afiqzudinhadi/zmk-steno-engine
1
0
Fork 0
Code Issues 3 Pull requests Projects Releases Packages Wiki Activity Actions

Fix partition + CHD: empirical bytes/entry, 1:1 bucket ratio

Browse source 
Partition: ratio-based split overloaded right side (113K entries →
875KB estimate → heavy trimming to 58K). Now uses 4.5 bytes/entry
from benchmark data → ~34K left, ~113K right, both fit budgets.

CHD: 1:1 bucket ratio (entry_count buckets) prevents hash collision
failures on large partitions. Most buckets have 1 key → trivial
placement. Displacement array grows but still small.
This commit is contained in:
afiqzudinhadi 2026-07-02 15:29:25 +08:00
parent 99bf19180e
commit 1339ec4525
1 changed files with 9 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

17
tools/compile_mphf.py
View file

@ -170,7 +170,7 @@ def build_chd(keys_and_bytes, entry_count):
displacements[bucket] = d value
slot_to_entry_idx[slot] = index into keys_and_bytes, or -1 if empty
"""
bucket_count = max(entry_count // 2, min(entry_count, 16))
bucket_count = max(entry_count, 16)
# Assign keys to buckets
buckets = defaultdict(list)
@ -271,7 +271,7 @@ def compile_mphf(entries, max_size=None, block_size=4096):
unique_count = len(unique_translations)
# Estimate size
bucket_count = max(entry_count // 2, min(entry_count, 16))
bucket_count = max(entry_count, 16)
est_value_bits = max(1, math.ceil(math.log2(max(unique_count, 2))))
est_disp_bits = 16 # conservative
est_disp_bytes = (bucket_count * est_disp_bits + 7) // 8
@ -318,7 +318,7 @@ def compile_mphf(entries, max_size=None, block_size=4096):
file=sys.stderr)
entry_count = len(keys_and_bytes)
bucket_count = max(entry_count // 2, min(entry_count, 16))
bucket_count = max(entry_count, 16)
print(f" Building CHD MPHF: {entry_count} entries, {bucket_count} buckets...",
file=sys.stderr)
@ -509,16 +509,17 @@ def partition_entries(entries, left_budget, right_budget):
"""Partition dict entries by importance into left (central) and right (peripheral).
Left gets highest-importance entries first (most common single-stroke words).
Right gets remaining entries. Split point is proportional to flash budgets.
compile_mphf's internal trimming handles any overshoot.
Right gets remaining entries. Uses empirical ~4 bytes/entry from MPHF benchmarks
to estimate how many entries each budget can hold.
Returns (left_entries, right_entries).
"""
sorted_entries = sorted(entries, key=lambda e: score_entry(e[0], e[1]))
total_budget = left_budget + right_budget
left_ratio = left_budget / total_budget
left_max = int(len(sorted_entries) * left_ratio)
# ~4 bytes/entry average from 583KB / 147K entries benchmark
# Use 4.5 for safety margin (smaller partitions have higher per-entry overhead)
est_bytes_per_entry = 4.5
left_max = int(left_budget / est_bytes_per_entry)
left_max = max(1, min(left_max, len(sorted_entries) - 1))
left_entries = sorted_entries[:left_max]