Symbian applications on the N810¶
9 November 2026.
The GPU desktop work provides the rendering path for another use: running Symbian applications on the N810's ARM processor.
The compatibility layer uses EKA2L1's application loader and system services, with a native ARM execution path and adapters for our MBX driver. It loads the application, resolves its library imports, handles requests and sends drawing operations to the GPU.
ChessGenius reaches its menus and board on the tablet. The current menu test averages about 1.2 seconds for a visible response and uses about 40 MiB of runtime memory. Input reaches the queue in about 15 milliseconds, so reducing redraw and runtime work remains the priority. The application needs faster responses before this is comfortable for everyday use.
Users supply compatible Symbian system libraries and application installers in
private storage. The runtime source is under platform/symbian/.
The managed launcher now accepts requests from Pocket's menu and Make. It releases the desktop's rendering context, runs the application, then restores the saved desktop. See Run Symbian applications.
For the tablet setup and graphical desktop, follow N810 setup and the GPU run guide.
10 November: resident GPU masks¶
Masked images now stay in GPU storage through composition and drawing. Native-size images reuse bounded staging storage, and font atlases pack the same glyphs more compactly. Physical pixel checks matched exactly.
Masked sprites can now share a GPU atlas page and one composition job. Exact image, mask and generation matches also reuse completed pages across draw barriers. The Nokia passed 230,400 pixel comparisons with unchanged output.
In an isolated 60-second ChessGenius test, mask pages reduced GPU submissions from 228 to 196 and CPU time from 25.87 to 24.74 seconds, with no mask readbacks. A separate matched test of completed-page reuse reduced CPU time another 3%. Input still takes seconds and varies with paging; reducing runtime handoffs and the remaining redraw work is the next priority. These measurements used completed GPU frames offscreen, with Pocket stopped and debug logging disabled.
10 November: testing the 480 MHz setup¶
We repeated the earlier ChessGenius test twice with the exact same native ARM runtime, Mesa library, 240×320 surface and Down/Up input sequence. Pocket and radio owners stayed stopped. Debug logging and sampling were off. The new kernel reported 480 MHz at both ends of each sample; no frequency setting changed.
| Measurement | Earlier 330 MHz setup | Two 480 MHz runs |
|---|---|---|
| CPU time in the 60-second application run | 24.46 s | 19.79 / 18.26 s |
| First completed image | 16.87 s | 12.89 / 10.91 s |
| Down → next completed image | 1.86 s | 1.86 / 1.41 s |
| Up → next completed image | 2.84 s | 1.19 / 0.97 s |
| Runtime peak RSS | 40.8 MiB | 52.5 / 52.6 MiB |
| Major page faults | 1,736 | 871 / 811 |
CPU time fell 19–25%, and Up responded faster in both runs. Down improved in one run. These are offscreen completed-image timings, not physical button-to-LCD latency or gameplay FPS. All 230,400 mask-fixture pixels matched exactly; both applications exited normally and the module unloaded without a reboot.
This compares complete setups, not frequency alone: the kernel and GPU service configuration also changed. GPU submissions rose from 191 to 259, and counters reported two readbacks totaling 462,720 bytes per run, versus none previously. Memory use also increased. Those discrepancies need tracing before calling this an unchanged rendering pipeline or enabling the newer cache optimizations.
11 November: tracing the extra reads¶
A bounded Nokia diagnostic traced both 231,360-byte reads to Mesa's ordinary render-target path. Runtime surface logging identified a 240×241 temporary made by age-based redraw-history compaction. That temporary does not fit the retained 128×128 or 240×320 image classes.
An opt-in runtime change keeps up to 64 stored commands instead of rasterizing that small history merely because it aged. The existing segment limit and large-history compaction remain. All 5,656 sanitized history cases preserved command order, and the ARM build passed. The change stays disabled by default.
The device comparison stopped before loading its test module because the kernel had recorded a zram decompression failure during swap-in. No reboot was issued; the GPU was unloaded and Pocket stayed stopped. Physical acceptance and the memory-growth investigation remain open. The sampled resident memory includes anonymous heaps and mapped code, so peak RSS alone does not establish an allocation leak or attribute the increase to those two reads.
Code, tests and build inputs¶
| Material | Repository files |
|---|---|
| Implementation | Native runtime (GitHub) |