Test/Source/Tests/Test_qtmul.cpp

   1 //
   2 //  Test_qtmul.cpp
   3 //  BulletTest
   4 //
   5 //  Copyright (c) 2011 Apple Inc.
   6 //
   7
   8
   9
  10 #include "LinearMath/btScalar.h"
  11 #if defined (BT_USE_SSE_IN_API) || defined (BT_USE_NEON)
  12
  13
  14 #include "Test_qtmul.h"
  15 #include "vector.h"
  16 #include "Utils.h"
  17 #include "main.h"
  18 #include <math.h>
  19 #include <string.h>
  20
  21 #include <LinearMath/btQuaternion.h>
  22
  23 #define BT_OP(a, b)     ((a) *= (b))
  24 // reference code for testing purposes
  25 static inline btQuaternion& qtmul_ref(btQuaternion& q1, btQuaternion& q2);
  26
  27 static inline btQuaternion& qtmul_ref(btQuaternion& q1, btQuaternion& q2)
  28 {
  29     float x,y,z,w;
  30     x = q1.w() * q2.x() + q1.x() * q2.w() + q1.y() * q2.z() - q1.z() * q2.y(),
  31     y = q1.w() * q2.y() + q1.y() * q2.w() + q1.z() * q2.x() - q1.x() * q2.z(),
  32     z = q1.w() * q2.z() + q1.z() * q2.w() + q1.x() * q2.y() - q1.y() * q2.x(),
  33     w = q1.w() * q2.w() - q1.x() * q2.x() - q1.y() * q2.y() - q1.z() * q2.z();
  34
  35     q1.setValue(x, y, z, w);
  36         return q1;
  37 }
  38
  39 #define LOOPCOUNT 1024
  40 #define NUM_CYCLES 1000
  41
  42 int Test_qtmul(void)
  43 {
  44     btQuaternion q1, q2, q3;
  45
  46     float x, y, z, w, vNaN;
  47
  48     // Init the data
  49     x = RANDF_01;
  50     y = RANDF_01;
  51     z = RANDF_01;
  52     w = RANDF_01;
  53     vNaN = BT_NAN;     // w channel NaN
  54     q1.setValue(x,y,z,w);
  55
  56     x = RANDF_01;
  57     y = RANDF_01;
  58     z = RANDF_01;
  59     w = RANDF_01;
  60     q2.setValue(x,y,z,w);
  61
  62         q3 = q1;
  63
  64     btQuaternion correct_res, test_res;
  65
  66     {
  67                 float vNaN = BT_NAN;
  68                 correct_res.setValue(vNaN, vNaN, vNaN, vNaN);
  69                 test_res.setValue(vNaN, vNaN, vNaN, vNaN);
  70                 correct_res = qtmul_ref(q1, q2);
  71                 test_res = BT_OP(q3,q2);
  72
  73                 if( fabsf(correct_res.x() - test_res.x()) +
  74                         fabsf(correct_res.y() - test_res.y()) +
  75                         fabsf(correct_res.z() - test_res.z()) +
  76                         fabsf(correct_res.w() - test_res.w()) > FLT_EPSILON*10 )
  77                 {
  78                         vlog( "Error - qtmul result error! "
  79                                         "\ncorrect = (%10.4f, %10.4f, %10.4f, %10.4f) "
  80                                         "\ntested  = (%10.4f, %10.4f, %10.4f, %10.4f) \n",
  81                                         correct_res.x(), correct_res.y(),
  82                     correct_res.z(), correct_res.w(),
  83                                         test_res.x(), test_res.y(),
  84                     test_res.z(), test_res.w());
  85
  86                         return 1;
  87                 }
  88         }
  89
  90 #define DATA_SIZE LOOPCOUNT
  91
  92         btQuaternion qt_arr1[DATA_SIZE];
  93         btQuaternion qt_arr2[DATA_SIZE];
  94
  95     uint64_t scalarTime;
  96     uint64_t vectorTime;
  97     size_t j, k;
  98
  99         {
 100         uint64_t startTime, bestTime, currentTime;
 101
 102         bestTime = -1LL;
 103         scalarTime = 0;
 104         for (j = 0; j < NUM_CYCLES; j++)
 105                 {
 106                         for( k = 0; k < DATA_SIZE; k++ )
 107                         {
 108                 x = RANDF_01;
 109                 y = RANDF_01;
 110                 z = RANDF_01;
 111                                 w = RANDF_01;
 112                                 qt_arr1[k].setValue(x,y,z,w);
 113
 114                 x = RANDF_01;
 115                 y = RANDF_01;
 116                 z = RANDF_01;
 117                                 w = RANDF_01;
 118                                 qt_arr2[k].setValue(x,y,z,w);
 119                         }
 120
 121             startTime = ReadTicks();
 122             for( k = 0; k < LOOPCOUNT; k++ )
 123                         {
 124                     qt_arr1[k] = qtmul_ref(qt_arr1[k], qt_arr2[k]);
 125                         }
 126                         currentTime = ReadTicks() - startTime;
 127             scalarTime += currentTime;
 128             if( currentTime < bestTime )
 129                 bestTime = currentTime;
 130         }
 131         if( 0 == gReportAverageTimes )
 132             scalarTime = bestTime;
 133         else
 134             scalarTime /= NUM_CYCLES;
 135     }
 136
 137     {
 138         uint64_t startTime, bestTime, currentTime;
 139
 140         bestTime = -1LL;
 141         vectorTime = 0;
 142         for (j = 0; j < NUM_CYCLES; j++)
 143                 {
 144                         for( k = 0; k < DATA_SIZE; k++ )
 145                         {
 146                 x = RANDF_01;
 147                 y = RANDF_01;
 148                 z = RANDF_01;
 149                                 w = RANDF_01;
 150                                 qt_arr1[k].setValue(x,y,z,w);
 151
 152                 x = RANDF_01;
 153                 y = RANDF_01;
 154                 z = RANDF_01;
 155                                 w = RANDF_01;
 156                                 qt_arr2[k].setValue(x,y,z,w);
 157                         }
 158
 159             startTime = ReadTicks();
 160             for( k = 0; k < LOOPCOUNT; k++ )
 161                         {
 162                                 qt_arr1[k] = BT_OP(qt_arr1[k], qt_arr2[k]);
 163                         }
 164                         currentTime = ReadTicks() - startTime;
 165             vectorTime += currentTime;
 166             if( currentTime < bestTime )
 167                 bestTime = currentTime;
 168         }
 169         if( 0 == gReportAverageTimes )
 170             vectorTime = bestTime;
 171         else
 172             vectorTime /= NUM_CYCLES;
 173     }
 174
 175     vlog( "Timing:\n" );
 176     vlog( "     \t    scalar\t    vector\n" );
 177     vlog( "    \t%10.4f\t%10.4f\n", TicksToCycles( scalarTime ) / LOOPCOUNT,
 178                                                                         TicksToCycles( vectorTime ) / LOOPCOUNT );
 179
 180     return 0;
 181 }
 182
 183 #endif //BT_USE_SSE