KNL Fortran 64 byte alignment

[Paulius_V_]

Hello all. 

I encountered an error after recompiling my code with -align array64byte and using !dir$ vector aligned directive. 

serial job  x86 - ok, mpi job x86 - ok , serial MIC - ok, MPI MIC - array index out of bounds. 

I suspect that this might be due to the MPI implementation using some C++ modules. Why is there no 64-byte alignment option in icc? Is there an easy way of addressing this issue through compiler flags? If not, and C++ alignment is indeed the cause of this failure then I can exclude such variables but this would be a labor intensive endeavour. 

Any suggestions?

Many thanks

[jimdempseyatthecove]

Paulis,

Consider: overload new technique as well as overload malloc.

See: http://stackoverflow.com/questions/16270891/can-we-overload-malloc

Jim Dempsey

[Paulius_V_]

Thanks for the tip, Jim. So if I use polymorphism to address this, I'd need to implement my own version of malloc that ends up allocating along the 64 bye boundary? How would I make it so that my malloc takes precendence over the standard malloc? Especially if I don't want to allocate everything to be algined? 

If that's not an option then would I not just use mm_malloc?

Also, do you have any insight as to why there is not compiler flag for this?

Thanks

[jimdempseyatthecove]

Both Windows and Linux have an API to perform aligned allocation (you also have the Intel mm_malloc as you have shown). You only need to write a (a few) shell function(s) that receives the malloc/free arguments and makes the appropriate call the aligned variants (e.g. mm_malloc). You can do the same with the C++ new, [] new, delete, [] delete.

Jim Dempsey

[jimdempseyatthecove]

>> How would I make it so that my malloc takes precendence over the standard malloc? Especially if I don't want to allocate everything to be aligned? 

The two have contradictory objectives.

Define your overload to suit your needs or use the alternate allocation/deallocation.

Jim Dempsey