Repository navigation
Test sub-group shuffles of a mixed-width tuple - #732
Merged
Merged
Conversation
On Gen12LP, IGC could place the byte that receives the Int8 of a shuffled Tuple{Bool, Int8, Float16, UInt16} at byte 3 of a dword, which the hardware doesn't write, so the Int8 came back as 0 (intel/intel-graphics-compiler#456). libigc_jll 2.41.5+2 carries the fix; NEO_jll's compat already allows it.
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #732 +/- ##
==========================================
- Coverage 85.08% 84.62% -0.47%
==========================================
Files 53 52 -1
Lines 4660 4533 -127
==========================================
- Hits 3965 3836 -129
- Misses 695 697 +2 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds a regression test for an IGC miscompile on Gen12LP (Tiger Lake, DG1, Alder Lake integrated graphics): a sub-group shuffle of a
Tuple{Bool, Int8, Float16, UInt16}returned 0 for theInt8. The shuffle packs the first three fields into one word. When unpacking, IGC could place the byte that receives theInt8at byte 3 of a dword. The hardware doesn't write a byte destination there, so the stale register contents got stored. Details and an OpenCL C reproducer are in intel/intel-graphics-compiler#456.This showed up in KernelInterface's shuffle tests on top of #731: the exception-reporting code on the bounds-check paths changed the register allocation. A plain kernel hits the bug on main too, though. The test reads and writes through pointers, so the bounds-check flag doesn't change the generated code. Without the fix, it fails on an Iris Xe (TGL) with and without
--check-bounds=yes.The fix is backported in JuliaPackaging/Yggdrasil#15280, which builds libigc_jll 2.41.5+2. No compat change is needed: NEO_jll's compat accepts any 2.41.5 build of libigc_jll, and Pkg resolves to the latest one. Until that build is registered, CI on Gen12LP hardware will fail this test. I ran the full test suite on TGL with that build: 22028 tests passed, 30 broken, none failed, including the new test.