Module sync 

Synchronization primitives (barriers, atomics). CPU- and CUDA-compatible wrappers for spirv_std::arch functions.

On target_arch = "spirv", these delegate to the real SPIR-V intrinsics. On target_arch = "nvptx64", these use LLVM NVVM intrinsics for CUDA. On the CPU, they provide functional equivalents using std::sync primitives.

Shader code should import from this module instead of spirv_std::arch directly to enable CPU and CUDA execution.

Functions

atomic_add_f32

Atomically adds a float value to the f32 stored (as u32 bits) at ptr.

atomic_add_i32

Atomically adds value to *ptr and returns the old value.

atomic_add_u32

Atomically adds value to *ptr (u32) and returns the old value.

atomic_add_u32_workgroup

Atomically adds value to *ptr (workgroup scope) and returns the old value.

atomic_and_u32

Atomically ANDs *ptr with value and returns the old value.

atomic_compare_exchange_u32

Atomically compares *ptr with comparator and, if equal, replaces with value. Returns the old value at *ptr.

atomic_exchange_u32

Atomically exchanges *ptr with value and returns the old value.

atomic_load_u32

Atomically loads the value at *ptr.

atomic_load_u32_shared

Atomically loads the value at *ptr from a shared reference.

atomic_load_u32_workgroup

Atomically loads *ptr (workgroup scope).

atomic_max_u32

Atomically computes max(*ptr, value) and returns the old value.

atomic_max_u32_workgroup

Atomically computes max(*ptr, value) (workgroup scope) and returns the old value.

atomic_min_u32

Atomically computes min(*ptr, value) and returns the old value.

atomic_or_u32

Atomically ORs *ptr with value and returns the old value.

atomic_store_u32_workgroup

Atomically stores value to *ptr (workgroup scope).

control_barrier

Control barrier with explicit execution scope, memory scope, and semantics.

subgroup_f_add

Subgroup (wave/warp) floating-point add reduction.

subgroup_f_max

Subgroup (wave/warp) floating-point max reduction.

workgroup_memory_barrier_with_group_sync

Workgroup memory barrier with group synchronization.