The end of 2022 is very close so I’m just in time for some self-promotion. As you may know, the ongoing collaboration between Valve and Igalia lets me and some of my colleagues work on improving the open-source Vulkan and OpenGL Conformance Test Suite. This work is essential to ship quality Vulkan drivers and, from the Khronos side, to improve the Vulkan standard further by, among other things, adding new functionality through API extensions. When creating a new extension, apart from reaching consensus among vendors about the scope and shape of the new APIs, CTS tests are developed in order to check the specification text is clear and vendors provide a uniform implementation of the basic functionality, corner cases and, sometimes, interactions with other extensions.
In addition to our CTS work, many times we review the Vulkan specification text from those extensions we develop tests for. We also do the same for other extensions and changes, and we also submit fixes and improvements of our own.
In 2022, our work was important to be able to ship a bunch of extensions you can probably see implemented in Mesa and used by VKD3D-Proton when playing your favorite games on Linux, be it on your PC or perhaps on the fantastic Steam Deck. Or maybe used by Zink when implementing OpenGL on top of your favorite Vulkan driver. Anyway, without further ado, let’s take a look.
This extension was created by our beloved super good coder Mike Blumenkrantz to be able to create a 2D view of a single slice of a 3D image. It helps emulate functionality which was already possible with OpenGL, and is used by Zink. Siru developed tests for this one but we reviewed the spec and are listed as contributors.
One of my favorite extensions shipped in 2022. Created by Hans-Kristian Arntzen to be used by Proton, this extension lets applications query identifiers (hashes, if you want to think about them like that) for existing VkShaderModule objects and also to provide said identifiers in lieu of actual VkShaderModule objects when creating a pipeline. This apparently simple change has real-world impact when downloading and playing games on a Steam Deck, for example.
You see, DX12 games ship their shaders typically in an intermediate assembly-like representation called DXIL. This is the equivalent of the assembly-like SPIR-V language when used with Vulkan. But Proton has to, when implementing DX12 on top of Vulkan, translate this DXIL to SPIR-V before passing the shader down to Vulkan, and this translation takes some time that may result in stuttering that, if done correctly, would not be present in the game when it runs natively on Windows.
Ideally, we would bypass this cost by shipping a Proton translation cache with the game when you download it on Linux. This cache would allow us to hash the DXIL module and use the resulting hash as an index into a database to find a pre-translated SPIR-V module, which can be super-fast. Hooray, no more stuttering from that! You may still get stuttering when the Vulkan driver has to compile the SPIR-V module to native GPU instructions, just like the DX12 driver would when translating DXIL to native instructions, if the game does not, or cannot, pre-compile shaders somehow. Yet there’s a second workaround for that.
If you’re playing on a known platform with known hardware and drivers (think Steam Deck), you can also ship a shader cache for that particular driver and hardware. Mesa drivers already have shader caches, so shipping a RADV cache with the game makes total sense and we would avoid stuttering once more, because the driver can hash the SPIR-V module and use the resulting hash to find the native GPU module. Again, this can be super-fast so it’s fantastic! But now we have a problem, you see? We are shipping a cache that translates DXIL hashes to SPIR-V modules, and a driver cache that translates SPIR-V hashes to native modules. And both are big. Quite big for some games. And what do we want the SPIR-V modules for? For the driver to calculate their hashes and find the native module? Wouldn’t it be much more efficient if we could pass the SPIR-V hash directly to the driver instead of the actual module? That way, the database translating DXIL hashes to SPIR-V modules could be replaced with a database that translates DXIL hashes to SPIR-V hashes. This can save space in the order of gigabytes for some games, and this is precisely what this extension allows. Enjoy your extra disk space on the Deck and thank Hans-Kristian for it! We reviewed the spec, contributed to it, and created tests for this one.
This one was written by Joshua Ashton and allows applications to put images in the special
VK_IMAGE_LAYOUT_ATTACHMENT_FEEDBACK_LOOP_OPTIMAL_EXT layout, in which they can both be used to render to and to sample from at the same time.
It’s used by DXVK 2.0+ to more efficiently support D3D9 games that read from active render targets.
We reviewed, created tests and contributed to this one.
I don’t need to tell you more about this one. You saw the Khronos blog post. You watched my XDC 2022 talk (and Timur’s). You read my slides. You attended the Vulkan Webinar. Important to actually have mesh shaders on Vulkan like you have in DX12, so emulation of DX12 on Vulkan was a top goal. We contributed to the spec and created tests.
“Hey, hey, hey!” I hear you protest. “This is just a rename of the VK_VALVE_mutable_descriptor_type extension which was released at the end of 2020.” Right, but I didn’t get to tell you about it then, so bear with me for a moment. This extension was created by Hans-Kristian Arntzen and Joshua Ashton and it helps efficiently emulate the raw D3D12 binding model on top of Vulkan. For that, it allows you to have descriptors with a type that is only known at runtime, and also to have descriptor pools and sets that reside only in host memory. We had reviewed the spec and created tests for the Valve version of the extension. Those same tests are the VK_EXT_mutable_descriptor_type tests today.
The final boss of dynamic state, which helps you reduce the number of pipeline objects in your application as much as possibly allowed. Combine some of the old and new dynamic states with graphics pipeline libraries and you may enjoy stutter-free gaming. Guaranteed!1 This one will be used by native apps, translation layers (including Zink) and you-name-it. We developed tests and reviewed the spec for it.
1Disclaimer: not actually guaranteed.