egui_winit_vulkano

This is an egui integration for winit and vulkano.

You'll need a Vulkano target image as an input to which the UI will be painted. The aim of this is to allow a simple enough API to separate UI nicely out of your renderer and make it easy to build your immediate mode UI with Egui.

Usage

Create your own renderer with Vulkano, and allow access to Vulkano's gfx queue Arc<Queue> and Vulkano's winit surface Arc<Surface<Window>>
Create Gui integration with the surface & gfx queue

// Has its own renderpass. Modify GuiConfig to determine image clear behavior etc.
let mut gui = Gui::new(&event_loop, renderer.surface(), renderer.queue(), renderer.swapchain_format(), GuiConfig::default());
// Or with subpass. This means that you must create the renderpass yourself. Egui subpass will then draw on your
// image.
let mut gui = Gui::new_with_subpass(&event_loop, renderer.surface(), renderer.queue(), renderer.swapchain_format(), subpass, GuiConfig::default());

Inside your event loop, update gui integration with WindowEvent

gui.update(&event);

Fill immediate mode UI through the integration in Event::RedrawRequested before you render

gui.immediate_ui(|gui| {
    let ctx = gui.context();
    // Fill egui UI layout here
});

// Or

gui.begin_frame();
// fill egui layout...


// And when you render with `gui.draw_on_image(..)`, this will finish the egui frame

Render gui via your renderer on any image or most likely on your swapchain images:

// Acquire swapchain future
let before_future = renderer.acquire().unwrap();
// Render gui by passing the acquire future (or any) and render target image (swapchain image view)
let after_future = gui.draw_on_image(before_future, renderer.swapchain_image_view());
// Present swapchain
renderer.present(after_future, true);
// ----------------------------------
// Or if you created the integration with subpass
let cb = gui.draw_on_subpass_image(framebuffer_dimensions);
draw_pass.execute(cb);

Note that Egui strongly prefers UNORM render targets, and passing an sRGB color space image is considered an error. See The correct color space for more details on how to deal with it.

See the examples directory for better usage guidance.

Remember, on Linux, you need to install following to run Egui

sudo apt-get install libxcb-render0-dev libxcb-shape0-dev libxcb-xfixes0-dev

Examples

./run_all_examples.sh

The correct color space

The key takeaway is that egui assumes any math in the shaders as well as alpha blending is done using colors in sRGB color space. Meanwhile, Vulkan assumes that one always works with colors in linear space, and will automatically convert sRGB textures to linear colors when using them in shaders. This is what we need to work around.

Longer explainer

What usually happens

The GPU converts all colors to linear values. sRGB textures are mainly an implementation detail that the GPU can use to store colors for human vision with fewer bits.

flowchart LR
    A[Shader A] -->|output linear colors| BB(UNORM Image)
    BB -->|reads linear colors| B[Shader B]

    C[Shader A] -->|GPU converts to sRGB| DD(sRGB Image)
    DD -->|GPU converts to linear colors| D[Shader B]

Meanwhile, Egui expects that the input and output colors are in sRGB color space. Very unusual, but happens to make sense for them.

So with Egui, we can use a we tricks.

We can always (ab)use a UNORM texture to store values without any conversions.

flowchart LR
    A[EGui] -->|output sRGB colors \n no conversion| BB(UNORM Image)
    BB -->|directly reads colors \n no conversion| B[Your shader gets sRGB values]

Your shader can then take the sRGB values and manually convert them to linear values.

The alternative, and better trick is using an image with multiple views.

flowchart LR
    I(Image)
    I1(UNORM Image View)
    I2(sRGB Image View)

    A[EGui] -->|output sRGB colors \n no conversion| I1

    I2 -->|converts sRGB to linear| B[Your shader gets linear values]

    I --- I1
    I --- I2

The same two techniques can also be applied to swapchains.

The simplest approach is to create the swapchain using a UNORM image format, as is done in most examples. The display engine will read the swapchain image, and interpret it as sRGB colors. This happens regardless of the swapchain format. When using this approach, Egui works as expected. However, all of your shaders that draw to the swapchain will need to manually convert pixels to the sRGB format.
You render into a UNORM image you allocated, and then copy that image onto your swapchain image using compute shaders or a fullscreen triangle. Your shader needs to read the UNORM image, which if you remember contains values in sRGB color space, manually convert it from sRGB to linear color space, and then write it to your sRGB swapchain image. This conversion is also why you cannot just blit the image, but need to use a shader in between.
You again render into an image you allocated with the VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT flag (the format may be UNORM or SRGB). This flag allows you to create image views with compatible but differing formats. Create two image views, one viewing your data as UNORM and one as sRGB (e.g. VK_FORMAT_R8G8B8A8_UNORM and VK_FORMAT_R8G8B8A8_SRGB) by modifying ImageViewCreateInfo::format parameter from the default value. Render your Egui into the UNORM image view, then you may blit it from the sRGB image view onto your framebuffer.
If your device supports VK_KHR_swapchain_mutable_format, which AMD, Nvidia and Intel do on both Windows and Linux, you may get around having to allocate an extra image and blit it completely. Enabling this extension and passing the VK_SWAPCHAIN_CREATE_MUTABLE_FORMAT_BIT_KHR flag when creating your swapchain, which should be in sRGB format, allows you to create image views of your swapchain with differing formats. So just like in 2, you may now create two image views, one as sRGB to use in your 3D rendering, and one as UNORM to use in a following Egui pass. But as you are rendering directly into your swapchain, there is no need for copying or blitting.

For desktop platforms, we would recommend approach 4 as it's supported by all major platforms and vendors, the most convenient to use, and saves both memory and memory bandwidth. If you intend to go for maximum compatibility, we recommend implementing approach 3 as it does not require any extra features or extensions, while still supporting approach 4 as it should be trivial to implement. Approach 2 is only interesting if you can combine the in-shader copy with some other post-processing shaders you would have to run anyway.

If you so wish, you may still draw to an image view in sRGB format, if you can accept discolorations when elements are alpha blended together. Doing so requires enabling GuiConfig::allow_srgb_render_target, as otherwise it is an error to draw to an sRGB image view. Normally, you would expect that drawing to the wrong color space will cause the entire UI to be discolored. But enabling that option also slightly changes the shader we are using for drawing UI elements, to only have discolorations in alpha blended areas instead of the entire image.

Notes

This integration would not have been possible without the examples from vulkano-examples or egui_winit_ash_vk_mem.

hakolao / egui_winit_vulkano