mbaret commented on pull request #6222:
URL: https://github.com/apache/incubator-tvm/pull/6222#issuecomment-672137931


   > It looks that this is a different from C source module and the json 
runtime module. Could you elaborate a bit why those two don't fit?
   
   The Ethos-N compiler compiles the graph down to a so-called 'command stream' 
which is a binary artifact that is directly consumed by the driver to execute 
the inference. It has encoded within it all the constants/weights. We're not 
calling out to a C library so the C source module wouldn't make sense here. 
Additionally, the JSON runtime wouldn't really add any useful functionality. We 
just want to load the artifact off the disk and wrap it in a packed function.
   
   > It seems that we are not really able to test them on HW in the CI. Is 
there any plan on this?
   
   We don't expect there to be HW in CI. CI will run SW testing including a 
mocked inference function so that we can test the compilation flow all the way 
through to the runtime. We will be running automated HW testing on a range of 
networks within Arm.
   
   To be explicit about the next 2 PRs, they are already written and only add 
more operators. Specifically, the next PR will add quantized conv2d and the 
final one will add quite a few more less complex operators. We wanted to split 
these up to make it easier to see how the codegen operates without the added 
complexity of all the other operators. The final PR will also include network 
tests for mobilenet, inceptionv3/4 and ssd mobilenet.


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to