[harry-z-kang]

What

Fixes #6119: guest format 133 (5_6_5_UNORM) fails when NVIDIA does not expose the packed format as a storage image. Formats 134 (5_5_5_1_UNORM) and 136 (4_4_4_4_UNORM) have the same capability gap.

Keep the packed image and native sampled/attachment formats, but write through a compatible R16_UINT storage view. Materialize storage descriptors with the original conversion format and pack stores into the original 16-bit guest layout, including half-precision inputs. Restrict each view's usage to the operations its format supports.

Packing reuses EmitFormatStoreComponent in SpirvBackend/src/SpirvMemory/SpirvFormatConvert.cpp and the existing packed texture channel mappings. Vulkan's format compatibility classes, extended image usage and normalized fixed-point conversion define the view compatibility and clamping/scaling. AMD/PS5 comparison of rounding ties and NaN handling remains unverified and is recorded with the existing packed-texture entry in docs/dev/debt/libSceAgcDriver.md.

The branch retains merged native format-43 repacking (#7271), converted 32-bit sampling (#5958), R8/RG8 sRGB copy handling (#6054), D16 SINT stores (#5899) and float-mode handling (#5943). No open PR is required.

Tested

Local validation on main 6ef5a1c4c and rebased PR head 4d26b03aa: Windows 11 (NT build 26300), NVIDIA GeForce RTX 5080, driver 617.42, WinLibs GCC 15.2.0, Release, BUILD_TESTING=ON, ANYPS5_ENABLE_SPIRV_TOOLS=OFF.

• Rebuilt the six focused GPU test targets and their driver/runtime/shader dependencies on main and the PR. This pass did not repeat the full default build or libs build; both passed in the previous validation on base 9b34cba9e.
• Ran the same nine focused cases on main's implementation and the PR with the identical regression harness. Both runs used ANYPS5_GPU=5080, ANYPS5_REQUIRE_VULKAN=1, ANYPS5_NO_SHADER_CACHE=1 and VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation.
• Command: ctest --test-dir build -R '^agc_driver_(image_(converted_unorm|packed16_load|store_sint|sint8|repacked_float|packed16_storage_(133|134|136))|fmask_image_reads)$' -V --parallel 1.
• Main passed the six existing cases and refused the three new packed-storage cases because their formats cannot be used as storage images. The PR passed 9/9 with zero skipped, including all three packed-storage cases; the final restored-PR run took 4.21 s.
• Used Khronos validation 1.4.365 from the official Windows CI artifact, commit aea761f4d54b86682be4099c428dba27b6609940, supplied through a workspace-local VK_LAYER_PATH. Loader logs confirm insertion in every test. Neither log contains a missing-layer warning or any VUID- error: zero on main, zero on the PR.
• The rebase applied cleanly. Both code/test commits have identical patches; the debt commit changes only its surrounding context. Conventions, whitespace and registration of all nine focused cases passed.

The wider suite was not repeated in this focused validation pass. Its previously reported image-table reuse failure was reproduced on the older main 60fb8078c.

Three new packed-format cases cover channel order, 2,048-texel bit-exact sampled-load/store round trips, full/partial masks, swizzles, 32-bit/half-precision stores, clamping and NaN/infinity inputs, sampling after repeated stores and packed guest-memory write-back.

Title and Linux validation reported by @snoozedDev, on the original main d92a8dad3 plus PR heads 8322164f1 and 628767b52: Linux kernel 7.2.5, GCC 16.2.1, RTX 3090, NVIDIA 610.57.04, Release with ANYPS5_ENABLE_SPIRV_TOOLS=ON.

• All 81 selected Linux tests passed on both heads, including the three packed-storage cases.
• Cyberpunk 2077 PPSA04029 v02.120.000 had no storage-format exception in six PR runs; both baseline runs reproduced it. The 628767b52 run recorded 645 presents and 1,497 completed draws before the later clip-control refusal.
• Later stops are the separate CLIP_DISABLE/PA_CL_CLIP_CNTL refusal and the intermittent watchdog tracked in Cyberpunk 2077 (PPSA04029) on Linux: EngineWatchdog still trips on main d592d54e after #4321; the sceKernelBatchMap2 threads are now in the GuestAllocations registry walks #4634. This PR fixes the reported storage-format stop; it does not establish complete title compatibility.

The packed-storage additions are unchanged. No new title run on the rebased head, AMD/PS5 packed-store comparison or lavapipe run is claimed.

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 under Silent stubs in their library's technical debt file
• 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
• Fixes: Fixes Cyberpunk 2077 (PPSA04029) on NVIDIA: guest storage texture format 4 (R5G6B5) cannot be used as a storage image #6119
• Unblocks: PPSA04029 Cyberpunk 2077 (PPSA04029) on Linux: sceKernelOpen("") opens a directory and the read throws EISDIR on main 961b25fe #3308 past the format-133 storage-image stop, confirmed by the issue reporter
• GPUs tested: NVIDIA GeForce RTX 5080, driver 617.42, Windows 11 (local, main/PR Khronos validation comparison); NVIDIA GeForce RTX 3090, driver 610.57.04, Linux kernel 7.2.5 (reported by @snoozedDev on the earlier heads above)
• AI-assisted: yes

[github-actions]

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
PR 6612	docs/dev/debt/libSceAgcDriver.md L185-186
PR 7085	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp L1798-1799
PR 8563	core/shader/recompiler/SpirvBackend/src/SpirvBufferFormat.cpp L57, L112
PR 8564	core/shader/recompiler/SpirvBackend/src/SpirvBufferFormat.cpp L57, L112
PR 8658	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp L1217-1218
PR 8898	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp L252-255, L1950
PR 8998	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp L1973-1974 core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp L277-278

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 907	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvImageEmitter.cpp
PR 1368	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 1395	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 2434	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 4448	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp docs/dev/debt/libSceAgcDriver.md
PR 4449	docs/dev/debt/libSceAgcDriver.md
PR 5021	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/SpirvBackend/src/SpirvImageEmitter.cpp
PR 5058	docs/dev/debt/libSceAgcDriver.md
PR 5371	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvImageEmitter.cpp
PR 5372	docs/dev/debt/libSceAgcDriver.md
PR 5431	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 5609	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 5925	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 6044	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6169	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6187	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6188	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6192	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6194	docs/dev/debt/libSceAgcDriver.md
PR 6197	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6202	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6209	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6223	docs/dev/debt/libSceAgcDriver.md
PR 6280	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvImageEmitter.cpp docs/dev/debt/libSceAgcDriver.md
PR 6292	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6300	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 6412	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/debt/libSceAgcDriver.md
PR 6481	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp docs/dev/debt/libSceAgcDriver.md
PR 6488	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvImageEmitter.cpp
PR 6501	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6539	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6548	docs/dev/debt/libSceAgcDriver.md
PR 6549	docs/dev/debt/libSceAgcDriver.md
PR 6615	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6642	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6788	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 6804	docs/dev/debt/libSceAgcDriver.md
PR 6808	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 6867	docs/dev/debt/libSceAgcDriver.md
PR 6868	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 6946	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp docs/dev/debt/libSceAgcDriver.md
PR 6960	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 6971	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 7201	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 7256	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp docs/dev/debt/libSceAgcDriver.md
PR 7275	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 7457	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 7550	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 7604	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp docs/dev/debt/libSceAgcDriver.md
PR 7642	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvImageEmitter.cpp docs/dev/debt/libSceAgcDriver.md
PR 7684	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 7689	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 7705	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 7910	docs/dev/debt/libSceAgcDriver.md
PR 8014	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 8077	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/RdnaDecoder/include/RdnaDecoder/RdnaDescriptorFormat.hpp core/shader/recompiler/RdnaDecoder/src/RdnaDescriptorFormat.cpp
PR 8126	core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 8246	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 8395	docs/dev/debt/libSceAgcDriver.md
PR 8484	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
PR 8527	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 8552	core/libs/prx/libSceAgcDriver/CMakeLists.txt
PR 8592	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 8625	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvImageEmitter.cpp
PR 8672	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 9020	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvBufferFormat.cpp docs/dev/debt/libSceAgcDriver.md
PR 9335	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvBufferFormat.cpp docs/dev/debt/libSceAgcDriver.md
PR 9387	docs/dev/debt/libSceAgcDriver.md
PR 9508	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp core/shader/recompiler/SpirvBackend/src/SpirvImageEmitter.cpp docs/dev/debt/libSceAgcDriver.md
PR 9533	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 9671	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/RdnaDecoder/include/RdnaDecoder/RdnaDescriptorFormat.hpp core/shader/recompiler/RdnaDecoder/src/RdnaDescriptorFormat.cpp
PR 9672	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/RdnaDecoder/include/RdnaDecoder/RdnaDescriptorFormat.hpp core/shader/recompiler/RdnaDecoder/src/RdnaDescriptorFormat.cpp docs/dev/debt/libSceAgcDriver.md
PR 9673	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/RdnaDecoder/include/RdnaDecoder/RdnaDescriptorFormat.hpp core/shader/recompiler/RdnaDecoder/src/RdnaDescriptorFormat.cpp docs/dev/debt/libSceAgcDriver.md
PR 9674	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/RdnaDecoder/include/RdnaDecoder/RdnaDescriptorFormat.hpp core/shader/recompiler/RdnaDecoder/src/RdnaDescriptorFormat.cpp docs/dev/debt/libSceAgcDriver.md
PR 9675	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/RdnaDecoder/include/RdnaDecoder/RdnaDescriptorFormat.hpp core/shader/recompiler/RdnaDecoder/src/RdnaDescriptorFormat.cpp docs/dev/debt/libSceAgcDriver.md
PR 9676	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/RdnaDecoder/include/RdnaDecoder/RdnaDescriptorFormat.hpp core/shader/recompiler/RdnaDecoder/src/RdnaDescriptorFormat.cpp docs/dev/debt/libSceAgcDriver.md
PR 9717	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 9742	core/libs/prx/libSceAgcDriver/CMakeLists.txt docs/dev/debt/libSceAgcDriver.md
PR 9748	core/libs/prx/libSceAgcDriver/CMakeLists.txt core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp core/shader/recompiler/SpirvBackend/src/SpirvBufferFormat.cpp docs/dev/debt/libSceAgcDriver.md

[harry-z-kang]

Reviewed both PRs in the overlap report. Neither is a prerequisite, so Depends on: none remains accurate.

• feat(shader): store D16 data to SINT storage images #5899 adds D16 stores to SINT images. This PR handles packed 16-bit UNORM images. Commit 628767b52f17a4dcdc39e559d912eddac6373fe2 keeps our half-to-float conversion in PackImageTexel, leaving StoreTexel unchanged from main and removing the conflict with feat(shader): store D16 data to SINT storage images #5899.
• feat(shader): sample converted unorm images in the shader #5958 adds sampling of converted 32-bit UNORM/float images; this PR preserves native sampling for the packed 16-bit formats. The conflict was between adjacent technical-debt entries; the packed-format entry has been moved without changing either entry's meaning.

Using the overlap checker's normalized merge procedure, the updated branch merges cleanly with #5899 at 8c1193744ac5 and #5958 at 8338defe54dd, in either order, and with both combined. No fixed merge order is required for those heads. These are merge checks; combined runtime builds were not run.

The updated standalone branch passed the full build, libs, all 98 related tests on the RTX 5080, and conventions/whitespace checks. Cyberpunk remains untested. The PR description now records the dependency review; the scheduled overlap report can refresh against the new head.

[Dextroid17]

Repeated @jotitadev's MoltenVK check on Linux, on the current head and plain main: 5d3a43196 merged onto main a62c724c3, lavapipe (llvmpipe, Mesa 26.2.4, LLVM 23.1.1), GCC 16.2.1, Release, ANYPS5_ENABLE_SPIRV_TOOLS=ON. lavapipe reports no optimal-tiling STORAGE_IMAGE for R5G6B5, A1R5G5B5 or B4G4R4A4_UNORM_PACK16 (it does for R16_UINT), so main refuses these three formats here, as on the RTX 3090 in #6119. By reading Texture.cpp, the PR sends them through the R16_UINT view on every device, not only on drivers without the bit.

• agc_driver_image_packed16_storage_{133,134,136}, agc_driver_image_packed16_load and agc_driver_image_converted_unorm: 15/15 passed (3 repeats).
• With the six non-test source files checked out from main (the PR's tests and CMakeLists.txt kept), the three packed-storage tests fail with the Cyberpunk 2077 (PPSA04029) on NVIDIA: guest storage texture format 4 (R5G6B5) cannot be used as a storage image #6119 error, guest storage texture format 4 cannot be used as a storage image for 133, and format 8 and 3 for 134 and 136. With the PR sources restored, all 5 pass again.
• Full ctest: 1 failure out of 1,451 (65 skipped), agc_compute_pipeline_replay (replay did not export a pipeline cache header). It fails the same way with those six files from main, and the PR does not touch it.

AI-assisted: yes

[DavidAmsellem]

First AMD-on-Linux (RADV) run of this PR. Merged onto main 8d5d4df96 (clean merge), built in full and compared with the same main, on the same machine.

AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 (Ubuntu 24.04 base), kernel 7.0.0-38, GCC 13.3, Release.

Full suite (ANYPS5_REQUIRE_VULKAN=1 ctest -j8) at b74c8abf4: 1641/1643 with this PR, against 1638/1640 on main. The two failures are the same on both, so this PR adds none: agc_driver_image_table_vouched_lookups (#7766, fixed by #7916) and agc_driver_guest_memory_lifetime (#7784).

The PR's three new ctest entries (agc_driver_image_packed16_storage_133, _134 and _136) all run and pass here; none is skipped. No other ctest entry is added, skipped or removed compared with main.

This checks correctness only; I did not measure performance.

GPUs tested: AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 / kernel 7.0.0-38

AI-assisted: yes

[klosixgaming-arch]

Cyberpunk 2077 PPSA03974 (v02.120.000, the same title as #6119 under another region's ID) on Windows 11 26200, RTX 5090, NVIDIA 617.42, WinLibs GCC 15.2.0.

• main b13183e stops 160 s in:

  [gpu] driver failure: compute shader 0x1aa4f07500: AGC graphics: guest storage texture format 4 cannot be used as a storage image
  

• This PR's merge ref a337b88 (head 4adf405 on main c70524a): that stop is gone and the title runs for 440 s, then stops with the Cyberpunk 2077 (PPSA04029): draws refused for PA_CL_CLIP_CNTL CLIP_DISABLE (register 0x204 = 0x00090000) once #6218 accepts the packed storage format #6806 line:

  unsupported PA_CL_CLIP_CNTL flags is unsupported (register 0x204 = 0x00090000)
  

  That base predates fix(agc): accept PA_CL_CLIP_CNTL CLIP_DISABLE without depth clamp #8046, which accepts this value on main now.

AI-assisted: yes. Written with AI assistance and reviewed by me.

[thuanquang]

Hardware data for the packed stores on RDNA 2: RX 6700 XT (gfx1031), raw ISA through the HIP runtime of AMD's Windows driver 32.0.21045.5002, so the conversion is the hardware's, not a Vulkan driver's.

Method: image_store v[4:7] ... dmask:0xf dim:SQ_RSRC_IMG_2D unorm glc through a linear 2D T# of each format (GFX10 codes), then image_load of the same texel through a 16_UINT T# over the same memory. First, one-channel stores of 1.0 locate each field. Then each channel gets 0, -0, 0x00000001, 0.25, 0.5, 0.5 ± 1 ulp, 1, 1 + 1 ulp, 2, -1, ±inf, ±NaN, a signalling NaN, 1/3 and 2/3, plus the f32 neighbours of every rounding midpoint (k + 0.5) / (2^n - 1) of its field. The same rows ran at IEEE 0 and 1, DX10_CLAMP 0 and 1, and with f32 denormals flushed and kept, and again with d16 half-precision data for the special values.

code	fields, low bit first	matches this PR's layout
133 5_6_5	R 0-4, G 5-10, B 11-15	yes
134	R 0-4, G 5-9, B 10-14, A bit 15	yes (5_5_5_1, W in bit 15)
135	R bit 0, G 1-5, B 6-10, A 11-15	not mapped by the PR
136 4_4_4_4	R 0-3, G 4-7, B 8-11, A 12-15	yes

Per field, in every mode, f32 and d16 alike:

• The stored value is the exact product x * (2^n - 1) rounded to nearest, not fl32(x * scale) rounded: 429 of 429 rows for 133, 252 of 252 for 136, and the same for 134/135 except the tie below. Rounding fl32(x * scale) misses 32 to 78 rows per format.
• NaN of either sign and signalling NaN store 0. -0, negative values and -inf store 0. +inf and values ≥ 1 store the field maximum. No bits outside the field are touched.
• The one exact tie, 0.5 into a 1-bit field, stores 1 (the A bit of 134 and the R bit of 135, with f32 0x3f000000 and with d16 0x3800). This is 2 of 354 rows per format.

So NormStoreBits' round-half-to-even differs from the hardware on that tie: it stores 0, the GPU stores 1. With f32 or f16 inputs, a 1-bit field at exactly 0.5 is the only tie that can occur: a tie needs x = (2k + 1) / (2 (2^n - 1)), which is a binary fraction only for n = 1. So rounding exact ties up would only change that case. The test's StoreField uses std::nearbyint, which rounds half to even too, and StoreValues contains 0.5, so the test expects 0 for that W bit as well.

This is desktop RDNA 2, not a PS5. I haven't run the PR's Vulkan tests here.

[DavidSamir]

First run on AMD's Windows Vulkan driver. It shows a different failure on main than on NVIDIA and lavapipe.

Setup: Windows 11 Pro 26200, AMD Radeon 680M (RDNA 2 iGPU, 1002:1681), AMD driver 32.0.21030.2001 (Vulkan 1.4.315), WinLibs GCC 15.2.0 r7, Release. main 1ff5c8ac5. This PR's head 4adf40588 merged into it cleanly. Khronos validation layer from SDK 1.4.363 (VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation), plus ANYPS5_REQUIRE_VULKAN=1 and ANYPS5_NO_SHADER_CACHE=1. The command is the PR's own nine-case ctest -R '^agc_driver_(image_(converted_unorm|packed16_load|store_sint|sint8|repacked_float|packed16_storage_(133|134|136))|fmask_image_reads)$' -V -j1. On the main side, main's implementation is built with only this PR's ImageConvertedUnorm.cpp and its test registration, so both sides run the same cases.

This driver reports optimal-tiling STORAGE_IMAGE for R5G6B5, B5G6R5, A1R5G5B5, R5G5B5A1, B4G4R4A4 and R4G4B4A4_UNORM_PACK16 (vulkaninfo --show-formats). So main does not refuse these formats here, as it does on the RTX 3090 and lavapipe. It uses them, and the result is wrong:

Case	main	this PR
agc_driver_image_packed16_storage_133	fails: packed image load/store changed guest bytes for format 133	passes
agc_driver_image_packed16_storage_134	fails: ... for format 134	passes
agc_driver_image_packed16_storage_136	fails: ... for format 136	passes
the six existing cases	pass	pass
total	6 of 9	9 of 9

No VUID- lines on either side, and no was not found but was requested by env var VK_INSTANCE_LAYERS.

On AMD's Windows driver, main doesn't stop on these formats the way it does on NVIDIA. It writes the wrong guest bytes without any error, so a title would render wrongly instead of stopping. This PR's route through the R16_UINT view gives the bit-exact round trip on this driver too.

AI-assisted: yes

[oneandonlydean]

@harry-z-kang a useful data point from @DavidSamir: on AMD's Windows driver, which reports storage support for these packed 16-bit formats, main silently writes the wrong guest bytes for formats 133, 134 and 136, and this PR's R16_UINT route round-trips them exactly. Thanks @DavidSamir; that makes this a correctness fix on that driver, not just a refusal-to-support change. The gate on current main is next.