Rebase: Unified SDIRK tableau structure and generic perform_step!#3030
Open
ChrisRackauckas-Claude wants to merge 5 commits intoSciML:masterfrom
Open
Rebase: Unified SDIRK tableau structure and generic perform_step!#3030ChrisRackauckas-Claude wants to merge 5 commits intoSciML:masterfrom
ChrisRackauckas-Claude wants to merge 5 commits intoSciML:masterfrom
Conversation
Rebase and cleanup of PR SciML#2779 onto current master. - Implement SDIRKTableau struct consolidating all Butcher coefficients - Add generic perform_step! reducing duplication across 20+ methods - Migrate ImplicitEuler, SDIRK2, Cash4, all ESDIRK/SFSDIRK methods - Preserve TRBDF2 special handling with unified coefficients - Unify KenCarp/Kvaerno tableaus with additive splitting support - Add Cash4Tableau_unified with proper embedding support - Add verbose parameter to alg_cache for upstream compatibility - Add regression tests for SDIRK methods with explicit first stages Note: Performance regression identified - generic perform_step! uses Vector{typeof(u)} for stage storage and dynamic dispatch which causes significant overhead vs the original flat-field specialized methods. This needs optimization before merging. Co-Authored-By: Krish Gaur <[email protected]> Co-Authored-By: curd <[email protected]> Co-Authored-By: Chris Rackauckas <[email protected]>
…uctors All unified SDIRK tableau constructors now call the original (correct) per-method tableau constructors and extract z-space coefficients from their fields. This fixes pervasive errors in the original refactor: - ESDIRK methods (Kvaerno3/4/5, KenCarp3/4/5/47/58, ESDIRK436L2SA2, ESDIRK437L2SA, ESDIRK547L2SA2, ESDIRK659L2SA, ESDIRK54I8L2SA, CFNLIRK3): A[1,1] was γ instead of 0 for explicit first stage; A matrix rows were incorrect; b was bhat instead of main weights; b_embed was bhat instead of btilde; wrong γ values in several cases - SDIRK2: A[2,1] was 1-γ instead of -1; b was wrong; b_embed missing - SSPSDIRK2: γ was 1-1/√2 instead of correct 1/4 - TRBDF2: b_embed was bhat instead of btilde - Hairer4/42: wrong γ and coefficients - SFSDIRK4-8: completely wrong coefficients and γ values Co-Authored-By: Chris Rackauckas <[email protected]>
Contributor
Author
Fixed: All unified SDIRK tableau coefficientsThe unified tableau constructors had pervasive errors causing 10x-50,000x performance regressions. All constructors now call the original (correct) per-method tableau constructors from Bugs fixed:ESDIRK methods (Kvaerno3/4/5, KenCarp3/4/5/47/58, ESDIRK436L2SA2, ESDIRK437L2SA, ESDIRK547L2SA2, ESDIRK659L2SA, ESDIRK54I8L2SA, CFNLIRK3):
SDIRK methods:
Test results:
|
…ia 1.10 LTS
The `@inferred init()` tests were failing on Julia 1.10 LTS because the
return type was inferred as Union{..PIController.., ..PIControllerCache..}
instead of a concrete type.
Root cause: Two code paths for creating controllers existed:
1. `default_controller_v7` returning new-style controllers
2. `legacy_default_controller` returning legacy controllers
Julia 1.10's type inference couldn't determine at compile time which
path would be taken, resulting in a Union type.
Changes:
- Always use `default_controller_v7` for the controller default parameter
- Enable `CompositeController` for composite algorithms instead of returning
`nothing` (which triggered the legacy fallback)
- Add `new_controller_from_legacy_params` to create new-style controllers
when legacy parameters are provided, ensuring type stability
- Replace `legacy_default_controller` call with new controller constructors
This fixes the Regression_II (LTS) test failure without breaking backwards
compatibility for users who specify legacy controller parameters.
Co-Authored-By: Chris Rackauckas <[email protected]>
Co-Authored-By: Claude Opus 4.5 <[email protected]>
Co-Authored-By: Chris Rackauckas <[email protected]>
The original ESDIRK659L2SA btilde values from sdirk_tableaus.jl had sum(btilde) ≈ -1.69, which violates the consistency requirement that sum(bhat - b) = 0 for embedded RK methods. This caused the algorithm to take 9979+ steps and hit MaxIters on simple problems like exponential decay. The fix corrects btilde9 so that sum(btilde) = 0. This makes the algorithm converge correctly (1322 steps, Success, error=1e-15 vs MaxIters, error=0.62). Note: The individual btilde coefficients may still be incorrect (based on how the error estimator behaves), but this fix is a significant improvement. The coefficients should be verified against the original Kennedy-Carpenter paper for optimal efficiency. Co-Authored-By: Chris Rackauckas <[email protected]> Co-Authored-By: Claude Opus 4.5 <[email protected]>
Contributor
Author
ESDIRK659L2SA FixFixed a critical bug in ESDIRK659L2SA where Before fix (master)
After fix
The fix corrects Note: The individual btilde coefficients may still be suboptimal (1322 steps vs ~10 steps for similar order methods). The coefficients should be verified against the original Kennedy-Carpenter paper for optimal efficiency. All SDIRK tests pass |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Rebase of #2779 onto current master with fixes:
SDIRKTableaustruct consolidating all Butcher coefficients usingStaticArraysperform_step!replacing 20+ method-specific implementationsCash4Tableau_unifiedwith proper dual-embedding supportverboseparameter compatibility with upstreamalg_cachesignaturesPerformance Regression Warning
Benchmarking revealed significant performance regressions in the generic implementation:
Root causes:
Vector{typeof(u)}for stage storage vs flat struct fieldsmax_iters, suggesting error estimation bugs in the generic code pathThese must be resolved before merging. The original PR #2779 had similar issues.
Test plan
🤖 Generated with Claude Code