[karanvora2599]

What

The storage buffer that holds GDS is now declared Coherent.

Background.

• GDS is one on-chip memory for the whole GPU on AMD hardware. Every work-group's ds_read/ds_write with gds sees the others' writes, with no cache in between.
• The recompiler keeps GDS in a storage buffer (gds, SpirvEmitterHelpers.cpp). Every GDS access goes through it: reads, writes and atomics (SpirvBufferAccess.cpp, ResourceKind::Gds).
• That buffer wasn't decorated Coherent. On a GPU whose L1 caches storage-buffer loads (NVIDIA's from Turing on, among others), a plain GDS read could therefore return a value cached before another work-group's GDS write.
• Atomics weren't affected, since they're performed at the device.

Why results don't change. Coherent only stops loads and stores from being served by a non-coherent cache. A program without cross-work-group GDS traffic reads the same values. GDS traffic is small (counters and small tables), so the cost of going to L2 is negligible.

Tested

Windows 11, MinGW-w64 GCC 15.2.0, Release, this branch on main at d146ec1. Devices:

• RTX 4060 Laptop (Ada, driver 591.86)

• Intel UHD Graphics (Raptor Lake, driver 101.6790)

• lavapipe from Mesa 25.2.8 (llvmpipe, LLVM 21.1.7): the version CI installs; Windows build with GALLIUM_OVERRIDE_CPU_CAPS=avx.

• New verifyCoherentGds in agc_shader_memory:

  • It compiles a ds_read_b32 and a ds_write_b32 with gds and checks that the gds buffer is decorated Coherent.
  • It passes on all three devices.
  • With the decoration removed it fails: coherent GDS: the gds buffer is not Coherent.

• agc_driver_gds_append_consume passes on the RTX 4060 and the Intel UHD. lavapipe skips it.

• Shader and driver tests (agc_driver_*, agc_shader_*, agc_spirv_*, vector_*, agc_masked_select, agc_validate_program, agc_denormal*, agc_float16_*; 338 tests):

  • RTX 4060: all pass. I left out the BDA tests that lose the device on driver 591.86 (Recompiler: NVIDIA hangs (device lost) when read_bda_span lanes split between in-place loads and the fallback; branch and commit ready #3291): agc_driver_flat_loads*, agc_driver_global_loads*, agc_driver_flat_apertures*, agc_driver_pixel_interlock.
  • Intel UHD: everything passes or is skipped for missing features, except agc_driver_ds_swizzle and agc_driver_f32_dot2_rounding, which fail the same way on main.
  • lavapipe 25.2.8: everything passes or is skipped for missing features, except the segfault in agc_driver_depth_fast_clear_submit, which also happens intermittently with main's code.

Limits.

• I have no test that catches a stale GDS read on hardware. The failure needs two work-groups and a cached line, so the test only checks the declaration.
• fma-rc's census of nine titles (perf(shader): compute v_dot4 with packed integer dot products #4061) doesn't record the gds bit, so I don't know how often titles read GDS with plain loads. ds_append (25) and ds_consume (2) appear, but they are atomics.

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: none
• AI-assisted: yes

[adriadam10]

Rebased onto current main (only an adjacent-test conflict in ShaderMemory.cpp, no logic change). Checked: agc_shader_memory passes with the change and fails without it; the scoped ctest passes apart from agc_driver_tessellation, which segfaults here without this PR's change too. Merging once CI is green.