mt/bt_matmul/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 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
 113    // ***************************** //
 114    // **** ADD YOUR CODE HERE ***** //
 115    // ***************************** //
 116    //
 117    // feel free to make a separate function for MI and MSI versions.
 118    int i, j, k;
 119    int temp0, temp1,temp2,temp3,temp4,temp5,temp6,temp7;
 120         int start = coreid*lda/2;
 121         int end = start + lda/2;
 122    int j_lda;
 123         int temp_i;
 124         int temp_A0, temp_A1, temp_A2, temp_A3 ;
 125
 126                  for ( i = start; i < end; i+=8){
 127                     for ( j = 0; j < lda; j++)
 128                     {
 129                                         j_lda = j*lda;
 130                                         temp0 = C[(i+0) + j_lda];
 131                                         temp1 = C[(i+1) + j_lda];
 132                                         temp2 = C[(i+2) + j_lda];
 133                                         temp3 = C[(i+3) + j_lda];
 134                                         temp4 = C[(i+4) + j_lda];
 135                                         temp5 = C[(i+5) + j_lda];
 136                                         temp6 = C[(i+6) + j_lda];
 137                                         temp7 = C[(i+7) + j_lda];
 138
 139
 140
 141                        for ( k = 0; k < lda; k+=4)
 142                        {
 143                                 temp_i = i;
 144                                 temp_A0         = A[j_lda + (k+0)] ;
 145                                 temp_A1         = A[j_lda + (k+1)] ;
 146                                 temp_A2         = A[j_lda + (k+2)] ;
 147                                 temp_A3 = A[j_lda + (k+3)] ;
 148
 149
 150                                 temp0 += temp_A0 * B[(k+0)*lda + temp_i];
 151                                 temp0 += temp_A1 * B[(k+1)*lda + temp_i];
 152                                 temp0 += temp_A2 * B[(k+2)*lda + temp_i];
 153                                 temp0 += temp_A3 * B[(k+3)*lda + temp_i];
 154                                 temp_i++;
 155
 156                                 temp1 += temp_A0 * B[(k+0)*lda + temp_i];
 157                                 temp1 += temp_A1 * B[(k+1)*lda + temp_i];
 158                                 temp1 += temp_A2 * B[(k+2)*lda + temp_i];
 159                                 temp1 += temp_A3 * B[(k+3)*lda + temp_i];
 160                                 temp_i++;
 161
 162                                 temp2 += temp_A0 * B[(k+0)*lda + temp_i];
 163                                 temp2 += temp_A1 * B[(k+1)*lda + temp_i];
 164                                 temp2 += temp_A2 * B[(k+2)*lda + temp_i];
 165                                 temp2 += temp_A3 * B[(k+3)*lda + temp_i];
 166                                 temp_i++;
 167
 168
 169                                 temp3 += temp_A0 * B[(k+0)*lda + temp_i];
 170                                 temp3 += temp_A1 * B[(k+1)*lda + temp_i];
 171                                 temp3 += temp_A2 * B[(k+2)*lda + temp_i];
 172                                 temp3 += temp_A3 * B[(k+3)*lda + temp_i];
 173                                 temp_i++;
 174
 175                                 temp4 += temp_A0 * B[(k+0)*lda + temp_i];
 176                                 temp4 += temp_A1 * B[(k+1)*lda + temp_i];
 177                                 temp4 += temp_A2 * B[(k+2)*lda + temp_i];
 178                                 temp4 += temp_A3 * B[(k+3)*lda + temp_i];
 179                                 temp_i++;
 180
 181                                 temp5 += temp_A0 * B[(k+0)*lda + temp_i];
 182                                 temp5 += temp_A1 * B[(k+1)*lda + temp_i];
 183                                 temp5 += temp_A2 * B[(k+2)*lda + temp_i];
 184                                 temp5 += temp_A3 * B[(k+3)*lda + temp_i];
 185                                 temp_i++;
 186
 187                                 temp6 += temp_A0 * B[(k+0)*lda + temp_i];
 188                                 temp6 += temp_A1 * B[(k+1)*lda + temp_i];
 189                                 temp6 += temp_A2 * B[(k+2)*lda + temp_i];
 190                                 temp6 += temp_A3 * B[(k+3)*lda + temp_i];
 191                                 temp_i++;
 192
 193
 194                                 temp7 += temp_A0 * B[(k+0)*lda + temp_i];
 195                                 temp7 += temp_A1 * B[(k+1)*lda + temp_i];
 196                                 temp7 += temp_A2 * B[(k+2)*lda + temp_i];
 197                                 temp7 += temp_A3 * B[(k+3)*lda + temp_i];
 198                                 temp_i++;
 199
 200                        }
 201
 202                                         C[i + j*lda] = temp0;
 203                                         C[(i+1) + j*lda] = temp1;
 204                                         C[(i+2) + j*lda] = temp2;
 205                                         C[(i+3) + j*lda] = temp3;
 206                                         C[(i+4) + j*lda] = temp4;
 207                                         C[(i+5) + j*lda] = temp5;
 208                                         C[(i+6) + j*lda] = temp6;
 209                                         C[(i+7) + j*lda] = temp7;
 210
 211                     }
 212                   }
 213 }
 214
 215 //--------------------------------------------------------------------------
 216 // Main
 217 //
 218 // all threads start executing thread_entry(). Use their "coreid" to
 219 // differentiate between threads (each thread is running on a separate core).
 220
 221 void thread_entry(int cid, int nc)
 222 {
 223    coreid = cid;
 224    ncores = nc;
 225
 226    // static allocates data in the binary, which is visible to both threads
 227    static data_t results_data[ARRAY_SIZE];
 228
 229         /*
 230    // Execute the provided, naive matmul
 231    barrier();
 232    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 233
 234
 235    // verify
 236    verify(ARRAY_SIZE, results_data, verify_data);
 237
 238    // clear results from the first trial
 239    size_t i;
 240    if (coreid == 0)
 241       for (i=0; i < ARRAY_SIZE; i++)
 242          results_data[i] = 0;
 243    barrier();
 244
 245    */
 246    // Execute your faster matmul
 247    barrier();
 248    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 249
 250 #ifdef DEBUG
 251    printArray("results:", ARRAY_SIZE, results_data);
 252    printArray("verify :", ARRAY_SIZE, verify_data);
 253 #endif
 254
 255    // verify
 256    verify(ARRAY_SIZE, results_data, verify_data);
 257    barrier();
 258
 259    exit(0);
 260 }