feat: Add 32-bit floating-point atomics (SHADER_FLOAT32_ATOMIC)

[AsherJingkongChen]

Connections

Resolves #1020

Description

The wgpu feature SHADER_FLOAT32_ATOMIC allows 32-bit floating point atomic operations on native backends:

• Metal ✅ (With version 3 or later + Apple 7+ or Mac 2 GPU) (Specification 2024-06-06)
• Vulkan ✅ (With the extension "VK_EXT_shader_atomic_float")
• DX12 ❌ (No, see issue #29 on HLSL specs)
• OpenGL ❌ (No, though there is extension "GL_EXT_shader_atomic_float")

The implemented or supported (natively) atomic operations by Metal and Vulkan are listed in the table (✅ is true, ❌ is false):

Function	Implemented?	Supported?	Implemented?	Supported?
f32	Metal	Metal	Vulkan	Vulkan
atomicLoad	✅	✅	✅	✅
atomicStore	✅	✅	✅	✅
atomicAdd	✅	✅	✅	✅
atomicSub	✅	✅	✅	❌
atomicMax	❌	❌	❌	✅
atomicMin	❌	❌	❌	✅
atomicAnd	❌	❌	❌	❌
atomicOr	❌	❌	❌	❌
atomicXor	❌	❌	❌	❌
atomicExchange	✅	✅	✅	✅
atomicCompareExchangeWeak	❌	✅	❌	❌

There are some floating-point atomic operations not implemented on some backends in this PR:

• Operations on pointers not in the storage address space: Metal 3 supports the atomic_float for device memory only. (Page 247)
• atomicAnd: This bitwise operation is designed for integer types, not floating-point ones.
• atomicOr: This bitwise operation is designed for integer types, not floating-point ones.
• atomicXor: This bitwise operation is designed for integer types, not floating-point ones.
• atomicMax: Metal does not support atomic_fetch_max_explicit (Page 250) on floating-point types.
• atomicMin: Metal does not support atomic_fetch_min_explicit (Page 250) on floating-point types.
• atomicCompareExchangeWeak: Vulkan only support OpAtomicCompareExchange on integer types.

There are some floating-point atomic operations implemented on all backends in this PR:

• Operations in the storage address space:
  • atomicLoad
  • atomicStore
  • atomicAdd
  • atomicSub
  • atomicExchange

There are some hacks in this PR implementation:

• atomicSub: Vulkan does not have subtraction for floating-point types (no OpAtomicFSub). We can simulate it by adding the negated value. They are mathematically equivalent since A - v == A + (-v).

Testing

• A naga snapshot atomicOps-float32 is added.
• A shader test numeric_builtins::float32_atomic is added.
• Relevant tests are passed on Apple M2 Pro (Metal 3.1+, Vulkan 1.2):

  • cargo xtask test snapshots float32_atomic
    

• Only 2 irrelevant tests failed on Apple M2 Pro (Vulkan 1.2):

  • cargo xtask test vertex_formats
    

Checklist

• Run cargo fmt.
• Run cargo clippy. If applicable, add:
  • --target wasm32-unknown-unknown
  • --target wasm32-unknown-emscripten
    • The crate winit does not compile on my device.
• Run cargo xtask test to run tests.
• Add change to CHANGELOG.md. See simple instructions inside file.
• Run cargo xtask validate to validate shaders.

[AsherJingkongChen]

Optional: Add fallback implementation for floating-point atomics

Description

Currently, the addition operation works on some recent chips like M1 Pro, NVIDIA RTX 4060, etc.
Without native supports, we can use more available operation, integer atomics, to simulate floating-point atomics.

By using atomic CAS loop, we can ensure the operation is atomic, but it is slower.

[teoxoy]

Sorry for taking so long to get to this.

[teoxoy]

Thank you for all the effort you put into thoroughly researching this and implementing it!

Besides the minor comments I left I have no other concerns.