This is part 9 of SoftGPU. Hardware runs hide many memory and synchronization defects until a rare schedule hits them. SoftGPU’s value for developers is to detect a declared class of defects with workgroup, wave, and lane context, without pretending to be an AMDGPU memory-model oracle.
NVIDIA’s Compute Sanitizer already exists for CUDA. SoftGPU is not a claim of parity with that tool. It is a path that can run without allocating a physical GPU to every CI job, on the execution model SoftGPU actually implements.
The sanitizer model
| Concept | SoftGPU meaning |
|---|---|
| Shadow | Per-byte allocated / uninitialized / initialized / freed for global and group arenas |
| Hard faults | Out-of-bounds, use-after-free, shadow-size limit — always fail closed |
| Soft findings | Uninitialized read, race, missing barrier, cross-workgroup race — fail-fast or collect |
| Happens-before | SoftGPU barrier generations under wave_barrier; same generation ⇒ no SoftGPU sync |
Modes: off, collect, fail_fast. Findings carry actor WorkItemId (workgroup, wave, lane, flat local) and an optional other actor.
Every finding can be packaged as softgpu-sanitizer-replay-v1 with program name, provenance, and exec config. Another engineer should be able to replay the finding without guessing the schedule.
Declared race subset
SoftGPU reports a conflict when two different workitems touch overlapping bytes with at least one non-atomic write, and:
- same workgroup, same barrier generation →
race(global) ormissing_barrier(group); - different workgroups on global →
cross_workgroup_race.
SoftGPU atomics on both sides of an access are treated as ordered by the sequential interpreter. There is no SoftGPU race between those atomics. That is an interpreter rule, not a proof about atomicAdd on silicon.
Blind spots (honest)
- Not AMDGPU acquire/release/scope evidence
- Not a claim about silent host-pointer aliasing outside SoftGPU arenas
- Not true hardware concurrency; schedule order alone does not define SoftGPU races
- Divergent barriers remain validate-time errors, not sanitizer findings
The suite covers OOB, use-after-free, uninitialized group reads, workgroup races, missing barriers, and cross-workgroup races in SFIR. It does not claim full GPU sanitizer parity with vendor tools, and it does not map findings to original HIP source lines.
cargo test -p softgpu-functional --test phase8_sanitize --locked
cargo run --locked -- run-functional --builtin tiny_add --sanitize collectNext: debugging a machine made of SoftGPU lanes.
Canonical: thanos.github.io.