I Created a Look_up table class but when I test it with sin() and cos() it is actually slower than stdlib math's sin. This surprises me by a lot.

inline double MATH_TABLES::GetSind(double& value)
	 return mp_Sind[(static_cast<int>(value/m_SinPrecisiond)];

mp_Sind is simply an array of doubles. This is the non-Lerp'd version and it is even slower than by an average of 1.5-2x than sin(x). I am quite surprised by this. Is there something in this code that stands out as a reason for it a table lookup being slower than a sin() function?

Test Code

const unsigned int NUMSINS = 100000;
double Sins[NUMSINS];
boost::timer Timer;
void TestSinChart()
	int counter = 0;
	double time_table = 0;
	double time_reference = 0;
	while( NUMSINS > counter++)
	time_table = Timer.elapsed();
	counter = 0;
	while(NUMSINS > counter++)
	time_reference = Timer.elapsed();
	std::cout << "Num Tests: " << counter-1 << "\n";
	std::cout << "Table Time: " << time_table*1000 << " ms \n";
	std::cout << "Reference Time: " << time_reference*1000 << " ms\n";
6 Years
Discussion Span
Last Post by mike_2000_17

It might have something to do with the way you are determining your index. I'm sure sin implementations vary widely across platforms and depending on the chip support underneath, but if I run a simple test where I do the following:

for( int i = 0; i < 10000; ++i)
   for (int j = 0; j < 100000; ++j)
      double d = get_value (j);

I get wild results just by changing how the index is generated.

return table[i];

I get a runtime of ~8 seconds. For

return table[static_cast<int>(i/1.0)];

I get a runtime of ~12 seconds. Seems to correspond to your 1.5x claim, though that is really more likely just a coincidence.


Its probably because of the casting and such. Unless your working in a 'tight' environment, generally, forget about lookup tables.


Well I'm looking into an embedded system project, that's why I was looking into lookup tables. I was going to write some a generic helper library while I wait on the chip because of Chinese New Years (Like a month from when I ordered it before they ship it).

However, I guess I'll have to do the test on the chip to determine whether it's because of code or hardware acceleration of the sin().


Converting a double to an integer is not a trivial operation because they are represented completely differently in binary. It is not like conversions between integer types where it's almost just a matter of changing the number of bits. Like L7Sqr has reported, I would expect the conversion from double to integer to be taking a lot of the computing time (we're only talking about a few clock-cycles here, of course). If you can get away with not using floating-point values at all (which is very common in embedded systems since many micro-controllers don't have floating-point modules on their CPU anyways, or very limited capabilities for floating-point operations), then it will be faster. If it is appropriate, consider perhaps using integers with a mapping like 1(integer) -> 0.001(floating-point) (or even a base 2 mapping) and do all your calculation with integers only and your transcendental functions with look-up tables.

And, I would expect that the code you have would be faster than the sin() on a micro-controller (if that sin() function exists).

Edited by mike_2000_17: n/a

This topic has been dead for over six months. Start a new discussion instead.
Have something to contribute to this discussion? Please be thoughtful, detailed and courteous, and be sure to adhere to our posting rules.