Soooo. It turns out that the bunch of smallish (~4-5% of runtime)
loads associated with Len(), Unit(), Rotated(), and so on... Were
actually more like 15% or more of computational effort. I first
figured this out by creating:
func (u Vec) Normal(v Vec) Vec
which gives you a vector normal to u->v. That consumed a lot
of CPU time, and was followed by .Unit().Scaled(imd.thickness / 2),
which consumed a bit more CPU time.
After some poking, and in the interests of avoiding UI cruft,
the final selection is
func (u Vec) Normal() Vec
This returns the vector rotated 90 degrees, which turns out to
be the most common problem.
We never actually need the "normal" value; it's an extra calculation
we didn't need, because ijNormal is the same value early on. It's
totally possible that we could further simplify this; there's a lot
of time going into the normal computations.
updateData()'s loops checking gt.Len() turns out to have been costing
significant computation, not least because each call then in turn
called gt.vs.Stride().
The computation including a call to Stride() can't be optimized away
safely because the compiler can't tell that Stride() is effectively
constant, but we know it won't change so we can make a slice pointing
at that part of the array.
CPU time for updateData goes from 26.35% to 18.65% in my test case.
A slice of points means copying every point into the slice, then
copying every point's data from the slice to TrianglesData. An
array of indicies lets the compiler make better choices.
For polyline, don't compute each normal twice; when we're going through a line,
the "next" normal for segment N is always the "previous" normal for segment
N+1, and we can compute fewer of them.
For internal operations (anything using getAndClearPoints), there's a
pretty good chance that the operation will repeatedly invoke something
like fillPolygon(), meaning that it needs to push "a few" points
and then invoke something that uses those points.
So, we add a slice for containing spare slices of points, and on the
way out of each such function, shove the current imd.points (as used
inside that function) onto a stack, and set imd.points to [0:0] of
the thing it was called with.
Performance goes from 11-13fps to 17-18fps on my test case.
It turns out that affine matrices are much simpler than the 3x3 matrices
they imply, and we can use this to dramatically streamline some code.
For a test program, this was about a 50% gain in frame rate just from
the cost of the applyMatrixAndMask calls in imdraw, which were calling
matrix.Project() many times. Simplifying matrix.Project, alone, got a
nearly 50% frame rate boost!
Also modify pixelgl's SetMatrix to copy the six values of a 3x2
Affine into the corresponding locations of a 3x3 matrix.
Removing the call to Alpha(1) and replacing it with an inline definition
produces measurable improvements. Replacing each instance of ZV with
Vec{} further improves things. We keep an inline RGBA because there
are circumstances (mostly when using pictures) where we don't want to
have to set colors to get default behavior.
For a fairly triangle-heavy thing, this reduces time spent in SetLen
from something over 10% of execution time to around 2.5% of execution
time.
Seebs
faiface
Michal Štrba