Threading Improvements, Double Precision Support

Upgrade Steps

Breaking Changes

• Breaking change in API as user specifies precision
• This leads to function handles accepting configured arguments in Numba functions, and updates to tests to reflect this
• Breaking change in configuration of the threading layer to use an OpenMP backend as default on Linux

New Features

• Double precision support

Bug Fixes

Improvements

• Better testing for linalg, kernel and surface modules
• Double precision support, can achieve > 10 digits of precision.

Other Changes

• N/A

Match Accuracy of ExaFMM and Add Gradients

Upgrade Steps

• Breaking change in API as target potentials now include gradients

Breaking Changes

• AdaptOctree version changed

New Features

• Gradient Calculations

Bug Fixes

• AdaptOctree Version 1.30, fixes a critical bug for non-uniform particle distributions.

Improvements

• exafmm-t and pyexafmm share the same accuracy
• Gradients are also calculated, making PyExaFMM useful in real applications.

Other Changes

• N/A

Fixing Accuracy Issues

Upgrade Steps

• Remove CUDA setup

Breaking Changes

• AdaptOctree version changed
• No more CUDA reliance

New Features

• N/A

Bug Fixes

• PR #80 details the various bugs caught that were ruining accuracy compared to theory

Improvements

• exafmm-t and pyexafmm share comparable accuracies.

Other Changes

• N/A

Speedy Operator Evaluation

Upgrade Steps

• Installation remains the same.
• Introduce new necessary environment variables required for speedy operators.

Breaking Changes

• Backends now customisable and configurable, creating breaking changes with old project structure

New Features

• Fully multithreaded P2M and M2L operator evaluation.
• Auto-vectorised ultra-fast kernel.
• Configurable compute backends (Only numba implemented for now)

Bug Fixes

• N/A

Improvements

• This release sees a 20X speedup in FMM performance from multithreading optimisations.
• A 1million particle benchmark problem (maximum of 100 particles per node) with O(5) expansions, is computed in 15 seconds c.p. with O(5minutes)

Other Changes

• N/A

Fully Adaptive FMM

Upgrade Steps

• Now built as a conda recipe, or installed from anaconda cloud directly

Breaking Changes

• New interface and configuration for Fmm class.
• New configuration and interface for operator precomputations
• Container classes Potential, Node, Density and related abstractions deleted
• Custom SVD deleted.

New Features

• Adaptive octree integrated with AdaptOctree
• Fully adaptive FMM
• M2L precomputations accelerated with CUDA and randomised SVD

Bug Fixes

• Working directory had to be source directory.

Improvements

• Radically easier experimental config
• Installation from the cloud
• A single HDF5 database used for all experimental data
• FMM Operator abstractions separated into modular functions, for easy optimisation
• FMM API solidified and separated from implementation of operators.

Other Changes

• N/A

Basic Non-Optimised KIFMM Implementation Added

• Add non-optimised 3D KIFMM operators and data structures
• Add Fmm core class, containing the main loop
• Add test code for operators and data structure
• Add test code for Fmm core class