Post Go back to editing

DpdReset (LUT_RESTORE / FULL) leaves the DPD actuator engaged with an all-zero LUT → Tx output goes to zero

Thread Summary

The user encountered issues with DPD tracking on an ADRV9025/ADRV9029 device, where the actuator LUT RAM was zeroed out after a `DpdReset(LUT_RESTORE)` call, causing the Tx output to disappear. The problem was resolved by downgrading the API and firmware to version 6.0, indicating a compatibility issue between the newer API/Firmware version 7.0 and the stream image. The user confirmed that the default 63-coefficient model with a single non-zero coefficient at (1,1,0) LUT9 = 1.0 + 0j worked correctly after the downgrade.
AI Generated Content
Category: Software
Product Number: ADRV9025

t()` and DPD tracking enabled but **no** `adi_adrv9025_DpdReset()` issued, the actuator is never programmed: its MMR block holds no model program, bit1 of the word at 0xA0820000 is set (our reading: actuator bypass, output = input, since `dpdMeanTxPower` ≈ `dpdMeanTuPower`), and tracking reports 0x340A (AM-AM outliers) on every iteration with `dpdUpdateCount = 0`.
2. Any `adi_adrv9025_DpdReset(TX1, ADI_ADRV9025_DPD_LUT_RESTORE)` or `ADI_ADRV9025_DPD_RESET_FULL` (tracking disabled ≥1 s beforehand, per the release-note workaround) programs the actuator from the loaded model **but the actuator LUT RAM reads all zero afterwards**. Tx output disappears (`dpdMeanTxPower` reads null/INT32_MIN, ORx −56 dBFS, error 0x3405 ORx signal too small). Tx attenuation stays at 10000 mdB, PA protection is disabled on this board and shows no flags, and CLGC-on/off ordering made no difference.
3. The condition is permanent at runtime (re-enabling tracking, reloading the model, repeating the reset in any mode/order, CLGC off, recovery-action changes — none recovered). The only recovery we found is writing bit1 of 0xA0820000 back by host DMA; Tx returns within 3 s and tracking resumes in the state of item 1.

We would like to understand what makes the firmware program the actuator with an empty LUT, and what the intended way is to obtain a non-bypass, non-zero actuator on this device.

## 2. Evidence

Register values were read with the private `adrv9025_CpuDmaMemRead()` at the addresses in `adrv9025_reg_addr_macros.h` (`ADRV9025_ADDR_DPD_ACT_MMR_DPD_CFG_TX1 = 0xA0820000`, `…_DPD_MOD_TX1 = 0xA0820004`, `…_LUT_OFF_TX1 = 0xA0820088`, `…_STAT_TX1 = 0xA08200F0`, `ADRV9025_EXT_DPD_ACTUATOR_TX1_LUT_BASEADDR = 0xA0800000`). We have no register description for these; interpretations below are ours.

### 2.1 Status (`adi_adrv9025_DpdStatusGet`, `adi_adrv9025_ClgcStatusGet`, `adi_adrv9025_TxAttenGet`)

| state | dpdErrorCode | meanTu | meanTx | meanORx | CLGC TxRms / OrxRms | Tx atten |
|---|---|---|---|---|---|---|
| tracking, no DpdReset ("healthy") | 0x340A, iterCount climbs, updateCount 0 | −18.9 dBFS | −19.0 dBFS | −16.5 dBFS | ~7300 / ~9500–10100 | 10000 mdB |
| 2 s after DpdReset(LUT_RESTORE) | 0 | −18.9 dBFS | null (INT32_MIN) | null | ~7070 / **83** | 10000 mdB |
| +20 s, tracking re-enabled | 0x3405 | −18.9 dBFS | null (INT32_MIN) | −56 dBFS | — / ~83 | 10000 mdB |

Pre-actuator level (meanTu) is unchanged; post-actuator level (meanTx) collapses. PA protection is disabled and no error flags are set; `TxAttenGet` and CLGC `activeTxAttenIndex` are unchanged.

### 2.2 Actuator MMR and LUT RAM (TX1)

| state | DPD_CFG (0xA0820000) | DPD_MOD (+0x04) | LUT_OFF (+0x88) | +0x08 | LUT RAM 0xA0800000… |
|---|---|---|---|---|---|
| fresh bring-up, before any reset | 0x00C00003 | 0 | 0 | 0 | non-zero (31663 of 32768 words over 128 KiB), contents change between two reads 10 s apart |
| right after LUT_RESTORE | 0x00C00001 | 0x0010001F | 0x9F8020E0 | 0x003FCF0F | **0 non-zero words of 8192 (first 32 KiB)** |
| +20 s, tracking on (muted) | 0x00C00001 | 0x0010001F | 0x9F8020E0 | 0x003FCF0F | **all zero across the full 128 KiB (0xA0800000–0xA081FFFF)** |
| never-enabled TX2 | 0 | 0 | 0xFFFFFFFF | — | first 32 KiB zero |

- `0x003FCF0F` at +0x08 (repeated at +0x28, +0x48) is exactly the LUT set used by our model (LUTs 0–3, 8–11, 14–21), so the reset does consult the loaded model when it programs the actuator.
- DPD_CFG bit1 is cleared by the reset. Writing `0x00C00003` back by host DMA restores Tx within 3 s (CLGC OrxRms 83 → 9857); in a second run the LUT RAM read all zero immediately before the write-back and Tx still returned. This is why we read bit1 as an actuator bypass/passthrough bit.
- No GPIO-based actuator bypass is configured: the stream image is the vendor-package one (`stream_options.txt`: `Gpio0..15Selection = 255`, `ExternalDpdEnabled = false`, no Tx Ant Cal GPIO) and `streamGpioCfg.streamGpInput0..15 = ADI_ADRV9025_GPIO_INVALID` in the post-MCS init structure. Tx/Rx use PIN enable mode, ORx SPI mode. So the state of bit1 is set by the firmware alone.
- Host DMA writes into the LUT RAM window return success but read back 0 (single word and 1 KiB fill on engaged TX1; single words on idle TX2). MMR writes do take effect.

### 2.3 Variations tried

| variation | result |
|---|---|
| RESET_FULL instead of LUT_RESTORE; LUT_RESTORE then FULL; GUI order (ModelConfigSet → LUT_RESTORE) | mutes |
| LUT_RESTORE on TX1→TX2→TX3→TX4 back to back (SDUG single-band sequence) | mutes |
| ADI firmware default model (63 features, read with `DpdModelConfigGet` before any load) instead of ours (95 features) | mutes |
| CLGC tracking disabled before the reset | mutes |
| `DpdRecoveryActionSet`: REVERT_LUTS_TO_UNITY + RESET_ADAPTATION_STATE on mean Tx/ORx power-low persistent metrics | did not fire within 43 iterations; cause not established |
| `minAvgSignalLevel = 8000` (≈ −12 dBFS, above meanTu) to force the U table | firmware reports 0x3407 (Tx signal too small), `dpdModelTable` stays M, DPD_CFG stays 0x00C00003 |
| `DpdModelConfigGet` | FW-default readback shows `txChannelMask = 0x0` (reserved field) |
| DPD init cal (`ADI_ADRV9025_DPD` = 0x00100000 in the init-cal mask) at runtime | "DPD hardware is being used by other cals" |
| DPD init cal in the boot init-cal mask | init cals pass, but `TrackingCalsEnableSet(DPD)` then returns rc=0 without the DPD bit sticking |
| External path delay cal (TX1) | passes every bring-up (fifoDelay 159–160) |
| ORx level sweep −16 … −36 dBFS, ORx saturation check, TxQEC off, direct-learning, staggered cals | no change to the 0x340A state |
| **FDD** 20 MHz 2T2R profile (`Adrv902X-44NLS-4.9GLR-20MHz-FDD-2T2R`), same carrier, PA continuously on | identical: 0x340A (with intermittent 0x3421 capture timeouts) and updateCount 0 while bypassed; LUT_RESTORE mutes with the same MMR signature and all-zero LUT RAM; bit1 write-back restores Tx. So TDD Tx-enable gating is not the cause. |

## 3. Questions

1. What is the expected content of the actuator LUT RAM after `DpdReset(LUT_RESTORE)` with a model that contains a single non-zero coefficient (ours: i=1, j=1, k=0, LUT 9, 1.0+0j; the other 94 features 0)? Is an all-zero LUT a legal outcome, and under what conditions does the firmware produce it?
2. Which M-table does LUT_RESTORE restore — the model most recently written with `DpdModelConfigSet`, or a coefficient set that only exists after a successful adaptation / DPD init cal? The comment on the reserved `txChannelMask` field in `adi_adrv9025_dfe_types.h` reads: "Reserved Param. Currently Tx channel mask is not supported. Settings are applied to the channel for which DPD init cal was run." Is the DPD init cal a prerequisite for LUT_RESTORE/FULL to produce a non-zero actuator? On our board it cannot run at runtime ("DPD hardware is being used by other cals") and when placed in the boot init-cal mask DPD tracking does not enable afterwards — what is the intended ordering with CFR and the other init cals?
3. Is bit1 of 0xA0820000 the actuator bypass / unity-model mux? Is it expected that tracking runs with bit1 = 1 for the whole run (20 s to several minutes in our tests) and never programs the actuator, reporting 0x340A on every iteration with `dpdUpdateCount = 0`? What does the adaptation compare in that state, given pre- and post-actuator are identical?
4. What do `0x0010001F` at +0x04 and `0x9F8020E0` at +0x88 encode? Is a register description of the actuator MMR block (0xA0820000–0xA0820140) available for support purposes?
5. Is host access to the actuator LUT RAM (0xA0800000, 128 KiB) intended to be read-only? Reads return data, but DMA writes read back 0.
6. Does the firmware zero-fill the LUT RAM on `DpdReset` by design before loading the model, so that an empty LUT means the model-expansion step failed silently? If so, is there any status/error we should see (`dpdErrorCode` reads 0 right after the reset)?
7. In this firmware's GMP convention, is i=1,j=1,k=0 (our unity seed) the identity term, or should the identity be i=0,j=0,k=0 (c·|u(n−i)|^k·u(n−j) form)? Could a model without an identity term be treated as a zero actuator?
8. Is there any sequence that leaves the bypass state with a non-zero actuator before an adaptation has succeeded (e.g. is `ADI_ADRV9025_DPD_MODEL_TABLE_U` selectable on demand)?

Thread Notes

  • Hi  ,

    We tested DPD on the same chip family using the default 63-coefficient model without DPD init cal and it worked — tracking converged, dpdUpdateCount climbed, Tx output was healthy.

    The default model starts with i=0, j=0, k=0 on LUT 0 — the GMP identity term (c · u(n)). Your model has no such term. Your only non-zero coefficient at i=1, j=1, k=0 is a one-sample delay, not unity. The firmware needs the direct-path feature to build a valid actuator LUT.
    Have you tried adding i=0, j=0, k=0 as the first feature in your model, then follow UG-1727 Table 263: DpdModelConfigSet → DpdReset(LUT_RESTORE) per channel → configure tracking → enable. No DPD init cal required.

    Regards,

    G

  • Thanks for the suggestion. We tested it, and it does not change the result on our board:

    1. Both models already contain the (0,0,0) LUT0 feature, with coefficient 0. The ADI FW default 63-feature model, read back with DpdModelConfigGet before any load, also has (1,1,0) LUT9 = 1.0 as its only non-zero coefficient. That is the same placement as ours.

    2. From a clean state, we loaded each model below at runtime (DpdModelConfigSet → DpdReset(LUT_RESTORE)), and read the actuator LUT RAM (0xA0800000, 128 KiB) directly:

      • FW default 63 with (0,0,0)=1.0 and (1,1,0)=0, which is your proposal
      • ours with (0,0,0)=1.0
      • FW default with both set to 1.0

      In all three cases the LUT RAM had 0 non-zero words, the actuator was engaged (bypass bit cleared), and Tx was muted (0x3405).

    3. The FW default model unmodified, and the UG-1727 Table 263 order with LUT_RESTORE on TX1–TX4, give the same result.

    So the empty LUT does not depend on model content. Could you share the details of your working setup?

    • API / ARM FW / DPD-core FW versions
    • the init-cal mask
    • whether DpdReset(LUT_RESTORE) was actually called before tracking converged
    • whether dpdUpdateCount advanced from the bypass state or after the reset

    Ours: API 7.0.0.14, ARM FW 7.0.0.11, TDD 100 MHz and FDD 20 MHz at 4.77 GHz (same result in both).

  • Hi  ,

    The procedure for testing is listed below:

    1. Start the system with the PA off.
    2. Select the appropriate calibration mask (IIO-Osc → IIO Device Attributes → global → calibrate_mask) - corresponding to the used Tx.
    3. Run the Tx LOL and Tx QEC initial calibrations (via IIO-Osc):
        a. Write 1 to the corresponding calibration attribute (e.g., global → calibrate_tx_lol_en).
        b. Then write 1 to the Calibrate property (global → calibrate).
    4. Select the ORx-to-Tx mapping (via IIO-Osc):
        a. IIO-Osc → IIO Device Attributes → debugfs → orx1_to_tx_mapping, orx2_to_tx_mapping, etc.
    5. Power on the PA.
    6. Run the Tx external path delay calibration (via IIO-Osc; make sure the correct Tx channel is selected using calibrate_mask):
       a. Same procedure as step 3.
    7. Run the Tx external LOL calibration (same procedure as steps 3 and 6).
    8. Load the desired model (done in the ZCU102 board's terminal):
        a. root@analog:/sys/bus/iio/devices/iio:device2# cat /boot/ADRV9025_DPD_63Coeff_Model_7.txt > dpd_coef_model

        Note: At the time of testing, this step had no effect

    9. Set the DPD channel mask (IIO-Osc → IIO Device Attributes → global → dpd_tx_mask).
    10. Apply the configuration (IIO-Osc → IIO Device Attributes → global → dpd_tracking_config_set).
    11. Enable DPD on the desired Tx channel (IIO-Osc → IIO Device Attributes → output_voltagex → dpd_en).
    12. Reset DPD (IIO-Osc → IIO Device Attributes → global → dpd_reset).
    13. Read the DPD status (via IIO-Osc or advanced DPD status in the board's terminal):
        a. IIO Device Attributes → debugfs → txx_dpd_status
        b. root@analog:/sys/kernel/debug/iio/iio:device2# cat tx2_advanced_dpd_status

    *To enable any attribute, write 1 to the specific field (e.g., dpd_reset).
    *The main branch of the Linux repository should be used.

    Regards,

    G

  • Thanks for the procedure. We reproduced it step by step on our board. The Tx output still disappears at step 12, so we'd appreciate help on the points below.

    Setup

    • Custom 4T4R board with ADRV9029 (ADRV9025 API), API 7.0.0.14, ARM FW 7.0.0.11, DPDCORE FW loaded. Carrier 4.77 GHz, 100 MHz TDD profile (also reproduced on a 20 MHz FDD 2T2R profile).
    • No ZCU102/IIO. We call the same API functions the adrv902x Linux driver (main branch, also API 7.0.0.14) calls for each attribute.
    • Tested on TX1, with ORx1 mapped to TX1, Tx attenuation 10 dB, and a JESD 100 MHz NR signal.

    Steps as executed

    Step Our call Result
    1–3 Init cals incl. Tx LOL / Tx QEC with PA off pass
    4 TxToOrxMappingSet(ORX1, TX1) ok
    5 PA on ok
    6 InitCalsRun(0x200000, TX1) external path delay pass, fifoDelay = 159
    7 InitCalsRun(0x800, TX1) external LOL pass
    8 Model load skipped (your note says it had no effect), so the FW default model is in use. DpdModelConfigGet returns 63 features with a single non-zero term: (i,j,k) = (1,1,0), LUT9 = 1.0 + 0j —
    9–10 DpdTrackingConfigSet (TX1; mu 50, regularization 20/35, 16384 samples, minAvgSignalLevel 512/512) ok
    11 TrackingCalsEnableSet(TRACK_TX1_DPD, ENABLE) ok. dpdErrorCode 0x340A, meanTx −18.9 dBFS, meanORx −16.7 dBFS, updateCount 0
    12 DpdReset(TX1, ADI_ADRV9025_DPD_RESET_FULL) with DPD tracking still enabled (as in the driver's dpd_reset) rc = 0, Tx output disappears
    13 DpdStatusGet(TX1) at +2 / +8 / +20 / +30 s 0x0 → 0x3405 (ORx signal too small). meanTx null, meanORx −56 dBFS, updateCount 0

    What changes at step 12 (read with adrv9025_CpuDmaMemRead)

    Before reset After reset
    DPD_CFG (0xA0820000) 0x00C00003 0x00C00001 (bit1 cleared)
    DPD_MOD (0xA0820004) 0 0x0010001F
    LUT_OFF (0xA0820088) 0 0x9F8020E0
    Actuator LUT RAM (0xA0800000, 128 KiB) 32728 / 32768 words non-zero all zero
    Model readback (1,1,0) LUT9 = 1.0 unchanged

    Writing bit1 of 0xA0820000 back to 1 brings the Tx output back within 3 s.

    Other variants, all with the same result (LUT RAM all zero after the reset, Tx muted, no API error):

    • LUT_RESTORE, RESET_FULL, and LUT_RESTORE followed by RESET_FULL.
    • With DPD tracking enabled or disabled before the reset.
    • On TX1 only, and on TX1–TX4 back to back.
    • The adrv9025_c_example main.c sequence and both Python scripts we received (Internal_DPD_Script_5p1p0p27_VVDN.py, Txall_DPD_working).
    • Unity models (1,1,0) LUT9 = 1.0 and (0,0,0) LUT0 = 1.0, the all-zero example model, and the FW default model. Every model reads back correctly before and after the reset.
    • TDD and FDD profiles; CLGC on and off.

    In every sequence where the external path delay cal runs after the reset (main.c, the Python scripts), it fails with "PathDelay: Measured TX power is too low". The same cal passes on every bring-up when it runs before the reset. So the firmware's own measurement also sees no Tx signal once the reset has been issued.

    Questions

    1. After DpdReset(RESET_FULL) with the FW default model (unity at (1,1,0) LUT9), what should the actuator LUT RAM contain? On our device it is all zero, and the actuator leaves bypass (DPD_CFG bit1 cleared), so the Tx output is multiplied by zero. Is an all-zero LUT ever a valid result of the reset, and under what condition does the firmware produce it?
    2. Could you run steps 1–13 on your ADRV9025 EVM with API 7.0.0.14 / FW 7.0.0.11 and share DPD_CFG (0xA0820000) and the first words of LUT RAM (0xA0800000) right after step 12? That would show directly whether our device behaves differently.
    3. Is there a firmware-side condition that must be met before the reset so that the LUT expansion from the model happens? For example, a prior DPD init cal or a minimum number of capture iterations. dpdErrorCode reads 0 right after the reset, so we get no indication of a failure.
    4. Minor: the Linux driver defaults dpdMThreshold to 2920, which is a linear amplitude for −21 dBFS. The API 7.0 header says this field takes the squared linear value, ((10^(dBFS/20))·32768)^2. Which one does the firmware expect?
  • Hi  ,

    Unfortunately, we don't have a setup build to test this functionality at the moment. Have you tried using the main branch for generating the Image. The LUT table loading works correctly there.

    Regards,

    G 

  • Hi,

    We had found the casue.

    It was come from the capablity between Ap/Frimware version and stream Image.

    We had upgrade Api/Firmware 7.0, but used the older stream image.

    After Api/Firmware downgrade 6.0, we can bring up with DPD.

    Thanks,

    Shinha.

  • Hi  

    We are glad that you solved the problem.

    Regards,

    G

Before You Switch


Switching languages will make ADI Explorer unavailable. Resume your session by switching back to English and reopening ADI Explorer.