cub::GridEvenShare
Defined in cub/grid/grid_even_share.cuh
-
template<typename OffsetT>
struct GridEvenShare GridEvenShare is a descriptor utility for distributing input among CUDA thread blocks in an “even-share” fashion.
Each thread block gets roughly the same number of input tiles.
- Overview
Each thread block is assigned a consecutive sequence of input tiles. To help preserve alignment and eliminate the overhead of guarded loads for all but the last thread block, to GridEvenShare assigns one of three different amounts of work to a given thread block: “big”, “normal”, or “last”. The “big” workloads are one scheduling grain larger than “normal”. The “last” work unit for the last thread block may be partially-full if the input is not an even multiple of the scheduling grain size.
Before invoking a child grid, a parent thread will typically construct an instance of GridEvenShare. The instance can be passed to child thread blocks which can initialize their per-thread block offsets using
BlockInit().
Public Functions
-
inline GridEvenShare()
Constructor.
-
inline void DispatchInit(OffsetT num_items_, int max_grid_size, int tile_items)
Dispatch initializer.
To be called prior prior to kernel launch.
- Parameters
num_items_ – Total number of input items
max_grid_size – Maximum grid size allowable (actual grid size may be less if not warranted by the the number of input items)
tile_items – Number of data items per input tile
-
template<int TILE_ITEMS>
inline void BlockInit(int block_id, Int2Type<GRID_MAPPING_RAKE>) Initializes ranges for the specified thread block index.
Specialized for a “raking” access pattern in which each thread block is assigned a consecutive sequence of input tiles.
-
template<int TILE_ITEMS>
inline void BlockInit(int block_id, Int2Type<GRID_MAPPING_STRIP_MINE>) Block-initialization, specialized for a “raking” access pattern in which each thread block is assigned a consecutive sequence of input tiles.
-
template<int TILE_ITEMS, GridMappingStrategy STRATEGY>
inline void BlockInit() Block-initialization, specialized for “strip mining” access pattern in which the input tiles assigned to each thread block are separated by a stride equal to the the extent of the grid.
-
template<int TILE_ITEMS>
inline void BlockInit(OffsetT block_offset, OffsetT block_end) Block-initialization, specialized for a “raking” access pattern in which each thread block is assigned a consecutive sequence of input tiles.
- Parameters
block_offset – [in] Threadblock begin offset (inclusive)
block_end – [in] Threadblock end offset (exclusive)