mt/am_matmul/matmul4.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 j;
 114   size_t k;
 115   size_t max_dim = 32*32;
 116   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};
 117  for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
 118     data_t element=A[i];
 119     data_t element2 = A[i+1];
 120     data_t element3 = A[i+2];
 121     data_t element4 = A[i+3];
 122     data_t element5 = A[i+4];
 123     data_t element6 = A[i+5];
 124     data_t element7 = A[i+6];
 125     data_t element8 = A[i+7];
 126     int row= (int)(i/32)*32;
 127     int column = i%32*32;
 128     int column2 = (i+1)%32*32;
 129     int column3 = (i+2)%32*32;
 130     int column4 = (i+3)%32*32;
 131     int column5 = (i+4)%32*32;
 132     int column6 = (i+5)%32*32;
 133     int column7 = (i+6)%32*32;
 134     int column8 = (i+7)%32*32;
 135
 136     for (j=0; j<32; j++){
 137       temp_mat[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];
 138     }
 139     if (i%32==24){
 140       for(k=0; k<32; k++){
 141         C[row+k]=temp_mat[k];
 142         temp_mat[k]=0;
 143       }
 144     }
 145     }*/
 146   int i,j,k,l;
 147   //data_t element11, element12, element13, element14, element21, element22, element23, element24;
 148   data_t element1, element2, element3, element4, element5, element6, element7, element8;
 149   int row, row2;
 150   //int column11, column12, column13, column14, column21, column22, column23, column24;
 151   int column1, column2, column3, column4, column5, column6, column7, column8;
 152   data_t temp[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};
 153   //data_t temp2[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};
 154   if (coreid == 0){
 155     for (i=0; i<32; i++){
 156       if (i==15){
 157         for (j=0; j<32; j+=4){
 158           row=15*32;
 159           element1 = A[row+j];
 160           element2 = A[row+j+1];
 161           element3 = A[row+j+2];
 162           element4 = A[row+j+3];
 163           column1 = j*32;
 164           column2 = (j+1)*32;
 165           column3 = (j+2)*32;
 166           column4 = (j+3)*32;
 167           for (k=0;k<32; k++){
 168             temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
 169           }
 170           if (j==28){
 171             for (l=0; l<32; l++){
 172               C[row+l]=temp[l];
 173               temp[l]=0;
 174             }
 175           }
 176         }
 177       }
 178       else{
 179         row = i*32;
 180         for (j=0; j<16; j+=4){
 181           element1 = A[i*32+j];
 182           element2 = A[i*32+j+1];
 183           element3 = A[i*32+j+2];
 184           element4 = A[i*32+j+3];
 185           column1 = j*32;
 186           column2 = (j+1)*32;
 187           column3 = (j+2)*32;
 188           column4 = (j+3)*32;
 189           for (k=0; k<32; k++){
 190             temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
 191           }
 192           if (j==12){
 193             for (l=0; l<32; l++){
 194               C[row+l]+=temp[l];
 195               temp[l]=0;
 196             }
 197           }
 198         }
 199       }
 200     }
 201   }
 202   else if (coreid==1){
 203     for (i=0; i<32; i++){
 204       row = (31-i)*32;
 205       if (row/32 != 15){
 206         for (j=16; j<32; j+=4){
 207           element1 = A[(31-i)*32+j];
 208           element2 = A[(31-i)*32+j+1];
 209           element3 = A[(31-i)*32+j+2];
 210           element4 = A[(31-i)*32+j+3];
 211           column1 = j*32;
 212           column2 = (j+1)*32;
 213           column3 = (j+2)*32;
 214           column4 = (j+3)*32;
 215           for (k=0; k<32; k++){
 216             temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
 217           }
 218           if (j==28){
 219             for (l=0; l<32; l++){
 220               C[row+l]+=temp[l];
 221               temp[l]=0;
 222             }
 223           }
 224         }
 225       }
 226     }
 227   }
 228    // ***************************** //
 229    // **** ADD YOUR CODE HERE ***** //
 230    // ***************************** //
 231    //
 232    // feel free to make a separate function for MI and MSI versions.
 233
 234 }
 235
 236 //--------------------------------------------------------------------------
 237 // Main
 238 //
 239 // all threads start executing thread_entry(). Use their "coreid" to
 240 // differentiate between threads (each thread is running on a separate core).
 241
 242 void thread_entry(int cid, int nc)
 243 {
 244    coreid = cid;
 245    ncores = nc;
 246
 247    // static allocates data in the binary, which is visible to both threads
 248    static data_t results_data[ARRAY_SIZE];
 249
 250
 251    // Execute the provided, naive matmul
 252    barrier();
 253    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 254
 255
 256    // verify
 257    verify(ARRAY_SIZE, results_data, verify_data);
 258
 259    // clear results from the first trial
 260    size_t i;
 261    if (coreid == 0)
 262       for (i=0; i < ARRAY_SIZE; i++)
 263          results_data[i] = 0;
 264    barrier();
 265
 266
 267    // Execute your faster matmul
 268    barrier();
 269    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 270
 271 #ifdef DEBUG
 272    printArray("results:", ARRAY_SIZE, results_data);
 273    printArray("verify :", ARRAY_SIZE, verify_data);
 274 #endif
 275
 276    // verify
 277    verify(ARRAY_SIZE, results_data, verify_data);
 278    barrier();
 279
 280    exit(0);
 281 }
 282