Add fast-path for gufunc (specifically matmul) #9482
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
This reduces the GUfunc overhead of a simple matmul from ~60% of the operation to probably 10%, I expect around 1/3 of the time if the matmul core itself is well optimized. Mainly, also fixes a bunch of bugs, unfortunately... this is rather complex code so it is hard to review. (I am sure it fixes a lot of bugs more than it opens, but it's hard to be sure everything is covered...) It does have one larger change (also a bug fix) that I am aware of: The old code incorrectly broadcast _core_ dimensions, this is not allowed by NumPy unless the dimension is specified as `n|1`. For matmul this just changes the error if someone actually relied on it, they would have to switch to `|1` on newer versions.
Since simple `cupy.matmul` calls, half of the overhead is still inside the gufuncs even with the previous PR, this suggests adding a fast-path mechanism. I would love to avoid the fast-path here, but I doubt it is worthwhile to carefully Cythonize the gufuncs and a 2-3x speeup is just still a bit underwhelming. (Fast-paths are always an opportunity for bugs after all...)
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Outdated suggestions cannot be applied.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.
Since simple
cupy.matmulcalls, half of the overhead is still inside the gufuncs even with the previous PR, this suggests adding a fast-path mechanism.I would love to avoid the fast-path here, but I doubt it is worthwhile to carefully Cythonize the gufuncs and a 2-3x speeup is just still a bit underwhelming.
(Fast-paths are always an opportunity for bugs after all...)
The core is to propose a mechanism of:
for "GUFuncs".
Draft as based on gh-9481 and I assume we can put this since @emcastillo already had a look.whoops, seems I forgot the draft...This maybe closes gh-8191, from my timings we are still at about 2x slower than
torch, though. That is lost in various places, the biggest chunk is probably arollaxis(creating a new array).