neurohacker

Zfp is lossy compression just FYI. Fpzip (an earlier invention by Lindstrom et al) has a lossless mode. They really push zfp more though as it has nice features like random access decoding and (except on my data) higher compression ratios.

If you want to downsample segmentation produced from convnet output using only numpy, I have this technique to contribute :)

https://towardsdatascience.com/countless-high-performance-2x...

To build on what you're saying: while true from the user perspective, out-of-order execution and instruction level parallelism means there are some types of parallelism happening at the single core level. However, these systems are designed to produce results to the end user as though no micro-level parallelism is occurring.

I mention this mainly because I've recently discovered the joy of SIMD intrinsics. While it's pretty difficult to gain anything from out-of-order execution (cough other than a spectre) it's possible to take advantage of SIMD through compiler autovectorization, intrinsics, or assembly coding. SIMD doesn't have the problem of race conditions though as the parallel operations are non-conflicting and the user view of the computation is synchronous. I imagine under the hood there are all kinds of tricks at play for complex instructions that involve different microcircuits that take different numbers of cycles.

EDIT: clarified last sentence

This rules, but until 3.8 is out I've been using http://semanchuk.com/philip/posix_ipc/ which unfortunately doesn't work on Windows.