Internal DLA error for Conv2d layer

[howkong]

I am running an onnx model exported from PyTorch on my Jetson AGX Xavier, using trtexec.
Uploading: capsnet.onnx…
However, trtexec complains about Internal DLA error for layer Conv_x for my Conv_2~Conv_9 layers.
I have cheked the guide here, and I think my Conv Layers should fit the requirements.

So how can I debug this?

[AastaLLL]

Hi,

We have verified the model shared in the below topic:
Tensorrt complains that DLA does not support dynamic shapes, even when I don't have any dynamic axes.

It works good but just some layers cannot be deployed on the DLA.

Thanks.

[howkong]

The vast majority of the layers cannot run on DLA, and running on DLA is what I am trying to do.
I am trying to eliminate the reasons why these layers failed to run on DLA, and I need some help on that.
Should I repost this question with the allowGPUFallback option disabled?

[AastaLLL]

Hi,

Please noted that there is a new JetPack release for Jetson this month:

Jetson Software Roadmap for 2H-2021 and 2022

NVIDIA continuously invests in JetPack and regularly ships software updates to enable performance improvements, new capabilities, and support for new modules and developer kits. In this announcement, we provide details about the current software roadmap for 2H-2021 and 2022. Note that this plan is subject to change. JetPack 4.6 is our next production release, currently scheduled for July 2021. It will bring updates to the CUDA, cuDNN, and TensorRT compute stack as well as a feature-rich BSP upd…

It will contain a new DLA library which adds some extra support to the operation.
Would you mind waiting for it and test your model with the new DLA library?

Thanks.