Assertion Error while building TensorRT engine for DLA

Hi All,

I get the following error while building TensorRT engine on DLA with INT8 precision. What does the error mean and how it can be resolved? When using FP16, the engine gets built without any issues on DLA.

[09/22/2020-13:14:52] [F] [TRT] Assertion failed: pitch == 1 && “requirements impossible to satisfy”

…/builder/tacticOptimizer.cpp:2356

Aborting…

Environment:
TensorRT version: 6.0.3-1
Cuda Version: 10.2
CUDNN Version: 7.6.6.33

Hi,

There is some limitation in the DLA+INT8 inference.
For example, deconv with stride != kernel is not supported.

Since you are using TensorRT 6.0, it’s recommended to upgrade your library to first.
Here is the support matrix of DLA for your reference:

Thanks.