innesco-rs
Rust clone of innesco (the IGNITE64 control/DAQ software, timespot/innesco
on baltig). This repo is a separate clean-room-style port: the original
Python/C repository is untouched, and its license/authorship is respected.
Deployment target: AMD Kria K26 SOM (Zynq UltraScale+ MPSoC, quad Cortex-A53, aarch64, Ubuntu 22.04) — the board buffering/timestamping the IGNITE64 detector data.
Layout
| crate | ported from | status |
|---|---|---|
innesco-protocol |
innesco/ecs/protocol.py |
ECS wire protocol (Read/Write/Ping, >BI headers) + TCP client |
innesco-decoding |
innesco/decoding/csrc/decoding.c |
13-byte frame decoder, DCO calibration, position map; golden tests |
innesco-cli |
— | ping, decode subcommands (zero-dependency) |
Zero external dependencies (pure std): cross-compiling for arm64 needs no
cross-linked C libraries, and release binaries are small (panic = "abort", LTO).
Build (x86_64 host)
cargo test --workspace # golden tests
cargo build --release # host binary
export CARGO_TARGET_DIR=/tmp/innesco-rs-target # keep artifacts off NFS
Build for arm64 (Kria K26)
scripts/build-arm64.sh --run-tests # static musl binary (recommended)
scripts/build-arm64.sh --gnu # dynamic, glibc (Ubuntu 22.04 on board)
The script bootstraps everything without root: rustup target add,
a ziglang pip venv as cross linker (via cargo-zigbuild), then
cargo zigbuild --release --target aarch64-unknown-linux-musl.
Artifacts land in dist/aarch64-<musl|gnu>/innesco-cli.
Deploy to the Kria: scp dist/aarch64-musl/innesco-cli <board>:~/ && ssh <board> ./innesco-cli ping localhost 8888.
Porting notes
See docs/porting-notes.md for the Python/C → Rust mapping and the
(deliberate) divergences from the C extension (64-bit arithmetic where the C
overflows 32-bit, strict socket reads where Python truncated silently,
skip-and-count instead of abort on malformed packets).
Open Points
(user-editable — tell me here what to focus on; details agreed with Lucio 2026-11-12)
- Storage layer (after #3):
storage.pyis a 185-lineParquetStoragethread (queue → parquet → S3/MinIO upload, prometheus counters). Rust port options: arrow-rs (lighter; arrow is SoA-native, closes the AoS/SoA divergence) vs polars-rs (feature parity with notebooks, heavy). Decisions needed: (a) arrow-rs vs polars, (b) S3 in-process vs local-only, (c) keep the exact Python column schema (notebook compatibility). - REST API (after #3): 7 endpoints in
rest_api.py(/ping,/metrics,/state,PUT /chunksize,POST /tag,POST /daq,GET /dataset). Port: axum + tokio in a separate crate with feature flag (library stays dependency-free)./stateand/daqdepend on #3; the other four can be ported independently. - Board/slow-control model (biggest item):
ecs/≈ 2,358 lines (asic_state 626, mattonella_state 569, quadrant 502, mattonella 426, …)Board.py268 +DataProcessor.py425. Suggested slices: (1) board+pixel 166 lines → (2) mattonella(+state) 995 → (3) asic_state 626 → (4) quadrant 502 → (5) daq+DataProcessor 574. Mirror-1:1 first (golden I2C tests comparable with Python), type-state validation later. Depends only on the (already ported) protocol crate.
- Real-board validation: no qemu here, no network to the boards.
(a) A raw frame dump from a real acquisition (even small) → decoder
validated on physical data as an official golden test — highest value,
zero prerequisites. (b) network access to a Kria running
data_server→ liveinnesco-cli ping+ harmless I2C read (0x7E). - CI runner: verified 2026-11-12 — no Actions runs on this Gitea repo
and runner admin endpoint is not accessible → no runner configured.
Self-hosting one here is impossible (no Docker daemon). Options: register
a runner on a machine with Docker, or rely on local
scripts/build-arm64.sh --run-testsuntil then.
Recommended order: 4(a) → 3 → 1 → 2 → 5.