If you use a straight forward loop to manipulate a CNiVector you end up using the overloaded [] operator - which peroforms bounds checking on every access. This is fine & dandy for VB applications, but these are suppose to be C++ tools! I first tried using the non-aggregrate functions ( overloaded +=) but they are not optimimally implemented. /Very/ far from it - actually they are _worse than using your own loop with the overloaded []! My program spends considerable time adding these vectors together, and it was taking an absurdly long time to do it with the provided methods.
If you're adding or otherwise manipulating CNiVectors, you
can increase the performance of the calculations SIXTEEN FOLD by using code similar to the following:
The code replaced was:
[source]
//272000 CPU Ticks
m_vDest += m_vSource;
[/source]
[source]
//17000 CPU Ticks
double* pdDest = m_vDest;
const double* pdStop = &m_vDest[n-1]; pdStop++; //don't ask I might explode - trust me that it's correct.
const double* pdSource = m_vSource[j];
for(; pdDest
{
*pdDest += *pdSource;
pdSource++;
}
[/source]
Both m_vDest and m_vSource are CNiReal64Vectors. I would expect similar performance increases for all the other vector types.
Ticks are from a PIII 700MHz (though the speed shouldn't matter, but the CPU type will - retrieved using RDTSC)
