T-Deck radio2 scheduling during display redraw

Measured on 18 October 2026 (America/Los_Angeles), using one ESP32-S3 T-Deck with 16 MiB flash and 8 MiB PSRAM. The receive-only harness uses the production CustomSX1262Wrapper, profile scheduler and ST7789LCDDisplay. It never sends radio packets or starts WiFi/Bluetooth. Source base: 22cdb7014, with the changes and application readback hash identified in the machine-readable results.

Change

T-Deck renders into a 153,600-byte RGB565 canvas allocated explicitly in PSRAM. It transfers at most 960 pixels per TFT transaction (384 microseconds of wire time at 40 MHz). Each transfer ends and releases chip select before servicing the radio. CPU rasterization also offers bounded cooperative checkpoints. The callback runs only the radio driver, without recursively entering the mesh dispatcher or UI. Pending-packet, active-reception and transmit guards remain.

The T-Deck radio also uses the existing buffered 8 MHz SPI HAL. The TFT and radio retain their shared HSPI peripheral and separate chip selects. Display wake restores the backlight without repeating controller initialization. Other boards keep their existing display path. If PSRAM allocation fails, the driver warns and falls back to direct rendering.

Same-image measurements

The two profiles were 909.5 MHz / SF7 / BW62.5 and 910.5 MHz / SF8 / BW500, both CR5 with explicit 32-symbol preambles. These are measurement controls, not a packet-delivery or preamble recommendation.

Each stress case ran three times for five seconds, requesting 20 redraws per second. Rendering could take longer than the requested frame interval; actual frame counts and elapsed times are retained in the results. Values below use all samples across the three trials, with worst observed maxima.

Display path	Radio SPI	Mean retune	Maximum retune	Maximum gap between radio service calls
Static screen	Bytewise, 2 MHz	497 us	525 us	584 us
Original direct redraw	Bytewise, 2 MHz	610 us	629 us	76,089 us
PSRAM canvas and bounded strips	Bytewise, 2 MHz	581 us	668 us	1,343 us
PSRAM canvas and bounded strips	Buffered, 8 MHz	468 us	581 us	1,237 us

At a normal requested refresh rate of one frame per second (two four-second trials per case), the original 2 MHz path's maximum service gap was 73,383 us; the new 8 MHz path's was 1,251 us. Reducing refresh frequency alone does not remove the individual blocking redraw.

Retune samples bracket the production retune hooks. They are separate from display frame duration and the gap between radio service calls. Histogram percentiles in the JSON are bucket upper bounds, not exact percentiles.

All 30 trials completed 7,025 profile switches with zero reported switch failures, failed retunes, BUSY timeouts or chip errors. Every final RX-mode and modulation-cache check passed. The minimum remaining loop stack was 3,340 bytes in this harness. No packets were decoded, so these measurements do not establish packet delivery under redraw, loaded-channel behavior, Bluetooth/ WiFi interference, or long-term reliability.

Reproduction and checks

Include tools/hil/tdeck_display_switch.ini in platformio.local.ini, build tdeck_display_switch, and flash the exact T-Deck. The board needs a ROM-compatible DIO boot header; the first QIO-header attempt failed before the application started. The supplied harness recipe selects DIO.

Run the collector with --fps 20 --seconds 5 --repeats 3, then with --fps 1 --seconds 4 --repeats 2. Only one tool may own the selected serial port, and PlatformIO commands must run sequentially.

Sixteen focused native display/R8/radio tests passed. They cover pixel parity, clipping, partial updates, strip bounds, released SPI before callbacks, callback recursion protection, PSRAM failure, mode toggling and sleep/wake. The new display regression is included in CI.

Full Companion smoke check

The optimized canonical Full Companion image v1.18.0.1-tdeck-radio2-fast-22cdb701 compiled and passed its unchanged RAM gate (230,680 internal bytes available; 164,896 required). It was installed on the same board. USB ASCII commands and Binary Companion device queries worked; Bluetooth reported on. Device queries retained 350 contacts and 40 channels. A Bluetooth client connection was not tested.

With the screen held on, a temporary receive-only secondary accumulated 759 switches with zero switch errors during a short runtime check. Its maximum successful active-pair retune was 1,013 us and the guarded budget was 1,115 us. One additional primary packet was received during that interval. This is a smoke check rather than a controlled packet-delivery trial. Returning to single-profile RX power saving raised the lifetime max diagnostic to 4,510 us; that teardown is outside the active pair's measured switching budget.

The board was left running Full Companion with Bluetooth on, radio2 off and its original button-pairing USB display mode. Seven hosted-document tests also passed after adding this report.