[NNC] fix support for FP16 in CudaCodgen #44209
Conversation
facebook-github-bot
added
the
oncall: jit
💊 CI failures summary and remediationsAs of commit 4409f54 (more details on the Dr. CI page):
ci.pytorch.org: 1 failedThis comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions on the GitHub issue tracker or post in the (internal) Dr. CI Users group. This comment has been revised 6 times. |
bertmaher
left a comment
bertmaher
left a comment
There was a problem hiding this comment.
I'm still a little nervous here because the HalfChecker mutation is going to turn loads of halfs into floats, but the expression tree has been constructed assuming all the arithmetic is half-typed. So this probably ends up working syntactically, but the types of the IR are potentially out of sync. I'm not really sure anything will break as a result of this though, so maybe we can push it on to a TODO stack.
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@nickgg has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
Codecov Report
@@ Coverage Diff @@
## master #44209 +/- ##
==========================================
- Coverage 69.24% 69.24% -0.01%
==========================================
Files 381 381
Lines 47573 47573
==========================================
- Hits 32943 32942 -1
- Misses 14630 14631 +1
Continue to review full report at Codecov.
|
4 participants
Fixes a bug where FP16 values could be incorrectly cast to a half type that doesn't have a cast operator by inserting the cuda specific cast to float during handling of the Cast node, not as a wrapper around printing Loads and Stores. Two main changes: the HalfChecker now inserts the casts to float explicitly in the IR, and the PrioritizeLoad mutator now consumes both Loads and a Cast which immediately preceded a load.
Tested with test_jit_fuser_te.py and test_tensorexpr.py, plus C++ tests obv.