GridModule

Data Structures
class	cub::GridBarrier
	GridBarrier implements a software global barrier among thread blocks within a CUDA grid. More...
class	cub::GridBarrierLifetime
	GridBarrierLifetime extends GridBarrier to provide lifetime management of the temporary device storage needed for cooperation. More...
struct	cub::GridEvenShare< OffsetT >
	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. More...
class	cub::GridQueue< OffsetT >
	GridQueue is a descriptor utility for dynamic queue management. More...

Enumerations
enum	cub::GridMappingStrategy { cub::GRID_MAPPING_RAKE, cub::GRID_MAPPING_STRIP_MINE, cub::GRID_MAPPING_DYNAMIC }
	cub::GridMappingStrategy enumerates alternative strategies for mapping constant-sized tiles of device-wide data onto a grid of CUDA thread blocks. More...

Functions
template<typename OffsetT >
__global__ void	cub::FillAndResetDrainKernel (GridQueue< OffsetT > grid_queue, OffsetT num_items)

Detailed Description

Enumeration Type Documentation

GridMappingStrategy

enum cub::GridMappingStrategy

cub::GridMappingStrategy enumerates alternative strategies for mapping constant-sized tiles of device-wide data onto a grid of CUDA thread blocks.

Enumerator
GRID_MAPPING_RAKE	An a "raking" access pattern in which each thread block is assigned a consecutive sequence of input tiles. Overview The input is evenly partitioned into p segments, where p is constant and corresponds loosely to the number of thread blocks that may actively reside on the target device. Each segment is comprised of consecutive tiles, where a tile is a small, constant-sized unit of input to be processed to completion before the thread block terminates or obtains more work. The kernel invokes p thread blocks, each of which iteratively consumes a segment of n/p elements in tile-size increments.
GRID_MAPPING_STRIP_MINE	An a "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. Overview The input is evenly partitioned into p sets, where p is constant and corresponds loosely to the number of thread blocks that may actively reside on the target device. Each set is comprised of data tiles separated by stride tiles, where a tile is a small, constant-sized unit of input to be processed to completion before the thread block terminates or obtains more work. The kernel invokes p thread blocks, each of which iteratively consumes a segment of n/p elements in tile-size increments.
GRID_MAPPING_DYNAMIC	A dynamic "queue-based" strategy for assigning input tiles to thread blocks. Overview The input is treated as a queue to be dynamically consumed by a grid of thread blocks. Work is atomically dequeued in tiles, where a tile is a unit of input to be processed to completion before the thread block terminates or obtains more work. The grid size p is constant, loosely corresponding to the number of thread blocks that may actively reside on the target device.

Definition at line 59 of file grid_mapping.cuh.

Function Documentation

FillAndResetDrainKernel()

template<typename OffsetT >

__global__ void cub::FillAndResetDrainKernel	(	GridQueue< OffsetT >	grid_queue,
		OffsetT	num_items
	)

Reset grid queue (call with 1 block of 1 thread)

Definition at line 203 of file grid_queue.cuh.