Hi,
I was wondering whether there is an academic/algorithm/whatever solution to the problem of data
dependency between threads. What i mean is whether there is some sort of way to convert, for example,
such code to multi-threaded/cuda code:
for ( int i = 0; i < 1000; i++ )
res[ i ] = res[ i - 1 ] + res[ i - 2 ];
thanks
eyal
Hello, why just not fill the line with Fibonacci numbers ;-)
Hi,
This is just a sample, not the real code. The real code is a loop where each iteration depends on the previous result.
I was wondering whether it can be parallized.
thanks
eyal
I believe it can. If it is some kind of matematical iterative formula, there are (usually) ways to convert it into a standalone formula that would not depend on the previous result.
Otherwise, maybe some scans algorithm?
To tell you more I think we need to see your exact problem.
Hi,
I’ve posted it in the past and made some progress however I’m still not happy with the results, and therefore I decided to ask a more general question to see if there
is indeed some sort of a parallel algorithm to do such thing.
Anyway here is the code and some insights:
The above loop/code cant be parallelized either as the results need to be checked serially.
iTimePlusCeeWindow - iTimeMinusCeeWindow can be assumed (at the moment) to be between 0…40
iTimeEnd - iTimeBegin can be 1…4000
int ceeWindow = ( int ) m_multifocParamSelectionWindow;
for( int iTime = iTimeBegin; iTime < iTimeEnd; iTime++ )
{
int iTimePlusCeeWindow = __min( iTime + ceeWindow, iTimeEnd-1 );
int iTimeMinusCeeWindow = __max( iTime - ceeWindow, 0 );
bool is_best = true;
double dCurrentCor = pCorTemp[ iTime ];
for ( int iSliceTime = iTimeMinusCeeWindow; iSliceTime <= iTimePlusCeeWindow; iSliceTime++ )
{
if ( m_ppCor_Temp_cre_angle_best[ i_rCre ][ iSliceTime ] > dCurrentCor )
{
is_best = false;
break;
}
}
if ( is_best )
{
for ( int iSliceTime = iTimeMinusCeeWindow; iSliceTime < iTimePlusCeeWindow; iSliceTime++ )
{
m_ppCor_Temp_cre_angle_best[ i_rCre ][ iSliceTime ] = pCorTemp[ iSliceTime ];
m_ppStack_Temp_cre_angle_best[ i_rCre ][ iSliceTime ] = pStack_Temp_cre_angle[ iSliceTime ];
}
}
}
thanks
eyal