Wgpu

Added:

* The wgpu backend can be forced using the `WGPU_BACKEND_TYPE` env variable.
Values can be e.g. "D3D12", "Metal", "Vulkan".
* Initial support for off-screen canvases.
* Adds `adapter.is_software` property.

Changed:

* The `GPUPresentationContext` class has been renamed to `GPUCanvasContext`.
* The functionality of the swap-chain has moved to the `GPUCanvasContext`.
* The now removed `GPUSwapChain` was used as a context manager. Instead,
the frame is presented (ala GL swapbuffers) automatically at the end of a draw.
* The `canvas.configure_swap_chain()` method has been removed. Instead,
`canvas.get_context()` should be used, to obtain a present/canvas context.
* The `adapter.request_device()` method has its arguments `non_guaranteed_features`
and `non_guaranteed_limits` replaced with `required_features` and `required_limits`.
* The enum field `StoreOp.clear` is now `StoreOp.discard`.
* The flag field `TextureUsage.SAMPLED ` is now `TextureUsage.TEXTURE_BINDING `.
* The flag field `TextureUsage.STORAGE ` is now `TextureUsage.STORAGE_BINDING `.
* The enum `InputStepMode` is now `VertexStepMode`.
* WGSL: `arrays` must be declared as `var` (not `let`) in order to allow dynamic indexing.
* WGSL: storage classes are written differently.

Changed:

* The backend selection is automatic by default. To force a backend, the `WGPU_BACKEND_TYPE` environment variable can be set to e.g. "Vulkan". It could be good to do this on Windows to prevent selection of DX12 for now.

Added:

* `adapter.properties` now has actual values, allowing inspection of the selected
GPU and backend.
* Added back support for filtering float32 textures by enabling a certain wgpu feature
by default.

Fixed:

* An error in the docs of `create_render_pipeline`.
* Vulkan backend is now forced to prevent DX12 being select and causing probems
because it's less mature.

This release uses a new version of wgpu-native which has changed quite a bit internally. There
is more validation (thus more restrictions). There are only a few changes to the API.
However, one big change is that shaders can now be provided as both SpirV and WGSL. Due to
the strict validation, most shaders compiled by PyShader are not usable anymore. We
recommend using WGSL instead.

Added:

* Added `GPUAdaper.properties` (the amount of information it contains will increase in the future).
* Added proper support for WGSL.

Changed:

* Renamed `renderpass.set_blend_color` -> `set_blend_constant`.
* Stricter validation of SpirV shaders.
* Float32 texture formats must now use a non-filtering sampler and texture-sample-type.
* Integer texture formats can no longer use a texture (use `textureLoad` instead).
* ... and more tighter restrictions.

Removed:

* The API concerning debug markers and groups is temporarily removed.
* Adapter and device features is temporarily removed.
* Adapter and device limits is temporarily removed.

- Crates:
- `wgpu-types`: common types between native and web targets
- `wgpu-core`: internal API for the native and remote wrappers
- Features:
- based on gfx-hal-0.5
- moved from Rendy to the new `gfx-memory` and `gfx-descriptor` crates
- passes are now recorded on the client side. The user is also responsible to keep all resources referenced in the pass up until it ends recording.
- coordinate system is changed to have Y up in the rendering space
- revised GPU lifetime tracking of all resources
- revised usage tracking logic
- all IDs are now non-zero
- Mailbox present mode
- Validation:
- active pipeline
- Fixes:
- lots of small API changes to closely match upstream WebGPU
- true read-only storage bindings
- unmapping dropped buffers
- better error messages on misused swapchain frames

wgpu-core-0.4.3 (2020-01-20)

- improved swap chain error handling

wgpu-core-0.4.2 (2019-12-15)

- fixed render pass transitions

wgpu-core-0.4.1 (2019-11-28)

- fixed depth/stencil transitions
- fixed dynamic offset iteration

This release represents about half a year of progress on the WebGPU API, so the API
has changed quite a bit. The wgpu-py API more closely reflects the webgpu API - wgpu-native does
not affect the API except for a few additional features.

Added:

* Added `GPUQueue.read_buffer` as extra API (next to `write_buffer` which is original WebGPU API).
* Added `GPUQueue.read_texture` as extra API.

y
Removed:

* Removed `GPUBuffer.read_data()`. Use `device.queue.read_buffer()` instead. Note that `usage` `MAP_READ` should be replaced with `COPY_SRC`.
* Removed `GPUBuffer.write_data()`. Use `device.queue.write_buffer()` instead. Note that `usage` `MAP_WRITE` should be replaced with `COPY_DST`.

Changed:

* `GPUCanvasContext.get_swap_chain_preferred_format()`: now takes an `adapter` instead of a `device`.
* `GPUAdapter.extensions`: is now called `features`.
* `GPUAdapter.request_device()`: the `extensions` and `limit` args are now `non_guaranteed_features` and `non_guaranteed_limits`.
* `GPUDevice.default_queue`: is now called `queue`.
* `GPUDevice.create_compute_pipeline()`: the `compute_stage` arg is now called `compute`.
* `GPUDevice.create_bind_group_layout()` has changed the required structure of the layout entry dicts.
* `GPUDevice.create_render_pipeline()` has changed *a lot* in terms of shape of input dicts. See new docs.
* `GPUTexture.create_view()`: args `mip_level_count` and `array_layer_count` are default `None` instead of `0`.
* `GPUCommandEncoder.begin_render_pass()`: the `color_attachments` and `depth_stencil_attachment` arguments have their `attachment` field renamed to `view`.
* `GPURenderEncoderBase.set_index_buffer()` has an extra argument (after the buffer) to specify the format. The index format is no longer specified in `device.create_render_pipeline()`.
* Flag `TextureUsage` has field OUTPUT_ATTACHMENT renamed to RENDER_ATTACHMENT.
* Enum `BindingType` is split up in different enums for buffer, sampler, sampled texture and storage texture.
* Enum `BlendFactor` has some of its field names changed.
* Enum `VertexFormat` has its field names changed, e.g. ushort2 -> uint16x2.
* The API is more restrictive in the use of buffer/texture usage combinations.
* The API is more restrictive in formats for storage buffers/textures.
* When copying from/to textures, the `bytes_per_row` must now be a multiple of 256.