Version 2.13

Features

• Implement Shader Model 6.8 min-spec
  • SV_StartInstanceLocation
  • SV_StartVertexLocation
  • WaveSize range
  • Implement Vulkan texturing catch-up features (esoteric comparison sampling functions)
• Implement interop for OpenVR / OpenXR on Proton
• Correctly support NULL index buffers with VK_KHR_maintenance6.
• Implement VK_MESA_image_alignment_control. Reduces memory bloat on AMD cards in particular.

Fixes

• Reimplement VK_NV_low_latency2 to fix some issues with heavy stuttering caused by non-monotonic frame IDs.
  Relies on a more recent dxvk-nvapi which can paper over API design issues in Reflex API.
  Requires a more recent NVIDIA driver which fixes some bugs exposed in this new code.
  On older NVIDIA drivers, it should run, but low-latency will not kick in as expected.
• Explicitly disable variable-rate shading when depth-stencil is written in shader.
  Fixes glitched hair rendering in Hellblade 2.
• Correctly expose MSAA features for depth-stencil. Fixes Arma Reforger.
• Fix bugs in MSAA resolve implementation when dealing with custom resolve formats. Fixes Arma Reforger.
• Fix validation error in internal query resolve shader.
• Fix some bugs in wave-ops where helper lanes participated where they were not supposed to.
  Fixes some WaveMatch / WaveMultiPrefix use-cases in the wild.
• Various dxil-spirv fixes to fix invalid control-flow as always.

Performance

• Tweak how we opt-in to ReBAR for UPLOAD heaps. Now, only > 8 GB cards will get it.
  On 8 GB cards, we were regularly hitting the upper limits of what the GPU could hold in VRAM,
  and using ReBAR would be detrimental to performance since there was risk of more important
  memory being demoted to system memory. Works well together with VK_MESA_image_alignment_control
  to free up significant amounts of VRAM. Performance gains from ReBAR on 8 GB were also found to be minimal
  compared to the larger GPUs since we quickly exhausted the limited 512 MiB budget anyway.
• Sub-allocate small image heaps. Avoids heavy stutter in Ghost of Tsushima on desktop.
  (Steam Deck code path does not seem to use small heaps to begin with).
• Improve performance with ROV when used with more complicated shader code patterns.

Workarounds

• Implement a crude workaround for depth-stencil sparse and MSAA sparse.
  • Just allocates a committed resource instead. Not correct, but good enough band-aid.
  • Allows SottR to run on RADV.
• Disable NV_dgcc on Halo Infinite on NV drivers.
• Workaround a missing barrier in AC: Mirage causing random corrupt geometry.

Misc

• Split vkd3d-proton shader cache up by .exe name when using a unified directory with VKD3D_SHADER_CACHE_PATH.
• Implement VK_EXT_device_address_binding_report.

Version 2.12

Features

• Implement support for NVIDIA Reflex through VK_NV_low_latency2. Thanks to NVIDIA for contributing implementation
• Implement D3D12 render pass API (tier 0)
• Implement ID3D12DeviceRemovedExtendedDataSettings stubs. Fixes some games that rely on this existing
• Implement VK_EXT_device_fault. Makes it possible to grab fault information and vendor binary if supported
• Implement VK_EXT_swapchain_maintenance1
  • Allows seamless transition between V-Sync and tearing present modes without stutter
  • Implemented on both Mesa and NV drivers
• Expose Shader Model 6.7 by default if
  VK_KHR_shader_maximal_reconvergence and VK_KHR_shader_quad_control are supported
• Add optimized descriptor copy path on Intel Arc GPUs that support VK_EXT_descriptor_buffer
• Implement fallback for compute shader derivatives on NVIDIA Pascal and older GPUs.
  Allows exposing Shader Model 6.7 by default on Pascal as well (albeit with some known cases where it does not work).
  The workaround is expected to work with any known use of SM 6.6 compute derivatives in the wild

Fixes

• Fix Atlas Fallen black screen due to edge case with MinLODClamp
• Correctly disable alpha-to-coverage if sampler mask is exported
• Fix format feature reports for DXGI_FORMAT_UNKNOWN
• Relax root signature compatibility rules when compiling Ray Tracing pipelines.
  Fixes GPU hang on NV in Warhammer: Darktide
• Fix GPU hang on NV in UE5 Lyra demo
• Explicitly validate stage IO signatures in PSO creation similar to native D3D12 runtime.
  Fixes some scenarios where a game attempts to create an invalid pipeline that should have failed creation
  on native D3D12

Workarounds

• Workaround crash in Resident Evil 4 RT mode when tessellation is enabled
• Workaround mesh shader glitches on NVIDIA in several UE5 titles
• Workaround GPU hang on NVIDIA in World of Warcraft when MSAA is enabled
• Disable RT by default in Persona 3 Reload on Deck

Performance

• Implement VK_NV_raw_access_chains. Significantly improves GPU performance on NV GPUs in some games.
  Games using DXBC instead of DXIL are expected to see more improvements.
  Not all games are expected to see an uplift
• Fix extremely poor GPU performance in some locations in Persona 3 Reload

Debug

• Add support for VKD3D_QUEUE_PROFILE, a simple system profiling method
  • Includes VK_NV_low_latency2 support to debug NVIDIA Reflex sleeps
• Root signature blobs are also dumped when dumping shaders
  • A simple CLI tool to inspect the root-sig blobs is included in programs/
• Misc improvements to breadcrumbs, debug ring, etc
• Pipeline creation failure now dumps PSO creation commands in log

Version 2.11.1

This release is a minor bug-fix release before the holidays.

• Implement COLOR -> STENCIL fallback copy on NVIDIA
• Implement SM 6.6 ResourceDescriptorHeap[] + UAV counters correctly on RADV
• Fix bugged implementation of DXBC resinfo instruction, affecting Avatar: Frontiers of Pandora
• Fix memory type used for DGC preprocess memory on NVIDIA (~5% performance, YMMV)
• Fix crash in Callisto Protocol when booting game with DXR support

More complete MSAA resolve implementation

• Add depth-stencil resolve
• Support typeless formats
• Add MIN/MAX resolve modes
• Implement missing code paths on NVIDIA

Workarounds

• Update workaround for GPU hang in CP77 when using DXR for patch 2.1.
• Remove workaround for NO_DGCC in Halo Infinite on NVIDIA.
• Workaround game bug in Pioneers of Pagonia causing GPU hangs on RADV.

Version 2.11

This release rolls up a bunch of features, perf improvements and bug fixes / workarounds as usual.

Features

DXR enabled by default

VKD3D_CONFIG=dxr is default now, and no longer needed.
There are some special cases where DXR is not enabled by default. The only such current example is
"Hellblade: Senua's Sacrifice" on Deck which force-enables DXR if it is supported, even on Deck.
New semantics are:

• dxr: Force-enable DXR, even when it is considered unsafe
• nodxr: Disable DXR
• dxr11: Removed. dxr already implied DXR 1.1 anyway

Sampler feedback

This feature was the last feature required for FL 12.2 and is implemented through emulation.
As demonstrated in the implementation docs, all
native implementations of this feature are fundamentally broken in some way.
There's also no known game that ships requiring this feature, so we just consider this a checkbox feature.

DX Ultimate (FL 12.2) now exposed by default

On RDNA2+ and Turing+ we can finally expose the DX Ultimate feature set!

Misc

• Implement a bunch of missing "Vulkan-on-D3D12" features
  • IndependentFrontAndBackStencilRefMaskSupported
  • TriangleFanSupported
  • DynamicIndexBufferStripCutSupported
  • DynamicDepthBiasSupported
  • NonNormalizedCoordinateSamplersSupported
  • MismatchingOutputDimensionsSupported
  • PointSamplingAddressesNeverRoundUp
  • RasterizerDesc2Supported
    • Explicit line rasterization mode
  • NarrowQuadrilateralLinesSupported
  • AnisoFilterWithPointMipSupported
• Implement missing MSAD instruction in DXIL, allowing FSR3 to run
• Implement some esoteric DXR features
  • Implement support for multiple mismatching global root signatures in DXR
    • Fixes crash in Battlefield V
  • Implement support for LOCAL_ON_EXTERNAL dependencies in DXR
    • Fixes DXR in Warhammer: Darktide
• Implement support for ExecuteIndirect + Mesh shaders with state changes
  • Currently unused by games

Performance

• Improve performance of NV_device_generated_commands and NV_device_generated_commands_compute by
  reordering and batching command preprocessing
  • We have observed 15% FPS gains in Halo Infinite on RADV
  • 1-2% in Starfield in some test locations
  • Needs pending Mesa work to land to take advantage of this improvement
• Tune memory allocation patterns for DGC preprocess buffers
  • Avoids a lot of allocation churn
  • Greatly reduces CPU overhead on NV

Workarounds

• Work around RADV bug causing GPU hang in RE4: Separate Ways DLC
• Work around RADV bug causing GPU hang in Lords of the Fallen
• Work around Witcher 3 bug causing broken shadows and GPU hangs when enabling DXR
• Work around Cyberpunk 2077 bug when RT is enabled, where game would cause spurious GPU hangs due to accessing descriptor heap out of bounds
• Work around Windjammers 2 bug causing random crashes on startup
• Add support for VK_EXT_image_compression_control to allow for more fine-grained workarounds for broken games running on RADV
• Enable NV_device_generated_commands_compute on latest NV beta drivers
  • 545.x drivers are still disabled until a fix can be confirmed on shipping drivers
• Remove CURB_MEMORY_PSO_CACHE workaround on Mesa 23.2+
  • Should reduce overhead in PSO creation

Fixes

• Misc dxil-spirv changes to fix various bugs in game shaders as usual
• Fix Jurassic World Evolution 2 crashing when enabling DXR
• Fix some deprecation warnings in Meson build system
  • Some submodule locations moved, which may cause minor disruption

Version 2.10

This release rolls up a ton of bug fixes, game and driver workarounds, and other improvements.

Features

DirectStorage MetaCommands

We can now make use of NV_memory_decompression to implement
GPU accelerated GDeflate compression in DirectStorage.
This is demonstrated to work in Ratchet & Clank: Rift Apart.

We also worked around an NV driver bug when using the fallback GDeflate shader.
The fallback works on RADV.

Enhanced Barriers

NOTE: This isn't all that well tested because there are no games shipping with this yet to our knowledge.

Device generated commands for compute

With NV_device_generated_commands_compute we can efficiently implement
Starfield's use of ExecuteIndirect which hammers multi-dispatch COMPUTE + root parameter changes.
Previously, we would rely on a very slow workaround.

NOTE: This feature is currently only enabled on RADV due to driver issues.

Misc

• Support Root Signature version 1.2
• Implement Shader Model 6.7
  • Includes all SM 6.7 features like AdvancedTextureOps, WaveOpsIncludeHelperLanes
  • Caveat: Technically not Vulkan spec compliant implementation, but works fine on at least NV and RADV. Currently implemented as an opt-in option for now in case some game relies on it to work
• Implement CreateSampler2
• Expose inverted viewport / height feature
• Implement RelaxedFormatCasting feature from Enhanced Barriers
• Implement support for adjacency topologies
• Support A8_UNORM format properly by using VK_KHR_maintenance5, allowing A8_UNORM UAVs to work correctly
• Handle range checked index buffers correctly with VK_KHR_maintenance5

New extension use

• VK_EXT_dynamic_rendering_unused_attachments
• VK_KHR_maintenance5
• VK_NV_device_generated_commands_compute

Performance

• Batch acceleration structure builds. Massively improves build performance on at least RADV.
• Massively improve ExecuteIndirect performance when using COMPUTE + root parameter changes when VK_NV_device_generated_commands_compute is enabled.

Fixes

• Fix root signature creation from DXIL library target (DXR) blobs
• Fix some dual source blending PSOs scenarios. Fixes Star Wars Battlefront II
• Implement wave operations in pixel shaders more strictly according to D3D12 rules
• Fix spurious hangs in Ashes of Singularity when using shared fences and wait-before-signal
• Fix PSO caching bug in mesh shaders. Fixes mesh shaders in Unreal Engine 5
• Fix udiv remainder in DXBC, which fixed some Xenia bugs
• Fix query heap tracking bug that was exposed by NV Streamline
• Various DXIL -> SPIR-V fixes as usual
• Rewrote descriptor set layouts to be more robust against application bugs
  • Motivated by Armored Core VI bug (see below)
  • Native D3D12 drivers are also robust against these application bugs :(

Workarounds

• Workaround bad ReBAR performance in Age of Wonders 4
• Remove workaround for KHR_present_wait on NV 535+ drivers
• Workaround Starfield memory corruption issue where it does not correctly query for 4 KiB alignment
• Disable ReBAR usage on Halo Infinite to workaround very poor CPU performance
• Workaround Street Fighter 6 bug causing spurious GPU hangs
  • Also appears to have worked around GPU hangs in Resident Evil 2
• Workaround Armored Core VI bug causing GPU hang on Balteus fight in chapter 1
• Workaround "firefly" glitches in Resident Evil 4 caused by dubious min16float usage
• Workaround "firefly" glitches in Monster Hunter Rise caused by dubious shader requiring particular precise math
• Workaround Unreal Engine 5 breaking if mesh shaders are exposed, but not barycentrics
• Workaround NV driver bug with TIMESTAMP query heaps that could cause spurious GPU hangs
• Workaround broken CFG code generation in Xenia's DXBC emitter

Version 2.9

This release rolls up various development happening over the last months.

d3d12core.dll split

Some games started assuming that the DLLs were laid out similar to AgilitySDK, where
d3d12.dll is just a loader, and d3d12core.dll contains the real implementation.
vkd3d-proton now implements this split as well. It is possible that various scripts must be updated
to accomodate both DLLs now. Once d3d12.dll is installed in a prefix,
only d3d12core.dll needs to be updated, as d3d12.dll is just a trivial shim either way.

Performance improvements

• Greatly reduce system memory requirements on the first run of an application.
  SPIR-V code was held in memory "just in case" it had to be recompiled later, but this is no longer the case.
• Use VK_EXT_graphics_pipeline_libraries to avoid shader compilation stutter in some extreme edge cases.
• Improve performance with certain bad occlusion query patterns in e.g. Elden Ring.
• Improve CPU performance of VK_EXT_descriptor_buffer even further with vendor-specific "ultra-fast" paths.
  • Our microbenchmark for single descriptor copies are now significantly faster than native D3D12 drivers on both RADV and NVIDIA.
  • Intel performance numbers are TBD, but we expect a win there as well.
• Improve VRAM oversubscription behavior when
  VK_EXT_pageable_device_local_memory is supported. (NVIDIA contribution, thanks!)
  This allows us to implement Evict and MakeResident APIs in a useful way.
  VK_EXT_memory_priority is also used for static priorities as a fallback.

Features

• Add VK_EXT_image_sliced_view_of_3d to support sliced 3D UAVs.
• Improve DXR 1.1 support with VK_EXT_pipeline_library_group_handles.
• Implement VK_EXT_fragment_shader_interlock. Completes FL 12.1.
• Move to Vulkan 1.3 as minimum version.
• D3D11on12 interoperability interfaces are now supported.
  It is compatible with DXVK 2.2, which actually implements 11on12.

Legacy swapchain removal

The old swapchain implementation is now gone.
DXVK 2.1 is required as we now share common code.
This means that there is no fallback for pre-2.1 DXVK versions or Wine DXGI anymore.

Native Linux swapchain support

The new swapchain can support Linux native surfaces.
The demo applications in demos/ with functional swapchain now builds on Linux as well.

Driver workarounds

• KHR_present_wait is currently disabled on NVIDIA drivers due to a bug that would occur on some PRIME setups.
  This is supposed to be fixed in the latest beta drivers, and this workaround will eventually be removed
  when we have a confirmed major version that fixes the issue.
• Workaround RADV bug causing memory bloat in shader caches. Can save several 100s of MBs of memory,
  which is important on certain memory hungry titles to avoid instability.
• Workaround NVIDIA bug with concurrent queue submissions using timeline semaphores.
  Fixed a bunch of inexplicable Xid 109 CTX_SWITCH_TIMEOUT errors in many different games.

Bugfixes and game workarounds

Various bugfixes for games as usual. Listing individual games is becoming impractical at this point,
and it's best to refer to other sources for compatibility information with specific games.
As usual, a bunch of fixes in dxil-spirv to fix shader bugs.

Misc

• On Wine, use winevulkan.dll rather than vulkan-1.dll if available.
  Works around some games that hook Vulkan despite using D3D12.
• Improve compatibility with games relying on certain AgilitySDK details.
• Improve build system compatibility with different widl versions.
• VKD3D_CONFIG=dxr now enables DXR 1.1 as well. dxr11 is kept for compat.
• Fix HDR Metadata MinLuminance value.
• Add VKD3D_LIMIT_TESS_FACTORS to work around excessive tessellation. Enabled for Wo Long.
• (Developers) Improve debugging support with more detailed breadcrumb information.
• (Developers) Insert more actionable information in captures and logs.
• (Developers) Log directly to Wine when available. Makes PROTON_LOG=1 more reliable.
• (Developers) Set thread names on Win32 build as well.
• (Developers) Use native Linux calling convention instead of ms_abi.

Version 2.8

This release rolls up some significant new developments before the holidays.

VK_EXT_descriptor_buffer support

This extension is significant in that it removes a ton of CPU overhead.
We already had most of this in place on RADV and Steam Deck,
but this will allow NVIDIA, Intel, Turnip, and other AMD driver implementations to hit the same optimal code paths.
GPU bound performance increases slightly since we can also remove some shader code that was required to workaround lack of descriptor buffers.

New extension requirements

To support descriptor buffers in the code base, these features are now required instead of optional.
Note that these features are widely supported already and is not expected to cause any problems.
If an implementation could support v2.7, it will support v2.8.

• VK_KHR_buffer_device_address
• VK_KHR_push_descriptor

Rewritten support for host accessible images

The entire API feature was rewritten from scratch to support more implementations and edge cases without
a lot of per-application hacks and workarounds.
As the most extreme example of weird API usage, Guardians of the Galaxy should (finally) run well on NVIDIA.

Rewritten swap chain

To most users, this change should be transparent.

• Allow more precise control on latency and frame pacing with VK_KHR_present_wait.
  • mesa-git supports this along with NVIDIA.
  • VKD3D_SWAPCHAIN_LATENCY_FRAMES=n allows user to force a specific amount of latency.
  • Implementation of DXGI latency fences is now correct.
• Reduce CPU overhead on the main thread that presents to swap chain.
• Fixes a spurious hang in Hitman III where game relies on asynchronous present in order to not lock up.
• Win32 specific DXGI code is handled by DXVK. A DXVK build from Experimental or later is required for this to work.
  This allows a native Linux implementation of vkd3d-proton, including swap chain.
• When VK_KHR_present_wait is not supported, behavior should be 1:1 with old implementation.

NOTE: The old swapchain implementation is still in the repository, and is expected to be removed in the next release.
For now, VKD3D_CONFIG=swapchain_legacy can be used to triage any potential issues with the new one.

NOTE: A driver crash was observed on NVIDIA 525.x drivers when running in some PRIME configurations.
For now, we disable use of present_wait on these drivers.

Fixes and workarounds

• Workaround GPU hangs in Spiderman Remastered: Miles Morales (same issue as the original).
• Fix rendering bug with gun damage in Borderlands 3 on RADV.
• Refactor how resizable BAR is used. GPUs with 4 GiB and lower will no longer attempt to use resizable BAR,
  which can avoid some out-of-memory situations.
• Fix GPU hang in Age of Empires IV.
• Fix some minor issues in mesh shader implementation.
• Fix some issues preventing RE: Village from booting on Arc.
• Some last minute frenzied fixes for Witcher 3 next-gen update.
  • All features except RT appears to work on RADV.
  • Hairworks is known to crash GPU on NVIDIA. More investigation is needed to root cause.
  • Some RT effects work on NVIDIA, others don't:
    • GI is reported to work.
    • AO crashes GPU. More investigation is needed to root cause.

Implement minor missing D3D12 features

An obscure feature was stubbed out and forgotten until now.
ID3D12Device1::SetEventOnMultipleFenceCompletion() is now implemented.

Also, implement SetEventOn(Multiple)FenceCompletion for shared D3D12 fences.
Fixes a regression in Gears 5 causing lockup on boot.

Version 2.7

This release rolls up a massive amount of work since the Steam Deck launch in late February
with mostly features and fixes.

Heightened driver requirements

Newer extensions are now required.

• VK_KHR_dynamic_rendering
• VK_EXT_extended_dynamic_state
• VK_EXT_extended_dynamic_state2 (no optional features required)
• VK_KHR_maintenance4

KHR_dynamic_rendering in particular requires Mesa 22.0 or NVIDIA 510 series, which should have
propagated to distributions a long time ago at this point.

NOTE: Proton 7.0 stable series will stick to v2.6 to avoid the hard driver requirement bump.
Proton Experimental and any future stable beyond 7.0 will stick to v2.7 and later.

KHR_dynamic_rendering fixes many previously unsolvable issues, but it required a rewrite,
and it was deemed impractical to support both legacy and modern paths.

Improved pipeline cache

v2.6 introduced support for pipeline libraries, but only for games which made correct use of the D3D12 API.
To improve the situation across the board,
vkd3d-proton now implements an internal "magic" disk cache to enable SPIR-V caching for all games.
It is possible to disable the magic cache and let applications manage the ID3D12PipelineLibrary itself if desired.

To further reduce on-disk footprint of the magic cache, we also make use of VK_EXT_shader_module_identifier
to reduce the vkd3d-proton cache by >95%, since there is no need to store actual SPIR-V data on-disk.

Optimizations

Mostly a lot of minor things this release.

• Slightly improve GPU performance for depth render passes.
• Improve GPU performance for certain floating-point images where UAV usage was enabled.
• Improve GPU performance for certain use cases of WriteBufferImmediate().
• Improve GPU performance for certain access patterns of root descriptors.
• Improve GPU performance for back-to-back buffer-image copies.
• Improve GPU performance when allocating large zero-cleared resources and heaps.
• Misc things here and there to reduce overhead.

New D3D12 features

Mesh shaders

VK_EXT_mesh_shader is required for this. Directly compatible with D3D12.

Advanced ExecuteIndirect

Uses VK_NV_device_generated_commands. Supported by both RADV and NVIDIA. Allows Halo Infinite to run.

DXR 1.1

Implement some missing features from DXR 1.1:

• AddToStateObject()
• ExecuteIndirect trace rays
• Various complex RTPSO features
• DXIL subobject parsing
• Misc query features

With these fixes in place, e.g. Cyberpunk 2077 DXR works. VK_KHR_raytracing_maintenance1 is required for some features.

NOTE: VKD3D_CONFIG=dxr11 is required to enable DXR 1.1 for now.

Shared resources

Basic shared resources and fences are now supported when running on Proton. Allows interop with DXVK.
Special thanks to Derek Lesho (@Guy1524) for implementation.

SV_Barycentrics

SM 6.1 barycentrics are now exposed through VK_KHR_fragment_shader_barycentric.

Preliminary HDR support

vkd3d-proton can take advantage of HDR now, assuming the system itself supports it.

Game fixes and workarounds

• Fix random GPU hangs in Hitman 3.
• Fix crash in Redout 2.
• Fix random GPU hang in F1 2021.
• Fix random flicker in Guardians of the Galaxy.
• Update some API checks required by latest AgilitySDK runtime features. Fix crash in F1 2022.
• Add various workarounds for game bugs in Halo Infinite.
• Add workaround for amdgpu kernel issue for certain games using imported host memory and multiple Vulkan devices.
• Workaround glitched rendering in F1 2020 due to game bug.
• Workaround certain games that violate placed resource API w.r.t. subresource initialization.
  Spiderman Remastered and Lost Judgment are affected. More games will likely surface.

DXIL support

Countless bug fixes for games released since last release. Too many to enumerate individually.

Misc

• Improve compatibility with Intel ANV driver.
• Improve correctness of GetFrameLatencyWaitableObject().
• Add BLOB PIX decoding.
• Improve stability when minimizing and alt-tabbing in and out of fullscreen in some games.
• Preparation for MIT re-license is underway.

Stronger debugging facilities

• For developers and power users, a breadcrumbs functionality is added to greatly aid GPU hang debugging.
  Requires either VK_AMD_buffer_marker or VK_NV_device_diagnostic_checkpoints.
• When capturing with RenderDoc, cached host memory is enabled by default to speed up capture and improve stability.
• Improve shader replacement system ease-of-use.

Version 2.6

It has been a long while since 2.5, and this release rolls up a lot of fixes, features and optimizations.

Fixes

• Fix black screen rendering bug in Horizon Zero Dawn after latest game updates.
• Fix crashes on startup in Final Fantasy VII: Remake and Warframe.
• Fix crashes in Guardians of the Galaxy when interacting with certain game objects.
• Fix hang on game shutdown in Elden Ring.
• Fix broken geometry rendering in Age of Empires: IV.

Optimization

• Improve generated shader code for vectorized load-store operations in DXIL.
• Greatly reduce CPU overhead for descriptor copy operations,
  which is a key contributor to CPU overhead in D3D12.

Features

Pipeline library rewrite

Support D3D12 pipeline libraries better where we can now also cache
generated SPIR-V from DXBC/DXIL.
Massively reduces subsequent load times in Monster Hunter: Rise,
and helps other titles like Guardian of the Galaxy and Elden Ring.
Also lays the groundwork for internal driver caches down the line for games which do not use this API.
Also, deduplicates binary blobs for reduced disk size requirements.

Shader models

Shader model 6.6 is now fully implemented. This includes support for:

• ResourceDescriptorHeap[] direct access
• 64-bit atomics
• IsHelperLane()
• Compute shader derivatives
• WaveSize attribute
• Packed math intrinsics

Minor features

• Handle API feature MinResourceLODClamp correctly if VK_EXT_image_view_min_lod is supported.
• Expose CastFullyTypedFormat feature.
• Expose some advanced shader features on Intel related to UAV formats (VK_KHR_format_feature_flags2).
• Support COLOR -> STENCIL copies.

Workarounds

• Workaround DEATHLOOP not emitting synchronization commands correctly. Fixes menu flicker on RADV.
• Workaround quirky API usage in Elden Ring. Removes many kinds of stutter and chug when traversing the scenery.
• Workaround certain environments failing to create Vulkan device if some VK_NVX_* extensions are enabled.
• Workaround glitched foliage rendering in Horizon Zero Dawn after latest game updates.
• Workaround some questionable UE4 shaders causing glitched rendering on RADV.

Note on future Vulkan driver requirements

2.6 is expected to be the last vkd3d-proton release before we require some newer Vulkan extensions.
VK_KHR_dynamic_rendering and VK_EXT_extended_dynamic_state
(and likely dynamic_state_2 as well) will be required.

VK_KHR_dynamic_rendering in particular requires up-to-date drivers and the legacy render pass path
will be abandoned in favor of it. Supporting both paths at the same time is not practical.
Moving to VK_KHR_dynamic_rendering allows us to fix some critical flaws with the legacy API
which caused potential shader compilation stutters and extra CPU overhead.

Version 2.5

This is a release with a little bit of everything!

Features

DXR progress

DXR has seen significant work in the background.

• DXR 1.1 is now experimentally exposed. It can be enabled with VKD3D_CONFIG=dxr11.
  Note that DXR 1.1 cannot be fully implemented in VK_KHR_ray_tracing's current form, in particular
  DispatchRays() indirect is not compatible yet,
  although we have not observed a game which requires this API feature.
• DXR 1.1 inline raytracing support is fully implemented.
• DXR 1.0 support is more or less feature complete.
  Some weird edge cases remain, but will likely not be implemented unless required by a game.
  VKD3D_CONFIG=dxr will eventually be dropped when it matures.

Some new DXR games are starting to come alive, especially with DXR 1.1 enabled,
but there are significant bugs as well that we currently cannot easily debug.
Some experimental results on NVIDIA:

• Control - already worked
• DEATHLOOP - appears to work correctly
• Cyberpunk 2077 - DXR can be enabled, but GPU timeouts
• World of Warcraft - according to a user, it works, but we have not confirmed ourselves
• Metro Exodus: Enhanced Edition -
  gets ingame and appears to work? Not sure if it looks correct.
  Heavy CPU stutter for some reason ...
• Metro Exodus (original release) - GPU timeouts when enabling DXR
• Resident Evil: Village - Appears to work, but the visual difference is subtle.

It's worth experimenting with these and others.
DXR is incredibly complicated, so expect bugs.
From here, DXR support is mostly a case of stamping out issues one by one.

NVIDIA DLSS

NVIDIA contributed integration APIs in vkd3d-proton which enables DLSS support in D3D12 titles in Proton.
See Proton documentation for how to enable NvAPI support.

Shader models

A fair bit of work went into DXIL translation support to catch up with native drivers.

• Shader model 6.5 is exposed.
  Shader model 6.6 should be straight forward once that becomes relevant.
• Shader model 6.4 implementation takes advantage of VK_KHR_shader_integer_dot_product when supported.
• Proper fallback for FP16 math on GPUs which do not expose native FP16 support (Polaris, Pascal).
  Notably fixes AMD FSR shaders in Resident Evil: Village (and others).
• Shader model 6.1 SV_Barycentric support implemented (NVIDIA only for now).
• Support shader model 6.2 FP32 denorm control.

Performance

Resizable BAR can improve GPU performance about 10-15% in the best case, depends a lot on the game.
Horizon Zero Dawn and Death Stranding in particular improve massively with this change.

By default, vkd3d-proton will now take advantage of PCI-e BAR memory types through heuristics
as D3D12 does not expose direct support for resizable BAR, and native D3D12 drivers are known to use heuristics as well.
Without resizable BAR enabled in BIOS/vBIOS, we only get 256 MiB which can help performance,
but many games will improve performance even more
when we are allowed to use more than that.
There is an upper limit for how much VRAM is dedicated to this purpose.
We also added VKD3D_CONFIG=no_upload_hvv to disable all uses of PCI-e BAR memory.

Other performance improvements:

• Avoid redundant descriptor update work in certain scenarios (NVIDIA contribution).
• Minor tweaks here and there to reduce CPU overhead.

Fixes and workarounds

• Fix behavior for swap chain presentation latency HANDLE. Fixes spurious deadlocks in some cases.
• Fix many issues related to depth-stencil handling, which fixed various issues in DEATHLOOP, F1 2021, WRC 10.
• Fix DIRT 5 rendering issues and crashes. Should be fully playable now.
• Fix some Diablo II Resurrected rendering issues.
• Workaround shader bugs in Psychonauts 2.
• Workaround some Unreal Engine 4 shader bugs which multiple titles trigger.
• Fix some stability issues when VRAM is exhausted on NVIDIA.
• Fix CPU crash in boot-up sequence of Far Cry 6 (game is still kinda buggy though, but gets in-game).
• Fix various bugs with host visible images. Fixes DEATHLOOP.
• Fix various DXIL conversion bugs.
• Add Invariant geometry workarounds for specific games which require it.
• Fix how d3d12.dll exports symbols to be more in line with MSVC.
• Fix some edge cases in bitfield instructions.
• Work around extreme CPU memory bloat on the specific NVIDIA driver versions which had this bug.
• Fix regression in Evil Genius 2: World Domination.
• Fix crashes in Hitman 3.
• Fix terrain rendering in Anno 1800.
• Various correctness and crash fixes.