Vectorisation in o-w collisions

[Jerboa-app]

We use marching squares/ tile based collision resolution, currently via branches.

• Can we do the same with a lookup table (function pointer valued?) for each tile code?

[Jerboa-app]

#include <vector>
#include <iostream>
#include <memory>
#include <cmath>
#include <iterator>

inline void doSomeStuff(double & y, double x, int i)
{	
	y = std::cos(x);
}

inline void doSomeOtherStuff(double & y, double x, int i)
{
	y = std::cos(x);
}

typedef void (*fptr) (double &, double, int);

void doWork(std::vector<double> & x, std::vector<double> & y)
{
        // vectorises
	for (int i = 0; i < 1024; i++)
	{
		//y[i] = std::sqrt(x[i]); nb sqrt requires -ffast-math
		doSomeStuff(y[i], x[i], i);
	}
}

void doWorkFptrs(std::vector<double> & x, std::vector<double> & y, std::vector<fptr> & fs)
{

        // missed: couldn't vectorize loop
        // missed: statement clobbers memory: _2 (_8, _3, i_22);
        //  missed: statement clobbers memory: _2 (_8, _3, i_22);
	for (int i = 0; i < 1024; i++)
	{
		fs[i](y[i], x[i], i);
	}
}


int main()
{
	std::vector<fptr> work;
	std::vector<double> x;
	std::vector<double> y(0.0, 1024);
	for (unsigned i = 0; i < 1024; i++)
	{
		if (std::cos(i) > 0.0)
		{
			x.push_back(3.14159*std::sqrt(i));
		}
		else
		{
			x.push_back(std::sqrt(i));
		}
		i % 2 == 0 ? work.push_back(&doSomeStuff) : work.push_back(&doSomeOtherStuff);
	}

	doWork(x, y);

	for (unsigned i = 0; i < 1024; i++)
	{
		std::cout << y[i] << "\n";
	}

	doWorkFptrs(x, y, work);

	for (unsigned i = 0; i < 1024; i++)
	{
		std::cout << y[i] << "\n";
	}

}

[Jerboa-app]

• function called must be inlineable for vectorisation
• so far function pointers appear to be an issue