Skip to content

Limitations

What the driver does not do, what it assumes where the datasheet is silent or contradicts itself, and what is not verified on hardware.

Datasheet sections this page relies on

From Texas Instruments SBASB22 -:

  • Open-wire safe mode: 7.3.7 Open Wire Safe Mode, p. 30
  • Block average, moving average and FIR filters: 7.3.11.2 to 7.3.11.4, p. 35
  • Daisy chain (not supported in v1): 7.3.14.2 Daisy Chain, p. 43
  • Channel registers AIN9 to AIN16 (reversed order): 8.3 AIN9 - AIN16 Channel Registers, p. 86

All of them, with the driver API for each: Datasheet map.

Out of scope: the daisy chain

The datasheet's daisy chain (section 7.3.14.2, Figure 7-29) shares CS and SCLK between several ADS9324 and passes each device's SDOUT to the next one's SDI. The driver leaves it out of version 1 and keeps the seam:

  • Why. In a chain a register read or write takes 32 SCLK per device instead of 24, so a single driver object cannot send one chip's frame; the frame belongs to the whole chain, with the position (DAISY_CHN_LOC) and the size (DAISY_CHN_NUM) programmed in every device. The chain works only with 1, 2 or 4 data lanes on D0 to D7, never with the data on SDOUT, which is the driver's default path. Doing it right needs a chain-level object and a host model of several chips, and neither exists yet.
  • What happens if you ask. Config::daisy exists, so the layout does not change when the support arrives; Validate() returns Error::Unsupported when daisy.enabled is set, EnsureInitialized() returns it before any I/O, the daisy-chain registers are never written, and CheckedConfig does not compile (a compile-fail test checks it).
  • What to do instead. Use one driver object per chip, each with its own chip select; the chips may share SCLK and SDI.

Behaviours to know

  • Continuous-mode calibration needs a running CONVST. CalibrationTrigger::Continuous waits 100,000 CONVST periods (Table 7-7) and then reads CALIB_BUSY_FLAG; the chip counts the periods only while CONVST toggles, so a timer must drive CONVST during the wait. With no CONVST the calibration starts at the next pulse and the data after it is not valid until it ends. CalibrationTrigger::SingleShot needs no timer.

  • The randomizer's XOR_BITS_CTL code. The register table gives 0x00F1; a note in the datasheet says 0x00C1. The driver writes 0x00F1 and only that value or 0x0000. If a bench read shows a chip that wants the other value, change reg::kXorBitsRandomizer and nothing else.

  • Coverage of the datasheet. Coverage matrix lists every parameter the datasheet specifies with the symbol that covers it; rows marked GAP are values the driver has no symbol for (limits and delays that are documented in Hardware only), and tools/check_matrix.py fails when a cited symbol is missing.

  • Block average, FIR and Convert(). Convert() is one CONVST pulse. With a block-average filter on the external clock, or a FIR filter, the chip delivers a result only after the block: MeasureOnce() pulses CONVST PulsesPerResult(Filter) times; with your own timer, pulse that many times and then call ReadFrame().

  • Stuck frames are errors. A frame of all 0x00 or all 0xFF is DataMismatch by default (Sampling::reject_stuck_frames): it is what a floating SDOUT reads. Turn it off for a test pattern that really produces it.
  • One calibration at a time with a filter off. RunAdcCalibration() refuses to run with a digital filter on.
  • OrMasked alarm. The sense of the ALRM_MASK bits is not verified (A41); the call needs Alarm::or_mask_acknowledged.
  • Only documented codes. The reserved register codes are unreachable: the enumerations hold the documented values only (for example TP_MODE 1 is not offered, A22).

Assumptions where the datasheet is silent or contradicts itself

Each row has a knob in the host-test chip model (tests/mock_ads9324.hpp, Knobs), so the alternatives are testable. The first column is the number used in the source comments.

# Topic What the driver does
A1 SPI mode Mode 0, derived from the timing diagram: SDI is sampled on the rising edge
A2 td_CKDO of 17 ns against 60 MHz kMaxSdoutClockHz (29 MHz) is the limit for any port that reads SDOUT in mode 0; kMaxClockHz is the 60 MHz of the table; examples run at 10 MHz
A3 Settle after reset: 30 ms or 1 ms Waits 30 ms (the safe one)
A4 RESET pulse width is not given 1 ms, a design choice
A5 Table 7-18 has no step 5 Treated as a numbering slip
A6 GEN_CFG1 reachability in the channel banks Always select the common bank for the command, then the target bank
A7 SDOUT after a non-read frame The driver never relies on it
A8 Register access during a conversion Allowed only outside the readout window
A9 Does SDI decode on a data frame The driver sends zeros while it reads data
A10 Frames shorter or longer than 24 bits The driver always sends 24 bits
A11 RESET_DETECT reset value: Table 7-13 says 1, Table 8-5 says 0 Set after a reset, cleared at start-up, either value tolerated
A12 EN_DIAG_FLAG or EN_STATUS_BITS The register bit name, EN_DIAG_FLAG
A13 XOR "TBD" with 24-bit words The randomizer is refused with 24-bit words
A14 Order of the PRBS and DIAG bytes DIAG first, then PRBS
A15 Meaning of the 24-bit low byte Exposed raw in Frame::low_byte
A16 DWC threshold reset values: register 0xFF00 against text 127 and -128 Thresholds and hysteresis are written before the window is enabled
A17 Glitch filter units A count of consecutive conversions
A18 DWC_RST edge or level Write 1, then 0
A19 Order of the AIN9 to AIN16 bank Mirrored: local index 15 - channel, from the register table
A20 Flag latch and clear Cleared with the group reset before use
A21 DWC_EN bit mapping Follows the bank mapping
A22 TP_MODE codes Only 0 and 2 are reachable
A23 TP_UPD_MODE Exposed as the ramp update point
A24 Test-pattern index add or OR Add, as Table 7-14 shows
A25 Gain equations g / 65536, clamped
A26 Offset sign Equation 1 (added), over the text "subtracted"
A27 Moving average per channel or common Common, one value
A28 DOUT_LENGTH 01 Reserved, unreachable
A29 Phase delay range and unit 0 to 255 CONVST periods
A30 SYSREF semantics 0, 1, 0 with at least one CONVST period
A31 AVG_MODE gate for FIR and moving average The OVR_EN and AVG_MODE pair is always written
A32 DRDY cadence with decimation One DRDY every N conversions
A33 FIR7 -3 dB printed 0.540 A documentation slip; no code depends on it
A34 Block-average clock default The configuration chooses explicitly
A35 INT_TRIG_MODE shared with the single-shot calibration Restored after the calibration
A36 CALIB_BUSY timing Wait 125 ms, then poll with a 25 ms margin
A37 Calibration with common-mode correction Mixed bank types are refused
A38 ADC calibration at power-up Optional, Calibration::run_at_start
A39 CONVST level at the end of the single shot High
A40 Standby wake time and register retention Wake waits 30 ms and re-verifies the registers
A41 ALRM_MASK bit map OrMasked needs the acknowledgement
A42 ALRM_DIS and the pulse width The pulse width is not given
A43 ALRM_POL on DRDY The driver applies the polarity to every source
A44 Reference switch settle Waits 30 ms
A45 td_CVDRDY_lo and CH_XTALK_LOW_SPEED Waits 350 ns; the low-speed bit is a Config field
A46 tPL_CV and tPH_CV names swapped Uses the figures: 50 ns high, 50 ns low at least
A47 Switching characteristics only at IOVDD 1.7 V to 1.9 V Validate() accepts up to 3.6 V
A48 Data-valid window and readout deadline Read within one conversion period

Some of the knobs describe the chip's behaviour only in the model (A9, A10, A17, A18, A20, A23, A25, A27, A29, A30, A32 to A34, A37, A42, A43, A45, A47): they record the assumption and no test depends on the alternative. A real chip that behaves differently shows up on the bench steps below, not in the host tests.

Not verified on hardware

Nothing in this driver is verified on a real ADS9324: the host tests run against a model written from the datasheet. The bench steps:

  1. EnsureInitialized() returns Ok (DEVICE_ID 0x0004) at your SPI clock; try 1 MHz first. This confirms mode 0 (A1) and the reset time (A3).
  2. RunDataPathSelfTest() passes with the inputs unconnected: the data format, the lanes and the mirrored channel mapping.
  3. Apply a known voltage to AIN9 and AIN16 and check microvolts on the right channel: the mirrored bank (A19).
  4. Run RunAdcCalibration() and check the code of a grounded input: the calibration timing (A36, A39).
  5. With 2, 4 or 8 lanes, the packing of your CaptureLanes() against SDOUT mode (the bit order is assumed).
  6. The window alert with a pulse alarm and the OrMasked source mask (A41).

MISRA C++:2023 CodeQL scanning and the ESP32 builds run in CI only; the repository's local checks run the other gates (Testing and CI).