Repository navigation
Mesa ANV: v_cmp_*_f16 treats f16 subnormals as zero (agc_driver_vopc_compare) #2861
Description
agc_driver_vopc_compare fails on Intel Arc (Meteor Lake, Mesa 26.2.3 ANV), as reported for other Intel GPUs in #2114:
vopc compares: thread 0 compare 36 is 0, expected 1
Every failing compare has an f16 subnormal operand (0x0001): -0 < 0x0001 and 0 < 0x0001 come out false. EmitF16BitsToF32 widens halves with UnpackHalf2x16, which may flush subnormals, and ANV does.
Fix: ThatsLeo/AnyPS5:fix/shader-f16-subnormal-widen (one commit, e59865e0, on main 57f96fd2). A subnormal half is converted as its fraction times 2^-24 with its sign, which is exact in f32; other halves still go through UnpackHalf2x16. All callers of EmitF16BitsToF32 get the exact conversion.
Tested: Fedora 44, GCC 16.2.1, Release. agc_driver_vopc_compare passes on Intel Arc and lavapipe. Full ctest on main (57f96fd2), on this branch and with all eight fixes listed in #2803-#2807 and here, on Intel Arc and lavapipe: no new failure. With all of them and #1250, all 698 tests pass on Intel Arc and on lavapipe. Dreaming Sarah (PPSA02929) runs with no visual difference.
AI-assisted: yes.
Activity
Silverash-i8 commentedon Oct 10, 2026
Tested this on the Windows Intel driver too, since it's the same failure I reported in #2822 for Arrow Lake.
Windows 11, WinLibs GCC 15.2.0 r7, Release, Intel Graphics in a Core Ultra 9 275HX (Arrow Lake, driver 32.0.101.8991),
ANYPS5_GPU=Intel. Onmain57f96fd2I stackede59865e0(this) and7df6c709(#2860) with the other Intel fixes (#2777's758eeb9f, #2815's377829ec, the LDS lock change from #2805, #1250 forvop1_float_unary), plus thefp16_overflowskip from my7217a416:agc_driver_vopc_comparepasses.ctest -j6: 2 failures out of 709 (46 skipped on this device). One wasagc_driver_gated_dispatch_capture, the timing flake from AGC: WAIT_REG_MEM gives up after 1 s while the awaited label is still pending on another queue, so gated work reads too early #2832, which passes on its own. The other waspatched_libraries, which passes on a re-run and doesn't touch shaders.So both fixes work on Arrow Lake. The difference is that #2822 adds
DenormPreserve 16toUnpackHalf2x16and only applies where the device has Float16, RoundingModeRTE and DenormPreserve, while this one converts subnormals exactly everywhere, at the cost of a few extra instructions per widen. Since this one doesn't depend on device support, I'd be happy for it to be the one that goes in, and I'll close #2822 if the maintainers agree.AI-assisted (Claude Code); I ran everything above on my machine.
oneandonlydean commentedon Oct 10, 2026
Closing as a duplicate of #2822. Your exact widening (
e59865e0) and test results would be useful there; #1463 (open) takes the same integer-decode route.