* improve fallback for vec types

* fix flake8

* Support M2L with FFT on ToyContext

* flake8 fixes

* Fix src_rscale, tgt_rscale ordering

* More tests with FFT

* Use sumpy.toys in test_translations

* flake8 fixes

* Fix enqueue_marker for NVIDIA CUDA

* update comment

* check for same queue

* Fix bad merge

Andreas Klöckner authored 1 year ago and Andreas Klöckner committed 1 year ago

This avoids long-lived references to CL kernels held by loopy caches

Isuru Fernando authored 1 year ago and Andreas Klöckner committed 1 year ago

gives a 40% performance boost in CUDA

* Use pytential branch

* Refactor E2P

* try new loopy branch

* fix formatting

* disable domains check

* register only if not found

* Move kernel_scaling to the outer kernel

* Refactor P2E

* Use loopy main

* re-enable implemented domains check

* Rename some loopy kernel handling functions

---------

Co-authored-by: Andreas Kloeckner <inform@tiker.net>

* Move derivative taker to a separate file

* Add fold markers

* fix docs