excute dnnl provider error

[jia-jidong]

I built onnxruntime use the commond:
./build.sh --config Release --build_shared_lib --use_dnnl

Then I code my app using dnnl provider, but I got the following error message:

2022-06-22 10:42:43.619136547 [E:onnxruntime:, sequential_executor.cc:368 Execute] Non-zero status code returned while running DNNL_12891192397758778371_0 node. Name:'DnnlExecutionProvider_DNNL_12891192397758778371_0_0' Status Message: could not construct a memory descriptor using a format tag
Non-zero status code returned while running DNNL_12891192397758778371_0 node. Name:'DnnlExecutionProvider_DNNL_12891192397758778371_0_0' Status Message: could not construct a memory descriptor using a format tag
debug-run
Segmentation fault (core dumped)

If use cpu provider everything is OK, but I want to use dnnl provider!

System information

• OS Platform and Distribution :Ubuntu 9.3.0-17ubuntu1~20.04
• ONNX Runtime installed from: source
• ONNX Runtime version: 1.11.1
• Python version:3.8.3
• Visual Studio version (if applicable): n/a
• GCC/Compiler version (if compiling from source): 7.5
• CUDA/cuDNN version: n/a
• GPU model and memory: n/a

To Reproduce
Run the model using dnnl provider.
Model download url: https://zenodo.org/record/6221127#.YrKJW3VByUk

Expected behavior
No error report using dnnl provider.

[faxu]

CC @jywu-msft

[jywu-msft]

@georgen117 , is this a known issue?

[georgen117]

No this is not a known issue. Thanks for bringing it to my attention. We will investigate.

[georgen117]

@jia-jidong
I ran the model with random (noise) input and it ran without any issues on my system.

You could share your input file or help with some debugging yourself.

a few things you can do that will give me more information.
set the following environment variables.

ORT_ENNL_DEBUG_LOG=1
ONEDNN_VERBOSE=2 (in older code this was DNNL_VERBOSE=2)

Share the output log.

Optional since I have generated this on my system.
ORT_DNNL_DUMP_SUBGRAPHS=1 <This will dump each subgraph generated by the dnnl ep on my system this would produce ~60 separate *.onnx files>

Find the subgraph matching the error DnnlExecutionProvider_DNNL_12891192397758778371_0_0.onnx and add it to the ticket. This will at least narrow down the error to a subset of the model.