Struct SubgroupOptimizer 

pub struct SubgroupOptimizer { /* private fields */ }

Vulkan subgroup operations optimizer.

Implementations

impl SubgroupOptimizer

pub const EMU_MAX: u32 = 256

Upper bound on the workgroup size supported by the emulated helpers.

The emulation scratch buffers are sized to this many f32/u32 elements; the n argument passed to each helper must not exceed it.

pub fn new(features: VulkanFeatures, config: VulkanOptimizationConfig) -> Self

Create a new subgroup optimizer.

pub fn optimize_shader(&self, shader_code: &str) -> String

Optimize shader code with subgroup operations.

Appends a set of subgroup_* WGSL helper functions to shader_code. The concrete implementation depends on the detected features:

• When VulkanFeatures::subgroup_arithmetic (respectively VulkanFeatures::subgroup_ballot) is true — i.e. the device was created with wgpu::Features::SUBGROUP — the helpers wrap the native WGSL subgroup built-ins (subgroupAdd, subgroupShuffle, subgroupBallot, …). These reduce/scan across the hardware subgroup (wave) and are the fast path.
• Otherwise the helpers fall back to a workgroup-shared-memory emulation built on workgroupBarrier(). The emulation reduces across the whole workgroup (not a subgroup) and is correct for a 1-D workgroup of up to Self::EMU_MAX invocations.

Every generated helper takes (…, lid: u32, n: u32) where lid is the caller’s local_invocation_index and n is the number of active invocations in the workgroup. The native helpers ignore lid/n; the emulated helpers use them to index the shared scratch buffers. This uniform signature lets shader authors switch paths without editing call sites.