学习 Wgpu

# News (Pre 0.12)

# Pong working on the web

This took a little while to figure out. I ended up using wasm-pack to create the wasm as I was having trouble with getting wasm-bindgen to work. I figured it out eventually but decided to keep using wasm-pack as I felt that the work flow would be more friendly to readers.

I would have released this sooner, but I wanted to add support for touch so that people on their phones could play the game. It appears that winit doesn't record touch events for WASM, so I shelved that idea.

If you want to check out the demo, just head over to the write up

# 0.12 further changes to Surface

SwapchainFrame is no longer a thing. Instead get_current_texture will return a SurfaceTexture directly. This means that getting a frame to draw to looks somethings like this:

let output = self.surface.get_current_texture()?;
let view = output
    .texture
    .create_view(&wgpu::TextureViewDescriptor::default());

Another change is that you must call SurfaceTexture::present() after you submit your render command buffers to the queue. It goes something like this:

self.queue.submit(iter::once(encoder.finish()));
output.present();

There a good deal of internal changes such as WebGL support (which I really need to cover). You can check out more on wgpu's changelog.

# Pong is fixed for 0.10

It wasn't actually that hard of a fix. I only really use the swapchain directly in the render module, and the only other change required me to include the spirv feature to wgpu in Cargo.toml.

# 0.10 SwapChain is dead, long live the Surface!

SwapChain and all related code has been removed from wgpu. All code pertaining to obtaining textures to draw to from the window will be available from the Surface instead. That means configuring SurfaceTextures will look something like this:

let config = wgpu::SurfaceConfiguration {
    usage: wgpu::TextureUsages::RENDER_ATTACHMENT,
    format: surface.get_preferred_format(&adapter).unwrap(),
    width: size.width,
    height: size.height,
    present_mode: wgpu::PresentMode::Fifo,
};
surface.configure(&device, &config);

Resizing the surface will use similar code:

if new_size.width > 0 && new_size.height > 0 {
    self.size = new_size;
    self.config.width = new_size.width;
    self.config.height = new_size.height;
    self.surface.configure(&self.device, &self.config);
}

Finally, getting a SurfaceTexture to draw to will use the surface directly.

let output = self.surface.get_current_texture()?;
let view = output
    .texture
    .create_view(&wgpu::TextureViewDescriptor::default());

The Pong and imgui examples are broken again. I may remove the imgui example as the corresponding crate already has examples on how to use it. I'm also considering reworking the Pong example, but I may end up just updating it.

# Pong and imgui demos are fixed!

The imgui_wgpu and wgpu_glyph crates have been updated to wgpu 0.8 so I was able to fixed the demos! They both still use GLSL, and I don't think I'll be changing that for now. I may switch them over to naga at some point.

# 0.8 and WGSL

# The GLSL shaders have been translated to WGSL

Originally I wanted to wait until the WGSL spec fully stabilized, but due to some issues with the GLSL code, I've decided to switch over the code now. While GLSL is supported by WebGPU, it's currently secondary to WGSL. I'll keep an example of how to use GLSL (and maybe add HLSL and Metal as well), but I'm going to use WGSL from now on.

# Shaderc has been removed

I've been thinking about doing this for a while now. Because shaderc is a c library, it often has to be redownloaded during builds. This has been slowing down my ability to add new content and maintain old content. I had been considering switching to naga earlier, but some of my shaders (notably the lighting ones) weren't compiling with naga as I was using features not available for compatibility reasons (inverse is not available in all languages targeting spirv).

Since I needed to make a bunch of changes to the code base to make the glsl, and because I wanted to switch the tutorial to WGSL anyways, I decided to bite the bullet and recode everything in WGSL and remove shaderc from the tutorials.

# Some of the showcase examples are broken

The wgpu_glyph, and imgui-wgpu crates currently depend on wgpu 0.7, which is causing the pong and imgui-demo to not compile. I decided to excluded them from the workspace until the underlying crates update to using wgpu 0.8. (Feel free to submit a issue or even PR when that happens!)

# Various API changes

  • The depth field is now depth_or_array_layers
  • wgpu::VertexFormat::Float3 is now wgpu::VertexFormat::Float32x3. Similar things apply to Float2 and Float4
  • CullMode is no longer a thing, instead PrimitiveState::cull_mode will require an Option<Face>
  • Added clamp_depth and conservative to PrimitiveState. Part of this means that DepthStencilState no longer has a clamp_depth field.
  • color_blend and alpha_blend have been moved into the new blend field with accepts an Option<wgpu::BlendState>
  • adapter.get_preferred_format() now returns an Option<wgpu::TextureFormat>
  • wgpu::RenderPassColorAttachmentDescriptor has been renamed wgpu::RenderPassColorAttachement and the attachment field has been renamed to view
  • wgpu::RenderPassDepthStencialAttachmentDescriptor also loses the Descriptor part of it's name. attachment gets renamed to view as well.
  • wgpu::TextureCopyView has been renamed to wgpu::ImageCopyTexture. This is a typedef for wgpu::ImageCopyTextureBase<T>
  • wgpu::TextureDataLayout is now wgpu::ImageDataLayout and it's bytes_per_row and rows_per_image now take NonZeroU32.
  • wgpu::ImageCopyBuffer is now wgpu::ImageCopyBuffer.

# 0.7

There were a lot of changes particularly to the RenderPipelineDescriptor. Most other things have not changed. You can check out the 0.9 PR for the full details.

# November 2020 Cleanup, Content Freeze, and Patreon

School is starting to ramp up, so I haven't had as much time to work on the site as I would like to. Because of that there were some issues piling up. I decided to tackle a bunch of them in one go. Here's a snapshot of what I did:

  • The tutorial now handles SurfaceError properly
  • I'm now using bytemuck's derive feature on all buffer data structs.
  • The instancing tutorial now uses vertex buffers instead of storage buffers.
  • build.rs now updates when individual shaders are changed, not whenever /src is changed.
  • Had some help from Github user @kanerogers to clean up the texturing tutorial.
  • I made a compute pipeline showcase that computes the tangent and bitangent for each vertex in a model.
  • I made a imgui showcase. It's very basic, but it should be a good starting point.

Now in the headline I mentioned a "Content Freeze". Wgpu is still a moving target. The migration from 0.4 to 0.5 was lot of work. The same goes for 0.5 to 0.6. I'm expected the next migration to be just as much work. As such, I won't be added much content until the API becomes a bit more stable. That being said, I still plan on resolving any issues with the content.

One more thing. This is actually quite awkward for me (especially since I'll be slowing down development), but I've started a patreon. My job doesn't give me a ton of hours, so things are a bit tight. You are by no means obligated to donate, but I would appreciate it.

You can find out more about contributing to this project on the introduction page

# 0.6

This took me way too long. The changes weren't difficult, but I had to do a lot of copy pasting. The main changes are using queue.write_buffer() and queue.write_texture() everywhere. I won't get into the nitty gritty, but you can checkout the pull request if you're interested.

# Added Pong Showcase

See it here

# Normal mapping

My perfectionism got in my way a bit with this one. I wasn't sure that what I was getting was "physically accurate", but it seems to look good.

# 0.5!

Too many things changed to make note of them here. Check out the 0.5 pull request if you're curious about specifics. That being said, 2 things are worth mentioning directly: the y-axis now points up like with DirectX and Metal, and requesting an adapter and creating a device now use Futures. The tutorials have been updated as well as the code.

# Reworked lighting tutorial

The lighting tutorial was not up to par, so I redid it.

# Added GIF showcase

Creating GIFs

# Updated texture tutorials

Up to this point, we created textures manually everytime. I've pulled out the texture creation code into a new texture.rs file and included it every tutorial from the textures tutorial onward.

# Fixed panics do to not specifying the correct usage

Wgpu has become more strict about what BufferUsagess and TextureUsagess are required when performing certain operations. For example int the Wgpu without a window example, the texture_desc only specified the usage to by COPY_SRC. This caused a crash when the texture was used as a render target. Adding OUTPUT_ATTACHMENT fixed the issue.

# Updating Winit from 0.20.0-alpha5 to 0.20

There were a lot of small changes to how the dpi stuff works. You can see all the changes in the changelog. That means that some of the tutorials had to change.

  • I've removed hidpi_factor from State entirely. They removed the hidpi_factor() method from winit::window::Window, and changed inner_size() to return PhysicalSize instead of LogicalSize, so we don't need to store the hidpi_factor anymore.
  • update_hidpi_and_resize is no more. Since ScaleFactorChanged passes in the windows new PhysicalSize, we can simply use resize().
  • State::size is now PhysicalSize<u32> instead of the pre 0.20 LogicalSize.
  • EventsCleared is now MainEventsCleared.

I may have missed a change, but I made sure that all the examples compile an run, so if you have trouble with your code you can use them as a reference.

# Changed tutorial examples to use a src directory

I wasn't using the traditional cargo binary folder setup. I've changed it to the standardized form now.

# Updating to 0.4 from 0.3

There are a few things that have changed:

  1. The use of Instance has been removed. Creating a Surface and requesting an Adapter are done as follows.
let surface = wgpu::Surface::create(window);
let adapter = wgpu::Adapter::request(&wgpu::RequestAdapterOptions {
    ..Default::default()
}).unwrap(); // needs to be unwrapped
  1. The request_device method now returns a (Device, Queue) tuple. This means that you can borrow the Queue mutably while using the Device immutably. Because of this change, submitting CommandBuffers to the queue uses the submit method on the Queue directly.
self.queue.submit(&[
    encoder.finish()
]);
  1. The create method on Surface takes in any struct that implements the HasRawWindow trait, instead of a RawWindowHandle. This means that the raw-window-handle = "0.3" line in Cargo.toml is no longer needed.

I don't know if this is a change from 0.4, but you use wgpu = "0.4" line in dependencies instead of the [dependencies.wgpu] as wgpu will determine the best back end for you.

# New/Recent Articles

Last Updated: 4/30/2022, 2:50:47 PM

←