mt/bc_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 #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 printArray( 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)) verify(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       int kk_start= (coreid == 0 ? 0 : LDA/2) ,kk_end = (coreid == 0 ? LDA/2 : LDA);
 139       for (kk = kk_start; kk < kk_end; kk += BLOCK_K) {
 140         //  for (ii = 0; ii < LDA; ii += BLOCK_I)
 141         for (j = jj; j < MIN(end, jj + BLOCK_J); j += REG_J) {
 142           Aj = A + j*LDA;
 143           Cj = C + j*LDA;
 144           for (i = 0; i < LDA/*, ii + BLOCK_I)*/; i += REG_I) {
 145             /* Load C in register blocks. */
 146             Bi = B + i;
 147             for (ri = 0; ri < REG_I; ri++) {
 148               for (rj = 0; rj < REG_J; rj++) {
 149                 c[ri][rj] = Cj[i + ri + ( rj)*LDA];
 150               }
 151             }
 152
 153
 154             for (k = kk; k < MIN(LDA, kk + BLOCK_K); k++) {
 155               for (ri = 0; ri < REG_I; ri++) {
 156                 b[ri] = Bi[k*LDA  + ri];
 157               }
 158               /* Compute C in register blocks. */
 159               for (rj = 0; rj < REG_J; rj++) {
 160                 a[rj] = Aj[(rj)*LDA + k];
 161                 for (ri = 0; ri < REG_I; ri++) {
 162                   c[ri][rj] += a[rj] * b[ri];
 163                 }
 164               }
 165             }
 166
 167             /* store C in register blocks. */
 168             for (ri = 0; ri < REG_I; ri++) {
 169               for (rj = 0; rj < REG_J; rj++) {
 170                 Cj[i + ri + ( rj)*LDA] = c[ri][rj];
 171               }
 172             }
 173           }
 174         }
 175         /* barrier(); */
 176
 177         /* kk_start= (coreid == 1 ? 0 : LDA/2); */
 178         /* kk_end = (coreid == 1 ? LDA/2 : LDA); */
 179         /* for (kk = kk_start; kk < kk_end; kk += BLOCK_K) { */
 180         /* //  for (ii = 0; ii < LDA; ii += BLOCK_I) */
 181         /* for (j = jj; j < MIN(end, jj + BLOCK_J); j += REG_J) { */
 182         /*   Aj = A + j*LDA; */
 183         /*   Cj = C + j*LDA; */
 184         /*   for (i = 0; i < LDA/\*, ii + BLOCK_I)*\/; i += REG_I) { */
 185         /*     /\* Load C in register blocks. *\/ */
 186         /*     Bi = B + i; */
 187         /*     for (ri = 0; ri < REG_I; ri++) { */
 188         /*       for (rj = 0; rj < REG_J; rj++) { */
 189         /*      c[ri][rj] = Cj[i + ri + ( rj)*LDA]; */
 190         /*       } */
 191         /*     } */
 192
 193
 194         /*     for (k = kk; k < MIN(LDA, kk + BLOCK_K); k++) { */
 195         /*       for (ri = 0; ri < REG_I; ri++) { */
 196         /*      b[ri] = Bi[k*LDA  + ri]; */
 197         /*       } */
 198         /*       /\* Compute C in register blocks. *\/ */
 199         /*       for (rj = 0; rj < REG_J; rj++) { */
 200         /*      a[rj] = Aj[(rj)*LDA + k]; */
 201         /*      for (ri = 0; ri < REG_I; ri++) { */
 202         /*        c[ri][rj] += a[rj] * b[ri]; */
 203         /*      } */
 204         /*       } */
 205         /*     } */
 206
 207             /* store C in register blocks. */
 208         /*     for (ri = 0; ri < REG_I; ri++) { */
 209     /*        for (rj = 0; rj < REG_J; rj++) { */
 210     /*          Cj[i + ri + ( rj)*LDA] = c[ri][rj]; */
 211     /*        } */
 212     /*      } */
 213     /*   } */
 214     /* } */
 215       }
 216     }
 217
 218
 219     //barrier();
 220     for (jj = start; jj < end; jj += BLOCK_J) {
 221       int kk_start= (coreid != 0 ? 0 : LDA/2), kk_end = (coreid != 0 ? LDA/2 : LDA);
 222       for (kk = kk_start; kk < kk_end; kk += BLOCK_K) {
 223         //  for (ii = 0; ii < LDA; ii += BLOCK_I)
 224         for (j = jj; j < MIN(end, jj + BLOCK_J); j += REG_J) {
 225           Aj = A + j*LDA;
 226           Cj = C + j*LDA;
 227           for (i = 0; i < LDA/*, ii + BLOCK_I)*/; i += REG_I) {
 228             /* Load C in register blocks. */
 229             Bi = B + i;
 230             for (ri = 0; ri < REG_I; ri++) {
 231               for (rj = 0; rj < REG_J; rj++) {
 232                 c[ri][rj] = Cj[i + ri + ( rj)*LDA];
 233               }
 234             }
 235
 236
 237             for (k = kk; k < MIN(LDA, kk + BLOCK_K); k++) {
 238               for (ri = 0; ri < REG_I; ri++) {
 239                 b[ri] = Bi[k*LDA  + ri];
 240               }
 241               /* Compute C in register blocks. */
 242               for (rj = 0; rj < REG_J; rj++) {
 243                 a[rj] = Aj[(rj)*LDA + k];
 244                 for (ri = 0; ri < REG_I; ri++) {
 245                   c[ri][rj] += a[rj] * b[ri];
 246                 }
 247               }
 248             }
 249
 250               /* store C in register blocks. */
 251             for (ri = 0; ri < REG_I; ri++) {
 252               for (rj = 0; rj < REG_J; rj++) {
 253                 Cj[i + ri + ( rj)*LDA] = c[ri][rj];
 254               }
 255             }
 256           }
 257         }
 258       }
 259     }
 260     /* We only care about performance for 32x32 matrices and 2 cores. Otherwise just naive mat_mul */
 261 } else {
 262       if (coreid > 0)
 263         return;
 264
 265       for ( i = 0; i < lda; i++ )
 266         for ( j = 0; j < lda; j++ )
 267           for ( k = 0; k < lda; k++ )
 268             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 269     }
 270   }
 271
 272 //--------------------------------------------------------------------------
 273 // Main
 274 //
 275 // all threads start executing thread_entry(). Use their "coreid" to
 276 // differentiate between threads (each thread is running on a separate core).
 277
 278 void thread_entry(int cid, int nc)
 279 {
 280    coreid = cid;
 281    ncores = nc;
 282
 283    // static allocates data in the binary, which is visible to both threads
 284    static data_t results_data[ARRAY_SIZE];
 285
 286
 287 //   /* // Execute the provided, naive matmul */
 288 //   barrier();
 289 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 290 //
 291 //
 292 //   // verify
 293 //   verify(ARRAY_SIZE, results_data, verify_data);
 294 //
 295 //   // clear results from the first trial
 296 //   size_t i;
 297 //   if (coreid == 0)
 298 //      for (i=0; i < ARRAY_SIZE; i++)
 299 //         results_data[i] = 0;
 300 //   barrier();
 301
 302
 303    // Execute your faster matmul
 304    barrier();
 305    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 306
 307 #ifdef DEBUG
 308    printArray("results:", ARRAY_SIZE, results_data);
 309    printArray("verify :", ARRAY_SIZE, verify_data);
 310 #endif
 311
 312    // verify
 313    verify(ARRAY_SIZE, results_data, verify_data);
 314    barrier();
 315
 316    exit(0);
 317 }
 318