block.cuh File Reference

CUDA device computation warp-level primitives, i.e. More...

#include <kat/on_device/collaboration/warp.cuh>
#include <kat/on_device/shared_memory/basic.cuh>
#include <kat/on_device/common.cuh>
#include <kat/on_device/math.cuh>
#include <kat/on_device/grid_info.cuh>
#include <type_traits>

Functions
template<typename T , bool Synchronize = true>
KAT_FD void	kat::collaborative::block::share_per_warp_data (T datum, T *__restrict__ where_to_make_available, unsigned writing_lane_id)
	Share one element of type T for each warp with the entire block - using a single array in shared memory for all shared values. More...
template<typename T , bool Synchronize = true>
KAT_FD void	kat::collaborative::block::share_per_warp_data (T datum, T *__restrict__ where_to_make_available)
	A variant of share_per_warp_data , with the writing lane index being decided dynamically in each lane based on who's actually active.
KAT_FD void	kat::collaborative::block::barrier ()
template<typename T , bool Synchronize = true, unsigned Dimensionality = 3>
KAT_FD T	kat::collaborative::block::get_from_thread (const T &value, kat::position_t source_thread_position)
	have all block threads obtain a value held by just one of the threads (and likely not otherwise easily accessible to the rest of the block's threads). More...
template<typename T , bool Synchronize = true>
KAT_FD T	kat::collaborative::block::get_from_first_thread (T &&value)
	have all block threads obtain a value held by the first thread in the block (and likely not otherwise easily accessible to the rest of the block's threads). More...
template<typename Function , typename Size = size_t>
KAT_FD void	kat::linear_grid::collaborative::block::at_block_stride (Size length, const Function &f)
	Have all threads in (one/some/all) blocks perform some action over the linear range of 0..length-1 - the same range for each block. More...
template<typename T , bool Synchronize = true>
KAT_FD void	kat::linear_grid::collaborative::block::share_per_warp_data (T datum, T *__restrict__ where_to_make_available, unsigned writing_lane_id)
	Share one element of type T for each warp with the entire block - using a single array in shared memory for all shared values. More...
template<typename T , bool Synchronize = true>
KAT_FD void	kat::linear_grid::collaborative::block::share_per_warp_data (T datum, T *__restrict__ where_to_make_available)
	A variant of share_per_warp_data , with the writing lane index being decided dynamically in each lane based on who's actually active.
KAT_FD void	kat::linear_grid::collaborative::block::barrier ()
template<typename T , bool Synchronize = true>
KAT_FD T	kat::linear_grid::collaborative::block::get_from_thread (T &&value, unsigned source_thread_id)
	have all block threads obtain a value held by just one of the threads (and likely not otherwise easily accessible to the rest of the block's threads). More...
template<typename T , bool Synchronize = true>
KAT_FD T	kat::linear_grid::collaborative::block::get_from_first_thread (T &&value)
	have all block threads obtain a value held by the first thread in the block (and likely not otherwise easily accessible to the rest of the block's threads). More...

Detailed Description

CUDA device computation warp-level primitives, i.e.

those involving interaction of many/all of each blocks's lanes, but no inter-block interaction.

Todo:

Some of these assume linear grids, others do not - sort them out

Function Documentation

at_block_stride()

template<typename Function , typename Size = size_t>

KAT_FD void kat::linear_grid::collaborative::block::at_block_stride	(	Size	length,
		const Function &	f
	)

Have all threads in (one/some/all) blocks perform some action over the linear range of 0..length-1 - the same range for each block.

Note

This function semi-assumes the block size is a multiple of the warp size; otherwise, it should still work, but - it'll be slow(ish).

Parameters

length	The length of the range (of integers) on which to act handle (serially)
f	The callable to execute for each element of the sequence.

get_from_first_thread() [1/2]

template<typename T , bool Synchronize = true>

KAT_FD T kat::collaborative::block::get_from_first_thread

(

T &&

value

)

have all block threads obtain a value held by the first thread in the block (and likely not otherwise easily accessible to the rest of the block's threads).

Note

uses shared memory for "broadcasting" the value

get_from_first_thread() [2/2]

template<typename T , bool Synchronize = true>

KAT_FD T kat::linear_grid::collaborative::block::get_from_first_thread

(

T &&

value

)

have all block threads obtain a value held by the first thread in the block (and likely not otherwise easily accessible to the rest of the block's threads).

Note

uses shared memory for "broadcasting" the value

get_from_thread() [1/2]

template<typename T , bool Synchronize = true, unsigned Dimensionality = 3>

KAT_FD T kat::collaborative::block::get_from_thread	(	const T &	value,
		kat::position_t	source_thread_position
	)

have all block threads obtain a value held by just one of the threads (and likely not otherwise easily accessible to the rest of the block's threads).

Note

uses shared memory for the "broadcast" by the thread holding the relevant value

get_from_thread() [2/2]

template<typename T , bool Synchronize = true>

KAT_FD T kat::linear_grid::collaborative::block::get_from_thread	(	T &&	value,
		unsigned	source_thread_id
	)

have all block threads obtain a value held by just one of the threads (and likely not otherwise easily accessible to the rest of the block's threads).

Note

uses shared memory for the "broadcast" by the thread holding the relevant value

share_per_warp_data() [1/2]

template<typename T , bool Synchronize = true>

KAT_FD void kat::collaborative::block::share_per_warp_data	(	T	datum,
		T *__restrict__	where_to_make_available,
		unsigned	writing_lane_id
	)

Share one element of type T for each warp with the entire block - using a single array in shared memory for all shared values.

Parameters

datum	a warp-specific (but not thread-specific) piece of data, one for each warp, which is to be shared with the whole block
where_to_make_available	the various warp-specific data will be stored here by warp index
writing_lane_id	which lane in each warp should perform write operations

share_per_warp_data() [2/2]

template<typename T , bool Synchronize = true>

KAT_FD void kat::linear_grid::collaborative::block::share_per_warp_data	(	T	datum,
		T *__restrict__	where_to_make_available,
		unsigned	writing_lane_id
	)

Share one element of type T for each warp with the entire block - using a single array in shared memory for all shared values.

Parameters

datum	a warp-specific (but not thread-specific) piece of data, one for each warp, which is to be shared with the whole block
where_to_make_available	the various warp-specific data will be stored here by warp index
writing_lane_id	which lane in each warp should perform write operations

Note

if different threads in a warp have different values, behavior is not guaranteed.