- multithreading/thread pool
- SIMD data types and back-ends
- processing of nD arrays
And now I've re-routed some of the vspline code lux uses to work with zimt. The main difference is with 'peeling': The original vspline code processes underfilled SIMD vectors with scalar code, whereas zimt 'stuffs' the unused lanes and then processes a full vector. This should not affect performance much, but it makes all the scalar 'eval' versions in the rendering code unnecessary, which should reduce the code's complexity.
Because this is a far-reaching change, I have created a
new branch 'zimt' in the repo and I've uploaded a
debian package built from this branch. Apart from the back-end changes, the binary is pretty much like 1.1.6, but I've made a better debian package, now with dependencies. Comments welcome!