HIPnosis
AMD Developer Hackathon · ACT II

Port CUDA to AMD ROCm —
and prove it still
computes the same numbers.

HIPnosis is an autonomous agent. Give it a CUDA repository; it compiles on a real MI300X, drains every build error, catches the wavefront-64 bugs a textual port ships silently, and hands back a verified port with a proof you can check yourself.

Launch live demo → How it works
387 tests · Gemma 3 27B local · ROCm 6.x · $0 API on the demo port
hipnosis · live run — bsw-cuda
HIPnosis dashboard: 8 build errors drained to 0, 100% resolved locally, 2 wave64 correctness bugs caught, verdict PASS on gfx942
Port Passport VERIFIED
diffsha256:8b3f…c10401
verdictPASS · 8→0
targetgfx942
Built on ROCmAMD Instinct MI300X Gemma 3 27BFireworks AI
The bug a translator can't see

A port that compiles can still be silently wrong.

NVIDIA warps are 32 lanes. AMD wavefronts are 64. Translate this line and it compiles cleanly on an MI300X — then returns corrupted numbers, with no error and no crash. HIPnosis is built to catch exactly this class of divergence, statically, before it ships.

// hipify translates it. hipcc accepts it. The result is wrong.
__ballot_sync(0xffffffff, tid < n); // 32-bit mask on a 64-lane wavefront
int lane = tid % 32; // warp-size arithmetic, off by 2×

// HIPnosis flags W01 · W05 → correctness, before verify.
Translation is step one of five

Most tools stop where the real work begins.

A translator hands you HIP source and calls it done. HIPnosis treats translation as a single phase, then compiles, patches, runs and verifies on real hardware until the numbers match.

Capability
Translation tools
HIPnosis
Translate CUDA → HIP syntax
✓ yes
✓ yes
Compile on a real MI300X, in a loop
✕ scripted / offline
✓ deterministic loop
Fixes measured by the compiler, reverted if worse
✕ none
✓ oracle-driven
Numerical parity vs. a reference (rtol/atol)
✕ none
✓ per-run
Wave64 silent-correctness detection
✕ not checked
✓ W01–W07
Verifiable, hash-signed provenance
✕ round numbers
✓ Port Passport
Automated test suite
✕ none published
✓ 387 tests
See it yourself

A verified port, in three steps.

The demo replays a recorded run of bsw-cuda (Smith-Waterman) end to end. No install, no sign-up — everything below runs in your browser.

INPUT · a CUDA repository URL
1
Phase · Build loop

Watch the loop drain

8 compiler errors fall to 0 across 4 iterations — each fix classified, committed, and measured by the build.

Open →
2
Phase · Static analysis

Find the silent bugs

2 wavefront-64 divergences flagged correctness — the exact bugs a textual port compiles clean and ships broken.

Inspect →
3
Phase · Verify

Verify the proof

The Port Passport recomputes sha256(diff) in your browser. Tamper one byte → TAMPERED.

Check →
✓  OUTPUT · a verified ROCm port + a Port Passport
Cross the border with papers

Every port ships with a proof.

Alongside the human-readable certificate, each run emits a hash-signed attestation — the diff, the verdict, the environment, the source and final commits. It follows the SLSA provenance model, and anyone can verify it with sha256sum. No trust required, no vendor lock-in on the proof itself.

Verify a passport →
Port Passport ● VERIFIED
repobsw-cuda · Smith-Waterman
diffsha256:8b3f1a9…c10401
verdictPASS · 8 → 0 errors
wave642 correctness bugs caught
targetgfx942 · MI300X
provenanceSLSA-L1 · unsigned
Recomputed in your browser. Change one byte of the port and this seal reads TAMPERED.

Radical honesty is the point. The demo replays a recorded run, and the badge always says which mode you're in — synthetic, mock, or live silicon. We never show a scripted terminal claiming a GPU result that didn't happen.

See it run →