commit 813d3ae10cbf7db994a44c05eaa8d61b1cc6abf8
author Stella Laurenzo <stellaraccident@gmail.com> Wed Oct 06 15:39:03 2021 -0700
committer GitHub <noreply@github.com> Wed Oct 06 15:39:03 2021 -0700
tree 13e11c1c1903274a9fa40d02d7cd8a1021b0c873
parent e63f67c3742b1b95cd87c9f12bd36fbbfe3309d3

[pydm] Defines the structure for the full numeric hierarchy. (#7274)

* [pydm] Defines the structure for the full numeric hierarchy.

* Full support modeled for signed/unsigned 8/16/32/64 bit integers, fp16/bf16/fp32/fp64, complex64/complex128, bool, weak integer, weak real, arbitrary precision integer.
* Actual support for everything is more limited. Using a frontend pass to squash all weak types to i32/f32 for now (type inference/analysis needs to come into play here before making such decisions).
* Numeric promotion is in-flux at the moment, but shooting for a combination of Numba/Cython/JAX reasoning about this. Key is that weak integer/real types exist and bind to the hierarchy in different ways. See: https://jax.readthedocs.io/en/latest/type_promotion.html
* This makes the generic runtime support a lot more complicated and required quite a few more lowerings and canonicalizations to achieve (i.e. the runtime library decodes the bit patterns in the type code to make numeric type decisions).
* The generated code is still a joke and not something we would ever use, but it does run: https://gist.github.com/stellaraccident/e9f41a09a3834465d7576312fc63c278
* Still holding off on any real optimizations beyond canonicalizations since generality is helpful at this stage. Most of what is there should melt away with some simple variable load/store analysis.