tree f0076251d896b5a3a02dfcae4609d3ff58d14ba0
parent 0df50f9a8e40ab1878479124b952f45d9f8c0650
author Jakub Kuderski <jakub@nod-labs.com> 1789130234 -0400
committer GitHub <noreply@github.com> 1789130234 -0400
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJqo/X6CRC1aQ7uu5UhlAAAN14QAKpWFwwxqlb2iCxaIEK6cfpV
 ilJ43d/5rMMGJRX7e/o5cCEVAxpEPpaL0FX6ekJt+7SVOV9hk9DwNJJeMsgW0q7U
 ikGA+L3xqU52OhjGuUQBVzuQkgs3gAigt5Sy/9cya1L0wcqqk8mJb9PcyUbUFWiW
 NWJF55NTE6pjA6qutS4+9DPUF4bYvx/t8OULiHMAgV79RwJ353zt/eYsWrosoHiP
 QHK83fsV8Ofz57JeKJKEAfSK+nnY832n/FFqrHYqRayPxv/X3277kNginKuDab1L
 of0lj5b5/OX45t/Yz86CNhtRGyRuINTOZxrR2RByW3xFHK62U5l0CH+Ide/20j5D
 SJDY+C3EDBx4sSpfq58M5GrDnDj34UnMzfjIHsY9a5Nf/msAR0fHvblowbsNH9PW
 QVXoWpUnwIbjyKZIWo42cEfWSBfUE9A8jj/aP45AYov/XVLFitzPiliXufFDBGAb
 fWdi837WO4k7SiLEQZu8+JKj5GJ/W6zfZMgHPLWY73/nJI39ifTwTtCcbV7EYX6T
 835WsdZjj4zW36AhTF59+DzulwqOJf1Y0lpQmWz3bUvP1gqVUAiPseZvDrwntuPJ
 pSnEZAfX5thPB1hTd1ERAeHHL6tW+62BqV6238Cqyaj9uGf7CAjndO11ovFFkYvm
 ufx8QZWu97/k8qfHAPin
 =AtJm
 -----END PGP SIGNATURE-----
 

[ROCM] Add rocjitsu e2e validation (#24913)

I want to enable e2e testing across all recent amdgpu / rcom targets via
rocjitsu, an amdgpu simulator, so that anyone with an x86 machine can
run those tests regardless of the local GPU or lack thereof.

Add a CPU-hosted ROCm e2e test helper for gfx942, gfx950, gfx1100,
gfx1201, and gfx1250. Reconfigure a regular IREE build and rebuild
iree-test-deps for each target, then run HIP tests with the simulator
from a pinned TheRock nightly.

Run each test in a fresh Mirage session with a watchdog and per-test
JUnit output. Propagate configure/build failures so stale artifacts
cannot produce a successful run.

Make the default native-test timeout configurable and label tests that
are too expensive to simulate. Keep those tests enabled for hardware
runs. Exclude TopK v2 from ROCm tests because its missing thread
distribution causes overlapping output writes on all five targets.

Use the target's preferred subgroup size when an executable does not
specify one, so gfx1250 defaults to wave32. Add regression coverage for
the default and explicit subgroup sizes.

Assisted-by: codex (gpt-6-astra)