mt/af_matmul/Ronald.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 //--------------------------------------------------------------------------
  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     int i,j,k,l;
 112     data_t element1, element2, element3, element4, element5, element6, element7, element8;
 113     int row, row2;
 114     int column1, column2, column3, column4, column5, column6, column7, column8;
 115     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};
 116     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};
 117     if (coreid == 0){
 118       for (i=0; i<32; i+=2){
 119         row = i*32;
 120         row2 = (i+1)*32;
 121         for (j=0; j<16; j+=4){
 122           element1 = A[row+j];
 123           element2 = A[row+j+1];
 124           element3 = A[row+j+2];
 125           element4 = A[row+j+3];
 126           column1 = j*32;
 127           column2 = (j+1)*32;
 128           column3 = (j+2)*32;
 129           column4 = (j+3)*32;
 130           element5 = A[row2+j];
 131           element6 = A[row2+j+1];
 132           element7 = A[row2+j+2];
 133           element8 = A[row2+j+3];
 134
 135           for (k=0; k<32; k+=4){
 136             temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
 137             temp[k+1]+=element1*B[column1+k+1]+element2*B[column2+k+1]+element3*B[column3+k+1]+element4*B[column4+k+1];
 138             temp[k+2]+=element1*B[column1+k+2]+element2*B[column2+k+2]+element3*B[column3+k+2]+element4*B[column4+k+2];
 139             temp[k+3]+=element1*B[column1+k+3]+element2*B[column2+k+3]+element3*B[column3+k+3]+element4*B[column4+k+3];
 140             temp2[k]+=element5*B[column1+k]+element6*B[column2+k]+element7*B[column3+k]+element8*B[column4+k];
 141             temp2[k+1]+=element5*B[column1+k+1]+element6*B[column2+k+1]+element7*B[column3+k+1]+element8*B[column4+k+1];
 142             temp2[k+2]+=element5*B[column1+k+2]+element6*B[column2+k+2]+element7*B[column3+k+2]+element8*B[column4+k+2];
 143             temp2[k+3]+=element5*B[column1+k+3]+element6*B[column2+k+3]+element7*B[column3+k+3]+element8*B[column4+k+3];
 144           }
 145           if (j==12){
 146             for (l=0; l<32; l++){
 147               C[row+l]+=temp[l];
 148               C[row2+l]+=temp2[l];
 149               temp[l]=0;
 150               temp2[l]=0;
 151             }
 152           }
 153         }
 154       }
 155     }
 156     else if (coreid==1){
 157       for (i=0; i<32; i+=2){
 158         row = (31-i)*32;
 159         row2 = (31-i-1)*32;
 160         for (j=16; j<32; j+=4){
 161           element1 = A[row+j];
 162           element2 = A[row+j+1];
 163           element3 = A[row+j+2];
 164           element4 = A[row+j+3];
 165           element5 = A[row2+j];
 166           element6 = A[row2+j+1];
 167           element7 = A[row2+j+2];
 168           element8 = A[row2+j+3];
 169           column1 = j*32;
 170           column2 = (j+1)*32;
 171           column3 = (j+2)*32;
 172           column4 = (j+3)*32;
 173           for (k=0; k<32; k+=4){
 174             temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
 175             temp[k+1]+=element1*B[column1+k+1]+element2*B[column2+k+1]+element3*B[column3+k+1]+element4*B[column4+k+1];
 176             temp[k+2]+=element1*B[column1+k+2]+element2*B[column2+k+2]+element3*B[column3+k+2]+element4*B[column4+k+2];
 177             temp[k+3]+=element1*B[column1+k+3]+element2*B[column2+k+3]+element3*B[column3+k+3]+element4*B[column4+k+3];
 178             temp2[k]+=element5*B[column1+k]+element6*B[column2+k]+element7*B[column3+k]+element8*B[column4+k];
 179             temp2[k+1]+=element5*B[column1+k+1]+element6*B[column2+k+1]+element7*B[column3+k+1]+element8*B[column4+k+1];
 180             temp2[k+2]+=element5*B[column1+k+2]+element6*B[column2+k+2]+element7*B[column3+k+2]+element8*B[column4+k+2];
 181             temp2[k+3]+=element5*B[column1+k+3]+element6*B[column2+k+3]+element7*B[column3+k+3]+element8*B[column4+k+3];
 182           }
 183           if (j==28){
 184             for (l=0; l<32; l++){
 185               C[row+l]+=temp[l];
 186               C[row2+l]+=temp2[l];
 187               temp[l]=0;
 188               temp2[l]=0;
 189             }
 190           }
 191         }
 192       }
 193       }
 194    // ***************************** //
 195    // **** ADD YOUR CODE HERE ***** //
 196    // ***************************** //
 197    //
 198    // feel free to make a separate function for MI and MSI versions.
 199
 200 }
 201 //--------------------------------------------------------------------------
 202 // Main
 203 //
 204 // all threads start executing thread_entry(). Use their "coreid" to
 205 // differentiate between threads (each thread is running on a separate core).
 206
 207 void thread_entry(int cid, int nc)
 208 {
 209    coreid = cid;
 210    ncores = nc;
 211
 212    // static allocates data in the binary, which is visible to both threads
 213    static data_t results_data[ARRAY_SIZE];
 214
 215
 216    // Execute the provided, naive matmul
 217    barrier(nc);
 218    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 219
 220
 221    // verify
 222    verifyMT(ARRAY_SIZE, results_data, verify_data);
 223
 224    // clear results from the first trial
 225    size_t i;
 226    if (coreid == 0)
 227       for (i=0; i < ARRAY_SIZE; i++)
 228          results_data[i] = 0;
 229    barrier(nc);
 230
 231
 232    // Execute your faster matmul
 233    barrier(nc);
 234    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 235
 236 #ifdef DEBUG
 237    printArrayMT("results:", ARRAY_SIZE, results_data);
 238    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 239 #endif
 240
 241    // verify
 242    verifyMT(ARRAY_SIZE, results_data, verify_data);
 243    barrier(nc);
 244
 245    exit(0);
 246 }