[PATCH] Optimize TestRayAASquare in Geometry.cpp

[Stan]

There is a todo line 244 of Geometry.cpp that says :

// Exactly like TestRaySquare with u=(1,0), v=(0,1)
// Assume the compiler is clever enough to inline and simplify all this
// (TODO: stop assuming that)

[Stan]

Fix the Todo.

[Stan]

Documented Version of Above Patch

[Stan]

Replace some scalar computations by the value to save some computations, add doxygen documentation, and rename some variables to make the code more readable.

[elexis]

Yeah, that's the right direction in my understanding.

• It would be better though to not rename variables, as they have the same name in the other code as well and the patch will be easier to review.

• Remove the au, av, bu, bv variables, as you can just reference u.X etc..

• The goal is to completely remove the constants u, v. So the u.Multiply(hw) things also need to be simplified / consolidated.

Those vector functions are defined in FixedVector2D.h and reflect the usual vector operations known from linear algebra.

[Stan]

Assert the previous comments, though we cannot remove u and v for now.

[Stan]

Here are the maths behind the changes as you can see there is not much to be done for s1,s3 as you cannot simply inver halfsize.X in one line same for halfsize.Y we could add an MultiplyCoordinate in CFixedVector2D function which would basically return (X*n,Y*m) that would fix it.

S0
(1;0) * X + (0;1) * Y - (B;C)
= (X;Y) - (B;C)
= (X-B;Y-C)
= (Halfsize - a)

S1
(1;0) * X - (0;1) * Y - (B;C)
= (X; - Y) - (B;C)
= (X-B;-Y-C)

S2
- (1;0) * X - (0;1) * Y - (B;C)
= -(X;Y) - (B;C)
= (-X-B;-Y-C)
= (-Halfsize - a)

S3
- (1;0) * X + (0;1) * Y - (B;C)
= (-X;Y) - (B;C)
= (-X-B;Y-C)

[Stan]

A proposal to simplify further the code.

[elexis]

An alpha 19 replay and it's profiling data showing the performance details. The match was a 40min 3v3 with 200 pop capacity and too many trees.

[elexis]

In 17500:

Pathfinder optimization, fixes #3541.
Simplify the math of Geometry::TestRayAASquare.

[elexis]

The patch removes two constants (0 and 1) and useless compuation (for example 1*x + 0*y).

Notice the last two peaks show 10% performance improvement. Otherwise not too much gain.