Extend Vulkan Depth Test Overlay

[Zorro666]

Description

Vulkan Replay support for Issue #2858

This is a PR in a sequence of PRs which all together complete the implementation of the feature.
The next PR will extend the automated overlay tests GL_Overlay_Test, D3D11_Overlay_Test, D3D12_Overlay_Test, VK_Overlay_Test.

The feature enables supporting shader exported depth for the Depth Test overlay by replaying the draw using the pixel shader from the capture instead of a replacement shader. Depth test passing pixels are identified by using the stencil buffer to generate a stencil mask.

If the depth buffer used for the draw does not have a stencil buffer then a new depth buffer is created with a depth+stencil format and a fullscreen pass is used to copy depth from the original depth buffer to the newly created depth+stencil buffer.

Tested on Windows using changes to Vk_Overlay_Test

The change makes the pixel shader write out depth 0.0 for a small rectangle just bottom-left of the centre

Before (using v1.29) : the rectangle of output depth 0.0 is shown as failing the depth test

Depth Buffer format D24_S8

Depth Buffer format D32F_S8

Depth Buffer format D16 MSAA

Depth Buffer format D24 MSAA

Depth Buffer format D32F MSAA

After : the rectangle of output depth 0.0 is shown as passing the depth test

Depth Buffer format D24_S8

Depth Buffer format D32F_S8

Depth Buffer format D16 MSAA

Depth Buffer format D24 MSAA

Buffer format D32F MSAA

[baldurk]

I think seeing as we don't have that many ConciseGraphicsPipelines that we create for internal pipelines I would rather update the struct to handle setting the new stencil elements that you need to change here. Otherwise this could quickly become unwieldy and annoying to work with.

Since stencil state is one of those things with many theoretical combinations but few in practice, you could replace the current VkStencilOp stencilOp with a StencilMode stencilOperations where StencilMode is an enum with the different configurations we want - I think currently just 'keep' and 'overwrite with 0'. You could add 'test equal to 1 and keep'. If that becomes difficult to sum up clearly in one enum you could split it into two enums for the test mode and the update mode.

[baldurk]

Looking at the changes here now I misspoke before, not every pipeline is explicitly writing zero as e.g. the FillWithDiscardPattern has a dynamic stencil ref, and so it's replacing with an arbitrary value (which it uses to fill the discard pattern in stencil)

I think it's better to keep the stencil op that was used here, REPLACE, rather than specialising to zero unnecessarily. It has the same effect but is more general.

[Zorro666]

Thank you - good spot. I have kept WRITE_ZERO also even though it can be achieved with REPLACE. It felt more descriptive in its intent.