AI Engine Experiments

AMD AI Engine development usually involves three pieces:

1. Kernel code (C++)
2. Graph code (C++ or Python)
3. Host code (C++ or Python)

For kernels, I have explored the lower-level side of the stack:

• Intrinsics: direct programming of AIE vector/MAC operations
• APIs: C++ wrappers around intrinsics from AMD

For graph design, the options are:

• Vitis graph flow in C++
• MATLAB/Simulink-based flow
• IRON in Python

Early work: intrinsics + MATLAB

My first AIE experiments started with intrinsics + MATLAB in this repo: amd_aie_matlab.

This repo includes small matrix/vector experiments such as gemv, gemv_opt, and gemv_opt_i8, along with MATLAB/Simulink files and AIE kernel code. The flow was simple: generate test data, run the model, and compare the simulated output against a golden result.

That work later evolved into this repository where some undegrad students work on optimizing intrinsics for GEMV and GEMM: aie-intrinsics-nn. This effort moved toward a more structured intrinsics-based workflow for small neural-network style building blocks. The focus was on reusable kernels and a clearer path from GemV-style experiments toward denser NN-style compositions.

Lightweight backend work for hls4ml

The next step was this repo: hls4ml-backend, for mapping small neural-network style computations onto AIE as an effort to create a backend for the popular python based tool hls4ml. It includes quickstart and run_workload.py, serving as a sandbox for experimenting with how small physics NN layers could be laid out across AIE resources.

Teaching material

I also worked as a Teaching Assistant (TA) for the AI Engine portion of Ryan’s CSE237C course. The public assignment and starter material are here:

• Assignment page
• Starter files

This material introduces the AIE programming model through:

• host / graph / kernel separation
• IRON-based Python graph construction
• tiled matrix multiplication
• single-core and whole-array examples
• small NN-style exercises built from these pieces

The assignment code includes files like basic_mm.py, matmul.cc, passthrough.py, and nn.py.