mt/bm_matmul/bm_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 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     int i, j, k;
 119     int space=lda/ncores;
 120     int max= space*coreid+space;
 121         data_t temp=0;
 122
 123         data_t temp1=0;
 124         data_t temp2=0;
 125         data_t temp3=0;
 126         data_t temp4=0;
 127
 128         data_t temp_1=0;
 129
 130         data_t temp1_1=0;
 131         data_t temp2_1=0;
 132         data_t temp3_1=0;
 133         data_t temp4_1=0;
 134
 135         data_t temp_2=0;
 136
 137         data_t temp1_2=0;
 138         data_t temp2_2=0;
 139         data_t temp3_2=0;
 140         data_t temp4_2=0;
 141
 142         data_t temp_3=0;
 143
 144         data_t temp1_3=0;
 145         data_t temp2_3=0;
 146         data_t temp3_3=0;
 147         data_t temp4_3=0;
 148
 149         if (coreid!=ncores-1){
 150         //main loop
 151                 for (i=space*coreid;i<max/4*4;i+=4)
 152                 {
 153                         for(j=0;j<lda;j+=4)
 154                         {
 155                                 temp1=C[j+i*lda];
 156                                 temp2=C[j+1+i*lda];
 157                                 temp3=C[j+2+i*lda];
 158                                 temp4=C[j+3+i*lda];
 159
 160                                 temp1_1=C[j+(i+1)*lda];
 161                                 temp2_1=C[j+1+(i+1)*lda];
 162                                 temp3_1=C[j+2+(i+1)*lda];
 163                                 temp4_1=C[j+3+(i+1)*lda];
 164
 165                                 temp1_2=C[j+(i+2)*lda];
 166                                 temp2_2=C[j+1+(i+2)*lda];
 167                                 temp3_2=C[j+2+(i+2)*lda];
 168                                 temp4_2=C[j+3+(i+2)*lda];
 169
 170                                 temp1_3=C[j+(i+3)*lda];
 171                                 temp2_3=C[j+1+(i+3)*lda];
 172                                 temp3_3=C[j+2+(i+3)*lda];
 173                                 temp4_3=C[j+3+(i+3)*lda];
 174                                 for (k=0;k<lda;k++)
 175                                 {
 176                                         temp=A[k+i*lda];
 177                                         temp1+=temp*B[j+k*lda];
 178                                         temp2+=temp*B[j+1+k*lda];
 179                                         temp3+=temp*B[j+2+k*lda];
 180                                         temp4+=temp*B[j+3+k*lda];
 181
 182                                         temp_1=A[k+(i+1)*lda];
 183                                         temp1_1+=temp_1*B[j+k*lda];
 184                                         temp2_1+=temp_1*B[j+1+k*lda];
 185                                         temp3_1+=temp_1*B[j+2+k*lda];
 186                                         temp4_1+=temp_1*B[j+3+k*lda];
 187
 188                                         temp_2=A[k+(i+2)*lda];
 189                                         temp1_2+=temp_2*B[j+k*lda];
 190                                         temp2_2+=temp_2*B[j+1+k*lda];
 191                                         temp3_2+=temp_2*B[j+2+k*lda];
 192                                         temp4_2+=temp_2*B[j+3+k*lda];
 193
 194                                         temp_3=A[k+(i+3)*lda];
 195                                         temp1_3+=temp_3*B[j+k*lda];
 196                                         temp2_3+=temp_3*B[j+1+k*lda];
 197                                         temp3_3+=temp_3*B[j+2+k*lda];
 198                                         temp4_3+=temp_3*B[j+3+k*lda];
 199
 200                                 }
 201                                 C[j+i*lda]=temp1;
 202                                 C[j+1+i*lda]=temp2;
 203                                 C[j+2+i*lda]=temp3;
 204                                 C[j+3+i*lda]=temp4;
 205
 206                                 C[j+(i+1)*lda]=temp1_1;
 207                                 C[j+1+(i+1)*lda]=temp2_1;
 208                                 C[j+2+(i+1)*lda]=temp3_1;
 209                                 C[j+3+(i+1)*lda]=temp4_1;
 210
 211                                 C[j+(i+2)*lda]=temp1_2;
 212                                 C[j+1+(i+2)*lda]=temp2_2;
 213                                 C[j+2+(i+2)*lda]=temp3_2;
 214                                 C[j+3+(i+2)*lda]=temp4_2;
 215
 216                                 C[j+(i+3)*lda]=temp1_3;
 217                                 C[j+1+(i+3)*lda]=temp2_3;
 218                                 C[j+2+(i+3)*lda]=temp3_3;
 219                                 C[j+3+(i+3)*lda]=temp4_3;
 220
 221                         }
 222
 223                 }
 224
 225
 226
 227         }
 228
 229         //second core
 230         else{
 231                 for (i=space*coreid;i<lda/4*4;i+=4)
 232                 {
 233                         for(j=0;j<lda;j+=4)
 234                         {
 235                                 temp1=C[j+i*lda];
 236                                 temp2=C[j+1+i*lda];
 237                                 temp3=C[j+2+i*lda];
 238                                 temp4=C[j+3+i*lda];
 239
 240                                 temp1_1=C[j+(i+1)*lda];
 241                                 temp2_1=C[j+1+(i+1)*lda];
 242                                 temp3_1=C[j+2+(i+1)*lda];
 243                                 temp4_1=C[j+3+(i+1)*lda];
 244
 245                                 temp1_2=C[j+(i+2)*lda];
 246                                 temp2_2=C[j+1+(i+2)*lda];
 247                                 temp3_2=C[j+2+(i+2)*lda];
 248                                 temp4_2=C[j+3+(i+2)*lda];
 249
 250                                 temp1_3=C[j+(i+3)*lda];
 251                                 temp2_3=C[j+1+(i+3)*lda];
 252                                 temp3_3=C[j+2+(i+3)*lda];
 253                                 temp4_3=C[j+3+(i+3)*lda];
 254                                 for (k=0;k<lda;k++)
 255                                 {
 256                                         temp=A[k+i*lda];
 257                                         temp1+=temp*B[j+k*lda];
 258                                         temp2+=temp*B[j+1+k*lda];
 259                                         temp3+=temp*B[j+2+k*lda];
 260                                         temp4+=temp*B[j+3+k*lda];
 261
 262                                         temp_1=A[k+(i+1)*lda];
 263                                         temp1_1+=temp_1*B[j+k*lda];
 264                                         temp2_1+=temp_1*B[j+1+k*lda];
 265                                         temp3_1+=temp_1*B[j+2+k*lda];
 266                                         temp4_1+=temp_1*B[j+3+k*lda];
 267
 268                                         temp_2=A[k+(i+2)*lda];
 269                                         temp1_2+=temp_2*B[j+k*lda];
 270                                         temp2_2+=temp_2*B[j+1+k*lda];
 271                                         temp3_2+=temp_2*B[j+2+k*lda];
 272                                         temp4_2+=temp_2*B[j+3+k*lda];
 273
 274                                         temp_3=A[k+(i+3)*lda];
 275                                         temp1_3+=temp_3*B[j+k*lda];
 276                                         temp2_3+=temp_3*B[j+1+k*lda];
 277                                         temp3_3+=temp_3*B[j+2+k*lda];
 278                                         temp4_3+=temp_3*B[j+3+k*lda];
 279
 280                                 }
 281                                 C[j+i*lda]=temp1;
 282                                 C[j+1+i*lda]=temp2;
 283                                 C[j+2+i*lda]=temp3;
 284                                 C[j+3+i*lda]=temp4;
 285
 286                                 C[j+(i+1)*lda]=temp1_1;
 287                                 C[j+1+(i+1)*lda]=temp2_1;
 288                                 C[j+2+(i+1)*lda]=temp3_1;
 289                                 C[j+3+(i+1)*lda]=temp4_1;
 290
 291                                 C[j+(i+2)*lda]=temp1_2;
 292                                 C[j+1+(i+2)*lda]=temp2_2;
 293                                 C[j+2+(i+2)*lda]=temp3_2;
 294                                 C[j+3+(i+2)*lda]=temp4_2;
 295
 296                                 C[j+(i+3)*lda]=temp1_3;
 297                                 C[j+1+(i+3)*lda]=temp2_3;
 298                                 C[j+2+(i+3)*lda]=temp3_3;
 299                                 C[j+3+(i+3)*lda]=temp4_3;
 300
 301                         }
 302
 303                 }
 304
 305
 306         }
 307
 308
 309 }
 310
 311 //--------------------------------------------------------------------------
 312 // Main
 313 //
 314 // all threads start executing thread_entry(). Use their "coreid" to
 315 // differentiate between threads (each thread is running on a separate core).
 316
 317 void thread_entry(int cid, int nc)
 318 {
 319    coreid = cid;
 320    ncores = nc;
 321
 322    // static allocates data in the binary, which is visible to both threads
 323    static data_t results_data[ARRAY_SIZE];
 324
 325
 326 //   // Execute the provided, naive matmul
 327 //   barrier();
 328 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 329 //
 330 //
 331 //   // verify
 332 //   verify(ARRAY_SIZE, results_data, verify_data);
 333 //
 334 //   // clear results from the first trial
 335 //   size_t i;
 336 //   if (coreid == 0)
 337 //      for (i=0; i < ARRAY_SIZE; i++)
 338 //         results_data[i] = 0;
 339 //   barrier();
 340
 341
 342    // Execute your faster matmul
 343    barrier();
 344    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 345
 346 #ifdef DEBUG
 347    printArray("results:", ARRAY_SIZE, results_data);
 348    printArray("verify :", ARRAY_SIZE, verify_data);
 349 #endif
 350
 351    // verify
 352    verify(ARRAY_SIZE, results_data, verify_data);
 353    barrier();
 354
 355    exit(0);
 356 }
 357