Dictionary: Improve Dictionary Search Efficiency

[nithin-aashik-mcw]

Changes Proposed

• Optimized the root channel lookup in create_search_channels() by replacing the repeated linear search with a lookup table indexed by (ciphone, ci2phone).
• Previously, each dictionary word performed an O(n_root_chan) scan to locate an existing root channel, increasing the time required to construct the search tree.
• Added a dynamically allocated lookup table based on the number of Context-Independent (CI) phones in the acoustic model, enabling O(1) root channel lookup.
• The optimization only changes the lookup mechanism used during search tree construction. The search tree structure and recognition behavior remain unchanged.

Validation

• Wrote a benchmark for Decoder.add_word(update=True) to evaluate the performance of rebuilding the search tree.
• The benchmark identified create_search_channels() as a performance bottleneck, specifically the repeated linear search for existing root channels.
• Reran the benchmark after implementing the optimization and verified the performance improvement.
• All existing functionality tests pass successfully with no functional regressions observed.

Performance Results

Benchmark	Metric	Before	After	Improvement
Decoder.add_word(update=True)	ms/call	62.080	51.899	16.4% faster

[lenzo-ka]

Nice optimization — replacing the O(n_root_chan) linear scan in create_search_channels() with an O(1) root_lookup[ciphone][ci2phone] table is a clean win. The lookup is logically equivalent to the original: the table is zero-initialized so NULL means "not found", it is populated exactly when a new root channel is created, and single-phone words continue before the lookup (ngram_search_fwdtree.c:202) so the ciphone/ci2phone indices are always valid CI phones in [0, n_ci). The root_chan_t *** type matches ckd_calloc_2d(n_ci, n_ci, sizeof(root_chan_t *)).

One issue to fix before merge, though: root_lookup is never freed.

It is allocated in init_search_tree() but deinit_search_tree() (which frees every other search-tree array) does not free it. deinit_search_tree() runs both at teardown (ngram_fwdtree_deinit) and before every rebuild (ngram_fwdtree_reinit), so the whole n_ci × n_ci table leaks on every search-tree build — including each Decoder.add_word(update=True), which is exactly the path this PR optimizes.

Confirmed under AddressSanitizer/LeakSanitizer (Linux/GCC), single decoder lifecycle via test_fwdtree:

Direct leak of 336 byte(s) in 1 object(s)      (row pointers, n_ci=42)
Indirect leak of 14112 byte(s) in 1 object(s)  (42 x 42 table)
SUMMARY: AddressSanitizer: 14448 byte(s) leaked in 2 allocation(s).
    #3 init_search_tree src/ngram_search_fwdtree.c:85

That is ~14 KB per build, and it grows without bound across repeated add_word(update=True).

The fix is one line, added to deinit_search_tree() alongside the other frees:

ckd_free_2d(ngs->root_lookup);
ngs->root_lookup = NULL;

With that added, the LeakSanitizer run comes back clean. deinit_search_tree() is the right place since it is paired with the init_search_tree() that allocates the table, covering both the rebuild and teardown paths.

[nithin-aashik-mcw]

Good catch. root_lookup was missing from deinit_search_tree(), causing a leak on every search-tree rebuild. I've added the corresponding ckd_free_2d() and verified that the leak is resolved.