[naga glsl-out] Differentiate between support for std140 and std430

[cloone8]

Description

When outputting code for older OpenGL versions, I noticed that Naga does not differentiate between not having support for explicit bindings/locations and outputting Uniform memory layouts. Furthermore, if no global binding was given, no attempt at writing an explicit memory layout is made at all. This PR attempts to improve that by making a distinction between:

• Supporting explicit locations
• Supporting std140 layout
• Supporting std430 layout

...And then outputting the memory layout if at all possible, even if no explicit location is supported/exists.

The result is that globals that have no explicitly given bindings now get a memory layout anyway.

Testing

I ran the whole test suite with cargo xtask test, of course, which showed a small amount of weird failures on my own system that were present on previous commits too. I also examined and validated (with glslang) the resulting changed shaders.

Furthermore, I reasoned (with my admittedly limited knowledge) that this change should be fine, because:

• Naga emits these layouts anyway for newer GLSL versions
• The previous versions with the default layouts required the use of reflection, which should still work perfectly with even with an explicit memory layout

But I admit I don't think I grasp the complete scope of how these changes could break things. Specifically, I saw some shader snapshots that were testing padding related things (naga/tests/out/glsl/wgsl-struct-layout.needs_padding_comp.Compute.glsl and naga/tests/out/glsl/wgsl-struct-layout.no_padding_comp.Compute.glsl) that worry me slightly, as changing memory layouts could break things if padding was manually inserted using different rules. In theory, however, it should be fine as the previous default was std140 anyway (for Vulkan, I believe) and shared for OpenGL, which leaves layout fully up to the driver and thus makes little sense to have padding inserted manually.

If these changes turn out to break other code, I'd be happy to take another look if someone points me in the right direction!

Checklist

• Run cargo fmt.
• Run taplo format.
• Run cargo clippy --tests. If applicable, add:
  • --target wasm32-unknown-unknown
• Run cargo xtask test to run tests.
• If this contains user-facing changes, add a CHANGELOG.md entry.

NOTE: I added the changelog entry to "new features" because I didn't know how you all differentiate between new features and changes (or if this is seen as a bug), but feel free to shuffle that around if you wish.

[teoxoy]

What is the specific issue that this change solves? As far as I know uniform buffers have std140 layout by default.

[cloone8]

What is the specific issue that this change solves? As far as I know uniform buffers have std140 layout by default.

As far as I can tell, this is true when the GLSL shader is used in Vulkan:

From https://github.com/KhronosGroup/GLSL/blob/d8fe3068fd869153ae2731d7aba907c6ea7534a0/extensions/khr/GL_KHR_vulkan_glsl.txt#L67

The following features are removed:
...

• shared and packed block layouts

From https://github.com/KhronosGroup/GLSL/blob/d8fe3068fd869153ae2731d7aba907c6ea7534a0/extensions/khr/GL_KHR_vulkan_glsl.txt#L693

Mapping of layouts
...
-> not shared, but std140 or std430

But not in OpenGL, from the version 4.6 core spec at https://registry.khronos.org/OpenGL/specs/gl/glspec46.core.pdf

Section 7.6.2.2 (page 146):

By default, uniforms contained within a uniform block are extracted from buffer
storage in an implementation-dependent manner. Applications may query the off-
sets assigned to uniforms inside uniform blocks with query functions provided by
the GL.

...And from the GLSL 4.60 spec at https://registry.khronos.org/OpenGL/specs/gl/GLSLangSpec.4.60.pdf

Section 4.4.5 (page 89):

The initial state of compilation when generating SPIR-V is as if the following were declared:
layout(std140, column_major) uniform;
layout(std430, column_major) buffer;

and later

The initial state of compilation when not generating SPIR-V is as if the following were declared:
layout(shared, column_major) uniform;
layout(shared, column_major) buffer;

Peeking through some of the old OpenGL specs, the default seems to have always been shared as far as I can tell.

[teoxoy]

I see, that makes sense, thanks for quoting the specs!

[teoxoy]

Could we error here instead?

[teoxoy]

Looks great, thank you!