[JIT][fuser] Improve precision of emitted code for prim::Constant #18817
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
facebook-github-bot
added
the
oncall: jit
jamesr66a
force-pushed
the
double_const
branch
7 times, most recently
from
2c7d20b to
07a25e8
Compare
zdevito
approved these changes
Apr 4, 2019
torch/csrc/jit/fuser/codegen.cpp
Outdated
There was a problem hiding this comment.
variableType is wrong. It should return 'int64_t' for IntType and 'double' for FloatType. There should be no need for the if statement here.
There was a problem hiding this comment.
Somebody made it so that the interface to fused kernels that take scalar floating point arguments is float, probably because the fusion compiler is biased toward GPU code. I can change that interface, but it would involve mucking with the GPU fuser
jamesr66a
force-pushed
the
double_const
branch
7 times, most recently
from
e272678 to
12846be
Compare
There was a problem hiding this comment.
@jamesr66a has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
|
@jamesr66a merged this pull request in 9b69f21. |
zhangguanheng66
pushed a commit
to zhangguanheng66/pytorch
that referenced
this pull request
May 6, 2019
Summary: Stacked on pytorch#18815 and pytorch#18811. This makes it so that we emit a higher-precision literal for float values in the fusion kernel, as well as assign that to a `double` variable. This prevents us from losing precision for values such as `pi`, but with the previous fixes this will also get downcasted to `float` if downstream operations require it. Therefore, we should not lose performance because of implicit promotions Pull Request resolved: pytorch#18817 Differential Revision: D14820842 Pulled By: jamesr66a fbshipit-source-id: 519671c6ca5e7adac746a4c4c72760a6d91e332f
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stacked on #18815 and #18811.
This makes it so that we emit a higher-precision literal for float values in the fusion kernel, as well as assign that to a
doublevariable. This prevents us from losing precision for values such aspi, but with the previous fixes this will also get downcasted tofloatif downstream operations require it. Therefore, we should not lose performance because of implicit promotions