mt/bc_matmul/bc_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 #define REG_I 8
  25 #define REG_J 2
  26 //#define BLOCK_I 32
  27 #define BLOCK_J 16
  28 #define BLOCK_K 16
  29 #define LDA 32
  30 #define NCORES 2
  31 #define MIN(X,Y) (X < Y ? X : Y)
  32
  33 //--------------------------------------------------------------------------
  34 // Input/Reference Data
  35
  36 typedef float data_t;
  37 #include "dataset.h"
  38
  39
  40 //--------------------------------------------------------------------------
  41 // Basic Utilities and Multi-thread Support
  42
  43 __thread unsigned long coreid;
  44 unsigned long ncores;
  45
  46 #include "util.h"
  47
  48 #define stringify_1(s) #s
  49 #define stringify(s) stringify_1(s)
  50 #define stats(code) do { \
  51     unsigned long _c = -rdcycle(), _i = -rdinstret(); \
  52     code; \
  53     _c += rdcycle(), _i += rdinstret(); \
  54     if (coreid == 0) \
  55       printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
  56              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); \
  57   } while(0)
  58
  59
  60 //--------------------------------------------------------------------------
  61 // Helper functions
  62
  63 void printArrayMT( char name[], int n, data_t arr[] )
  64 {
  65    int i;
  66    if (coreid != 0)
  67       return;
  68
  69    printf( " %10s :", name );
  70    for ( i = 0; i < n; i++ )
  71       printf( " %3ld ", (long) arr[i] );
  72    printf( "\n" );
  73 }
  74
  75 void __attribute__((noinline)) verifyMT(size_t n, const data_t* test, const data_t* correct)
  76 {
  77    if (coreid != 0)
  78       return;
  79
  80    size_t i;
  81    for (i = 0; i < n; i++)
  82    {
  83       if (test[i] != correct[i])
  84       {
  85          printf("FAILED test[%d]= %3ld, correct[%d]= %3ld\n",
  86             i, (long)test[i], i, (long)correct[i]);
  87          exit(-1);
  88       }
  89    }
  90
  91    return;
  92 }
  93
  94 //--------------------------------------------------------------------------
  95 // matmul function
  96
  97 // single-thread, naive version
  98 void __attribute__((noinline)) matmul_naive(const int lda,  const data_t A[], const data_t B[], data_t C[] )
  99 {
 100    int i, j, k;
 101
 102    if (coreid > 0)
 103       return;
 104
 105    for ( i = 0; i < lda; i++ )
 106       for ( j = 0; j < lda; j++ )
 107       {
 108          for ( k = 0; k < lda; k++ )
 109          {
 110             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 111          }
 112       }
 113
 114 }
 115
 116
 117
 118 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 119 {
 120
 121    // ***************************** //
 122    // **** ADD YOUR CODE HERE ***** //
 123    // ***************************** //
 124    //
 125    // feel free to make a separate function for MI and MSI versions.
 126
 127   int i, j, k, ri, rj, ii, jj, kk;
 128   data_t *Aj, *Cj, *Bi;
 129   data_t c[REG_I][REG_J], a[REG_J], b[REG_I];
 130   size_t start = coreid * (LDA / NCORES), end = (coreid == NCORES - 1 ? LDA : (coreid + 1) * (LDA / NCORES));
 131
 132   /* if (coreid > 0) { */
 133   /*   return; */
 134   /* } */
 135   /* start = 0, end = lda; */
 136   if (ncores == NCORES && lda == LDA) {
 137     for (jj = start; jj < end; jj += BLOCK_J)
 138       for (kk = 0; kk < LDA; kk += BLOCK_K)
 139         //for (ii = 0; ii < LDA; ii += BLOCK_I)
 140         for (j = jj; j < MIN(end, jj + BLOCK_J); j += REG_J) {
 141           Aj = A + j*LDA;
 142           Cj = C + j*LDA;
 143           for (i = 0; i < LDA; i += REG_I) {
 144             /* Load C in register blocks. */
 145             Bi = B + i;
 146             for (ri = 0; ri < REG_I; ri++) {
 147               for (rj = 0; rj < REG_J; rj++) {
 148                 c[ri][rj] = Cj[i + ri + ( rj)*LDA];
 149               }
 150             }
 151
 152
 153             for (k = kk; k < MIN(LDA, kk + BLOCK_K); k++) {
 154               /* Load a,b in register blocks. */
 155               /*          for (rj = 0; rj < REG_J; rj++) {
 156                           a[rj] = A[(j + rj)*LDA + k];
 157                           }*/
 158               /* for (ri = 0; ri < REG_I; ri++) { */
 159               /*        b[ri] = Bi[k*LDA  + ri]; */
 160               /* } */
 161               /* /\* Compute C in register blocks. *\/ */
 162               /* for (rj = 0; rj < REG_J; rj++) { */
 163               /*        a[rj] = Aj[( rj)*LDA + k]; */
 164               /*        for (ri = 0; ri < REG_I; ri++) { */
 165               /*          c[ri][rj] += a[rj] * b[ri]; */
 166               /*        } */
 167               /* } */
 168               a[0] = Aj[k];
 169               a[1] = Aj[k + LDA];
 170               b[0] = Bi[k*LDA];
 171               b[1] = Bi[k*LDA + 1];
 172               b[2] = Bi[k*LDA + 2];
 173               b[3] = Bi[k*LDA + 3];
 174               b[4] = Bi[k*LDA + 4];
 175               b[5] = Bi[k*LDA + 5];
 176               b[6] = Bi[k*LDA + 6];
 177               b[7] = Bi[k*LDA + 7];
 178
 179
 180               c[0][0] += b[0] * a[0];
 181               c[0][1] += b[0] * a[1];
 182               c[1][0] += b[1] * a[0];
 183               c[1][1] += b[1] * a[1];
 184               c[2][0] += b[2] * a[0];
 185               c[2][1] += b[2] * a[1];
 186               c[3][0] += b[3] * a[0];
 187               c[3][1] += b[3] * a[1];
 188               c[4][0] += b[4] * a[0];
 189               c[4][1] += b[4] * a[1];
 190               c[5][0] += b[5] * a[0];
 191               c[5][1] += b[5] * a[1];
 192               c[6][0] += b[6] * a[0];
 193               c[6][1] += b[6] * a[1];
 194               c[7][0] += b[7] * a[0];
 195               c[7][1] += b[7] * a[1];
 196
 197
 198               /* c[0][0] +=  b[0] * a[0];              */
 199               /* c[1][1] +=  b[1] * a[1];              */
 200               /* c[2][0] +=  b[2] * a[0];              */
 201               /* c[3][1] +=  b[3] * a[1];              */
 202               /* c[4][0] +=  b[4] * a[0];              */
 203               /* c[5][1] +=  b[5] * a[1];              */
 204               /* c[6][0] +=  b[6] * a[0];              */
 205               /* c[7][1] +=  b[7] * a[1];              */
 206               /* c[0][0] +=  b[0] * a[0];              */
 207               /* c[1][1] +=  b[1] * a[1];              */
 208               /* c[2][0] +=  b[2] * a[0];              */
 209               /* c[3][1] +=  b[3] * a[1];              */
 210               /* c[4][0] +=  b[4] * a[0];              */
 211               /* c[5][1] +=  b[5] * a[1];              */
 212               /* c[6][0] +=  b[6] * a[0];              */
 213               /* c[7][1] +=  b[7] * a[1];              */
 214
 215             }
 216
 217             /* store C in register blocks. */
 218             for (ri = 0; ri < REG_I; ri++) {
 219               for (rj = 0; rj < REG_J; rj++) {
 220                 Cj[i + ri + (rj)*LDA] = c[ri][rj];
 221               }
 222             }
 223           }
 224
 225
 226
 227
 228         }
 229     /* We only care about performance for 32x32 matrices and 2 cores. Otherwise just naive mat_mul */
 230   } else {
 231     if (coreid > 0)
 232       return;
 233
 234     for ( i = 0; i < lda; i++ )
 235       for ( j = 0; j < lda; j++ )
 236         for ( k = 0; k < lda; k++ )
 237           C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 238   }
 239 }
 240
 241 //--------------------------------------------------------------------------
 242 // Main
 243 //
 244 // all threads start executing thread_entry(). Use their "coreid" to
 245 // differentiate between threads (each thread is running on a separate core).
 246
 247 void thread_entry(int cid, int nc)
 248 {
 249    coreid = cid;
 250    ncores = nc;
 251
 252    // static allocates data in the binary, which is visible to both threads
 253    static data_t results_data[ARRAY_SIZE];
 254
 255
 256 //   /* // Execute the provided, naive matmul */
 257 //   barrier(nc);
 258 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 259 //
 260 //
 261 //   // verify
 262 //   verifyMT(ARRAY_SIZE, results_data, verify_data);
 263 //
 264 //   // clear results from the first trial
 265 //   size_t i;
 266 //   if (coreid == 0)
 267 //      for (i=0; i < ARRAY_SIZE; i++)
 268 //         results_data[i] = 0;
 269 //   barrier(nc);
 270
 271
 272    // Execute your faster matmul
 273    barrier(nc);
 274    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 275
 276 #ifdef DEBUG
 277    printArrayMT("results:", ARRAY_SIZE, results_data);
 278    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 279 #endif
 280
 281    // verify
 282    verifyMT(ARRAY_SIZE, results_data, verify_data);
 283    barrier(nc);
 284
 285    exit(0);
 286 }
 287