TIRx Harness: An Open Compiler Harness for Agentic GPU Programming

TIRx Harness is a compiler-facing environment designed to make agent-driven GPU kernel development predictable, reusable, and measurable. It pairs a minimal, PTX-close intermediate representation with tooling that surfaces numerical, synchronization, and data-race diagnostics, a kernel zoo of over 60 concrete implementations, and remote benchmarking via KCoral to centralize and de-noise measurements. The design responds to practical failures observed when coding agents optimize kernels: long compilation paths that obscure semantics, noisy timing from shared GPUs, and repeated relearning of optimization patterns. By keeping the IR thin and exposing hardware capabilities directly, providing simulators and profilers for richer feedback, and preserving concrete, validated implementations, the harness shifts agent effort from debugging compiler uncertainty to exploring algorithmic and mapping choices.

On evaluated attention-family workloads, the harness-enabled agents produced sizable speedups: family-level geometric means ranged from 1.33× to 6.84×, with Kimi Delta Attention hitting 2.94× over FlashKDA (forward) and 6.84× over Flash Linear Attention (backward). Traces show concrete wins and diagnostics: a PTX output-lane mask raised GPU frequency for a 2-3% gain; moving several ops onto Tensor Cores yielded ~15% latency reduction; hierarchical inverse decomposition cut latency ~17%; adapting FlashAttention’s mixed exp strategy improved latency ~2.9%; synchronization analysis detected a latent mbarrier race that passed functional tests. The harness thus enables discovery, transfer of strategies across kernels, and safer, repeatable evaluation for agentic kernel optimization.