Struct WarpState 

pub struct WarpState { /* private fields */ }

Per-warp shared state used to emulate warp-level operations.

In a real GPU each warp executes in lock-step and has hardware support for cross-lane communication. On the CPU we emulate this by having all threads in a “warp” share a WarpState and synchronise explicitly via barriers.

Implementations

impl WarpState

pub fn new() -> Self

Create a new warp state with all lanes active.

pub fn set_lane_active(&self, lane_id: u32)

Set a lane as active.

pub fn set_lane_inactive(&self, lane_id: u32)

Set a lane as inactive.

pub fn active_mask(&self) -> u32

Get the current active mask.

pub fn is_lane_active(&self, lane_id: u32) -> bool

Returns true if the specified lane is currently active.

pub fn shuffle(&self, lane_id: u32, value: u32, src_lane: u32) -> u32

Emulate __shfl_sync: read the value from src_lane.

The caller (at lane_id) first writes its own value, then after a barrier reads from src_lane. In a single-threaded emulation context, the caller can pre-populate all lanes and then read.

Arguments

• lane_id - The calling thread’s lane within the warp (0..31)
• value - The value this lane contributes
• src_lane - The lane to read from

Returns the value from src_lane, or this lane’s own value if src_lane is out of range.

pub fn shuffle_xor(&self, lane_id: u32, value: u32, lane_mask: u32) -> u32

Emulate __shfl_xor_sync: read from lane_id ^ lane_mask.

pub fn shuffle_up(&self, lane_id: u32, value: u32, delta: u32) -> u32

Emulate __shfl_up_sync: read from lane_id - delta. If the source lane would be negative, return the caller’s own value.

pub fn shuffle_down(&self, lane_id: u32, value: u32, delta: u32) -> u32

Emulate __shfl_down_sync: read from lane_id + delta. If the source lane would be >= WARP_SIZE, return the caller’s own value.

pub fn shuffle_f32(&self, lane_id: u32, value: f32, src_lane: u32) -> f32

Shuffle an f32 value (reinterpret bits through u32).

pub fn shuffle_xor_f32(&self, lane_id: u32, value: f32, lane_mask: u32) -> f32

Shuffle XOR with f32.

pub fn shuffle_up_f32(&self, lane_id: u32, value: f32, delta: u32) -> f32

Shuffle up with f32.

pub fn shuffle_down_f32(&self, lane_id: u32, value: f32, delta: u32) -> f32

Shuffle down with f32.

pub fn vote_all(&self, lane_id: u32, predicate: bool) -> bool

Emulate __all_sync: returns true if all active lanes have predicate == true.

pub fn vote_any(&self, lane_id: u32, predicate: bool) -> bool

Emulate __any_sync: returns true if any active lane has predicate == true.

pub fn ballot(&self, lane_id: u32, predicate: bool) -> u32

Emulate __ballot_sync: returns a bitmask where bit i is set if lane i is active and its predicate is true.

pub fn reduce_sum_f32(&self, lane_id: u32, value: f32) -> f32

Warp-level sum reduction using shuffle_down (butterfly pattern).

Assumes all 32 lanes call this with their value. Returns the sum at lane 0; other lanes get a partial result.

pub fn reduce_max_f32(&self, lane_id: u32, value: f32) -> f32

Warp-level max reduction.

pub fn reduce_min_f32(&self, lane_id: u32, value: f32) -> f32

Warp-level min reduction.

pub fn popc_ballot(&self, lane_id: u32, predicate: bool) -> u32

Count the number of active lanes with a true predicate (popcount of ballot).

Trait Implementations

impl Default for WarpState

fn default() -> Self

Returns the “default value” for a type.