Where things stand

Defining the hardware in Vivado

I built the block design — the Zynq processor plus GPIO peripherals for the LED and buttons — generated the HDL wrapper, and wrote the constraints. When I went to export the hardware, Vivado refused: "cannot write hardware definition file — no generated IPI blocks." My first instinct was that I'd typed the export command wrong. I hadn't. Checking the project's actual run status told the real story — nothing had ever been synthesized. The wrapper and constraints existed, the project looked finished, and none of it had actually been built yet. I launched synthesis and implementation properly and got a real bitstream.

Before writing any device-tree code, I had a plan in my head for how the LEDs were wired: through EMIO, the standard PS-to-fabric GPIO route. Reasonable assumption. Instead of building on it, I unzipped the exported hardware file and checked. Vivado's own automation hadn't used EMIO at all — it had used two separate AXI GPIO cores, a different, equally valid approach, sitting at addresses I could read directly out of the file. Nothing was broken. My assumption just wasn't what got built, and I caught it before it cost me anything downstream.

What I ended up with: a verified .xsa — a real bitstream, real hardware handoff data, real processor-init data, all confirmed present rather than assumed.

Yocto setup — and a real dependency maze

This is the part that took longest, and it wasn't a single bug — it was a wrong turn at the tooling level. I matched my Yocto release to a Xilinx layer branch (confirmed against what each layer actually declares, not just what a README claims), and started registering layers. Each one kept surfacing a dependency I hadn't added yet, which is normal for a real board support package — until I followed a deprecated layer's own advice toward what's now AMD's recommended migration path, and watched it cascade four layers deep into virtualization, TPM, and security infrastructure that has nothing to do with getting Linux to boot on a small board with two GPIO cores.

I stepped back. Rather than keep pulling in unrelated infrastructure because a deprecated README pointed there, I checked whether the older, simpler tool for consuming a .xsa directly was still around. It was — deprecated, but fully functional. After tracking down one missing dependency (a submodule that needed initializing) and one confusing tool-path issue (an argument that wanted an installation root, not the binary path I'd correctly located), it worked: a real Yocto machine, generated directly from my own verified hardware.

nRF5340 Auracast bring-up

Before any of the Zynq work makes sense, there needs to be something to receive. I adapted Nordic's official nRF Audio application to a custom nRF5340 board — gateway broadcasting LC3 over Auracast (BIS), headset receiving and presenting the decoded audio as a USB sound card. Nine real bugs fixed along the way, from toolchain mismatches to a mono hardcode in the decode path that silently dropped one stereo channel. Custom board files at github.com/a-teresa/broadcast.

System architecture

Signal flow

The gateway starts broadcasting two LC3 streams at 48kHz, 96kbps — one per stereo channel:

The headset syncs to both BIS streams (bitfield 0x0003) and the ISO RX SDU counters show zero bad packets:

The nRF5340 enumerates on the Linux host as a standard USB audio card:

To verify the full path end-to-end, I fed a 1 kHz sine wave into the gateway's I2S microphone input and recorded the output from the USB audio card in Audacity. Recognisable and periodic — residual jitter is a known open item (clock-domain mismatch on the USB sink path, no drift compensation yet), but the wireless → USB path is confirmed working:

USB host + ALSA on the Zynq

Device tree for the ChipIdea USB host port plus committed kernel config fragments for the host stack and snd-usb-audio. With the board on its 12 V supply, the nRF5340 receiver enumerates and /proc/asound/cards lists it as card 0, attached at usb-ci_hdrc.0-1.

Streaming verified, and a live dashboard

Audio confirmed flowing mic → Auracast → Zynq, and aurascope-dash built to watch it: ALSA capture, FFT, and a Canvas dashboard over Server-Sent Events showing a waveform per stereo channel and a spectrum, plus 30 s of level history.

eBPF tracing in the same dashboard

kprobes on snd_pcm_period_elapsed and __snd_pcm_xrun via ply, feeding a period-interval latency histogram and an xrun counter into the same live feed. 9350 period events observed, all in the 16–32 ms bucket, zero xruns.