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 2022/07/13 00:05:05 UTC
[GitHub] [tvm] tkonolige opened a new pull request, #12079: [AUTOSCHEDULER,FIX] Calculate arithmetic intensity without log scale
tkonolige opened a new pull request, #12079:
URL: https://github.com/apache/tvm/pull/12079
In autoscheduler's featurization, arithmetic intensity was incorrectly calculated as log(FLOPs) / log(bytes). This change removes the logs so arithmetic intensity is FLOPs / bytes.
@junrushao1994 @merrymercy @octoJon
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] octoJon commented on pull request #12079: [AUTOSCHEDULER,FIX] Calculate arithmetic intensity without log scale
Posted by GitBox <gi...@apache.org>.
octoJon commented on PR #12079:
URL: https://github.com/apache/tvm/pull/12079#issuecomment-1183705384
> Pretty sure @octoJon just needs to subtract the two values to get the feature they want.
Right, but the issue is that I can't compute log(A/B) --which is what I want -- from log(A)/log(B) -- which is what we have before Tristan's change. I will be able to get it from log2p(A/B), so I am good with this change.
A different option would have been to change the logic lower down in the file to essentially replace all the occurrences of log(FLOPs) / log(bytes) with log(FLOPs) - log(bytes), but Tristan's version is simpler and makes it less likely that we'd screw up the interpolation when attempting to make that change.
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] tkonolige commented on pull request #12079: [AUTOSCHEDULER,FIX] Calculate arithmetic intensity without log scale
Posted by GitBox <gi...@apache.org>.
tkonolige commented on PR #12079:
URL: https://github.com/apache/tvm/pull/12079#issuecomment-1183569122
Is the scale of values in arithmetic intensity so large that it needs to be log scaled? And even if it needs to be log scaled, then shouldn't the log scaling be applied after the arithmetic intensity is computed and not before?
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] merrymercy commented on pull request #12079: [AUTOSCHEDULER,FIX] Calculate arithmetic intensity without log scale
Posted by GitBox <gi...@apache.org>.
merrymercy commented on PR #12079:
URL: https://github.com/apache/tvm/pull/12079#issuecomment-1183565979
I use `log` to normalize these values so the machine learning-based cost model can deal with these feature vectors better.
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] AndrewZhaoLuo commented on pull request #12079: [AUTOSCHEDULER,FIX] Calculate arithmetic intensity without log scale
Posted by GitBox <gi...@apache.org>.
AndrewZhaoLuo commented on PR #12079:
URL: https://github.com/apache/tvm/pull/12079#issuecomment-1183697645
Pretty sure @octoJon just needs to subtract the two values to get the feature they want.
As Log(A) - Log(B) = Log(A / B)
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] AndrewZhaoLuo merged pull request #12079: [AUTOSCHEDULER,FIX] Calculate arithmetic intensity without log scale
Posted by GitBox <gi...@apache.org>.
AndrewZhaoLuo merged PR #12079:
URL: https://github.com/apache/tvm/pull/12079
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] merrymercy commented on pull request #12079: [AUTOSCHEDULER,FIX] Calculate arithmetic intensity without log scale
Posted by GitBox <gi...@apache.org>.
merrymercy commented on PR #12079:
URL: https://github.com/apache/tvm/pull/12079#issuecomment-1183574411
The case is right, but we need to find a way to normalize the values.
Simply applying a log after arithmetic intensity is computed may not work, because the arithmetic intensity can be less than 1 in rare cases.
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] merrymercy commented on pull request #12079: [AUTOSCHEDULER,FIX] Calculate arithmetic intensity without log scale
Posted by GitBox <gi...@apache.org>.
merrymercy commented on PR #12079:
URL: https://github.com/apache/tvm/pull/12079#issuecomment-1183575932
We can use this shifted log https://github.com/apache/tvm/blob/c811fb85248550495ac27be88ed3e2fdd7c3bece/src/auto_scheduler/feature.cc#L1110
--
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: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org