#313

D3D12 → Vulkan layer: critical path, who builds what, open methods

OpenKoertLichtendonkopened 15h ago

The Xbox D3D12 → Vulkan layer in one place: what's done, who builds what, and what waits for what. It's the busiest area of the project (about 20 items and 10 people), so it gets its own page. Rules and the claim commands are in #1380; claims for D3D- items can be posted here or there. The plan and milestones are in docs/d3d12_vulkan.md.

Updated 2026-11-22 12:05 UTC by the coordinator (@KoertLichtendonk); corrections are welcome as comments.

Critical path

Most layer work builds on these PRs. Reviewing and testing them first (Windows, MSVC and a Vulkan GPU) unblocks the most people.

  1. #119 (@Ryx69): the Vulkan-backed device, copies, presentation and the base of the compute path. Partly merged; its newer commits (head 5796549) are what most layer PRs below build on.
  2. #291 (@Ryx69): compute path and graphics draws, milestone M1 (D3D-3, D3D-4). Depends on #119; about +3,900 lines in 23 files.
  3. Built on #119/#291: #286 heaps and placed resources (D3D-16), DXGI shim in Ryx69/Leonida#2 (D3D-2), device wiring of the bindless descriptors (D3D-5; #230 merged), copy and compute queues (#306, D3D-21), mesh shaders (D3D-29), render passes (D3D-27).

Up for grabs

Unclaimed D3D- items, most useful first. ⭐ = good first task. None of them needs game files.

IDWhatWhy it mattersNeedsSize
D3D-165d3d12_present fails on Intel UHD 770: Present returns VK_ERROR_OUT_OF_DATE_KHR, then -1, and the readback is wrong (came with the presenter commits e876fc83, 147f9300) (found by @kapitche) (see finding)presenting is how the game shows frames; NVIDIA passes, so the swapchain recreate path is likely missingan Intel GPU; C++; no game filesS
D3D-27Render passes: BeginRenderPass/EndRenderPass on dynamic rendering (earlier: @Ryx69; no code found)BeginRenderPass/EndRenderPass on top of #291's graphics pipelines, now on main via #1083C++; a Vulkan GPU; no game files; @Ryx69 claimed it before #291 landed and has first refusalM
D3D-29Mesh shaders on the layer: CreatePipelineState (pipeline state streams) and DispatchMesh (earlier: @Ryx69; no code found)mesh-shader PSOs (pipeline state streams) and DispatchMesh, now that #291's PSOs are on main via #1083C++; a GPU with VK_EXT_mesh_shader; no game files; @Ryx69 claimed it before #291 landed and has first refusalM

Who builds what

IDItemOwnerPRStatus
D3D-2DXGI swapchain shim as a thin layer over the presenter@Vyral1978—rescoped on top of #742 (merged): a dxgi.dll shim table so the game reaches the layer's factory, and a factory before the device; no code until agreed
D3D-3Compute path: shader lookup by FSSM hash, root signatures, GPU address arena, compute PSOs, Dispatch@Ryx69, @TheHoracio, @AlxcBTW#291 merged, #819, #831, #848, #873, #1489 draft, #907#291 with D3D-4; mergeable; the split is agreed in #313: #291 lands first, set-0 hookup follows from @aslan2142 (Depends on: #291, #328)
D3D-8CBV/SRV/UAV/sampler views, GetGPUVirtualAddress, command list state setters@0xTehuti#157PR open (conflicts); per @aslan2142, its buffer V# code fits #328's buffer branch; #334 also writes CBV/SRV/UAV views
D3D-17ExecuteIndirect with simple command signatures (draw, draw-indexed, dispatch, mesh dispatch, count buffers)@mrb-eel#317, #743 merged, #1339 mergeddraft #317 (@mrb-eel): fixed-count indirect dispatch; @totonicpu's #743 adds count buffers (not claimed); @Vonsek: on #291's merge candidate, a graphics ExecuteIndirect reaches replay without the draw state and without the buffer's suballocation offset (finding)
D3D-21Copy and compute command queues (D3D12_COMMAND_LIST_TYPE_COPY/COMPUTE) with cross-queue fences@IAmNotSiob#306PR open
D3D-32Variable rate shading: RSSetShadingRate/RSSetShadingRateImage on VK_KHR_fragment_shading_rate@ezwinxdlmao, @0xTehuti#469, #519#469; @0xTehuti closed #519 for it, tested and approved #469; waits for #291
D3D-34Pipeline library and shader cache: CreatePipelineLibrary, CreateShaderCacheSession, ShaderCacheControl on VkPipelineCache@mrb-eel#367 draftdraft #367: use the pipeline cache the build produces (cache seeding only)
D3D-38Per-type descriptor slot sizes and a heap-handle registry (GetDescriptorHandleIncrementSize, per-type stride in CopyDescriptors)@kontaktjakubjagiello#333PR open
D3D-39CBV/SRV/UAV views as GFX10 descriptors (#334) and a bindless index allocator with descriptor-table binding commands (#335, draft), stacked on #333@kontaktjakubjagiello#334, #335 draftoverlaps #157, #328 and #291; @Ryx69 invited a review of #291 against #335 and will take first_draw_test if it adds coverage
D3D-45Compare clears with the host runtime: ClearRenderTargetView across UNORM/SNORM/SRGB/FLOAT/UINT/SINT formats, ClearDepthStencilView depth/stencil, with and without rects@NWX10, @luisbaker#521, #525 merged#521 (@NWX10) and #525 (@luisbaker); @NWX10 proposes keeping #521 and asks for an Intel run of it
D3D-47Compare resolves with the host runtime: ResolveSubresource/ResolveSubresourceRegion per format and sample count@ToxicmaggotRP#493 closed#493 closed by its author (no merge); @ToxicmaggotRP posted done: unclear
D3D-48Compare fences and queues with the host runtime: Signal/Wait/GetCompletedValue/SetEventOnCompletion, cross-queue waits, signalling an already-passed value@luisbaker#497PR open
D3D-49Same errors as the host runtime for invalid arguments: CreateCommittedResource, CreateDescriptorHeap, CreateCommandQueue, CreateFence, Map/Unmap ranges and heap types@omar-akermi#558draft; 157 cases against the host on an RX 9070 XT, an integrated Radeon and WARP
D3D-50CheckFeatureSupport: same HRESULTs as the host for wrong structure sizes and unknown features (answers may differ by GPU)@luisbaker#459PR open
D3D-51Command-list rules like the host runtime: Close twice, Reset while recording, executing an open list, reusing an allocator in flight@Aminecsmc#495#495; @NWX10 dropped their branch, will review and suggest the undefined-list-type case
D3D-52GetResourceAllocationInfo with a host comparison test (again)@ray3242#450 closed, #517 closed, #522 closed, #526 closed, #631 closed#327 already implements these three methods with a host comparison and is merged (D3D-23)
D3D-53M1.5: draw a translated RDNA2 VS+PS through the layer (PSO looked up by hash), follow-up to #335@AlxcBTW#699 closed, #907draft #699: the first draw with translated VS and PS (Depends on: #291)
D3D-54Synthetic draw from an UPLOAD-heap vertex/index buffer and an upload-backed descriptor: check the -heap_base bias (open question 17)@VictorJulianiR#653#653; @woreksami closed #648, tested #653 on an RTX 3060 and offered its descriptor-table case
D3D-55Arena high-water mark: log GetGPUVirtualAddress results and descriptor bases, print the peak at device release (open question 5)@ismailchaka1#612#612; @0xTehuti closed #642
D3D-57Primitive restart: the PSO's IBStripCutValue on dynamic primitive-restart state@ezwinxdlmao#572#572; Depends on: #291, #469
D3D-58Wave64/wave32 subgroup size control in PSOs (required subgroup size per stage, full subgroups for compute)@AlxcBTW#609 closeddraft #609 (Depends on: #291)
D3D-631D textures and 1D arrays: CheckFeatureSupport reports TEXTURE1D support, but CreateCommittedResource refuses the dimension@Ayinaki#561 closed, #693PR open
D3D-68B5G5R5A1_UNORM sampled textures and copy footprints (split from D3D-61)@VictorJulianiR#617 closed#617, stacked on #562
D3D-69ClearRenderTargetView on a TYPELESS resource through a view of another class (UINT/SINT views) ignores the view's format@NWX10#621#621, stacked on #521 and #589
D3D-71Default-heap buffers are not zero-filled yet: D3D12 committed resources start zeroed unless D3D12_HEAP_FLAG_CREATE_NOT_ZEROED@ismailchaka1#655#655 (@ismailchaka1, Depends on: #291); @SpositoDouglas also claimed it
D3D-73Custom sampler border colours on VK_EXT_custom_border_color (transparent black today)@0xTehuti#760#760 (@0xTehuti)
D3D-74Min/max reduction samplers on samplerFilterMinmax (they blend today)@luisbaker—claimed in #313 (@luisbaker, 01:27); @guimsolutions also claimed it (01:30)
D3D-78Unordered-access views of 3D, 1D and 1D-array textures (only 2D and 2D-array today)@guimsolutions#773PR open
D3D-79Depth-stencil views of 1D and 1D-array textures (after #561 adds 1D textures; the 2D MS array part is D3D-83)@ismailchaka1#684 draft#684 (@ismailchaka1, draft, Depends on: #561); @lordnyx0 closed #703 and released it
D3D-85Castable formats: answer D3D12_OPTIONS12 (RelaxedFormatCastingSupported) and accept castable-format lists on mutable-format Vulkan images (refused today)@Ayinaki#693, #561 closed#693 (committed resources); the placed half is D3D-90
D3D-86CheckFeatureSupport: OPTIONS2–5, 7, 9, 11, 13–16, 18–21 (the rest is split off), compared with the host on the same adapter@JavierBlanco23, @Proudnoob4, @AlxcBTW#775 merged, #745 merged, #469, #715 merged, #693, #706 merged, #816 merged, #855 mergedPR open
D3D-87B8G8R8X8_UNORM/_SRGB/_TYPELESS (split from D3D-61)@SpositoDouglas#759PR open
D3D-88Aliasing: placed textures that share a heap range, and aliasing barriers (open question 18)@BACONMANXDGOD#770#770 (Depends on: #286); 14/14 answers match native D3D12 on WARP, NVIDIA and Intel
D3D-99Keep root arguments when the same root signature is set again (a #291 review finding; a PR against #291's branch)@lahlousaad4—claimed in #634
D3D-102D3D-70 follow-up regression coverage for RGBA8 typeless RTV subresources (merged #630).@73N1X—claimed in the tracker (mentions #630)
D3D-103new D3D item on #291's pipelines — DepthClipEnable = FALSE is logged as "not modelled (depth clamp)" and ignored, so geometry beyond the near/far planes is clipped wher@ismailchaka1#813PR open
D3D-106Runtime: stage block-compressed copies one vkCmdCopyImage cannot express@adonizm#860PR open
D3D-107new TEST item, a CPU benchmark of the D3D12 → Vulkan replay: how long the layer itself takes per command to record, check (prepare) and write (emit) a frame's command@traeb0n#878 merged, #893, #901PR open
D3D-108eight D3D12 GPU tests are built but never registered with ctest, so ctest (and CI) never runs them: d3d12_clear_test, d3d12_com_test, d3d12_descriptor_test, `d3d1@quackottu#883 closedPR open
D3D-112sampled color Texture2DArray SRVs ignore ResourceMinLODClamp on main 067d3e8.@Vonsek#974 closedPR open
D3D-114d3d12: map the six narrow signed IA vertex formats (Depends on #291)@dea7hsmiless#899PR open
D3D-132D3D12_QUERY_HEAP_TYPE_PIPELINE_STATISTICS1 (mesh and amplification counts) when VK_EXT_mesh_shader's meshShaderQueries is available@quackottu—claimed in the tracker
D3D-136Console-order command-list proxy (docs/d3d12_command_objects.md §3.5a): device slots 12/51 wrap the PC list in a proxy with the measured vtable order@quackottu#1325 draft, #1270 mergedopens as a draft with Depends on: BUILD-231's PR (@AimLogin's queue/fence proxies)
D3D-149depth ↔ colour copies on the Vulkan layer (CopyResource / CopyTextureRegion between D32_FLOAT ↔ R32_* and D16_UNORM ↔ R16_*), which host D3D12 and WARP perfor@LloydNicholson#1324claimed in the tracker
D3D-162Occlusion resolves during an active pipeline-statistics query: replay refuses them because its internal dispatch would change the compute-invocation count@hadagalberto#1519PR open
D3D-163CopyTiles on the command list (tiled resources), not implemented@quackottu—claimed in the tracker
D3D-164RSSetDepthBias (ID3D12GraphicsCommandList9, the console RSSetDepthBiasX's public form per docs/d3d12_xs_surface.md §5.1) as dynamic depth-bias state on the Vulkan lay@abilliontokens#1451PR open
D3D-166M2 synthetic gate (docs/d3d12_vulkan.md §6).@AlxcBTW—claimed in the tracker
D3D-169xbox_present fails under LEONIDA_VK_VALIDATION=1: the validation layer checks the sRGB swapchain with an implied STORAGE usage that runtime/src/vulkan_present.h never requests@bobcatchris#1524#1524 (@bobcatchris): skips only when the NVIDIA swapchain VUID is the sole validation error; needs an NVIDIA run
D3D-171CheckFeatureSupport(FORMAT_SUPPORT): item 3 of @zneel's RX 9070 XT comparison with native D3D12@zneel#1511PR open
D3D-17213 DXGI formats native D3D12 creates are refused by the layer (R16_UINT, R8_UINT/SINT/SNORM, R16G16_UNORM/SNORM, B8G8R8X8_UNORM(_SRGB), A4B4G4R4_UNORM, ...): missing from vulkan_format() (d3d12_vulkan_device.cpp:84)@Dom-in, @bobcatchris#1493PR open
D3D-174D24_UNORM_S8_UINT/R24G8_TYPELESS on AMD: FORMAT_SUPPORT returns Support1 = 0 and MSAA quality 0, though creating them works through the D24-to-D32 fallback (#737); use the same fallback in FORMAT_SUPPORT@hosienzade45#1513PR open
D3D-178AMD-format BVH builder for the Vulkan layer (docs/d3d12_vulkan.md §4.8, path 2): a CPU reference encoder and validator for the box16/box32/triangle node layout that `im@romainveurrier#1500PR open
D3D-179RGB9E5 multisampled resource creation and resolve on devices that expose the native color attachment.@Vonsek—claimed in the tracker (mentions #1450)
D3D-180Runtime: PresentX ImmediateThresholdPercent — a slightly late frame tears instead of waiting a vblank (VK_KHR_swapchain_maintenance1)@AlxcBTW#1523PR open
D3D-181ExecuteIndirect with MaxCommandCount 0 makes the command list fail Close (E_INVALIDARG) on the Vulkan layer, while the host gives S_OK on an RTX 5070 and WARP.@quackottu—claimed in the tracker

Comparison tests

Once a method exists, the next step is checking it against the host PC runtime (or WARP), as GetCopyableFootprints, GetResourceAllocationInfo and the UINT clears already are: the same answers, the same errors, the same pixels. Those comparison items are in Up for grabs above; they need Windows and a GPU, no game files.

Method coverage

D3D12 methods of the interfaces the runtime implements, counted from the SDK's d3d12.h against STDMETHODCALLTYPE overrides in runtime/ on main and in open PRs. "Open" methods log as unimplemented and return a failure. In the objects table, a PR stacked on another one (for example on #291) counts the objects it carries along.

InterfaceMethodsOn mainIn open PRsOpen
ID3D12Device1177221639
ID3D12GraphicsCommandList879173626
ID3D12CommandQueue16835
ID3D12CommandAllocator6501
ID3D12Fence19801
ID3D12DescriptorHeap8701
ID3D12Resource12831

Objects the layer has to create, beyond those above:

InterfaceNeeded forOn mainIn open PRs
ID3D12PipelineStatePSOs (D3D-3/D3D-4)no#367, #335, #317, #291, #286
ID3D12RootSignatureroot signaturesno#367, #317, #291, #286
ID3D12QueryHeapqueries (D3D-15)no#278
ID3D12Heapheaps and placed resources (D3D-16)no#286
ID3D12CommandSignatureExecuteIndirect (D3D-17)no#317
ID3D12PipelineLibrary1pipeline libraries (D3D-34)no—
ID3D12StateObjectray tracing (not planned yet)no—

Not planned until the game's trace shows them: shared handles and OpenExistingHeap*, protected resource sessions, meta commands, sampler feedback, lifetime trackers, RemoveDevice, CheckDriverMatchingIdentifier, stream output (SOSetTargets, waits for SHADER-16's counts), CopyTiles, and the ray-tracing calls (CreateStateObject, BuildRaytracingAccelerationStructure, DispatchRays, milestone M6, see RT-4 and ASSET-23).

Decisions for the maintainer

  • Review order: #119 first, then #291. Most of the layer PRs build on them.
  • #138's root-signature packer (merged) and the root-signature code in #291: keep one or both.

Activity

30
Ryx6915h ago

i will take D3D-29 and D3D-27 for now!

KoertLichtendonk15h ago

Descriptors: four overlapping PRs. @kontaktjakubjagiello @aslan2142 @0xTehuti @Ryx69, there are now several PRs for the same pieces:

PieceAlready on main or openNew
GFX10 T#/S#/V# codec, bindless set 0, CopyDescriptors#230 (merged, @aslan2142)—
Texture SRV/UAV views and samplers into the bindless set#328 (@aslan2142)#334
Buffer views (CBV/SRV/UAV) as V#s#157 (@0xTehuti); @aslan2142 suggested it fits #328's buffer branch#334
Root descriptor tables and SetDescriptorHeaps in the command list, bindless index allocation#291 (@Ryx69), with set 0 binding from D3D-5#335 (draft)
Per-type descriptor sizes and a CPU-handle → heap registry—#333

#333 looks new. @kontaktjakubjagiello, could you compare #334 and #335 with #328, #157 and #291 here before going further, and keep what they lack? The rest could become reviews of those PRs. One suggestion for the split: #328 for textures and samplers, #157's V# code in #328's buffer branch for buffer views, #291 for the command-list binding, and #333 on top of all three.

73N1X15h ago

Coordination note: my RGBA8 typeless RTV clear work is already claimed in #156: https://github.com/ItsNotRexed/Leonida/issues/156#issuecomment-7070653793

Phase 1 is locally implemented and validated, but no PR is open: it waits for D3D-36 / #318 to be integrated, then will be based directly on main. It preserves explicit UNORM/SRGB/UINT RTV formats on one-mip, one-layer, single-sample 2D RGBA8 typeless resources. It does not implement BeginRenderPass/EndRenderPass (D3D-27), DispatchMesh (D3D-29), or bindless SRV/UAV/sampler descriptors.

Compatibility check against the current #333 diff: my TargetView format field still fits its 32-byte RTV/DSV slots. I do not change descriptor strides, heap registration or CopyDescriptors. If #333 lands first, the typeless patch will retain its per-type slot checks and registry rather than restore the old shared-size implementation. The shared edit surface is d3d12_vulkan_device.cpp; the scopes are different.

The clear implementation uses a short dynamic-rendering scope internally when the RTV reinterprets the image format. I will compare that path with D3D-27 when its implementation is available, so its render-pass work can reuse the same typed-view interpretation without duplicating it.

Mip/array RTV expansion will be a separate phase after phase 1 is integrated. No phase-2 implementation PR is being stacked on the pending bases. Please flag any overlap not covered by the current tables.

aslan214215h ago

@KoertLichtendonk The suggested split works for me: #328 for texture views and samplers, #157's V# code in its buffer branch, #291 for the command-list binding, #333 on top.

@kontaktjakubjagiello #333 keeps every slot at 32 bytes, so it does not change what #328 writes (a T# is 32 bytes, an S# is 16 bytes in a 32-byte slot). The two only conflict textually in d3d12_vulkan_device.cpp. I can rebase #328 if #333 lands first.

@Ryx69 #291 still refuses shaders with set-0 bindings in two places. Once #328 is in, the hookup is rt::d3d12::bindless_set_layout in the pipeline layout and bindless_set bound before the dispatch or draw. I can send that as a follow-up with an end-to-end test, or you can fold it into #291.

AI-assisted (Claude Code, Claude Fable 5.2).

KoertLichtendonk15h ago

@kontaktjakubjagiello @Ryx69: since the comment above, #335 has also gained the draw path: SetPipelineState, root signatures, DrawInstanced/DrawIndexedInstanced and replay. That's the core of #291 (D3D-3/D3D-4, @Ryx69), which is mergeable and is the critical path everything else builds on. @aslan2142 agreed to the split above. @kontaktjakubjagiello, could you hold #335 at the bindless allocator and the table scan, compare the rest with #291 here, and offer what #291 lacks (your first_draw_test, for example) as a review or a test of #291?

Ryx6915h ago

Thanks both, the split works for me.

@aslan2142 please send the set-0 hookup as a follow-up (Depends on: #291, #328) rather than folding it into #291. That way #291 can land now without waiting on #328. The two refusals you mean are the uses set-0 bindings checks in d3d12_vulkan_device.cpp, one in compute PSO creation and one in graphics. Your PR can drop both once the layout carries set 0.

@kontaktjakubjagiello a review of #291 against #335 is welcome. If first_draw_test covers something #291's tests don't, I'm happy to take it into #291 as a test or have it land right after.

73N1X14h ago

done: D3D-41 #388

#318 is integrated, so phase 1 is now rebased directly onto main (d7cbc3e) and submitted as one typeless-only commit. Full runtime build and CTest 19/19 passed, along with 13 typeless WARP readbacks, 19 CPU scenarios, 12 invalid descriptor rejections and six GPU regression executables under validation. The PR links the original #156 claim and the #333/D3D-27 compatibility note. Mip/array expansion remains outside #388 and will wait for this phase to be integrated.

KoertLichtendonk14h ago

Order for the five PRs editing d3d12_vulkan_device.cpp, as @ItsNotRexed asked in the merge-round note (posted in #156). A suggestion based on the split agreed above; authors, please say here if you'd rather have it otherwise:

  1. #288 (@Parakeet-Wheat, WriteToSubresource/ReadFromSubresource): small and independent of the descriptor work, so it rebases first.
  2. #328 (@aslan2142, texture views and samplers into the bindless set): the base of the agreed split.
  3. #333 (@kontaktjakubjagiello, per-type descriptor sizes and the handle registry): "on top" in the split; @aslan2142 noted it keeps 32-byte slots, so it only conflicts textually with #328.
  4. #334 (@kontaktjakubjagiello): only what #328 and #157 don't already have, after those two. Buffer views go into #328's buffer branch with #157's V# code.
  5. #329 (@0xTehuti): #327 was merged this round with the same three GetResourceAllocationInfo methods, so #329 can probably close. If it has something #327 lacks, a small follow-up on top of main is easier.

Separately, #291 (@Ryx69) stays the critical path and lands independently of these. @aslan2142's set-0 hookup follows after #291 and #328.

Parakeet-Wheat14h ago

@KoertLichtendonk thanks for the order, it works for me.

#288 (D3D-22, WriteToSubresource/ReadFromSubresource) is rebased onto main at d937b7f and mergeable again (head 74c99af). Texture creation now uses #327's texture_create_info and adds the host transfer usage on top. Retested on an RTX 3070: ctest 18/18, both copy paths (host image copy and staging) and the allocation info comparison pass with zero validation errors. @aslan2142, #328 can go next whenever you're ready.

#258 (D3D-18, SetPredication) is rebased too (head 96f113a). It doesn't touch the descriptor work. Its replay test is now part E, after main's new barrier test (part D).

AI-assisted (Claude Code).

KoertLichtendonk14h ago

New work for the layer: comparison tests. Almost every D3D12 item here is claimed, so Up for grabs above now has seven items (D3D-45 to D3D-51) that check what already exists against the host PC runtime or WARP: the same pixels for clears, copies and resolves; the same behaviour for fences and queues; the same errors for invalid arguments, CheckFeatureSupport and command-list misuse. GetCopyableFootprints, GetResourceAllocationInfo and the UINT clears already work this way (create_host_device() in runtime/test/d3d12_vulkan_device_test.cpp). Windows and a GPU are enough; no game files. Four of them are small first tasks.

The method coverage also gained an objects table: none of pipeline state, root signature, query heap, heap or command signature exists on main yet. They come with #291, #286, #278 and #317, and the pipeline library and state object exist nowhere.

ezwinxdlmao13h ago

claiming: new item (D3D): primitive restart on the Vulkan layer.

#291's graphics path logs a PSO's IBStripCutValue as index-buffer strip cuts (primitive restart) are not modelled, and docs/d3d12_vulkan.md lists primitive restart as a known gap. No claim or PR covers it.

Plan:

  • Mapping: the PSO's IBStripCutValue (0xFFFF / 0xFFFFFFFF) maps to Vulkan primitive restart, set at each draw through the dynamic VK_DYNAMIC_STATE_PRIMITIVE_RESTART_ENABLE (core 1.3).
  • When it's on: for strip topologies, when the cut value matches the bound index format. That is the all-ones index Vulkan restarts on.
  • Unclear cases: the mismatched-format and list-topology cases get the host D3D12 runtime's answers, measured on an RTX 3050 and a Radeon 680M.
  • Test: a draw test reads back strips with cuts.

It builds on #291's draw path (Depends on: #291). @Ryx69, if you'd rather keep this inside #291, say so and I'll hand it over. No game files.

KoertLichtendonk13h ago

D3D-32 (variable rate shading): @0xTehuti, who claimed it first, now has a runtime/d3d12-variable-rate-shading branch, and @ezwinxdlmao has #469 for the same item. Could you two compare and keep one PR? The other could review it or add tests. Both touch d3d12_com.cpp, vulkan_backend.h and the command-list files, so two PRs would conflict anyway.

New in Up for grabs here: D3D-54 (a synthetic draw from upload-heap buffers to check the -heap_base bias, open question 17), D3D-55 (log the arena high-water mark, open question 5) and D3D-56 (a GPU benchmark of the descriptor hash table). None needs game files.

KoertLichtendonk13h ago
  • D3D-45: #521 (@NWX10, who claimed it at 00:07) and #525 (@luisbaker) both compare clears with the host runtime and add SNORM/SINT targets. Could you two compare and keep one? The other's extra cases could go in as a review suggestion.
  • D3D-32: #519 (@0xTehuti) is open now as well as #469 (@ezwinxdlmao), so the same item has two PRs. Please agree which one continues.
  • #291: after this round's merges it conflicts with main. @Ryx69, a rebase would let it land next. @ToxicmaggotRP's test of its snapshot (with #119) passed on an RTX 4060 Ti with validation on (comment).
  • New IDs: D3D-57 primitive restart (@ezwinxdlmao), D3D-58 subgroup size control in PSOs (@AlxcBTW). Both depend on #291.
AlxcBTW13h ago

New item, claiming: Xbox presentation through PresentX and frame events (answers open question 6 in docs/d3d12_vulkan.md §8.2, "DXGI, PresentX, or both")

Today's plan (§4.5) is "implement the DXGI path and stub PresentX". Public Microsoft code says a GDKX title presents without a DXGI swapchain. All of the following are facts from public, MIT-licensed sources; no SDK material.

  • Sources:

    • microsoft/Xbox-GDK-Samples SimpleDeviceAndSwapChain. Its readme says: "Instead of using DXGI for presentation, this uses the new PresentX API".
    • DirectXTK12 wiki, DeviceResources: "The GDKX platform uses PresentX, so the method WaitForOrigin has to be called each frame".
    • The GDKX DeviceResources template, and SDL2's Xbox D3D12 backend, follow the same flow.
  • Call sequence, as in those samples:

    1. D3D12XboxCreateDevice creates the device.
    2. Back buffers are ordinary CreateCommittedResource with D3D12_HEAP_FLAG_ALLOW_DISPLAY, initial state PRESENT.
    3. DXGI is used only to find the output: device.As(IDXGIDevice1) → GetAdapter → EnumOutputs(0).
    4. Then device->SetFrameIntervalX(output, D3D12XBOX_FRAME_INTERVAL_60_HZ, backBufferCount - 1, FLAG_NONE) and device->ScheduleFrameEventX(D3D12XBOX_FRAME_EVENT_ORIGIN, 0, nullptr, FLAG_NONE).
    5. Each frame:
      • device->WaitFrameEventX(ORIGIN, INFINITE, nullptr, FLAG_NONE, &token);
      • render, then ExecuteCommandLists;
      • queue->PresentX(1, &plane, nullptr), with plane = {Token, ResourceCount = 1, ppResources = &backbuffer} (HDR: 2 planes, with pSrcViewRects and ColorSpace).
    6. On suspend: queue->SuspendX(0). On resume: queue->ResumeX(), then frame events are registered again. At shutdown: PresentX(0, nullptr, nullptr) for a blank screen.
    7. Pacing statistics come from GetFrameStatisticsX.

    This matches the UMD names in docs/d3d12_xs_surface.md §7: PresentX, SuspendX and ResumeX on the queue, the frame events on the device, at most 2 planes.

  • What it means: if the game follows the GDKX pattern, which is very likely, the DXGI swapchain shim (D3D-2) is not on the boot path; only IDXGIDevice1/Adapter/Output enumeration is. Nothing in runtime/ implements PresentX or the frame events yet. Without them no frame ever reaches the screen.

Plan, slot-independent first, because the vtable slots of the X methods are still unmeasured:

  1. Docs PR: answer §8.2 Q6 with these sources, update §4.5 and the M0 row, and list the method shapes taken from the public call sites.
  2. A presentation core on #119's presenter (Depends on: #119). It covers:
    • PresentX with 0–2 planes, presenting plane 0 into the Vulkan swapchain;
    • SetFrameIntervalX / ScheduleFrameEventX / WaitFrameEventX on a vblank timer, with the token carried through to the present;
    • GetFrameStatisticsX (minimal), and SuspendX / ResumeX;
    • unit tests that call the methods directly, with a fake clock.
  3. Slot wiring once the slots are measured. I'll coordinate with D3D-2 (@Vyral1978) and #119 (@Ryx69) so the presenter is shared, not duplicated.

AI-assisted (Claude Code).

aslan214213h ago

It's night for ryx69, so i don't think #119 is gonna be merged anytime soon

KoertLichtendonk13h ago

D3D-59: @AlxcBTW's PresentX and frame-events claim. @Vyral1978 (D3D-2), this shares #119's presenter with your DXGI shim, so please coordinate there.

New in Up for grabs: texture formats the layer still refuses.

  • D3D-60 (⭐): the integer and typeless formats left after D3D-45.
  • D3D-61: packed and special formats.
  • D3D-62: 96-bit RGB32 textures.
  • D3D-63: 1D textures. CheckFeatureSupport already reports them as supported.
  • D3D-64: render-target and depth views on 3D textures.

None of them needs game files.

NWX1013h ago

D3D-51: leaving it to #495, and I'll review it. My fix/command-list-rules branch was never opened as a PR. It makes the same three changes as #495 (allocator Reset while a list records, CreateCommandList and list Reset on an allocator another list records into). It has one case #495 doesn't: CreateCommandAllocator and CreateCommandList with an undefined list type, which are E_INVALIDARG on the PC runtime and S_OK in the layer on main. @Aminecsmc, I'll post that as a suggestion on #495 and won't open a PR.

D3D-45: I'd keep #521. From #525's description, the SNORM and SINT targets and the SINT conversion (NaN to 0, clamp, truncate toward zero) are also in #521. #521 adds tie rounding (0.5 on an 8-bit channel is 128 on WARP and 127 from NVIDIA's clear), 15 formats instead of 4, the R8/R16 formats, sRGB and the four depth formats: 178 cases against WARP on an RTX 3060 Ti.

What #525 has that #521 doesn't is a run on an Intel GPU, and that is worth keeping. @luisbaker, could you run d3d12_clear_formats_test from #521 on your Intel graphics and post the result on #521? You saw the sRGB .5 boundary differ there, and #521 leaves the sRGB colour curve to the GPU, so an Intel result is the one thing I can't show. If a case differs I'll fix it in #521.

ezwinxdlmao13h ago

done: D3D-57 #572

I measured the host D3D12 runtime first, on a Radeon 680M, an RTX 3050 and WARP. A strip is cut only where an index equals the cut value. So the layer turns Vulkan's primitive restart on per draw only when the cut value is all ones for the bound index format.

Two cases can't follow every driver; both are in the docs and the PR:

  • 0xFFFF with 32-bit indices: D3D12 cuts at 0x0000FFFF, which Vulkan can't express.
  • Lists: AMD's driver ignores the cut, while NVIDIA's and WARP apply it. The layer follows AMD and vkd3d-proton.
AlxcBTW13h ago

status: Xbox presentation (PresentX + frame events)

  • Step 1 (docs): #540.
  • Step 2: a presentation core on #119 (FramePacer, PresentX with 0–2 planes, SuspendX/ResumeX, GetFrameStatisticsX). It is built and tested on both GPUs with 0 validation errors, and is in review fixes; a draft PR follows with Depends on: #119.

status: wave64 subgroup size control in PSOs

  • Built on #291 and green on both GPUs; native wave64 compute is exact on the 880M with required 64.
  • In review fixes: adding a wave64 pixel-shader end-to-end test and reading the driver's actual subgroup size back. A draft PR follows with Depends on: #291.

status: D3D-46 (copies)

  • 39 host-comparison cases, 0 differences across Vulkan AMD/NV × host NV/AMD/WARP. Two layer fixes: six missing formats, and BC copies at the padded mip edge.
  • In review fixes (separate source/destination edge clipping); the PR follows.
kevineduardo13h ago

But why go to Vulkan, if DirectX 12 it's native on Windows? WHYYYY?