Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[viogpu3d] Virtio GPU 3D acceleration for windows #943

Open
wants to merge 2 commits into
base: master
Choose a base branch
from

Conversation

max8rr8
Copy link

@max8rr8 max8rr8 commented Jul 18, 2023

Hello! This series of changes spanning across multiple repositories introduce support for 3d accelerated virtiogpu windows guests.

Demo image

Wglgears window is rendered with wgl on virgl and window below it is cube rendered with d3d10umd on virgl.

How to test

NOTE: This driver does have some rendering glitches and might crash. Try at your own risk.
0. Create qemu windows VM with VirtIO GPU with 3d acceleration enabled. It is highly recommended to use "disposable" virtual machine to test, loss of data might occur.

  1. Use patched version of virglrenderer from this repo branch viogpu_win
  2. Compile from source OR download pre-built drivers.
  3. Install drivers on target VM. Note: if drivers were not signed you need to manually select them in device manager.

Known issues

  • FIXED: Frames displayed on screen are lagging behind

  • FIXED: D3d10 clearing color is not supported

  • FIXED: D3d10 applications using DXGI_SWAP_EFFECT_DISCARD and DXGI_SWAP_EFFECT_SEQUENTIAL are not displayed.

  • Rendering glitches in WinUI3 apps.

    There are some rendering glitches in apps based on WinUI3 (maybe other apps too), best way to see them is to install WinUI3 Gallery from microsoft store and navigate around it. Haven't yet invistigated.

  • Vscode (possibly other electron apps) does not render

    Black window. Requires implementation of PIPE_QUERY_TIMESTAMP_DISJOINT in virglrenderer.

  • No preemption

    Kernel-mode driver does not implement preemption, and i am very confised about how to implement it in WDDM. VioGpu3D disables preemption systemwide to workaround lack of preemption implementation, but this is not ideal. Would appreciate some help.

Siblings

@vrozenfe
Copy link
Collaborator

@max8rr8
Hi Max,

Thank for a very impressive work you've done.
Please give us some time to go through your code and
see how to integrate it into our upstream repository.

Nice work!
All the best,
Vadim.

5. Build and install mesa binaries: `ninja install`
6. Go to `viogpu` directory of this repository
7. Run: `.\build_AllNoSdv.bat`
8. Compiled drivers will be available in `viogpu\viogpu3d\objfre_win10_amd64\amd64\viogpu3d`
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @max8rr8,

This instruction is not clear to me. Can you please describe it more with directory examples? What dependencies should be installed and how configured? (Example: https://github.com/virtio-win/kvm-guest-drivers-windows/wiki/Building-the-drivers-using-Windows-11-21H2-EWDK).
I looked into the mesa compilation guide and it looks completely different.

meson setup builddir/
meson compile -C builddir/
sudo meson install -C builddir/

Several questions:

  1. What is %MESA_PREFIX% and where it defined?
  2. Can we use any precompiled MESA for Windows (for example https://fdossena.com/?p=mesa/index.frag)

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is %MESA_PREFIX% and where it defined?

%MESA_PREFIX% environment variable is set during build process, it points to directory where mesa installs its files.

Can we use any precompiled MESA for Windows

I don't think so, as pointed out in mesa MR when building user-mode driver we have to build it with specific mesa flags to build only virgl driver (to avoid conflicts)

Copy link
Author

@max8rr8 max8rr8 Jul 20, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Building instruction

NOTE 1: this is valid for now, until changes are not in upstream repositories
NOTE 2: this assumes that all build dependencies(meson, WDK, ninja, etc...) are installed

Part 1: Virglrenderer

On host machine it is required that patched version of virglrenderer is used.

  1. Acquire source code git clone --branch viogpu_win https://gitlab.freedesktop.org/max8rr8/virglrenderer && cd virglrenderer
  2. Create install directory mkdir install and build directory: mkdir build && cd build
  3. Configure build meson --prefix=$(pwd)/../install (we set prefix to install libvirglrenderer not globally but to previously created dir install)
  4. Compile and install ninja install
    Now ensure that qemu loads libvirglrenderer from install directory, this can be done by setting LD_LIBRARY_PATH to something like /some/path/to/starter_dir/virglrenderer/install

Part 2: Build mesa

Now inside virtual machine with build tools installed create working directory, then inside it (this assumes use of Powershell):

  1. Create mesa prefix dir mkdir mesa_prefix and set env MESA_PREFIX to its path: $env:MESA_PREFIX="$PWD\mesa_prefix"
  2. Get patched mesa source code git clone --depth 10 --branch viogpu_win https://gitlab.freedesktop.org/max8rr8/mesa and then cd into it cd mesa
  3. Create build directory mkdir build && cd build
  4. Configure build meson .. --prefix=$env:MESA_PREFIX -Dgallium-drivers=virgl -Dgallium-d3d10umd=true -Dgallium-wgl-dll-name=viogpu_wgl -Dgallium-d3d10-dll-name=viogpu_d3d10 -Db_vscrt=mt, build options explained:
  • --prefix=$env:MESA_PREFIX set installation path to dir created in step 1
  • -Dgallium-drivers=virgl build only virgl driver
  • -Dgallium-d3d10umd=true build DirectX 10 user-mode driver (opengl one is build by default)
  • -Dgallium-d3d10-dll-name=viogpu_d3d10 name of generated d3d10 dll to viogpu_d3d10.dll
  • -Dgallium-wgl-dll-name=viogpu_wgl name of generated wgl dll to viogpu_wgl.dll
  • -Db_vscrt=mt use static c runtime (see this comment)
  1. Build and install (to mesa prefix): ninja install

Part 3: Build driver

Now that mesa is build and installed into %MESA_PREFIX% viogpu3d will be built (in case %MESA_PREFIX is not set viogpu3d inf generation is skipped)

  1. Acquire source code git clone --branch viogpu_win https://github.com/max8rr8/kvm-guest-drivers-windows and cd into it cd kvm-guest-drivers-windows
  2. Go to viogpu cd viogpu
  3. (optional, but very useful) setup test code signning from visual studio
  4. Call build .\build_AllNoSdv.bat

Part 4: Installation

Now copy kvm-guest-drivers-windows\viogpu\viogpu3d\objfre_win10_amd64\amd64\viogpu3d to target VM and install it.

EDIT: Added gallium-windows-dll-name to mesa parameters.
EDIT2: More changes related to dll naming in mesa parameters

@max8rr8
Copy link
Author

max8rr8 commented Jul 21, 2023

Changes:

  • Fixed crash caused by use of PAGED_CODE at DISPATCH_LEVEL in flip timer. Instead of using ExTimer, separate flipping thread is used.
  • Building system now checks for viogpu related dll's specifically when deciding whether to compile viogpu3d
  • Renamed DLL's and updated build instructions due to changes requested by mesa

@max8rr8
Copy link
Author

max8rr8 commented Jul 24, 2023

Changes:

  • Re-implemented RotateResourceIdentites. This fixed biggest issue of frames lagging behind, now using driver becomes pretty smooth experience.
  • Implemented clear_render_target similarly to i915
  • Implemented support for staging resources and kernel-mode Present BLT which fixes window display in applications using DXGI_SWAP_EFFECT_DISCARD and DXGI_SWAP_EFFECT_DISCARD.

Pre-built driver provided in description was updated with these changes.

@mincore
Copy link

mincore commented Aug 1, 2023

I've tested the driver on win10, and got a black screen. But if change the guest os to ubuntu(same qemu command line), glxgears performs well. (glxinfo shows that the renderer is NVIDIA gpu)

env

guest os:
win 10 enterprise ltsc + viogpu3d

host os:
fedora38, qemu 8.0.3 with virglrenderer

qemu command line

-display egl-headless,rendernode=/dev/dri/card1 -device virtio-vga-gl -trace enable="virtio_gpu*" -D qemu.log

virglrenderer.log

gl_version 46 - core profile enabled
GLSL feature level 460
vrend_check_no_error: context error reported 3 "" Unknown 1282
context 3 failed to dispatch DRAW_VBO: 22
vrend_decode_ctx_submit_cmd: context error reported 3 "" Illegal command buffer 786440
GLSL feature level 460
vrend_check_no_error: context error reported 5 "" Unknown 1282
context 5 failed to dispatch DRAW_VBO: 22
vrend_decode_ctx_submit_cmd: context error reported 5 "" Illegal command buffer 786440
GLSL feature level 460
GLSL feature level 460
GLSL feature level 460
GLSL feature level 460
GLSL feature level 460
vrend_check_no_error: context error reported 9 "" Unknown 1282
context 9 failed to dispatch DRAW_VBO: 22

qemu_strace.log

virtio_gpu_features virgl 0
virtio_gpu_features virgl 0
virtio_gpu_cmd_get_edid scanout 0
virtio_gpu_cmd_get_display_info
virtio_gpu_cmd_ctx_create ctx 0x1, name
virtio_gpu_cmd_ctx_create ctx 0x2, name
virtio_gpu_cmd_res_create_3d res 0x1, fmt 0x1, w 1280, h 1024, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x2, res 0x1
virtio_gpu_cmd_res_back_attach res 0x1
virtio_gpu_cmd_res_create_3d res 0x2, fmt 0x1, w 1280, h 1024, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x2, res 0x2
virtio_gpu_cmd_res_create_2d res 0x3, fmt 0x2, w 1280, h 1024
virtio_gpu_cmd_res_back_attach res 0x3
virtio_gpu_cmd_set_scanout id 0, res 0x1, w 1280, h 1024, x 0, y 0
virtio_gpu_cmd_res_flush res 0x1, w 1280, h 1024, x 0, y 0
virtio_gpu_cmd_set_scanout id 0, res 0x1, w 1280, h 1024, x 0, y 0
virtio_gpu_cmd_res_flush res 0x1, w 1280, h 1024, x 0, y 0
virtio_gpu_cmd_ctx_create ctx 0x3, name
virtio_gpu_cmd_ctx_destroy ctx 0x3
virtio_gpu_cmd_ctx_create ctx 0x3, name
virtio_gpu_cmd_res_create_3d res 0x4, fmt 0xb1, w 48, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x4
virtio_gpu_cmd_res_create_3d res 0x5, fmt 0xb1, w 4000, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x5
virtio_gpu_cmd_res_create_3d res 0x6, fmt 0xb1, w 16, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x6
virtio_gpu_cmd_res_create_3d res 0x7, fmt 0xb1, w 48, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x7
virtio_gpu_cmd_res_create_3d res 0x8, fmt 0xb1, w 240012, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x8
virtio_gpu_cmd_res_create_3d res 0x9, fmt 0xb1, w 102400, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x9
virtio_gpu_cmd_res_create_3d res 0xa, fmt 0xb1, w 144, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0xa
virtio_gpu_cmd_res_create_3d res 0xb, fmt 0xb1, w 160000, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0xb
virtio_gpu_cmd_res_create_3d res 0xc, fmt 0xb1, w 16000, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0xc
virtio_gpu_cmd_res_create_3d res 0xd, fmt 0xb1, w 240000, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0xd
virtio_gpu_cmd_res_create_3d res 0xe, fmt 0xb1, w 192, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0xe
virtio_gpu_cmd_res_create_3d res 0xf, fmt 0xb1, w 16, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0xf
virtio_gpu_cmd_res_create_3d res 0x10, fmt 0xb1, w 272, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x10
virtio_gpu_cmd_res_create_3d res 0x11, fmt 0xb1, w 240, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x11
virtio_gpu_cmd_res_create_3d res 0x12, fmt 0xb1, w 272, h 1, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x12
virtio_gpu_cmd_res_create_3d res 0x13, fmt 0x1, w 50, h 50, d 1
virtio_gpu_cmd_ctx_res_attach ctx 0x3, res 0x13
virtio_gpu_cmd_ctx_create ctx 0x4, name
virtio_gpu_cmd_res_back_attach res 0x13
virtio_gpu_cmd_res_back_attach res 0x4
virtio_gpu_cmd_res_back_attach res 0x5
virtio_gpu_cmd_res_back_attach res 0x6
virtio_gpu_cmd_res_back_attach res 0xa
virtio_gpu_cmd_res_back_attach res 0xe
virtio_gpu_cmd_res_back_attach res 0xf
virtio_gpu_cmd_res_back_attach res 0x10
virtio_gpu_cmd_res_back_attach res 0x11
virtio_gpu_cmd_res_back_attach res 0x12
virtio_gpu_cmd_ctx_submit ctx 0x3, size 74220

@max8rr8
Copy link
Author

max8rr8 commented Aug 2, 2023

I've tested the driver on win10, and got a black screen. But if change the guest os to ubuntu(same qemu command line), glxgears performs well. (glxinfo shows that the renderer is NVIDIA gpu)

It seems to be an bug in virglrenderer on nvidia related to GL_PRIMITIVE_RESTART_NV. You can try to apply this diff to fix it, but it's a bit hacky solution.

@CE1CECL
Copy link

CE1CECL commented Aug 4, 2023

What versions of windows does it go down to?
In my case, I am experimenting with https://github.com/CE1CECL/qemu-vmvga and it works with vista (no 3d yet), trying to get 3D to work right now.

@foxlet
Copy link

foxlet commented Aug 6, 2023

I tried the pre-built driver on Windows 10 22H2 and it only showed a black screen before hard-locking and resetting. Also tested the same config with Ubuntu which had working virgl.

Using -display gtk,gl=on -device virtio-vga-gl on an AMD R9 6900HS + RX 6700S host.
virglrenderer was compiled from git (considering that it's already been merged in).
qemu is version 7.2.4

qemu.log

@Conan-Kudo
Copy link

Can we validate that this works with Windows 7 too? It's a fairly common virtualization guest for playing older games, and it's still WDDM class.

@max8rr8
Copy link
Author

max8rr8 commented Aug 13, 2023

What versions of windows does it go down to?

Current driver theoretically supports windows 8.1, but it is built for and tested only on Windows 10 22H2. But it is important to note that support for blob resources which are required to improve performance and support vulkan will require using WDDM 2 which lift minimum windows version to 10.

Can we validate that this works with Windows 7 too?

I doubt it will work with current code, it might be possible to adapt it for windows 7, but i do not have interest in doing that (though wouldn't mind if someone else will adapt code). Plus at some point driver will have to use WDDM 2 which will splitting codebases or more likely require either dropping support for older OS.

I tried the pre-built driver on Windows 10 22H2 and it only showed a black screen before hard-locking and resetting. Also tested the same config with Ubuntu which had working virgl.

Nothing seems wrong in attached qemu.log. I don't see any lines like GLSL feature level 460 that should be printed to stdout, can you also attach stdout of qemu.
Although it could be issue in kernel driver then the only way to know what's wrong is to attach kernel debugger to vm.

@Torinde
Copy link

Torinde commented Aug 14, 2023

That's great @max8rr8!
It seems your focus is on modern software (Win8.1/10) and you plan going upwards in support (D3D12/Vulkan/venus). In that regards - do you plan adding:

  • video encoding/decoding acceleration
  • lvp/llvmpipe/softpipe - fallback when no compatible GPU is present on the host or to run the graphics on a high-corecount CPU instead of its weak integrated GPU? Or if possible to combine the two (augment the performance of a weak GPU with llvmpipe running on the CPU cores)

For the retro direction, above was mentioned Win7 and in addition I think such virgl/venus GPU will be very useful also for:

@RedGreenBlue09
Copy link

RedGreenBlue09 commented Aug 19, 2023

I compiled and tried it, but I got the BSOD DRIVER_CORRUPTED_EXPOOL.
Host: Windows 11 Pro (10.0.22621.2070), NVIDIA driver 536.67
Guest: Windows 10 Enterprise (10.0.16299.15)

First, I compiled QEMU in MSYS2 like this:

cp /c/Program\ Files\ \(x86\)/Windows\ Kits/10/Include/10.0.22621.0/um/WinHv*\
 /ucrt64/x86_64-w64-mingw32/include/

./configure \
--target-list=x86_64-softmmu,i386-softmmu,arm-softmmu \
--cpu=x86_64 \
--enable-lto \
--enable-malloc=jemalloc \
--enable-avx2 \
--enable-dsound \
--enable-hax \
--enable-iconv \
--enable-lzo \
--enable-opengl \
--enable-png \
--enable-sdl \
--enable-sdl-image \
--enable-spice \
--enable-spice-protocol \
--enable-tcg \
--enable-whpx \
--enable-virglrenderer \
--disable-docs 

I compiled virglrenderer (upstream) like your comment above, then replaced the dll in QEMU dir with the newly built one.

Then I compiled mesa with VS (MinGW GCC prints a bunch of errors related to the WDK headers, so I gave up):

meson setup build/ -Dgallium-drivers=virgl -Dgallium-d3d10umd=true -Dgallium-wgl-dll-name=viogpu_wgl -Dgallium-d3d10-dll-name=viogpu_d3d10 --backend=vs

It outputs to my root folder (C:\) so I copied bin to a directory named mesa_prefix.
I opened cmd, set MESA_PREFIX to the mesa_prefix directory then compiled viogpu.sln with msbuild (Configuration="Win10 Release", Platform=x64).

I signed the output in viogpu\viogpu3d\objfre_win10_amd64\amd64\viogpu3d:

for /R . %a in (*.exe, *.sys, *.dll) do signtool sign /f "Surface.pfx" /t "http://timestamp.sectigo.com" /fd certHash "%a"

Then I booted up the VM, enabled test signing and then install the driver. The screen instantly went black and Windows crashes.

WinDBG, qemu logs and VM command line attached. I tried using serial kernel debug but it just hangs forever at boot so I analyzed the dump file instead.

windbg.txt
qemu.log
cmdline.txt

@RedGreenBlue09
Copy link

Additionally, I tried booting Linux Mint 20.2 live cd but QEMU itself crashed on some heap corruption issue. I replaced the original MSYS2 virglrenderer dll which fixed it. The Windows 10 guest BSOD remains.

@max8rr8
Copy link
Author

max8rr8 commented Aug 20, 2023

Hi @RedGreenBlue09, it seems that kernel crash happened during driver unloading (VioGpu3DRemoveDevice in backtrace) and while it is a problem, it is not the root cause for driver not working on your system as for some unknown reason windows triggers driver unload.

You should try to attach windbg to running vm (try to use network debugging instead of serial, it works for me). Additionally when connecting to vm from windbg enable break on connection and run following command

bp  watchdog!WdLogEvent5_WdError "k; g"

This windbg command should add backtrace logging to all errors happening in dxgkrnl. After you ran that command continue kernel execution with windbg command g or a button in gui. Then get full log and send it there, so we can analyze what actually went wrong (why does unload happen).

@RedGreenBlue09
Copy link

Hi @RedGreenBlue09, it seems that kernel crash happened during driver unloading (VioGpu3DRemoveDevice in backtrace) and while it is a problem, it is not the root cause for driver not working on your system as for some unknown reason windows triggers driver unload.

You should try to attach windbg to running vm (try to use network debugging instead of serial, it works for me). Additionally when connecting to vm from windbg enable break on connection and run following command

bp  watchdog!WdLogEvent5_WdError "k; g"

This windbg command should add backtrace logging to all errors happening in dxgkrnl. After you ran that command continue kernel execution with windbg command g or a button in gui. Then get full log and send it there, so we can analyze what actually went wrong (why does unload happen).

I will test again later. Also, if I use VGA + virtio-gpu-gl setup the driver don't load and it have error 49 in device manager. The error message is something like "Windows unloaded the driver because it has reported problems"

@RedGreenBlue09
Copy link

... Network debug also hangs like serial. Interestingly that doesn't happen with ReactOS.

@RedGreenBlue09
Copy link

RedGreenBlue09 commented Aug 20, 2023

I have no luck with live kernel debug. I tried using accel tcg, official qemu, none of these helps. If
I enable boot debug, Windows boot manager itself freezes. Absolutely no idea.

Edit: Even ditching OVMF for SeaBIOS, it still freezes.

@RedGreenBlue09
Copy link

@max8rr8 Okay, I'm sorry for the rant. It doesn't hang forever, just 30 minutes :((. I tried your command and got an error:

0: kd> bp  watchdog!WdLogEvent5_WdError "k; g"
Bp expression 'watchdog!WdLogEvent5_WdError ' could not be resolved, adding deferred bp

@RedGreenBlue09
Copy link

RedGreenBlue09 commented Aug 21, 2023

After reloading symbols, the error is gone. Here is the new log.

windbg2.txt

YanVugenfirer pushed a commit that referenced this pull request Dec 2, 2024
Based on #943

Signed-off-by: Max Ramanouski <[email protected]>
Signed-off-by: Kostiantyn Kostiuk <[email protected]>
YanVugenfirer pushed a commit that referenced this pull request Dec 2, 2024
Based on #943

Signed-off-by: Max Ramanouski <[email protected]>
Signed-off-by: Kostiantyn Kostiuk <[email protected]>
YanVugenfirer pushed a commit that referenced this pull request Dec 2, 2024
Based on #943

Signed-off-by: Max Ramanouski <[email protected]>
Signed-off-by: Kostiantyn Kostiuk <[email protected]>
YanVugenfirer pushed a commit that referenced this pull request Dec 2, 2024
Based on #943

Signed-off-by: Max Ramanouski <[email protected]>
Signed-off-by: Kostiantyn Kostiuk <[email protected]>
YanVugenfirer pushed a commit that referenced this pull request Dec 2, 2024
Based on #943

Signed-off-by: Max Ramanouski <[email protected]>
Signed-off-by: Kostiantyn Kostiuk <[email protected]>
YanVugenfirer pushed a commit that referenced this pull request Dec 2, 2024
This changes will be used in 3D driver.

Based on #943

Signed-off-by: Max Ramanouski <[email protected]>
Signed-off-by: Kostiantyn Kostiuk <[email protected]>
@SimonFair
Copy link

Are there any time frames when this might be a usable published driver?

@YanVugenfirer
Copy link
Collaborator

Are there any time frames when this might be a usable published driver?

Unfortunately driver with mesa might not be the right production solution.
So we don’t have the timeline

@RushingAlien
Copy link

I wonder if windows guets can make use of virtio-native-contex? Cause you know, the contexts are DRM ioctls, meanwhile windows has something different.

@DemiMarie
Copy link

Are there any time frames when this might be a usable published driver?

Unfortunately driver with mesa might not be the right production solution. So we don’t have the timeline

I suspect that Mesa is the only reasonable solution. In particular, it is mandatory for GPU virtualization with untrusted guests on consumer hardware. This is because the shader compiler must run in the guest (not the host), and Mesa is the only open-source shader compiler available.

@AlfCraft07
Copy link

AlfCraft07 commented Dec 21, 2024

Are there any time frames when this might be a usable published driver?

Unfortunately driver with mesa might not be the right production solution. So we don’t have the timeline

I suspect that Mesa is the only reasonable solution. In particular, it is mandatory for GPU virtualization with untrusted guests on consumer hardware. This is because the shader compiler must run in the guest (not the host), and Mesa is the only open-source shader compiler available.

I would like this being ported to Windows 7.

@kostyanf14
Copy link
Member

I suspect that Mesa is the only reasonable solution. In particular, it is mandatory for GPU virtualization with untrusted guests on consumer hardware. This is because the shader compiler must run in the guest (not the host), and Mesa is the only open-source shader compiler available.

There is also Proton (Steam solution to play Windows games on Linux) that works fine and can convert DX12 API to Vulkan and it will be more simple than implementing DX12 UMD on top of MESA.
vkd3d-proton provide shared compiler to.

@kostyanf14
Copy link
Member

I would like this being ported to Windows 7.

First of all, Windows 7 is dead. For the second one, we dropped Windows 7 support from all other drivers. In the end, GPU drivers for Win7 and Win10 are different from an infrastructure point of view.

@kasperk81
Copy link

I suspect that Mesa is the only reasonable solution. In particular, it is mandatory for GPU virtualization with untrusted guests on consumer hardware. This is because the shader compiler must run in the guest (not the host), and Mesa is the only open-source shader compiler available.

There is also Proton (Steam solution to play Windows games on Linux) that works fine and can convert DX12 API to Vulkan and it will be more simple than implementing DX12 UMD on top of MESA. vkd3d-proton provide shared compiler to.

would user be able to play games on Windows 11 arm running under UTM on macOS arm64 (utmapp/UTM#4028)?

HansKristian-Work/vkd3d-proton#1889 suggests it's not possible, so perhapsDX12 UMD on top of MESA may help there?

@kasperk81
Copy link

also https://github.com/microsoft?q=directx not sure if and how useful https://github.com/microsoft/DirectXShaderCompiler could be for this driver

@sharkautarch
Copy link

sharkautarch commented Dec 21, 2024

Ermm, in terms of shader compilers, to clear stuff up, there’s actually multiple separate issues wrt the need to use the mesa driver in the guest:

  • IIRC, Shaders on windows d3d9/d3d11/d3d12 are first compiled to an intermediate representation (IR) and then that IR is fed to the graphics driver. D3d11 uses DXIL IR https://github.com/Microsoft/DirectXShaderCompiler/blob/main/docs/DXIL.rst. Whereas Vulkan only uses SPIRV. Tho supposedly d3d11 may move to SPIRV in the future https://devblogs.microsoft.com/directx/directx-adopting-spir-v/
  • Putting aside the shader IR, viogpu3d still has to "speak" to the windows-guest in terms of d3d api.
  • Afaik, mesa drivers are designed to support being used as a user-mode guest driver on windows, wherein it can communicate in terms of d3d api. It also has other facilities that may become handy in the future for viogpu such as virtio native context & what not.
  • Finally, dxvk requires several Vulkan features be supported for it to work, which I’m not sure how feasible that’d be for viogpu3d: https://github.com/doitsujin/dxvk/wiki/Driver-support
    IG some of those required features could be supported, but I have no clue how viogpu3d would be able to support the required VK_EXT_transform_feedback feature

So yes, it really does seem like using mesa in the guest is the most straightforward path

@kasperk81
Copy link

Whereas Vulkan only uses SPIRV.

dxcompiler.dll seem to support both SPIRV and DXIL with HSLS as a go-between language https://github.com/microsoft/DirectXShaderCompiler/blob/main/docs/SPIR-V.rst. readme says dxcompiler.dll is also part of windows sdk

@kasperk81
Copy link

kasperk81 commented Dec 21, 2024

So yes, it really does seem like using mesa in the guest is the most straightforward path

https://github.com/microsoft/wslg/blob/main/README.md#opengl-accelerated-rendering-in-wslg WSLg also seem to be using mesa to run Linux GUI applications, so that's another +1. mesa is a submodule in that repo. the solution is not lightweight tho

@v-fox
Copy link

v-fox commented Dec 22, 2024

I suspect that Mesa is the only reasonable solution. In particular, it is mandatory for GPU virtualization with untrusted guests on consumer hardware. This is because the shader compiler must run in the guest (not the host), and Mesa is the only open-source shader compiler available.

There is also Proton (Steam solution to play Windows games on Linux) that works fine and can convert DX12 API to Vulkan and it will be more simple than implementing DX12 UMD on top of MESA. vkd3d-proton provide shared compiler to.

Proton is just wine fork with a bunch of haphazardly implemented hacks, most of which will never go upstream. And wine has a bunch of fundamental problems (ironically, pretty much all of them non-graphics related) that obliterate performance or even ability for an app to work at all. Such as atrocious sync implementation, replacement for which is being continuously sandbagged by kernel's bureaucrats. It may take decades or never to fix entire list of fundamental compatibility issues. And if you need a working application now, running mostly Windows native stack is the only option. But running that crap on bare hardware is lunacy.

So, no, running broken wine and waiting for years for maintainers to wake up and approve patches, that may or may not appear, is neither simple or fast. It really either this or for everyone to build their hardware with dual GPUs just to passthrough the main one to Windows for native driver to take over.

@DemiMarie
Copy link

Proton is just wine fork with a bunch of haphazardly implemented hacks, most of which will never go upstream.

Which hacks are you referring to?

@kasperk81
Copy link

another practical example: godotengine/godot#70315 they used a variant of https://gitlab.freedesktop.org/mesa/mesa/-/tree/main/src/microsoft/spirv_to_dxil after trying out the alternative HansKristian-Work/dxil-spirv#103 (comment). we can learn from their experiences and stay put with mesa choice to complete this pr

@Torinde
Copy link

Torinde commented Dec 22, 2024

Another example of driver using Mesa is SoftGPU:

I would like this being ported to Windows 7

SoftGPU supports Win9x/NT4 and maybe can be extended to XP (which may cover Vista/7 as well? But it's not a WDDM driver)

@CryptoManiac
Copy link

Another example of driver using Mesa is SoftGPU:

I would like this being ported to Windows 7

SoftGPU supports Win9x/NT4 and maybe can be extended to XP (which may cover Vista/7 as well? But it's not a WDDM driver)

Vista might be able to work with the older driver model but it won't provide the composition features in such case.

@binarymaster
Copy link

Latest drivers from VirtualBox Guest Additions currently work fine (including DWM acceleration) in both Vista and 7, so you may want to look into the source of that.

@yuq
Copy link

yuq commented Dec 23, 2024

I'm a mesa developer watching this PR for a while. My opinion is you need a DX to VK translation layer anyway, either in host or guest because host (Linux) only provide Vulkan API.

Do this in guest would be much simpler. You just need to implement a virtio-gpu based vulkan driver (either venus or native context), then use DXVK and VKD3D for DX API, Zink for OpenGL API, RustiCL for OpenCL API (these translation layers are build on top of Vulkan). All these components don't need to build from scratch, just need to adopt to the windows environment. This way you get all GPU APIs at once with reasonable effort. Otherwise you have to implement every API UMD from scratch for guest which is a big dev and maintain effort.

Windows WSLg uses a native context way (WDDM level) to run DX12 UMD as guest UMD on top of host windows KMD first, then uses translation layer in guest (DZN for Vulkan and D3D12 for OpenGL) to translate Vulkan/OpenGL to DX12.

@Usulyre
Copy link

Usulyre commented Dec 23, 2024

I'm a mesa developer watching this PR for a while. My opinion is you need a DX to VK translation layer anyway, either in host or guest because host (Linux) only provide Vulkan API.

Do this in guest would be much simpler. You just need to implement a virtio-gpu based vulkan driver (either venus or native context), then use DXVK and VKD3D for DX API, Zink for OpenGL API, RustiCL for OpenCL API. All these components don't need to build from scratch, just need to adopt to the windows environment. This way you get all GPU APIs at once with reasonable effort. Otherwise you have to implement every API UMD from scratch for guest which is a big dev and maintain effort.

Windows WSLg uses a native context way (WDDM level) to run DX12 UMD as guest UMD on top of host windows KMD first, then uses translation layer in guest (DZN for Vulkan and D3D12 for OpenGL).

Hi,

Would you be able to implement something like that?

@yuq
Copy link

yuq commented Dec 23, 2024

Would you be able to implement something like that?

I think so.

@kasperk81
Copy link

bottom line: this is not the first time this problem is being solved. this pr is using mesa in its current state, is it the best approach? maybe not, but if multiple similar projects are using mesa, it's the one which works. pragmatic way is to move this pr forward and merged. first version doesn't have to be perfect, it (like any software) can improve later on

@v-fox
Copy link

v-fox commented Dec 24, 2024

Proton is just wine fork with a bunch of haphazardly implemented hacks, most of which will never go upstream.

Which hacks are you referring to?

Like, all of them. If it wasn't mostly made up from un-upstream'able hacks then it wouldn't be a fork of an active massive project.
Either way, userspace translation framework, that is full of stubs and perpetual half-baked temporary solutions, is mostly irrelevant for full-blown kernel-level emulator running an entire OS.

But it does use vkd3d/dxvk, just like vanilla wine, which would have to be used with Mesa anyway. Pretty much the only part that is relevant and relatively polished out.

So ironic that with wine the graphics are fine yet everything else may fatally glitch out for your app but with qemu/kvm it's the exact opposite: you run "stable" native stack but graphics are even worse than running CPU-only (Mesa's llvmpipe) dxvk under vanilla wine. And, yes, I've tested that.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.