mt/ap_matmul/ap_matmul.c

   1 //**************************************************************************
   2 // Multi-threaded Matrix Multiply benchmark
   3 //--------------------------------------------------------------------------
   4 // TA     : Christopher Celio
   5 // Student: ME STEPHANIE TUNG
   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, kk;
 119    int block = lda / ncores;
 120    int leftover = lda % ncores;
 121    int start = block * coreid;
 122
 123
 124
 125    for ( j = start; j < (start+block); j++ )
 126       for ( k = 0; k < lda; k++ )
 127       {
 128          for ( i = 0; i < lda; i++ )
 129          {
 130             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 131          }
 132       }
 133 /*
 134
 135
 136    for ( j = coreid; j < lda; j += ncores )
 137       for ( k = 0; k < lda; k++ )
 138       {
 139          for ( i = 0; i < lda; i++ )
 140          {
 141             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 142          }
 143       }
 144
 145 /*
 146    if (coreid > 0) {
 147       return;
 148    }
 149
 150    for ( j = (lda - leftover); j < lda; j++ )
 151       for ( i = 0; i < lda; i++ )
 152       {
 153          for ( k = 0; k < lda; k++ )
 154          {
 155             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 156          }
 157       }
 158
 159
 160
 161 if (coreid > 0) {
 162         return;
 163 }
 164
 165
 166
 167 for (jj = start; jj < start+block; jj += 4) {
 168    for (kk = 0; kk < lda; kk += 4) {
 169       for (ii = 0; ii < lda; ii += 4) {
 170          for (i = ii; i < ii+4; i += 4) {
 171             //float * p = B + i;
 172             for (j = jj; j < jj+4; j++) {
 173                for (k = kk; k < kk+4; k++) {
 174
 175                   float a = A[k + j*lda];
 176
 177                   C[i + j*lda] += a * B[k*lda + i];
 178                   C[i + j*lda + 1] += a * B[k*lda + i + 1];
 179                   C[i + j*lda + 2] += a * B[k*lda + i + 2];
 180                   C[i + j*lda + 3] += a * B[k*lda + i + 3];
 181                }
 182             }
 183          }
 184       }
 185    }
 186 }
 187
 188 */
 189
 190 }
 191
 192 //--------------------------------------------------------------------------
 193 // Main
 194 //
 195 // all threads start executing thread_entry(). Use their "coreid" to
 196 // differentiate between threads (each thread is running on a separate core).
 197
 198 void thread_entry(int cid, int nc)
 199 {
 200    coreid = cid;
 201    ncores = nc;
 202
 203    // static allocates data in the binary, which is visible to both threads
 204    static data_t results_data[ARRAY_SIZE];
 205
 206
 207 //   // Execute the provided, naive matmul
 208 //   barrier();
 209 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 210 //
 211 //
 212 //   // verify
 213 //   verify(ARRAY_SIZE, results_data, verify_data);
 214 //
 215 //   // clear results from the first trial
 216 //   size_t i;
 217 //   if (coreid == 0)
 218 //      for (i=0; i < ARRAY_SIZE; i++)
 219 //         results_data[i] = 0;
 220 //   barrier();
 221
 222
 223    // Execute your faster matmul
 224    barrier();
 225    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 226
 227 #ifdef DEBUG
 228    printArray("results:", ARRAY_SIZE, results_data);
 229    printArray("verify :", ARRAY_SIZE, verify_data);
 230 #endif
 231
 232    // verify
 233    verify(ARRAY_SIZE, results_data, verify_data);
 234    barrier();
 235
 236    exit(0);
 237 }
 238