ashutosh-arm opened a new issue, #16301:
URL: https://github.com/apache/tvm/issues/16301

   #### Context
   At present, CMSIS-NN pass `extract_constant()` kicks out tensor constants 
into the main() function. This was needed as TIR didn't support buffer 
allocation for constants. With the new AllocateConst API in TIR, tensor 
constants can reside in the composite functions and can still be translated 
into TIR.
   
   #### Desired changes
   1. Example of AllocateConsts usage: 
https://github.com/apache/tvm/blob/506eff23b034cfd88dda44cf5040ab8b40683155/src/relay/backend/contrib/cmsisnn/relay_to_tir.cc#L125
   2. Remove the pass `extract_constants()` 
https://github.com/apache/tvm/blob/506eff23b034cfd88dda44cf5040ab8b40683155/src/relay/backend/contrib/cmsisnn/extract_constants.cc#L241
   3. When the pass is removed, RelayToTIR would require changes to generate 
AllocateConsts corresponding to the constants existing within the partitioned 
function. Pre-requisite is to identify the constants within the emitted 
operator.
   4. Remove following set of tests: 
https://github.com/apache/tvm/blob/main/tests/python/contrib/test_cmsisnn/test_extract_constants.py
   
   
   ### Environment
   
   TVM 0.14.0
   
   ### Steps to reproduce
   
   Any test from 
https://github.com/apache/tvm/blob/506eff23b034cfd88dda44cf5040ab8b40683155/tests/python/contrib/test_cmsisnn/test_conv2d.py#L241
   
   
   


-- 
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.

To unsubscribe, e-mail: [email protected]

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

Reply via email to