fix(recompiler): emit a formatted buffer access through one function per resource

[Raghav314]

What

Since #639 a shader is compiled once to a generic module and specialized per descriptor at its first draw or dispatch. A formatted buffer load or store whose format is a specialization constant (a prepared descriptor) emitted the switch over all 77 buffer formats inline at every access, so a shader with many formatted accesses produced a very large generic module, and specializing it at first use took long.

• The switch is now emitted once per resource, component count and direction, as a SPIR-V function marked not to be inlined (like the recompiler's BDA helpers) that takes the index, the offset, soffset, the record count and the storage buffer offset (and the data for a store). Each access calls it. The descriptor fields it needs are the same specialization constants as before.
• Specialization still folds the switch to the descriptor's format; nothing changes in what an access computes.
• Accesses through GPU-selected descriptors, and typed accesses (format known at compile time), are unchanged.
• SpirvModule gets a second function section (BeginHelperFunction/EndHelperFunction) that Finalize appends after the entry point.

On the console the buffer unit decodes the V# and does the addressing for each access in hardware, so one implementation serves every access; this emits the addressing once per resource in the same way.

PPSA36082

One compute shader has 16 formatted loads from one descriptor and 8 formatted stores to another. Its generic module was 2,851,193 words, with 99.8% of its instructions inside the 24 inlined switches, and its specialization took about 200 ms on the queue worker at about 3.6 s, while the loader waits on a GPU fence before it frees its 512 MiB upload pool. With this change the generic module is 245,872 words, the specialized one 51,338 (was 681,294), and the specialization takes about 16 ms.

On main f145139 with #861, #1367, #1249, #1905, #1961 (import at map) and #1960 (imports leave pages unprotected) it reaches the main menu in 10/10 runs (40 s each); before this change, on the same stack without #1960, 0/4: the pool overflows at about 3.7 s; two stack snapshots and a perf profile of failing runs put the queue worker in SpecializeSpirv for this shader.

Tested

Linux (CachyOS), x86-64, GCC, 7800X3D / RTX 4090.

• agc_formatted_buffer_once (new, no device): against a shader without formatted loads, one formatted load adds one call and one function; two more loads from the same descriptor add two calls, no function, no format switch and under 10% to the module. It passes with and without ANYPS5_ENABLE_SPIRV_TOOLS (with it, the specialized switch is folded away). A formatted store to a second descriptor gets its own function, and two more loads and stores add only calls. The function is marked not to be inlined. Emitting the switch inline at every access, or only at the first, or leaving the function inlinable, fails it in both builds.
• agc_driver_buffer_format, agc_driver_format_d16 and agc_driver_format_d16_store run the new function on the device (loads and stores, 1 to 4 components, d16) and pass.
• spirv-val (Vulkan 1.3) on the 37 modules these tests produce: all valid.

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

[MarcZX10]

Linux CI finished with 497/498 tests passing. The new agc_formatted_buffer_once test fails at its first assertion: it emits 0 format switches for 3 loads, while the test expects one shared switch. This suggests the fixture is not exercising the generic prepared-descriptor path described in the PR, or the expected SPIR-V shape is wrong. Please fix the fixture or assertion and rerun the full checks; I will re-review the updated head.

[Raghav314]

Thanks, The fixture was fine, my assertion was the problem. CI builds with SPIRV-Tools on, and there the specialized module goes through dead-branch elimination, which folds the format switch away completely. So it ends up with 0 switches whether there's 1 load or 3. I only ran it locally without SPIRV-Tools.

The shared function is still there either way (one call per load, same number of functions), so I changed the test to check that instead: 3 loads shouldn't add any switches or functions compared to 1 load, just one call each. It passes with and without SPIRV-Tools now, and if I put the inline switch back it fails in both.

[MarcZX10]

All checks now pass, but the updated test still permits the original failure mode: switches <= 1 accepts zero switches. It also compares only one-load and three-load outputs, so equal function counts do not establish that either output contains the shared helper; the two additional calls alone do not prove the first load took that path. Please add a zero-load baseline and assert that one dynamic formatted load emits the expected helper and format dispatch, while three loads add two calls but no additional helper or dispatch. Then rerun the checks. I will re-review the result.

[MarcZX10]

Thanks for clarifying that SPIRV-Tools folds the format switch away; I withdraw the requirement to assert a switch. One coverage gap remains: comparing only 1-load and 3-load modules proves the two extra loads add two calls, but does not prove the first load itself calls the shared helper. Please include a zero-load baseline and assert that 1 load adds one call, then 3 loads add two more, with no extra helper functions. That will verify every load uses the shared function both with and without SPIRV-Tools.

[Raghav314]

Thanks again. I added a baseline with no formatted loads, so now one load has to add exactly one call and one function, and three loads only add two more calls with no extra function or switch. Passes with and without SPIRV-Tools.

I also checked it actually catches the thing u pointed out, made only the first load emit the switch inline and the rest call the helper, and the test fails in both builds ("one formatted load emitted 0 calls and 1 functions more than none").

[MarcZX10]

The zero-load baseline closes the remaining coverage gap: one load must add one call and helper, and two more loads add two calls without another helper. The reported inline-first-load mutation fails this assertion, which checks the intended regression directly. I re-reviewed the full diff; Linux, Windows and Conventions pass on this head. Approved.

[Raghav314]

The relinker (windows-latest) failure here isn't from this PR: main fails the same way at 3bd74c1. hardware_oracle now runs AssemblyTests from #1962, which only skips when clang/ld.lld are missing. The Windows runner's clang is there but has no AMDGPU target ("unknown target triple 'amdgcn-amd-amdhsa'"), so it errors instead of skipping. Linux and macOS pass.

[popcoder565]

Native Windows run, since no Windows results were posted. Setup: Windows 10 (19045), RTX 3060 Laptop (NVIDIA 616.92), MinGW-w64 GCC 15.2.0 (WinLibs r7), Release. Head f5e118aa, merged onto main bc40a6b together with #1960 and #2124.

• This PR's tests pass: agc_formatted_buffer_once.
• Full ctest: 609 of 609 in two of three runs. In the other, amd64_only_windows (a relinker test) segfaulted once under -j 8; it then passed 10 of 10 on its own.

No title was run.

🤖 Addressed by Claude Code

[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
#1640	core/libs/prx/libSceAgcDriver/CMakeLists.txt L857

[adriadam10]

Checked on Linux (RADV): agc_formatted_buffer_once passes and fails without the change; agc_driver_buffer_format, agc_driver_format_d16, agc_driver_format_d16_store, agc_driver_formatted_store, agc_driver_tbuffer_format_d16 and agc_driver_bda_device pass. The function key covers resource, component count, direction, d16, coherent and kind, and the call site adds memory.offset as before. Native Windows run by @popcoder565 also passes the new test. Merging.