Skip to content

Conversation

F-Stuckmann
Copy link
Collaborator

Remove Memory Size constraints for PostIncrement Selection in AIE2P.

There are no regressions in performance or PM in QoR.

@andcarminati
Copy link
Collaborator

Could we have a test?

@F-Stuckmann F-Stuckmann force-pushed the stuckmann.2d.3d.postinc.size branch 3 times, most recently from 28480a3 to a4b6e19 Compare September 4, 2025 08:29
@martien-de-jong
Copy link
Collaborator

martien-de-jong commented Sep 12, 2025

I'm curious as to the rationale. Does it block a useful optimization?
I can also imagine that it is a first step to share all G_AIE opcodes and bring more generic functionality to AIEBaseInstrInfo.cpp. Which I would applaud.

@@ -313,8 +311,6 @@ std::optional<unsigned> AIE2PInstrInfo::getCombinedPostIncOpcode(
case TargetOpcode::G_INTRINSIC:
switch (cast<GIntrinsic>(PostIncI).getIntrinsicID()) {
case Intrinsic::aie2p_add_2d:
if (Size >= 1024)
return {};
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also weird that this wasn't Size > 512

@F-Stuckmann F-Stuckmann force-pushed the stuckmann.2d.3d.postinc.size branch from a4b6e19 to 2cf21dd Compare September 15, 2025 12:12
@F-Stuckmann F-Stuckmann enabled auto-merge (rebase) September 15, 2025 12:12
@F-Stuckmann
Copy link
Collaborator Author

I'm curious as to the rationale. Does it block a useful optimization? I can also imagine that it is a first step to share all G_AIE opcodes and bring more generic functionality to AIEBaseInstrInfo.cpp. Which I would applaud.

GEMM now performs vloads of 2048 bits and thus this optimization enables postincrements on these loads.

@F-Stuckmann F-Stuckmann merged commit 7f4f732 into aie-public Sep 15, 2025
7 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants