mt/ai_matmul/ai_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 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 printArrayMT( 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)) verifyMT(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 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 110 {
 111
 112    // ***************************** //
 113    // **** ADD YOUR CODE HERE ***** //
 114    // ***************************** //
 115    //
 116    // feel free to make a separate function for MI and MSI versions.
 117
 118 //----------MSI--------------
 119 ///*
 120    int i,j,k;
 121    barrier(ncores);
 122    for(j = coreid*lda/ncores; j < coreid*lda/ncores + lda/ncores; j++) {
 123         for(i = 0; i < lda; i+=4) {
 124                 data_t Cval0 = 0;
 125                 data_t Cval1 = 0;
 126                 data_t Cval2 = 0;
 127                 data_t Cval3 = 0;
 128                 for(k = 0; k < lda; k++) {
 129                         Cval0 += A[j*lda+k]*B[k*lda+i];
 130                         Cval1 += A[j*lda+k]*B[k*lda+i+1];
 131                         Cval2 += A[j*lda+k]*B[k*lda+i+2];
 132                         Cval3 += A[j*lda+k]*B[k*lda+i+3];
 133                 }
 134                 C[j*lda+i] = Cval0;
 135                 C[j*lda+i+1] = Cval1;
 136                 C[j*lda+i+2] = Cval2;
 137                 C[j*lda+i+3] = Cval3;
 138         }
 139    }
 140 //*/
 141
 142 //------------------MI-------------------
 143 /*
 144    int i,j,k;
 145    barrier(nc);
 146    for(j = coreid*lda/ncores; j < coreid*lda/ncores + lda/ncores; j++) {
 147         for(i = 0; i < lda; i+=4) {
 148                 data_t Cval0 = 0;
 149                 data_t Cval1 = 0;
 150                 data_t Cval2 = 0;
 151                 data_t Cval3 = 0;
 152                 if(coreid == 0) {
 153                         for(k = 0; k < lda; k++) {
 154                                 Cval0 += A[j*lda+k]*B[k*lda+i];
 155                                 Cval1 += A[j*lda+k]*B[k*lda+i+1];
 156                                 Cval2 += A[j*lda+k]*B[k*lda+i+2];
 157                                 Cval3 += A[j*lda+k]*B[k*lda+i+3];
 158                         }
 159                 } else {
 160                         for(k = lda-1; k >= 0; k--) {
 161                                 Cval0 += A[j*lda+k]*B[k*lda+i];
 162                                 Cval1 += A[j*lda+k]*B[k*lda+i+1];
 163                                 Cval2 += A[j*lda+k]*B[k*lda+i+2];
 164                                 Cval3 += A[j*lda+k]*B[k*lda+i+3];
 165                         }
 166                 }
 167                 C[j*lda+i] = Cval0;
 168                 C[j*lda+i+1] = Cval1;
 169                 C[j*lda+i+2] = Cval2;
 170                 C[j*lda+i+3] = Cval3;
 171         }
 172    }
 173 */
 174 }
 175
 176 //--------------------------------------------------------------------------
 177 // Main
 178 //
 179 // all threads start executing thread_entry(). Use their "coreid" to
 180 // differentiate between threads (each thread is running on a separate core).
 181
 182 void thread_entry(int cid, int nc)
 183 {
 184    coreid = cid;
 185    ncores = nc;
 186
 187    // static allocates data in the binary, which is visible to both threads
 188    static data_t results_data[ARRAY_SIZE];
 189
 190
 191 //   // Execute the provided, naive matmul
 192 //   barrier(nc);
 193 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 194 //
 195 //
 196 //   // verify
 197 //   verifyMT(ARRAY_SIZE, results_data, verify_data);
 198 //
 199 //   // clear results from the first trial
 200 //   size_t i;
 201 //   if (coreid == 0)
 202 //      for (i=0; i < ARRAY_SIZE; i++)
 203 //         results_data[i] = 0;
 204 //   barrier(nc);
 205
 206
 207    // Execute your faster matmul
 208    barrier(nc);
 209    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 210
 211 #ifdef DEBUG
 212    printArrayMT("results:", ARRAY_SIZE, results_data);
 213    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 214 #endif
 215
 216    // verify
 217    verifyMT(ARRAY_SIZE, results_data, verify_data);
 218    barrier(nc);
 219
 220    exit(0);
 221 }
 222