AMD Kernel Autopilot
Autotuning that locks in the fastest correct GPU kernel.
AMD/ROCm lacks a robust kernel-autotuning ecosystem, especially on consumer RDNA2 and Windows. Autopilot removes the guesswork: point it at a parameterized kernel and it generates candidates, verifies each against a CPU reference, benchmarks them, and keeps the winner — fast-but-wrong results are rejected.
$ autopilot tune wgrad-rt --execute
scanning 18 candidates · correctness-gated
✔ best config verified · rel-L2 3.1e-4 (f32)
before 4.13 s → after 1.26 s (3.27× faster)
✔ locked to autopilot.lock
Representative output — not a live instance.
●Parameterized kernel templates — 13 canonical patterns (elementwise, reduction, GEMM, softmax, and validated RWKV-7 kernels) swept over tile size, block dim, unroll, wave width and LDS usage.
●Correctness-gated autotuning — every candidate is verified against a CPU reference before it's benchmarked; fast-wrong results never win.
●Multi-backend execution — HIP for AMD and wgpu/Vulkan compute for portability across AMD, NVIDIA, Intel and Apple.
●A real, measured win — the four RWKV-7 backward patterns it ships were measured at 4.13s → 1.26s (3.27×, rel-L2 ≤ 1e-3) on the author's RDNA2 workload; an imported benchmark, not one the dry-run tool re-measures.
●Safe by default — dry-run codegen with no GPU by default; opt-in --execute; hard timeouts and process isolation to survive GPU driver hangs.
3.27×
measured RWKV-7 backward speedup
0
kernel templates
0
unit tests · clippy clean
≤1e-3
rel-L2 correctness bound
Rust · 9 cratesHIP (hipcc)wgpu / WGSLclap · serdeRDNA2 (gfx1031) validated
🔒
Architecture and measured results shown. Source and kernel IP are private — this is a capability showcase, not a distribution.
Need GPU code
that's actually fast?
I profile, tune, and verify GPU kernels — correctness first, then speed. Tell me what's bottlenecked.
Start a project →