Skip to content

nRF52 SAADC held-input functional contract

This original MIT slice makes configured SAADC conversions respond to held analog input levels. It does not qualify micro:bit microphone audio capture.

The register facts and conversion equation follow Nordic's SAADC product specification for the shared nRF52 SAADC IP. Buffer/scan semantics additionally follow nRF52833 Product Specification v1.7 (manufacturer PDF mirrored by ANU), §§6.21.3–6.21.4, pp577–579. The nRF52833 instance is SAADC at 0x40007000, IRQ7. The micro:bit Foundation pinmap assigns MIC_IN to P0.05/AIN3 and RUN_MIC to P0.20. A microphone fixture must drive AIN3 through SAADC, rather than the PDM peripheral.

Input and code mapping

Peripheral::set_adc_channel_input(ain, millivolts) holds one physical AIN0..7. adc_channel_count() returns 8. clear_adc_channel_input(ain) releases it. Invalid physical input indices return false without altering an input. The WASM set_adc_channel_millivolts("saadc", 3, millivolts) and clear_adc_channel("saadc", 3) use these generic hooks. The existing STM32 ADC downcast remains the compatibility fallback.

CH[n] is a converter configuration slot, not an AIN or GPIO index. PSELP/PSELN values 1..8 select AIN0..7, so any CH[n].PSELP=4 selects the microphone pad. PSELP=0 disables a slot; invalid selectors are skipped. PSELN is ignored in single-ended mode. In differential mode NC is modeled as ground. Selector9 uses the explicit modeled 3300mV supply, also used by VDD/4 reference.

The configured gain choices are 1/6, 1/5, 1/4, 1/3, 1/2, 1, 2 and 4. Reference is 600mV internal or 825mV modeled VDD/4. Integer rational arithmetic computes (VP-VN) * gain/reference * 2^(resolution-m), where m=0 for single-ended mode and m=1 for differential mode. Division truncates toward zero. Single-ended results saturate to 0..2^N-1; differential results saturate to -2^(N-1)..2^(N-1)-1. DMA stores signed little-endian 16-bit samples.

Un-driven AINs are explicitly modeled at ground, including after releasing the last held input. There is no synthetic 3000mV fallback and no conversion for disconnected slots. Nordic Tier-1 fixtures explicitly select VDD (PSELP=9) and expect 3754 at 12-bit / 938 at 10-bit with the modeled 3300mV supply.

Functional START/scan contract and remaining work

START requires ENABLE and latches PTR/MAXCNT, resets AMOUNT, arms acquisition and generates STARTED. Edits to PTR/MAXCNT affect the next START, not the active buffer. Each accepted SAMPLE visits enabled CH[n] slots once in ascending order and appends their results. A partial final scan stops at MAXCNT without writing past it; END fires only when the buffer fills, then disarms acquisition. SAMPLE before START, after END or while another scan is pending is ignored. MAXCNT is bounded to 32767 samples; zero capacity and no valid enabled slots generate no samples or conversion events. Existing event latches require software clear.

A delay-zero scheduler event or the bare-bus tick drains each scan, not an entire buffer. Only one scheduler entry is announced per trigger; generation tokens prevent canceled entries from completing newer triggers. STOP/disable cancel pending scans and disarm, preserving partial AMOUNT. A new START resets AMOUNT and cancels any previous pending scan. Input/configuration is evaluated when the scan drains; this zero-delay functional model is not acquisition-time sampling. Invalid/non-RAM PTR and DMA bus-write errors remain unsupported; successful RAM writes are proven only with valid RAM buffers.

DONE and RESULTDONE latch on nonempty scans; END latches only at buffer-full. Their enabled bits raise the level IRQ. Nordic event bits use (event offset-0x100)/4, so STARTED=0, END=1, DONE=2, RESULTDONE=3, CALIBRATEDONE=4 and STOPPED=5. Event clear and INTENCLR deassert the line. Acquisition time, oversampling/BURST, local timed sampling, calibration, limit events and resistor networks are not modeled.

START/scan unit gates cover bare bus and scheduled completion, sparse channels, partial buffers/sentinels, pointer edits, amount/events/IRQ and canceled-event restart. Hosted native source proof passed at 18a8cf70: 26 SAADC unit tests, the new ARM scan guest, four tick512 EasyDMA tests and three native WASM routing tests (run36713727900). The receipt records the exact tested merge ref separately from the PR head and the missing ADC ELF artifact limitation. Main landing and the separate whole-chip CorePerf baseline gate remain pending; this is not browser ADC or microphone/audio qualification. The earlier local exact-source stub harness is not engine proof. Rebuild migrated fixture blobs with scripts/tier1/build_nordic_rp2040.sh --nordic-only.

A second native run at ee80c19b preserved the actual hosted ADC ELF, verified SHA256 3875d7473117d105a34abf561c37e1a087915ea775fa99c7eb687d32d3cbc5b0. The same 297 selected functional test executions passed, but the old CPU-base motion benchmark failed at median 0.865205x / min 0.862687x. Its failed-baseline receipt and original motion/GPIO JSONs are retained separately from the earlier green host result. The stacked CPU/GPIO/SAADC candidate is not qualified by either historical run; it needs fresh combined hosted proof.

The stacked candidate at e734e675 now passed native combined qualification (run36720954929): 311 core plus three native WASM selected test executions, including all 26 SAADC unit tests, the actual ARM scan guest and four tick512 EasyDMA tests. On the recorded AMD EPYC 9V45 runner, active median was 6.326197x and motion median/min 1.996539x/1.988887x. The combined receipt distinguishes the PR head from tested merge cfd20eb6, retains original JSONs, complete selected functional log and runner context, and records the downloaded ADC ELF SHA256 f07e2f81dadc606e1d958c198807db291c57367804f82542c455cd3c4803ae5e. This qualifies that native candidate, not browser ADC, microphone capture or the separate whole-chip baseline gate. Host variation prevents interpreting cross-run score changes as a measured SAADC optimization.

Before qualifying microphone capture, implement and prove sample timing/PPI/oversampling, RUN_MIC gating and an explicitly bounded waveform source. Microphone bias, sound-pressure calibration and analog noise require separate board fixtures. No vendor firmware or analog hardware model is added by this slice.