J784S4XEVM: Custom Operators with int32 Input Fail to Compile [SDK 11.2]

Mahmoud Hamed

Part Number: J784S4XEVM

I am working with ti-processor-sdk-rtos-j784s4-evm-11_02_00_06 and running into a compilation issue while compiling an ONNX model that contains custom operators whose first input is of type int32. In this case, compilation fails with an unsupported data type error, which appears to indicate that int32 may not be supported for custom operators.

========================== [Memory Planning Started] ==========================


------------------------- Network Compiler Traces ------------------------------
Wrong data type for layer input 9, 30
Error : Error Code = <ERR_UNSUPPORTED_DATA_TYPE>
[TIDL Import]  FATAL ERROR: Network Compiler failed to execute - Memory planning failed with return code - 65280 -- [tidl_import_core.cpp, 1068]
[TIDL Import] Aborting

One example scenario is when the TopK indices output is fed into a custom Tile operator. A model reproducing this scenario, along with the generated artifacts is attached. topk_tile.zip

Please let us know whether this case is a known limitation in SDK 11.2. If so, is support for it expected in a future SDK 11.2 release?

5 months ago

0 Mahmoud Hamed 5 months ago

Prodigy 60 points

Hello Brijesh Jadav,

Could you please support us on this topic?

Regards,

Mahmoud

0 Brijesh Jadav 5 months ago in reply to Mahmoud Hamed

TI__Guru**** 501675 points

Hi Mohmoud,

Let me check and get back to you in few days.

Regards,

Brijesh

0 Vaibhav Kumar 5 months ago in reply to Mahmoud Hamed

TI__Intellectual 2160 points

Hi,
As per TIDL team, 32 bit inputs are not supported for custom layers. There are plans for future but no confirmation.
Regards

0 Vaibhav Kumar 4 months ago in reply to Vaibhav Kumar

TI__Intellectual 2160 points

Hi, would you able to share your custom implementation files for this model ?

Regards

0 Ghassen souissi 4 months ago in reply to Vaibhav Kumar

Intellectual 320 points

Hello Vaibhav ,

Thank you for your reply.

To avoid sharing sensitive details in the public thread, we reproduced the same scenario using a custom Cast operator instead. You can find the model, associated artifacts, and the custom implementation here:

custom_cast.zip topk_cast_model.zip

For this scenario, we are testing a potential workaround to proceed with the model. However, we are encountering the same limitation in another case.

When using a test model that includes a Cast node (custom operator) with an int64 input, the issue persists. From the visualization, it appears that the int64 inputs are automatically quantized to int32, which leads eventually to an error during model compilation:

I have attached the model used to reproduce this problem: cast_model.zip

Could you please take a look and let us know if there is a recommended workaround or best practice to handle these 2 issues?

Best regards

Ghassen

0 Vaibhav Kumar 4 months ago in reply to Ghassen souissi

TI__Intellectual 2160 points

Hi, if the the compilation is not working in this configuration then which configuration is the svg visualisation you attached from ?

Regards

0 Ghassen souissi 4 months ago in reply to Vaibhav Kumar

Intellectual 320 points

Hello Vaibhav,

These artifacts are generated during the compilation process before the failure.

Regards

Ghassen

Processors

Processors forum

J784S4XEVM: Custom Operators with int32 Input Fail to Compile [SDK 11.2]