Depth clamp/clip improvements

[teoxoy]

Based on gpuweb/gpuweb#2100 (comment) & gpuweb/gpuweb#3638 (comment)

DX12

Clamps depth & has DepthClipEnable which corresponds to !unclippedDepth

wgpu/wgpu-hal/src/dx12/device.rs

Line 1345 in 17143c1

DepthClipEnable: BOOL::from(!desc.primitive.unclipped_depth),

TODO:

• Nothing.

Metal

Has setDepthClipMode available on: iOS 11.0+ & macOS 10.11+

Note: I think if setDepthClipMode is not available, Metal will still clip by default.

if setDepthClipMode is available:

• expose the DEPTH_CLIP_CONTROL feature
• if unclippedDepth { set MTLDepthClipModeClamp } else { set MTLDepthClipModeClip and clamp frag_depth in shader }

else:

• clamp frag_depth in shader

TODO:

• Update the following to check for versions instead and rename the capability to supports_set_depth_clip_mode.
  

  wgpu/wgpu-hal/src/metal/adapter.rs

  Lines 737 to 739 in 17143c1

  	//Depth clipping is supported on all macOS GPU families and iOS family 4 and later
  	supports_depth_clip_control: os_is_mac
  	|| device.supports_feature_set(MTLFeatureSet::iOS_GPUFamily4_v1),

• Implement frag_depth clamping in the shader

Vulkan

Has depthClamp & VK_EXT_depth_clip_enable

if both are available:

• expose the DEPTH_CLIP_CONTROL feature
• set depthClampEnable
• set depthClipEnable based on unclippedDepth

else if depthClamp is available:

• expose the DEPTH_CLIP_CONTROL feature
• if unclippedDepth { set depthClampEnable } else { clamp frag_depth in shader }

else:

• clamp frag_depth in shader

TODO:

• Shader clamping was implemented in [spv-out] option to clamp frag_depth naga#1431 but wgpu is not setting the flag and it should also clamp between viewport.minDepth and viewport.maxDepth (not 0 and 1)

• Implement the flow above; too many changes to list here (might be useful to undo part of Implement depth-clip-control using depthClamp #3892)

[JMS55]

I noticed this on a stress test of my renderer (VK backend). When rendering directional light shadow maps with a fragment shader to output unclamped depth, Nvidia NSight says I would get a ~3x speedup in rasterization if ZCULL was enabled, which would require removing the fragment shader and using the pipeline-level depth clamping control.