mt/af_matmul/bestattemptthusfar2.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 data_t mult(data_t x, data_t y)
  87 { data_t result = 0;
  88         size_t i;
  89         for (i=0; i < x; i++) {
  90                 result += y;
  91         }
  92         return result;
  93 }
  94 //--------------------------------------------------------------------------
  95 // matmul function
  96
  97 // single-thread, naive version
  98 void __attribute__((noinline)) matmul_naive(const int lda,  const data_t A[], const data_t B[], data_t C[] )
  99 {
 100    int i, j, k;
 101
 102    if (coreid > 0)
 103       return;
 104
 105    for ( i = 0; i < lda; i++ )
 106       for ( j = 0; j < lda; j++ )
 107       {
 108          for ( k = 0; k < lda; k++ )
 109          {
 110             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 111          }
 112       }
 113
 114 }
 115
 116
 117 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 118 {
 119
 120   size_t i, j, k, l;
 121   int row,row2, column, column2, column3, column4, column5, column6, column7, column8;
 122   data_t element, element2, element3, element4, element5, element6, element7, element8;
 123         data_t B1, B2, B3, B4;
 124   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};
 125   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};
 126         int local_lda = lda;
 127   //for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
 128   for (l=coreid*local_lda/ncores; l<local_lda*(1+coreid)/ncores; l+=2){
 129     row=l*32;
 130     row2=(l+1)*32;
 131                 //element = A[row];
 132                 //element5 = A[row2];
 133     for (i=0; i<local_lda; i+=4){
 134       element = A[row+i];
 135       element2 = A[row+i+1];
 136       element3 = A[row+i+2];
 137       element4 = A[row+i+3];
 138
 139       element5 = A[row2+i];
 140       element6 = A[row2+i+1];
 141       element7 = A[row2+i+2];
 142       element8 = A[row2+i+3];
 143
 144       column=i*local_lda;
 145       column2=(i+1)*local_lda;
 146       column3=(i+2)*local_lda;
 147       column4=(i+3)*local_lda;
 148
 149                         B1 = B[column];
 150                         B2 = B[column2];
 151                         B3 = B[column3];
 152                         B4 = B[column4];
 153
 154       for (j=0; j<lda; j+=4){
 155                                 temp_mat[j]+=element*B1+element2*B2+element3*B3+element4*B4;
 156                                 temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1];
 157                                 temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2];
 158                                 temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3];
 159
 160                                 temp_mat2[j]+=element5*B1+element6*B2+element7*B3+element8*B4;
 161                                 temp_mat2[j+1]+=element5*B[column+j+1]+element6*B[column2+j+1]+element7*B[column3+j+1]+element8*B[column4+j+1];
 162                                 temp_mat2[j+2]+=element5*B[column+j+2]+element6*B[column2+j+2]+element7*B[column3+j+2]+element8*B[column4+j+2];
 163                                 temp_mat2[j+3]+=element5*B[column+j+3]+element6*B[column2+j+3]+element7*B[column3+j+3]+element8*B[column4+j+3];
 164
 165                                 B1 = B[column+j+4];
 166                                 B2 = B[column2+j+4];
 167                                 B3 = B[column3+j+4];
 168                                 B4 = B[column4+j+4];
 169
 170                                 }
 171                         //element = A[row+i+4];
 172                         //element5 = A[row2+i+4];
 173       }
 174
 175                         for(k=0; k<local_lda; k++){
 176                                 C[row+k]=temp_mat[k];
 177                                 temp_mat[k]=0;
 178                                 C[row2+k]=temp_mat2[k];
 179                                 temp_mat2[k]=0;
 180
 181                                 }
 182
 183
 184   }
 185
 186    // ***************************** //
 187    // **** ADD YOUR CODE HERE ***** //
 188    // ***************************** //
 189    //
 190    // feel free to make a separate function for MI and MSI versions.
 191
 192 }
 193 //--------------------------------------------------------------------------
 194 // Main
 195 //
 196 // all threads start executing thread_entry(). Use their "coreid" to
 197 // differentiate between threads (each thread is running on a separate core).
 198
 199 void thread_entry(int cid, int nc)
 200 {
 201    coreid = cid;
 202    ncores = nc;
 203
 204    // static allocates data in the binary, which is visible to both threads
 205    static data_t results_data[ARRAY_SIZE];
 206
 207
 208    // Execute the provided, naive matmul
 209    barrier(nc);
 210    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 211
 212
 213    // verify
 214    verifyMT(ARRAY_SIZE, results_data, verify_data);
 215
 216    // clear results from the first trial
 217    size_t i;
 218    if (coreid == 0)
 219       for (i=0; i < ARRAY_SIZE; i++)
 220          results_data[i] = 0;
 221    barrier(nc);
 222
 223
 224    // Execute your faster matmul
 225    barrier(nc);
 226    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 227
 228 #ifdef DEBUG
 229    printArrayMT("results:", ARRAY_SIZE, results_data);
 230    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 231 #endif
 232
 233    // verify
 234    verifyMT(ARRAY_SIZE, results_data, verify_data);
 235    barrier(nc);
 236
 237    exit(0);
 238 }