[SleglJan]

What

v_cubeid_f32, v_cubesc_f32, v_cubetc_f32 and v_cubema_f32 refuse any output modifier (mul:2, mul:4, div:2) with VOP3 source modifiers are not implemented. PPSA11386 (007 First Light) stops there on current main, within seconds of start, on v_cubeid_f32 v, v, -v, -v div:2.

Measured on RDNA2 (Ryzen 9 7950X3D iGPU, gfx1036, not the PS5's gfx1013) with tools/hw-oracle, 2048 inputs (edge cases and random), all four instructions with each modifier, in all 16 combinations of IEEE mode, the four f32 denormal values and f16/f64 denormals flushed or kept, with round to nearest even and DX10_CLAMP=1:

• With IEEE mode off the modifier is applied in all eight denormal combinations, f32 output denormals kept included, while v_mul_f32_e64 ... div:2 in the same kernels follows fix(shader): apply the output modifier only in the float modes where the hardware does #1851's rule (ignored with f32 output denormals kept). With IEEE mode on it is ignored.
• Where it is applied, the result is main's bit-level f32 output modifier result on every row: of the 24,576 results per mode, 21,768 differ from the unmodified value and all of them match; NaN and infinity pass unchanged, denormal and zero results become +0, div:2 below 2^-125 gives a signed zero, mul overflow a signed infinity. clamp applies after the modifier. The negated-source form 007 uses behaves the same.
• With fix(shader): quiet signaling NaN cube results in IEEE mode #2043's quieting of signaling NaN results (IEEE mode on only) and the modifier (IEEE mode off only), the two never act on the same result; composed, they match the measurement on all 524,288 results of the 16 modes, 1,088 of them signaling NaNs.

Changes:

• The decoder accepts the output modifier on the four instructions.
• The f32 output modifier scaling moves into scaleF32Result. floatCube applies it whenever IEEE mode is off and then writes with the modifier cleared, so clamp is still applied after it.
• floatCube does not use the generic gate (outputModifierApplies), which drops the modifier with IEEE mode off and f32 output denormals kept; for these four the hardware applies it there. In the usual mode (IEEE off, f32 denormals flushed) the two agree.
• The refusal check in Vop3CubeClamp.cpp becomes an execution test, and its run and check helpers take the kernel, rows and names instead of being duplicated. The output modifier entry in docs/dev/debt/shader.md records the cube rule.

Only round to nearest even is modelled. The same 2048 rows in the other three round modes (IEEE mode off, f32 denormals flushed and kept) differ from round to nearest even where a result overflows to infinity: the hardware rounds the overflow as the mode says (to the largest finite value when rounding toward zero), in v_cubema_f32's own result as well as in mul:2 and mul:4. The translation gives infinity there, as main already does for v_cubema_f32 without a modifier. The only other difference is the sign of three zero v_cubema_f32 results with a NaN input when rounding toward negative infinity. The debt entry records both.

Overlap: #2463 (closed by the intake pause) accepted the modifier with the generic gate.

Tested

• Linux (CachyOS, kernel 7.2.9, GCC 16.2.1), Release with BUILD_TESTING=ON. The measurements below were made with this branch on main 23e93c0; it was then rebased without conflicts onto main 145cb82, where tools/check_conventions.py --base upstream/main and git diff --check pass and the 19 vop3 execution tests, agc_driver_vop3_cube_clamp included, pass on lavapipe.
• Full ctest on lavapipe (Mesa 26.2.4, LLVM 23.1.1; Arch Linux container, GCC 16.2.1) with ANYPS5_ENABLE_SPIRV_TOOLS=ON as in CI: 1107/1121; main 23e93c0 built the same way fails the same 14 (hardware_oracle, guest_syslog, guest_allocator_failures, agc_driver_bc_block_copy, agc_driver_depth_compare_copy, agc_compute_pipeline_replay, autorun_exit_status, guest_none_relocations, guest_module_args, linux_executable_exports, linux_entry_argv, linux_program_header_order, linux_input_program_headers, linux_guest_stack).
• Full ctest on the RTX 3080 (NVIDIA 615.71.09): no failures out of 1113; 8 of them skip themselves (seven VideoOut tests and agc_driver_stencil_ref_export).
• agc_driver_vop3_cube_clamp: a 16-instruction kernel (all four instructions with each modifier, the negated-source form, clamp with each modifier) over 32 rows (two of them with a signaling NaN input), compiled with no float mode, IEEE=0 with f32 denormals flushed (0xc0) and kept (0xf0), and IEEE=1 with both, against two tables captured on gfx1036 (IEEE=0 and IEEE=1; the hardware gives the same results with f32 denormals flushed or kept). With only the test applied to main 23e93c0 it fails on the decoder refusal. Mutations caught: the generic f32 gate (fix(shader): accept cube output modifiers #2463's behaviour; fails at IEEE=0 with f32 denormals kept), applying the modifier in IEEE mode, never applying it, and skipping the signaling NaN quieting when a modifier is present (fails on a signaling NaN row with IEEE=1).
• PPSA11386 v01.018.000 in a bubblewrap sandbox from a cold shader cache, 3 runs each, no environment overrides: on main 23e93c0 every run ends on the decoder refusal (terminate with VOP3 source modifiers are not implemented). With this change every run gets past it, compiles about a thousand shaders and stops at its first presentation on bindless image table has 18960 entries (0 materials), limit 16. Along the way 63 to 72 shaders are logged as not prepared after registration; none of the reasons names a cube instruction or an output modifier (GetImageResource dword 0 is not a valid runtime value, the clip distance limit, static bindless image capacity, bindless storage image tables).

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
• Unblocks: PPSA11386 007 First Light (PPSA11386) #5714
• GPUs tested: NVIDIA RTX 3080, driver 615.71.09, CachyOS (Linux 7.2.9); lavapipe (Mesa 26.2.4, LLVM 23.1.1), Arch Linux container; the reference tables come from an AMD Ryzen 9 7950X3D iGPU (gfx1036, RDNA2) through tools/hw-oracle
• AI-assisted: yes (Claude Code)

🤖 Generated with Claude Code

[jotitadev]

Run on an Apple GPU, as a third device next to the RTX 3080 and lavapipe.

GPUs tested: Apple M2 Max, MoltenVK 1.4.2 (Vulkan 1.4.357), macOS 27.2, x86-64 under Rosetta

main does not run the GPU tests on macOS yet, so I tested on macos-port at b7eb6b6b (mugurc/AnyPS5) with this PR's diff applied, not on main. Release build, LunarG Vulkan SDK 1.4.363.0. The suite is ctest -R agc -E 'flip|display|lds_atomics64' -j1 --timeout 30. Without the PR, 22 tests fail on this branch, all of them known macOS failures listed in #1. Head 2acac0828. One conflict, in RdnaVectorOpDecoder.cpp (supportsNativeVop3ResultModifiers): the branch doesn't have isNativeVop3F16TernaryOpcode in that condition yet, so I kept only this PR's || isCubeOpcode(opcode). Everything else applies cleanly, except docs/.

• agc_driver_vop3_cube_clamp passes. With only the test file applied to the unchanged branch, it fails with VOP3 source modifiers are not implemented, so the new cases do run on this device.
• Suite: the same 22 failures with and without the PR. None are new and none are fixed. agc_driver_f32_output_modifier is among them on this branch without the PR, and the cube modifier cases here still match the gfx1036 tables.

Written with AI assistance (Claude Code). I ran the builds and tests above on my machine.

[github-actions]

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 4072	core/shader/recompiler/Translation/include/Translation/TranslationContext.hpp core/shader/recompiler/Translation/src/FloatInstructions.cpp
PR 5021	core/shader/recompiler/Translation/src/FloatInstructions.cpp docs/dev/debt/shader.md
PR 5371	docs/dev/debt/shader.md
PR 5426	docs/dev/debt/shader.md
PR 5500	docs/dev/debt/shader.md
PR 5941	docs/dev/debt/shader.md
PR 6009	docs/dev/debt/shader.md
PR 6235	docs/dev/debt/shader.md
PR 6280	docs/dev/debt/shader.md
PR 6412	docs/dev/debt/shader.md
PR 6488	docs/dev/debt/shader.md
PR 6977	docs/dev/debt/shader.md
PR 7085	docs/dev/debt/shader.md
PR 7395	docs/dev/debt/shader.md
PR 7484	docs/dev/debt/shader.md
PR 7684	docs/dev/debt/shader.md
PR 7888	docs/dev/debt/shader.md
PR 7922	docs/dev/debt/shader.md
PR 8004	docs/dev/debt/shader.md
PR 8077	docs/dev/debt/shader.md
PR 8267	docs/dev/debt/shader.md
PR 8281	docs/dev/debt/shader.md

[Scramblrrrr]

Windows cross-test on current main 4524944: author-preserved e32f5e1 port at df0e8a7, followed by #6317 and #6332, integration head 00abe7e. GCC 15.2 / Release / SPIRV-Tools OFF / RTX 5070 Ti, Vulkan required. Fresh patched libraries and fixtures built successfully.

agc_driver_vop3_cube_clamp PASS 0.66 s, exercising the included oracle expectations with IEEE on/off and f32 denormals flushed/kept. Seven adjacent controls (shader memory, SPIR-V module, graphics, shared writes, write tracking, position invariance, constant phi) also pass: 8/8 total 6.24 s, no skips. This is a combined-stack Windows cross-test, not a full-suite result or new independent hardware oracle. Source/conventions clean; public integration branch: https://github.com/Scramblrrrr/AnyPS5/tree/codex/first-light-oct16-current-main .

007 is currently preparing a cold shader cache without the earlier cube stop; no rendered-gameplay or FPS claim yet. AI-assisted: Codex; title data stays local.

[oneandonlydean]

The Windows run (@Scramblrrrr, RTX 5070 Ti, agc_driver_vop3_cube_clamp plus seven adjacent controls) and the Apple M2 Max run (@jotitadev, red without the PR and green with it) are a third and fourth device next to the RTX 3080 and lavapipe. Both are on a port or a combined stack, not this head on current main, so we'll run it there.