Display Pipeline

From asset to AMOLED

PNG / bitmap atlas
  └─ convert_assets.py
       └─ game_assets.bin + index
            └─ AssetRenderer (RGB565)
                 └─ GameRenderer into PSRAM canvas A
                      └─ diff / dirty regions against canvas B
                           └─ LVGL / BSP flush
                                └─ CO5300 QSPI AMOLED

Web and firmware share source PNGs and palette, but firmware never decodes general PNG at runtime. The converter packs binary alpha and RGB565 plus a validated index; Pillow 11 and 12 must produce byte-identical output.

Why two full-screen canvases

• Compare pixels/regions without reading AMOLED.
• Separate the next render from the current submitted frame.
• Compare partial output to a full host reference pixel by pixel.
• Fall back to one full frame on page changes or complex state.

The cost is 868,624 bytes of PSRAM, acceptable with 8 MiB but not an invitation to add a third canvas.

Draw order

1. Pure black ground.
2. Navigation and structural arcs.
3. Underlay FX.
4. Page body and Makers.
5. Edge gesture / STAR RUSH.
6. Particles.
7. Overlay FX.
8. Callouts and messages.

Dirty regions cover old and new layers. Moving objects and sparse FX use previous ∪ current bounds so colored trails do not remain on black.

RGB565 parity

The browser is RGB888 and the panel RGB565. Both are mapped back to logical CONSOLE-RGB16 colors before layout/semantic comparison. In fixed initial/endgame states at now=0, all six v6 page images matched 114,546 tested pixels per page.

Partial refresh

Stable frames update core, Makers, numbers, meters, particles and feedback. Fingerprint changes and navigation use a full frame. Typical v6 dirty area was about 40.6K pixels/frame with a periodic full refresh around 15/150 frames; stable v7 dirty area is about 15K–18K with roughly 10% periodic full frames.

Before and after

WEB V4DEVICE V7 SNAPSHOT

Drag to compare WEB V4 / DEVICE V7 SNAPSHOT