benchmarks/mt-vvadd/mt-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 double data_t;
  28 #include "dataset.h"
  29
  30
  31 //--------------------------------------------------------------------------
  32 // Basic Utilities and Multi-thread Support
  33
  34 __thread unsigned long coreid;
  35
  36 #include "util.h"
  37
  38 #define stringify_1(s) #s
  39 #define stringify(s) stringify_1(s)
  40 #define stats(code) do { \
  41     unsigned long _c = -rdcycle(), _i = -rdinstret(); \
  42     code; \
  43     _c += rdcycle(), _i += rdinstret(); \
  44     if (coreid == 0) \
  45       printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
  46              stringify(code), _c, _c/DATA_SIZE, 10*_c/DATA_SIZE%10, _c/_i, 10*_c/_i%10); \
  47   } while(0)
  48
  49 //--------------------------------------------------------------------------
  50 // vvadd function
  51
  52 //perform in-place vvadd
  53 void __attribute__((noinline)) vvadd(int ncores, size_t n, data_t* __restrict__ x, const data_t* __restrict__ y)
  54 {
  55    size_t i;
  56
  57    // interleave accesses
  58    for (i = coreid; i < n; i+=ncores)
  59    {
  60       x[i] = x[i] + y[i];
  61    }
  62 }
  63
  64 void __attribute__((noinline)) vvadd_opt(size_t n, data_t* __restrict__ x, const data_t* __restrict__ y)
  65 {
  66    // ***************************** //
  67    // **** ADD YOUR CODE HERE ***** //
  68    // ***************************** //
  69 }
  70
  71 //--------------------------------------------------------------------------
  72 // Main
  73 //
  74 // all threads start executing thread_entry(). Use their "coreid" to
  75 // differentiate between threads (each thread is running on a separate core).
  76
  77 void thread_entry(int cid, int nc)
  78 {
  79    coreid = cid;
  80
  81    // static allocates data in the binary, which is visible to both threads
  82    static data_t results_data[DATA_SIZE];
  83
  84    // because we're going to perform an in-place vvadd (and we're going to run
  85    // it a couple of times) let's copy the input data to a temporary results
  86    // array
  87
  88    size_t i;
  89    if (coreid == 0)
  90    {
  91       for (i = 0; i < DATA_SIZE; i++)
  92          results_data[i] = input1_data[i];
  93    }
  94
  95
  96    // Execute the provided, terrible vvadd
  97    barrier(nc);
  98    stats(vvadd(nc, DATA_SIZE, results_data, input2_data); barrier(nc));
  99
 100
 101    // verify
 102    int res = verifyDouble(DATA_SIZE, results_data, verify_data);
 103    if (res)
 104       exit(res);
 105
 106 #if 0
 107    // reset results from the first trial
 108    if (coreid == 0)
 109    {
 110       for (i=0; i < DATA_SIZE; i++)
 111          results_data[i] = input1_data[i];
 112    }
 113    barrier(nc);
 114
 115    // Execute your faster vvadd
 116    barrier(nc);
 117    stats(vvadd_opt(DATA_SIZE, results_data, input2_data); barrier(nc));
 118
 119 #ifdef DEBUG
 120    printDoubleArray("results: ", DATA_SIZE, results_data);
 121    printDoubleArray("verify : ", DATA_SIZE, verify_data);
 122 #endif
 123
 124    // verify
 125    res = verifyDouble(DATA_SIZE, results_data, verify_data);
 126    if (res)
 127       exit(res);
 128    barrier(nc);
 129 #endif
 130
 131    exit(0);
 132 }