[milescoviello]

What

Crimson Desert (PPSA25286) stops at its first draw:

[videoout] presentation failed: AGC graphics: fragment shader must export float4 colors to its attachments
terminate called after throwing an instance of 'std::runtime_error'

A throwaway print at the refusal in ValidateShaders (ShaderValidation.cpp) shows the draw:

[fsexport] location 0 signature vertex:u32x4 attachments 1 blends 0 outputs: 0:vertex:u32x4 colors:

The draw binds no color target (no blends, no colors), probably a depth-only pass. Its pixel shader still exports a uint4 at MRT0. The check goes over every output below max(blends, 1) and requires f32x4, or u32x4 for integer and packed targets. At location 0 it finds no target, so it falls back to f32x4 and refuses the draw.

A location without a color target has no attachment in the pipeline. With no blends, colorAttachmentCount is 0, and a gap between written slots is VK_ATTACHMENT_UNUSED. Vulkan discards an output that has no attachment, whatever its type. The check now applies only where a color target exists. The location still goes into the returned output set as before, so masks and pipeline keys of draws that already ran don't change.

Tested

Linux (Ubuntu 24.04, GCC 13.3, Release), RTX 4070 Ti, NVIDIA 595.99.

• agc_driver_graphics: the packed-export shader there exports a uint4 at location 0. It's now validated against a state with no color targets and no blends, and against one with a target at slot 1 and slot 0 unused, the gap a draw can actually have. (@jotitadev pointed out that my first unused-slot state, no colors with one blend, is one the pipeline refuses before validation.) Both are accepted, and validating it against a float target still fails with float4 colors. With main's ShaderValidation.cpp the test fails with the title's line, fragment shader must export float4 colors to its attachments. It passes 3 of 3 with this change.
• ctest (^agc_driver_|^agc_vertex_barrier|recompiler|shader_cache|request_serializer|spirv, 468 tests): 467 pass. agc_driver_store_over_untracked fails the same way on main 5d9dcc4 here; test(agc): keep waiting for the neighbour copy case's result #6385 addresses it.

Crimson Desert, scripted pad input, with #6612, #6613, #6706 and #6178 (for memory), which it needs to reach this draw:

Build	Result
without this PR	stops at its first draw, 597 s in, on the refusal above
with this PR	gets past it with no stop: 22 draws by 591 s, while it is still building pipelines (1 to 3 s per draw), until my 16 GiB memory cap ends the run

Nothing is presented by then, so I haven't seen this draw's result. The test above covers the validation.

Validated on Windows 11 with an AMD Radeon 860M (RDNA 3.5, driver 32.0.13058.8001) by @LeandroLP23 (comment), in one build of main 6efb726 with this head and 13 other pull requests: agc_driver_graphics and agc_driver_store_over_untracked pass 3 runs of 3. 713 of 716 agc_* tests pass; the 3 that fail fail the same way on main there.

Not checked on Intel or on a console.

• Khronos validation layer (LunarG SDK 1.4.363), NVIDIA RTX 4070 Ti (595.99.02), Ubuntu 24.04: VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation ANYPS5_NO_SHADER_CACHE=1 ANYPS5_REQUIRE_VULKAN=1 ctest -R '^agc_' -V on main e2a2914 and on main + this PR (3d5335a, a local merge). main: 815 of 815 pass; with this PR: 815 of 815 pass. Neither log reports the layer missing. VUIDs on main: VUID-VkImageViewCreateInfo-usage-02275 ×6, VUID-VkPipelineColorBlendStateCreateInfo-pAttachments-00605 ×18, VUID-vkCmdCopyBuffer-srcBuffer-00118 ×2, VUID-vkCmdDraw-multisampledRenderToSingleSampled-07285 ×2, VUID-vkCmdWaitEvents-srcStageMask-01158 ×2, VUID-vkSetEvent-event-09543 ×2; with this PR: VUID-VkImageViewCreateInfo-usage-02275 ×6, VUID-VkPipelineColorBlendStateCreateInfo-pAttachments-00605 ×18, VUID-vkCmdCopyBuffer-srcBuffer-00118 ×2, VUID-vkCmdDraw-multisampledRenderToSingleSampled-07285 ×2, VUID-vkCmdWaitEvents-srcStageMask-01158 ×2, VUID-vkSetEvent-event-09543 ×2. None is new with this PR.

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 (Crimson Desert reaches this draw only with feat(agc): draw tessellation followed by an NGG geometry stage #6612, fix(shader): emit nothing for a vertex s_barrier that precedes every memory access #6613 and perf(shader): emit runtime V# byte writes through the looping helper #6706)
• Unblocks: none
• GPUs tested: NVIDIA GeForce RTX 4070 Ti, driver 595.99, Ubuntu 24.04; AMD Radeon 860M, driver 32.0.13058.8001, Windows 11 (tests, by @LeandroLP23)
• AI-assisted: yes

🤖 Generated with Claude Code

[MarcZX10]

I also found a hole in the targetless case. ValidateShaders still loops only up to max(blends.size(), 1) and validates the export type when no ColorTarget exists at that location. So with colors.empty() but blends.size() == 1 (an unused attachment slot), a uint4 export still hits the fallback f32x4 check. The new test resizes blends but uses a packed export? It should exercise this exact state with a write-enabled blend entry; skipping type validation when find_if(state.colors, exportIndex) returns end fixes it while retaining the location in the returned set. I reproduced the path from the current function; the full graphics test build is currently blocked in unrelated GCC 12 parsing at SrtDescriptorEvaluation.cpp:183.

[MarcZX10]

Correction: my previous comment was a false positive. I checked the PR diff again and its Require(color == state.colors.end() || ...) already skips type validation for a location with no ColorTarget, including when blends has an unused slot. The added tests cover the no-blend and unused-slot cases, so the extra write-mask case I suggested is redundant. Sorry for the noise.

[jotitadev]

Static review of head d5b49fd (no build, no GPU run).

The fix looks right to me: a location with no ColorTarget maps to VK_ATTACHMENT_UNUSED (or no attachment at all) in Pipeline.cpp, so skipping the type check there matches what Vulkan does with that output.

One note on the second new test case. targetless.colors.clear(); targetless.blends.resize(1); is a state the Pipeline constructor rejects:

Require(state.blends.size() == (state.colors.empty() ? 0u : state.colors.back().exportIndex + 1u) && ...,
        "blend states do not match decoded color state");

With no colors, blends.size() has to be 0, so a draw never reaches ValidateShaders with that state's pipeline. The unused-slot shape a real draw can have is a target at exportIndex 1 with slot 0 empty (colors = {exportIndex 1}, blends.size() == 2), with the shader still exporting at location 0. The change covers it too (the find_if misses location 0), but neither case in the test builds it. Swapping the second case for that state would test the gap the description talks about.

Method: read the diff against main 5d9dcc4 and modelled both Requires in a small Python script; nothing was built.

Written with AI assistance (Claude Code). No builds or tests were run for this note; only the standalone snippet above.

[milescoviello]

Thanks @jotitadev, good catch: the pipeline would refuse my unused-slot state before validation. f663675 replaces it with your shape, a target at slot 1 with slot 0 unused and the shader exporting at location 0. It fails on main's ShaderValidation.cpp with the title's line, and passes 3 of 3 with the change.

AI-assisted: yes

[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 1395	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 4040	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 4072	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp
PR 5441	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5499	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5500	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5925	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6146	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6188	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6194	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6235	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6280	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6290	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp
PR 6481	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6548	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6636	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6683	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6808	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6947	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp
PR 6970	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6972	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7085	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7256	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7332	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7642	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp
PR 7668	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7862	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7888	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp
PR 8099	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp

[LeandroLP23]

AMD on Windows: the tests this PR names pass on an RDNA 3.5 GPU.

Radeon 860M (RDNA 3.5, driver 32.0.13058.8001), Windows 11 26300, WinLibs GCC 15.2.0 r7, Release. main 6efb7261b with this PR (3d5335ab3) merged, together with #7296, #7446, #7377, #7445, #6748, #6785, #6706, #6613, #7449, #7391, #7489, #7484, #7514. They were built once, so the result is for the combination; each merge was clean.

• agc_driver_graphics, agc_driver_store_over_untracked: pass, 3 runs of 3.
• Every agc_* test on the combined build: 713 of 716 pass. The 3 that fail, agc_driver_host_import_lifetime, agc_driver_quad_pixel_mask and agc_driver_layered_volume_vertex_no_gpl, fail the same way on unchanged main here, so none comes from these PRs.

AI-assisted

[milescoviello]

Thanks @LeandroLP23 for the AMD run on Windows. I added it to the description.

AI-assisted: yes (Claude Code)

[adriadam10]

Re-checked against current main: the merge is clean, ShaderValidation.cpp is untouched by main since the last review, and the two new test cases (no targets, and a gap below a target) are still in place after main's Graphics.cpp changes. The one-line change skips the export type check where no ColorTarget exists, which matches the pipeline (no attachment / VK_ATTACHMENT_UNUSED), and the location stays in the returned set. The new cases fail on main with the float4 colors message and a float target still fails. Merged.