joel falcou wrote:
I am actually interested about the iterators performance in multi-array too, since their iterators seem to just wrap indexes. That _seems_ like it would be bad, but Idunno.
iterator or pointer almost leads to the same performances (ina 5 to 10% margin maybe). The real error people keep doing is performing test on huge dataset without any cache friendly algorithm
Did you mean to say 'index or pointer'? It seems like the trend in image processing is to avoid using indexes in the inner loop if at all possible. As far as the cache-friendly issue, the example used the same memory and the same access pattern. I repeated the test on my system with mingw g++ 4.3.1 and the -O2 flag, and I got similarly bad results ... BUT I figured out why the performance was so bad in that case! First, the poster did not use BOOST_DISABLE_ASSERTS, so the multi_array was doing the extra checks. Second, the results of the operation were never read, so the compiler removed the whole 'Pointers' test. With that macro defined, both loops are completely removed by the compiler. If I make the result volatile, then the running times are: Pointer....: 532 Multi_array: 593 HOWEVER, when I switch to using iterators & pointers for the inner loop the penalty is a factor of 10 (not 10%, but 1000%) Pointer....: 516 Multi_array: 5141 Again, same access pattern, same memory. --John //---modified code---- #define BOOST_DISABLE_ASSERTS #include "windows.h" #include <iostream> #include "boost\multi_array.hpp" using namespace std; using namespace boost; const size_t DimX = 2000; const size_t DimY = 1500; const size_t DimZ = 1; volatile int iVal; //gbl volatile in an attempt to avoid optimizing away.... template<class It> void test_helper(It begin, It end){ while (begin != end) iVal = *begin++; } void testPointer(multi_array<int,3> &V3) { int * pInt = V3.data(); for (size_t Z=0; Z<DimZ; ++Z) for (size_t Y=0; Y<DimY; ++Y) test_helper(pInt + (Z*DimY+Y)*DimX, pInt + (Z*DimY+Y)*DimX + DimX); } void testMultiArray(multi_array<int,3> &V3) { for (size_t Z=0; Z<DimZ; ++Z) for (size_t Y=0; Y<DimY; ++Y) test_helper(V3[Z][Y].begin(), V3[Z][Y].end()); } int main() { multi_array<int,3> V3(extents[DimZ][DimY][DimX]); unsigned long time1, time2; time1 = GetTickCount(); for (int q = 0; q < 100; ++q) testPointer(V3); time2 = GetTickCount(); cout << "Pointer....: " << time2-time1 << endl; time1 = GetTickCount(); for (int q = 0; q < 100; ++q) testMultiArray(V3); time2 = GetTickCount(); cout << "Multi_array: " << time2-time1 << endl; return 0; }