Mesa's Venus Vulkan Driver Updated To Allow QEMU Support

[Quackdoc]

Thanks to everyone that replied.

According to sebastianlacuesta 's link, they're currently using a d3d10umd port to virgl, but I'm wondering if venus can be ported and then zink + DXVK + VKD3D used instead? IIUC, DXVK and VKD3D implement the Direct3D API rather than the DDI, which means they cannot be installed as drivers in Windows, but can still be used by just replacing Microsoft's d3d DLLs with their own?

Also, I vaguely remember reading in another Phoronix thread that zink + venus should be faster than virgl, because of something about virgl being an emulated GPU using OpenGL as a back end while venus just passing through Vulkan calls to the host. If so, wouldn't that mean that DXVK/VKD3D + venus would be faster?

DXVK/VKD3D aren't umd drivers. you would still need to port D3D10UMD to zink to have a usable desktop using vulkan. You could load individual games using DXVK and VKD3D but for general use no. you need UMD drivers.

but yes, zink + DXVK/VKD3D would be, and is much faster then wined3d, even on linux.

[rabcor]

If the VirtIO GPU is ported to Windows, could this be used with VKD3D and/or DXVK to get faster Windows VMs in QEMU?

Yes but the gtk display you will normally be using in qemu would still be limited to 30hz

Also ur not really gonna be gaming on virtio unless it's a game from the 90s, virtio gl performance is embarrasingly bad, like actual software renderers (llvmpipe/lavapipe) are faster lol.

[pong]

It is being ported to Windows, and theoretically yes, but at first the speed up will be minor. The current state is enough for slightly more responsive 2D for the Windows Display Manager equivalent.

It will be a bit before accelerated 3D works or well. And I don't remember if they are going to be able to support Direct X acceleration, or just OpenGL, and Vulkan if we are lucky.

That'd be awfully nice to be able to run even simple 2D/3D/GPU dependent stuff in VMs for those legacy productivity / utility tools that one needs sporadically and then have to go through the nonsense of getting multiple GPUs to pass one through VFIO just to run some application while one has a perfectly good CPU and GPU on the host.

Well, that and universal useful working SR-IOV on all mid-range "gamer" GPUs and up to handle acceleration even more natively.
What's the point of having these 16 core 128 GBy etc. and $700 GPUs systems if one practically even run a lousy VM or two that actually works
for 2D/3D/compute on the GPU to the level of a 5 year old APU laptop or something.

[dragorth]

That'd be awfully nice to be able to run even simple 2D/3D/GPU dependent stuff in VMs for those legacy productivity / utility tools that one needs sporadically and then have to go through the nonsense of getting multiple GPUs to pass one through VFIO just to run some application while one has a perfectly good CPU and GPU on the host.

Well, that and universal useful working SR-IOV on all mid-range "gamer" GPUs and up to handle acceleration even more natively.
What's the point of having these 16 core 128 GBy etc. and $700 GPUs systems if one practically even run a lousy VM or two that actually works
for 2D/3D/compute on the GPU to the level of a 5 year old APU laptop or something.

The Intel iGPUs can do the basics like this, and the i915 driver is increasingly being used to do SR-IOV, not to mention the Intel ARC 7xx something. Level1Tech has some videos and forum posts on how to get it working. It isn't ready for mainstream yet, but it is getting there.

[qarium]

That'd be awfully nice to be able to run even simple 2D/3D/GPU dependent stuff in VMs for those legacy productivity / utility tools that one needs sporadically and then have to go through the nonsense of getting multiple GPUs to pass one through VFIO just to run some application while one has a perfectly good CPU and GPU on the host.
Well, that and universal useful working SR-IOV on all mid-range "gamer" GPUs and up to handle acceleration even more natively.
What's the point of having these 16 core 128 GBy etc. and $700 GPUs systems if one practically even run a lousy VM or two that actually works
for 2D/3D/compute on the GPU to the level of a 5 year old APU laptop or something.

in the past amd open-source driver people told us that SR-IOV does not do what people think.
SR-IOV plain and simple is not made to run games in a VM..........
who ever started this meme that SR-IOV is made to make VM magically run like playing games in windows 11 running in a VM on linux...

well thats the reason why cards like the AMD PRO w7900 do not even have SR-IOV,,,

as bridgman told us in the past SR-IOV is a server hardware feature used for compute tasks

with nearly zero use for any workstation or desktop means you can not play games in a VM ...

[Quackdoc]

who ever started this meme that SR-IOV is made to make VM magically run like playing games in windows 11 running in a VM on linux....

you... you do know SR-IOV has been tested to do exactly this and it works great right? no, ofc not, because that would involve actual research.

for anyone who isn't missing half of their brain, Cloud gaming is a great usecase for SR-IOV and has been tested by folk at level1tech, and plenty of others, it's not as good as Nvidia's solution granted, but it's still pretty nice and for sure. If you have a supported intel igpu SR-IOV is still a nice way to play some older games that break in wine for one reason or another.

[qarium]

you... you do know SR-IOV has been tested to do exactly this and it works great right? no, ofc not, because that would involve actual research.

Intel maybe implemented it that you can run games in a VM but technically this has nothing to do with SR-IOV

for anyone who isn't missing half of their brain, Cloud gaming is a great usecase for SR-IOV and has been tested by folk at level1tech, and plenty of others, it's not as good as Nvidia's solution granted, but it's still pretty nice and for sure. If you have a supported intel igpu SR-IOV is still a nice way to play some older games that break in wine for one reason or another.

you have a deep misunderstanding here about cloud gaming service providers because these cloud gaming service providers have no display output... thats the reason why SR-IOV has no meaning in running a game in a VM on your local desktop/workstation.

well i do not say that intel did not invest a lot of money and resources to bring up their VM drivers to make something like this possible.

but you have to unterstand that SR-IOV is not designed to share GPU resources between host and guest in a VM environment with display output.

what intel has here is most likely a software solution (the consumer does not care if software or hardware)

and even if intel use some functionality from the SR-IOV hardware it does not mean that the software solution would not do it as well the the same result.

as bridgman told us multible times AMD will not bring SR-IOV hardware to any notebook/desktop/workstation cards or hardware...

they have the opinion that a pure software solution is the right way to go ...

[Quackdoc]

Intel maybe implemented it that you can run games in a VM but technically this has nothing to do with SR-IOV



you have a deep misunderstanding here about cloud gaming service providers because these cloud gaming service providers have no display output... thats the reason why SR-IOV has no meaning in running a game in a VM on your local desktop/workstation.

well i do not say that intel did not invest a lot of money and resources to bring up their VM drivers to make something like this possible.

but you have to unterstand that SR-IOV is not designed to share GPU resources between host and guest in a VM environment with display output.

what intel has here is most likely a software solution (the consumer does not care if software or hardware)

and even if intel use some functionality from the SR-IOV hardware it does not mean that the software solution would not do it as well the the same result.

as bridgman told us multible times AMD will not bring SR-IOV hardware to any notebook/desktop/workstation cards or hardware...

they have the opinion that a pure software solution is the right way to go ...

Gaming VMs often do not need a display out, you can use DMA to copy the frames, or do what I do, which is use moonlight. A pure software implementation will always be inferior to a decently implemented SR-IOV

[qarium]

Gaming VMs often do not need a display out, you can use DMA to copy the frames, or do what I do, which is use moonlight. A pure software implementation will always be inferior to a decently implemented SR-IOV

i think your statement is clearly wrong here because from the rational standpoint its all about the transistor count SR-IOV need a lot of transistors to make this pay out you really need to run like 32 VM instances

in laptop/desktop/workstation you many time only run 1 instance means 1 host and 1 guest... the performance you lose in such a scenario is less than what you need if you invest the tranistor count on stuff like more shader units.

this means the transistor count is better used in more shaders and the pure software solution has less overhead that the performance win with more shader units.

the tranistor count used for SR-IOV only pay out if you use 32 instances (typical scenario of a cloud VM gpu)

this means from a rational point of view a rational GPU maker will never implement SR-IOV for a senario of 1 host and 1 guest to play a game in a VM running windows on top of KVM/QEMU...

but we all know Intel does not act rational in the GPU market they perform subsidies on their GPUs sell the chips with a loss just to get some marketshare they implement high-end-cloud-gpu features ln the consumer gpus who does not help them at all with consumer related szenarios and all this did make intel already lose 6.4 billion dollars just to enter the GPU market.

keep in mind i have a Intel ARC A380 like you have and this card can not even compete with my 6-7 years old EOL Vega64 in performance and driver support even to the fact that the driver of the vega64 is EOL ... and intel actively supports its card.

intel acts in a way that they will not make money on GPUs the next 5+ years any tranistors they put in like for features like SR-IOV costs money and the customers have to pay for it (or they lose even more money on subsidies...) and the people more or less buy hardware who can not compete in performance.

[Quackdoc]

i think your statement is clearly wrong here because from the rational standpoint its all about the transistor count SR-IOV need a lot of transistors to make this pay out you really need to run like 32 VM instances

in laptop/desktop/workstation you many time only run 1 instance means 1 host and 1 guest... the performance you lose in such a scenario is less than what you need if you invest the tranistor count on stuff like more shader units.

this means the transistor count is better used in more shaders and the pure software solution has less overhead that the performance win with more shader units.

the tranistor count used for SR-IOV only pay out if you use 32 instances (typical scenario of a cloud VM gpu)

this means from a rational point of view a rational GPU maker will never implement SR-IOV for a senario of 1 host and 1 guest to play a game in a VM running windows on top of KVM/QEMU...

but we all know Intel does not act rational in the GPU market they perform subsidies on their GPUs sell the chips with a loss just to get some marketshare they implement high-end-cloud-gpu features ln the consumer gpus who does not help them at all with consumer related szenarios and all this did make intel already lose 6.4 billion dollars just to enter the GPU market.

keep in mind i have a Intel ARC A380 like you have and this card can not even compete with my 6-7 years old EOL Vega64 in performance and driver support even to the fact that the driver of the vega64 is EOL ... and intel actively supports its card.

intel acts in a way that they will not make money on GPUs the next 5+ years any tranistors they put in like for features like SR-IOV costs money and the customers have to pay for it (or they lose even more money on subsidies...) and the people more or less buy hardware who can not compete in performance.

DXVK/VKd3d can only be so good, and they don't implement UMD drivers.