How arc-research-lab/aim works
Project Type and Purpose
• This is an FPGA-centric heterogeneous computing research framework (“AIM”) for accelerating arbitrary-precision integer multiplication on AMD Versal ACAP.
• It comprises:
• A code‐generation/template engine (in “templete/”) that emits host, AIE graph, and PL kernel code based on user configs.
• Several reference applications under “application/” (IntegerMultiplication, RSA, Mandelbrot) demonstrating the generated accelerator.
• Build scripts (Makefile, build.sh), configuration files (.cfg), and Python tools (AIM.py, checkfreq.py, gen_array.py) to automate generation, synthesis, and frequency‐tuning.
File-Structure Analysis
• Top‐level:
• .github/workflows – CI for traffic plotting
• application/… – three sample apps, each with:
– src/: AIE kernels (aie_aim_*), graph (aie_graph.cpp/h), PL kernel, host.cpp
– tools/: frequency checker, array generator, Jinja templates
– system.cfg, Makefile, utils.mk
• templete/: the master template project with identical structure to application/, plus AIM.conf, AIM.py, build.sh
• assets/, images/: documentation assets
• Architectural pattern:
• Host (ARM CPU in SoC) ↔ AIE array (vector compute engines) ↔ PL fabric (fine-grain carry‐prop)
• Code generation (Jinja templates) for parameterized bitwidths and resource constraints
• Build pipeline uses Vitis/Vitis HLS toolchain and PetaLinux SDK
Key Architectural Insights from README
• AIM divides large‐integer multiply into many 32-bit chunks computed on AIE array; carry propagation offloaded to PL.
• Three phases: code generation (Python + Jinja) → host + AIE + PL synthesis → runtime execution.
• Users specify bitwidth and resource constraints in AIM.conf; template scripts handle timing/frequency adjustments.