ereader_m33 Directory Reference

Directory dependency graph for ereader_m33:

Directories
	inc
	src

Detailed Description

The #150 power-saving model: an e-reader spends almost all its time idle on a rendered page, so holding that page on the M85 @ 1 GHz is waste. This app hands the reader to the Cortex-M33 @ 250 MHz and parks the M85, then runs the full mode-switch cycle – the M85 clock-gates and drops into WFI, the M33 holds the page and polls touch, and on each page turn the M33 pokes the M85 over IPC to do the heavy next-page work before it re-parks.

What it teaches

• The render runs on the secondary core, through the production gfx stack. The M33 validates a baked RABOOK1 blob, walks the chapter DOM iteratively (no recursion) through the header-only book.h accessors, and renders the page into an RGB565 framebuffer in external SDRAM with the real ra8_gfx text path. ra8_gfx is dependency-clean, zero-heap and scalar (no Helium), so it links into the freestanding M33 image with no logging backend, no panel driver and no malloc.
• The framebuffer lives in SDRAM at 0x68000000, in a NOLOAD .sdram_bss section the CPU1 linker script pins at the SDRAM base – the same arena-in-SDRAM pattern compile_on_m33 uses – and the M33 publishes its base, geometry, format, glyph count and CRC-32 into the shared SRAM mailbox.
• The M33 CRCs its own pixels, and that is not an accident. Under emulation the two cores share only the on-chip SRAM mailbox; each core's external-SDRAM window is a separate mapping, so the parked M85 cannot read the bytes the M33 wrote. The M33 therefore reads its own framebuffer back to fold the CRC – reading the pixels back is itself the proof they landed – and publishes the value for the M85 to narrate. On silicon the single physical SDRAM is shared, so an M85 re-read would match.
• The wake path is real. The M33 bumps a turn request and pokes IPC0 (ra8_ipc_send_event, the same #149 wake compile_on_m33 uses); the woken M85 restores its clock, does the work the fast core owns, acks, and re-parks; the M33 re-renders the held page, re-folding the identical CRC. The CRC staying stable across every re-render is the assertion.

Blocked on

The control flow is fully exercisable off-target – the M85 really parks and is really woken out of WFI by the M33's poke. What no emulator can show is the power delta of the park (the clock-gate and WFI are functionally exercised, but current draw is not modelled) and real touch input (an emulated page-dwell stands in for a GT911 poll). Both need the bench (#30). The remaining #150 display work – pointing the GLCDC scan-out plane at the M33's framebuffer for a true display-plane handoff – is HIL-bound for the same reason.