v0.0.7

0.0.7 (2026-01-20)

Features

• add runtime logging infrastructure for GPU kernels and memory transfers (#289) (431425d)
• add unified save_variables and summarise_variables parameters to solver interface (#342) (c7d7531)
• cellml-generated systems now cached (#510) (92b21e0)
• enable driver interpolator profiling in all_in_one.py (#419) (1574ec9)
• File-based caching implemented (#491) (1bfe68b)
• MultipleInstanceCUDAFactory subclass (and matching config) now handle cases like newton_atol and krylov_atol when instantiating multiple of the same base class (d674bcd)
• scaled norm function now available as a CUDAFactory for repeated use (d674bcd)
• Scaled tolerance in Newton-Krylov solver (#473) (d674bcd)
• Solve functions now save the final value on loop exit when no timing parameters are given. (d9fb64b)
• Time-domain save settings now decoupled from summary metric settings (d9fb64b)

Bug Fixes

• build_grid now takes None parameters/initial conditions in solver.py (be4bab9)
• chunking fails when VRAM is limited due to stride incompatibility (#487) (b8a486e), closes #438
• codegen hashing now session-independent (and so working... this time) (59bb488)
• correct false circular dependency error in topological_sort (#422) (97c13be)
• default neumann preconditioner order set to 2 (2434358)
• DIRK codegen pipeline now decoupled from rosenbrock cache planning. (80171b5)
• dummy-kernel based compile time logging removed (it doubled compile time) (431425d)
• Internal code generation variables prefixed to avoid name clashes (#466) (90e8ca3), closes #373
• load_cellml_model surfaced to toplevel import (54b05e0)
• loop now exits on irrecoverable-error status codes (ba800a0)
• map CellML time variable to standard 't' symbol (#425) (261c109)
• Newton-krylov solver no longer propagates krylov non-convergence or max_backtrack errors if it recovers (1411135)
• Parsed system definition now hashed properly so generated code is properly cached (655e54a)
• patch event.query() bug in numba by swapping handle (d43ae8f)
• Refactor BatchGridBuilder -> BatchInputHandler (#437) (ffb8478)
• repeated CSE calls no longer raise warning about an already-used CSE symbol. (16785f3)
• Rosenbrock step config's hash now deterministic for caching (105befb)
• Runs with impossible time settings now raise sensible errors (#465) (d46a7b2), closes #440
• skip codegen on cache hit in get_solver_helper (#512) (cfc6b0b)
• state-aware derivative detection to avoid misinterpreting d-prefixed auxiliaries (#468) (fee6d70)
• Stride incompatibility fixed when array is sliced to fit in VRAM (b8a486e)
• timelogger printouts now include cache hit/miss messaging (#497) (6323746)

Performance Improvements

• Convert whole-module imports to explicit imports in CUDAFactory files (#443) (9cebe45)

Miscellaneous Chores

• release 0.0.7 (40a0d75)

v0.0.6

0.0.6 (2025-12-27)

Features

• raw output type added to output device array copies with no processing (22cef9f)
• add time logging to cellml import (#257) (6a220f8)
• Additional summary output metrics added (#212) (daccbae)
• Buffer indexing, sizing, and locating now consolidated into a BufferSettings object (480df1a)
• Buffer memory locations on GPU now user-selectable (between local and shared) (480df1a)
• build_grid() surfaced to user API (#338) (45cfa90)
• CellML to Cubie adapter layer added (#221) (b8f448e)
• CUDAFactory.update() now updates nested dicts and attrs classes (480df1a)
• device buffers are now togglable (local, shared, local persistent) and managed centrally by buffer_registry (39107a9)
• Global stores now have cache write-through hints, closes #291 (480df1a)
• load_from_cellml updated to parse complicated models (#238) (50341f6)
• newton and linear solvers now fully-functional CUDAFactory subclasses (39107a9)
• optional configuration parameters are no longer explicit in inits, they are filtered and collected in kwargs (39107a9)
• py39 compatibility removed due to a Numba update. (480df1a)
• Solver API now skips extra memory and grid-building work when possible (#324) (5f225a0)
• summary metrics combined (eg. extrema, [mean, std, rms]) to reduce buffer space (daccbae)
• summary metrics now respect numerical precision (daccbae)
• Sympy inputs from CellML now go Sympy->Sympy instead of through strings (#259) (23e201d)
• time logging added to parsing, codegen, and CUDA compilation (#256) (5d8f75b)
• Trajectories with errors (nonzero status codes) now return NaNs in solveresult (#333) (6068ebc)
• update() methods now unpack settings dicts provided to them (#332) (f016e4d)
• Warp-friendly FSAL caching implented, redundant accumulation removed (#211) (96a9dd0)

Bug Fixes

• _ensure_context() method added to avoid segfaults in CI (#265) (60acecc)
• shift value in standard deviation calcs now updates after each save. (073d406)
• Adaptive step controllers now sum errors correctly, closes #302 (480df1a)
• add set_stride_order method for access from solver (e4b80d3)
• All algorithms now exit when the next save time is > end time (480df1a)
• buffer aliasing logic now child-location-agnostic (377ad10)
• Compile-timing kernels now use device arrays (16bf1d6)
• consolidate timing parameter sets for fewer fixture builds (#350) (4698e9e)
• contiguous arrays now marked as such for the compiler to do its grim work (#276) (fe97291)
• Controller-algorithm compatibility enforced (4ab0230)
• correct FIRK algorithm implementation (#408) (39107a9)
• correct fsal warp-vote implementation (60448f1)
• Counters array added to ERK signature (99db833)
• CUDA-breaking simsafe 'local' module adapter removed (d205979)
• CUDAFactories now only precompile if timelogging is on (836da2a)
• dead code and duplications pruned from codegenned device functions (#266) (de0dda8)
• default precision types have been removed from all but entry points (39107a9)
• fast path through batchgridbuilder for arrays provided verbatim added (080ab66)
• FIRK no longer does an extra f(x) calculation after performing its nonlinear solve (39107a9)
• Fixed steppers now use a slower but more sensible save time incrementing technique (480df1a)
• FSAL warp test now doesn't break everything (2e68692)
• gustafsson controller prev error ratio flipped, various test edits for GPU runs (3e04438)
• host arrays now pinned to facilitate asynchronous transfers (480df1a)
• infinite loops in adaptive steppers under dummy compile now finite (b0547fd)
• lineinfo toggle for CUDA compilation now conditional on CUDASIM status (#280) (6d5bf57)
• Loop iterators now 32-bit (480df1a)
• loops now accumulate time in f64, no longer get stuck when dt < 1e-7 * time (#281) (480df1a), closes #272
• make shape validator np.int compatible in BaseArrayManager.py (e77d800)
• make stage_increment buffer persistent in rosenbrock (#411) (4b0b6bc)
• Missing iteration_counters type added to device signature in BatchSolverKernel (6eefd67)
• move stream sync function to after chunked queue (e4b80d3)
• n_saves calculation now includes start time, closes #282 (480df1a)
• numeric literals now wrapped with precision() or int32 in CUDA code generation (#258) ([21850f2](21850f293b6743e2fe677b4fc...

v0.0.5

0.0.5 (2025-11-04)

Features

• "Instrumented" device steps added for diagnostics (472f960)
• additional ERK, DIRK, Rosenbrock tableaus added (ed866d7)
• Algorithms now set their own step control defaults (#139) (2a0efed)
• Array managers now support heterogeneous arrays within the same container (0905422)
• Compile-settings updates now won't force a rebuild if the value hasn't changed. (5a9e281)
• DIRK and DIRKTableaus added (472f960)
• Explicit and Diagonally-Implicit Runge Kutta algorithms added (#151) (472f960)
• Explicit RK and Tableau added, closes #83 (472f960)
• Fully-Implicit Runge-Kutta (FIRK) Methods implemented (#162) (ed866d7)
• Generic Butcher Tableau implemented (472f960)
• Generic Rosenbrock-W methods added (#148) (6abfed2)
• minimal FSAL caching added to DIRK, ERK, Rosenbrock (472f960)
• N-stage flattened linear operators, preconditioners, nonlinear residual codegen added (ed866d7)
• Parser now processes indexed arrays as variables (#152) (7fe800b)
• rodasnp methods, dop853, tsit5, vern7 (87b53bc)
• Rosenbrock methods now 100% more rosenbrock (#157) (35641e6)
• status codes now aggregated by batchSolverKernel (0905422)
• Steps and step controllers now have a unified argument-filtering factory (2a0efed)
• Tableau libraries and tableau resolvers/getters added (472f960)
• There are now auxiliary-cached jacobian functions for reusing some computational work. (6abfed2)
• Third and fourth order SDIRK tableaus added (91ee1e9)
• Time-derivative helpers added for symbolic functions and interpolated arrays (35641e6)
• Very rough caching of jvp nodes implemented for rosenbrock solvers. (7fe800b)
• working arrays and quantities in algorithms and solvers now draw from a memory "pool", allowing easier reuse (b0f5b6f)

Bug Fixes

• add error to sdirk 4, correct controllers for loop tests (47a6cb1)
• Added (non-CI) testing for DIRK loops added. (7fe800b)
• batchgridbuilder class now has static wrappers for batchgridbuilder helper functions. (74c08fa)
• BatchsolverKernel's compile settings now features the compile-critical settings it always should have had. (f383157)
• Buffer footprints reduced by aliasing vectors with disjoint lifetimes (472f960)
• correct off-by-one error and datatype discrepancy in cpu driver evaluator (472f960)
• DIRK now accumulates rhs's and scales after stages, reducing round-off (7fe800b)
• faulty implementation of ode23s removed, dop853 tableau amended (d5a784c)
• matplotlib now spelled correctly in pyproject.toml (3773ff8)
• Meaty loop tests confined to test_ode_loop.py (7fe5677)
• numerous numerical errors amended after instrumenting steps (472f960)
• precision type hints now correct; no more pesky yellow lines. (2a0efed)
• Rosenbrock buffer footprint reduced by 2n (ed866d7)
• settings passing from solver now less cramped and hopefully more robust (2a0efed)
• Step controllers no longer mutate error vector (7fe800b)

Documentation

• Add autodocs subpages for manual project structure docs (c2c0d7c)
• add copilot instructions (#161) (2bfe6f2)
• added "buffer map" comments to generic algorithms to demystify aliasing (472f960)
• autodocs param lists now format one line per param (cb3d74b)
• Batchsolving module source files now better documented in numpydocs format (97a0ab8)
• de-computer some language in api reference, rejig indexes (b558994)
• increase index depth, force one-param-per-line printing (b822f4c)
• more docs organisation (5631989)
• more docs organisation (cd98043)
• refactor api structure; remove autosummary, implement manual docs (881ce4a)
• Refs in "getting started" are now actually refs not highlighted garbage. (6abfed2)
• top-level batchsolving docs added (6cfddfa)

Miscellaneous Chores

• release 0.0.5 (93a7255)

v0.0.4

0.0.4 (2025-10-05)

Features

• Adaptive step-size controllers added : i (traditional), pi, pid, gustafsson acceleration (1d903f2)
• Adaptive time-step controllers now have a programmable dead-band. (b0fafd9)
• AGENTS.md extended and partially updated to summarise entire project for ai agents (1d903f2)
• arbitrary drivers can now be looped or clamped to zero (smoothly). (b0fafd9)
• Backwards Euler implicit fixed-step method added (with and without predictor-corrector mechanism), closes #114. (1d903f2)
• Codegen for residual functions, jvps, and various solver helper functions created (1d903f2)
• Crank-Nicolson trapezoidal adaptive-step algorithm implemented. (1d903f2)
• cuda simulation patches consolidated for cuda-free environment tests. (1d903f2)
• Forcing (driver) terms now adaptive-step friendly (#132) (b0fafd9)
• matrix free solvers added (1dffd94)
• Nonlinear Newton-Krylov iterative solver with preconditiong added for implicit methods, closes #101, #102, #111 (1d903f2)
• plotting added to driver interpolator - keep an eye on what the machine is doing. (b0fafd9)
• shared memory padding now closer to optimal (nothing to be done about 64-bit values), closes #86. (1d903f2)

Bug Fixes

• Array "chunking" logic now respects "unchunkable" arrays in allocation (53141df)
• CPU test step controllers now raise dt_too_small errors (19af8ff)
• Crank-Nicolson and adaptive controllers now 50% more idiot-error free. (b0fafd9)
• CUDAFactory now updates underscored config variables as intended (1d903f2)
• Many edits to precision settings and flow throughout system (making it work) (1d903f2)
• Observables calculation now occurs in sync with state for adaptive-step loops (#130) (19af8ff)
• pypi version tag trigger now using correct syntax (3b3caa1)
• some sign confusion and bogus gains corrected in step controllers (19af8ff)

Documentation

• Batch arrays re-docstringed in keeping with rest of library. (b0fafd9)
• fix circular import for docs building (64d8933)
• manual docs added for integrators, memory modules. submodules of systems, outputhandling documented. (1d903f2)
• step controller comparison added to docs/examples (19af8ff)
• top-level summaries of odesystems and outputhandling (a9cb4c0)
• update docstrings in outputhandling and odesystems root directorys (87bb133)

Miscellaneous Chores

• release 0.04 (ed5045a)

v0.0.3

0.0.3 (2025-09-04)

Features

• Parser accepts and translates sympy and user-provided functions (#108) (25f3c5d)
• Symbolic input parsing added (14577c1)
• symbolic interface and analytical jacobian generation added (14577c1)

Bug Fixes

• BaseArrayManager.py once again contains all of it's methods, after half of the class was sausage-fingered clean off. (96ba1cb)
• buggy regex removed from pyproject (bc5606a)
• ignore generated python files (14577c1)
• implement four-byte padding to reduce shared memory conflicts (944066e)
• jacobian product codegen now does a simple dead-code removal sweep at the expression level (5a46559)
• metric function compilation now deferred until fn requested. (fc42de7)
• sympy piecewise printing patched in subclass (ca20f84)
• SystemValues now interpreting sympy symbols correctly (0549bea)
• SystemValues, BaseODE now have comprehensible repros (0549bea)

Documentation

• batchsolving module now has all docstrings in numpydocs-friendly format (8e6e8cf)
• conf.py path reverted for Sphinx build (5383de6)
• docs updated and thinned to match structure (9df6978)
• first-pass narrative docs added (14577c1)
• get sphinx-build working again and ReadTheDocs themed (6df9a82)
• insert google verification tag, cross-link repo and docs (0d61cd2)
• integrators section docstrings brought in line with numpydocs format (8b22790)
• memory section docstrings brought into numpy format (d495551)
• output_functions section docstrings brought into numpy format (4e7b2c0)
• pypi version badge added (8b22790)
• readme now has code coverage badge (8e6e8cf)
• systemmodels section docstrings brought into numpy format (9df6978)

Miscellaneous Chores

• re-release 0.0.3 (359abe3)
• release 0.0.3 (684ef91)

v0.0.2

0.0.2 (2025-08-18)

Features

• BaseArrayManager class added to unify approach to allocating/deallocating device arrays through the memory manager (416e363)
• BatchConfigurator now accepts extra user input types to match usage (a345679)
• BatchInputArrays and BatchOutputArrays now subclass BaseArrayManager (808695e)
• Memory Manager extended to queue and process requests from multiple objects (67e66bf)
• MemoryManager implemented (3387142)
• Solver interface now set up for real people to use (I think) (2b05c1e)
• UserArrays now handles delivery from device arrays to end user for inspection (9d93a7a)

Bug Fixes

• BatchSolverKernel now uses blocksize to calculate dynamic shared memory correctly, and reduces blocksize if it's > limit (5d8263f)
• bug in previosu commit: BatchSolverKernel now uses blocksize to calculate dynamic shared memory correctly, and reduces blocksize if it's > limit (afda7cb)
• fix circular imports introduced in 636b5e3 (2865160)
• force odd shared memory size per run to minimise clashes. (fc12c76)
• forward declarations no longer cause circular imports (e3841ff), closes #73
• Newly-initialised memory manager no longer breaks in CUDA sim (db99862)
• output config flags now treated as derived quantities instead of attributes (9d93a7a)
• pyproject.toml now points at correct license and readme files for building. (f7bd7f0)
• pyproject.toml now points at correct license file (but for real this time) (93591ed)
• SolverKernel now solves and summarises accurately. (879da10)
• SolverKernel tests made CUDA-simulator-friendly (ad702bb)
• UserArrays now SolveResult, and works with array managers to produce a sensible output (994cd7c)
• UserArrays.as_numpy now returns copies rather than mappedarrays (01af17d)

Documentation

• Docs don't mention CuMC anymore, and we shall never speak of it again (fff2e00)
• Docs updated to reflect cubie refactor (205e748)
• Properties which expose lower-level attributes are now docstringed as such (a97d368)

Miscellaneous Chores

• release 0.0.2 (60b3cd1)

v0.0.1

0.0.1 (2025-08-01)

Features

• BatchConfigurator implemented and passing tests. (0bc00f2)
• BatchConfigurator implemented and passing tests. (f4859e4)
• BatchConfigurator implemented and passing tests. (edf60fa)
• first attempt at a batch configurator (4d723ed)
• Initial dev version release. Not all features documented in changelog; some commit messages poorly named. Expect better changelogs in subsequent releases (1a7691e)

Bug Fixes

• Add a nozeros toggle to array sizes (1300f3f)
• ci: Fix release-please label permissions (fa5d77f)
• complete GPU tests on release tag (f622a03)
• complete GPU tests on release tag (f622a03)
• correct import error in ODEData.py (cdd3144)
• Corrected ancient typo in SystemValues that made all parameter setting invalid, confirmed test coverage (c05f532)
• fix circular dependency, improve arraysizes interface (98ecf54)
• git: close issue #41 (da83993)
• git: remove local junk from repo #41 (a3cb122)
• Implemented adapters for array size and allocation classes (7a4da6a)
• Improve adapters and access to output_sizes objects through higher objects (649e8cf)
• Output array indices now gated by boolean flags to avoid memory access errors (4f5628c)
• OutputHandling: switched indexing order in 2d arrays to match intended striding. (86a1d2f)
• Plumbing now works between lower-level modules (d79cfdb)
• remove bug introduced in time-saving (ce96d23)
• remove doto #19 (d3457bd)
• Remove todos to close issues (f2e7638)
• Set buffer height methods in output_functions to properties to follow the convention for the rest of the module (3a5c387)
• SingleIntegratorRun and children now re-validate timing after an update (0baad64)
• Swat rename bug in summary_metrics testing (cffd94d)

Miscellaneous Chores

• release 0.0.1 (73c85c5)