Intel® oneAPI Threading Building Blocks
Ask questions and share information about adding parallelism to your applications when using this threading library.

Slow memory access

tofuny
Beginner
230 Views
Hi!!, I have to access the array p in loop of k index of below code to save any calculated value. But it takes much time to access the array. I checked access time in my machine (VS2008, release mode, i7(6 cores), windows 7):
- there is empty loop(no calculation and no array access): 17ms
- there is the calculation part of (1): 17ms
- no calculation part and there is just "p = 1000" in the loop : 21ms
- calculation part (1) + "p = 1000" : 57ms
- calculation part (1) + "p = val" : 123ms
I cannot understand why it takes so long time to access the array "p" in the for loop.
=======================================
#include "tbb/tbb.h"
#include "tbb/blocked_range.h"
#include "tbb/blocked_range2d.h"
#include "tbb/parallel_for.h"
#include "tbb/tick_count.h"
#include "tbb/scalable_allocator.h"
using namespace std;
using namespace tbb;
/////
class Average : public CPartLoop
{
public:
...
int hh, ww;
void operator() (const blocked_range2d& range) const;
void ParallelProcessing(int hh, int ww);
};
void Average::operator() (const blocked_range2d& range) const
{
register int i,j,k;
float *p = scalable_allocator().allocate(100);
for(i=range.rows().begin(); i != range.rows().end(); ++i)
{
int ii = (int)(i/50);
int jj = i-ii*50;
for(j=range.cols().begin(); j != range.cols().end(); ++j)
{
int dd = (int)(j/36);
int ss = j-dd*36;
for(k=0; k
{
//// (1) math calculation part ////
int nr = (int)kp.r;
int nc = (int)kp.c;
if( ii+nr>hh-1 || ii+nr<0 || jj+nc > ww-1 || jj+nc<0) continue;
int index1 = (ii+nr)*ww;
int tr = A[index1+(jj+nc)];
int tc = B[index1+(jj+nc)];
float te = C[index1+(jj+nc)];
float aa = dd + (kp.d*r2d);
float bb = kp.e;
float val = (tr+tc)*cos(aa*d2r)*(bb * te);
////// end of (1) ///////
p = val; // 1000 // (2) array access part
}
}
}
tbb::scalable_allocator().deallocate(p,100);
}
void Average::ParallelProcessing(int hh, int ww)
{
this->hh = hh;
this->ww = ww;
parallel_for(blocked_range2d(0,50*50,4,0,360,4), *this);
}
0 Kudos
2 Replies
RafSchietekat
Valued Contributor III
230 Views
Before we look at anything else, try to "hoist" the range boundary "out of the loop" (just in the inner loop may well be enough), by assigning it to a value that can then be used in the condition instead. The compiler may then be able to perform better optimisations.

(Added 2012-04-03) Well, maybe not here...
0 Kudos
SergeyKostrov
Valued Contributor II
230 Views

Could you provide a complete Test-Case? Here is a list of compilation errors:

...
error C2504: 'CPartLoop' : base class undefined
error C2065: 'kp' : undeclared identifier
error C2065: 'A' : undeclared identifier
error C2065: 'B' : undeclared identifier
error C2065: 'C' : undeclared identifier
error C2065: 'r2d' : undeclared identifier
error C2065: 'd2r' : undeclared identifier
...

0 Kudos
Reply