Handle Vector types in G_FADD using G_FPEXT #557
Closed
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
1. Use a custom legalizer for bf16(only works for smaller than 32 elements) 2. Use underlying G_FPEXT to first convert bf16 -> f32 then perform add. G_FADD(G_FPEXT(V1), G_FPEXT(V2)) Convert back to original types and shapes if needed. 3. Can see fully vectorized code which wasn't seen before.
Break <64xbf16> into 2 chunks of <32xbf16>.
Pending check: Not sure about how to verify pad and unpad logic, seems it's unrolling into a lot of boilerplate code.
Sameeranjoshi
requested review from
F-Stuckmann,
SagarMaheshwari99,
abhinay-anubola,
abnikant,
andcarminati,
katerynamuts,
khallouh,
konstantinschwarz,
martien-de-jong,
niwinanto and
stephenneuendorffer
as code owners
Less than Vectors of f32 = 16, 32 gets converted to 64xf32 as those are legal. Vectors of bf16 = 32xbf16 is a custom case converts other sizes into this corresponding vector.
Sameeranjoshi
added a commit
to Sameeranjoshi/iree-amd-aie
that referenced
this pull request
Jul 21, 2025
Patch adds tests to make sure all the tests with `vector.multi_reduction` generate successfully pass Peano legalizer and generate efficient vectorized code. This patch checks only the IREE side to keep the dependency minimun on Peano. (Depends on Peano: 1. Xilinx/llvm-aie#548 2. Xilinx/llvm-aie#557 ) 1. `reassociateFpReductions=true` is must else code is scalarized. This flag could be added into the IREE vectorization pipeline to trigger automatically. 2. bf16/i32/f32 all types with different sizes work now.
Sameeranjoshi
added a commit
to Sameeranjoshi/iree-amd-aie
that referenced
this pull request
Jul 21, 2025
Patch adds tests to make sure all the tests with `vector.multi_reduction` generate successfully pass Peano legalizer and generate efficient vectorized code. This patch checks only the IREE side to keep the dependency minimun on Peano. (Depends on Peano: 1. Xilinx/llvm-aie#548 2. Xilinx/llvm-aie#557 ) 1. `reassociateFpReductions=true` is must else code is scalarized. This flag could be added into the IREE vectorization pipeline to trigger automatically. 2. bf16/i32/f32 all types with different sizes work now.
Sameeranjoshi
added a commit
to Sameeranjoshi/llvm-aie
that referenced
this pull request
Jul 22, 2025
…d code. This patch is dependent on Xilinx#548 and Xilinx#557. Previously bf16 and f32 failed to generate fully vectorized code and used to scalarize, this test makes sures different types and vector sizes work and are fully vectorized. This is a supplementary patch for verifying below pipeline: Part 1: `vector.multi_reduction` to `vector.reduction` to `llvm.vector.reduce.fadd.*` nod-ai/iree-amd-aie#1336 Part 2: Further lowers to AIE2P instructions.(This patch)
| @@ -0,0 +1,56 @@ | |||
| ; RUN: llc -mtriple=aie2p -O0 -stop-after=legalizer %s -o - 2>&1 | FileCheck %s | |||
There was a problem hiding this comment.
For the tests, it is better to run just the pass of interest. For example, you can create an MIR test including just the ilegal type operation and run llc with -run-pass=legalizer. In this way we can easily spot the specific legalization change in action.
There was a problem hiding this comment.
The good part is that you can use llvm/utils/update_mir_test_checks.py to update the tests.
|
Squashed into #604 |
mgehre-amd
pushed a commit
that referenced
this pull request
Aug 21, 2025
[AutoBump] Merge with fixes of 8388040 (Jan 23) (19)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
Depends on #548