@Kangz PTAL, thanks!

LGTM as mentioned offline, but we need other browsers to take a look too.

@Jiawei-Shao if you did the research, please provide the links to save us time verifying it.

I don't think we ever considered that format since it wasn't in the original list @JusSn added in #211. D16 isn't enough precision for many usecases so I think all of us kinda forgot about it to focus on the issues supporting depth-stencil formats.

One interesting note from the Metal tables: depth16unorm is always filterable, unlike depth32float. I wonder if that's something we can guarantee across apis.

@kvark sorry about that. Let me attach my investigations here.

Depth16Unorm can be supported on D3D feature level 11_0 (item No. 55) , all Metal feature sets (Page 9 of 11) and all Vulkan devices according to Vulkan SPEC (Table 48) .

In addition, it is a depth format that is required to support "Filter" on Metal and VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT on Vulkan (on D3D12 we can create a texture with the typeless format R16_TYPELESS and create an SRV on it with the format R16_UNORM and a DSV with D16_UNORM).

Sorry for my fault that I realized I referred a wrong document when writing this patch.

Although depth16unorm is available on all Metal devices, it requires macOS 10.12 and iOS 13, so we cannot bring this format into core WebGPU SPEC.

Good candidate for an extensions then.
Let's roll this back from the spec though... Edit: nvm, it wasn't merged

I'd like to reconsider adding this to core as the min iOS version is probably acceptable. So we need to decide whether we want to support macOS 10.11 (most importantly, hardware that didn't get updates to 10.12).

I would be surprised if we can otherwise-support that ten year old hardware. We should strongly consider leaving it behind.

There is a not insignificant portion of Chrome users on macOS on 10.11 (and even 10.10 which stays the min_deployement_target for Chrome on Mac) so we should carefully weight whether we want to drop WebGPU support for them. And D16 is imho not enough to do that, because it doesn't have a huge advantage compared to D32 that is in core WebGPU.

Do you trust those first-gen metal drivers? :) It would be nice to support them but I suppose I'm inclined to write them off as likely too hard to support.

I suppose we can add it later, but we have to make the call whether leaving that door open for now is worth whatever sacrifices it forces us to make to our spec. I would like to skate towards where the puck is going, here. (but I also don't feel strongly about it)

Discussed at the 2020-08-24 meeting.

Resolution: Approved! macOS 10.11 is EOL in Nov 2018 and not expected to receive Safari feature updates - would be limited to Chrome/Firefox.

We agreed this format is valuable, especially for memory usage on lower end devices, and copyability.

I don't remember why it was a new PR but this was fixed in #1026.