README rewrite (c't style) + COMPARISON.md + AMX SIGILL fix#103
Conversation
…ection Root cause: amx_available() used CPUID leaf 0xD (what CPU supports) instead of _xgetbv(0) (what OS actually enabled). Hypervisors that advertise AMX in CPUID but don't enable tile state caused SIGILL. Fix — 4-step detection: 1. CPUID.07H bits 24+25 → CPU has AMX-TILE + AMX-INT8? 2. CPUID.01H bit 27 → OS supports XSAVE? 3. _xgetbv(0) bits 17+18 → OS ACTUALLY enabled tile state? 4. prctl(ARCH_REQ_XCOMP_PERM, 18) → process has tile permission? VNNI hierarchy documented: avx512vnni (EVEX zmm, 64 MACs) → checked first avxvnniint8 (VEX ymm, 32 MACs) → only if avx512vnni absent Before: 1612 pass + SIGILL crash After: 1612 pass, 0 fail, 36 ignored Cherry-picked from 282daf7 (claude/continue-lance-graph-ndarray-Ld786) https://claude.ai/code/session_017ZN5PNEf8boFBgorUZVrFU
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f68fefa514
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| // multiple times. | ||
| #[cfg(target_os = "linux")] | ||
| { | ||
| const SYS_PRCTL: i64 = 157; // x86_64 syscall number for prctl |
There was a problem hiding this comment.
On Linux x86_64, ARCH_REQ_XCOMP_PERM is an arch_prctl operation, but this code issues syscall 157 (prctl). That call returns EINVAL for option 0x1023, so amx_available() will return false whenever this branch runs, disabling AMX even on hosts where AMX is actually usable. The request should be made via arch_prctl (syscall 158) with the same option/feature arguments.
Useful? React with 👍 / 👎.
2 participants
Summary
AMX SIGILL fix detail
Root cause: amx_available() checked CPUID (what CPU supports) not _xgetbv(0) (what OS enabled). Hypervisors advertising AMX without enabling tile state caused SIGILL on LDTILECFG.
Fix: 4-step detection (CPUID → OSXSAVE → _xgetbv(0) → prctl). Before: SIGILL crash. After: 1612 pass, 0 fail.
Test plan
https://claude.ai/code/session_017ZN5PNEf8boFBgorUZVrFU