benchmarks/vec-vvadd/vec_vvadd_main.c

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