Thanks + findings from porting Jet to ESP32-P4 (RISC-V, dual core)

[indecenti]

Hi! We're using Jet (commit c56dfc0) as the rasteriser core of "Vertice", the 3D engine of NucleoOS, an OS for the ESP32-P4 (dual RISC-V @360 MHz, PSRAM). Credit and the MIT notice are kept in the repo. A few findings that might be useful upstream:

1. RV32 has no int64->float instruction. Evaluating row-start attributes (z, u, v, 1/z, brightness) from the int64 edge weights costs ~18 soft conversions per scanline (~3800 cycles/row on the P4, more than 2/3 of raster time). Anchoring screen-space plane equations once per triangle (a0 + ax*dx + ay*dy in float) removes them.
2. PSRAM writes are latency-bound on the P4 (~33 MB/s per core for cached CPU writes). Rendering into small row tiles in internal SRAM (colour + depth), then copying each finished tile out with the 2D DMA, works much better than rendering straight into a PSRAM framebuffer. Both cores pull tiles from an atomic counter; the render queue is binned per tile once per frame.
3. Soft particles need a fill budget: a puff right in front of the camera became a 97x97 sprite with an integer divide per pixel.

Happy to share details or patches if any of this is interesting. Thanks for Jet!

[indecenti]

For context, the project is public: https://github.com/indecenti/NucleoOS-P4

Vertice lives in components/vertice (the Jet core is in components/vertice/core, MIT notice kept), with a small kart game built on it in apps/vxgp. If you find it interesting, a star would be much appreciated. Thanks again!