mt/ae_matmul/ae_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 printArrayMT( 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)) verifyMT(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    // ***************************** //
 113    // **** ADD YOUR CODE HERE ***** //
 114    // ***************************** //
 115    //
 116    // feel free to make a separate function for MI and MSI versions.
 117
 118
 119
 120         data_t *b1;
 121         data_t *b2;
 122         data_t *b3;
 123         data_t *b4;
 124         data_t c1;
 125         data_t c2;
 126         data_t c3;
 127         data_t c4;
 128         data_t a1;
 129         data_t a2;
 130         data_t a3;
 131         data_t a4;
 132         data_t a5;
 133         data_t a6;
 134         data_t a7;
 135         data_t a8;
 136         int i, j, k;
 137         static data_t BB[1024];
 138
 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(ncores);
 150
 151         for ( i = 0; i < lda; i+=4 ) {
 152                 for ( j = coreid*(lda/ncores); j < (coreid+1)*(lda/ncores); j++ )  {
 153                         c1 = 0; c2 = 0; c3 = 0; c4 = 0;
 154                         b1 = &BB[(i+0)*lda];
 155                         b2 = &BB[(i+1)*lda];
 156                         b3 = &BB[(i+2)*lda];
 157                         b4 = &BB[(i+3)*lda];
 158                         for ( k = 0; k < lda; k+=8 ) {
 159
 160                                 a1 = A[j*lda + k+0];
 161                                 a2 = A[j*lda + k+1];
 162                                 a3 = A[j*lda + k+2];
 163                                 a4 = A[j*lda + k+3];
 164                                 a5 = A[j*lda + k+4];
 165                                 a6 = A[j*lda + k+5];
 166                                 a7 = A[j*lda + k+6];
 167                                 a8 = A[j*lda + k+7];
 168
 169                                 c1 += a1 * b1[k+0];
 170                                 c1 += a2 * b1[k+1];
 171                                 c1 += a3 * b1[k+2];
 172                                 c1 += a4 * b1[k+3];
 173                                 c1 += a5 * b1[k+4];
 174                                 c1 += a6 * b1[k+5];
 175                                 c1 += a7 * b1[k+6];
 176                                 c1 += a8 * b1[k+7];
 177
 178                                 c2 += a1 * b2[k+0];
 179                                 c2 += a2 * b2[k+1];
 180                                 c2 += a3 * b2[k+2];
 181                                 c2 += a4 * b2[k+3];
 182                                 c2 += a5 * b2[k+4];
 183                                 c2 += a6 * b2[k+5];
 184                                 c2 += a7 * b2[k+6];
 185                                 c2 += a8 * b2[k+7];
 186
 187                                 c3 += a1 * b3[k+0];
 188                                 c3 += a2 * b3[k+1];
 189                                 c3 += a3 * b3[k+2];
 190                                 c3 += a4 * b3[k+3];
 191                                 c3 += a5 * b3[k+4];
 192                                 c3 += a6 * b3[k+5];
 193                                 c3 += a7 * b3[k+6];
 194                                 c3 += a8 * b3[k+7];
 195
 196                                 c4 += a1 * b4[k+0];
 197                                 c4 += a2 * b4[k+1];
 198                                 c4 += a3 * b4[k+2];
 199                                 c4 += a4 * b4[k+3];
 200                                 c4 += a5 * b4[k+4];
 201                                 c4 += a6 * b4[k+5];
 202                                 c4 += a7 * b4[k+6];
 203                                 c4 += a8 * b4[k+7];
 204
 205
 206                         }
 207                         C[i+0 + j*lda] = c1;
 208                         C[i+1 + j*lda] = c2;
 209                         C[i+2 + j*lda] = c3;
 210                         C[i+3 + j*lda] = c4;
 211                 }
 212         }
 213
 214 }
 215
 216 //--------------------------------------------------------------------------
 217 // Main
 218 //
 219 // all threads start executing thread_entry(). Use their "coreid" to
 220 // differentiate between threads (each thread is running on a separate core).
 221
 222 void thread_entry(int cid, int nc)
 223 {
 224    coreid = cid;
 225    ncores = nc;
 226
 227    // static allocates data in the binary, which is visible to both threads
 228    static data_t results_data[ARRAY_SIZE];
 229
 230 /*
 231    // Execute the provided, naive matmul
 232    barrier(nc);
 233    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 234
 235
 236    // verify
 237    verifyMT(ARRAY_SIZE, results_data, verify_data);
 238
 239    // clear results from the first trial
 240    size_t i;
 241    if (coreid == 0)
 242       for (i=0; i < ARRAY_SIZE; i++)
 243          results_data[i] = 0;
 244    barrier(nc);
 245 */
 246
 247
 248    // Execute your faster matmul
 249    barrier(nc);
 250    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 251
 252 #ifdef DEBUG
 253    printArrayMT("results:", ARRAY_SIZE, results_data);
 254    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 255 #endif
 256
 257    // verify
 258    verifyMT(ARRAY_SIZE, results_data, verify_data);
 259    barrier(nc);
 260
 261    exit(0);
 262 }
 263