mycuda_reduce Namespace Reference

Reduction operators and kernels. More...

Data Structures
class	SUM
class	MAX
class	MIN

Functions
template<typename T , typename R >
__device__ void	reduce_dev (T *x, int n, R op)
	Reduces a vector x of length n. More...
template<typename T , typename R >
__global__ void	reduce (T *xsum, T *x, int n, int stride, R op)
	Reduces an M x N matrix. Given a K x N grid of blocks, returns a K x N array of sums. More...

Detailed Description

Reduction operators and kernels.

Function Documentation

template<typename T , typename R >

__global__ void mycuda_reduce::reduce	(	T *	xsum,
		T *	x,
		int	n,
		int	stride,
		R	op
	)

Reduces an M x N matrix. Given a K x N grid of blocks, returns a K x N array of sums.

Notes

• Uses dynamic allocation for shared memory. Allocation size is determined by configuration of kernel launch.

Requires

• blocksize >=32
• shared_memory_size = 2*blocksize*sizeof(T)

Definition at line 111 of file mycuda_reduce.h.

template<typename T , typename R >

__device__ void mycuda_reduce::reduce_dev	(	T *	x,
		int	n,
		R	op
	)

Reduces a vector x of length n.

Notes

• x should be stored in shared memory for best performance.
• sum is computed in place and returned in x[0].
• reduction operator implemented as a functor class.

Requires

• n <= 2*blocksize
• blocksize >= 32

Definition at line 73 of file mycuda_reduce.h.