Crate warp_types 

Warp Types: Type-safe GPU warp programming via linear typestate.

Prevents warp divergence bugs at compile time using linear typestate. A diverged warp literally cannot call shuffle — the method doesn’t exist.

Core Idea

use warp_types::*;

let warp: Warp<All> = Warp::kernel_entry();
let data = data::PerLane::new(42i32);

// OK: shuffle on full warp
let _shuffled = warp.shuffle_xor(data, 1);

// After diverge, shuffle is gone from the type:
let (evens, odds) = warp.diverge_even_odd();
// evens.shuffle_xor(data, 1);  // COMPILE ERROR — method not found
let merged: Warp<All> = merge(evens, odds);

Module Overview

• active_set — Lane subset types (All, Even, Odd, …) and complement proofs
• warp — Warp<S> type parameterized by active set
• data — Value categories: PerLane<T>, Uniform<T>, SingleLane<T, N>
• diverge — Split warps by predicate (produces complementary sub-warps)
• merge — Rejoin complementary sub-warps (compile-time verified)
• shuffle — Shuffle/ballot/reduce (restricted to Warp<All>) + permutation algebra
• fence — Fence-divergence interactions (§5.6) — type-state write tracking
• block — Block-level: shared memory ownership, inter-block sessions, reductions
• proof — Soundness proof sketch (progress + preservation)
• platform — CPU/GPU platform trait for dual-mode algorithms
• gradual — DynWarp ↔ Warp<S> bridge for gradual typing (§9.4)
• gpu — GPU intrinsics for nvptx64 and amdgpu targets
• cub — Typed CUB-equivalent warp primitives (reduce, scan, broadcast)
• sort — Typed warp-level bitonic sort
• tile — Cooperative Groups: thread block tiles with typed shuffle safety
• dynamic — Data-dependent divergence with structural complement guarantees
• simwarp — Multi-lane warp simulator with real shuffle semantics (testing)

Re-exports

pub use active_set::ActiveSet;

pub use active_set::All;

pub use active_set::CanDiverge;

pub use active_set::ComplementOf;

pub use active_set::ComplementWithin;

pub use active_set::Empty;

pub use active_set::Even;

pub use active_set::EvenHigh;

pub use active_set::EvenLow;

pub use active_set::HighHalf;

pub use active_set::Lane0;

pub use active_set::LowHalf;

pub use active_set::NotLane0;

pub use active_set::Odd;

pub use active_set::OddHigh;

pub use active_set::OddLow;

pub use block::BlockId;

pub use block::SharedRegion;

pub use block::ThreadId;

pub use data::LaneId;

pub use data::PerLane;

pub use data::Role;

pub use data::SingleLane;

pub use data::Uniform;

pub use data::WarpId;

pub use dynamic::DynDiverge;

pub use fence::Fenced;

pub use fence::FullWrite;

pub use fence::GlobalRegion;

pub use fence::PartialWrite;

pub use fence::Unwritten;

pub use fence::WriteState;

pub use gradual::DynWarp;

pub use merge::merge;

pub use merge::merge_within;

pub use platform::CpuSimd;

pub use platform::GpuWarp32;

pub use platform::GpuWarp64;

pub use platform::Platform;

pub use platform::SimdVector;

pub use shuffle::BallotResult;

pub use shuffle::Compose;

pub use shuffle::HasDual;

pub use shuffle::Identity;

pub use shuffle::Permutation;

pub use shuffle::RotateDown;

pub use shuffle::RotateUp;

pub use shuffle::ShuffleSafe;

pub use shuffle::Xor;

pub use tile::Tile;

pub use warp::Warp;

Modules

active_set

Active set types: compile-time lane subset tracking.

block

Block-level types: shared memory ownership and inter-block sessions.

cub

Typed CUB-equivalent warp primitives.

data

GPU data types: uniform vs per-lane value distinction.

diverge

Diverge operations: splitting a warp by predicate.

dynamic

Data-dependent divergence with structural complement guarantees.

fence

Fence-divergence interaction types (§5.6).

gpu

GPU intrinsics for nvptx64 and amdgpu targets.

gradual

Gradual typing: DynWarp ↔ Warp<S> bridge.

merge

Merge operations: reconverging diverged warps.

platform

Platform abstraction for CPU/GPU unified targeting

prelude

Convenience prelude — import everything needed for typical usage.

shuffle

Shuffle operations and permutation algebra.

simwarp

SimWarp: multi-lane warp simulator with real shuffle semantics.

sort

Typed warp-level bitonic sort.

tile

Cooperative Groups: thread block tiles with typed shuffle safety.

warp

The core Warp<S> type: a warp with compile-time tracked active lanes.

Constants

WARP_SIZE

Number of lanes per warp/wavefront.

Traits

GpuValue

Marker trait for types that can live in GPU registers.

Functions

zero_overhead_butterfly

Zero-overhead benchmark: 5 shuffle permutations + butterfly reduction.

zero_overhead_diverge_merge

Diverge-merge round trip: the type system’s core mechanism.

Attribute Macros

warp_kernel

Mark a function as a GPU kernel entry point.