HIPCC_VERIFY: detect IGC-dropped kernels at compile time

[pvelesko]

Summary

Adds HIPCC_VERIFY (default ON): after every successful hipcc compile, a new post-compile tool chip-kernel-verify invokes ocloc and diffs the SPIR-V OpEntryPoint set against the kernels present in the native binary. Mismatches — silent entry-point drops by IGC — fail the compile with a named list of the missing symbols.

Tracks: intel/intel-graphics-compiler#403 — on dg2 IGC's SIMD32 register-pressure retry path silently drops template instantiations from the native binary while ocloc reports success. At runtime chipStar then fails with "Failed to find kernel via kernel name" far from the root cause. This check catches that at build time.

What's in the box

• tools/hipcc-verify/chip-kernel-verify — extracts the SPIR-V from a hipcc output (reusing tools/spirv-extractor/), runs ocloc compile -spirv_input -device <...> per target in HIPCC_VERIFY_DEVICES, then ocloc disasm to enumerate the native kernels, diffs, reports.
• CMake option HIPCC_VERIFY (ON) + cache var HIPCC_VERIFY_DEVICES (default dg2). Baked into hipcc.bin as HIPCC_VERIFY_DEFAULT.
• HIPCC driver hook (paired branch: CHIP-SPV/HIPCC hipcc-verify) — after a successful compile, invokes ${HIP_PATH}/bin/chip-kernel-verify <output>. Submodule pointer bump included here.
• Test fixture + ctest: tests/hipcc-verify/ ships a real 1.9 MB SPIR-V dump (via CHIP_DUMP_PROCESSED_SPIRV) of rocThrust set_difference_by_key. Under IGC-with-#403 the reproducer ctest fails and names the 3 dropped lookback_set_op_kernel instantiations (ff/ii/jj element types). When IGC is eventually fixed the test flips to passing — that's the signal to retire the fixture.

Env overrides

Var	Default	Effect
HIPCC_VERIFY	built-in ON	0/off = skip; warn = print only; else = fail on mismatch
HIPCC_VERIFY_DEVICES	dg2	comma-separated ocloc -device targets
HIPCC_VERIFY_OCLOC	ocloc	binary path override

If ocloc is absent, the verifier prints a one-line notice and exits 0 — never fails a build just because the tool isn't installed.

Example failure output

[chip-kernel-verify] device=dg2: 3 kernel(s) missing from native binary
  (IGC likely dropped them; see https://github.com/intel/intel-graphics-compiler/issues/403 ):
    - ...lookback_set_op_kernel<...default_set_operations_config<10000, float, float>, true, ...>
    - ...lookback_set_op_kernel<...default_set_operations_config<10000, int,   int>,   true, ...>
    - ...lookback_set_op_kernel<...default_set_operations_config<10000, uint,  uint>,  true, ...>

Test plan

• Standalone build of chip-kernel-verify — smoke-tested off/fail/warn modes, missing-file, non-fatbin ELF, missing-ocloc paths.
• Full chipStar build with LLVM 21 (meatloaf). hipcc produces working binaries; verifier fires per compile.
• End-to-end: hipcc samples/2_vecadd/VecAdd.cpp with HIPCC_VERIFY_DEVICES=nonsense-device → hipcc exits 1; same with HIPCC_VERIFY=warn → exits 0; HIPCC_VERIFY=0 → fully silent.
• ctest -R hipcc_verify_ — hipcc_verify_detects_dropped_kernels FAILS under current IGC (expected: demonstrates IGC #403); hipcc_verify_off_mode_silent passes.
• Full check.py dgpu level0 — verify no regressions from default-on verification on the existing suite.
• Full check.py dgpu opencl.
• Salami / pastrami (aarch64 / macOS where ocloc is absent) — confirm graceful skip.

Notes

• Runtime Level Zero code is unchanged — this is purely a build-time check.
• tools/spirv-extractor/spirv-extractor.hh has an unbounded 1 MB scan on its raw-data path; the verifier works around it by padding the input buffer before calling extractSPIRVModule. Separate cleanup for another PR.
• The reproducer fixture is 1.9 MB — consistent with existing multi-MB .ll fixtures under tests/compiler/promoteInt/.

References

• IGC upstream issue: SIMD32: lookback_set_op_kernel (Lb1E) variants dropped from native object after RetryManager; ocloc still succeeds intel/intel-graphics-compiler#403
• Paired HIPCC branch: https://github.com/CHIP-SPV/HIPCC/tree/hipcc-verify