Repository navigation
16 code review - #17
Merged
Merged
Conversation
See #16 also add pixi files and other minor changes
Build CSC-style column pointers (col_ptr/col_idx) when the sparsity pattern is set, and group pointers (grp_ptr/grp_cols) when the partition is set, instead of scanning the whole pattern for every column (count/pack over icol) and every group (over ngrp). * compute_indices now takes n and also builds the column index (a stable counting sort, so element order and results are unchanged). * new compute_group_index, called after dsm, for a user-supplied partition, and for the dense partition. * compute_jacobian_standard, compute_jacobian_partitioned and columns_in_partition_group now use index slices; the latter no longer grows arrays by concatenation. * compute_sparsity_dense builds the indices after filling irow/icol. * set_sparsity_pattern now also rejects irow<1 or icol<1. Tridiagonal test problem (n=m=20000, central diffs, -O2): unpartitioned 2.31s -> 0.24s, partitioned 1.75s -> 0.001s. Test output (jacobians, partitions, function counts) is unchanged.
* the 2-point methods used for sparsity_mode=4 are now looked up once in set_sparsity_mode (new sparsity_class_meths component) rather than on every compute_sparsity_random_2 call. * initialize(classes=...) now looks up each distinct class once instead of once per variable.
maxlst = sum(row_nnz**2)/n can be 0 for very sparse patterns, in which case the column selection loop never runs and jcol keeps a stale value. Use max(1,...) so at least one column is always examined.
On a partial cache hit, the missing functions were computed into the same f array that held the cached values. Since f is intent(out) in the user function, the cached values could be lost (e.g., if the user function initializes f), producing wrong Jacobian elements. The missing functions are now computed into a temporary array and merged into f. Results are also no longer cached if the user function raised an exception.
With sparsity_mode=3, compute_sparsity is null until the user calls set_sparsity_pattern. compute_jacobian called it unconditionally, so computing a Jacobian first crashed. It now raises an exception (code 30) instead, and also checks for exceptions raised while computing the sparsity pattern.
divide_interval returns fewer points than requested for large num_points (the top points are clamped to the upper bound and the duplicates removed), so compute_sparsity_random_2 read past the end of the coefficient array for num_sparsity_points >~ 164. Zero also crashed. num_sparsity_points must now be in [1, max_num_sparsity_points] (a new public constant, 50); other values raise an exception. All points are distinct and interior up to ~80, so outputs for allowed values are unchanged. Add divide_interval_test with assertions for the documented example, n=1..50, and the initialize validation.
The sparsity tolerances were compared absolutely, so functions of small magnitude (e.g. 1e-20) lost real dependencies, and large ones could pick up round-off noise. * equal_within_tol has a new optional `relative` argument. * sparsity_mode=2 function values and the linear-slope checks now use relative comparisons. * the sparsity_mode=4 zero check now tests whether the change in the function (|J|*dx) is within function_precision_tol*|f(x)| at every point (the same test as mode 2), instead of |J| against an absolute linear_sparsity_tol. * add sparsity_scaling_test: f, 1e-20*f, and 1e20*f must give the same pattern in modes 2 and 4.
The hash table is allocated 0:isize-1, but print looped over 1..size(me%c), skipping slot 0 and reading past the end of the table. Loop over lbound..ubound instead. Add cache_test, which prints the cache to a scratch file and checks that exactly the occupied slots are printed (including slot 0 of a one-slot cache), plus the uninitialized-cache message.
New diff_test calls diff directly for orders 1-3 over symmetric and one-sided cases (both directions), x0=0, absolute/relative/zero eps, relative/absolute accr, very wide and very narrow intervals, and functions with high-frequency terms. For each case, the actual error must be within the estimated error and ifail must be consistent with the requested tolerance. Also covers invalid inputs (ifail=2), too-small intervals (ifail=3), constant functions, a kink, the overflow guard, and user termination at every function evaluation (ifail=-1, no further evaluations, next call unaffected). The only uncovered statement is unreachable in double precision.
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## master #17 +/- ##
===========================================
+ Coverage 66.30% 86.08% +19.78%
===========================================
Files 5 5
Lines 1938 2027 +89
===========================================
+ Hits 1285 1745 +460
+ Misses 653 282 -371 🚀 New features to boost your workflow:
|
The DSM partition was computed from the nonlinear pattern only, so columns sharing a row through a linear (constant) element could be grouped and perturbed together, silently corrupting the nonlinear elements in that row. * dsm_wrapper now partitions on the nonlinear + linear pattern. * set_sparsity_pattern stores the linear pattern before partitioning. * a user-supplied partition is checked for consistency against the full pattern (new exception 33). * linear indices < 1 are now rejected. Add linear_partition_test: checks the partitioned jacobian against the analytic one (forward, central, class 3), that it is identical to the unpartitioned jacobian, and that inconsistent user partitions are rejected.
dfunc called the numdiff_type terminate (a no-op, since the exception was already raised) instead of the diff_func one, so diff kept evaluating the function after a user termination (and with the cache, the termination status was replaced by a DIFF error). It also did not check for a termination from the info function before evaluating the function. Add terminate_test: terminates at every function evaluation and every info call for diff and finite difference modes (including during the sparsity computation), and checks that no further evaluations occur.
A chunk_size of 0 made expand_vector allocate a size-0 array and then write past its end. expand_vector now grows by max(1,chunk_size), and numdiff_type and function_cache store max(1,...) as well. Add chunk_size_test: chunk sizes 0, -3, 1, 5, 100 through expand_vector, unique, the function cache, and a sparsity computation.
options_test covers jacobian_methods/classes, perturb modes 1-3 and set_dpert (checking the actual steps), class method selection at the bounds (with/without partitioning, including when no method fits), sparsity bounds and dpert_for_sparsity, diff outside the bounds, and that the cache does not change the results (including a partial hit). validation_test covers each reachable input error code, error status, formulas, printing a method, and returning/printing the sparsity pattern with linear elements and a partition. Fix: with partitioning enabled, a function that does not depend on x (empty sparsity pattern) raised an exception, since dsm rejects an empty pattern. All columns are now put in one group.
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
1 participant
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.
No description provided.