mt/as_matmul/as_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 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
 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
 119    int i, j, k, n, m;
 120
 121
 122    //matmul_naive(32, input1_data, input2_data, results_data); barrier(nc): 957424 cycles, 29.2 cycles/iter, 3.6 CPI
 123    //matmul(32, input1_data, input2_data, results_data); barrier(nc): 340408 cycles, 10.3 cycles/iter, 1.8 CPI
 124
 125    for (n = 0; n < lda; n += 1) {
 126       for (m = 0; m < lda; m += 1) {
 127          bTranspose[lda*m + n] = B[lda*n + m];
 128          bTranspose[lda*n + m] = B[lda*m + n];
 129       }
 130    }
 131    barrier(ncores);
 132
 133    for ( j = coreid; j < lda; j += 2*ncores ) {
 134       for ( i = 0; i < lda; i += 1 ){
 135          c1 = 0;     //global vars c1, c2
 136          c2 = 0;
 137          for ( k = 0; k < lda; k += 1 ) {
 138             c1 += A[j * lda + k] * bTranspose[i*lda + k];
 139             c2 += A[(j+2) * lda + k] * bTranspose[i*lda + k];
 140
 141             //barrier(nc);
 142          }
 143
 144          C[i + j * lda] = c1;
 145          C[i + (j+2) * lda] = c2;
 146          barrier(ncores);
 147       }
 148    //barrier(nc);
 149    }
 150
 151
 152
 153
 154    //matmul_naive(32, input1_data, input2_data, results_data); barrier(nc): 983609 cycles, 30.0 cycles/iter, 3.7 CPI
 155    //matmul(32, input1_data, input2_data, results_data); barrier(nc): 389942 cycles, 11.9 cycles/iter, 2.5 CPI
 156
 157    /*
 158    for ( j = coreid; j < lda; j += 2*ncores ) {
 159       for ( i = 0; i < lda; i += 1 ){
 160          c1 = 0;     //global vars c1, c2
 161          c2 = 0;
 162          for ( k = 0; k < lda; k += 1 ) {
 163             c1 += A[j * lda + k] * B[k*lda + i];
 164             c2 += A[(j+2) * lda + k] * B[k*lda + i];
 165
 166             //barrier(nc);
 167          }
 168
 169          C[i + j * lda] = c1;
 170          C[i + (j+2) * lda] = c2;
 171          barrier(nc);
 172       }
 173    //barrier(nc);
 174    }
 175    */
 176
 177    // matmul_naive(32, input1_data, input2_data, results_data); barrier(nc): 973781 cycles, 29.7 cycles/iter, 3.7 CPI
 178    // matmul(32, input1_data, input2_data, results_data); barrier(nc): 461066 cycles, 14.0 cycles/iter, 3.5 CPI
 179    // for ( k = 0; k < lda; k += 1 ) {
 180    //    for ( j = coreid; j < lda; j += 2*ncores ) {
 181    //       for ( i = 0; i < lda; i += 1 ){
 182    //          C[i + j * lda] += A[j * lda + k] * B[k*lda + i];
 183    //          C[i + (j+2) * lda] += A[(j+2) * lda + k] * B[k*lda + i];
 184    //          //barrier(nc);
 185    //       }
 186    //       barrier(nc);
 187    //    }
 188    // //barrier(nc);
 189    // }
 190
 191
 192    // matmul_naive(32, input1_data, input2_data, results_data); barrier(nc): 965136 cycles, 29.4 cycles/iter, 3.7 CPI
 193    // matmul(32, input1_data, input2_data, results_data); barrier(nc): 513779 cycles, 15.6 cycles/iter, 3.2 CPI
 194
 195    // for ( j = coreid; j < lda; j += 2*ncores ) {
 196    //    for ( i = 0; i < lda; i += 1 ){
 197    //       for ( k = 0; k < lda; k += 1 ) {
 198    //          C[i + j * lda] += A[j * lda + k] * B[k*lda + i];
 199    //          C[i + (j+2) * lda] += A[(j+2) * lda + k] * B[k*lda + i];
 200
 201    //          //barrier(nc);
 202    //       }
 203    //       barrier(nc);
 204    //    }
 205    // //barrier(nc);
 206    //}
 207
 208
 209    // matmul_naive(32, input1_data, input2_data, results_data); barrier(nc): 937892 cycles, 28.6 cycles/iter, 3.6 CPI
 210    // matmul(32, input1_data, input2_data, results_data); barrier(nc): 576478 cycles, 17.5 cycles/iter, 3.5 CPI
 211
 212    // for ( i = 0; i < lda; i += 1 ){
 213    //    for ( j = coreid; j < lda; j += 2*ncores ) {
 214    //       for ( k = 0; k < lda; k += 1 ) {
 215    //          C[i + j * lda] += A[j * lda + k] * B[k*lda + i];
 216    //          C[i + (j+2) * lda] += A[(j+2) * lda + k] * B[k*lda + i];
 217
 218    //          //barrier(nc);
 219    //       }
 220    //       barrier(nc);
 221    //    }
 222    //    //barrier(nc);
 223    // }
 224
 225    //for ( i = coreid; i < lda; i += ncores ){
 226    //   for ( j = coreid; j < lda; j += ncores ) {
 227    //      for ( k = coreid; k < lda; k += ncores ) {
 228    //         C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 229    //      }
 230          //barrier(nc);
 231    //   }
 232    //}
 233 }
 234
 235 //--------------------------------------------------------------------------
 236 // Main
 237 //
 238 // all threads start executing thread_entry(). Use their "coreid" to
 239 // differentiate between threads (each thread is running on a separate core).
 240
 241 void thread_entry(int cid, int nc)
 242 {
 243    coreid = cid;
 244    ncores = nc;
 245
 246    // static allocates data in the binary, which is visible to both threads
 247    static data_t results_data[ARRAY_SIZE];
 248
 249
 250 //   // Execute the provided, naive matmul
 251 //   barrier(nc);
 252 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 253 //
 254 //
 255 //   // verify
 256 //   verifyMT(ARRAY_SIZE, results_data, verify_data);
 257 //
 258 //   // clear results from the first trial
 259 //   size_t i;
 260 //   if (coreid == 0)
 261 //      for (i=0; i < ARRAY_SIZE; i++)
 262 //         results_data[i] = 0;
 263 //   barrier(nc);
 264
 265
 266    // Execute your faster matmul
 267    barrier(nc);
 268    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 269
 270 #ifdef DEBUG
 271    printArrayMT("results:", ARRAY_SIZE, results_data);
 272    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 273 #endif
 274
 275    // verify
 276    verifyMT(ARRAY_SIZE, results_data, verify_data);
 277    barrier(nc);
 278
 279    exit(0);
 280 }
 281