Vectorization revamp

mentioned in issue #4 (closed)

mentioned in issue #20 (closed)

mentioned in issue #1 (closed)

mentioned in issue #26 (closed)

added refactor label

vectorize function should receive the config. default_float_type should not be needed

Revamping the vectorization is not an easy feat. There are multiple tasks ahead:

Abstraction of Vectorizer The vectorization should transform the AST to a vectorized one, ignorant of the target architecture. Translating the vectorized AST to a specific architecture is the step afterward. The printer should be simple!
pymbolic sympy is notoriously a double-edged sword. On one hand, we love its capabilities. On the other, we hate, when it does stuff without being asked. This comes more often than we would think. This issue can be handled aport from vectorization revamp, but not vice versa. Our strategy is to replace sympy with pymbolic at a certain point in the generation process, possibly right after most of the simplifications, or right after the type system. pymbolic is basically a braindead sympy. It doesn't do anything without being asked. With pymbolic AST transformation on the expression level should be way safer.
Use Marco's sb-simd Marco spent a lot of time creating an abstract layer for vectorization for his Petalisp. We should use his work with cl4py. Ideally, we would do once a transition from lisp to pyton/json. This would result in having the abstraction and we would have all the information nicely represented for various architectures (SSE, AVX, ...)
Pretty-printer The backend, the pretty-printer, should not have any logic! This should be done previously. See pow, mul, etc.

mentioned in merge request !344 (merged)

mentioned in merge request !359 (closed)

mentioned in issue #73

marked this issue as related to #73

changed milestone to %Release 2.0

This project is moving forward at #104, so I'm closing this issue.

closed

Child items