This extension adds functionality to avoid having to pass down complete SPIR-V to shaders in situations where we speculate that an implementation already has a pipeline blob in cache and conversion to SPIR-V is not needed to begin with.

Problem Statement

In some applications, SPIR-V is generated on-the-fly, usually by translating from some other representation. API translation libraries and emulators in particular frequently run into these problems.

In such applications, the overhead required to obtain valid SPIR-V before any pipeline creation call can be problematic. Especially in graphics API translation layering efforts, applications expect that compilation with hot caches is "instant", as that is how a native driver would behave. Translating to SPIR-V can therefore become a performance problem. There are two common problems:

  • Applications compile PSOs late → stutter! → but we are expected to mitigate

  • Applications compile a lot of PSOs early → good! but can lead to excessive load times even on subsequent runs of application

For translation layers, there are currently two options we can consider to mitigate the issue:

  • Optimize the translation

  • Cache converted SPIR-V on disk

Neither option may be good enough. Disk requirements for large application caches can be impractical on some platforms, since we might end up having to store in the order of 100k SPIR-V modules, easily in the gigabyte range. Optimizing the translation might not be enough when faced with tens of thousand pipelines being compiled at once.

Solution Space

The solution this extension addresses is the minimum viable approach to fix the problem. The main idea is that when pipeline caches are primed, SPIR-V modules are largely useless, since most implementations are likely to only hash, and never look at the SPIR-V again. We can just hand back the hash to the implementation instead.

The extension is designed to work on top of VK_EXT_pipeline_creation_cache_control. We can reuse the main idea of a "non-blocking" compile where we return early if pipeline compilation is required, and the translation layer can build SPIR-V as needed. Next time, we are likely to hit in cache.

An important consideration here is that this solution is intended to aid internal implementation caching, i.e. a "magic disk cache", which most desktop implementations of graphics APIs are expected to have.

For explicit application side caching mechanisms, larger cache sizes are reasonable and expected, but we are more constrained with internal caches. These should be as lean and mean as possible, but internal caches are also more "fuzzy" in nature. Spurious failure is okay, a "best effort" approach is suitable for this use case.

One could extend this idea to full PSO keys as well, but that is better left to other proposals.

Example use case

One scenario where this extension has been found to be particularly useful is D3D12 to Vulkan translation. The translation layers need to translate DXBC and DXIL code to SPIR-V, which is then translated to GPU ISA. SPIR-V to ISA translation is cached by Vulkan pipeline caches or in-driver caches, but the DXBC/DXIL → SPIR-V cache is not covered by the API.

We can store SPIR-V on-disk and reload that in response to a pipeline creation call, but the overhead of storing SPIR-V on disk, validating it, decompressing it, etc, is a significant overhead that can be avoided, with storage space being the most significant problem. If the final ISA is present in pipeline caches, we do not really need the SPIR-V at all.

We have observed >95% disk savings with this scenario, and this is transformative since it makes it practical to share this cache across different machines. This hypothetically allows an end-user to never observe shader compilation stutter or excessive load times on first run of a game.


Querying identifier

After the application has converted a shader to SPIR-V and compiled a pipeline, vkGetShaderModuleIdentifierEXT is used to obtain a shader identifier. This identifier can be stored on-disk for later use. (vkGetShaderModuleCreateInfoIdentifierEXT can be used as an object-less alternative.) VkPhysicalDeviceShaderModuleIdentifierPropertiesEXT::shaderModuleIdentifierAlgorithmUUID is also needed so applications know if we need to throw away any caches using the identifier. This should only happen on different driver implementations. Different versions of the same driver are not expected to change hashing algorithms. For drivers sharing the same framework (e.g. Mesa), the module hashing algorithm could even be the same one.

To make the API friendly to applications, there is a small upper bound on how large an identifier may be, so that the identifiers can be retrieved without memory allocation.

VK_NULL_HANDLE module proxy

On subsequent runs of an application, we speculate that the driver caches (or VkPipelineCaches) are primed, and thus having SPIR-V is not useful anymore. We then set VkPipelineShaderStageCreateInfo::module to VK_NULL_HANDLE and chain in VkPipelineShaderStageModuleIdentifierCreateInfoEXT as a proxy. This allows a driver to generate the same internal PSO key that it would generate if we passed in actual SPIR-V. VK_PIPELINE_CREATE_FAIL_ON_COMPILE_REQUIRED_BIT must be set in this situation, since this is a speculative compile by definition.

Handling fallbacks

In a situation where we do not have the pipeline cached, we receive VK_PIPELINE_COMPILE_REQUIRED, and fall back to re-creating SPIR-V as usual.

Soft guarantees of successfully compiling pipelines

The proposal as-is states that implementations may fail compilation for any reason. This is a defensive measure to make it possible for this extension to interoperate with layers, validation, debug tooling, etc., without too many problems. In most such layers, there is a need to parse the SPIR-V itself to figure out information required for correct operation. While the ICD might recognize an identifier, a layer might not, and therefore they might need the escape hatch where they can spuriously fail compilation.

This effectively makes the spec somewhat vague, and it becomes a quality-of-implementation issue on what ICDs do. This is not different from what implementations already do either way. After all, you may or may not have a PSO in disk cache and that is okay.


RESOLVED: Should applications be allowed to specify their own shader module identifier?


It is plausible that applications might want to generate their own keys instead of using driver-generated keys. For this to be useful, an application will need to generate a key which depends on input data/shaders, the revision of the code which performs runtime conversion to SPIR-V, and potentially, the driver kind or any configuration options which affect shader conversion. A typical problem which comes up when doing forward hashing like this is that hashes can change for every revision of the application, even if the resulting SPIR-V ends up being identical. This will easily contribute to pipeline cache bloat, since the exact same pipelines might end up in cache with different hashes. Implementations can be defensive about this and introduce extra identifier indirections, e.g. have an extra hashmap for application identifier to driver identifier, but ideally, this extension should not introduce extra implementation complexity to support it well.

Applications could also hash the resulting SPIR-V and ensure non-duplicated identifiers this way, but this is not meaningfully different from just using the driver identifier, and also avoids added implementation complexity.

RESOLVED: How does this interact with VK_KHR_ray_tracing_pipeline, VK_KHR_pipeline_library and VK_EXT_graphics_pipeline_library?


When using pipeline libraries, there are two scenarios where pipeline creation can fail if we only have an identifier, at creation time of the library, and the consumption of that library.

There are at least three possibilities an implementation could consider when building libraries and consuming them:

  • Generate final code when creating library, link step is trivial. Ray tracing pipeline libraries may be implemented like this.

  • Generate code when creating library, but allow link-time optimization for later. Graphics pipeline libraries is a common case here.

  • Just retain a reference to the shader module, perform actual compilation during linking. Another strategy for ray tracing libraries.

In the latter two scenarios, it is reasonable to assume that compilation may happen during the final pipeline build and compilation would spuriously fail if the source module was only defined by identifier and the final PSO did not exist in cache. If we do not allow compilation to fail with VK_PIPELINE_CREATE_FAIL_ON_COMPILE_REQUIRED_BIT here, it would not be safe to return VK_SUCCESS from the library creation step, which would be unfortunate.

For scenarios where the implementation may generate code later, we require that any pipeline libraries which were created with identifiers inherit the requirement of using VK_PIPELINE_CREATE_FAIL_ON_COMPILE_REQUIRED_BIT. This allows applications to speculatively create link-time optimized pipelines from identifiers only as well as ray-tracing pipelines from libraries.

RESOLVED: Should there be stronger guarantees on when pipeline compilation with identifier must succeed?


The existing proposal gives a lot of lee-way for implementations to spuriously fail compilation when module is VK_NULL_HANDLE. It might be possible to give stronger guarantees with tighter spec language?

CTS testing will report quality warnings if identifiers cannot be used with VkPipelineCache, as there is no good excuse why an implementation should not be able to satisfy those pipelines.

Further Functionality