mt/af_matmul/failedattempt.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
 110 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 111 {
 112   size_t i;
 113         size_t i2;
 114   size_t j;
 115         size_t j2;
 116         size_t k;
 117         size_t k2;
 118   size_t max_dim = lda*lda;
 119         size_t block_size = lda/2;
 120   data_t temp_mat[16] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 121         if (coreid == 0) {
 122                 //making a 16x16 block
 123                 //First block: Top 16x16 block left of A and top left of B = top left of C
 124                 //Second block: top right 16x16 right block of A and top right of B = top right of C
 125                 for (j2= 0; j2 < 2; j2++) {
 126                   for (i2 = 0; i2 < 2; i2++) {
 127                                 //for (j2= 0; j2 < 2; j2++) {
 128                                         //K represents which row of A and C
 129                                         for (k = 0; k < block_size; k++) {
 130                                                 int rowIndex = k*32;
 131                                                 for (i = i2*block_size; i < i2*block_size+block_size; i++) {
 132                                                         int elementA = A[rowIndex+i];
 133                                                         int columnIndex = i%32*32;
 134                                                         for (j = 0; j < block_size; j++) {
 135                                                                 temp_mat[j] += elementA*B[columnIndex+j+j2*block_size];
 136                                             }
 137                                                 }
 138                                                 //Put temp_mat into actual result Matrix
 139                                                 for (k2 = 0; k2 < block_size; k2++) {
 140                                                         C[rowIndex+k2+j2*block_size] += temp_mat[k2];
 141                                                         temp_mat[k2] = 0;
 142                                                 }
 143                                         }
 144                                 }
 145                         }
 146         } else {
 147                 for (j2= 0; j2 < 2; j2++) {
 148                         for (i2 = 0; i2 < 2; i2++) {
 149                                 //for (j2= 0; j2 < 2; j2++) {
 150                                 //K represents which row of A and C
 151                                         for (k = block_size; k < lda; k++) {
 152                                                 int rowIndex = k*32;
 153                                                 for (i = i2*block_size; i < i2*block_size+block_size; i++) {
 154                                                         int elementA = A[rowIndex+i];
 155                                                         int columnIndex = i%32*32;
 156                                                         for (j = 0; j < block_size; j++) {
 157                                                                 temp_mat[j] += elementA*B[columnIndex+j+j2*block_size];
 158                                                   }
 159                                                 }
 160                                                 //Put temp_mat into actual result Matrix
 161                                                 for (k2 = 0; k2 < block_size; k2++) {
 162                                                         C[rowIndex+k2+j2*block_size] += temp_mat[k2];
 163                                                         temp_mat[k2] = 0;
 164                                                 }
 165                                         }
 166                                 }
 167                         }
 168         }
 169
 170
 171   //size_t half_lda = lda/2;
 172         // k = which pair of row we're on
 173
 174
 175
 176
 177
 178
 179 /*
 180   for (k = coreid*lda/ncores; k < (lda/ncores + coreid*lda/ncores); k += 2) {
 181           //printf("%d", k);
 182         for (i = 0; i < lda ; i++) {
 183                         int elementA = A[32*k+i];
 184                         int elementA2 = A[i + 32*(k+1)];
 185                         int column = i%32*32;
 186             for (j = 0; j < lda; j++) {
 187                                 C[32*k + j] += elementA*B[column+j];
 188                                 C[32*(k+1) + j] += elementA2*B[column+j];
 189                         }
 190                 }
 191
 192         }
 193 */
 194
 195 /*
 196     data_t element=A[i];
 197     data_t element2 = A[i+1];
 198     data_t element3 = A[i+2];
 199     data_t element4 = A[i+3];
 200     data_t element5 = A[i+4];
 201     data_t element6 = A[i+5];
 202     data_t element7 = A[i+6];
 203     data_t element8 = A[i+7];
 204     int row= (int)(i/32)*32;
 205     int row2 = (i+1)/32*32;
 206     int row3 = (i+2)/32*32;
 207     int row4 = (i+3)/32*32;
 208     int row5 = (i+4)/32*32;
 209     int row6 = (i+5)/32*32;
 210     int row7 = (i+6)/32*32;
 211     int row8 = (i+7)/32*32;
 212     int column = i%32*32;
 213     int column2 = (i+1)%32*32;
 214     int column3 = (i+2)%32*32;
 215     int column4 = (i+3)%32*32;
 216     int column5 = (i+4)%32*32;
 217     int column6 = (i+5)%32*32;
 218     int column7 = (i+6)%32*32;
 219
 220         */
 221
 222     //int column8 = (i+7)%32*32;
 223
 224   /*
 225     for (j=0; j < lda; j++) {
 226         sum = B[
 227       C[row+j]+=element*B[column+j];
 228       C[row2+j]+=element2*B[column2+j];
 229       C[row3+j]+=element3*B[column3+j];
 230       C[row4+j]+=element4*B[column4+j];
 231       C[row5+j]+=element5*B[column5+j];
 232       C[row6+j]+=element6*B[column6+j];
 233       C[row7+j]+=element7*B[column7+j];
 234       C[row8+j]+=element8*B[column8+j];
 235       C[row+j]+=element*B[column+j]+element2*B[column2+j]+element3*B[column3+j]+element4*B[column4+j]+element5*B[column5+j]+element6*B[column6+j]+element7*B[column7+j]+element8*B[column8+j];
 236     }
 237   }
 238         */
 239
 240
 241
 242
 243
 244
 245    // ***************************** //
 246    // **** ADD YOUR CODE HERE ***** //
 247    // ***************************** //
 248    //
 249    // feel free to make a separate function for MI and MSI versions.
 250
 251 }
 252
 253 //--------------------------------------------------------------------------
 254 // Main
 255 //
 256 // all threads start executing thread_entry(). Use their "coreid" to
 257 // differentiate between threads (each thread is running on a separate core).
 258
 259 void thread_entry(int cid, int nc)
 260 {
 261    coreid = cid;
 262    ncores = nc;
 263
 264    // static allocates data in the binary, which is visible to both threads
 265    static data_t results_data[ARRAY_SIZE];
 266
 267
 268    // Execute the provided, naive matmul
 269    barrier();
 270    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 271
 272
 273    // verify
 274    verify(ARRAY_SIZE, results_data, verify_data);
 275
 276    // clear results from the first trial
 277    size_t i;
 278    if (coreid == 0)
 279       for (i=0; i < ARRAY_SIZE; i++)
 280          results_data[i] = 0;
 281    barrier();
 282
 283
 284    // Execute your faster matmul
 285    barrier();
 286    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 287
 288 #ifdef DEBUG
 289    printArray("results:", ARRAY_SIZE, results_data);
 290    printArray("verify :", ARRAY_SIZE, verify_data);
 291 #endif
 292
 293    // verify
 294    verify(ARRAY_SIZE, results_data, verify_data);
 295    barrier();
 296
 297    exit(0);
 298 }