DXVK (DirectX over Vulkan) - info a vše okolo

[Krteq]

DXVK 2.4.1

General improvements

• The size for memory chunks is now determined dynamically based on the amount of memory the application has already allocated.
  This should improve the out-of-the-box behaviour in various game launchers; the dxvk.maxChunkSize option was removed accordingly.
• Fixed an issue where the Vulkan swap chain would not always be recreated appropriately in a native Wayland environment. (PR #4252)
• Fixed an issue with descriptor pools growing too large on Nvidia cards in some situations. (PR #4166)
• Changed default shader code generation for DXBC mad instructions to work around flickering issues in games that use different vertex shaders to render the same geometry in multiple render passes. The d3d11.longMad and d3d9.longMad options were removed accordingly.

D3D8 / D3D9

• Improved overall robustness of the D3D8 implementation with error handling and fixed various memory leaks.
• Slightly improved performance by optimizing vertex buffer uploads. (PR #4275)
• Significantly improved performance in some games using Software Vertex Processing. (PR #4274)
• Fixed full-screen presentation issues in some Game Maker games. (PR #4245)
• Removed VK_IMAGE_USAGE_SAMPLED_BIT from textures that don't need it. (#4176)
• Enabled strict float emulation by default on AMDVLK now that this driver optimizes the pattern emitted by DXVK. (PR #4203). Do note that this was already the case on RADV.
• Chrome: Gold Edition: Enabled 60 FPS limit to work around game issues at higher frame rates. (PR #4260)
• GTA: San Andreas: Fix regression that made it crash with DXVK 2.4 (#4141, PR #4151)
• Operation Racoon City: Worked around a game crash. (#4172, PR #4175)
• Prince of Persia (2008): Enabled 240 FPS limit to work around game issues at very high frame rates. (PR #4281)
• Rayman 3: Fixed missing geometry. (PR #4143)
• Serious Sam 2: Fix in-game videos (#4158, PR #4161)
• Splinter Cell: Pandora Tomorrow: Enabled 60 FPS limit to fix broken input and physics. (PR #4254)
• The First Templar: Fixed lighting issues. (#4291, PR #4292)
• The Sims 2: Fix regression that made it crash with earlier DXVK versions (PR #4277)

D3D11 / DXGI

• Worked around a general issue wherein some games default to an integrated GPU based on the amount of available video memory.
• Worked around an issue where various games would misbehave when the amount of available VRAM is reported as a power of two.
• Changed shader code generation for dot product instructions to be more in line with code generation on native drivers.
  This fixes a water rendering issue in Trails through Daybreak on Nvidia. (#4162)
• Implemented a basic dead code elimination pass to avoid generating invalid SPIR-V shaders in rare situations.
  Note: All these shader-related changes will invalidate fossilize databases.
• Games will now leave fullscreen mode when the game window loses focus.
  This is closer to Windows behaviour and fixes various issues with games hanging on some setups. (PR #2675)
• Made sure the built-in options in DXVK for a game take precedence over the options enabled/disabled by DXVK-NVAPI.
  This fixes a performance regression in God of War on Nvidia GPUs.
• Batman Arkham Knight: Worked around an issue where the game does not start when it detects an Intel GPU. (PR #4152)
• Hunt Showdown: Fixed a crash caused by DXVK reporting a very high UMD driver version.
• Kena: Bridge of Spirits: Worked around a water rendering issue when the game detects an Intel GPU. (PR #4207)
• Microsoft Flight Simulator: Fixed garbled screen on start-up (#4276)
• Senran Kagura Peach Ball: Fixed character models being rendered the wrong way around. (#4216)
• The Sims 4: Work around a crash caused by a use-after-free bug in the game.

GitHub - doitsujin - DXVK

[sam_217]

Další zajímavá věc se starší hrou - nyní testuji přechod z RX 6800 XT na RTX 3080 Ti, v AC Odyssey jakožto AMD supported hře se té RX 6800 XT daří o dost lépe, než 3080 Ti (111 vs 88 fps AVG v built-in Benchmarku)... viděl jsem, že ale je tam nějaký problém s frame pacingem a GPU usage taky haprovalo, takže bylo jasné, že důvod proč má RX 6800 XT tady navrch proti 3080 Ti není zrovna kvůli HW... ¨

Tak jsem zkusil DXVK a ejhle, problémy se stuttery/frame pacingem na 3080 Ti jsou pryč a výsledek v benchmarku se vyrovnal výsledku 6800 XT (tj. plichta)... ještě jsem nekontroloval vizuálně pod lupou, jestli je vše v pořádku jak má být, ale nastavení detailů ve hře bylo vždy maximální a zběžně OK.

[Krteq]

DXVK 2.5

Memory managment
Resource and memory management were completely rewritten in order to use allocated video memory more efficiently:

• Reduced fragmentation may reduce peak memory usage in games such as God of War by up to 1 GiB in extreme cases.
• Memory defragmentation is now performed periodically to return some unused memory back to the system.
  The goal is not to reduce VRAM usage at all costs; instead this is done conservatively if the system is under memory pressure, or if a significant amount of allocated memory is unused. Keeping some unused memory is useful to quickly service subsequent allocations.

Note: Defragmentation is currently disabled on Intel's ANV driver, see #4434. The dxvk.enableMemoryDefrag config option can be set to enable or disable this feature via the the Configuration file.

Driver support
While technically not required, the new memory management works best on drivers that support both VK_EXT_memory_budget and VK_KHR_maintenance5. The Driver Support page was updated accordingly.

D3D8 / D3D9
Software cursor
Support for emulated cursors was implemented for the D3D9 cursor API, which allows games to set an arbitrary image as the mouse cursor. This fixes an issue in Dungeon Siege 2 (#3020) and makes the cursor appear correctly in Act of War and various older D3D8 games. (PR #4302)



Sampler pool
Unreal Engine 3 games using D3D9 have a quirk in that they pass a seemingly uninitialized value as the mipmap LOD bias. In order to avoid creating more Vulkan sampler objects than the driver supports, previous versions of DXVK would round the LOD bias to a multiple of 0.5, which could introduce visual inaccuracies. As a more correct soluition, DXVK will now destroy unused Vulkan samplers on the fly and use the correct LOD bias.

Note: The aforementioned workaround was never needed or used in the D3D11 implementation, it only affected D3D9.

Bug fixes and Improvements

• On Nvidia driver version 565.57.01 and newer, strict float emulation is enabled by default for improved correctness.
  Games for which this option was already enabled may see a small performance uplift on this driver.
• Made various changes to potentially improve performace on certain mobile GPUs. (includes PR #4358)
• Display modes are now ordered by refresh rate to be more consistent with wined3d and fix issues with some games picking the wrong display mode.
• Fixed a large number of wine test failures.
• Ascension to the Throne: Fixed old regression that would cause parts of the ground to render black. (#4338, PR #4341)
• Command & Conquer: Generals: Fixed performance issue caused by a missing D3D8 entry point. (PR #4342)
• King's Bounty: Warriors of the North: Fixed water rendering issue. (#4344, PR #4350)
• Tomb Raider: Legend: Fixed flickering geometry with strict float emulation. (#4319, PR #4442)
• Rayman 3: Fixed a regression that caused rendering issues. (#4422, PR #4423)

D3D11 / DXGI
Resource management changes
In order to reduce system memory pressure and improve stability in 32-bit games, creating, uploading and discarding resources is now throttled if the amount of temporary staging memory allocations exceed a certain threshold. This fixes crashes in Total War: Rome II and a number of other games. Additionally, large DYNAMIC textures commonly used for video playback will no longer use a staging buffer.

The d3d11.maxDynamicImageBufferSize and d3d11.maxImplicitDiscardSize options were removed accordingly; affected games such as Total War: Warhammer III and Ryse: Son of Rome should now perform well by default, without excessive memory usage.

Note: These changes negatively affect CPU-bound performance in a number of games, including Shadow Warrior 2.

Bug fixes and Improvements

• SEQUENTIAL swap effects are now implemented for DXGI swap chains, which allows games to read previously presented backbuffers. This fixes an issue wherein savegame thumbnails would appear black in certain visual novels. (ValveSoftware/Proton#7017)
• Devirtualized some D3D11 method calls to improve compatibility with Special K.
• Fixed incorrect shader code generation for EvaluateAttributeSnapped.
• Lock contention is reduced in certain games that use Deferred Contexts for rendering.
  This may improve performance on older CPUs in Sekiro: Shadows Die Twice and some other games.
• Call of Duty: Modern Warfare 2 Campaign Remastered: Fixed a possible GPU hang. (#3884)
• Diablo 4: Work around an issue where the game does not start if an integrated GPU is exposed.
• The Sims 4: Work around a use-after-free bug in the game's D3D11 renderer for real this time. (#4360)
• Vindictus: Work around potential rendering issues caused by uninitialized constant buffer data. (#4405, #4406)
• Yakuza 0 and Yakuza Kiwami: Fixed a regression introduced in DXVK 2.4.1 that would cause these games to lock up on start. (PR #4297)

Miscellaneous changes

• An SDL3 backend was added for dxvk-native. (PR #4326, #4404)
• Fixed an issue introduced in DXVK 2.4.1 which would lead to error messages about failed buffer creation.
• Fixed a long-standing issue where overlapping occlusion queries would lead to incorrect Vulkan usage. (#2698)
• Fixed a rare issue wherein timestamp queries would not be tracked correctly and could read incorrect data.
• Fixed various other issues that led to Vulkan validation errors in games such as Dishonored 2, Tales of Arise and The Sims 4.
• Fixed various issues with MSVC builds. (PR #4444)
• Disabled a workaround for boken render target clears on Nvidia drivers prior to version 560.28.03 on unaffected drivers.
• If supported, VK_EXT_pageable_device_local_memory is now used to enable better driver-side memory management.

GitHub - doitsujin - DXVK

[Krteq]

DXVK 2.6

D3D10 / D3D11
Nvidia Reflex support
When running Proton Experimental and an Nvidia driver with VK_NV_low_latency2 support, Reflex can now be enabled in D3D11 games that support this feature. This includes God of War, Overwatch 2, Quake Champions and some others. Approximate input latency can be displayed in these games by setting DXVK_HUD=latency.

Note: Reflex does not work in most Unreal Engine 4 games in D3D11 mode since LatencySleep and other related functions never get called. This cannot be worked around inside DXVK; prefer using D3D12 for affected games.

Bug fixes and Improvements

• Worked around a common game bug that would cause garbled output on Nvidia when using MSAA. (#4480, PR #4731)
  This affects games such as Assassin's Creed 3 / Black Flag, Watch Dogs, Stalker: Clear Sky, and others.
  Note: This may have a small impact on performance and memory usage in affected games, including on AMD GPUs, where this issue did not manifest in rendering issues due to hardware details.
• Potentially improved GPU-bound performance in games that use NVAPI UAVOverlap feature, such as Baldurs Gate 3.
  While this feature was supported before, our previous implementation was too conservative. (PR #4691)
  Note: Like other vendor extensions, this does not work on Windows.
• Removed d3d11.dcSingleUseMode option. Submitting deferred command lists multiple times will now behave as expected by default.
  Note: This may increase memory usage in some games. Fixes rendering issues in Cardfight!! Vanguard Dear Days 2.
• Slightly improved GPU-bound performance in modern games that extensively use pixel shader UAVs, such as Trine 5. (PR #4693)
• Slightly reduced CPU overhead in some games that use resource binding methods inefficiently, such as God of War.
• Fixed layered D3D11 video processor view creation. (PR #4651)
• Added support for the ID3D11On12Device1 interop interface.
• Clanfolk: Worked around a game bug that would cause missing geometry on Intel. (PR #4716)
• Kingdom Come: Deliverance: Fixed a case of invalid Vulkan usage. (PR #4684)
• Watch Dogs 2: Worked around a game bug that would cause the sky to flicker on RDNA3 GPUs. (PR #4717)

D3D8 / D3D9

• Fixed a race condition that could lead to crashes in games doing multi-threaded asset loading. (#4613, PR #4646)
• Fixed various issues that would occur when disabling shader support via the d3d9.shaderModel option. (#4625, PR #4629)
• Fixed an issue where games using both software and hardware cursors would not transition between the two smoothly. (PR #4670)
• Fixed a regression that would cause textures to not be bound correctly in some games. (PR #4694)
• Fixed an issue where games that request half-rate Vsync would run at quarter-rate instead.
• Fixed minor issues that would cause wine test failures. (PR #4732)
• CivCity: Rome: Added vendor ID override to make the game render shadows on Intel GPUs. (PR #4739)
• Global Operations: Fixed rendering issues (#4138, PR #4678)
• Need For Speed: Hot Pursuit 2: Fixed missing fog. (#4113, PR #4653)
• Silent Hill 2: Fixed an issue where the menu would render incorrectly with the Silent HIll 2 Enhancements mod. (#3943, PR #3948)
• Tom Clancy's Splinter Cell: Worked around alt+tab and shadow rendering issues. (PR #4656, PR #4660)
• Trainz v1.3: Worked around alt+tab issue. (#4116, PR #4659)

General Improvements

• Reworked Vulkan swapchain implementation to be more robust. (PR #4609)
• Improved efficiency of multisample resolves on tiling GPUs.
• Fixed a regression introduced in DXVK 2.5 that would cause memory allocation errors on certain unified memory setups, including the Qualcomm proprietary driver.
  Note: This does not mean that DXVK will officially support Android, or proprietary mobile drivers.
• Fixed a crash if an application unloads the SetupAPI DLL. This issue affected various AdvHD-based visual novels. (#4662, PR #4668)
• Fixed an issue where software rasterizers such as Lavapipe would not be considered when setting DXVK_FILTER_DEVICE_NAME. (PR #4716)
• Resource initialization is now always performed on the asynchronous transfer queue if possible, and no longer uses additional memory on drivers that support Vulkan sparse residency features. This may save up to 64 MiB of VRAM in some games.
• Worked around an std::regex bug that would cause games to instantly crash when using Japanese locale. (PR #4669)
• Worked around potential 32-bit crashes caused by stack alignment issues in GCC-compiled binaries. (PR #4627, PR #4633)

The Developer guidelines wiki page was updated to reflect current implementation details.

GitHub - doitsujin - DXVK