Add NEON GF(256) kernels for FEC encode/decode on ARM targets - #2
Open
wkumik wants to merge 3 commits into
Open
Conversation
The FEC codec's only optimized path was 32-bit x86 assembly, so all real targets (Pi, Radxa, OpenIPC) ran the scalar byte-at-a-time table lookup loops. Add NEON implementations of addmul1/mul1 using nibble-split vtbl table lookups, processing 16 bytes per iteration, for both aarch64 and armv7 NEON. The NEON path is enabled only after a one-time self test in fec_init verifies it matches the scalar reference exactly; platforms without NEON (e.g. Pi Zero W, ARMv6) keep the scalar path. Add test_fec (make test_fec) that encodes, erases and decodes blocks and verifies recovery; verified on x86 (scalar), aarch64 (NEON), armv7+NEON and armv6 (scalar) via qemu-user. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TWFTzQ3aWucyVgKpyf6HcT
- Consumer wait: replace the sem_trywait + fixed 200us sleep + retry poll with a blocking sem_clockwait on CLOCK_MONOTONIC honoring the caller's timeout (glibc >= 2.30; other libcs keep the previous behavior). The consumer now wakes immediately on sem_post instead of finishing a fixed sleep, and does not burn CPU polling while idle. The previously commented-out sem_timedwait attempt had a unit bug (timeout scaled by 1e6 instead of 1e3) and used CLOCK_REALTIME, which jumps on NTP sync. - Enqueue ordering: fill the queue slot (payload + metadata) before advancing the published write index, so a consumer woken between the index update and the memcpy can no longer read a partially written slot. - Rx thread: map poll fd indexes to radio interface indexes when building the pollfd array instead of rescanning all interfaces for every ready fd. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TWFTzQ3aWucyVgKpyf6HcT
The video output FIFO to the streamer is opened blocking, so a stalled video player made write() block the router main loop, delaying retransmission requests and all radio/IPC processing (the 'main loop took too long' overloads). Route the writes through a 512KB ring buffer drained by a dedicated writer thread: - The router loop now only memcpys into the ring (never blocks on the player). If the player stalls long enough to fill the ring, new data is dropped and reported as a truncated write, feeding the existing IO error alarm logic. Writer IO errors (EPIPE etc) are reported back on the next output call, feeding the same restart logic as before. - The writer thread polls POLLOUT with a short timeout and writes at most PIPE_BUF bytes per wakeup, so it never blocks indefinitely and stops promptly when the pipe is closed or reopened. Also fix the pipe open retry loop, which did not stop on success: it opened the FIFO up to 200 times, leaking the first 199 descriptors, and pass a mode to open(O_CREAT) as required. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TWFTzQ3aWucyVgKpyf6HcT
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The FEC codec's only optimized path was 32-bit x86 assembly, so all real
targets (Pi, Radxa, OpenIPC) ran the scalar byte-at-a-time table lookup
loops. Add NEON implementations of addmul1/mul1 using nibble-split vtbl
table lookups, processing 16 bytes per iteration, for both aarch64 and
armv7 NEON. The NEON path is enabled only after a one-time self test in
fec_init verifies it matches the scalar reference exactly; platforms
without NEON (e.g. Pi Zero W, ARMv6) keep the scalar path.
Add test_fec (make test_fec) that encodes, erases and decodes blocks and
verifies recovery; verified on x86 (scalar), aarch64 (NEON), armv7+NEON
and armv6 (scalar) via qemu-user.
Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01TWFTzQ3aWucyVgKpyf6HcT