matrix/webgpu_notes.txt

Learning WebGPU
	Request an adaptor from navigator.gpu
	Request a device from the adapter
		Devices have features and (numeric) limits
			So do adapters, but the device ones are the important ones
		Devices have a destroy method
	Create objects from the device
		Pipelines (optionally with async) and pipeline layouts
		Shader modules (programs)
		Textures and samplers
			You need a sampler to sample a texture in a shader
			Texture has a createView() method, these views get bound to bind groups and passes
			Has destroy() method
		Buffers
			Has destroy() method
		Bind groups and bind group layouts
			Bind groups are the interfaces defined between the CPU and a pipeline
		Command encoders
			Used to create passes, like a render pass
				begin, set, set, set, end, finish, submit to device queue
					It should be possible to reuse command buffers
		Render bundle encoders [What are these?]
		Query sets
			Useful for measuring stats that come from the GPU
			[Is this needed to detect errors?]


You do not need a canvas to webgpu.
You DO need a canvas to display what you're rendering
	Get canvas's "webgpu" context
	Configure context to point at the device
	Adjust it when you resize
		detect in the RAF when the presentation size isn't the canvas size
	Create a view of the context's current texture, and reference it in the render pass's color attachments
		[When does that currentTexture change?]

Textures were never resizable, you simply forgot
	Screen-size textures have to be destroyed and recreated
	Transfer the data over if you need to

Bind groups let you bind a bunch of resources at once
Render bundles let you reissue commands

You can only use up to FOUR bind groups on your laptop's device!
Here's an annoying gotcha: uniform buffers are like classic UBOs, and have esoteric ("std140") layout requirements!

Texture lookup in vertex shader
	You can only sample textures in fragment stages
	But you can use textureLoad anywhere to get a specific texel, no sampler needed