mt/bt_matmul/bt_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
 119           int i, j, k , jj , kk;
 120                 int start_i = coreid*lda/2;
 121                 int end_i = start_i + lda/2;
 122                 int step_j, step_k;
 123                 int start_k, end_k, start_j, end_j;
 124                 int j_lda;
 125                 int pos_A , pos_B, pos_C;
 126                 data_t temp00, temp01,temp02,temp03,temp04,temp05,temp06,temp07;
 127                 data_t temp10, temp11,temp12,temp13,temp14,temp15,temp16,temp17;
 128                 data_t temp_A0, temp_A1, temp_A2, temp_A3, temp_A4, temp_A5, temp_A6, temp_A7;
 129
 130                 if (coreid == 0)
 131                 {
 132                         step_k = 1;
 133                         start_k= 0;
 134                         end_k = lda;
 135
 136                         step_j = 2;
 137                         start_j= 0;
 138                         end_j = lda;
 139
 140                 }else
 141                 {
 142
 143                         step_k = -1;
 144                         start_k = lda-1;
 145                         end_k = -1;
 146
 147                         step_j = -2;
 148                         start_j= lda-2;
 149                         end_j = -2;
 150                 }
 151
 152                 for( kk = start_k ; kk!= end_k ; kk+=(step_k*16) )
 153                 {
 154                         for( jj = start_j ; jj!= end_j ; jj+=(step_j*8) )
 155                         {
 156                                 for ( i = start_i; i < end_i; i+=8 )
 157                                 {
 158                                         //pos_C = i + jj*lda;
 159                                         for ( j = jj; j != (jj+(step_j*8)) ; j+=step_j )
 160                                         {
 161
 162                                                         pos_C = i + j*lda;
 163                                                         temp00 = C[(pos_C + 0)];
 164                                                         temp01 = C[(pos_C + 1)];
 165                                                         temp02 = C[(pos_C + 2)];
 166                                                         temp03 = C[(pos_C + 3)];
 167                                                         temp04 = C[(pos_C + 4)];
 168                                                         temp05 = C[(pos_C + 5)];
 169                                                         temp06 = C[(pos_C + 6)];
 170                                                         temp07 = C[(pos_C + 7)];
 171
 172                                                         //pos_C += lda;
 173                                                         pos_C = i + (j+1)*lda;
 174
 175                                                         temp10 = C[(pos_C + 0)];
 176                                                         temp11 = C[(pos_C + 1)];
 177                                                         temp12 = C[(pos_C + 2)];
 178                                                         temp13 = C[(pos_C + 3)];
 179                                                         temp14 = C[(pos_C + 4)];
 180                                                         temp15 = C[(pos_C + 5)];
 181                                                         temp16 = C[(pos_C + 6)];
 182                                                         temp17 = C[(pos_C + 7)];
 183
 184                                                 pos_B = kk*lda + i;
 185                                                 pos_A = j*lda + kk;
 186                                                 for ( k = kk; k != (kk+(step_k*16)) ; k+=step_k )
 187                                                 {
 188                                                         temp_A0 = A[ pos_A ] ;
 189                                                         temp_A1 = A[pos_A +lda];
 190
 191                                                         temp00 += temp_A0 * B[(pos_B + 0)];
 192                                                         temp01 += temp_A0 * B[(pos_B + 1)];
 193                                                         temp02 += temp_A0 * B[(pos_B + 2)];
 194                                                         temp03 += temp_A0 * B[(pos_B + 3)];
 195                                                         temp04 += temp_A0 * B[(pos_B + 4)];
 196                                                         temp05 += temp_A0 * B[(pos_B + 5)];
 197                                                         temp06 += temp_A0 * B[(pos_B + 6)];
 198                                                         temp07 += temp_A0 * B[(pos_B + 7)];
 199
 200                                                         temp10 += temp_A1 * B[(pos_B + 0)];
 201                                                         temp11 += temp_A1 * B[(pos_B + 1)];
 202                                                         temp12 += temp_A1 * B[(pos_B + 2)];
 203                                                         temp13 += temp_A1 * B[(pos_B + 3)];
 204                                                         temp14 += temp_A1 * B[(pos_B + 4)];
 205                                                         temp15 += temp_A1 * B[(pos_B + 5)];
 206                                                         temp16 += temp_A1 * B[(pos_B + 6)];
 207                                                         temp17 += temp_A1 * B[(pos_B + 7)];
 208
 209                                                         pos_B += (lda*step_k) ;
 210                                                         pos_A += step_k;
 211                                                 }
 212                                                 //barrier();
 213
 214                                                 C[(pos_C + 0)] = temp10;
 215                                                 C[(pos_C + 1)] = temp11;
 216                                                 C[(pos_C + 2)] = temp12;
 217                                                 C[(pos_C + 3)] = temp13;
 218                                                 C[(pos_C + 4)] = temp14;
 219                                                 C[(pos_C + 5)] = temp15;
 220                                                 C[(pos_C + 6)] = temp16;
 221                                                 C[(pos_C + 7)] = temp17;
 222                                                 //barrier();
 223
 224                                                 pos_C = i + j*lda;
 225                                                 //pos_C -= lda;
 226                                                 C[(pos_C + 0)] = temp00;
 227                                                 C[(pos_C + 1)] = temp01;
 228                                                 C[(pos_C + 2)] = temp02;
 229                                                 C[(pos_C + 3)] = temp03;
 230                                                 C[(pos_C + 4)] = temp04;
 231                                                 C[(pos_C + 5)] = temp05;
 232                                                 C[(pos_C + 6)] = temp06;
 233                                                 C[(pos_C + 7)] = temp07;
 234                                                 //barrier();
 235                                                 //pos_C += step_j * lda;
 236                                         }
 237                                         //barrier();
 238                                 }
 239                                 //barrier();
 240
 241                         }
 242                         //barrier();
 243                 }
 244 }
 245
 246 //--------------------------------------------------------------------------
 247 // Main
 248 //
 249 // all threads start executing thread_entry(). Use their "coreid" to
 250 // differentiate between threads (each thread is running on a separate core).
 251
 252 void thread_entry(int cid, int nc)
 253 {
 254    coreid = cid;
 255    ncores = nc;
 256
 257    // static allocates data in the binary, which is visible to both threads
 258    static data_t results_data[ARRAY_SIZE];
 259
 260         /*
 261    // Execute the provided, naive matmul
 262    barrier();
 263    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 264
 265
 266    // verify
 267    verify(ARRAY_SIZE, results_data, verify_data);
 268
 269    // clear results from the first trial
 270    size_t i;
 271    if (coreid == 0)
 272       for (i=0; i < ARRAY_SIZE; i++)
 273          results_data[i] = 0;
 274    barrier();
 275
 276    */
 277    // Execute your faster matmul
 278    barrier();
 279    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 280
 281
 282
 283 #ifdef DEBUG
 284    printArray("results:", ARRAY_SIZE, results_data);
 285    printArray("verify :", ARRAY_SIZE, verify_data);
 286 #endif
 287
 288    // verify
 289    verify(ARRAY_SIZE, results_data, verify_data);
 290    barrier();
 291
 292
 293         //printf("input1_data");
 294 exit(0);
 295
 296 }