DXVK-GPLALL 2.6.9-1 (DXVK-GPLAsync-LowLatency 2.6.9-1)
Based on DXVK 2.6.2 and commits from DXVK 2.7 and later, DXVK GPLAsync 2.6.2 and DXVK GPLAsync 2.7.1-patch, DXVK Low Latency 2.7.1-commit cfc119b.
It consists of:
- DXVK-GPLALL-GCC-WinMacLinux-SSE2-O3-LTO 2.6.9-1
- DXVK-GPLALL-GCC-WinMacLinux-SSE4.2-O3-LTO-ZEN3 2.6.9-1
- DXVK-GPLALL-GCC-WinMacLinux-SSE4.2-O3-LTO-SKYLAKE 2.6.9-1
- DXVK-GPLALL-MSVC-Windows-SSE2-O1_Oi_Ob3-LTCG 2.6.9-1
- DXVK-GPLALL-MSVC-Windows-AVX2-O1_Oi_Ob3-LTCG-AMD64 2.6.9-1
- DXVK-GPLALL-MSVC-Windows-SSE4.2-O1_Oi_Ob3-LTCG-INTEL64 2.6.9-1
- DXVK Native-GPLALL-GCC-SSE2-O3-LTO 2.6.9-1
- DXVK Native-GPLALL-GCC-SSE4.2-O3-LTO-ZEN3 2.6.9-1
- DXVK Native-GPLALL-GCC-SSE4.2-O3-LTO-SKYLAKE 2.6.9-1
- Source Code
- dxvk.conf
Release Notes:
Seventh maintenance release of DXVK-GPLALL 2.6.x branch.
IMPORTANT: Please refer to Quick Start Guide and Builds Reference Guide to achieve the best results on the end-user system.
- Significant changes:
DXVK: 8b751b0 - [dxvk] Disable secondary command buffers on desktop- Improves stability and performance on drivers that do not have proper support for usage of secondary command buffers.
- Specifically fixes bugs on AMD Windows Vulkan Driver on RDNA GPUs in Assassin's Creed Origins - #109 - doitsujin#5813 - doitsujin#5356
- Thanks to chaosophy for reporting the issue and helping to identify the cause quickly.
- Specifically fixes bugs on AMD Windows Vulkan Driver on RDNA GPUs in Assassin's Creed Origins - #109 - doitsujin#5813 - doitsujin#5356
- Improves stability and performance on drivers that do not have proper support for usage of secondary command buffers.
GPLALL: Reduced amount ofLoggercalls, by commenting them out.- Lowers logging overhead, slightly reduces DLLs size.
MSVC: Added a group of optimization compiler flags.- Slightly improves CPU-bound performances in some applications.
GCC: Adjusted tuning settings for each type of GCC build.- Set
-mtune=amdfam10forSSE2build,-mtune=znver3forSSE4.2-O3-LTO-ZEN3build,-mtune=skylakeforSSE4.2-O3-LTO-SKYLAKEbuild, which results in better CPU-bound perfomance for all CPUs that may actually run DXVK. Please refer to Builds Reference Guide for details.
- Set
GCC: Added a group of safeguard compiler flags, which were already default for MSVC.- Ensures consistent behavior across different build types.
GCC: Added a group of optimization compiler flags, with a focus on getting all benefits of-O3and more, with reduced size from-Os- similar to what was done for MSVC previously.- As a result,
GCCbuilds are now at least as fast asMSVCbuilds, and at most are faster thanMSVC, even on Windows. SinceGCCbuilds are the most performant, they are the recommended ones to use.
- As a result,
Native: Changed GitHub Action build container forDXVK Nativebuilds fromSteamRT3-SnipertoSteamRT4.- Changes
GCCversion used to build DXVK-GPLALL from GCC 10 to GCC 14.2.0.- Improves CPU-bound performance of
DXVK Nativebuilds, due to the compiler improvements and support of new CPUs and compiler flags.
- Improves CPU-bound performance of
- Changes
GPLALL: ImplementedSSE2SIMD intrinsics in appropriate functions, where applicable and possible.- Scalar fallbacks for non-x86 are present. Unaligned operations are used due to inherent problems with alignment in DXVK.
AVXis not used because of inherent problems withAVXin DXVK and questionable efficiency.- Improves CPU-bound performance and energy efficiency in D3D8/D3D9 applications.
- Scalar fallbacks for non-x86 are present. Unaligned operations are used due to inherent problems with alignment in DXVK.
DXVK: Clean up device and instance creation and Rework the way device properties are accessed- Allows to backport further improvements from upstream DXVK.
- Thanks to EPOCHvoyager for testing these significant changes.
- Allows to backport further improvements from upstream DXVK.
DXVK: Implement shared resources using the D3DKMT API- Allows Windows applications on Linux to use the new Shared Resources implementation in Proton 11/Wine 11 and later.
- Thanks to Laitinlok for testing these significant changes.
- Allows Windows applications on Linux to use the new Shared Resources implementation in Proton 11/Wine 11 and later.
GPLALL: Optimized existing sleep logic, increased amount of time between wake-up checks in busy-wait phase, set timer resolution (sleepGranularity) to 500 us (0.5 ms, 2000 Hz).- Significantly improves power efficiency and precision, when using DXVK FPS Limiter.
GPLALL: Changed FPS Limiter algorithm to deadline-based, which is based on upstream DXVK Deadline FPS Limiter with some improvements.- Reason: Deadline-based FPS Limiter is the only one, that does work properly under heavy stress, and specific situations, such as usage of
d3d9.maxFrameLatency=1. It is not as smooth as the alternatives, but it is compensated by improvements to the sleep logic as much as possible.
- Reason: Deadline-based FPS Limiter is the only one, that does work properly under heavy stress, and specific situations, such as usage of
GPLALL: Implemented various improvements to the DXVK-GPLALL's implementation of MSAA shader.- Significantly improves GPU-bound performance in applications that use MSAA.
- Fully hand-written paths for every mask and 1x/2x/4x/8x sample count, optimized and simplified logic in every possible place. As a result, shader's size increased, but it actually compiles faster than the previous, and shows significant performance improvements, which are higher for higher sample counts. Basically: MSAA 2x - as fast as possible in performance; MSAA 4x - similar to previous MSAA 2x in performance; MSAA 8x - similar to previous MSAA 4x in performance. And all of that without quality loss.
- Note: Non-standard sample counts, i.e. not MSAA 1x/2x/4x/8x, are using generic paths, so they will be slower, and are not recommended to use.
- Fully hand-written paths for every mask and 1x/2x/4x/8x sample count, optimized and simplified logic in every possible place. As a result, shader's size increased, but it actually compiles faster than the previous, and shows significant performance improvements, which are higher for higher sample counts. Basically: MSAA 2x - as fast as possible in performance; MSAA 4x - similar to previous MSAA 2x in performance; MSAA 8x - similar to previous MSAA 4x in performance. And all of that without quality loss.
- Significantly improves GPU-bound performance in applications that use MSAA.
GPLALL: Implemented changes that improve Fullscreen Exclusive (FSE) behaviour on Windows. Author - KaitouJoker, backport - Digger1955.- Fullscreen Exclusive (FSE) is converted to an emulated Fullscreen Exclusive (eFSE) that uses Independent Flip (iFlip). As a result, enabling Fullscreen Exclusive should not break any functionality of application and provide the best latency and performance.
- Important: Unity Engine have broken FSE implementation. Do not enable FSE in Unity Engine applications.
- Fullscreen Exclusive (FSE) is converted to an emulated Fullscreen Exclusive (eFSE) that uses Independent Flip (iFlip). As a result, enabling Fullscreen Exclusive should not break any functionality of application and provide the best latency and performance.
DXVK: Deduplicate D3D8/D3D9 device lock- Improves CPU-bound performance in D3D8/D3D9 applications.
DXVK: Rework staging buffer logic for D3D11 resource initialization + Part of Rework D3D9 shader constant management- Improves CPU-bound performance.
DXVK Low Latency: cfc119b - [dxvk] Drop 64-bit requirement for GPU timestamps- Any GPU may now use
dxvk.framePace = "low-latency"ordxvk.framePace = "low-latency-vrr-x"modes.
- Any GPU may now use
GPLALL: Implemented various improvements from DXVK-Salkim. Author - nanomatters, backport - Digger1955.- Signifcant improvements to HDR and Nvidia Reflex behaviour and changes that improve CPU-bound performance.
GPLALL: Implemented various improvements from DXVK-Skyrim Community Shaders. Author - doodlum, backport - Digger1955.- Improves CPU-bound performance across various applications, with the most signficant improvements in D3D11 applications.
- Note: Slightly increases VRAM usage.
- Improves CPU-bound performance across various applications, with the most signficant improvements in D3D11 applications.
GPLALL: Changed default DXVK-GPLALL settings in order to provide low latency indxvk.framePace = "max-frame-latency"mode.dxvk.latencySleep = True- enables upstream DXVK latency reduction algorithm, which works well enough without performance regressions;dxvk.allowFse = True- after porting KaitouJoker improvements, it is safe to enable by default;d3d9.maxFrameLatency = 1/dxgi.maxFrameLatency = 1- improves input latency without performance regressions.
GPLALL: Addeddxvk.minSwapchainLatencyoption to prevent forced triple-buffering. Author - KaitouJoker, backport - Digger1955.- Minimizes swapchain latency by allocating only
minImageCount(double-buffering) instead of theminImageCount + 1(triple-buffering). Reduces input-to-display latency by up to ~1 frame (16.67ms at 60fps).- May have positive effect on GPU-bound performance with VSync disabled. Default is
dxvk.minSwapchainLatency = False, due to possibly unstable frame pacing with VSync enabled in GPU-bound applications. Please refer to dxvk.conf Options Guide for details. Additional Info: https://www.4rknova.com/blog/2025/09/12/triple-buffering - https://docs.vulkan.org/samples/latest/samples/performance/hpp_swapchain_images/README.html
- May have positive effect on GPU-bound performance with VSync disabled. Default is
- Minimizes swapchain latency by allocating only
GPLALL: Added set of non-default config options todxvk.conf, which are safe to apply for the vast majority of applications, in order to improve application's performance/pacing/latency . Author - Digger1955.
Next maintenance releases will have less changes, due to more significant incompatibilities between DXVK 2.6.x and next DXVK versions.
List of ported and adapted upstream DXVK commits:
8b751b0 - [dxvk] Disable secondary command buffers on desktop
7fedbb7 - [d3d8] Properly handle W11V11U10 support
6c7162f - [d3d8] Rework legacy D3D compatibility mode handling
eab8045 - [d3d8] Clean up CopyRects logging - [d3d8] Assorted debug loggers cleanup
6227b63 - [d3d8] Warn on unhandled tokens in TranslateVertexShader8 - [d3d8] Assorted debug loggers cleanup
bee7816 - [d3d9] Handle D3D9Ex checks with the D3D compatibility flags
87803a3 - [meta] Switch CI Arch builder image
adeda66 - [meta] Use Arch container directly in CI
c68b929 - [dxvk] Add new device capability class for device creation - Clean up device and instance creation
10f14b5 - [dxvk] Add device capability info to adapter - Clean up device and instance creation
b01a7ae - [dxvk] Read adatper info from new device capability class - Clean up device and instance creation
ad8f984 - [dxvk] Use new capability class for device creation - Clean up device and instance creation
e4eb72d - [d3d11] Remove now unused GetDeviceFeatures method - Clean up device and instance creation
21832ae - [d3d9] Remove GetDeviceFeatures method - Clean up device and instance creation
00a361c - [dxvk] Use proper extension list for extension providers - Clean up device and instance creation
65b3292 - [dxvk] Rework instance creation and device filtering - Clean up device and instance creation
f530591 - [dxvk] Remove old extension enablement code - Clean up device and instance creation
ca1aa20 - [dxvk] Fix device creation issues on devices without sparse queue
10c9ef0 - [dxvk] Call providers' initDeviceExtensions with initialized adapters
b9b5ca5 - Fix usage of copyFeature
9cba329 - [dxvk] Fix device typo
60978eb - [dxvk] Re-add workaround for failed device creation with NVX extensions
e664345 - [dxvk] Query queue ownership transfer properties - Misc opts
9157dcc - [dxvk] Expose queue ownership transfer properties in device - Misc opts
c42bd32 - [dxvk] Fix OOB indexing by queue.family when queue type is absent
b6dd5d1 - [dxvk] Re-query device properties on device creation - Rework the way device properties are accessed
c5ae0c3 - [d3d9,d3d11] Always query properties from device when possible - Rework the way device properties are accessed
a607990 - [d3d9] Take format into account when checking MSAA compatibility - [d3d9] Take format into account when checking MSAA compatibility
53c8df7 - [d3d9] Overhaul sample count validation when creating textures - [d3d9] Take format into account when checking MSAA compatibility
bf8f0cf - [d3d9] Only check hardware sample support if we create vulkan image - Fix black squares in Medieval: Total War 2.
cad8bad - [d3d9] Query adapter capabilities separately - Rework the way device properties are accessed
1945226 - [d3d11] Query device features and properties from actual Vulkan device - Rework the way device properties are accessed
26a8268 - [dxvk] Introduce limite adapter properties - Rework the way device properties are accessed
9e2de40 - [dxvk] Use limited adapter info in OpenVR interop - Rework the way device properties are accessed
caadb5c - [dxvk] Use limited adapter info for device enumeration - Rework the way device properties are accessed
9d521a4 - [dxgi] Use limited adapter info for device enumeration - Rework the way device properties are accessed
90a6daf - [dxgi] Use limited adapter info for monitor enumeration - Rework the way device properties are accessed
27bf88b - [d3d9] Use limited adapter info for 9on12 device selection - Rework the way device properties are accessed
0978983 - [d3d9] Use limited adapter info to get texture memory - Rework the way device properties are accessed
f591b89 - [dxvk] Do not expose features and properties in DxvkAdapter - Rework the way device properties are accessed
7ffd67a - [dxvk] Add config option to disable unified layout support - Implement unified image layouts
9e89133 - [dxvk] Add option to disable NV_raw_access_chains - Disable raw access chains for Granblue Relink
3408045 - [dxvk] Allow querying enabled instance extensions - Update to VK_KHR_present_id2/wait2
df15de2 - [dxvk] Enable KHR_surface_maintenance1 instance extension - Update to VK_KHR_present_id2/wait2
1e599e9 - [dxvk] Enable KHR_swapchain_maintenance1 if supported - Update to VK_KHR_present_id2/wait2
3fac2fa - [dxvk] Open adapter D3DKMT handle from the device LUID - Implement shared resources using the D3DKMT API
c2ca319 - [dxvk] Try creating D3DKMT device handles from their adapter - Implement shared resources using the D3DKMT API
e3a85e5 - [dxvk] Open the D3DKMT semaphore local handle on creation - Implement shared resources using the D3DKMT API
482f2ce - [dxvk] Open the D3DKMT resource local handle on allocation - Implement shared resources using the D3DKMT API
5e425a2 - [d3d11] Use the D3DKMT local handles to allocate shared handles - Implement shared resources using the D3DKMT API
2c2d1bf - [d3d9] Use D3DKMTEscape to set shared resource metadata - Implement shared resources using the D3DKMT API
a26bf9e - [d3d9] Validate shared handle D3DKMT descriptor on import - Implement shared resources using the D3DKMT API
38d61a0 - [d3d11] Use D3DKMTEscape to set shared resource metadata - Implement shared resources using the D3DKMT API
df2cc56 - [d3d11] Use the D3DKMT API to open shared resources - Implement shared resources using the D3DKMT API
e3f0efd - [d3d11] Move mutex support checks into AcquireSync / ReleaseSync - Implement shared resources using the D3DKMT API
38c5d6a - [d3d11] Implement keyed mutexes with the D3DKMT API - Implement shared resources using the D3DKMT API
796208b - [d3d11] Avoid accessing keyed mutex pointer if not available
326aff0 - [include] Add E_ABORT, DXGI_ERROR_ALREADY_EXISTS and use last one - Add E_ABORT, DXGI_ERROR_ALREADY_EXISTS and use last one
8833f14 - [include] remove redefinition of 3 typedefs: UINT64, LPWSTR, LPCSTR - Add E_ABORT, DXGI_ERROR_ALREADY_EXISTS and use last one
83f167d - [d3d11] Load Intel vendor hack DLL
0f939f3 - [d3d9] Rename FormatColor to Format - D3D9: Small cleanup to format code + mini fix
19dd2bb - [d3d9] Use fixed-up format for fine-grained MSAA support check - D3D9: Small cleanup to format code + mini fix
d276233 - [d3d9] Ignore zero extent stretch rect
1aecb5d - [dxvk] Add Kosmickrisp to tilers
b07a478 - [d3d9] Set non-empty sampler LOD range for single-mip textures - [d3d9] Set non-empty sampler LOD range for single-mip textures
9370ef1 - [dxgi] Reserve a hot-patchable entry on the DXGI entry points overlays hook - [dxgi] Reserve a hot-patchable entry on the DXGI entry points overlays hook
fd666da - [dxgi] Fix DXVK_HOTPATCHABLE definition - [dxgi] Fix DXVK_HOTPATCHABLE definition
0e117c3 - [util] Add missing includes - [util] Use hybrid_patchable attribute for DXVK_HOTPATCHABLE on ARM64EC
b749c26 - [dxvk] Explicitly cast this to void * to silence -Wnontrivial-memcall warnings - [util] Use hybrid_patchable attribute for DXVK_HOTPATCHABLE on ARM64EC
5bbce04 - [util] Use hybrid_patchable attribute for DXVK_HOTPATCHABLE on ARM64EC - [util] Use hybrid_patchable attribute for DXVK_HOTPATCHABLE on ARM64EC
6c70fd8 - Annotate ResizeTarget(), ResizeBuffers1() as hybrid_patchable - Annotate ResizeTarget() as __declspec(hybrid_patchable)
1ac57a0 - Fix whitespace - Annotate ResizeTarget() as __declspec(hybrid_patchable)
375c522 - [dxvk] Fix dynamic multisample state not getting applied in some cases
fe8a4b5 - [d3d11] Guard against failed video image relocation
28a3211 - [d3d11] Guard against failed image relocation in ClearUAV
5502fe6 - [dxvk] Guard against failed image relocation in copyImageToBufferCs
708fd7d - [dxvk] Do not always print error messages if image relocation fails
75ecd6e - [dxvk] Properly check device support for sample locations
9cd6755 - [dxvk] Add config option to disable CUDA interop extensions - More CUDA interop workarounds
d647fdf - [d3d9] Pass validation for imported D3D11 textures.
d7ac258 - [d3d9] Ensure to always submit if ThrottleAllocation hits threshold
c833ea7 - [d3d8] Extend legacy discard handling to handle legacy buffer mapping
6bc6d9e - [d3d9] Update window handle during swapchain reset
6bbca91 - [util] Add common code implementation of D3D9 device lock - Deduplicate D3D8/D3D9 device lock
e5bf902 - [d3d9] Use new ScopedDeviceLock for device lock - Deduplicate D3D8/D3D9 device lock
624635c - [d3d8] Use new ScopedDeviceLock for device lock - Deduplicate D3D8/D3D9 device lock
878473b - [util] Enable cached dynamic resources for Skyrim SE
911f1dc - [meta] Set working directory as safe for git in CI
260237c - [util] Use one-arg pthread_setname_np on macOS
dc7f97d - [d3d9] Return -1 from D3DPERF_BeginEvent/EndEvent when not annotating
3dbf33c - [d3d11] Fix incorrect swapchain resize dimensions
efeb15e - [dxvk] Fix lack of forward progress guarantee in presenter - Fix possible presenter deadlock (hopefully)
b128f9f - [dxvk] Work around gamescope bug w.r.t. present fence signaling
d495b30 - [dxvk] Replace frame queue with fixed-size ring buffer - Present timing
3b40b91 - [dxvk] Use GetSurfaceCapabilities2 functionality if available - Present timing
40e0164 - [dxvk] End render pass before recording frame capture markers
b419e80 - [d3d9] Initialize the device child device pointer - [d3d8] Cleanup of shader declaration/function translation
96dfda2 - [d3d8] Use standard masks and shifts in encodeDestRegister - [d3d8] Cleanup of shader declaration/function translation
0abbe8a - [d3d8] Encode float constant destination registers as D3DSPR_CONST - [d3d8] Cleanup of shader declaration/function translation
3c40ca1 - [dxvk] Move EXT_swapchain_colorspace to instance extensions
5cbf108 - [dxvk] Disable dual-source blending if fragment shader only has one export
271ad46 - [d3d9] Clean up repeated 9Ex usage checks - Misc. mini changes for D3D9
a25e33e - [d3d9] Improve GetRenderTargetData validation - Misc. mini changes for D3D9
e2c890c - [d3d9] Don't clear imported textures - Misc. mini changes for D3D9
c51f443 - [dxvk] Remove tryInvalidateDeviceLocalBuffer - Misc. mini changes for D3D9
ec93112 - [d3d9] Fix oversight with systemmem surfaces - [d3d9] Fix oversight with systemmem surfaces
6853015 - [d3d9] Check the correct surface for multisampling in GetRTData - [d3d9] Fix oversight with systemmem surfaces
4bdfb9a - [util] Enable more workarounds for Total War Pharaoh
9b6cdd1 - [dxvk] Handle duplicate signals in signal tracker - Rework staging buffer logic for D3D11 resource initialization
35f1123 - [d3d9] Always align constant buffer slices to full cache lines - Rework D3D9 shader constant management
9d0f002 - [d3d9] Add fast path for recursive device lock - Rework D3D9 shader constant management
ba2b59c - [d3d9] Explicitly call ExecuteFlush from locked contexts - Rework D3D9 shader constant management
94967a3 - [dxvk] Add support for partial asynchronous buffer uploads - Rework D3D9 shader constant management - Rework D3D9 constant buffers
850a013 - [d3d11] Revise locking logic in device initializer - Rework staging buffer logic for D3D11 resource initialization
97a2061 - [d3d11] Bump limits for staging memory in flight - Rework staging buffer logic for D3D11 resource initialization
d6e0e94 - [d3d11] Serialize staging buffer allocation requests in initializer - Rework staging buffer logic for D3D11 resource initialization
3200bb1 - [d3d11] Fix infinite loop in AllocStagingBuffer
984bc31 - [d3d11] Fix submission synchronization with VR interop interfaces - Frame memes
f20f363 - [dxvk] Avoid recreating resolve image on first use - Frame memes
ac912ba - [dxvk] Also allocate reflex frame ID on rendersubmit start - Reflex memes
5b94142 - [dxvk] Handle TRIGGER_FLASH reflex marker - Reflex memes
0e10259 - [d3d11] Support converting d3d12 texture info in D3D11Device::ConvertRuntimeDescriptor().
852d454 - [dxvk] Improve hang detection
71ec7b5 - [dxvk] Fix broken negative timeout for keyed mutex
2c94c5f - [util] Only spoof vendor id in New Vegas
List of ported and adapted upstream DXVK Low Latency commits:
7f35548 - [dxvk] Fix edge case in ringbuffer allocator
e394d0c - Revert "[dxvk] Optimize DxvkCsChunkPool"
cfc119b - [dxvk] Drop 64-bit requirement for GPU timestamps
List of ported and adapted DXVK-GPLALL fork commits:
791fd08 - [perf] util: Optimize timer resolution to 0.5ms and reduce sleep threshold for Windows
4661d2e - [wsi] Conditionally exclude non-Win32 WSI sources on Windows builds
aed8771 - [wsi] win32: Add WS_EX_NOREDIRECTIONBITMAP to bypass DWM composition in fullscreen/borderless
ddc5004 - [presenter] Default allowFse to true and apply WS_POPUP for Windows Independent Flip (iFlip)
d48768b - [dxvk] presenter: Add dxvk.minSwapchainLatency option to prevent forced triple-buffering
List of ported and adapted upstream DXVK's forks commits:
a08a24d - [wsi] Query monitor advanced color state.
90c221f - [wsi] Select monitors from the client area.
4d2e22b - [dxgi] Detect HDR state per monitor.
9a72818 - [d3d11] Align Reflex latency reports
8a4b12b - [wsi] Read HDR metadata from all CTA extensions
fc32e39 - [dxgi] Export CompatValue with no compatibility overrides
d36d8aa - [dxvk] Avoid worker self-waits during queue error recovery
6c7ca87 - [dxvk] Check Reflex sleep and semaphore wait results
990267a - [dxvk] Preserve Reflex frame mappings when reusing ring slots
abb7b7b - [dxvk] Preserve unavailable Reflex timestamps
3f240ba - [util] Handle long environment variable values
596df02 - [presenter] Skip unchanged HDR metadata
3b8cbfe - [dxvk] Acquire swapchain images on the application thread
4488d28 - [dxvk] Restore early swapchain image acquisition
f3673ea - perf(d3d11): cut Skyrim render-thread translation cost
6f9fa87 - wip(perf): session-4 exterior optimizations (keep-alive, array-batch cache, adopt) - Note: Reverted - [dxvk] Port DXVK Low Latency changes - [dxvk] Optimize DxvkSharedAllocationCache::freeAllocation - 9e07b86
0b50e2a - fix(d3d11): use-after-free crash during cell streaming (heap corruption)
42ea6b7 - perf(d3d11): non-owning bindings, made safe via deferred wrapper destruction
d7f4cc0 - refactor(dxvk): production cleanup + minimize FSE divergence - Note: Ported part of commit
24440bb - docs(dxvk): genericize comments for standalone use - Note: Ported part of commit
44539f9 - perf(memory): deepen small dynamic-alloc recycling (BatchCapacity 64->256)
f089c5f - perf(cs): tune DxvkCsChunkSize to the swept sweet spot (64K->32K)
8f7153a - [d3d9] Fix Left 4 Dead 2 dynamic buffer uploads