(int cid, int nc)
 151 {
 152    coreid = cid;
 153    ncores = nc;
 154
 155    // static allocates data in the binary, which is visible to both threads
 156    static data_t results_data[ARRAY_SIZE];
 157
 158
 159 //   // Execute the provided, naive matmul
 160 //   barrier();
 161 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 162 //
 163 //
 164 //   // verify
 165 //   verify(ARRAY_SIZE, results_data, verify_data);
 166 //
 167 //   // clear results from the first trial
 168 //   size_t i;
 169 //   if (coreid == 0)
 170 //      for (i=0; i < ARRAY_SIZE; i++)
 171 //         results_data[i] = 0;
 172 //   barrier();
 173
 174
 175    // Execute your faster matmul
 176    barrier();
 177    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 178
 179 #ifdef DEBUG
 180    printArray("results:", ARRAY_SIZE, results_data);
 181    printArray("verify :", ARRAY_SIZE, verify_data);
 182 #endif
 183
 184    // verify
 185    verify(ARRAY_SIZE, results_data, verify_data);
 186    barrier();
 187
 188    exit(0);
 189 }