Add VK_EXT_image_2d_view_of_3d support

[squidbus]

Continuation of #2332 as it has not been active for a few months now, completing support for the imageView2DOn3DImage portability feature and VK_EXT_image_2d_view_of_3d.

Differences from that PR:

• Remaining review comments have been addressed.
• Support for sampler2DViewOf3D has been fixed. It seems aliased 2D texture does not synchronize properly with changes to 3D texture without write usage, which is why only storage tests were passing.
• Support for 2D array views has been fixed. Instead of depth, arrayLength needs to be set to the layer count.
• Minor formatting fixes.

Ran CTS with MVK_CONFIG_USE_MTLHEAP=1 as it is required for this.

Passes dEQP-VK.pipeline.monolithic.image_2d_view_3d_image.*:

Test run totals:
  Passed:        24/48 (50.0%)
  Failed:        0/48 (0.0%)
  Not supported: 24/48 (50.0%)
  Warnings:      0/48 (0.0%)
  Waived:        0/48 (0.0%)

And now supports more tests from dEQP-VK.pipeline.monolithic.render_to_image.*, to demonstrate imageView2DOn3DImage portability feature support:

• Before:

Test run totals:
  Passed:        840/1245 (67.5%)
  Failed:        0/1245 (0.0%)
  Not supported: 405/1245 (32.5%)
  Warnings:      0/1245 (0.0%)
  Waived:        0/1245 (0.0%)

• After:

Test run totals:
  Passed:        927/1245 (74.5%)
  Failed:        0/1245 (0.0%)
  Not supported: 318/1245 (25.5%)
  Warnings:      0/1245 (0.0%)
  Waived:        0/1245 (0.0%)

Additionally, I modified the CTS tests for image_2d_view_3d_image to use 2D arrays, and verified they are working as expected.

[CLAassistant]

All committers have signed the CLA.

[K0bin]

So you implement this by aliasing a 3D texture with a 2D one. Does Metal actually guarantee that this works or is it more in the realm of UB that just happens to work?

[squidbus]

So you implement this by aliasing a 3D texture with a 2D one. Does Metal actually guarantee that this works or is it more in the realm of UB that just happens to work?

As I understand aliasing image memory is valid in Metal but this relies on the underlying memory layout which is not defined in spec. Please refer to the original PR as I did not author the basic approach.

[billhollings]

Thanks for taking up the mantle and getting this done!

LGTM.

Please add VK_EXT_image_2d_view_of_3d to MoltenVK_Runtime_UserGuide.md with a note (see other entries) that MVK_CONFIG_USE_MTLHEAP must be enabled.

Other than that, just a couple of minor nits.

Does this addition now make it time to enable MVK_CONFIG_USE_MTLHEAP by default? Once this is pulled in, I'll run a full CTS to see if there are any obvious issues with that, but I wanted to get your input from the experience of working with this.

[squidbus]

Comments have been addressed. Regarding switching MVK_CONFIG_USE_MTLHEAP on by default, I personally have not had any issues with it on M-series GPUs. Related to that, I believe there are some places where heap usage is always disabled (e.g. depth-stencil images) that were disabled back on Intel GPUs, which may be worth re-testing on newer drivers / M-series GPUs.

[billhollings]

Thanks! LGTM now.