mt/bf_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    int j, k;
 119    data_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
 120    data_t temp8, temp9, temp10, temp11, temp12, temp13, temp14, temp15;
 121    if(coreid == 0) {
 122       //16*0:16*(0+1) ;; 16*1+16*(1+1)
 123       //0:16 ;; 16:32
 124
 125       //complete Q1
 126       for(j = 0; j < 16; j++) {
 127          temp0  = C[j*lda];
 128          temp1  = C[1  + j*lda];
 129          temp2  = C[2  + j*lda];
 130          temp3  = C[3  + j*lda];
 131          temp4  = C[4  + j*lda];
 132          temp5  = C[5  + j*lda];
 133          temp6  = C[6  + j*lda];
 134          temp7  = C[7  + j*lda];
 135          temp8  = C[8  + j*lda];
 136          temp9  = C[9  + j*lda];
 137          temp10 = C[10 + j*lda];
 138          temp11 = C[11 + j*lda];
 139          temp12 = C[12 + j*lda];
 140          temp13 = C[13 + j*lda];
 141          temp14 = C[14 + j*lda];
 142          temp15 = C[15 + j*lda];
 143          for(k = 0; k < 32; k++) {
 144             temp0  += A[j*lda + k] * B[k*lda];
 145             temp1  += A[j*lda + k] * B[1+k*lda];
 146             temp2  += A[j*lda + k] * B[2+k*lda];
 147             temp3  += A[j*lda + k] * B[3+k*lda];
 148             temp4  += A[j*lda + k] * B[4+k*lda];
 149             temp5  += A[j*lda + k] * B[5+k*lda];
 150             temp6  += A[j*lda + k] * B[6+k*lda];
 151             temp7  += A[j*lda + k] * B[7+k*lda];
 152             temp8  += A[j*lda + k] * B[8+k*lda];
 153             temp9  += A[j*lda + k] * B[9+k*lda];
 154             temp10 += A[j*lda + k] * B[10+k*lda];
 155             temp11 += A[j*lda + k] * B[11+k*lda];
 156             temp12 += A[j*lda + k] * B[12+k*lda];
 157             temp13 += A[j*lda + k] * B[13+k*lda];
 158             temp14 += A[j*lda + k] * B[14+k*lda];
 159             temp15 += A[j*lda + k] * B[15+k*lda];
 160          }
 161          C[j*lda] = temp0;
 162          C[1  + j*lda] = temp1;
 163          C[2  + j*lda] = temp2;
 164          C[3  + j*lda] = temp3;
 165          C[4  + j*lda] = temp4;
 166          C[5  + j*lda] = temp5;
 167          C[6  + j*lda] = temp6;
 168          C[7  + j*lda] = temp7;
 169          C[8  + j*lda] = temp8;
 170          C[9  + j*lda] = temp9;
 171          C[10 + j*lda] = temp10;
 172          C[11 + j*lda] = temp11;
 173          C[12 + j*lda] = temp12;
 174          C[13 + j*lda] = temp13;
 175          C[14 + j*lda] = temp14;
 176          C[15 + j*lda] = temp15;
 177       }
 178       for(j = 16; j < 32; j++) {
 179          temp0  = C[j*lda];
 180          temp1  = C[1  + j*lda];
 181          temp2  = C[2  + j*lda];
 182          temp3  = C[3  + j*lda];
 183          temp4  = C[4  + j*lda];
 184          temp5  = C[5  + j*lda];
 185          temp6  = C[6  + j*lda];
 186          temp7  = C[7  + j*lda];
 187          temp8  = C[8  + j*lda];
 188          temp9  = C[9  + j*lda];
 189          temp10 = C[10 + j*lda];
 190          temp11 = C[11 + j*lda];
 191          temp12 = C[12 + j*lda];
 192          temp13 = C[13 + j*lda];
 193          temp14 = C[14 + j*lda];
 194          temp15 = C[15 + j*lda];
 195          for(k = 0; k < 32; k++) {
 196             temp0  += A[j*lda + k] * B[k*lda];
 197             temp1  += A[j*lda + k] * B[1+k*lda];
 198             temp2  += A[j*lda + k] * B[2+k*lda];
 199             temp3  += A[j*lda + k] * B[3+k*lda];
 200             temp4  += A[j*lda + k] * B[4+k*lda];
 201             temp5  += A[j*lda + k] * B[5+k*lda];
 202             temp6  += A[j*lda + k] * B[6+k*lda];
 203             temp7  += A[j*lda + k] * B[7+k*lda];
 204             temp8  += A[j*lda + k] * B[8+k*lda];
 205             temp9  += A[j*lda + k] * B[9+k*lda];
 206             temp10 += A[j*lda + k] * B[10+k*lda];
 207             temp11 += A[j*lda + k] * B[11+k*lda];
 208             temp12 += A[j*lda + k] * B[12+k*lda];
 209             temp13 += A[j*lda + k] * B[13+k*lda];
 210             temp14 += A[j*lda + k] * B[14+k*lda];
 211             temp15 += A[j*lda + k] * B[15+k*lda];
 212          }
 213          C[j*lda] = temp0;
 214          C[1  + j*lda] = temp1;
 215          C[2  + j*lda] = temp2;
 216          C[3  + j*lda] = temp3;
 217          C[4  + j*lda] = temp4;
 218          C[5  + j*lda] = temp5;
 219          C[6  + j*lda] = temp6;
 220          C[7  + j*lda] = temp7;
 221          C[8  + j*lda] = temp8;
 222          C[9  + j*lda] = temp9;
 223          C[10 + j*lda] = temp10;
 224          C[11 + j*lda] = temp11;
 225          C[12 + j*lda] = temp12;
 226          C[13 + j*lda] = temp13;
 227          C[14 + j*lda] = temp14;
 228          C[15 + j*lda] = temp15;
 229       }
 230    }
 231    //16*(2-1) : 16*2 ;; 16*(1-1) : 16*1
 232    //16:32 ;; 0:16
 233    if(coreid == 1)   {
 234          //complete Q3
 235       for(j = 16; j < 32; j++) {
 236          temp0  = C[16+j*lda];
 237          temp1  = C[17+j*lda];
 238          temp2  = C[18+j*lda];
 239          temp3  = C[19+j*lda];
 240          temp4  = C[20+j*lda];
 241          temp5  = C[21+j*lda];
 242          temp6  = C[22+j*lda];
 243          temp7  = C[23+j*lda];
 244          temp8  = C[24+j*lda];
 245          temp9  = C[25+j*lda];
 246          temp10 = C[26+j*lda];
 247          temp11 = C[27+j*lda];
 248          temp12 = C[28+j*lda];
 249          temp13 = C[29+j*lda];
 250          temp14 = C[30+j*lda];
 251          temp15 = C[31+j*lda];
 252          for(k = 0; k < 32; k++) {
 253             temp0  += A[j*lda + k] * B[16+k*lda];
 254             temp1  += A[j*lda + k] * B[17+k*lda];
 255             temp2  += A[j*lda + k] * B[18+k*lda];
 256             temp3  += A[j*lda + k] * B[19+k*lda];
 257             temp4  += A[j*lda + k] * B[20+k*lda];
 258             temp5  += A[j*lda + k] * B[21+k*lda];
 259             temp6  += A[j*lda + k] * B[22+k*lda];
 260             temp7  += A[j*lda + k] * B[23+k*lda];
 261             temp8  += A[j*lda + k] * B[24+k*lda];
 262             temp9  += A[j*lda + k] * B[25+k*lda];
 263             temp10 += A[j*lda + k] * B[26+k*lda];
 264             temp11 += A[j*lda + k] * B[27+k*lda];
 265             temp12 += A[j*lda + k] * B[28+k*lda];
 266             temp13 += A[j*lda + k] * B[29+k*lda];
 267             temp14 += A[j*lda + k] * B[30+k*lda];
 268             temp15 += A[j*lda + k] * B[31+k*lda];
 269          }
 270          C[16 + j*lda] = temp0;
 271          C[17 + j*lda] = temp1;
 272          C[18 + j*lda] = temp2;
 273          C[19 + j*lda] = temp3;
 274          C[20 + j*lda] = temp4;
 275          C[21 + j*lda] = temp5;
 276          C[22 + j*lda] = temp6;
 277          C[23 + j*lda] = temp7;
 278          C[24 + j*lda] = temp8;
 279          C[25 + j*lda] = temp9;
 280          C[26 + j*lda] = temp10;
 281          C[27 + j*lda] = temp11;
 282          C[28 + j*lda] = temp12;
 283          C[29 + j*lda] = temp13;
 284          C[30 + j*lda] = temp14;
 285          C[31 + j*lda] = temp15;
 286       }
 287       //complete Q4
 288       for(j = 0; j < 16; j++) {
 289          temp0  = C[16 + j*lda];
 290          temp1  = C[17 + j*lda];
 291          temp2  = C[18 + j*lda];
 292          temp3  = C[19 + j*lda];
 293          temp4  = C[20 + j*lda];
 294          temp5  = C[21 + j*lda];
 295          temp6  = C[22 + j*lda];
 296          temp7  = C[23 + j*lda];
 297          temp8  = C[24 + j*lda];
 298          temp9  = C[25 + j*lda];
 299          temp10 = C[26 + j*lda];
 300          temp11 = C[27 + j*lda];
 301          temp12 = C[28 + j*lda];
 302          temp13 = C[29 + j*lda];
 303          temp14 = C[30 + j*lda];
 304          temp15 = C[31 + j*lda];
 305          for(k = 0; k < 32; k++) {
 306             temp0  += A[j*lda + k] * B[16 + k*lda];
 307             temp1  += A[j*lda + k] * B[17 + k*lda];
 308             temp2  += A[j*lda + k] * B[18 + k*lda];
 309             temp3  += A[j*lda + k] * B[19 + k*lda];
 310             temp4  += A[j*lda + k] * B[20 + k*lda];
 311             temp5  += A[j*lda + k] * B[21 + k*lda];
 312             temp6  += A[j*lda + k] * B[22 + k*lda];
 313             temp7  += A[j*lda + k] * B[23 + k*lda];
 314             temp8  += A[j*lda + k] * B[24 + k*lda];
 315             temp9  += A[j*lda + k] * B[25 + k*lda];
 316             temp10 += A[j*lda + k] * B[26 + k*lda];
 317             temp11 += A[j*lda + k] * B[27 + k*lda];
 318             temp12 += A[j*lda + k] * B[28 + k*lda];
 319             temp13 += A[j*lda + k] * B[29 + k*lda];
 320             temp14 += A[j*lda + k] * B[30 + k*lda];
 321             temp15 += A[j*lda + k] * B[31 + k*lda];
 322          }
 323          C[16 + j*lda] = temp0;
 324          C[17 + j*lda] = temp1;
 325          C[18 + j*lda] = temp2;
 326          C[19 + j*lda] = temp3;
 327          C[20 + j*lda] = temp4;
 328          C[21 + j*lda] = temp5;
 329          C[22 + j*lda] = temp6;
 330          C[23 + j*lda] = temp7;
 331          C[24 + j*lda] = temp8;
 332          C[25 + j*lda] = temp9;
 333          C[26 + j*lda] = temp10;
 334          C[27 + j*lda] = temp11;
 335          C[28 + j*lda] = temp12;
 336          C[29 + j*lda] = temp13;
 337          C[30 + j*lda] = temp14;
 338          C[31 + j*lda] = temp15;
 339       }
 340    }
 341
 342
 343
 344 }
 345
 346 //--------------------------------------------------------------------------
 347 // Main
 348 //
 349 // all threads start executing thread_entry(). Use their "coreid" to
 350 // differentiate between threads (each thread is running on a separate core).
 351
 352 void thread_entry(int cid, int nc)
 353 {
 354    coreid = cid;
 355    ncores = nc;
 356
 357    // static allocates data in the binary, which is visible to both threads
 358    static data_t results_data[ARRAY_SIZE];
 359
 360
 361 //   // Execute the provided, naive matmul
 362 //   barrier();
 363 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 364 //
 365 //
 366 //   // verify
 367 //   verify(ARRAY_SIZE, results_data, verify_data);
 368 //
 369 //   // clear results from the first trial
 370 //   size_t i;
 371 //   if (coreid == 0)
 372 //      for (i=0; i < ARRAY_SIZE; i++)
 373 //         results_data[i] = 0;
 374 //   barrier();
 375
 376
 377    // Execute your faster matmul
 378    barrier();
 379    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 380
 381 #ifdef DEBUG
 382    printArray("results:", ARRAY_SIZE, results_data);
 383    printArray("verify :", ARRAY_SIZE, verify_data);
 384 #endif
 385
 386    // verify
 387    verify(ARRAY_SIZE, results_data, verify_data);
 388    barrier();
 389
 390    exit(0);
 391 }
 392