|
|
A beautiful sunny Sunday in Seattle. Outside my window, ships go through the
locks, salmon swim up the ladder, a bagpipe band plays in the garden and I am
inside coding. Actually I love coding. And I have kept it down to two hours.
NEW_mandelx_pattern was identified all along as a hack. Perfectly adequate for
demonstrating the improved power computations, but for timing purposes
representing a worst-case promise rather than any basis for quantitative
comparison.
NEWER_mandelx_pattern is a speed-optimized version. Ready for quantitative
comparisons. My main avenue of attack was to unroll my awful interior loops.
Short, badly branch-predicted. Ideal candidates. The measured result is
posted among the images.
Also I have extended NEWER_mandelx_pattern to treat exponents up to 255.
The old mandel_pattern. A fine code. I would have said speed optimal, in fact.
I was hoping to come within 10% of its speed. But the measurements are
showing NEWER_mandelx_pattern to be 6.8% FASTER, for some unexpected reason. I
will not be satisfied until I have tracked down why this is.
Will post NEWER_mandelx_pattern after I understand it vs. mandel_pattern and
prune its code way down.
Algo
Post a reply to this message
|
|