Support texture-compression-bc-sliced-3d in wgpu

[mehmetoguzderin]

Fixes #5750

Since it has been a while since I created a PR for wgpu, I might have overlooked some aspects. Your guidance would be appreciated. Thank you!

[teoxoy]

I do have a larger concern about this though; with the availability of ASTC/ETC formats on Vulkan gpuweb/gpuweb#3183 (comment) and it seems ETC formats on Metal as well (contrary to their docs) gpuweb/gpuweb#3183 (comment).

[mehmetoguzderin]

@teoxoy right, seems like further investigation is necessary as to only allow BCn or also ASTC (ETC2 seems harder).

[teoxoy]

The macOS runner seems to be failing a lot of tests (looks like the ones that actually execute something on the queue) right after we start executing clear_texture_compressed_bcn.

        SLOW [> 45.000s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_compressed_bcn
        SLOW [> 45.000s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_uncompressed
        SLOW [> 45.000s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_uncompressed_gles
        PASS [  60.282s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_uncompressed
        PASS [  61.647s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_compressed_bcn
...
 TERMINATING [> 90.000s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_uncompressed_gles
     TIMEOUT [  90.003s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_uncompressed_gles
...
     Summary [ 881.135s] 382 tests run: 360 passed (15 slow), 17 failed, 5 timed out, 0 skipped
     TIMEOUT [  90.003s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_uncompressed_gles
        FAIL [  60.095s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::compute_pass_ownership::compute_pass_resource_ownership
        FAIL [  60.116s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::occlusion_query::occlusion_query
        FAIL [  60.150s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::push_constants::partial_update
        FAIL [  60.115s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::regression::issue_3349::multi_stage_data_binding
     TIMEOUT [  90.008s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::regression::issue_4122::clear_buffer_range_respected
        FAIL [  60.143s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::scissor_tests::scissor_test_custom_rect
        FAIL [  60.111s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::scissor_tests::scissor_test_full_rect
        FAIL [  60.115s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader::data_builtins::pack4x_i8
        FAIL [  60.126s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader::data_builtins::pack4x_u8
        FAIL [  60.104s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader::data_builtins::unpack4x_i8
        FAIL [  60.102s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader::data_builtins::unpack4x_u8
        FAIL [  60.155s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader::numeric_builtins::numeric_builtins
     TIMEOUT [  90.010s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader::struct_layout::push_constant_input
     TIMEOUT [  90.010s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader::struct_layout::storage_input
        FAIL [  61.269s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader::struct_layout::uniform_input
        FAIL [  60.112s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader_primitive_index::draw
        FAIL [  60.126s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader_primitive_index::draw_indexed
        FAIL [  60.185s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::shader_view_format::reinterpret_srgb
     TIMEOUT [  90.010s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::vertex_indices::vertex_indices
        FAIL [  60.075s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::write_texture::write_texture_subset_2d
        FAIL [  60.112s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::write_texture::write_texture_subset_3d

For reference, this is how long the clear texture ones take on trunk:

        PASS [   0.347s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_uncompressed
        PASS [   0.730s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_compressed_bcn
        PASS [   2.736s] wgpu-test::wgpu-test [Executed] [Metal/Apple Paravirtual device/0] wgpu_test::clear_texture::clear_texture_uncompressed_gles

This is odd since the BC test eventually succeeds but there is some internal state getting messed up making the device unusable even across tests?

[mehmetoguzderin]

@teoxoy On my local device (just baseline M1), I get the following runtime for BCn clear tests:

PASS [   3.343s] wgpu-test::wgpu-test [Executed] [Metal/Apple M1/1] wgpu_test::clear_texture::clear_texture_compressed_bcn
PASS [   3.295s] wgpu-test::wgpu-test [Executed] [Vulkan/Apple M1/0] wgpu_test::clear_texture::clear_texture_compressed_bcn

So, unfortunately, I am unable to reproduce the CI's situation locally. BTW, I just checked the validation layers on my macOS environment, and they are all set to valid paths, which seem to work across many apps, including Vulkan Configurator itself, yet I still get the same [Vulkan/Apple M1/0] wgpu_test::shader::struct_layout::uniform_input failure (the only test fail on my system).

[teoxoy]

That sounds like another curious issue tracked by #5665.

[mehmetoguzderin]

@teoxoy apologies for the sudden ping during these busy times. As I was looking into migrating to extension, I got curious, although the tests pass with my local, non-paravirtualized M1 drivers, in case restructuring into extension keeps throwing errors for paravirtual driver, is it reasonable to tentatively blocklist this extension for Apple platforms? Is there a policy to allow partial forward progress with corner cases like this? All the best.

[teoxoy]

No worries. Yes there is a .expect_fail() method on TestParameters.

[mehmetoguzderin]

@teoxoy thank you for the insight! I will be conscious of it as I update the PR.

[mehmetoguzderin]

(Merging trunk made this branch go green with CI, interesting! Hopefully that'll hold valid by the time I update into extension)

[mehmetoguzderin]

@teoxoy sorry for the ping again. I finally got around to looking at migrating to the extension structure. First of all, apologies for the increase in surface area of this PR due to the delay in putting the support behind a feature. I tried my best to:

1. Comment that the BC base feature guarantees things,
2. Cover backends where possible (DX12, Metal, Vulkan, GL).

Furthermore, I noticed that the feature comment for BC was outdated; nowadays, all Apple9 iOS and iPadOS devices, as well as some Apple7 and Apple8 iPadOS devices, support BC as mobile alongside desktops.

However, the changes might have been less than optimal for a few aspects:

1. I set the flag to 1 << 11. I don't know if that's optimal for the project's future; there might be too few bits and too many features at some point, potentially.
2. I use the require_features and map the error to reject if the feature is not enabled. However, it is the only use of that kind for texture creation, and I am OK with updating if there's a better way.

I just enabled the SLICED_3D feature for testing, as BC simply guarantees it.

I'd appreciate it if you could take a look at your convenience. I'll try to improve in case there's a less-than-desirable aspect in the current shape of PR depending on review. On the good news, CI got green once again (which might let us put things in TODO if need be and go incremental)!

[teoxoy]

No problem, thanks for updating the PR!

For point 1, I think we should keep all the feature flags ordered, there is also this comment that needs updating:

wgpu/wgpu-types/src/lib.rs

Lines 418 to 420 in f6a3eef

	// Bits 11-19 available for webgpu features. Should you chose to use some of them for
	// for native features, don't forget to update `all_webgpu_mask` and `all_native_mask`
	// accordingly.

For point 2, I think it's fine as it is.

[mehmetoguzderin]

@teoxoy thank you very much for your kind review.

Good catch, sorry for the overlook! The latest commit addresses that and makes the flag counting in order as they appear in spec.

I appreciate the feedback, please let me know if there's anything further to refine this PR to make sure it is a reasonable commit for the project, and thank you and the team for all the efforts to keep wgpu a great project.

[teoxoy]

Thanks for your kind words and your contributions! :)

[mehmetoguzderin]

@teoxoy Likewise, thank you for the merge!