Follow-up to #1452, claimed in the coordination issue: claim.
What and why
--modes emulate exists to run a wave64 program as two 32-lane halves. With --wave 32 the selection therefore contains no program that emulate can run, and the tool ended like this:
$ python -I tools/crosslane_fuzz.py --wave 32 --modes emulate --programs 2
summary:
$ echo $?
0summary: with nothing under it, exit 0 — indistinguishable from a clean pass, and the summary line is the only place the program count appears. It is the same false pass #1452 fixed for --programs 0, reached through a combination of two valid flags instead of a count.
The refusal sits with the existing --oracle validation, after the --self-test early return (so a self-test still works with any flags) and before the missing-harness check, so it is reported as a usage error whatever the state of the build:
if not backends or any(t not in ("translator", "hw") for t in backends):
ap.error("--oracle takes translator, hw or translator,hw")
+ if args.wave == "32" and "emulate" in args.modes:
+ ap.error("--modes emulate needs wave64 programs: with --wave 32 nothing runs (use --wave 64 or both)")--wave 32 --modes native, --wave 64|both --modes emulate|native|both and the defaults (--wave both, --modes native,emulate) are unchanged.
Tested
CPU only, no GPU, no game data; Windows 11, CPython 3.12.13.
python -m pytest tests/test_crosslane_fuzz.py -q-> 37 passed, 1 skipped (35 before this PR).- Direct:
--wave 32 --modes emulate --programs 2->crosslane_fuzz.py: error: --modes emulate needs wave64 programs: with --wave 32 nothing runs (use --wave 64 or both), exit 2, no stdout; and with--exe no-such-build/vk_compute_testit still reports the usage error, not the missing-harness line. - Unchanged:
--wave 64 --modes emulate --self-testand--wave both --modes emulate --self-test->self-test OK, exit 0. - Negative control with only
tools/crosslane_fuzz.pyreverted tomain(99f99890) and the new tests kept -> 1 failed, 36 passed, 1 skipped: the refusal test fails, the wave64/both test passes on both sides as it should. python -m compileall -q tools tests-> clean.
Not covered: a real --wave 64 --modes emulate comparison run, which needs a Vulkan GPU.
Checklist
- No game data, keys or decryption (CONTRIBUTING.md), including in logs or reports quoted here.
- A new translator rule is in the contract table of
docs/design.md, with how it was confirmed. (N/A, tool CLI only.)