mt/aj_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 void matrix_sub(int size, data_t A[], data_t B[], data_t C[]) {
  87         if (coreid != 0)
  88                 return;
  89
  90         for(int i = 0; i < size; i++){
  91                 C[i] = A[i] + B[i];
  92         }
  93 }
  94
  95 void matrix_add(int size, data_t A[], data_t B[], data_t C[]) {
  96         if (coreid != 0)
  97                 return;
  98
  99         for(int i = 0; i < size; i++){
 100                 C[i] = A[i] - B[i];
 101         }
 102 }
 103
 104 void strassen_mult(int dime, const data_t sA[],  const data_t sB[], data_t sC[]) {
 105
 106         if (coreid != 0)
 107                 return;
 108
 109         int height, width;
 110         int sub_size = dime*dime/4;
 111
 112 //      data_t  A_11[sub_size], B_11[sub_size], C_11[sub_size],
 113 //                      A_12[sub_size], B_12[sub_size], C_12[sub_size],
 114 //                      A_21[sub_size], B_21[sub_size], C_21[sub_size],
 115 //                      A_22[sub_size], B_22[sub_size], C_22[sub_size];
 116
 117         data_t *A_11 = malloc(sub_size*sizeof(data_t));
 118         data_t *A_12 = malloc(sub_size*sizeof(data_t));
 119         data_t *A_21 = malloc(sub_size*sizeof(data_t));
 120         data_t *A_22 = malloc(sub_size*sizeof(data_t));
 121         data_t *B_11 = malloc(sub_size*sizeof(data_t));
 122         data_t *B_12 = malloc(sub_size*sizeof(data_t));
 123         data_t *B_21 = malloc(sub_size*sizeof(data_t));
 124         data_t *B_22 = malloc(sub_size*sizeof(data_t));
 125
 126         for(height=0; height < dime/2; height++) {
 127                 for(width= 0; width < dime/2; width++) {
 128                         A_11[width+(height*dime/2)] = sA[width + height*dime];
 129                         B_11[width+(height*dime/2)] = sB[width + height*dime];
 130
 131                         A_12[width+(height*dime/2)] = sA[dime/2 + width + height*dime];
 132                         B_12[width+(height*dime/2)] = sB[dime/2 + width + height*dime];
 133
 134                         A_21[width+(height*dime/2)] = sA[(dime*dime)/2 + width + height*dime];
 135                         B_21[width+(height*dime/2)] = sB[(dime*dime)/2 + width + height*dime];
 136
 137                         A_22[width+(height*dime/2)] = sA[(dime*dime)/2 + dime/2 + width + height*dime];
 138                         B_22[width+(height*dime/2)] = sB[(dime*dime)/2 + dime/2 + width + height*dime];
 139                 }
 140         }
 141
 142 //      data_t          H_1[sub_size], H_2[sub_size], H_3[sub_size], H_4[sub_size], H_5[sub_size],
 143 //                      H_6[sub_size], H_7[sub_size], H_8[sub_size], H_9[sub_size], H_10[sub_size],
 144 //                      H_11[sub_size], H_12[sub_size], H_13[sub_size], H_14[sub_size],
 145 //                      H_15[sub_size], H_16[sub_size], H_17[sub_size], H_18[sub_size];
 146
 147         data_t *H_1 = malloc(sub_size*sizeof(data_t));
 148         data_t *H_2 = malloc(sub_size*sizeof(data_t));
 149         data_t *H_3 = malloc(sub_size*sizeof(data_t));
 150         data_t *H_4 = malloc(sub_size*sizeof(data_t));
 151         data_t *H_5 = malloc(sub_size*sizeof(data_t));
 152         data_t *H_6 = malloc(sub_size*sizeof(data_t));
 153         data_t *H_7 = malloc(sub_size*sizeof(data_t));
 154         data_t *H_8 = malloc(sub_size*sizeof(data_t));
 155         data_t *H_9 = malloc(sub_size*sizeof(data_t));
 156         data_t *H_10 = malloc(sub_size*sizeof(data_t));
 157
 158         matrix_add(sub_size, A_11, A_22, H_1); //Helper1
 159         matrix_add(sub_size, B_11, B_22, H_2); //Helper2
 160         matrix_add(sub_size, A_21, A_22, H_3); //Helper3
 161         matrix_sub(sub_size, B_12, B_22, H_4); //Helper4
 162         matrix_sub(sub_size, B_21, B_11, H_5); //Helper5
 163         matrix_add(sub_size, A_11, A_12, H_6); //Helper6
 164         matrix_sub(sub_size, A_21, A_11, H_7); //Helper7
 165         matrix_add(sub_size, B_11, B_12, H_8); //Helper8
 166         matrix_sub(sub_size, A_12, A_22, H_9); //Helper9
 167         matrix_add(sub_size, B_21, B_22, H_10); //Helper10
 168
 169         free(A_12);
 170         free(A_21);
 171         free(B_12);
 172         free(B_21);
 173
 174         A_12 = NULL;
 175         A_21 = NULL;
 176         B_12 = NULL;
 177         B_21 = NULL;
 178
 179 //      data_t          M_1[sub_size], M_2[sub_size], M_3[sub_size], M_4[sub_size],
 180 //                      M_5[sub_size], M_6[sub_size], M_7[sub_size];
 181
 182         data_t *M_1 = malloc(sub_size*sizeof(data_t));
 183         data_t *M_2 = malloc(sub_size*sizeof(data_t));
 184         data_t *M_3 = malloc(sub_size*sizeof(data_t));
 185         data_t *M_4 = malloc(sub_size*sizeof(data_t));
 186         data_t *M_5 = malloc(sub_size*sizeof(data_t));
 187         data_t *M_6 = malloc(sub_size*sizeof(data_t));
 188         data_t *M_7 = malloc(sub_size*sizeof(data_t));
 189
 190         if (sub_size == 1) {
 191                 M_1[0] = H_1[0]*H_2[0];
 192                 M_2[0] = H_3[0]*B_11[0];
 193                 M_3[0] = A_11[0]*H_4[0];
 194                 M_4[0] = A_22[0]*H_5[0];
 195                 M_5[0] = H_6[0]*B_22[0];
 196                 M_6[0] = H_7[0]*H_8[0];
 197                 M_7[0] = H_9[0]*H_10[0];
 198         } else {
 199                 strassen_mult(dime/2, H_1, H_2, M_1);
 200                 strassen_mult(dime/2, H_3, B_11, M_2);
 201                 strassen_mult(dime/2, A_11, H_4, M_3);
 202                 strassen_mult(dime/2, A_22, H_5, M_4);
 203                 strassen_mult(dime/2, H_6, B_22, M_5);
 204                 strassen_mult(dime/2, H_7, H_8, M_6);
 205                 strassen_mult(dime/2, H_9, H_10, M_7);
 206         }
 207
 208         free(A_11);
 209         free(A_22);
 210         free(B_11);
 211         free(B_22);
 212
 213         A_11 = NULL;
 214         A_22 = NULL;
 215         B_11 = NULL;
 216         B_22 = NULL;
 217
 218         free(H_1);
 219         free(H_2);
 220         free(H_3);
 221         free(H_4);
 222         free(H_5);
 223         free(H_6);
 224         free(H_7);
 225         free(H_8);
 226         free(H_9);
 227         free(H_10);
 228
 229         H_1 = NULL;
 230         H_2 = NULL;
 231         H_3 = NULL;
 232         H_4 = NULL;
 233         H_5 = NULL;
 234         H_6 = NULL;
 235         H_7 = NULL;
 236         H_8 = NULL;
 237         H_9 = NULL;
 238         H_10 = NULL;
 239
 240         data_t *H_11 = malloc(sub_size*sizeof(data_t));
 241         data_t *H_12 = malloc(sub_size*sizeof(data_t));
 242         data_t *H_13 = malloc(sub_size*sizeof(data_t));
 243         data_t *H_14 = malloc(sub_size*sizeof(data_t));
 244
 245         data_t *C_11 = malloc(sub_size*sizeof(data_t));
 246         data_t *C_12 = malloc(sub_size*sizeof(data_t));
 247         data_t *C_21 = malloc(sub_size*sizeof(data_t));
 248         data_t *C_22 = malloc(sub_size*sizeof(data_t));
 249
 250         matrix_add(sub_size, M_1, M_4, H_11);
 251         matrix_add(sub_size, M_5, M_7, H_12);
 252         matrix_sub(sub_size, H_11, H_12, C_11);
 253
 254         matrix_add(sub_size, M_3, M_5, C_12);
 255
 256         matrix_add(sub_size, M_2, M_4, C_21);
 257
 258         matrix_sub(sub_size, M_1, M_2, H_13);
 259         matrix_add(sub_size, M_3, M_6, H_14);
 260         matrix_add(sub_size, H_13, H_14, C_22);
 261
 262         free(H_11);
 263         free(H_12);
 264         free(H_13);
 265         free(H_14);
 266
 267         H_11 = NULL;
 268         H_12 = NULL;
 269         H_13 = NULL;
 270         H_14 = NULL;
 271
 272
 273         for(height=0; height < dime/2; height++) {
 274                 for(width= 0; width < dime/2; width++) {
 275                         sC[width + height*dime]                                 = C_11[width+(height*dime/2)];
 276                         sC[dime/2 + width + height*dime]                        = C_12[width+(height*dime/2)];
 277                         sC[(dime*dime)/2 + width + height*dime]                 = C_21[width+(height*dime/2)];
 278                         sC[(dime*dime)/2 + dime/2 + width + height*dime] = C_22[width+(height*dime/2)];
 279                 }
 280         }
 281
 282         free(C_11);
 283         free(C_12);
 284         free(C_21);
 285         free(C_22);
 286
 287         C_11 = NULL;
 288         C_12 = NULL;
 289         C_21 = NULL;
 290         C_22 = NULL;
 291
 292 }
 293
 294 //--------------------------------------------------------------------------
 295 // matmul function
 296
 297 // single-thread, naive version
 298 void __attribute__((noinline)) matmul_naive(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 299 {
 300    int i, j, k;
 301
 302    if (coreid > 0)
 303       return;
 304
 305    for ( i = 0; i < lda; i++ )
 306       for ( j = 0; j < lda; j++ )
 307       {
 308          for ( k = 0; k < lda; k++ )
 309          {
 310             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 311          }
 312       }
 313
 314 }
 315
 316
 317
 318 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 319 {
 320
 321    // ***************************** //
 322    // **** ADD YOUR CODE HERE ***** //
 323    // ***************************** //
 324    //
 325    // feel free to make a separate function for MI and MSI versions.
 326
 327         if (coreid > 0)
 328         return;
 329
 330         strassen_mult(lda, A, B, C);
 331
 332 }
 333
 334 //--------------------------------------------------------------------------
 335 // Main
 336 //
 337 // all threads start executing thread_entry(). Use their "coreid" to
 338 // differentiate between threads (each thread is running on a separate core).
 339
 340 void thread_entry(int cid, int nc)
 341 {
 342    coreid = cid;
 343    ncores = nc;
 344
 345    // static allocates data in the binary, which is visible to both threads
 346    static data_t results_data[ARRAY_SIZE];
 347
 348
 349 //   // Execute the provided, naive matmul
 350 //   barrier();
 351 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 352 //
 353 //
 354 //   // verify
 355 //   verify(ARRAY_SIZE, results_data, verify_data);
 356 //
 357 //   // clear results from the first trial
 358 //   size_t i;
 359 //   if (coreid == 0)
 360 //      for (i=0; i < ARRAY_SIZE; i++)
 361 //         results_data[i] = 0;
 362 //   barrier();
 363
 364
 365    // Execute your faster matmul
 366    barrier();
 367    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 368
 369 #ifdef DEBUG
 370    printArray("results:", ARRAY_SIZE, results_data);
 371    printArray("verify :", ARRAY_SIZE, verify_data);
 372 #endif
 373
 374    // verify
 375    verify(ARRAY_SIZE, results_data, verify_data);
 376    barrier();
 377
 378    exit(0);
 379 }
 380