vkd3d seems to be ignoring maxTexelBufferElements in some games, leading to visual corruption

[pzanoni-intel]

Hello

Original bug report: https://gitlab.freedesktop.org/mesa/mesa/-/issues/9963

On Intel graphics with the Mesa drivers we're observing that some games such as Marvel's Spider-Man Remastered and Assassin's Creed: Valhalla seem to be emitting vkGetDescriptorEXT() calls that ignore VUID-VkDescriptorGetInfoEXT-type-09427, which relates to VkPhysicalDeviceLimits::maxTexelBufferElements.

https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkDescriptorGetInfoEXT.html#VUID-VkDescriptorGetInfoEXT-type-09427

Since vkGetDescriptorEXT() doesn't have a way to return errors, I tried to change anv to completely disable VK_EXT_descriptor_buffer. With this, we have the same kind of problem, but now with vkCreateBufferView() calls violating VkBufferViewCreateInfo-range-00930, which is basically the same thing.

https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkBufferViewCreateInfo.html#VUID-VkBufferViewCreateInfo-range-00930

Now, vkCreateBufferView() is able to return errors, but if I make it return VK_ERROR_VALIDATION_FAILED_EXT on the relevant cases then Spider-Man GPU hangs on the main menu. But an interesting thing is that, when I do this, the white parts of his uniform (the spider in his chest and the back of his hands), which previously showed rendering corruption, are now completely black. So this change really seems to have an effect on the current visual corruption.

Please notice: in Release builds the anv driver doesn't do any checking related to this at all, but as a result we end up programming the hardware with wrong values. In debug builds we catch this with an assertion deep in our call stack:

https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/intel/isl/isl_surface_state.c?ref_type=heads#L997

... as a result, the driver ignores the errors. On the other hand, both games show some rendering corruption. Still, we shouldn't be issuing invalid Vulkan commands.

Is there something that can be done about this issue in the vkd3d-proton layer?

Thanks in advance!

[doitsujin]

What kind of numbers are we talking about here? D3D12 requirements are at least 2^27 texels regardless of format.

There's also the problematic case of games using typed descriptors but passing non-typed buffer descriptors, which is an extremely common app bug that works on all D3D12 drivers that we need to work around by creating typed views. If that's what's happening here then the best we can do is to just not create the typed view and have the game be broken entirely, or create a smaller view and most likely still completely break the game, but at least that way it wouldn't trigger Vulkan validation. Emulating larger views via multiple smaller views is not possible for us.

[pzanoni-intel]

@doitsujin:

What kind of numbers are we talking about here?

For Anv, maxTexelBufferElements is 134217728 (aka 2^27).

On Spider-Man, we see:

error: num_elements 209715200 > 134217728 (size:838860800 stride:4)
error: num_elements 209715200 > 134217728 (size:838860800 stride:4)
error: num_elements 209715200 > 134217728 (size:838860800 stride:4)
error: num_elements 419430400 > 134217728 (size:838860800 stride:2)

Edit:

On Assassin's creed:

(gdb) print num_elements
$1 = 268435456
(gdb) print buffer_size
$2 = 536870912
(gdb) print info->stride_B
$3 = 2

format = ISL_FORMAT_R16_UINT

[doitsujin]

I'll have a look next week at what the game is doing, but the fact that there's an R16 format involved is a relatively strong indication that this is just coming directly from the app and is just expected to work somehow.

[pzanoni-intel]

For Spider-Man the formats are:

ISL_FORMAT_R8G8B8A8_UNORM
ISL_FORMAT_R10G10B10A2_UNORM
ISL_FORMAT_R16G16_SINT
ISL_FORMAT_R16_UINT

[doitsujin]

Confirmed that it's the app itself creating these views.

Also confirmed that these work as expected both on native AMD/Nvidia drivers, as well as the respective Vulkan drivers, despite NV also reporting 2^27 as an upper limit, as in, querying the size and accessing the last element of the view works just fine. I strongly suspect that the real limit for this kind of hardware is something along the lines of 2GB for texel buffers rather than a specific element count, and 2^27 is conservative enough to cover R32G32B32A32 formats.

D3D12 does not have runtime validation for this or any sort of query for the app, so we really can't do much here.

[pzanoni-intel]

Thank you for your assessment here!

So I'm not sure what to do here other than remove the assert() in the Anv driver, but at least now I have better information. If you have any ideas/suggestions, please let me know.

[werman]

For the record, the same issue is present on Turnip, it leads to bad UVs on different objects in Spider-Man Remastered.