[adriadam10]

What

This implements the six buffer atomics still missing. It also fixes the existing 32-bit buffer_atomic_fmin/fmax.

Opcode	Semantics (measured on an RDNA2 GPU)
buffer_atomic_fcmpswap (0x3e) / _x2 (0x5e)	mem = (mem == cmp) ? src : mem, compared as floats: −0 equals +0, NaN equals nothing, denormals compare exactly
buffer_atomic_fmin_x2 (0x5f) / fmax_x2 (0x60)	IEEE minNum/maxNum on f64: −0 < +0. A signalling NaN operand gives that NaN quieted (src first). A quiet NaN loses to a number. Two quiet NaNs give src
buffer_atomic_inc_x2 (0x5c) / dec_x2 (0x5d)	as the 32-bit inc/dec from #321, on u64

Bug: the existing 32-bit buffer_atomic_fmin/fmax (EmitFloatAtomicReplacement) kept the old value whenever either operand was a NaN or both were zeros. The hardware follows the same rule as the f64 forms above:

• fmin(qNaN, 1.0) = 1.0
• fmin(+0, −0) = −0
• fmin(1.0, sNaN 0x7fa00000) = 0x7fe00000

2048/2048 vectors on hardware. The helper is now unused and is removed.

Implementation:

• AtomicUpdate's compare-exchange loop is generalized to a type (AtomicUpdateTyped), and BufferAtomic64Update runs it on the u64 storage-buffer view that feat(shader): implement buffer_atomic_inc/dec and the 64-bit integer buffer atomics #321 added.
• AtomicFloatBits does the float compare and minNum/maxNum on the raw bits, for 32 and 64 bits, so the driver's denormal/NaN float controls don't matter.

Measuring: float atomics are not performed on fine-grained system memory. The first capture returned zeros, while the integer atomics were fine. The vectors were captured with the buffers in the GPU's coarse-grained pool.

Tests, all values captured on the RDNA2 GPU (ROCm HSA, raw kernels):

• agc_driver_buffer_atomics_float: fcmpswap (32/64) returned and stored values, fmin/fmax 32 and 64 across ±0, ±inf, quiet and signalling NaNs with payloads, and denormals.
• agc_driver_buffer_atomics_inc_dec64: the wrap edges (0, limit, limit ± 1, 2^32, 2^64−1).

Tested

Linux, full ctest (101) passes.

• RADV (Radeon 660M) and lavapipe: both tests pass.
• With the old 32-bit replacement: fmin_mem0 is 0x7fa00000, expected 0x7fe00000. Without the sNaN quieting: fmin_x2_mem0 is 0xbf800000, expected 0x00000001.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: ci: run the shader execution tests on lavapipe #251 (test harness), feat(shader): implement buffer_atomic_inc/dec and the 64-bit integer buffer atomics #321 (64-bit buffer atomics); review only the last commit
• AI-assisted: yes

[github-actions]

Progress report (2cf901b → 7f291b7)

📈 GPU shader instructions: 81.73% (+1.46%, +17 implemented instructions)

✅ 17 implemented

Encoding	Name
MUBUF	BUFFER_ATOMIC_ADD_X2
MUBUF	BUFFER_ATOMIC_AND_X2
MUBUF	BUFFER_ATOMIC_CMPSWAP_X2
MUBUF	BUFFER_ATOMIC_DEC
MUBUF	BUFFER_ATOMIC_DEC_X2
MUBUF	BUFFER_ATOMIC_FCMPSWAP
MUBUF	BUFFER_ATOMIC_FCMPSWAP_X2
MUBUF	BUFFER_ATOMIC_FMAX_X2
MUBUF	BUFFER_ATOMIC_FMIN_X2
MUBUF	BUFFER_ATOMIC_INC
MUBUF	BUFFER_ATOMIC_INC_X2
MUBUF	BUFFER_ATOMIC_SMAX_X2
MUBUF	BUFFER_ATOMIC_SMIN_X2
MUBUF	BUFFER_ATOMIC_SUB_X2
MUBUF	BUFFER_ATOMIC_UMAX_X2
MUBUF	BUFFER_ATOMIC_UMIN_X2
MUBUF	BUFFER_ATOMIC_XOR_X2

[adriadam10]

Rebased on main. main now has SharedAtomicUpdate for the LDS atomics, so this PR no longer carries its own SharedFloatAtomic: the shared min/max path is main's, unchanged. The buffer side (AtomicFloatBits, BufferAtomic64Update) is as before. Still stacked on #321; review only the last commit.

[oneandonlydean]

This passed my PPSA21564 check too; it needs a rebase once #321 is rebased, since #319 to #331 are now on main. Then it's ready to merge from my side.

[adriadam10]

Rebased on the rebased #321 and current main (same MUBUF table conflict, both sides kept). BufferAtomicsFloat, BufferAtomicsIncDec64 and BufferAtomics pass on RADV and lavapipe.

[oneandonlydean]

Thanks for the rebase. Both commits have exactly the content I ran through my PPSA21564 check (the first is #321's head), and agc_driver_buffer_atomics_float, agc_driver_buffer_atomics_inc_dec64 and agc_driver_buffer_atomics pass here on current main. Ready to merge right after #321.