mt/af_matmul/bestattemptthusfar.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 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+=4){
 118     row=l*lda;
 119     row2=(l+1)*lda;
 120     for (i=0; i<lda; i+=4){
 121       element = A[row+i];
 122       element2 = A[row+i+1];
 123       element3 = A[row+i+2];
 124       element4 = A[row+i+3];
 125
 126       element5 = A[row2+i];
 127       element6 = A[row2+i+1];
 128       element7 = A[row2+i+2];
 129       element8 = A[row2+i+3];
 130
 131       column=i*lda;
 132       column2=(i+1)*lda;
 133       column3=(i+2)*lda;
 134       column4=(i+3)*lda;
 135
 136       for (j=0; j<lda; j+=4){
 137                                 temp_mat[j]+=element*B[column+j]+element2*B[column2+j]+element3*B[column3+j]+element4*B[column4+j];
 138                                 temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1];
 139                                 temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2];
 140                                 temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3];
 141
 142                                 temp_mat2[j]+=element5*B[column+j]+element6*B[column2+j]+element7*B[column3+j]+element8*B[column4+j];
 143                                 temp_mat2[j+1]+=element5*B[column+j+1]+element6*B[column2+j+1]+element7*B[column3+j+1]+element8*B[column4+j+1];
 144                                 temp_mat2[j+2]+=element5*B[column+j+2]+element6*B[column2+j+2]+element7*B[column3+j+2]+element8*B[column4+j+2];
 145                                 temp_mat2[j+3]+=element5*B[column+j+3]+element6*B[column2+j+3]+element7*B[column3+j+3]+element8*B[column4+j+3];
 146       }
 147
 148       }
 149
 150                         for(k=0; k<32; k++){
 151                                 C[row+k]=temp_mat[k];
 152                                 temp_mat[k]=0;
 153                                 C[row2+k]=temp_mat2[k];
 154                                 temp_mat2[k]=0;
 155
 156
 157                                 }
 158   }
 159
 160    // ***************************** //
 161    // **** ADD YOUR CODE HERE ***** //
 162    // ***************************** //
 163    //
 164    // feel free to make a separate function for MI and MSI versions.
 165
 166 }
 167 //--------------------------------------------------------------------------
 168 // Main
 169 //
 170 // all threads start executing thread_entry(). Use their "coreid" to
 171 // differentiate between threads (each thread is running on a separate core).
 172
 173 void thread_entry(int cid, int nc)
 174 {
 175    coreid = cid;
 176    ncores = nc;
 177
 178    // static allocates data in the binary, which is visible to both threads
 179    static data_t results_data[ARRAY_SIZE];
 180
 181
 182    // Execute the provided, naive matmul
 183    barrier();
 184    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 185
 186
 187    // verify
 188    verify(ARRAY_SIZE, results_data, verify_data);
 189
 190    // clear results from the first trial
 191    size_t i;
 192    if (coreid == 0)
 193       for (i=0; i < ARRAY_SIZE; i++)
 194          results_data[i] = 0;
 195    barrier();
 196
 197
 198    // Execute your faster matmul
 199    barrier();
 200    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 201
 202 #ifdef DEBUG
 203    printArray("results:", ARRAY_SIZE, results_data);
 204    printArray("verify :", ARRAY_SIZE, verify_data);
 205 #endif
 206
 207    // verify
 208    verify(ARRAY_SIZE, results_data, verify_data);
 209    barrier();
 210
 211    exit(0);
 212 }