mt/am_matmul/matmul3.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   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};
 118   //for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
 119    for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores)/2; i+=8){
 120     data_t element=A[i];
 121     data_t element2 = A[i+1];
 122     data_t element3 = A[i+2];
 123     data_t element4 = A[i+3];
 124     data_t element5 = A[i+4];
 125     data_t element6 = A[i+5];
 126     data_t element7 = A[i+6];
 127     data_t element8 = A[i+7];
 128     data_t elementA2 = A[i+32*8];
 129       data_t elementA21 = A[i+32*8+1];
 130       data_t elementA22 = A[i+32*8+2];
 131       data_t elementA23 = A[i+32*8+3];
 132       data_t elementA24 = A[i+32*8+4];
 133       data_t elementA25 = A[i+32*8+5];
 134       data_t elementA26 = A[i+32*8+6];
 135       data_t elementA27 = A[i+32*8+7];
 136     int row= (int)(i/32)*32;
 137     int row2 = row+8*32;
 138     int column = i%32*32;
 139     int column2 = (i+1)%32*32;
 140     int column3 = (i+2)%32*32;
 141     int column4 = (i+3)%32*32;
 142     int column5 = (i+4)%32*32;
 143     int column6 = (i+5)%32*32;
 144     int column7 = (i+6)%32*32;
 145     int column8 = (i+7)%32*32;
 146
 147     for (j=0; j<32; j++){
 148       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];
 149
 150       temp_mat2[j]+=elementA2*B[column+j]+elementA21*B[column2+j]+elementA22*B[column3+j]+elementA23*B[column4+j]+elementA24*B[column5+j]+elementA25*B[column6+j]+elementA26*B[column7+j]+elementA27*B[column8+j];
 151     }
 152     if (i%32==24){
 153       for(k=0; k<32; k++){
 154         C[row+k]=temp_mat[k];
 155         C[row2+k]=temp_mat2[k];
 156         temp_mat[k]=0;
 157         temp_mat2[k]=0;
 158
 159       }
 160     }
 161     }
 162
 163
 164
 165
 166
 167    // ***************************** //
 168    // **** ADD YOUR CODE HERE ***** //
 169    // ***************************** //
 170    //
 171    // feel free to make a separate function for MI and MSI versions.
 172
 173 }
 174
 175 //--------------------------------------------------------------------------
 176 // Main
 177 //
 178 // all threads start executing thread_entry(). Use their "coreid" to
 179 // differentiate between threads (each thread is running on a separate core).
 180
 181 void thread_entry(int cid, int nc)
 182 {
 183    coreid = cid;
 184    ncores = nc;
 185
 186    // static allocates data in the binary, which is visible to both threads
 187    static data_t results_data[ARRAY_SIZE];
 188
 189
 190    // Execute the provided, naive matmul
 191    barrier();
 192    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 193
 194
 195    // verify
 196    verify(ARRAY_SIZE, results_data, verify_data);
 197
 198    // clear results from the first trial
 199    size_t i;
 200    if (coreid == 0)
 201       for (i=0; i < ARRAY_SIZE; i++)
 202          results_data[i] = 0;
 203    barrier();
 204
 205
 206    // Execute your faster matmul
 207    barrier();
 208    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 209
 210 #ifdef DEBUG
 211    printArray("results:", ARRAY_SIZE, results_data);
 212    printArray("verify :", ARRAY_SIZE, verify_data);
 213 #endif
 214
 215    // verify
 216    verify(ARRAY_SIZE, results_data, verify_data);
 217    barrier();
 218
 219    exit(0);
 220 }
 221