tiovxdlcolorblend: needs to support int64

[aniket-l]

Some semantic segmentation models have output tensor data type as int64. Needs to be supported by colorblend.
Add it to the application/x-tensor-tiovx caps as enum number 8 and enable support in colorblend.

[m-herrera]

Hello @aniket-l we added support for this new data_type in the element in the branch feature/support-int64-in-tiovxdlcolorblend. From the Gstreamer side of things the element should now support int64; however, during our testing it seems that the vxCreateTensor does not support this format. We get the following error.

VX_ZONE_ERROR:[vxCreateTensor:262] Invalid data_type for the tensor.

We were also able to reproduce this with the modules by changing the data_type in the test from VX_TYPE_UINT8 to VX_TYPE_INT64, which produces the same error.

Please test the branch on your end with your use case to see if you come across the same issue and let us know how to proceed.

[m-herrera]

@aniket-l we tried this feature in the new SDK; we no longer get the invalid data_type error, however the pipeline gives a segmentation fault.

I also tried using VX_TYPE_INT64 on the modules, but I get the following error:

Running DL color-blend module test
MEM: ERROR: Alloc failed with status = 22 !!!
Incorrect data_type/bit-depth!
 
Incorrect data_type/bit-depth!
 
MEM: ERROR: Failed to free memory at virt addr = (nil) !!!
All tests complete!

I just wanted to know if there is anything I am missing from TI side to get it running on the modules. Then I will continue debugging the gstreamer side.