Files

T

Max Gramser 08197c4ae1 Free memory on the CYD and drop the old manual profile (0.2.50)

The CYD restarted under load: a slider fill rounded less than its track made
LVGL draw the fill into a buffer of its own size on every redraw (15-20 KB), and
when that allocation failed the refresh retried until the task watchdog fired.
Track and fill now share one radius, and tests/test_layer_free.py keeps every
layer-making style out of the profiles, the packages and the firmware headers.

A tile keeps only the fields every card uses; climate modes, select options,
weather, sun and timer times, the media title and the vacuum rows moved to an
Extra block that exists only while a state needs one, and the state copy became
a fingerprint. runtime_tiles::model went from 24,296 to 8,216 bytes.

The manual profile that bound tiles to fixed entities in the YAML is gone:
331 TILE* substitutions, DYNAMIC_TILES/TILE_COUNT/DIRECT_ACTIONS, the 80 Home
Assistant subscriptions, its tap handlers and scripts, its own vacuum card, its
guides and helper scripts, and the five stale header copies in the repository
root that two C++ tests were compiling instead of the components. esphome config
of both Easy profiles is unchanged by that cleanup.

tile_scroll and the wake overlay no longer light up on a tap, and the debug
sensor min_free reports the lowest free heap since boot. On the bench CYD static
RAM went from 32.0% to 22.8% and free heap from 64-77 KB to about 101 KB, with
78 KB as the lowest point over thirteen hours without a restart.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

2026-09-16 08:06:41 +02:00

2.8 KiB

Raw Permalink Blame History

Test results app 0.2.50 / firmware 0.2.43 (2026-09-15/16)

Memory for the CYD: layer-free sliders, a leaner tile model, and the old manual profile out of both board profiles. See CHANGELOG 0.2.50 and RELEASING "Compatibility 0.2.50".

Automated

Python (.venv-portal): 175 tests OK. New: tests/test_layer_free.py (no style that makes LVGL draw into a buffer of its own, and every slider fill keeps its track's radius, in both profiles, both packages and the firmware headers) and tests/test_easy_package.py (packages current, no secrets or local paths, the manual profile gone from profiles and packages, runtime parts present, and the generator refuses a fixed Home Assistant subscription).
C++ (clang++ -std=c++17 -Wall -Wextra -Werror -I . -I components/smart_display): 13/13 PASS, with test_runtime_model.cpp and test_tile_controls.cpp also clean under -fsanitize=address,undefined. test_cyd_ui.cpp and test_light_controls.cpp were compiling the stale header copies in the repository root; they include the component headers now and still pass.
generate_packages.py --check green.
esphome config of both Easy profiles is identical before and after the manual-profile cleanup, apart from two reordered lines: the firmware that was tested on the bench is the firmware released.

Build sizes (ESPHome 2026.6.2)

Profile	RAM 0.2.49	RAM 0.2.50	Flash 0.2.49	Flash 0.2.50
easy-cyd-device.yaml	32.0%	22.8% (74,668 B)	78.6%	75.5%
easy-guition-device.yaml	33.0%	23.8% (77,892 B)	19.0%	18.3%

runtime_tiles::model (twenty tile slots) shrank from 24,296 to 8,216 bytes.

Bench boards

Both boards were flashed over the air with firmware 0.2.43 and ran from 19:08 to 08:00 the next morning without a restart: last_boot unchanged, Synced, no memory warning in the log.

	CYD before (0.2.42)	CYD after	Guition before	Guition after
Free heap after the sync	64–77 KB	101–103 KB	51–65 KB	84–97 KB
Lowest free heap since boot	not measured	78.1 KB over 13 hours	not measured	62.8 KB
Largest free block	48 KB	48 KB	31.7 KB	36.9–57.3 KB

The owner used both screens for about ten minutes right after the update: the brightness slider of a light tile and of the light card, the vacuum card with its cleaning settings, cards and page swipes. Home Assistant confirmed every action within about 0.3 s. No lv_draw_buf_create_ex: No memory, no watchdog reset, no rollback. On the Guition, LVGL reported 216–276 ms for single operations while a card was being rebuilt, as it did before this release.

Not tested / for the owner

A screen that still ran the old manual profile. There are none left; the route is gone.
The USB touch calibration was not measured again; its code and documentation are unchanged.

2.8 KiB Raw Permalink Blame History Unescape Escape