mt/af_matmul/keeptrying3.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 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   size_t i, j, k, l;
 120   int row,row2, row3, row4, column, column2, column3, column4, column5, column6, column7, column8;
 121   data_t element, element2, element3, element4, element5, element6, element7, element8;
 122   data_t element9, element10, element11, element12, element13, element14, element15, element16;
 123         data_t elementB1,elementB2,elementB3,elementB4;
 124   data_t temp_mat[128]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 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   //for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
 127   for (l=coreid*32/ncores; l<32*(1+coreid)/ncores; l+=4){
 128     row=l*lda;
 129     row2=(l+1)*lda;
 130     row3=(l+2)*lda;
 131     row4=(l+3)*lda;
 132     for (i=0; i<lda; i+=4){
 133       element = A[row+i];
 134       element2 = A[row+i+1];
 135       element3 = A[row+i+2];
 136       element4 = A[row+i+3];
 137
 138       element5 = A[row2+i];
 139       element6 = A[row2+i+1];
 140       element7 = A[row2+i+2];
 141       element8 = A[row2+i+3];
 142
 143                 element9 = A[row3+i];
 144       element10 = A[row3+i+1];
 145       element11 = A[row3+i+2];
 146       element12 = A[row3+i+3];
 147
 148       element13 = A[row4+i];
 149       element14 = A[row4+i+1];
 150       element15 = A[row4+i+2];
 151       element16 = A[row4+i+3];
 152
 153       column=i*lda;
 154       column2=(i+1)*lda;
 155       column3=(i+2)*lda;
 156       column4=(i+3)*lda;
 157
 158
 159       for (j=0; j<lda; j+=4){
 160
 161                                 temp_mat[j]+=element*B[column+j]+element2*B[column2+j]+element3*B[column3+j]+element4*B[column4+j];
 162                                 temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1];
 163                                 temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2];
 164                                 temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3];
 165
 166                                 temp_mat[j+lda]+=element5*B[column+j]+element6*B[column2+j]+element7*B[column3+j]+element8*B[column4+j];
 167                                 temp_mat[j+1+lda]+=element5*B[column+j+1]+element6*B[column2+j+1]+element7*B[column3+j+1]+element8*B[column4+j+1];
 168                                 temp_mat[j+2+lda]+=element5*B[column+j+2]+element6*B[column2+j+2]+element7*B[column3+j+2]+element8*B[column4+j+2];
 169                                 temp_mat[j+3+lda]+=element5*B[column+j+3]+element6*B[column2+j+3]+element7*B[column3+j+3]+element8*B[column4+j+3];
 170
 171                                 temp_mat[j+2*lda]+=element9*B[column+j]+element10*B[column2+j]+element11*B[column3+j]+element12*B[column4+j];
 172                                 temp_mat[j+1+2*lda]+=element9*B[column+j+1]+element10*B[column2+j+1]+element11*B[column3+j+1]+element12*B[column4+j+1];
 173                                 temp_mat[j+2+2*lda]+=element9*B[column+j+2]+element10*B[column2+j+2]+element11*B[column3+j+2]+element12*B[column4+j+2];
 174                                 temp_mat[j+3+2*lda]+=element9*B[column+j+3]+element10*B[column2+j+3]+element11*B[column3+j+3]+element12*B[column4+j+3];
 175
 176                                 temp_mat[j+3*lda]+=element13*B[column+j]+element14*B[column2+j]+element15*B[column3+j]+element16*B[column4+j];
 177                                 temp_mat[j+1+3*lda]+=element13*B[column+j+1]+element14*B[column2+j+1]+element15*B[column3+j+1]+element16*B[column4+j+1];
 178                                 temp_mat[j+2+3*lda]+=element13*B[column+j+2]+element14*B[column2+j+2]+element15*B[column3+j+2]+element16*B[column4+j+2];
 179                                 temp_mat[j+3+3*lda]+=element13*B[column+j+3]+element14*B[column2+j+3]+element15*B[column3+j+3]+element16*B[column4+j+3];
 180
 181
 182       }
 183
 184       }
 185
 186                         for(k=0; k<32; k++){
 187                                 C[row+k]=temp_mat[k];
 188                                 temp_mat[k]=0;
 189                                 C[row2+k]=temp_mat[k+lda];
 190                                 temp_mat[k+lda]=0;
 191                                 C[row3+k]=temp_mat[k+2*lda];
 192                                 temp_mat[k+2*lda]=0;
 193                                 C[row4+k]=temp_mat[k+3*lda];
 194                                 temp_mat[k+3*lda]=0;
 195
 196
 197                                 }
 198
 199
 200   }
 201    // ***************************** //
 202    // **** ADD YOUR CODE HERE ***** //
 203    // ***************************** //
 204    //
 205    // feel free to make a separate function for MI and MSI versions.
 206
 207 }
 208 //--------------------------------------------------------------------------
 209 // Main
 210 //
 211 // all threads start executing thread_entry(). Use their "coreid" to
 212 // differentiate between threads (each thread is running on a separate core).
 213
 214 void thread_entry(int cid, int nc)
 215 {
 216    coreid = cid;
 217    ncores = nc;
 218
 219    // static allocates data in the binary, which is visible to both threads
 220    static data_t results_data[ARRAY_SIZE];
 221
 222
 223    // Execute the provided, naive matmul
 224    barrier();
 225    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 226
 227
 228    // verify
 229    verify(ARRAY_SIZE, results_data, verify_data);
 230
 231    // clear results from the first trial
 232    size_t i;
 233    if (coreid == 0)
 234       for (i=0; i < ARRAY_SIZE; i++)
 235          results_data[i] = 0;
 236    barrier();
 237
 238
 239    // Execute your faster matmul
 240    barrier();
 241    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 242
 243 #ifdef DEBUG
 244    printArray("results:", ARRAY_SIZE, results_data);
 245    printArray("verify :", ARRAY_SIZE, verify_data);
 246 #endif
 247
 248    // verify
 249    verify(ARRAY_SIZE, results_data, verify_data);
 250    barrier();
 251
 252    exit(0);
 253 }