You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2021/05/25 18:08:11 UTC

[GitHub] [tvm] tkonolige opened a new pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

tkonolige opened a new pull request #8128:
URL: https://github.com/apache/tvm/pull/8128


   We handle lowering rank 0 tensors by lowering them to rank 1 buffers with a single element.
   
   @junrushao1994 @Hzfengsy @vinx13 @MasterJH5574 @tqchen 
   


-- 
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



[GitHub] [tvm] junrushao1994 commented on pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
junrushao1994 commented on pull request #8128:
URL: https://github.com/apache/tvm/pull/8128#issuecomment-848107931


   @Hzfengsy would you like to take a look at this corner case? Basically it is due to the fact that 0-dimension tensor in Relay is lowered to a block with 0 IterVars, which causes ambiguity with the "opaque" block.


-- 
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



[GitHub] [tvm] tqchen edited a comment on pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
tqchen edited a comment on pull request #8128:
URL: https://github.com/apache/tvm/pull/8128#issuecomment-848218371






-- 
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



[GitHub] [tvm] junrushao1994 commented on pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
junrushao1994 commented on pull request #8128:
URL: https://github.com/apache/tvm/pull/8128#issuecomment-848192731


   @tqchen The main problem is the ambiguity introduced by the definition of an opaque block. An alternative approach is that we just add a dummy iter-var in the block with unit extent. What do you think?


-- 
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



[GitHub] [tvm] tqchen edited a comment on pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
tqchen edited a comment on pull request #8128:
URL: https://github.com/apache/tvm/pull/8128#issuecomment-848190602


   Perhaps we should think about proper rank 0 buffer support instead to change to rank 1. 
   
   We already support such rank-0 buffer through out the previous TE pipeline, and would be great to confirm and make things work through the current pipeline as well


-- 
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



[GitHub] [tvm] tqchen commented on pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
tqchen commented on pull request #8128:
URL: https://github.com/apache/tvm/pull/8128#issuecomment-848218371


   I still think it would be great to have first-class 0-rank buffer and compute support, so we do not need to do special casing to handle the items. Such first-class support will bring a lot of benefit later, for example, in the cases such as adding a batching to the entire loop computations


-- 
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



[GitHub] [tvm] tqchen commented on pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
tqchen commented on pull request #8128:
URL: https://github.com/apache/tvm/pull/8128#issuecomment-848190602


   Perhaps we should think about proper rank 0 buffer support instead to change to rank 1


-- 
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



[GitHub] [tvm] junrushao1994 merged pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
junrushao1994 merged pull request #8128:
URL: https://github.com/apache/tvm/pull/8128


   


-- 
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



[GitHub] [tvm] tqchen commented on pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
tqchen commented on pull request #8128:
URL: https://github.com/apache/tvm/pull/8128#issuecomment-854711103


   Suggest close this for now and instead move to a tracking issue that tracks the 0-rank support in various places of tensorIR.


-- 
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



[GitHub] [tvm] junrushao1994 commented on pull request #8128: [TE/TIR] Fix create_prim_func to properly handle rank 0 tensors.

Posted by GitBox <gi...@apache.org>.
junrushao1994 commented on pull request #8128:
URL: https://github.com/apache/tvm/pull/8128#issuecomment-854831387


   Given this PR does help some workloads (some subgraphs in BERT) temporarily, and doesn't introduce new problems, let's merge it in and start a tracking issue


-- 
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