[GitHub] [incubator-tvm] anijain2305 edited a comment on issue #4277: [ARM][Topi] Improving Int8 Perf in Spatial Conv2D schedule.

[GitBox <gi...@apache.org>]

anijain2305 edited a comment on issue #4277: [ARM][Topi] Improving Int8 Perf in Spatial Conv2D schedule.
URL: https://github.com/apache/incubator-tvm/pull/4277#issuecomment-552169896
 
 
   @jackwish Thanks for the comment.
   
   If we go legalize way, following kind of transformation would happen
   
   
   ~~~
   Original Relay graph --> conv2d(%int8_data, %int8_kernel)
   
   AfterLegalize Relay graph --> 
   
   %int16_data = cast(%int8_data, dtype="int16")
   %int16_kernel = cast(%int8_kernel, dtype="int16")
   conv2d(%int16_data, %int16_kernel)
   ~~~
   So, in this case, even if we want to use tensorize, we will only see int16 inputs in conv. Does that answer you question?
   
   I am sort of inclining towards going for schedule, because it is contained in one place for now. And one can easily change if one decides to use tensorize. Legalize might throw schedule developers off who want to see int8 inputs, and are surprised to see int16 inputs. 

----------------------------------------------------------------
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:
users@infra.apache.org


With regards,
Apache Git Services