mt/ab_matmul/ab_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    // I think I've got a way for this to not need the "shared" state to work nicely, so no MSI version
 120    int i, j, k, lda_over_2;
 121    lda_over_2 = lda/2;
 122
 123    if(coreid > 1)
 124       return;
 125    // left side of c
 126    if(coreid == 0)
 127    {
 128       // first half of topleft corner
 129       for(i = 0; i < lda_over_2; i++) {
 130        for(j = 0; j < lda_over_2; j++) {
 131         for(k = 0; k < lda_over_2; k++) {
 132          C[i*lda + j] += A[i*lda + k]*B[k*lda + j];
 133         }
 134        }
 135       }
 136       // second half of topleft corner
 137       for(i = 0; i < lda_over_2; i++) {
 138        for(j = 0; j < lda_over_2; j++) {
 139         for(k = lda_over_2; k < lda; k++) {
 140          C[i*lda + j] += A[i*lda + k]*B[k*lda + j];
 141         }
 142        }
 143       }
 144       // second half of bottomleft corner
 145       for(i = lda_over_2; i < lda; i++) {
 146        for(j = 0; j < lda_over_2; j++) {
 147         for(k = lda_over_2; k < lda; k++) {
 148          C[i*lda + j] += A[i*lda + k]*B[k*lda + j];
 149         }
 150        }
 151       }
 152       // first half of bottomleft corner
 153       for(i = lda_over_2; i < lda; i++) {
 154        for(j = 0; j < lda_over_2; j++) {
 155         for(k = 0; k < lda_over_2; k++) {
 156          C[i*lda + j] += A[i*lda + k]*B[k*lda + j];
 157         }
 158        }
 159       }
 160    }
 161    else  // coreid == 1
 162    {
 163       // first half of bottomright corner
 164       for(i = lda_over_2; i < lda; i++) {
 165        for(j = lda_over_2; j < lda; j++) {
 166         for(k = 0; k < lda_over_2; k++) {
 167          C[i*lda + j] += A[i*lda + k]*B[k*lda + j];
 168         }
 169        }
 170       }
 171       // second half of bottomright corner
 172       for(i = lda_over_2; i < lda; i++) {
 173        for(j = lda_over_2; j < lda; j++) {
 174         for(k = lda_over_2; k < lda; k++) {
 175          C[i*lda + j] += A[i*lda + k]*B[k*lda + j];
 176         }
 177        }
 178       }
 179       // second half of topright corner
 180       for(i = 0; i < lda_over_2; i++) {
 181        for(j = lda_over_2; j < lda; j++) {
 182         for(k = lda_over_2; k < lda; k++) {
 183          C[i*lda + j] += A[i*lda + k]*B[k*lda + j];
 184         }
 185        }
 186       }
 187       // first half of topright corner
 188       for(i = 0; i < lda_over_2; i++) {
 189        for(j = lda_over_2; j < lda; j++) {
 190         for(k = 0; k < lda_over_2; k++) {
 191          C[i*lda + j] += A[i*lda + k]*B[k*lda + j];
 192         }
 193        }
 194       }
 195    }
 196
 197    return;
 198 }
 199
 200 //--------------------------------------------------------------------------
 201 // Main
 202 //
 203 // all threads start executing thread_entry(). Use their "coreid" to
 204 // differentiate between threads (each thread is running on a separate core).
 205
 206 void thread_entry(int cid, int nc)
 207 {
 208    coreid = cid;
 209    ncores = nc;
 210
 211    // static allocates data in the binary, which is visible to both threads
 212    static data_t results_data[ARRAY_SIZE];
 213
 214
 215 //   // Execute the provided, naive matmul
 216 //   barrier(nc);
 217 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 218 //
 219 //
 220 //   // verify
 221 //   verifyMT(ARRAY_SIZE, results_data, verify_data);
 222 //
 223 //   // clear results from the first trial
 224 //   size_t i;
 225 //   if (coreid == 0)
 226 //      for (i=0; i < ARRAY_SIZE; i++)
 227 //         results_data[i] = 0;
 228 //   barrier(nc);
 229
 230
 231    // Execute your faster matmul
 232    barrier(nc);
 233    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 234
 235 #ifdef DEBUG
 236    printArrayMT("results:", ARRAY_SIZE, results_data);
 237    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 238 #endif
 239
 240    // verify
 241    verifyMT(ARRAY_SIZE, results_data, verify_data);
 242    barrier(nc);
 243
 244    exit(0);
 245 }
 246