mt/ak_matmul/ak_matmul.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
 113    // ***************************** //
 114    // **** ADD YOUR CODE HERE ***** //
 115    // ***************************** //
 116    //
 117    // feel free to make a separate function for MI and MSI versions.
 118   int i, j, k, ii, jj, bsize;
 119   bsize = 16;
 120   for ( jj = bsize*coreid; jj < lda; jj += bsize*ncores) {
 121     for ( ii = 0; ii < lda; ii += bsize) {
 122       for ( j = jj; j < lda && j < jj + bsize; j++) {
 123         for ( i = ii; i < lda && i < ii + bsize; i += 8) {
 124           data_t c1 = C[i + j*lda];
 125           data_t c2 = C[i + j*lda + 1];
 126           data_t c3 = C[i + j*lda + 2];
 127           data_t c4 = C[i + j*lda + 3];
 128           data_t c5 = C[i + j*lda + 4];
 129           data_t c6 = C[i + j*lda + 5];
 130           data_t c7 = C[i + j*lda + 6];
 131           data_t c8 = C[i + j*lda + 7];
 132           for ( k = 0; k < lda; k+=4 ) {
 133             for (int x = 0; x < 4; x++) {
 134               data_t a = A[j*lda + k+x];
 135               data_t b1 = B[(k+x)*lda + i];
 136               data_t b2 = B[(k+x)*lda + i + 1];
 137               data_t b3 = B[(k+x)*lda + i + 2];
 138               data_t b4 = B[(k+x)*lda + i + 3];
 139               data_t b5 = B[(k+x)*lda + i + 4];
 140               data_t b6 = B[(k+x)*lda + i + 5];
 141               data_t b7 = B[(k+x)*lda + i + 6];
 142               data_t b8 = B[(k+x)*lda + i + 7];
 143               c1 += a * b1;
 144               c2 += a * b2;
 145               c3 += a * b3;
 146               c4 += a * b4;
 147               c5 += a * b5;
 148               c6 += a * b6;
 149               c7 += a * b7;
 150               c8 += a * b8;
 151             }
 152           }
 153           C[i + j*lda] = c1;
 154           C[i + j*lda + 1] = c2;
 155           C[i + j*lda + 2] = c3;
 156           C[i + j*lda + 3] = c4;
 157           C[i + j*lda + 4] = c5;
 158           C[i + j*lda + 5] = c6;
 159           C[i + j*lda + 6] = c7;
 160           C[i + j*lda + 7] = c8;
 161         }
 162       }
 163     }
 164   }
 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 }
 213