GEMM produces incorrect output for certain MNK tilings on AIE2 (XDNA1) with latest wheels

[Fangtangtang]

Hi, I'm encountering incorrect GEMM results when using the latest wheels build of mlir-aie targeting AIE2 (XDNA1). The same example and configurations worked correctly with release v1.0.

I'm following setup instructions in commit e952e58d8c1ec7c2759696318731527de8dcc0b5. I install mlir-aie with

python3 -m pip install mlir_aie==0.0.1.2025122304+e952e58 -f https://github.com/Xilinx/mlir-aie/releases/expanded_assets/latest-wheels-2/

and llvm-aie wheel with

python3 -m pip install llvm-aie -f https://github.com/Xilinx/llvm-aie/releases/expanded_assets/nightly

When running programming_examples/basic/matrix_multiplication/single_core on RyzenAI-npu1, I observe incorrect computation results for certain MNK tiling configurations.

For example:

• when m = 16, k = 64, n = 16, the output matrix is all zeros

the same example and tiling configurations produced correct results when I was using release v1.0. I'm trying to understand whether this is caused by a misconfiguration on my side or any recent compiler change.

Thanks in advance!

[hunhoffe]

Hello! I will look into this issue; as an interim solution, have you tried using a release version? Do you have issues with release wheels?

Installation instructions for the latest release are below:

# Get the latest release version
latest_tag_with_v=$(curl -s "https://api.github.com/repos/Xilinx/mlir-aie/releases/latest" | jq -r '.tag_name')
latest_tag="${latest_tag_with_v#v}"

# Install IRON library and mlir-aie from the latest stable release
python3 -m pip install mlir_aie==${latest_tag} -f https://github.com/Xilinx/mlir-aie/releases/expanded_assets/${latest_tag_with_v}

If using a release version does not give you a stable experience, please do let me know! In the meantime, I will also track down the cause of the issue you are experiencing.

[hunhoffe]

Debugging progress:

• I was able to replicate the issue
• I tried installing from 011d965 with the latest release, and still got the same error
• Can just change m = 16 and that is enough to break it

I'm still looking for the root cause and/or an interim workaround.