[Omp] aligned memory erratic behaviour when using parallel for, Help!

[Alchemist]

Hi ,

I just want ask, I am using openmp for distributing the threads in my program. I am using the parallel for directive for a 2 iterations for loop (I could easily use sections directive here) each of those two iterations contain big chunk of calculations , Each iteration has the responsibility of calculating collision detection and response , a particle system + other. Each iteration has its own data. Now I found that by defining and initialising the objects by using STL vector the program runs predictably but not so if I use aligned malloc with 16bytes allignment . 
ex. (For the STL version)
//Definition , where CDemo is our class 
vector< CDemo > demo;

//first free the vector if it has something
//Initialisation
for(int i = 0; i < 2; i++)
	demo.push_back(tempDemo);    //Were temp demo has the initial
					//values of the demo class defined
//as "CDemo tempDemo;"

//RunTime
#pragma omp parallel for
for(int i = 0; i < 2; i++)
 	demo[i].calculate();

But when I define the objects as a pointer and initialise the memory by using the "_aligned_malloc" and "aligned_free", to allocate aligned  memory blocks, I am getting a more erratic behaviour from my program.

ex. (for the aligned memory version)
//Definition
CDemo *demo;

//free if demo if not NULL by "using _aligned_free(demo)"
//allocate memory
demo = (CDemo*)_aligned_malloc(2*sizeof(CDemo), 16);

//using crt optimised memcpy "#pragma function(memcpy)", for 
//aligned memory chunks
for(int i = 0; i < 2; i++)
	memcpy(&demo[i], &tempDemo, sizeof(CDemo));

//RunTime is the same as in the Vector example 

//------------------------------------------------------------//
Am I doing something wrong? do I have to change my schedule chunk to 16 or something for the aligned memory part? 

I am using the OpenMP that is provided with the VS2005 beta version, but I have compiled the program with Intel Compiler as well in WinXP Pro and I am getting similar results.  

I would appreciate any help. 
  

-- 
Best regards,
 Emmanouil Hatjissavvas                          mailto:my at daedalus.plus.com