Skip to content

Releases

VKD3D-Proton macOS 6.0

Choose a tag to compare

@github-actions github-actions released this 16 Nov 06:19

VKD3D-Proton macOS 6.0 is built from main (2178b90, #22 and two later commits). Most of it is about speed: pipelines compile without blocking, and shaders converted by Metal Shader Converter are kept on disk. ExecuteIndirect stays on converted pipelines and runs only the counted draws and dispatches. Tessellation of DXIL shaders runs in the converter's emulation. Fence waits wake as soon as the GPU signals.

Pipelines and shaders

  • Pipelines compile while their creation returns. Deferred pipelines are on by default: the creating thread starts each one and Metal's completion handler finishes it, so no thread waits on Metal. Metal compiles with every core.
  • Converted shaders are kept on disk beside the pipeline cache. They are built for this GPU family and macOS, without DXIL debug information. When the in-memory cache fills, it drops only its oldest quarter. A DXBC shader is converted to DXIL once and reused with every root signature.
  • Tessellation of DXIL shaders (vertex, hull and domain) goes through Metal Shader Converter's object and mesh emulation as a mesh pipeline. Before, it ran SPIR-V compute passes and restarted the render pass for every tessellated draw. VKD3D_MSC_TESSELLATION=0 keeps it on SPIR-V. Isolines, DXBC tessellation and ExecuteIndirect with such a PSO still use SPIR-V.

ExecuteIndirect

  • Root arguments keep converted pipelines: command signatures write root constants and root descriptor addresses into Metal Shader Converter's top-level argument buffer. Before, such an ExecuteIndirect fell back to the PSO's SPIR-V pipeline, compiled at first use.
  • Count buffers: indirect draws run only the counted draws, encoded into an indirect command buffer by a kernel outside the render pass. Dispatch-only signatures run only the counted dispatches.
  • No render pass restarts: ExecuteIndirect inside a render pass no longer ends it. Before, every attachment was stored and reloaded once per ExecuteIndirect.
  • 25 or more push-constant-only draws run from one indirect command buffer. VKD3D_MSC_EXECUTE_INDIRECT=0 turns converted ExecuteIndirect off.

Synchronization, encoding and present

  • Fence waits wait on the timeline semaphore's Metal shared event directly instead of through a listener on another thread (MVK_CONFIG_DIRECT_SEMAPHORE_WAIT=0 keeps the listener).
  • Barriers: transitions between dispatches, and barriers from compute into other stages, keep the compute encoder. Enhanced barriers keep the render pass open, as ResourceBarrier does. CPU-only descriptor heaps are host-only pools.
  • Command buffer prefill (encoding a command list when its recording ends, on the recording thread) is opt-in. As a default it stalled Subnautica 2's GPU queue. Recording more command lists than half the Metal queue before submitting no longer freezes vkEndCommandBuffer.
  • SyncInterval 2 to 4 now holds frames (VK_EXT_present_timing relative targets). Before, it presented every refresh, like 1.
  • MVK_CONFIG_RAY_TRACING_MIN_APPLE_FAMILY and MVK_CONFIG_CONSERVATIVE_RASTERIZATION=0 can hide two slow emulated features: ray tracing below a given GPU family, and conservative rasterization. Neither option is set by default, so both features are still reported.

DLSS and DXVK

  • DLSS Super Resolution passes its bias mask (MetalFX's reactive mask), the MVJittered flag and the subrect bases of each input and the output to MetalFX. MetalFX temporal scalers are created without stopping the queue, within the GPU's input scale range.
  • Assetto Corsa with Custom Shaders Patch no longer crashes with MSAA under FEX. Its detour of ResolveSubresource landed on ARM64EC code; DXVK's hot-patchable entry, backported, gives it an x86_64 one.
  • The ARM64X NGX core and NVAPI carry Wine's builtin marker, as on x86_64.

artifacts/evidence/load-time-and-encoding-costs.md has the functional harness (139 of 156, unchanged) and perf probe results against the MetalSharp Wine 7.0 lane.


Built from 2178b90. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v6.0-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll) as x86_64 Wine builtins, which run DLSS Super Resolution and Ray Reconstruction on MetalFX in an x86_64 Wine (MetalSharp's, under Rosetta). Stage them together.
  • vkd3d-proton-v6.0-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll, ARM64X), which run DLSS Super Resolution and Ray Reconstruction on MetalFX.
  • moltenvk-v6.0-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex, mesh, amplification and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off), and DirectXShaderCompiler's DXBC to DXIL converter (libdxbcconverter.dylib, NCSA), which hands it DXBC shaders too (METALSHARP_MSC_DXBC=0 keeps them on SPIR-V). It runs MetalFX's scalers for VKD3D-Proton's and DXVK's DLSS and swap chain upscaling (mvk_metalfx.h).

VKD3D-Proton macOS 5.4

Choose a tag to compare

@github-actions github-actions released this 10 Nov 03:54
6a1326c

VKD3D-Proton macOS 5.4 is built from main (6a1326c, #20 and #21). Most of it fixes CONTROL Resonant on the VKD3D-Proton lane: materials, window interiors, mesh shaders and ray tracing now render, and its GPU faults are gone. ExecuteIndirect now works when a command signature changes a root CBV, and timestamp queries cost less.

CONTROL Resonant

  • Streamed materials rendered as block noise, and the game hit a GPU fault (#20): MoltenVK gave every sparse texture ShaderWrite, which puts it on Metal's sparse tier 1, and tier 1 tiles of block-compressed formats aren't the rectangles Metal reports. Compressed sparse textures now stay on tier 2, whose tiles match. Binds that mixed tier 1 and tier 2 textures stalled Metal's queue for 4 s, which the game reported as a GPU fault; each bind now maps tier 1 textures first, then tier 2.
  • Window interiors glowed magenta and yellow (#21): Metal (macOS 27.2, placement sparse tier 2) loses mips 6 and 7 of the tail of some block-compressed sparse textures whose height isn't a power of two, such as 2048×1080 BC6H and 1920×1080 BC7. The first sparse image of each shape now runs a GPU self-test. An image whose tail doesn't hold gets an ordinary texture, bound as the sparse one. Unmapped blocks of it read zeros, both when it's created and after an unmap. tiled_bc now includes both shapes.
  • Mesh shaders failed with E_OUTOFMEMORY (#20): SPIRV-Cross copied an array output per vertex or primitive (float2 TEXCOORD[2]) whole into each flattened member, which Metal rejects. All 49 of the game's mesh shaders now compile; 5 failed before.
  • Ray tracing hung with a GPU fault (#20): vkCmdTraceRaysKHR now accepts miss, hit and callable tables with a stride of 0, which read their first record for every index.

ExecuteIndirect

  • A command signature that changes a root CBV (#21) failed with E_NOTIMPL ("Root parameter is not a raw VA"). Root CBVs are push descriptors on MoltenVK (Metal constant buffers, which read faster than through a device address). A new MoltenVK token, VK_INDIRECT_COMMANDS_TOKEN_TYPE_PUSH_DESCRIPTOR_BUFFER_MVK, now binds each sequence's address to the descriptor from the indirect command buffer. DirectX-Graphics-Samples' D3D12ExecuteIndirect renders the same as on D3DMetal.

Performance and debugging

  • Timestamps (#20) are written on the next render or compute pass instead of a blit pass each. In Subnautica 2, Xcode's API insights fell from 689 to 112.
  • GPU captures of frames with sparse binds now replay. METALSHARP_TRACE=gpu reports sparse binds that are slow or stuck.

Known open

  • A dark floor decal in CONTROL Resonant still flickers.

Built from 6a1326c. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v5.4-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll) as x86_64 Wine builtins, which run DLSS Super Resolution and Ray Reconstruction on MetalFX in an x86_64 Wine (MetalSharp's, under Rosetta). Stage them together.
  • vkd3d-proton-v5.4-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll, ARM64X), which run DLSS Super Resolution and Ray Reconstruction on MetalFX.
  • moltenvk-v5.4-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex, mesh, amplification and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off), and DirectXShaderCompiler's DXBC to DXIL converter (libdxbcconverter.dylib, NCSA), which hands it DXBC shaders too (METALSHARP_MSC_DXBC=0 keeps them on SPIR-V). It runs MetalFX's scalers for VKD3D-Proton's and DXVK's DLSS and swap chain upscaling (mvk_metalfx.h).

VKD3D-Proton macOS 5.3

Choose a tag to compare

@github-actions github-actions released this 07 Nov 21:22
2103382

VKD3D-Proton macOS 5.3 is built from main (2103382). It fixes black 3D in DXVK games that read constant buffers out of range, and the x86_64 package now carries the lane's own signed NGX core and NVAPI, so DLSS runs in an x86_64 Wine.

Fixes

  • Black 3D on DXVK (MoltenVK's SPIRV-Cross, #19): a bounds-checked constant buffer load took the address of a vector component (&cb.m[i] for a float4 m). Metal rejects that ("address of vector element requested"), so those shaders didn't compile and DXVK dropped their draws. PEAK (Unity, D3D11) rendered its 3D black, with 27 failed pipelines. Only an access that ends on the array element the bounds check covers now reads from the zero constant by address; anything else selects on the loaded value. PEAK renders with 0 pipeline failures, and 24 of its shaders still use the faster address path.

Packages

  • x86_64 NGX and NVAPI (#18): vkd3d-proton-v5.3-x86_64 now includes the lane's nvngx.dll and nvapi64.dll as x86_64 Wine builtins, signed by this repository's CI. They run DLSS Super Resolution and Ray Reconstruction on MetalFX in an x86_64 Wine (MetalSharp's, under Rosetta).

Built from 2103382. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v5.3-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll) as x86_64 Wine builtins, which run DLSS Super Resolution and Ray Reconstruction on MetalFX in an x86_64 Wine (MetalSharp's, under Rosetta). Stage them together.
  • vkd3d-proton-v5.3-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll, ARM64X), which run DLSS Super Resolution and Ray Reconstruction on MetalFX.
  • moltenvk-v5.3-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex, mesh, amplification and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off), and DirectXShaderCompiler's DXBC to DXIL converter (libdxbcconverter.dylib, NCSA), which hands it DXBC shaders too (METALSHARP_MSC_DXBC=0 keeps them on SPIR-V). It runs MetalFX's scalers for VKD3D-Proton's and DXVK's DLSS and swap chain upscaling (mvk_metalfx.h).

VKD3D-Proton macOS 5.2

Choose a tag to compare

@github-actions github-actions released this 07 Nov 03:32
1dcaa3e

VKD3D-Proton macOS 5.2 is built from main (1dcaa3e, #17). It moves most of VKD3D-Proton's remaining SPIR-V pipelines onto Apple's Metal Shader Converter, fixes accuracy bugs found along the way, and makes draws, dispatches and pipeline creation cheaper. It also fixes Subnautica 2 and God of War Ragnarök pipeline failures.

Across the 147 tests/perf scenes (x86_64 under FEX, Apple M6), the geometric mean against 5.1 is −23% CPU, −21% GPU and −9% peak memory.

What the Metal Shader Converter now takes

In 5.1, shaders of these kinds stayed on SPIR-V. In 5.2 they go through the converter:

  • Typed buffers (Buffer<T>, RWBuffer<T>). The one exception is three-component 32-bit typed SRVs. Each D3D12 view gets an exact Metal texel view, so the converter no longer reads from element 0.
  • Mesh shaders, with or without an amplification shader (Nanite).
  • Depth-only passes: pipelines with no pixel shader, or with nothing to write at all.
  • Dual-source blending.
  • SM 6.8 draw parameters.
  • DXBC (SM 4.x/5.x) shaders, through DirectXShaderCompiler's dxilconv. This needs the new libdxbcconverter.dylib. Set METALSHARP_MSC_DXBC=0 to keep DXBC on SPIR-V.
  • The depth bounds test and min/max sampler filtering, on Apple10 GPUs with macOS 26.

Games

  • Subnautica 2: its mesh pipelines that pass a texture with a min-LOD clamp to a helper function compile now. SPIRV-Cross didn't pass the clamp word along, so those pipelines failed: 0 compile failures now.
  • God of War Ragnarök: its tessellation pipelines compile now. Before, the game quit around 80 s in with "The system is out of memory. Try increasing the size of the PageFile." That message was VKD3D-Proton's E_OUTOFMEMORY for a pipeline that failed to compile. The vertex shader packed its outputs for the hull shader in a layout the hull shader doesn't read.

Accuracy fixes

  • Depth clip: nothing in front of the near plane was clipped in any D3D game. MoltenVK ran every render pass in Metal's OpenGL mode. It now uses that mode only for the draws that need it.
  • ResourceMinLODClamp: views were moved to their clamped mip, so textures came out blurrier and GetDimensions was wrong. MoltenVK now has VK_EXT_image_view_min_lod, and both paths clamp, comparison samples included.
  • Sampler MipLODBias: it was ignored on M1 to M4 and on macOS 14 and 15. It is now applied on both paths.
  • Min/max samplers on converted shaders before Apple10: they read as linear. Those PSOs now stay on SPIR-V.
  • f32tof16 rounds toward zero and saturates to 65504. It used to round to nearest and return +inf, which can turn bright HDR values into NaNs.
  • D24 depth bias is no longer 8× too small at depth 0.1, which showed as shadow acne.
  • Cull distances: the converter drops SV_CullDistance, so shaders that write it stay on SPIR-V.
  • SV_StencilRef: the converter ignores it, so pixel shaders that write it stay on SPIR-V.
  • Barycentrics: wrong when they were a pixel shader's only input, because of an Apple driver bug. SPIRV-Cross now works around it.
  • EvaluateAttributeSnapped no longer reads 1/16 pixel low.
  • R10G10B10A2_UINT vertex attributes now make a pipeline.
  • Geometry shaders:
    • correct primitive IDs and winding on cut strips;
    • one-pixel points;
    • cut triangle strips with adjacency;
    • several streams.
  • pow2 tessellator partitioning now rounds factors up to a power of 2.
  • Stripped DXBC (no reflection) is bridged now.
  • Fixes from reviewing the branch: a use-after-free in shadow descriptor entries, and stale views keeping freed buffers' memory.

Performance (tests/perf, Apple M6, against the 5.1 lane)

  • Predication: each predicated draw used to end the render pass. 2000 predicated draws: 53 → 0.49 ms CPU and 67 → 0.91 ms GPU.
  • 64-bit atomics used a hashed spin lock. Max and min are now native, and add, sub and bitwise ops split into two 32-bit atomics. atomics64_contended: 384 → 0.08 ms CPU and 381 → 0.22 ms GPU.
  • Sampler feedback skips the lock when the write would change nothing. sampler_feedback: 22.2 → 0.32 ms GPU.
  • Geometry shaders run up to 32 input primitives per threadgroup. xfb_gs_dxbc: 11.3 → 1.0 ms GPU. Cube shadow maps: −36% to −51% GPU.
  • Cheaper bridge draws and dispatches:
    • the SPIR-V bindings are no longer flushed;
    • arguments identical to the last push aren't pushed again;
    • dispatch arguments come from a ring buffer;
    • typed views are shared between descriptors of the same range;
    • conversions are cached by blob.
  • Slower: out-of-bounds buffer loads on the converter take +50% to +100% GPU, because it loads through a null pointer. Typed views at offsets that are never reused cost more too, mostly from Metal validating each view.

Tests

  • More than 50 new tests/d3d12 tests.
  • An op-level DXIL/DXBC differential test (opdiff).
  • A converter feature sweep, written up in docs/msc-feature-coverage.md.
  • tests/run/matrix.sh runs the suite on Metal 4 and Metal 3 (METALSHARP_METAL_VERSION=3 acts as macOS 14 and 15), with and without the bridge, on native arm64 and on x86_64 under FEX. Every combination passes except d3d11on12 on native arm64, which needs the lane's DXGI.

Known gaps

  • Geometry and tessellation shaders stay on SPIR-V. The converter's emulation contract is written up in docs/msc-geometry-emulation.md.
  • Isoline tessellation makes no pipeline.
  • Tessellation followed by a geometry shader skips its draws.
  • Pipeline statistics queries read 0.
  • Aliasing barriers between draws end the render pass. Apple GPUs have no memory barriers inside a render pass.
  • ResourceMinLODClamp isn't applied to gathers.

Built from 1dcaa3e. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v5.2-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows). Stage them together.
  • vkd3d-proton-v5.2-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll, ARM64X), which run DLSS Super Resolution and Ray Reconstruction on MetalFX.
  • moltenvk-v5.2-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex, mesh, amplification and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off), and DirectXShaderCompiler's DXBC to DXIL converter (libdxbcconverter.dylib, NCSA), which hands it DXBC shaders too (METALSHARP_MSC_DXBC=0 keeps them on SPIR-V). It runs MetalFX's scalers for VKD3D-Proton's and DXVK's DLSS and swap chain upscaling (mvk_metalfx.h).

VKD3D-Proton macOS 5.1

Choose a tag to compare

@github-actions github-actions released this 06 Nov 07:51
a3c9d40

VKD3D-Proton macOS 5.1 is built from main (a3c9d40, #15). It makes Minecraft Bedrock's Ray Traced mode render correctly on the arm64 lane, gets Unreal Engine 5 Nanite games (Subnautica 2) running and drawing the right way up, finishes the geometry-shader fixes from #11, and adds a DLSS render scale override.

Ray tracing (Minecraft Bedrock, Ray Traced mode)

Before 5.1, the game hit GPU page faults and an MMU fault, then corrupted chunks (missing surfaces, sky-coloured mirror materials), then GPU watchdog kills (kIOGPUCommandBufferCallbackErrorImpactingInteractivity) during ray-query traversal.

  • Acceleration-structure copies get memory of their own. The size an app is told for a compacted or cloned BLAS is the structure's own size. It doesn't include the heap size and placement alignment. A compacted BLAS placed in the app's memory could run into the next one, and VKD3D-Proton's D3D12 games pack them tightly. Copies now write to memory of their own.
  • Null and unknown structure addresses in TLAS instances and ray queries now find an empty structure that every ray misses. Before, they read memory that holds no structure, which caused the GPU page faults.
  • BLAS vertices Metal can't read in place (16-bit components at offsets or strides that aren't multiples of 4, a zero stride) are repacked by a small kernel.
  • A null RTAS descriptor keeps the null RTAS's address. It stored 0 before, so a ray query through it read address 0. This bug is in upstream VKD3D-Proton too.
  • Shaders with ray queries stay on SPIR-V. Metal Shader Converter's modules have no binding for acceleration structures, so these shaders no longer go through it.
  • The "no heap is bound" root-signature warning is logged once instead of every frame.

Unreal Engine 5 mesh shaders (Subnautica 2, UE 5.6)

  • Nanite's mesh pipelines convert. SPIRV-Cross ran out of recompile passes on shaders whose helpers or 64-bit atomics turned up one per pass ("Maximum compilation loops detected"). VKD3D-Proton then crashed after the game's loading screen. All 64 of UE 5.6's Nanite hardware raster pipelines now convert, and the game runs.
  • Mesh shaders draw the right way up. dxil-spirv declares D3D12's SV_Position as a variable of its own, and SPIRV-Cross's vertical flip skipped it. Nanite's hardware rasterizer drew the world upside down and faceted. It now matches the compute rasterizer and the vertex shaders.

Geometry shaders (#11)

  • A vertex output wider than the geometry shader's input now compiles. Forza Horizon 3 failed with assigning to 'float2' from incompatible type 'float4' (TEXCOORD, the float3 NextNormal, fog parameters). The value is now narrowed or zero-padded and converted to the input's type. This is the fix from the 5.1 alpha shared in #11.

DLSS

  • METALSHARP_DLSS_SCALE sets the render scale for every DLSS mode. It takes a mode name (ultra_performance, performance, balanced, quality, ultra_quality, dlaa) or a fraction from 0.25 to 1, for games that pick the mode themselves. Minecraft's ray tracing resets its upscaler to Quality; with Performance it traces about 44% fewer rays and was much faster in game.

Tests

  • New: gs_narrow_dxil and gs_narrow_dxbc.
  • tests/d3d12 on the VKD3D-Proton lane (under FEX): 5.0's MoltenVK fails 4 tests (19 checks); 5.1 fails 1 (tiled3d, 1 check, failing the same way in both).
  • The compaction fix was checked with an offline Vulkan reproduction. Before it, the neighbouring BLAS broke in 7 of 18 slot and size layouts; with it, in 0 of 18.
  • In game:
    • Minecraft Bedrock 1.26.62, Ray Traced with DLSS: no chunk corruption, GPU faults or watchdog kills.
    • Subnautica 2: renders upright and past the loading screen.

CI

  • Metal source compilation runs on macos-15, because GitHub no longer runs macos-14 jobs.

Known gaps

Still open from 5.0:

  • Per-draw binding cost
  • Predication
  • Transform feedback
  • DispatchRays
  • Isolines
  • Fence completion under heavy CPU load
  • On Subnautica 2 this lane runs at about half Mullion's frame rate.

Built from a3c9d40. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v5.1-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows). Stage them together.
  • vkd3d-proton-v5.1-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll, ARM64X), which run DLSS Super Resolution and Ray Reconstruction on MetalFX.
  • moltenvk-v5.1-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off). It runs MetalFX's scalers for VKD3D-Proton's and DXVK's DLSS and swap chain upscaling (mvk_metalfx.h).

VKD3D-Proton macOS 5.0

Choose a tag to compare

@github-actions github-actions released this 03 Nov 09:03

VKD3D-Proton macOS 5.0 is built from main (54537b6, #12). It brings DLSS Frame Generation on MetalFX, Metal pipelines compiled on background threads, the fixes reported in #11, and the roadmap's command-stream work.

Fixes for #11

  • Packed geometry shader inputs (Forza Horizon 3's "duplicate member in1", Crimson Desert's packed geometry payload): SPIRV-Cross names geometry-as-mesh payload members per location and component, and reads the right component.
  • GetAttributeAtVertex / PerVertexKHR (Crimson Desert): dxil-spirv's 8-bit vertex index fix from upstream (patches/dxil-spirv/0001). SPIRV-Cross reads per-vertex inputs through Metal's vertex_value where a one-triangle self-test shows it works, and falls back to an interpolated approximation with a warning where it doesn't (macOS 27.2 on M4).
  • Stall report: with METALSHARP_TRACE=sync, after 5 s without a Present, VKD3D-Proton logs the fences being waited on.

Performance (tests/perf scenes, Apple M4)

  • Constant buffer rows out of range read from a zero constant: cbv_rows 410 → 20 ms.
  • WriteBufferImmediate markers batched within a render pass: draw_markers 186 → 1.5 ms.
  • ALL_COMMANDS barriers between dispatches stay in the compute encoder: alias_dispatches 6.5 → 0.66 ms.
  • Counted ExecuteIndirect dispatches run from a compute indirect command buffer: 13.7 → 2.4 ms.
  • One timestamp sample per pool per batch: timestamps 29 → 2.9 ms, queries 29 → 1.8 ms.
  • The implicit pipeline cache is saved outside its lock. Before, the cache file doubled on every launch and never hit.
  • Background pipeline compiling: MoltenVK compiles Metal pipelines on its own threads (MVK_CONFIG_DEFERRED_PIPELINES, on unless 0).

Correctness

  • Timestamps resolved in the same command list that wrote them read their values. The resolve runs on the host, and timeline semaphores never go back.
  • R32G32B32 typed buffer SRVs (Buffer<float3> and the like) work without such Metal texel formats. dxil-spirv reads them as three single-component texels.
  • Metal Shader Converter bridge pipelines get their static samplers. Logic operations and conservative rasterization stay on SPIR-V.
  • A SPIRV-Cross recompile-limit failure is fixed. It left a God of War Ragnarök compute shader unconverted ("D3D Device Crashed").

MetalFX and DLSS

  • DLSS Frame Generation (NGX feature 11) on MetalFX's frame interpolator. This needs macOS 27 and the game's own nvngx_dlssg.dll. NVAPI answers the DRS session and Reflex calls Streamline makes.
  • Ray Reconstruction and Frame Generation are offered only where the game ships their libraries.
  • MetalFX temporal upscales take motion vectors at the output resolution (Hogwarts Legacy).
  • MetalFX spatial upscaling of the swap chain is asked for when the back buffers are larger than the window too. Before, ELDEN RING stayed drawn in a corner of the screen.
  • VKD3D_METALFX_SPATIAL=1/0 overrides vkd3d.conf. VKD3D_CONFIG_FILE may be a Unix path.
  • The arm64 package ships vkd3d.conf and dxvk.conf, and the installer sets them up with the NGX core and NVAPI. This is for Wines other than MetalSharp's.
  • nvngx.dll is signed with the Mullion NGX key, whose certificate the lane's DXGI trusts in the prefix.
  • MoltenVK's frame interpolator is built with the macOS 26 SDK and turned on at run time on macOS 27.

Tests

New D3D12 tests:

  • aliaschain
  • executeindirect_dispatch_count
  • timestamps (including same-list resolves)
  • static_sampler_lod
  • typedbuffer_rgb32 (+ _dxbc)
  • view_kinds_dxil/_dxbc
  • minmax
  • wave_first_helpers

Each change was checked with its own tests and scenes. The full suite was not rerun.

Known gaps

These are not covered in 5.0:

  • Per-draw binding cost
  • Predication
  • Transform feedback
  • DispatchRays
  • Isolines
  • Fence completion under heavy CPU load (about twice Mullion's)

Built from 54537b6. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v5.0-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows). Stage them together.
  • vkd3d-proton-v5.0-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll, ARM64X), which run DLSS Super Resolution and Ray Reconstruction on MetalFX.
  • moltenvk-v5.0-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off). It runs MetalFX's scalers for VKD3D-Proton's and DXVK's DLSS and swap chain upscaling (mvk_metalfx.h).

VKD3D-Proton MacOS v4.9

Choose a tag to compare

@github-actions github-actions released this 02 Nov 01:10

VKD3D-Proton macOS 4.9 is built from main (2f7120c). It brings DLSS on MetalFX and MetalFX upscaling of the swap chain to the VKD3D-Proton lane, with its own NGX core and NVAPI.

What's new

DLSS through MetalFX (Direct3D 12)

  • DLSS Super Resolution runs on MetalFX's temporal scaler, and DLSS Ray Reconstruction on its denoised scaler (macOS 26 or later).
  • It works for games that call NGX directly and for games that go through Streamline (The Witcher 3, God of War Ragnarök).
  • The lane ships its own NGX core and NVAPI (nvngx.dll, nvapi64.dll, ARM64X, built from ngx/). They import only Windows' DLLs and DXGI.
    • NVAPI reports the NVIDIA GPU that DXGI shows (its adapters, LUIDs and description), so Streamline's GPU check matches.
    • Both load only in programs whose DXVK configuration shows an NVIDIA GPU (dxgi.customVendorId = 10de).
  • VKD3D-Proton's command lists record the scales through their own private interface, ID3D12GraphicsCommandListMetalFX.
  • DXGI (DXVK 2.7.1) sets up DLSS for those programs: it adds the NVIDIA driver's NGXCore registry values and trusts the certificate NGX checks its core against.

MetalFX upscaling at present

  • With d3d12.metalSpatialUpscale = True (or d3d11.metalSpatialUpscale) in vkd3d.conf, the swap chain keeps the back buffers' size. MoltenVK then upscales it to the display with MetalFX's spatial scaler instead of stretching it. vkd3d.conf is named by VKD3D_CONFIG_FILE, otherwise it's the working directory's.
  • The output is the layer's size on screen, including in native macOS fullscreen at an emulated display mode.
  • Wine (MetalSharp Wine 0108–0110, 0113) launches the game at the display divided by the factor. It also stops reporting these smaller swap chains as suboptimal, which had made VKD3D-Proton recreate them every frame.

MoltenVK (mvk_metalfx.h)

  • MetalFX temporal and denoised scales ride on vkCmdPipelineBarrier2 (VkMetalFXScaleInfoMVK). The swap chain upscaling is requested with VkSwapchainCreateInfoKHR (VkSwapchainMetalFXSpatialCreateInfoMVK).
  • On 64-bit, Wine's Vulkan layer hands both structures to the driver as they are, so no new Vulkan entry points are needed.
  • Support is announced in driverInfo: metalfx, and metalfx-denoised on macOS 26 or later.

Use it with

  • MetalSharp Wine main cbc408d or later. It sets up DLSS games, the launch size and the lane's DLL loading.
  • MetalSharp with metalsharp/MetalSharp#755, which writes the MetalFX lines into vkd3d.conf and dxvk.conf.

Not yet

  • DLSS Frame Generation.
  • Motion vectors at the output resolution (games that render them that way get a warning).
  • 32-bit programs: Wine's 32-bit Vulkan thunks drop the MetalFX structures.
  • Direct3D 11 games on this lane: they launch at the MetalFX size, but the lane's DXVK 2.7.1 doesn't upscale at present. They get MetalFX on the DXVK lane (DXVK-MacOS v3.1.1-macos.2).

Packages

Built from 2f7120c. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v4.9-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows). Stage them together.
  • vkd3d-proton-v4.9-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows), and the lane's own NGX core and NVAPI (nvngx.dll, nvapi64.dll, ARM64X), which run DLSS Super Resolution and Ray Reconstruction on MetalFX.
  • moltenvk-v4.9-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off). It runs MetalFX's scalers for VKD3D-Proton's and DXVK's DLSS and swap chain upscaling (mvk_metalfx.h).

VKD3D-Proton MacOS v4.8

Choose a tag to compare

@github-actions github-actions released this 31 Oct 03:38

VKD3D-Proton macOS 4.8 is built from main (9daabea). It is the VKD3D-Proton lane MetalSharp ships: the same VKD3D-Proton patch set as 4.0, with MoltenVK and SPIRV-Cross fixes on top. main was reset to that base on Oct 30, so 4.8 does not contain everything 4.5 did (see below).

What's in 4.8

VKD3D-Proton (v3.0.1-485, the 30 patches of 4.0):

  • feature level 12_2 on MoltenVK;
  • D3D11On12;
  • the Metal Shader Converter bridge for compute, vertex and pixel shaders (METALSHARP_MSC=0 turns it off);
  • the MetalSharp trace flag;
  • render pass and per-draw cost fixes;
  • the five-game fixes.

MoltenVK and SPIRV-Cross:

  • Dragon's Dogma 2 launches: MoltenVK's residency set counts each allocation's adds and removes. Before, freed resources were left in the set and the driver faulted at the first vkQueueSubmit2. SPIRV-Cross derives a vertex shader's SubgroupLocalInvocationId from the active-lane mask, so vertex shaders with wave operations convert.
  • Fixes DXVK 3.1 needs (the GRAPHICS_BACKEND=dxvk lane shares this MoltenVK): D3D9 goes from 4 of 8 to 8 of 8 smoke tests (fixed-function, textured, render target feedback and shadow map draws had failed to compile). Ray tracing stays off by default.

The MoltenVK package is universal (arm64 + x86_64).

Not in 4.8 (were in 4.5)

  • Metal Shader Converter bridge: a null descriptor reads a zeroed buffer and texture instead of address and resource ID 0.
  • Metal Shader Converter bridge: shaders with application quirks (Dragon's Dogma 2's heap robustness, RE Engine's culling), and pixel shaders with wave operations, stay on SPIR-V. On the bridge, waterfall loops over WaveReadLaneFirst with helper lanes could hang the GPU.
  • MoltenVK blending RGB9E5 on Apple GPUs.

Both packages carry DXVK 2.7.1 (x86_64) for dxgi, d3d11 and d3d10core, not the DXVK-MacOS 3.1 of the unreleased tag v4.6. MetalSharp's GRAPHICS_BACKEND=dxvk lane ships DXVK-MacOS 3.1 on its own.

Packages

Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v4.8-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows). Stage them together.
  • vkd3d-proton-v4.8-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine). It has VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows).
  • moltenvk-v4.8-universal: the universal, ad hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on. It includes Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex and pixel shaders for VKD3D-Proton.

Smoke-tested before tagging, from the same commit's CI build, on MetalSharp Wine's VKD3D lane (Apple M4): draws, descriptor tables, DXBC and DXIL sampling, blending and depth, MSAA, compute, mesh shaders, ray queries, ExecuteIndirect, queries and a Unity-style frame.

VKD3D-Proton MacOS v4.5

Choose a tag to compare

@github-actions github-actions released this 28 Oct 06:48
153e9fb

VKD3D-Proton macOS 4.5 gets Dragon's Dogma 2 running on the VKD3D-Proton lane. It also makes the Metal Shader Converter bridge leave shaders it can't run correctly on SPIR-V.

What's new since 4.0

Dragon's Dogma 2

  • MoltenVK residency sets now count each allocation's adds and removes. Before, the heap an allocation was in at removal time decided whether it left the set. That left freed resources in the set, and the driver faulted at the game's first vkQueueSubmit2.
  • SPIRV-Cross: a vertex shader's SubgroupLocalInvocationId now comes from the active-lane mask, because Metal vertex functions have no [[thread_index_in_simdgroup]]. The game's depth-only vertex shaders use wave compaction. They failed to convert, and the game exited with E_OUTOFMEMORY.

Metal Shader Converter bridge

  • Shaders with application quirks stay on SPIR-V. The quirks are fixes from upstream's tables: descriptor heap robustness for out-of-range bindless indices, fair scheduling for persistent threads, and wave size clamps. Only dxil-spirv applies them, so the converter compiled those shaders without them. Barrier quirks still apply to converted shaders. Dragon's Dogma 2 has heap robustness on every shader, so it now runs entirely on SPIR-V, as upstream intends.
  • Pixel shaders with wave operations stay on SPIR-V. D3D leaves helper lanes out of a pixel shader's wave operations, but the converted code counts them. Waterfall loops over WaveReadLaneFirst then never ended, and the GPU hung (kIOGPUCommandBufferCallbackErrorHang).
  • Null descriptors now read a zeroed 64 KiB buffer and a 1×1 zeroed texture instead of address 0 and resource ID 0. CBV/SRV/UAV heaps start out filled with them.

MoltenVK: RGB9E5Float render targets blend on Apple GPUs, where the write mask is all channels or none anyway; an M4 measured 0.5 + 0.25 = 0.75. Blending stays off for other Mac GPUs. Before, VKD3D-Proton warned at every RGB9E5 blend state, and additive draws overwrote the target instead of adding.

Results

On an M4 MacBook Air in MetalSharp Wine 4.0, Dragon's Dogma 2 now launches, gets through shader compiling and reaches gameplay on this lane. The bridge changes target the posterized and missing textures, missing geometry, black frames and GPU hangs that a traced run on 4.0 showed; whether they clear all of them is not yet confirmed in game. A full play-through on 4.5 hasn't been done yet; Mullion is the tested lane for this game.

Third-party licenses

Unchanged from 4.0. Each package carries the licenses of what it contains: VKD3D-Proton (LGPL-2.1), DXVK (zlib), and MoltenVK and SPIRV-Cross (Apache-2.0). moltenvk-v4.5-universal also carries Apple's Metal Shader Converter library, unmodified and Apple-signed, under Apple's license, for shader conversion only. See the README's third-party licenses.

Packages

Built from 153e9fb. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v4.5-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows). Stage them together.
  • vkd3d-proton-v4.5-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows).
  • moltenvk-v4.5-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off).

VKD3D-Proton MacOS v4.0

Choose a tag to compare

@github-actions github-actions released this 27 Oct 23:29
25c81a0

VKD3D-Proton macOS 4.0 compiles D3D12 shaders with Apple's Metal Shader Converter, gets Elden Ring and The Witcher 3 (DX12) through their pipeline failures, and fixes what traces of five games showed. It also adds one flag, METALSHARP_TRACE, for tracing VKD3D-Proton and MoltenVK together.

What's new since 3.0

Metal Shader Converter (on by default; METALSHARP_MSC=0 turns it off, METALSHARP_MSC=compute keeps graphics shaders on SPIR-V)

  • Compute shaders, and graphics pipelines of a DXIL vertex and pixel shader, go from DXIL straight to a Metal library in 1–4 ms. Before, they went DXIL → SPIR-V → MSL text, which Metal then compiled in 0.4–1.7 s per large shader. Those compiles were 85% of God of War Ragnarök's pipeline time.
  • Shader Model 6.6 shaders that index the descriptor heaps directly are converted too.
  • Shaders this path can't express stay on SPIR-V, as before. That covers DXBC, local root signatures, other graphics stages, emulated depth bounds, and shaders that read typed buffers: the converter reads typed buffers only from the start of their view.
  • The MoltenVK package carries Apple's libmetalirconverter.dylib, unmodified and Apple-signed, with its license; the installers put it next to libMoltenVK.dylib.
  • Integer division by zero in converted shaders gives the converter's result, not DXBC's 0xffffffff.

Fixes from traces of five games

  • The Witcher 3: a pass with no attachments asked for 17 GiB every frame. It now uses one layer, and MoltenVK caps pixels × layers at 2^30.
  • Ghostrunner: draws with a tessellation + geometry pipeline stopped its queue; they are skipped, with one message.
  • Red Dead Redemption 2 and God of War Ragnarök memory growth: indexed indirect tessellation sized its buffers for a million vertices (about 1.3 GiB per draw) and now uses the bound index buffer. Temporary buffers of 16 MB or more are released after use, keeping up to 512 MB for reuse. God of War's extra Metal memory went from 6.7 GB to about 3 GB.
  • Binary archives now save. They used to fail every time: they leave out pipelines Metal can't serialize, wait until the game has compiled nothing for 2 s, and save every minute and at exit.
  • SPIRV-Cross: Ghostrunner's geometry pipelines and Red Dead Redemption 2's hull shaders compile.
  • MoltenVK:
    • Descriptor sets no longer write past their allocation at the end of VKD3D-Proton's sampler heap.
    • A draw after a converted draw in the same render pass binds its own descriptor sets again. Before, it read the converter's heaps instead, which caused GPU page faults in God of War Ragnarök.

Elden Ring and The Witcher 3 (DX12)

  • D3D12 device creation works on MoltenVK again. A driver ID read under Wine hid MoltenVK, so every D3D12CreateDevice failed.
  • Elden Ring no longer crashes creating its device: pipeline statistics query heaps record nothing on devices without pipeline statistics, instead of failing.
  • Apple's shader compiler no longer crashes on aliased buffer arrays in mesh stages. This cost Elden Ring 9 geometry pipelines per session.
  • Geometry pipelines fetch packed and per-instance vertex attributes. Tessellation with point or isoline output gets a winding order, so The Witcher 3's characters render.
  • Pipelines with dynamic vertex strides no longer clash with MoltenVK's buffer indexes.
  • The Metal HUD heading shows VKD3D-Proton's feature level. Ray tracing state objects log their exports, variants and the hit groups they leave out.

Tracing: METALSHARP_TRACE=<categories>[+heavy] logs present, sync, gpu, memory, shader and more from VKD3D-Proton and MoltenVK into one folder per game (docs/tracing.md). With gpu, a GPU fault names the shaders of the draws that ran before it.

Results

On an M4 MacBook Air in MetalSharp Wine:

  • The Witcher 3 (next-gen, DX12): starts, compiles all its pipelines except one isoline tessellation pipeline (Metal has no equivalent), and its characters render.
  • Elden Ring: creates its device, has 0 failed pipelines and no compiler crashes on a fresh cache, and runs with ray tracing on.
  • Converter tests: the bridge's native tests (tests/msc-bridge) pass, with and without robustness. They cover compute, heaps, vertex and pixel shaders, and converted and SPIR-V draws in one render pass.

Third-party licenses

Each package carries the licenses of what it contains: VKD3D-Proton (LGPL-2.1), DXVK (zlib), and MoltenVK and SPIRV-Cross (Apache-2.0).

New in this release, moltenvk-v4.0-universal includes Apple's Metal Shader Converter library, libmetalirconverter.dylib (© Apple Inc.). It is proprietary, licensed under Apple's Software License Agreement for Metal Shader Converter, which ships next to it with Apple's acknowledgements.

  • What it does here: MoltenVK loads it at runtime to convert D3D12 DXIL shaders into Metal libraries, and uses it for nothing else.
  • Why it can ship: section 2B of the agreement allows distributing the library "for the sole purpose of shader conversion".
  • How it ships: only the library, unmodified and still signed by Apple (section 2E forbids modifying it), in a macOS-only package, with its notices (sections 2C and 2D).

The README's third-party licenses section has the details.

Packages

Built from 25c81a0. Each package holds its files, their licenses and SHA256SUMS.

  • vkd3d-proton-v4.0-x86_64: VKD3D-Proton's d3d12.dll and d3d12core.dll, with DXVK's dxgi.dll, d3d11.dll and d3d10core.dll, for x86_64 Windows programs (x86_64-windows). Stage them together.
  • vkd3d-proton-v4.0-arm64: for an arm64 Wine that runs x86_64 programs with FEX (MetalSharp Wine): VKD3D-Proton's d3d12.dll and d3d12core.dll as ARM64X (aarch64-windows, loaded by x86_64 programs as ARM64EC), with DXVK's x86_64 dxgi.dll, d3d11.dll and d3d10core.dll (x86_64-windows).
  • moltenvk-v4.0-universal: the universal (arm64 + x86_64), ad-hoc signed libMoltenVK.dylib and MoltenVK_icd.json both lanes run on, with Apple's Metal Shader Converter library (libmetalirconverter.dylib, unmodified, Apple-signed), which MoltenVK uses to convert DXIL compute, vertex and pixel shaders for VKD3D-Proton (METALSHARP_MSC=0 turns it off).