Benchmark results comparing `image from value` variants

The benchmarks compare the current image from value against a new variant using the new xz folding of the blitter.


Parent ↗

Summary

The current/old form is about 1.1 to 1.3 times slower than the new implementation, for a test image of 800x600x3 (0.5 MP). Extending the test image to 2 MP (1600x1200x3) shifts things up, if only a bit. Around 1.15. Going to 16 MP (4608x3456x3) it drops again to around 1.1.
Good enough to switch as many blitters as we can to this mode. Definitely good enough in light of the fact that such xz loops have a good chance of being simd/vectorizable. See google's highway for a possibility.
- Addendum (Dec 3, 2025). The code of the highway/simd experiment can be found in branch highway-simd-experiment. Also the Results
Summary: A 4-unrolled (super)scalar loop beats the SIMD-based loop, which has only 2 lanes in the vector unit for doubles. And unrolling the SIMD-based loop is counterproductive.
Details of the underlying differences
- With the new xz axis of the blitter the image bands are folded into the columns, allowing a single loop to run over them all, instead of using many separate z-loops each going over the not many bands of a column.

For the 16 MP case.