Separates the Vulkan feature sets
`VkPhysicalDeviceShaderFloat16Int8Features` and
`VkPhysicalDevice16BitStorageFeatures`, which previously were used
"together, or not at all".
This commit should not change any behavior yet, but I'd like to run full
CI tests on it for now. If the CI tests pass, I'll use this separation
to enable the `shader_int8` feature separately from the rest of the
features to enable optimizations of `[un]pack4x{I,U}8[Clamp]` on SPIR-V.
Changes the MSL and HLSL backends to support writing only a single entry
point, and uses them that way in wgpu-hal.
This is working towards a fix for #5885.
* Increase the limit in test_stack_size
The `metal` crate is currently unsound regarding unknown/future enum
variants, see https://github.com/gfx-rs/metal-rs/issues/209 and
https://github.com/rust-lang/rfcs/pull/3803.
`objc2-metal` fixes this by emitting C enums as a newtype + constants
for each variant, but that prevents us from importing the
variants/constants. So this commit converts to a pattern that works with
that in preparation for the migration.
A lot of Metal types are prefixed with MTL, which makes it quite clear
where they're coming from. This means that we don't loose any clarity if
we import them instead of having them prefixed with `metal::`.
This will make it easier to migrate to `objc2-metal` since that crate is
named differently from the `metal` crate.
This makes texture view destruction cheap since users of wgpu-hal are required to keep resources referenced by encoders alive for at least as long as the encoder is alive.
This is also a prerequisite to implement rendering to slices of 3D textures (by creating temporary texture views).
On the other hand the cache won't be as effective but that is probably ok, we can reevaluate the implementation if it turns out to be a problem.
Imageless framebuffers are not needed, we know which views will be attached to the render pass already. Even if we wanted to more aggressively cache imageless framebuffers that wouldn't be possible since they still require specifiying view usage, texture usage and view formats.
Removing usage of imageless framebuffers simplifies the code substantially.
The D3D12 API doesn't guarantee that it returns `E_OUTOFMEMORY` in high memory pressure situations; drivers/kernel will happily start swapping objects that were in VRAM to RAM and then RAM to DISK, slowing down the system to a crawl if done in a loop.