mt/bm_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 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
 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         int i, j, k;
 119     int space=lda/ncores;
 120     int max= space*coreid+space;
 121         static data_t B1[32*32];
 122         if (coreid==ncores-1){
 123                 for (i=0; i<lda*lda/2;i++)
 124                 {
 125                         B1[i]=B[i];
 126                 }
 127         }
 128         else{
 129                 for (i=lda*lda/2;i<lda*lda;i++)
 130                         B1[i]=B[i];
 131         }
 132         data_t temp=0;
 133         data_t temp1=0;
 134         data_t temp2=0;
 135         data_t temp3=0;
 136         data_t tempB=0;
 137
 138         data_t temp_1=0;
 139         data_t temp1_1=0;
 140         data_t temp2_1=0;
 141         data_t temp3_1=0;
 142         data_t tempB_1=0;
 143
 144         data_t temp_2=0;
 145         data_t temp1_2=0;
 146         data_t temp2_2=0;
 147         data_t temp3_2=0;
 148         data_t tempB_2=0;
 149
 150         data_t temp_3=0;
 151         data_t temp1_3=0;
 152         data_t temp2_3=0;
 153         data_t temp3_3=0;
 154         data_t tempB_3=0;
 155         barrier(nc);
 156         if (coreid!=ncores-1){
 157                 for (i=space*coreid;i<max/4*4;i+=4)
 158                 {
 159                         for(j=0;j<lda/4*4;j+=4)
 160                         {
 161                                 temp=C[j+i*lda];
 162                                 temp1=C[j+(i+1)*lda];
 163                                 temp2=C[j+(i+2)*lda];
 164                                 temp3=C[j+(i+3)*lda];
 165                                 temp_1=C[j+1+i*lda];
 166                                 temp1_1=C[j+1+(i+1)*lda];
 167                                 temp2_1=C[j+1+(i+2)*lda];
 168                                 temp3_1=C[j+1+(i+3)*lda];
 169                                 temp_2=C[j+2+i*lda];
 170                                 temp1_2=C[j+2+(i+1)*lda];
 171                                 temp2_2=C[j+2+(i+2)*lda];
 172                                 temp3_2=C[j+2+(i+3)*lda];
 173                                 temp_3=C[j+3+i*lda];
 174                                 temp1_3=C[j+3+(i+1)*lda];
 175                                 temp2_3=C[j+3+(i+2)*lda];
 176                                 temp3_3=C[j+3+(i+3)*lda];
 177                                 for (k=0;k<lda;k++)
 178                                 {
 179                                         tempB=B[j+k*lda];
 180                                         temp+=A[k+i*lda]*tempB;
 181                                         temp1+=A[k+(i+1)*lda]*tempB;
 182                                         temp2+=A[k+(i+2)*lda]*tempB;
 183                                         temp3+=A[k+(i+3)*lda]*tempB;
 184
 185                                         tempB_1=B[j+1+k*lda];
 186                                         temp_1+=A[k+i*lda]*tempB_1;
 187                                         temp1_1+=A[k+(i+1)*lda]*tempB_1;
 188                                         temp2_1+=A[k+(i+2)*lda]*tempB_1;
 189                                         temp3_1+=A[k+(i+3)*lda]*tempB_1;
 190
 191                                         tempB_2=B[j+2+k*lda];
 192                                         temp_2+=A[k+i*lda]*tempB_2;
 193                                         temp1_2+=A[k+(i+1)*lda]*tempB_2;
 194                                         temp2_2+=A[k+(i+2)*lda]*tempB_2;
 195                                         temp3_2+=A[k+(i+3)*lda]*tempB_2;
 196
 197                                         tempB_3=B[j+3+k*lda];
 198                                         temp_3+=A[k+i*lda]*tempB_3;
 199                                         temp1_3+=A[k+(i+1)*lda]*tempB_3;
 200                                         temp2_3+=A[k+(i+2)*lda]*tempB_3;
 201                                         temp3_3+=A[k+(i+3)*lda]*tempB_3;
 202                                 }
 203                                 C[j+i*lda]=temp;
 204                                 C[j+(i+1)*lda]=temp1;
 205                                 C[j+(i+2)*lda]=temp2;
 206                                 C[j+(i+3)*lda]=temp3;
 207
 208                                 C[j+1+i*lda]=temp_1;
 209                                 C[j+1+(i+1)*lda]=temp1_1;
 210                                 C[j+1+(i+2)*lda]=temp2_1;
 211                                 C[j+1+(i+3)*lda]=temp3_1;
 212
 213                                 C[j+2+i*lda]=temp_2;
 214                                 C[j+2+(i+1)*lda]=temp1_2;
 215                                 C[j+2+(i+2)*lda]=temp2_2;
 216                                 C[j+2+(i+3)*lda]=temp3_2;
 217
 218                                 C[j+3+i*lda]=temp_3;
 219                                 C[j+3+(i+1)*lda]=temp1_3;
 220                                 C[j+3+(i+2)*lda]=temp2_3;
 221                                 C[j+3+(i+3)*lda]=temp3_3;
 222
 223                         }
 224                 }
 225         }
 226         else{
 227                 for (i=space*coreid;i<lda/4*4;i+=4)
 228                 {
 229                         for(j=0;j<lda/4*4;j+=4)
 230                         {
 231                                 temp=C[j+i*lda];
 232                                 temp1=C[j+(i+1)*lda];
 233                                 temp2=C[j+(i+2)*lda];
 234                                 temp3=C[j+(i+3)*lda];
 235                                 temp_1=C[j+1+i*lda];
 236                                 temp1_1=C[j+1+(i+1)*lda];
 237                                 temp2_1=C[j+1+(i+2)*lda];
 238                                 temp3_1=C[j+1+(i+3)*lda];
 239                                 temp_2=C[j+2+i*lda];
 240                                 temp1_2=C[j+2+(i+1)*lda];
 241                                 temp2_2=C[j+2+(i+2)*lda];
 242                                 temp3_2=C[j+2+(i+3)*lda];
 243                                 temp_3=C[j+3+i*lda];
 244                                 temp1_3=C[j+3+(i+1)*lda];
 245                                 temp2_3=C[j+3+(i+2)*lda];
 246                                 temp3_3=C[j+3+(i+3)*lda];
 247                                 for (k=0;k<lda;k++)
 248                                 {
 249                                         tempB=B1[j+k*lda];
 250                                         temp+=A[k+i*lda]*tempB;
 251                                         temp1+=A[k+(i+1)*lda]*tempB;
 252                                         temp2+=A[k+(i+2)*lda]*tempB;
 253                                         temp3+=A[k+(i+3)*lda]*tempB;
 254
 255                                         tempB_1=B1[j+1+k*lda];
 256                                         temp_1+=A[k+i*lda]*tempB_1;
 257                                         temp1_1+=A[k+(i+1)*lda]*tempB_1;
 258                                         temp2_1+=A[k+(i+2)*lda]*tempB_1;
 259                                         temp3_1+=A[k+(i+3)*lda]*tempB_1;
 260
 261                                         tempB_2=B1[j+2+k*lda];
 262                                         temp_2+=A[k+i*lda]*tempB_2;
 263                                         temp1_2+=A[k+(i+1)*lda]*tempB_2;
 264                                         temp2_2+=A[k+(i+2)*lda]*tempB_2;
 265                                         temp3_2+=A[k+(i+3)*lda]*tempB_2;
 266
 267                                         tempB_3=B1[j+3+k*lda];
 268                                         temp_3+=A[k+i*lda]*tempB_3;
 269                                         temp1_3+=A[k+(i+1)*lda]*tempB_3;
 270                                         temp2_3+=A[k+(i+2)*lda]*tempB_3;
 271                                         temp3_3+=A[k+(i+3)*lda]*tempB_3;
 272                                 }
 273                                 C[j+i*lda]=temp;
 274                                 C[j+(i+1)*lda]=temp1;
 275                                 C[j+(i+2)*lda]=temp2;
 276                                 C[j+(i+3)*lda]=temp3;
 277
 278                                 C[j+1+i*lda]=temp_1;
 279                                 C[j+1+(i+1)*lda]=temp1_1;
 280                                 C[j+1+(i+2)*lda]=temp2_1;
 281                                 C[j+1+(i+3)*lda]=temp3_1;
 282
 283                                 C[j+2+i*lda]=temp_2;
 284                                 C[j+2+(i+1)*lda]=temp1_2;
 285                                 C[j+2+(i+2)*lda]=temp2_2;
 286                                 C[j+2+(i+3)*lda]=temp3_2;
 287
 288                                 C[j+3+i*lda]=temp_3;
 289                                 C[j+3+(i+1)*lda]=temp1_3;
 290                                 C[j+3+(i+2)*lda]=temp2_3;
 291                                 C[j+3+(i+3)*lda]=temp3_3;
 292
 293                         }
 294                 }
 295         }
 296
 297
 298
 299
 300 }
 301
 302 //--------------------------------------------------------------------------
 303 // Main
 304 //
 305 // all threads start executing thread_entry(). Use their "coreid" to
 306 // differentiate between threads (each thread is running on a separate core).
 307
 308 void thread_entry(int cid, int nc)
 309 {
 310    coreid = cid;
 311    ncores = nc;
 312
 313    // static allocates data in the binary, which is visible to both threads
 314    static data_t results_data[ARRAY_SIZE];
 315
 316
 317 //   // Execute the provided, naive matmul
 318 //   barrier(nc);
 319 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 320 //
 321 //
 322 //   // verify
 323 //   verifyMT(ARRAY_SIZE, results_data, verify_data);
 324 //
 325 //   // clear results from the first trial
 326 //   size_t i;
 327 //   if (coreid == 0)
 328 //      for (i=0; i < ARRAY_SIZE; i++)
 329 //         results_data[i] = 0;
 330 //   barrier(nc);
 331
 332
 333    // Execute your faster matmul
 334    barrier(nc);
 335    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 336
 337 #ifdef DEBUG
 338    printArrayMT("results:", ARRAY_SIZE, results_data);
 339    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 340 #endif
 341
 342    // verify
 343    verifyMT(ARRAY_SIZE, results_data, verify_data);
 344    barrier(nc);
 345
 346    exit(0);
 347 }
 348