#include <allreduce.h>

Inheritance diagram for AllReduceThreads:

Public Member Functions
	AllReduceThreads (AllReduceThreads *root, const size_t ptotal, const size_t pnode, bool quiet=false)

	AllReduceThreads (const size_t ptotal, const size_t pnode, bool quiet=false)

virtual	~AllReduceThreads ()

template<class T , void(*)(T &, const T &) f>
void	all_reduce (T *buffer, const size_t n)

Public Member Functions inherited from AllReduce
	AllReduce (size_t ptotal, const size_t pnode, bool pquiet=false)

virtual	~AllReduce ()=default

Private Attributes
AllReduceSync *	m_sync

bool	m_syncOwner

Additional Inherited Members
Public Attributes inherited from AllReduce
const size_t	total

const size_t	node

bool	quiet

Detailed Description

Definition at line 123 of file allreduce.h.

Constructor & Destructor Documentation

◆ AllReduceThreads() [1/2]

AllReduceThreads::AllReduceThreads	(	AllReduceThreads *	root,
		const size_t	ptotal,
		const size_t	pnode,
		bool	quiet = `false`
	)

Definition at line 53 of file allreduce_threads.cc.

     : AllReduce(ptotal, pnode, pquiet), m_sync(root->m_sync), m_syncOwner(false)
 {
 }

◆ AllReduceThreads() [2/2]

AllReduceThreads::AllReduceThreads	(	const size_t	ptotal,
		const size_t	pnode,
		bool	quiet = `false`
	)

Definition at line 58 of file allreduce_threads.cc.

     : AllReduce(ptotal, pnode, pquiet), m_sync(new AllReduceSync(ptotal)), m_syncOwner(true)
 {
 }

◆ ~AllReduceThreads()

AllReduceThreads::~AllReduceThreads ( )

virtual

Definition at line 63 of file allreduce_threads.cc.

References m_sync, and m_syncOwner.

 {
   if (m_syncOwner)
   {
     delete m_sync;
   }
 }

Member Function Documentation

◆ all_reduce()

template<class T , void(*)(T &, const T &) f>

void AllReduceThreads::all_reduce	(	T *	buffer,
		const size_t	n
	)

inline

Definition at line 137 of file allreduce.h.

References AllReduceSync::buffers, f, and AllReduceSync::waitForSynchronization().

   {  // register buffer
     T** buffers = (T**)m_sync->buffers;
     buffers[node] = buffer;
     m_sync->waitForSynchronization();
 
     size_t blockSize = n / total;
     size_t index;
     size_t end;
 
     if (blockSize == 0)
     {
       if (node < n)
       {
         index = node;
         end = node + 1;
       }
       else
       {  // more threads than bytes --> don't do any work
         index = end = 0;
       }
     }
     else
     {
       index = node * blockSize;
       end = node == total - 1 ? n : (node + 1) * blockSize;
     }
 
     for (; index < end; index++)
     {  // Perform transposed AllReduce to help data locallity
       T& first = buffers[0][index];
 
       for (size_t i = 1; i < total; i++) f(first, buffers[i][index]);
 
       // Broadcast back
       for (size_t i = 1; i < total; i++) buffers[i][index] = first;
     }
 
     m_sync->waitForSynchronization();
   }

Member Data Documentation

◆ m_sync

AllReduceSync* AllReduceThreads::m_sync

private

Definition at line 126 of file allreduce.h.

Referenced by ~AllReduceThreads().

◆ m_syncOwner

bool AllReduceThreads::m_syncOwner