In order to support external textures, we must be able to map a single
external texture resource binding to multiple Vulkan descriptors. This
means we must be able to override the `Binding` and `DescriptorSet`
values for global variables when generating SPIR-V, rather than simply
passing through the group and binding values from Naga IR.
This patch extends the existing SPIR-V Naga backend's `BindingMap` to
contain a descriptor set and binding value in addition to the existing
array size. When creating BindGroupLayouts/BindGroups we use a
sequentially incrementing value for each entry's binding value,
continuing to just use the bind group index as the descriptor set
value.
The Naga backend looks up each resource in the map when emitting its
`Binding` and `DescriptorSet` decorations. If the entry cannot be
found in the map, it will either error or emit fake bindings based on
its configuration.
This is a refactoring of Vulkan BindGroupLayout and BindGroup creation
in preparation for implementing external texture support.
Currently when creating a BindGroupLayout the Vulkan backend creates a
list of the vk::DescriptorType for each entry, as well as the count, for
binding arrays. Then when creating the BindGroup, it iterates through
this list and does whatever it needs to do for each entry based on these
values.
In order to support external textures, we will have to create multiple
descriptors for each BindingType::ExternalTexture. This means we cannot
map each binding type to a single Vulkan descriptor type. Instead, store
the list of BindGroupLayoutEntries on the BindGroupLayout and use those
when creating the BindGroup, using the same "layout_and_entry_iter"
idiom used by other backends. In subsequent patches this will allow us
to create multiple descriptors for a single resource binding.
I think what happened here is that another reference to 2.9.3 was added
to Cargo.lock in parallel with the renovate change that bumped the
version to 2.9.4.
In the future `transfer` will want to use the `Arc` versions of these
types. I have exported them from `wgpu_core::command::ffi`, to document
that these are used for FFI purposes, and from `wgpu_core::command`, for
backwards compatibility, although I also change the in-tree uses to
use them from `wgpu_types` instead of from `wgpu_core::command`.
It is emptied by `reset_queries` at the end of every render pass, so
it's just keeping an allocation alive, not holding any state. It seems
unlikely that there is sufficient performance gain from reusing the
memory allocation to justify the complexity of additional state at
higher layers.
Previously, the check was skipped if the copy was a single row, which is
not correct. The check should be made whenever bytes_per_row is
specified. It is permissible not to specify bytes_per_row if the copy is
a single row, but if it is specified, it must be aligned.
Also removes a redundant check of the `offset` alignment.
Since the offset and bytesPerRow alignment checks are not part of
"validating linear texture data", I chose to remove that instance of
them. These checks are now in `validate_texture_buffer_copy`, which
does not correspond 1:1 with the spec, but has a comment explaining how
it does correspond.
This also allow to remove the MESA_VERSION and VULKAN_SDK_VERSION
duplication between the ci workflow and the install action.
To pass down the version in each steps, I added action inputs
because there is no way as of today to define environments
variables that are shared between steps in composite actions
(no top level `env:` like for workflows).
TransferError now has separate variants for texture copy formats that
are only forbidden in combination with specific aspects
(CopyFrom/ToForbiddenTextureFormatAspect), and texture copy formats that
are always forbidden, irrespective of the aspect
(CopyFrom/ToForbiddenTextureFormat).
This produces a less confusing error message by not mentioning the
aspect it is not relevant.
Although the operation of these functions is defined in terms of f16
semantics, the input/output types are not f16, and they are generally
available even when native `f16` support is not. But in at least one
case, they are only available with `f16` support, so add a new downlevel
flag that is cleared when these functions are not available.
Add some infrastructure to simplify testing of missing
capabilities/extensions, and add tests for a few more kinds of f16
usage.
Co-authored-by: Erich Gubler <erichdongubler@gmail.com>
