[Codegen][GPU] Let integer range optimization narrow GPU computations to i32 #19473
+354
−72
Conversation
This file contains 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
benvanik
reviewed
Dec 12, 2024
Comment on lines
+72
to
+73
| "Flag indicating if computations that can be performed with 32 bits shuld be." | ||
| " Mainly used for GPU code generation to not waste registers"> |
There was a problem hiding this comment.
style nit: needs a terminating .
krzysz00
force-pushed
the
index-narrowing
branch
from
d9af342 to
562f3ba
Compare
…ree-org#19361) This reverts commit cb5be1d. Compaled to the previous revision, this one works around a correctness bug in dataflow analysis that's being fixed by removing the analysis after SCF->CF. --- First, this patch implements InferIntRangeInterface for hal.interface.workgroup.{size,id,count} using a local upper_bound attribute. Then, it adds a -iree-codegen-gpu-propagate-dispatch-size-bounds pass that adds these upper_bounds identifiers to the interface.workgroup operations and to gpu.thread_id based on static information available late in the codegen pipeline. Then, it uses -optimize-int-arithmetic to optimize indexing after -lower-affine, getting rid of a bunch of "if the input's negative" logic that isn't actually needed in many of our kernels. It also ensures that these upper_bound values propagate to LLVM.
krzysz00
force-pushed
the
index-narrowing
branch
from
562f3ba to
4dee634
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Note: This PR is stacked on top of #19372, and so looks bigger than it is. The relevant changes are in the last commit.
Add an option to -iree-util-optimize-int-arithmetic to have it perform computations in i32 where possible, which is enabled when optimizing arithmetic for GPU codegen. This allows LLVM co correctly conclude that various computations don't need to be done at full 64-bit precision, thus saving registers and instructions. (LLVM has some rewrites for this, but they're, for example, gated on only having one use of the potentially-truncated value, which means that shared math stays in an over-wide data type).