Usage of OpenMP reduction clause with nested loops - c++

Question

I have the current version of a function:
void*
function(const Input_st *Data, Output_st *Image)
{
int i,j,r,Offset;
omp_set_num_threads(24);
#pragma omp parallel for schedule(static) shared(Data,Image),\
private(i,j,r,Offset)
for (i = 0; i < Data->NX; i++)
{
for (j = 0; j < (Data->NZ); j++)
{
for (r = 0; r < Data->NR; r++)
{
Offset = i*Data->NR*Data->NZ + j*Data->NR + r;
Image->pTime[Offset] = function2()
}
}
}
return NULL;
}
It works very well, however I wanted to remove the calculation of the variable Offset and use of a pointer pointing to the member Image->pTimeR and then increment, which can look like following:
void*
function(const Input_st *Data, Output_st *Image)
{
int i, j, r;
double *pTime = Image->pTime;
omp_set_num_threads(24);
#pragma omp parallel for schedule(static) shared(Data,Image),\
private(i,j,r)
for (i = 0; i < Data->NX; i++)
{
for (j = 0; j < (Data->NZ); j++)
{
for (r = 0; r < Data->NR; r++)
{
*pTime = function2()
pTime++;
}
}
}
return NULL;
}
I get Seg Fault. I assume I need to use the reduction clause like reduction(+:pTime).
First, the purpose here is to speed up the function and I am wondering if such change would significantly speed up? (Like less cache memory used?)
Second, well I tried to benchmark it and failed to do so! I think the problem here can be solved by using a reduction clause, but since loops are nested the problem is not that straightforward to me.

Answer 1

There's no need of any sort of reduction clause here. However,at the moment, all threads use the same pointer and update the same memory location (with race conditions in the value assigned to pTime, hence the crashes I suspect).
So you need to define your pointer in a private way (typically by declaring it within the parallel region, and to set it individually per thread to a meaningful value. Then it can be incremented the way you want.
Here is what the code could look like once fixed (not tested obviously):
void* function( const Input_st *Data, Output_st *Image ) {
#pragma omp parallel for schedule( static ) num_threads( 24 )
for ( int i = 0; i < Data->NX; i++ ) {
double *pTime = Image->pTime + i * Data->NR * Data->NZ;
for ( int j = 0; j < Data->NZ; j++ ) {
for ( int r = 0; r < Data->NR; r++ ) {
*pTime = function2();
pTime++;
}
}
}
return NULL;
}