[Research][Autocomp] Automated Ukernel Synthesis Agent

[copparihollmann]

Context

Writing optimized micro-kernels for new architectures (e.g., SpacemiT X60, RISC-V V-Extension) is labor-intensive. Leveraging the "Autocomp" paper (UC Berkeley), we will build an agent that synthesizes verified micro-kernels by searching over an "Optimization Menu" (Unroll, Vectorize, Reorder).

Objective

Implement the Autocomp Beam Search agent (src/tools/autocomp/) to generate high-performance MLIR/C micro-kernels.

Scope of Work

1. Optimization Menu: Define a structured list of transformations (e.g., Loop Unroll, Accumulator Split, Vectorize) applicable to our target.
2. Planner & Coder:
   • Phase 1 (Plan): LLM selects a sequence of menu items.
   • Phase 2 (Implement): LLM generates the specific C/Intrinsic code.
3. Verification Loop:
   • Compiles candidate kernels using builder.py --profile riscv_spacemit.
   • Runs correctness checks on QEMU/Hardware.
4. Feedback: Use cycle counts (from iree-benchmark-module) to guide the Beam Search.

Acceptance Criteria

• Test 1: Input a scalar matmul. Agent successfully applies "Vectorize" plan and produces valid RVV intrinsics.
• Test 2: Agent discovers a kernel for riscv_spacemit that outperforms the generic llvm-cpu lowering by >1.5x.