Skip to contentSkip to conversation

drm/etnaviv: detecting disabled Vivante GPU?

DISCUSSION Andre Przywara 4 participants 6 messages Sep 3, 2025 – Sep 10, 2025 Arm/Allwinner sunXi SoC support

lore
Share & tools
lore
#0 lore raw

Hi,

the Allwinner A523/A527/T527 family of SoCs feature a Vivante "VIP9000"(?) NPU, though it seems to be disabled on many SKUs. See https://linux-sunxi.org/A523#Family_of_sun55iw3 for a table, the row labelled "NPU" indicates which model has the IP. We suspect it's all the same die, with the NPU selectively fused off on some packages.

Board vendors seem to use multiple SKUs of the SoC on the same board, so it's hard to say which particular board has the NPU or not. We figured that on unsupported SoCs all the NPU registers read as 0, though, so were wondering if that could be considered as a bail-out check for the driver?
At the moment I get this, on a SoC with a disabled NPU:

[    1.677612] etnaviv etnaviv: bound 7122000.npu (ops gpu_ops)
[    1.683849] etnaviv-gpu 7122000.npu: model: GC0, revision: 0
[    1.690020] etnaviv-gpu 7122000.npu: Unknown GPU model
[    1.696145] [drm] Initialized etnaviv 1.4.0 for etnaviv on minor 0
[    1.953053] etnaviv-gpu 7122000.npu: GPU not yet idle, mask: 0x00000000

Chen-Yu got this on his board featuring the NPU:

    etnaviv-gpu 7122000.npu: model: GC9000, revision: 9003

If I get the code correctly, then etnaviv_gpu_init() correctly detects the "unsupported" GPU model, and returns -ENXIO, but load_gpu() in etnaviv_drv.c then somewhat ignores this, since it keeps looking for more GPUs, and fails to notice that *none* showed up:
/sys/kernel/debug/dri/etnaviv/gpu is empty in my case.

Quick questions:
- Is reading 0 from VIVS_HI_CHIP_IDENTITY (or any other of the ID registers) an invalid ID, so we can use that to detect those disabled NPUs? If not, can any other register used to check this? The whole block seems to be RAZ/WI when the NPU is disabled.

- Would it be acceptable to change the logic to error out of the driver's init or probe routine when no GPU/NPU has been found, at best with a proper error message? As it stands at the moment, the driver is loaded, but of course nothing is usable, so it keeps confusing users.

Happy to provide a patch, but just wanted to test the waters.

Cheers,
Andre

#1 lore raw
Quoted 10 lines · 在 2025-09-04星期四的 00:29 +0100,Andre Przywara写道:

Hi,

the Allwinner A523/A527/T527 family of SoCs feature a Vivante "VIP9000"(?) NPU, though it seems to be disabled on many SKUs. See https://linux-sunxi.org/A523#Family_of_sun55iw3 for a table, the row labelled "NPU" indicates which model has the IP. We suspect it's all the same die, with the NPU selectively fused off on some packages.

Board vendors seem to use multiple SKUs of the SoC on the same board,

In the case of Amlogic things, they seem to provide different DT's, see arch/arm64/boot/dts/amlogic/meson-g12b-a311d-bananapi-m2s.dts and arch/arm64/boot/dts/amlogic/meson-g12b-s922x-bananapi-m2s.dts ; the A311D is NPU-equipped SKU. (Although I don't know why they included different DTSI's and then still enable &npu node in a311d variant of board DT)

Quoted 40 lines

so it's hard to say which particular board has the NPU or not. We figured that on unsupported SoCs all the NPU registers read as 0, though, so were wondering if that could be considered as a bail-out check for the driver?
At the moment I get this, on a SoC with a disabled NPU:
[    1.677612] etnaviv etnaviv: bound 7122000.npu (ops gpu_ops) [    1.683849] etnaviv-gpu 7122000.npu: model: GC0, revision: 0 [    1.690020] etnaviv-gpu 7122000.npu: Unknown GPU model [    1.696145] [drm] Initialized etnaviv 1.4.0 for etnaviv on minor 0 [    1.953053] etnaviv-gpu 7122000.npu: GPU not yet idle, mask:
0x00000000

Chen-Yu got this on his board featuring the NPU:
    etnaviv-gpu 7122000.npu: model: GC9000, revision: 9003

If I get the code correctly, then etnaviv_gpu_init() correctly detects
the "unsupported" GPU model, and returns -ENXIO, but load_gpu() in etnaviv_drv.c then somewhat ignores this, since it keeps looking for more
GPUs, and fails to notice that *none* showed up:
/sys/kernel/debug/dri/etnaviv/gpu is empty in my case.

Quick questions:
- Is reading 0 from VIVS_HI_CHIP_IDENTITY (or any other of the ID registers) an invalid ID, so we can use that to detect those disabled
  NPUs? If not, can any other register used to check this? The whole block seems to be RAZ/WI when the NPU is disabled.

- Would it be acceptable to change the logic to error out of the driver's init or probe routine when no GPU/NPU has been found, at best with a proper error message? As it stands at the moment, the driver is loaded, but of course nothing is usable, so it keeps confusing users.

Happy to provide a patch, but just wanted to test the waters.

Cheers,
Andre

#2 lore raw
In reply to Andre Przywara

Hi

Quoted 19 lines

the Allwinner A523/A527/T527 family of SoCs feature a Vivante "VIP9000"(?) NPU, though it seems to be disabled on many SKUs. See https://linux-sunxi.org/A523#Family_of_sun55iw3 for a table, the row labelled "NPU" indicates which model has the IP. We suspect it's all the same die, with the NPU selectively fused off on some packages.

Board vendors seem to use multiple SKUs of the SoC on the same board, so it's hard to say which particular board has the NPU or not. We figured that on unsupported SoCs all the NPU registers read as 0, though, so were wondering if that could be considered as a bail-out check for the driver?
At the moment I get this, on a SoC with a disabled NPU:

[    1.677612] etnaviv etnaviv: bound 7122000.npu (ops gpu_ops)
[    1.683849] etnaviv-gpu 7122000.npu: model: GC0, revision: 0
[    1.690020] etnaviv-gpu 7122000.npu: Unknown GPU model
[    1.696145] [drm] Initialized etnaviv 1.4.0 for etnaviv on minor 0
[    1.953053] etnaviv-gpu 7122000.npu: GPU not yet idle, mask: 0x00000000
Quoted 9 lines

Chen-Yu got this on his board featuring the NPU:

    etnaviv-gpu 7122000.npu: model: GC9000, revision: 9003

If I get the code correctly, then etnaviv_gpu_init() correctly detects the "unsupported" GPU model, and returns -ENXIO, but load_gpu() in etnaviv_drv.c then somewhat ignores this, since it keeps looking for more GPUs, and fails to notice that *none* showed up:
/sys/kernel/debug/dri/etnaviv/gpu is empty in my case.

Looks fine to me - no wrong behavior.

Quoted 6 lines

Quick questions:
- Is reading 0 from VIVS_HI_CHIP_IDENTITY (or any other of the ID registers) an invalid ID, so we can use that to detect those disabled NPUs? If not, can any other register used to check this? The whole block seems to be RAZ/WI when the NPU is disabled.

Quoted 6 lines

- Would it be acceptable to change the logic to error out of the driver's init or probe routine when no GPU/NPU has been found, at best with a proper error message? As it stands at the moment, the driver is loaded, but of course nothing is usable, so it keeps confusing users.

From an application standpoint, it’s not confusing since there is no etnaviv device to interact with. The user might wonder about the kernel messages, but that’s actually caused by an incorrect device tree. If the SoC doesn’t have an NPU, it shouldn’t be enabled in the DTS.

Signature · 5 lines
greets
--
Christian Gmeiner, MSc

https://christian-gmeiner.info/privacypolicy
#3 lore raw
Quoted 46 lines · On Thu, 4 Sep 2025 12:10:30 +0200 Christian Gmeiner…

Hi

Quoted 19 lines

the Allwinner A523/A527/T527 family of SoCs feature a Vivante "VIP9000"(?) NPU, though it seems to be disabled on many SKUs. See https://linux-sunxi.org/A523#Family_of_sun55iw3 for a table, the row labelled "NPU" indicates which model has the IP. We suspect it's all the same die, with the NPU selectively fused off on some packages.

Board vendors seem to use multiple SKUs of the SoC on the same board, so it's hard to say which particular board has the NPU or not. We figured that on unsupported SoCs all the NPU registers read as 0, though, so were wondering if that could be considered as a bail-out check for the driver?
At the moment I get this, on a SoC with a disabled NPU:

[    1.677612] etnaviv etnaviv: bound 7122000.npu (ops gpu_ops)
[    1.683849] etnaviv-gpu 7122000.npu: model: GC0, revision: 0
[    1.690020] etnaviv-gpu 7122000.npu: Unknown GPU model
[    1.696145] [drm] Initialized etnaviv 1.4.0 for etnaviv on minor 0
[    1.953053] etnaviv-gpu 7122000.npu: GPU not yet idle, mask: 0x00000000
Quoted 9 lines

Chen-Yu got this on his board featuring the NPU:

    etnaviv-gpu 7122000.npu: model: GC9000, revision: 9003

If I get the code correctly, then etnaviv_gpu_init() correctly detects the "unsupported" GPU model, and returns -ENXIO, but load_gpu() in etnaviv_drv.c then somewhat ignores this, since it keeps looking for more GPUs, and fails to notice that *none* showed up:
/sys/kernel/debug/dri/etnaviv/gpu is empty in my case.

Looks fine to me - no wrong behavior.

Quoted 6 lines

Quick questions:
- Is reading 0 from VIVS_HI_CHIP_IDENTITY (or any other of the ID registers) an invalid ID, so we can use that to detect those disabled NPUs? If not, can any other register used to check this? The whole block seems to be RAZ/WI when the NPU is disabled.

Quoted 6 lines

- Would it be acceptable to change the logic to error out of the driver's init or probe routine when no GPU/NPU has been found, at best with a proper error message? As it stands at the moment, the driver is loaded, but of course nothing is usable, so it keeps confusing users.

Quoted 4 lines

From an application standpoint, it’s not confusing since there is no etnaviv device to interact with. The user might wonder about the kernel messages, but that’s actually caused by an incorrect device tree. If the SoC doesn’t have an NPU, it shouldn’t be enabled in the DTS.

You have a point there, but as I mentioned above, that sounds tricky to
do: I have two boards that looks otherwise identical, but one has an A527, the other an T527. And still both don't have the NPU, since only some T527s feature it. So putting this on the user to use the right DT (or U-Boot defconfig) does not sound very nice.

And in contrast to many other devices described in DTs, we *can* safely detect the existence of this NPU: each of the SoCs have all the clock gates and resets, and accesses to the MMIO frame do not fault - and the kernel code apparently can cope with this situation already. So yeah, we could smear something into U-Boot, to put a status = "disabled"; in there, but I would like to avoid that, especially if the kernel is almost there already.

Cheers,
Andre

#4 lore raw
In reply to Andre Przywara

Hi Andre,

Am Donnerstag, dem 04.09.2025 um 00:29 +0100 schrieb Andre Przywara:

Quoted 29 lines

Hi,

the Allwinner A523/A527/T527 family of SoCs feature a Vivante "VIP9000"(?) NPU, though it seems to be disabled on many SKUs. See https://linux-sunxi.org/A523#Family_of_sun55iw3 for a table, the row labelled "NPU" indicates which model has the IP. We suspect it's all the same die, with the NPU selectively fused off on some packages.

Board vendors seem to use multiple SKUs of the SoC on the same board, so it's hard to say which particular board has the NPU or not. We figured that on unsupported SoCs all the NPU registers read as 0, though, so were wondering if that could be considered as a bail-out check for the driver?
At the moment I get this, on a SoC with a disabled NPU:

[    1.677612] etnaviv etnaviv: bound 7122000.npu (ops gpu_ops)
[    1.683849] etnaviv-gpu 7122000.npu: model: GC0, revision: 0
[    1.690020] etnaviv-gpu 7122000.npu: Unknown GPU model
[    1.696145] [drm] Initialized etnaviv 1.4.0 for etnaviv on minor 0
[    1.953053] etnaviv-gpu 7122000.npu: GPU not yet idle, mask: 0x00000000

Chen-Yu got this on his board featuring the NPU:

    etnaviv-gpu 7122000.npu: model: GC9000, revision: 9003

If I get the code correctly, then etnaviv_gpu_init() correctly detects the "unsupported" GPU model, and returns -ENXIO, but load_gpu() in etnaviv_drv.c then somewhat ignores this, since it keeps looking for more GPUs, and fails to notice that *none* showed up:
/sys/kernel/debug/dri/etnaviv/gpu is empty in my case.

Quoted 6 lines

Quick questions:
- Is reading 0 from VIVS_HI_CHIP_IDENTITY (or any other of the ID registers) an invalid ID, so we can use that to detect those disabled NPUs? If not, can any other register used to check this? The whole block seems to be RAZ/WI when the NPU is disabled.

Yes, 0 is not a valid core ID.

Quoted 5 lines

- Would it be acceptable to change the logic to error out of the driver's init or probe routine when no GPU/NPU has been found, at best with a proper error message? As it stands at the moment, the driver is loaded, but of course nothing is usable, so it keeps confusing users.

Yes, I think it would be fine to not expose the DRM device at all if no usable GPU core has been found.

However, I think it would still be good if your bootloader would fix up the DT to avoid probing the driver at all if possible. It may work on your system but access to a disabled GPU MMIO is not RAZ/WI everywhere.

On i.MX8M* we have a similar situation with disabled peripherals. There the bootloader can detect which peripherals are active by looking at the SoC fuses and fixes up the DT status property accordingly. Not sure if you may have something similar on the Allwinner.

Regards,
Lucas

#5 lore raw
Quoted 32 lines · On Wed, 10 Sep 2025 09:28:19 +0200 Lucas Stach <[email protected]> wrote:

Hi Andre,

Am Donnerstag, dem 04.09.2025 um 00:29 +0100 schrieb Andre Przywara:

Quoted 29 lines

Hi,

the Allwinner A523/A527/T527 family of SoCs feature a Vivante "VIP9000"(?) NPU, though it seems to be disabled on many SKUs. See https://linux-sunxi.org/A523#Family_of_sun55iw3 for a table, the row labelled "NPU" indicates which model has the IP. We suspect it's all the same die, with the NPU selectively fused off on some packages.

Board vendors seem to use multiple SKUs of the SoC on the same board, so it's hard to say which particular board has the NPU or not. We figured that on unsupported SoCs all the NPU registers read as 0, though, so were wondering if that could be considered as a bail-out check for the driver?
At the moment I get this, on a SoC with a disabled NPU:

[    1.677612] etnaviv etnaviv: bound 7122000.npu (ops gpu_ops)
[    1.683849] etnaviv-gpu 7122000.npu: model: GC0, revision: 0
[    1.690020] etnaviv-gpu 7122000.npu: Unknown GPU model
[    1.696145] [drm] Initialized etnaviv 1.4.0 for etnaviv on minor 0
[    1.953053] etnaviv-gpu 7122000.npu: GPU not yet idle, mask: 0x00000000

Chen-Yu got this on his board featuring the NPU:

    etnaviv-gpu 7122000.npu: model: GC9000, revision: 9003

If I get the code correctly, then etnaviv_gpu_init() correctly detects the "unsupported" GPU model, and returns -ENXIO, but load_gpu() in etnaviv_drv.c then somewhat ignores this, since it keeps looking for more GPUs, and fails to notice that *none* showed up:
/sys/kernel/debug/dri/etnaviv/gpu is empty in my case.

Quoted 7 lines
Quoted 6 lines

Quick questions:
- Is reading 0 from VIVS_HI_CHIP_IDENTITY (or any other of the ID registers) an invalid ID, so we can use that to detect those disabled NPUs? If not, can any other register used to check this? The whole block seems to be RAZ/WI when the NPU is disabled.

Yes, 0 is not a valid core ID.

Ah, thanks, that helps to identify the situation (see below).

Quoted 6 lines
Quoted 5 lines

- Would it be acceptable to change the logic to error out of the driver's init or probe routine when no GPU/NPU has been found, at best with a proper error message? As it stands at the moment, the driver is loaded, but of course nothing is usable, so it keeps confusing users.

Quoted 6 lines

Yes, I think it would be fine to not expose the DRM device at all if no usable GPU core has been found.

However, I think it would still be good if your bootloader would fix up the DT to avoid probing the driver at all if possible. It may work on your system but access to a disabled GPU MMIO is not RAZ/WI everywhere.

Yes, I agree, Allwinner is "nice" by using RAZ/WI if possible, but indeed most other platforms SERROR out when touching an address without a device.

Quoted 4 lines

On i.MX8M* we have a similar situation with disabled peripherals. There the bootloader can detect which peripherals are active by looking at the SoC fuses and fixes up the DT status property accordingly. Not sure if you may have something similar on the Allwinner.

We have code for DT patching in U-Boot, and use that already for some Allwinner boards. I don't know if we can identify the actual fuse used. We could use the SoC package identifier (first word of the SID eFUSE device, cf. the second row in here [1]), but that would require maintenance, as new IDs will show up, and we don't know every single one even now. Another method would be to just ungate the clock and de-assert the reset, then read the NPU ID register (0x18?) to see if it's zero or not. Probably the most robust, but involves a bit more code, as we probably would need some pseudo-device to fetch those reset/clock details from the DT.

I guess the current situation works as well, for now, but I will look at a U-Boot patch, just wanted to check if we can avoid that.

Cheers,
Andre