mt/ap_vvadd/ap_vvadd.c

   1 //**************************************************************************
   2 // Vector-vector add benchmark
   3 //--------------------------------------------------------------------------
   4 // Author  : Andrew Waterman
   5 // TA      : Christopher Celio
   6 // Student :
   7 //
   8 // This benchmark adds two vectors and writes the results to a
   9 // third vector. The input data (and reference data) should be
  10 // generated using the vvadd_gendata.pl perl script and dumped
  11 // to a file named dataset.h
  12
  13 // to print out arrays, etc.
  14 //#define DEBUG
  15
  16 //--------------------------------------------------------------------------
  17 // Includes
  18
  19 #include <string.h>
  20 #include <stdlib.h>
  21 #include <stdio.h>
  22
  23
  24 //--------------------------------------------------------------------------
  25 // Input/Reference Data
  26
  27 typedef float data_t;
  28 #include "dataset.h"
  29
  30
  31 //--------------------------------------------------------------------------
  32 // Basic Utilities and Multi-thread Support
  33
  34 __thread unsigned long coreid;
  35 unsigned long ncores;
  36
  37 #include "util.h"
  38
  39 #define stringify_1(s) #s
  40 #define stringify(s) stringify_1(s)
  41 #define stats(code) do { \
  42     unsigned long _c = -rdcycle(), _i = -rdinstret(); \
  43     code; \
  44     _c += rdcycle(), _i += rdinstret(); \
  45     if (coreid == 0) \
  46       printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
  47              stringify(code), _c, _c/DATA_SIZE, 10*_c/DATA_SIZE%10, _c/_i, 10*_c/_i%10); \
  48   } while(0)
  49
  50
  51 //--------------------------------------------------------------------------
  52 // Helper functions
  53
  54 void printArray( char name[], int n, data_t arr[] )
  55 {
  56   int i;
  57   if (coreid != 0)
  58      return;
  59
  60   printf( " %10s :", name );
  61   for ( i = 0; i < n; i++ )
  62     printf( " %4ld ", (long) arr[i] );
  63   printf( "\n" );
  64 }
  65
  66 void __attribute__((noinline)) verify(size_t n, const data_t* test, const data_t* correct)
  67 {
  68    if (coreid != 0)
  69       return;
  70
  71    size_t i;
  72    for (i = 0; i < n; i++)
  73    {
  74       if (test[i] != correct[i])
  75       {
  76          printf("FAILED test[%d]= %4ld, correct[%d]= %4ld\n",
  77             i, (long) test[i], i, (long)correct[i]);
  78          exit(-1);
  79       }
  80    }
  81
  82    return;
  83 }
  84
  85 //--------------------------------------------------------------------------
  86 // vvadd function
  87
  88 //perform in-place vvadd
  89 void __attribute__((noinline)) vvadd(size_t n, data_t* __restrict__ x, const data_t* __restrict__ y)
  90 {
  91    size_t i;
  92
  93    // interleave accesses
  94    for (i = coreid; i < n; i+=ncores)
  95    {
  96       x[i] = x[i] + y[i];
  97    }
  98 }
  99
 100 void __attribute__((noinline)) vvadd_opt(size_t n, data_t* __restrict__ x, const data_t* __restrict__ y)
 101 {
 102    // ***************************** //
 103    // **** ADD YOUR CODE HERE ***** //
 104    // ***************************** //
 105
 106    size_t i, j;
 107
 108    size_t blocksize = n / ncores;
 109    size_t start = coreid * blocksize;
 110    size_t leftover = n % ncores;
 111
 112 //   int i, j;
 113
 114    for (i = start; i < (start + blocksize); i++) {
 115       x[i] = x[i] + y[i];
 116    }
 117
 118    for (j = (n - leftover) + coreid; j < n; j += ncores) {
 119       x[j] = x[j] + y[j];
 120    }
 121
 122 }
 123
 124 //--------------------------------------------------------------------------
 125 // Main
 126 //
 127 // all threads start executing thread_entry(). Use their "coreid" to
 128 // differentiate between threads (each thread is running on a separate core).
 129
 130 void thread_entry(int cid, int nc)
 131 {
 132    coreid = cid;
 133    ncores = nc;
 134
 135    // static allocates data in the binary, which is visible to both threads
 136    static data_t results_data[DATA_SIZE];
 137
 138    // because we're going to perform an in-place vvadd (and we're going to run
 139    // it a couple of times) let's copy the input data to a temporary results
 140    // array
 141
 142    size_t i;
 143    if (coreid == 0)
 144    {
 145       for (i = 0; i < DATA_SIZE; i++)
 146          results_data[i] = input1_data[i];
 147    }
 148
 149
 150    // Execute the provided, terrible vvadd
 151    barrier();
 152    stats(vvadd(DATA_SIZE, results_data, input2_data); barrier());
 153
 154
 155    // verify
 156    verify(DATA_SIZE, results_data, verify_data);
 157
 158    // reset results from the first trial
 159    if (coreid == 0)
 160    {
 161       for (i=0; i < DATA_SIZE; i++)
 162          results_data[i] = input1_data[i];
 163    }
 164    barrier();
 165
 166
 167    // Execute your faster vvadd
 168    barrier();
 169    stats(vvadd_opt(DATA_SIZE, results_data, input2_data); barrier());
 170
 171 #ifdef DEBUG
 172    printArray("results: ", DATA_SIZE, results_data);
 173    printArray("verify : ", DATA_SIZE, verify_data);
 174 #endif
 175
 176    // verify
 177    verify(DATA_SIZE, results_data, verify_data);
 178    barrier();
 179
 180    exit(0);
 181 }
 182