mt/af_matmul/keeptrying2.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  size_t i, j, k, l;
 112   int row,row2, column, column2, column3, column4, column5, column6, column7, column8;
 113   data_t element, element2, element3, element4, element5, element6, element7, element8;
 114   data_t temp_mat[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 115   data_t temp_mat2[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 116   //for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
 117         for (l=coreid*32/ncores; l<32*(1+coreid)/ncores; l+=2){
 118           row=l*lda;
 119           row2=(l+1)*lda;
 120                 if (coreid == 0) {
 121                                 for (i=0; i<lda; i+=4){
 122                           element = A[row+i];
 123                           element2 = A[row+i+1];
 124                           element3 = A[row+i+2];
 125                           element4 = A[row+i+3];
 126
 127                           element5 = A[row2+i];
 128                           element6 = A[row2+i+1];
 129                           element7 = A[row2+i+2];
 130                           element8 = A[row2+i+3];
 131
 132                           column=i*lda;
 133                           column2=(i+1)*lda;
 134                           column3=(i+2)*lda;
 135                           column4=(i+3)*lda;
 136
 137                           for (j=0; j<lda; j+=4){
 138                                         temp_mat[j]+=element*B[column+j]+element2*B[column2+j]+element3*B[column3+j]+element4*B[column4+j];
 139                                         temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1];
 140                                         temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2];
 141                                         temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3];
 142
 143                                         temp_mat2[j]+=element5*B[column+j]+element6*B[column2+j]+element7*B[column3+j]+element8*B[column4+j];
 144                                         temp_mat2[j+1]+=element5*B[column+j+1]+element6*B[column2+j+1]+element7*B[column3+j+1]+element8*B[column4+j+1];
 145                                         temp_mat2[j+2]+=element5*B[column+j+2]+element6*B[column2+j+2]+element7*B[column3+j+2]+element8*B[column4+j+2];
 146                                         temp_mat2[j+3]+=element5*B[column+j+3]+element6*B[column2+j+3]+element7*B[column3+j+3]+element8*B[column4+j+3];
 147                           }
 148
 149                           }
 150
 151                                 for(k=0; k<32; k++){
 152                                         C[row+k]=temp_mat[k];
 153                                         temp_mat[k]=0;
 154                                         C[row2+k]=temp_mat2[k];
 155                                         temp_mat2[k]=0;
 156
 157
 158                                         }
 159                         } else {
 160
 161                         for (i=0; i<lda; i += 4){
 162                           element = A[row-i+lda-1];
 163                           element2 = A[row-i-1+lda-1];
 164                           element3 = A[row-i-2+lda-1];
 165                           element4 = A[row-i-3+lda-1];
 166
 167                           element5 = A[row2-i+lda-1];
 168                           element6 = A[row2-i-1+lda-1];
 169                           element7 = A[row2-i-2+lda-1];
 170                           element8 = A[row2-i-3+lda-1];
 171
 172                           column=(-i+lda-1)*lda;
 173                           column2=(-i-1+lda-1)*lda;
 174                           column3=(-i-2+lda-1)*lda;
 175                           column4=(-i-3+lda-1)*lda;
 176
 177                           for (j=0; j<lda; j+=4){
 178                                         temp_mat[j]+=element*B[column+j]+element2*B[column2+j]+element3*B[column3+j]+element4*B[column4+j];
 179                                         temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1];
 180                                         temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2];
 181                                         temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3];
 182
 183                                         temp_mat2[j]+=element5*B[column+j]+element6*B[column2+j]+element7*B[column3+j]+element8*B[column4+j];
 184                                         temp_mat2[j+1]+=element5*B[column+j+1]+element6*B[column2+j+1]+element7*B[column3+j+1]+element8*B[column4+j+1];
 185                                         temp_mat2[j+2]+=element5*B[column+j+2]+element6*B[column2+j+2]+element7*B[column3+j+2]+element8*B[column4+j+2];
 186                                         temp_mat2[j+3]+=element5*B[column+j+3]+element6*B[column2+j+3]+element7*B[column3+j+3]+element8*B[column4+j+3];
 187                           }
 188
 189                           }
 190
 191                                 for(k=0; k<32; k++){
 192                                         C[row+k]=temp_mat[k];
 193                                         temp_mat[k]=0;
 194                                         C[row2+k]=temp_mat2[k];
 195                                         temp_mat2[k]=0;
 196
 197
 198                                         }
 199                 }
 200         }
 201
 202    // ***************************** //
 203    // **** ADD YOUR CODE HERE ***** //
 204    // ***************************** //
 205    //
 206    // feel free to make a separate function for MI and MSI versions.
 207
 208 }
 209 //--------------------------------------------------------------------------
 210 // Main
 211 //
 212 // all threads start executing thread_entry(). Use their "coreid" to
 213 // differentiate between threads (each thread is running on a separate core).
 214
 215 void thread_entry(int cid, int nc)
 216 {
 217    coreid = cid;
 218    ncores = nc;
 219
 220    // static allocates data in the binary, which is visible to both threads
 221    static data_t results_data[ARRAY_SIZE];
 222
 223         /*
 224    // Execute the provided, naive matmul
 225    barrier(nc);
 226    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 227
 228
 229    // verify
 230    verifyMT(ARRAY_SIZE, results_data, verify_data);
 231
 232    // clear results from the first trial
 233    size_t i;
 234    if (coreid == 0)
 235       for (i=0; i < ARRAY_SIZE; i++)
 236          results_data[i] = 0;
 237    barrier(nc);
 238
 239    */
 240    // Execute your faster matmul
 241    barrier(nc);
 242    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 243
 244 #ifdef DEBUG
 245    printArrayMT("results:", ARRAY_SIZE, results_data);
 246    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 247 #endif
 248
 249    // verify
 250    verifyMT(ARRAY_SIZE, results_data, verify_data);
 251    barrier(nc);
 252
 253    exit(0);
 254 }