A class for determining thread frontiers for all basic blocks. More...

#include <ThreadFrontierAnalysis.h>

Inheritance diagram for analysis::ThreadFrontierAnalysis:

Collaboration diagram for analysis::ThreadFrontierAnalysis:

Public Types
typedef ir::ControlFlowGraph::ConstBlockPointerVector	BlockVector
typedef ir::ControlFlowGraph::const_iterator	const_iterator
typedef unsigned int	Priority
typedef std::unordered_map < const_iterator, Priority >	PriorityMap
Public Member Functions
	ThreadFrontierAnalysis ()
	Create the analysis.
void	analyze (ir::IRKernel &kernel)
	Computes an up to date set of thread frontiers.
BlockVector	getThreadFrontier (const_iterator block) const
	Get the blocks in the thread frontier of a specified block.
Priority	getPriority (const_iterator block) const
	Get the scheduling priorty of a specified block.

Detailed Description

A class for determining thread frontiers for all basic blocks.

Thread frontiers defines a scheduling order of threads and identifies re-convergence points. The approach has two main steps: 1) Assign a priority order to basic blocks. 2) For each basic block, determine all other possible blocks where stalled threads may be waiting.

The intersection of this set and a branch target's predecessors need to be checked for re-convergence. The algorithm used here is taken directly from [1].

[1] - "SIMD Reconvergence at Thread Frontiers" by Diamos et al.