mt/ba_matmul/matmul_mi.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 /*
  93    int i, j, k;
  94
  95    if (coreid > 0)
  96       return;
  97
  98    for ( i = 0; i < lda; i++ )
  99       for ( j = 0; j < lda; j++ )
 100       {
 101          for ( k = 0; k < lda; k++ )
 102          {
 103             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 104          }
 105       }
 106 */
 107 }
 108
 109
 110
 111 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 112 {
 113         size_t c_start = lda / ncores * coreid;
 114         size_t c_row;
 115         size_t c_col;
 116         size_t colSplit = 0;
 117         size_t i;
 118         size_t useSplit = 0;
 119         data_t a1;
 120         data_t a2;
 121         data_t a3;
 122         data_t a4;
 123         data_t a5;
 124         data_t a6;
 125         data_t a7;
 126         data_t a8;
 127         data_t c1;
 128         data_t c2;
 129         data_t c3;
 130         data_t c4;
 131         data_t c5;
 132         data_t c6;
 133         data_t c7;
 134         data_t c8;
 135         size_t block;
 136         for (block = 0; block < 2; block++) {
 137                 for (colSplit = 0; colSplit < 4; colSplit++) {
 138                         useSplit = (coreid == 0) ? colSplit : (colSplit + 2 ) % 4;
 139                         for (c_row = c_start + block * 8; c_row < c_start + block * 8 + 8; c_row += 2) {
 140                                 for (c_col = 0; c_col < lda; c_col+=4) {
 141                                         c1 = C[c_row*lda+c_col];
 142                                         c2 = C[(c_row+1)*lda+c_col];
 143                                         c3 = C[c_row*lda+c_col+1];
 144                                         c4 = C[(c_row+1)*lda+c_col+1];
 145                                         c5 = C[c_row*lda+c_col+2];
 146                                         c6 = C[(c_row+1)*lda+c_col+2];
 147                                         c7 = C[c_row*lda+c_col+3];
 148                                         c8 = C[(c_row+1)*lda+c_col+3];
 149                                         for (i = useSplit * lda / 4; i < (useSplit + 1) * lda / 4; i+=4) {
 150                                                 a1 = A[c_row*lda+i];
 151                                                 a2 = A[(c_row+1)*lda+i];
 152                                                 a3 = A[c_row*lda+i+1];
 153                                                 a4 = A[(c_row+1)*lda+i+1];
 154                                                 a5 = A[c_row*lda+i+2];
 155                                                 a6 = A[(c_row+1)*lda+i+2];
 156                                                 a7 = A[c_row*lda+i+3];
 157                                                 a8 = A[(c_row+1)*lda+i+3];
 158
 159                                                 c1 += a1 * B[i*lda+c_col];
 160                                                 c2 += a2 * B[i*lda+c_col];
 161
 162                                                 c1 += a3 * B[(i+1)*lda+c_col];
 163                                                 c2 += a4 * B[(i+1)*lda+c_col];
 164
 165                                                 c1 += a5 * B[(i+2)*lda+c_col];
 166                                                 c2 += a6 * B[(i+2)*lda+c_col];
 167
 168                                                 c1 += a7 * B[(i+3)*lda+c_col];
 169                                                 c2 += a8 * B[(i+3)*lda+c_col];
 170
 171                                                 c3 += a1 * B[i*lda+c_col+1];
 172                                                 c4 += a2 * B[i*lda+c_col+1];
 173
 174                                                 c3 += a3 * B[(i+1)*lda+c_col+1];
 175                                                 c4 += a4 * B[(i+1)*lda+c_col+1];
 176
 177                                                 c3 += a5 * B[(i+2)*lda+c_col+1];
 178                                                 c4 += a6 * B[(i+2)*lda+c_col+1];
 179
 180                                                 c3 += a7 * B[(i+3)*lda+c_col+1];
 181                                                 c4 += a8 * B[(i+3)*lda+c_col+1];
 182
 183                                                 c5 += a1 * B[i*lda+c_col+2];
 184                                                 c6 += a2 * B[i*lda+c_col+2];
 185
 186                                                 c5 += a3 * B[(i+1)*lda+c_col+2];
 187                                                 c6 += a4 * B[(i+1)*lda+c_col+2];
 188
 189                                                 c5 += a5 * B[(i+2)*lda+c_col+2];
 190                                                 c6 += a6 * B[(i+2)*lda+c_col+2];
 191
 192                                                 c5 += a7 * B[(i+3)*lda+c_col+2];
 193                                                 c6 += a8 * B[(i+3)*lda+c_col+2];
 194
 195                                                 c7 += a1 * B[i*lda+c_col+3];
 196                                                 c8 += a2 * B[i*lda+c_col+3];
 197
 198                                                 c7 += a3 * B[(i+1)*lda+c_col+3];
 199                                                 c8 += a4 * B[(i+1)*lda+c_col+3];
 200
 201                                                 c7 += a5 * B[(i+2)*lda+c_col+3];
 202                                                 c8 += a6 * B[(i+2)*lda+c_col+3];
 203
 204                                                 c7 += a7 * B[(i+3)*lda+c_col+3];
 205                                                 c8 += a8 * B[(i+3)*lda+c_col+3];
 206                                         }
 207
 208                                         C[c_row*lda+c_col] = c1;
 209                                         C[(c_row+1)*lda+c_col] = c2;
 210
 211                                         C[c_row*lda+c_col+1] = c3;
 212                                         C[(c_row+1)*lda+c_col+1] = c4;
 213
 214                                         C[c_row*lda+c_col+2] = c5;
 215                                         C[(c_row+1)*lda+c_col+2] = c6;
 216
 217                                         C[c_row*lda+c_col+3] = c7;
 218                                         C[(c_row+1)*lda+c_col+3] = c8;
 219                                 }
 220                         }
 221                 }
 222         }
 223 }
 224
 225 //--------------------------------------------------------------------------
 226 // Main
 227 //
 228 // all threads start executing thread_entry(). Use their "coreid" to
 229 // differentiate between threads (each thread is running on a separate core).
 230
 231 void thread_entry(int cid, int nc)
 232 {
 233    coreid = cid;
 234    ncores = nc;
 235
 236    // static allocates data in the binary, which is visible to both threads
 237    static data_t results_data[ARRAY_SIZE];
 238
 239
 240    // Execute the provided, naive matmul
 241 //   barrier();
 242 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 243
 244
 245    // verify
 246 //   verify(ARRAY_SIZE, results_data, verify_data);
 247
 248    // clear results from the first trial
 249 //   size_t i;
 250 //   if (coreid == 0)
 251 //      for (i=0; i < ARRAY_SIZE; i++)
 252 //         results_data[i] = 0;
 253 //   barrier();
 254
 255
 256    // Execute your faster matmul
 257    barrier();
 258    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 259
 260 #ifdef DEBUG
 261    printArray("results:", ARRAY_SIZE, results_data);
 262    printArray("verify :", ARRAY_SIZE, verify_data);
 263 #endif
 264
 265    // verify
 266    verify(ARRAY_SIZE, results_data, verify_data);
 267    barrier();
 268
 269    exit(0);
 270 }
 271