benchmarks/vec-vvadd/vec_vvadd_main.c

   1 // See LICENSE for license details.
   2
   3 //**************************************************************************
   4 // Vector-Thread Vector-vector add benchmark
   5 //--------------------------------------------------------------------------
   6 //
   7 // This benchmark uses adds to vectors and writes the results to a third
   8 // vector. The input data (and reference data) should be generated using the
   9 // vvadd_gendata.pl perl script and dumped to a file named dataset.h.
  10
  11 // Choose which implementation you wish to test... but leave only one on!
  12 // (only the first one will be executed).
  13 //#define SCALAR_C
  14 //#define SCALAR_ASM
  15 #define VT_ASM
  16
  17 //--------------------------------------------------------------------------
  18 // Input/Reference Data
  19
  20 //#include "dataset_test.h"
  21 #include "dataset.h"
  22
  23 //--------------------------------------------------------------------------
  24 // Helper functions
  25
  26 int verify( int n, float test[], float correct[] )
  27 {
  28   int i;
  29   for ( i = 0; i < n; i++ ) {
  30 //    if ( test[i] != correct[i] ) {
  31     if ( test[i] > 1.02*correct[i]
  32        || test[i] < 0.98*correct[i]) {
  33 #if HOST_DEBUG
  34       printf("    test[%d] : %3.2f\n", i, test[i]);
  35       printf("    corr[%d] : %3.2f\n", i, correct[i]);
  36 #endif
  37       // tell us which index fails + 2
  38       // (so that if i==0,i==1 fails, we don't
  39       // think it was a 'not-finished yet' or pass)
  40 //      return i+10;
  41       return 2;
  42     }
  43   }
  44   return 1;
  45 }
  46
  47 void finishTest( int correct, long long num_cycles, long long num_retired )
  48 {
  49    int toHostValue = correct;
  50 #if HOST_DEBUG
  51   if ( toHostValue == 1 )
  52     printf( "*** PASSED ***\n" );
  53   else
  54     printf( "*** FAILED *** (tohost = %d)\n", toHostValue );
  55   exit(0);
  56 #else
  57   // we no longer run in -testrun mode, which means we can't use
  58   // the tohost register to communicate "test is done" and "test results"
  59   // so instead we will communicate through print* functions!
  60   if ( correct == 1 )
  61   {
  62     printstr( "*** PASSED *** (num_cycles = 0x" );
  63     printhex(num_cycles);
  64     printstr( ", num_inst_retired = 0x");
  65     printhex(num_retired);
  66     printstr( ")\n" );
  67   }
  68   else
  69   {
  70     printstr( "*** FAILED *** (num_cycles = 0x");
  71     printhex(num_cycles);
  72     printstr( ", num_inst_retired = 0x");
  73     printhex(num_retired);
  74     printstr( ")\n" );
  75   }
  76   exit();
  77 #endif
  78 }
  79
  80
  81 // deprecated - cr10/stats-enable register no longer exists
  82 void setStats( int enable )
  83 {
  84 #if ( !HOST_DEBUG && SET_STATS )
  85   asm( "mtpcr %0, cr10" : : "r" (enable) );
  86 #endif
  87 }
  88
  89 long long getCycles()
  90 {
  91    long long cycles = 1337;
  92 #if ( !HOST_DEBUG && SET_STATS )
  93   __asm__ __volatile__( "rdcycle %0" : "=r" (cycles) );
  94 #endif
  95   return cycles;
  96 }
  97
  98 long long getInstRetired()
  99 {
 100    long long inst_retired = 1338;
 101 #if ( !HOST_DEBUG && SET_STATS )
 102   __asm__ __volatile__( "rdinstret %0" : "=r" (inst_retired) );
 103 #endif
 104   return inst_retired;
 105 }
 106
 107 //--------------------------------------------------------------------------
 108 // vvadd function
 109
 110 // scalar C implementation
 111 void vvadd( int n, float a[], float b[], float c[] )
 112 {
 113   int i;
 114   for ( i = 0; i < n; i++ )
 115     c[i] = a[i] + b[i];
 116 }
 117
 118 // assembly implementations can be found in *_asm.S
 119
 120 //--------------------------------------------------------------------------
 121 // Main
 122
 123 int main( int argc, char* argv[] )
 124 {
 125   float results_data[DATA_SIZE];
 126   long long start_cycles = 0;
 127   long long stop_cycles = 0;
 128   long long num_cycles;
 129   long long start_retired = 0;
 130   long long stop_retired = 0;
 131   long long num_retired;
 132
 133   // Output the input array
 134
 135 #if HOST_DEBUG
 136   printArray( "input1", DATA_SIZE, input1_data );
 137   printArray( "input2", DATA_SIZE, input2_data );
 138   printArray( "verify", DATA_SIZE, verify_data );
 139 #endif
 140
 141   // --------------------------------------------------
 142   // If needed we preallocate everything in the caches
 143
 144 #if PREALLOCATE
 145
 146 #ifdef SCALAR_C
 147   vvadd( DATA_SIZE, input1_data, input2_data, results_data );
 148 #else
 149 #ifdef SCALAR_ASM
 150   scalar_vvadd_asm( DATA_SIZE, input1_data, input2_data, results_data );
 151 #else
 152 #ifdef VT_ASM
 153   vt_vvadd_asm( DATA_SIZE, input1_data, input2_data, results_data );
 154 #endif
 155 #endif
 156 #endif
 157
 158 #endif
 159
 160   // --------------------------------------------------
 161   // Do the vvadd
 162   start_cycles = getCycles();
 163   start_retired = getInstRetired();
 164
 165 #ifdef SCALAR_C
 166   vvadd( DATA_SIZE, input1_data, input2_data, results_data );
 167 #else
 168 #ifdef SCALAR_ASM
 169   #if HOST_DEBUG==0
 170   scalar_vvadd_asm( DATA_SIZE, input1_data, input2_data, results_data );
 171   #endif
 172 #else
 173 #ifdef VT_ASM
 174   #if HOST_DEBUG==0
 175   vt_vvadd_asm( DATA_SIZE, input1_data, input2_data, results_data );
 176   #endif
 177 #endif
 178 #endif
 179 #endif
 180
 181   stop_cycles  = getCycles();
 182   stop_retired = getInstRetired();
 183   num_cycles = stop_cycles - start_cycles;
 184   num_retired = stop_retired - start_retired;
 185
 186 //  printstr("stop_cycles: "); printhex(stop_cycles); printstr("\n");
 187 //  printstr("star_cycles: "); printhex(start_cycles); printstr("\n");
 188
 189   // --------------------------------------------------
 190   // Print out the results
 191
 192 #if HOST_DEBUG
 193   printArray( "results", DATA_SIZE, results_data );
 194 #endif
 195
 196   // --------------------------------------------------
 197   // Check the results
 198   int correct = verify( DATA_SIZE, results_data, verify_data );
 199   finishTest(correct, num_cycles, num_retired);
 200 }