FrozenGene commented on issue #4828: [QNN][TFLite] TFLite rounding mode support
URL: https://github.com/apache/incubator-tvm/pull/4828#issuecomment-583285949
 
 
   > I thought little more about the bit exact problem. One source of 
discrepancy for certain is QNN add, and QNN concatenate ops. These call 
Requantize internally, and they will have default rounding in C++ (UPWARD). For 
testing, @LiangHao151941 , you can set the C++ default to TFLIte to see if it 
helps. Meanwhile, we can also think how to can make that C++ rounding visible 
at python level.
   
   Maybe we could use `@register_func` to register one function named as 
`@register_func(qnn.requantize.rounding)` in TFLite parser, then we could get 
the value in `RequantizeQnnCanonicalize` from C++. However, this will have one 
problem that if users specify the rounding value, it won't work, because we 
don't know whether the rounding value is setting by default or by user 
explicitly. So, if we support it this, we will constraint we will only one 
rounding setting by TFLite parser.
   
   Another way is we provide `rounding` args for `qnn.add / qnn.mul / 
qnn.conconcate`, because they use `requantize` in fact too, so they need 
`rounding`.

----------------------------------------------------------------
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]


With regards,
Apache Git Services

Reply via email to