Automatically resize / recreate depth buffer

[marcdownie]

The with swap_chain as current_texture_view: trick causes swap_chain to recreate its native swap chain if the canvas it was produced from has been resized since the last __enter__. This makes resizing the canvas magically work --- so well you have to go looking for the resizing code.

All is well. But then you want to have a depth texture view associated with your canvas which needs to be resized at the same time or panic. Any additional sugar to be poured on this? An argument should_resize to draw_frame ? Should we have a swap_chain.size_has_changed() ? Something else that's already there that I'm missing?

[almarklein]

Hi, thanks for reporting this. I see what you mean! I guess I'll need to play with depth buffers myself and see how to best adjust the API. For now you may have to query the canvas size for changes, I suppose ...

[almarklein]

In visvis2 pygfx, we actually do have a depth buffer already. We simply do something like:

        physical_size = self._canvas.get_physical_size()  
        ...
        if self._depth_texture_size != physical_size:
            self._depth_texture_size = physical_size
            self._depth_texture = device.create_texture(
                size=(physical_size[0], physical_size[1], 1),
                usage=wgpu.TextureUsage.OUTPUT_ATTACHMENT,
                dimension="2d",
                format=wgpu.TextureFormat.depth32float,
            )

[almarklein]

Will consider how we can improve the API though. Not having to keep track of the current size would already help a lot, I think.

[marcdownie]

That code's working great. This is solely a question concerning the amount of pythonic-sugar you want to pour on this here (and a report of some surprise).

(In general, I'm finding that WGPU is sharing Vulkan's need for dependency tracking among high level "objects" as soon as anything can change: shader needs reloading? well, that means that the pipeline needs rebuilding etc. and so on.)

[almarklein]

Yeah, wgpu-py remains a rather low-level API, but I hope we can make it feel consistent and predictable. So feel free to report surprising behavior :)

[almarklein]

In #117, I add a size property to the texture view object. I considered something like swap_chain.has_size_changed, but then you have to consider some things: Compared to when? What if some other code calls that method too?

With the current change you can do something like:

        with swap_chain as texture_view:
            if texture_view.size != getattr(self, "_swap_chain_size", None):
                self._swap_chain_size = texture_view.size
                self._depth_texture = device.create_texture(
                    size=texture_view.size,
                    usage=wgpu.TextureUsage.OUTPUT_ATTACHMENT,
                    dimension="2d",
                    format=wgpu.TextureFormat.depth32float,
                )

So still some tracking of sizes, but it feels more natural because you're comparing directly with the render texture view.