mt/ae_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 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 110 {
 111
 112         data_t a1;
 113         data_t a2;
 114         data_t a3;
 115         data_t a4;
 116         data_t a5;
 117         data_t a6;
 118         data_t a7;
 119         data_t a8;
 120         data_t *b1;
 121         data_t *b2;
 122         data_t *b3;
 123         data_t *b4;
 124         data_t *b5;
 125         data_t *b6;
 126         data_t *b7;
 127         data_t *b8;
 128         data_t c1;
 129         data_t c2;
 130         data_t c3;
 131         data_t c4;
 132         data_t c5;
 133         data_t c6;
 134         data_t c7;
 135         data_t c8;
 136         int i, j, k;
 137         int start, end;
 138         static data_t BB[1024];
 139
 140
 141         //transpose B
 142         if (coreid == 0 | coreid == 1 ) {
 143                 for ( k = 0; k < lda; k++) {
 144                         for ( i = coreid*(lda/2); i < (coreid+1)*(lda/2); i++ )  {
 145                                 BB[i*lda + k] = B[k*lda + i];
 146                         }
 147                 }
 148         }
 149         barrier();
 150
 151         for ( int x = 0; x < ncores; x++) {
 152                 //split the i values into two chunks so the threads don't interfere on the B loads
 153                 //this could be generalized if needed, but I won't bother since it would be tricky
 154                 //and we already know the size and numthreads
 155                 start = coreid == x ? 0 : 16;
 156                 end = coreid == x ? 16 : 32;
 157                 for ( i = start; i < end; i+=8 ) {
 158                         for ( j = coreid*(lda/ncores); j < (coreid+1)*(lda/ncores); j++ )  {
 159                                 c1=0;c2=0;c3=0;c4=0;c5=0;c6=0;c7=0;c8=0;
 160                                 b1 = &BB[(i+0)*lda];
 161                                 b2 = &BB[(i+1)*lda];
 162                                 b3 = &BB[(i+2)*lda];
 163                                 b4 = &BB[(i+3)*lda];
 164                                 b5 = &BB[(i+4)*lda];
 165                                 b6 = &BB[(i+5)*lda];
 166                                 b7 = &BB[(i+6)*lda];
 167                                 b8 = &BB[(i+7)*lda];
 168
 169                                 for ( k = 0; k < lda; k+=8 ) {
 170                                         a1 = A[j*lda + k+0];
 171                                         a2 = A[j*lda + k+1];
 172                                         a3 = A[j*lda + k+2];
 173                                         a4 = A[j*lda + k+3];
 174                                         a5 = A[j*lda + k+4];
 175                                         a6 = A[j*lda + k+5];
 176                                         a7 = A[j*lda + k+6];
 177                                         a8 = A[j*lda + k+7];
 178
 179                                         c1 += a1 * b1[k+0];
 180                                         c1 += a2 * b1[k+1];
 181                                         c1 += a3 * b1[k+2];
 182                                         c1 += a4 * b1[k+3];
 183                                         c1 += a5 * b1[k+4];
 184                                         c1 += a6 * b1[k+5];
 185                                         c1 += a7 * b1[k+6];
 186                                         c1 += a8 * b1[k+7];
 187
 188                                         c2 += a1 * b2[k+0];
 189                                         c2 += a2 * b2[k+1];
 190                                         c2 += a3 * b2[k+2];
 191                                         c2 += a4 * b2[k+3];
 192                                         c2 += a5 * b2[k+4];
 193                                         c2 += a6 * b2[k+5];
 194                                         c2 += a7 * b2[k+6];
 195                                         c2 += a8 * b2[k+7];
 196
 197                                         c3 += a1 * b3[k+0];
 198                                         c3 += a2 * b3[k+1];
 199                                         c3 += a3 * b3[k+2];
 200                                         c3 += a4 * b3[k+3];
 201                                         c3 += a5 * b3[k+4];
 202                                         c3 += a6 * b3[k+5];
 203                                         c3 += a7 * b3[k+6];
 204                                         c3 += a8 * b3[k+7];
 205
 206                                         c4 += a1 * b4[k+0];
 207                                         c4 += a2 * b4[k+1];
 208                                         c4 += a3 * b4[k+2];
 209                                         c4 += a4 * b4[k+3];
 210                                         c4 += a5 * b4[k+4];
 211                                         c4 += a6 * b4[k+5];
 212                                         c4 += a7 * b4[k+6];
 213                                         c4 += a8 * b4[k+7];
 214
 215                                         c5 += a1 * b5[k+0];
 216                                         c5 += a2 * b5[k+1];
 217                                         c5 += a3 * b5[k+2];
 218                                         c5 += a4 * b5[k+3];
 219                                         c5 += a5 * b5[k+4];
 220                                         c5 += a6 * b5[k+5];
 221                                         c5 += a7 * b5[k+6];
 222                                         c5 += a8 * b5[k+7];
 223
 224                                         c6 += a1 * b6[k+0];
 225                                         c6 += a2 * b6[k+1];
 226                                         c6 += a3 * b6[k+2];
 227                                         c6 += a4 * b6[k+3];
 228                                         c6 += a5 * b6[k+4];
 229                                         c6 += a6 * b6[k+5];
 230                                         c6 += a7 * b6[k+6];
 231                                         c6 += a8 * b6[k+7];
 232
 233                                         c7 += a1 * b7[k+0];
 234                                         c7 += a2 * b7[k+1];
 235                                         c7 += a3 * b7[k+2];
 236                                         c7 += a4 * b7[k+3];
 237                                         c7 += a5 * b7[k+4];
 238                                         c7 += a6 * b7[k+5];
 239                                         c7 += a7 * b7[k+6];
 240                                         c7 += a8 * b7[k+7];
 241
 242                                         c8 += a1 * b8[k+0];
 243                                         c8 += a2 * b8[k+1];
 244                                         c8 += a3 * b8[k+2];
 245                                         c8 += a4 * b8[k+3];
 246                                         c8 += a5 * b8[k+4];
 247                                         c8 += a6 * b8[k+5];
 248                                         c8 += a7 * b8[k+6];
 249                                         c8 += a8 * b8[k+7];
 250                                 }
 251                                 C[i+0 + j*lda] += c1;
 252                                 C[i+1 + j*lda] += c2;
 253                                 C[i+2 + j*lda] += c3;
 254                                 C[i+3 + j*lda] += c4;
 255                                 C[i+4 + j*lda] += c5;
 256                                 C[i+5 + j*lda] += c6;
 257                                 C[i+6 + j*lda] += c7;
 258                                 C[i+7 + j*lda] += c8;
 259                         }
 260                 }
 261         }
 262 }
 263
 264 //--------------------------------------------------------------------------
 265 // Main
 266 //
 267 // all threads start executing thread_entry(). Use their "coreid" to
 268 // differentiate between threads (each thread is running on a separate core).
 269
 270 void thread_entry(int cid, int nc)
 271 {
 272    coreid = cid;
 273    ncores = nc;
 274
 275    // static allocates data in the binary, which is visible to both threads
 276    static data_t results_data[ARRAY_SIZE];
 277
 278 /*
 279    // Execute the provided, naive matmul
 280    barrier();
 281    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 282
 283
 284    // verify
 285    verify(ARRAY_SIZE, results_data, verify_data);
 286
 287    // clear results from the first trial
 288    size_t i;
 289    if (coreid == 0)
 290       for (i=0; i < ARRAY_SIZE; i++)
 291          results_data[i] = 0;
 292    barrier();
 293 */
 294
 295
 296    // Execute your faster matmul
 297    barrier();
 298    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 299
 300 #ifdef DEBUG
 301    printArray("results:", ARRAY_SIZE, results_data);
 302    printArray("verify :", ARRAY_SIZE, verify_data);
 303 #endif
 304
 305    // verify
 306    verify(ARRAY_SIZE, results_data, verify_data);
 307    barrier();
 308
 309    exit(0);
 310 }
 311