[PATCH] d3d12: Check for VK_KHR_external_fence_capabilities before using VkPhysicalDeviceIDProperties.

[Zebediah Figura (she/her)]

On 2/25/21 5:33 PM, Philip Rebohle wrote:
> 
> 
> Am 25.02.21 um 23:17 schrieb Zebediah Figura (she/her):
>> On 2/25/21 3:59 PM, Philip Rebohle wrote:
>>> Am 25.02.21 um 17:38 schrieb Zebediah Figura (she/her):
>>>> On 2/25/21 10:18 AM, Joshua Ashton wrote:
>>>>> On 2/25/21 4:01 PM, Zebediah Figura (she/her) wrote:
>>>>>> On 2/25/21 7:00 AM, Joshua Ashton wrote:
>>>>>>> There is no reason to. This is core in Vulkan 1.1.
>>>>>>>
>>>>>>> You should just use a newer API version.
>>>>>>> There is no relevant driver that doesn't support this.
>>>>>>
>>>>>> Mesa's v3dv is an example of a driver that does not support Vulkan
>>>>>> 1.1.
>>>>>
>>>>> That is not relevant hardware:
>>>>> It is not capable of supporting D3D12.
>>>>
>>>> If it's capable of supporting vkd3d, it's capable of supporting
>>>> d3d12 ;-)
>>>>
>>>> (Granted, the former may actually not be the case, for lack of
>>>> KHR_shader_draw_parameters. I don't know if that's possible to
>>>> implement.)
>>>>
>>>> More to the point, though, there are more salient reasons to support
>>>> Vulkan 1.0: drivers not yet written (e.g. there is not yet a free
>>>> NVidia
>>>> or Mali driver, nor a conformant software driver), drivers whose latest
>>>> versions are not accessible (e.g. the last version of MoltenVK that can
>>>> be built for MacOS Mojave only supports Vulkan 1.0), hardware which may
>>>> not be capable of supporting Vulkan 1.1 (if any exists), and less
>>>> painful bisection of Vulkan drivers down to old versions. Compared with
>>>> the cost of adding a couple lines, and that only because I happened to
>>>> run the validation layers, I'm not particularly inclined to believe
>>>> dropping support for Vulkan 1.0 is a worthwhile trade.
>>>
>>> If anyone ever asked me why vkd3d-proton exists as a fork, this would be
>>> a perfect example.
>>>
>>> The point Josh is trying to make here is that any new-ish driver project
>>> will of course be aiming for Vulkan 1.0 conformance with a limited
>>> feature set before adding all the optional stuff that was added later,
>>> simply because it's easier and faster to bring up. v3dv reached this
>>> point three months ago. It already supports a bunch of features that
>>> were promoted to Vulkan 1.1 such as the external memory stuff, but is
>>> still missing some others. It'll get there eventually.
>>>
>>> Thing is, the feature set that these young implementations provide is
>>> barely enough to support a D3D9 feature set, and that is if you emulate
>>> basic functionality like block-compressed textures because the Vulkan
>>> implementation doesn't even support those (v3dv can't due to hardware
>>> limitations, Swiftshader could but doesn't, etc.). If we're *very*
>>> generous, we're looking at a subset of D3D11 FL11_0 functionality with
>>> loads of inefficient hacks to make things like D3D11_DSV_READ_ONLY_DEPTH
>>> work, recompiling compute shaders with the exact view formats in mind
>>> because shaderStorageWriteWithoutFormat isn't supported, and the list
>>> could go on forever, there's so much that these drivers just do not
>>> support yet in any reasonable way. And god beware if anyone actually
>>> tries to use geometry shaders.
>>>
>>> For D3D12 you can't even emulate all the missing stuff. The binding
>>> model just doesn't give you enough information at the time command lists
>>> are being recorded to work around any of this. That is, if you actually
>>> implement it correctly and support update-after-bind semantics and
>>> descriptor indexing, which are not optional in D3D12.
>>>
>>> Has anyone other than Józef actually done any sort of research into how
>>> the API works and evaluated what kind of baseline feature set is
>>> required to drive a translation layer? The "Vkd3d known issues" page on
>>> winehq itself provides more than enough insight on why a baseline Vulkan
>>> 1.0 implementation simply cannot run this API even remotely reasonably.
>>> There are solutions to most of them nowadays, but a lot of the upstream
>>> vkd3d code has been written before those were available.
>>>
>>> I can sort of understand MoltenVK on Mojave as a somewhat valid reason
>>> to keep an old code path around, provided that there actually some D3D12
>>> content working on that combination in the first place, but instead of
>>> making D3D work properly on platforms which *do* provide a good enough
>>> Vulkan environment, now we're discussing a feature that was promoted to
>>> the core Vulkan spec *three years ago* to cater to platforms that can't.
>>>
>>> And then people are wondering why we don't try to work with upstream
>>> anymore.
>>
>> Cards on the table, then: I really don't know what I'm talking about.
>>
>> I look at this code and say, "I don't know the consequences of making
>> that extension, or Vulkan 1.1, mandatory, but checking for the
>> extension's presence looks like a right solution and is easy."
>> vkd3d-proton exists as a fork because, for better or for worse, there
>> are those of us who are not willing to settle for that.
>>
>> My mistake here is trying to defend my choice on what I think are the
>> appropriate devil's-advocate grounds. Trouble is: I really don't know
>> what I'm talking about. I certainly haven't done that research you're
>> asking about. But as far as I see, vkd3d-proton also exists because
>> both sides got fed up with the other not responding to their concerns
>> and decided to stop interacting. Given the choice I'll make the
>> mistake of trying to fix that, at the risk of making some pretty bad
>> arguments. [Because someone has to, and no one else will.]
>>
>> The concern of changes like requiring 1.1 (and, more importantly,
>> dropping all the 1.0-specific code and feature checks) making it
>> harder to bisect was the only one that was actually raised to me as a
>> salient point; the others were my own points and should probably be
>> ignored. If nothing else, maybe we can argue about that part?
> 
> Harder to bisect what though? I guess for MVK you do have a point there
> since Vulkan 1.1 support is still somewhat recent there, but the value
> of this will deminish as time passes (also, there's a boatload of other
> issues with trying to run D3D12 on Metal, but that's being discussed on
> the MVK issue tracker so I'm not going to go into detail on that front;
> safe to say that Metal itself is unfortunately not a very good API for
> layering purposes).
> 
> For Mesa it's already not really useful to bisect very old issues in my
> experience; ancient Mesa versions are hard to build due to the LLVM
> build-time requirement as well and the drivers have seen significant
> changes since then as well (new compiler, tons of issues being caused by
> LLVM itself, etc). Unless a regression has happened within the last two
> or so major releases, it's probably more productive to just post a
> RenderDoc capture or small test case to reproduce an issue with and move
> on.

I don't think I can offer any meaningful comment here. I do get the
impression that most of the vkd3d contributors (current and past?) have
preferred debugging strategies that don't rely on
renderdoc/apitrace/etc., but I really can't evaluate the relative
merits. I don't necessarily think it's right to say "it's not worth
supporting bisects even if you prefer them", though.

> 
> Now arguably your patch doesn't really do anything that would mandate
> switching to 1.1; what really blew my fuse here is that in order to work
> around a validation layer bug you were asked to check for a ubiquitous
> extension from over four years ago instead of an equally ubiquitous
> extension from over three and a half years ago which also happens to be
> the one that validation is actually complaining about.
> 
> One might ask if it's worth working around a validation bug that's
> likely to get fixed properly at some point in the first place, but in
> all fairness, there's arguments to be made for it especially if it has
> consequences other than some annoying log spew.
> 
I'm not sure it's reasonable to object to that. I think the general
policy of "don't drop support for anything without a strong reason", for
some definition of "strong", is worth reëvaluating, in this case if
nothing else. But under the bounds of that policy, I think Henri's
review is right (and the validation layers *are* broken—and I don't
think we really care about fixing the warnings per se as much as fixing
the correctness issues they find. Which, for an API like Vulkan... I'd
be perfectly willing to believe that the driver takes a missing
extension like this as license to randomly crash 5 minutes later into
the program.)

> FWIW, 1.1 turned out to be necessary to support Shader Model 6 (which is
> actively being used in newer D3D12 content, and was actually the last
> thing we *tried* to upstream with Hans-Kristian's dxil-spirv library, go
> figure), subgroup operations in particular, since there's no equivalent
> extension you can just enable on a 1.0 device, which left all the 1.0
> fallback code there to rot and be a maintenance burden. Vulkan is still
> an evolving ecosystem and D3D12 is a moving target with new things being
> tacked on in every major Windows release, compromises for
> driver/hardware support will have to be made at some point as newer
> features get adopted.

In case it was unclear, I think (or at least assume) the intent for
vkd3d is to conditionally request newer versions depending on the
requested feature level, and then leave 1.0-compatible paths (or
whatever) for any lower feature level, whereas the SM6 translation would
only be allowed if the relevant feature level is enabled. That's roughly
the way wined3d's GL backend works, minus the part where extensions and
versions have to be explicitly enabled.

I suspect that vkd3d-proton (and maybe even DXVK) operates in general on
lines of thought like "we are sure that the AMD and NVidia GPUs/drivers
all support this (or soon will), and we don't care about any others,
therefore we will require it even if the game doesn't ask for it",
possibly paired with "we are sure that no game will not ask for it".
Possibly this is paired with a policy to implement fallback paths if
those assertions turn out to be wrong, or possibly just "it's not worth
supporting this for your GPU, get a better one" (I certainly hope the
latter isn't the case, but from interacting with the DXVK community I
get a strong impression that it is. But it's not obvious to me that the
former is actually a bad idea. Or maybe it really depends on Vulkan
being so new.)

I similarly suspect that wined3d is operating on years of users
complaining when things get broken, and has therefore adopted a
defensive position of "just don't drop support for things; instead
restructure to allow newer features to give you fast paths". I mean, I
know this is true for the broader wine project—it probably strikes those
in the DXVK community as bizarre how much we care about regressions. But
I can see disadvantages: this does create a maintenance burden, one way
or another, and it of course risks being pointlessly defensive.

It would be really nice if we can find a middle ground. I'm not trying
to evangelize compromise here qua compromise—there are a couple
high-level things about wine's development process I like a *lot* better
than DXVK/vkd3d-proton, and in some cases even think are objectively
better. Like, even if I didn't have a severe emotional investment in the
feud [not that I'm sure which way that biases me—believe me when I say
that both sides carry significant blame], and even if I could contribute
to DXVK without possible licensing problems, I know which project I
would prefer contributing to.

But in a case like this—and I mean the general disagreement of whether
to aggressively keep legacy code around or aggressively require the
latest drivers/GPU, I can see the arguments for both sides, or some
middle ground. And I find it tragic that the expertise and experience
that all of the people here who aren't me have is just not respected,
and that everyone's reaction to disagreement is to immediately
disengage. I find it especially tragic because I'm not sure that under
Henri couldn't be convinced to just require Vulkan 1.1—but nobody's
really tried. If I'm trying to do anything here it's to get these issues
out in the open and discussed.

> 
> - Philip
> 
>>
>>>
>>> - Philip
>>>
>>>>
>>>>>
>>>>> - Joshie 🐸✨
>>>>>
>>>>>>
>>>>>>>
>>>>>>> - Joshie 🐸✨
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>>
>>>
>>
> 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature
Type: application/pgp-signature
Size: 495 bytes
Desc: OpenPGP digital signature
URL: <http://www.winehq.org/pipermail/wine-devel/attachments/20210226/d4e022db/attachment-0001.sig>