[Copak731]

What and why

Three half-pair defects, found by a GPU cross-check against an exact binary16 reference (one rounding, FMA included).

• The swizzle of a constant operand was ignored. AHalfLanes took a constant as is, so HMUL2 R37, R37.H0_H0, c[0x0] [0x168].H0_H0 (a half scaled by one half of a uniform) gave the high lane the constant's other half. nvdisasm names the swizzle in all three slots (A 74..75, the 32 slot 60..61, the 64 slot 81..82); only an immediate has none. This also affected HSETP2 / HSET2 / HMNMX2 through AHalfSecond.
• Swizzle value 1 is .F32, the operand's own 32-bit float (nvdisasm), not .H1_H1. Its exact rounding on NVIDIA is not documented; the float itself is the closer reading. MUFU.F16, whose bits 60..61 pick one half, keeps its reading (MufuF16 caught an early version of this change).
• HSETP2 / HSET2 .FTZ (bit 80) was dropped. A subnormal half must compare as zero, as AFtz already does for FSETP / FSET; but the lanes are unpacked to 32 bits, where a half's subnormals are normal floats, so AHalfFtz flushes under the half's own bound (2^-14). .H_AND keeps the reading NAK's OpHSetP2.horizontal documents.

Reference entries: docs/reference/shaders.md, "Half-pair swizzle on a constant, and .F32" and "HSETP2 / HSET2 .FTZ".

Checked

• Tests: DMSH_TEST_XLATE=<translator> python3 dm/gfx/tests/test_dmsh_ampere_alu.py, new HalfSwizzle (3) and HalfCompareFtz (3); the five checks of the fix fail on main. All test_dmsh_*.py pass.
• GPU cross-check (The cross-checks translate real sm_86 encodings (an nvdisasm 13.5-validated corpus) with dmsh_xlate using the production options (debug and reduced on, as dmsh_build.py), convert with spirv-cross 1.4.363.0 at MSL 2.3, run them on the GPU and compare every bit with a reference written from the oracle, PTX's documentation and Mesa NAK where it applies.): HADD2 / HMUL2 / HFMA2 / HSETP2 / HSET2 / HMNMX2, 1,200 instructions: 89.7 % identical on main, 96.8 % with the change (86 constant-swizzle mismatches in HADD2 / HMUL2 / HFMA2 and 15 in HSETP2 to none; 4 FTZ compares in 400 subnormal inputs to none). Left: arithmetic .FTZ on subnormals (dropped by design), .RELU on NaN, and NaN payloads.
• Devices and OS: MacBook, Apple M5, macOS 26.6.2, Xcode 27.2.
• Games run: none (no game files were used).

Docs

• The finding is in its docs/reference/shaders.md entry, and python3 dm/tools/docs_symptoms.py was run (the symptom index)
• No game files, keys or data extracted from a game in the change or the attachments

🤖 Generated with Claude Code

[seedds]

Landed as 9dc8195 with you as co-author (HalfSwizzle and HalfCompareFtz fail before, pass after). Thank you.