topic Please amend the in Intel® oneAPI Math Kernel Library

mkl_realloc alignment

DarioMangoni — Tue, 01 Sep 2015 12:20:00 GMT

Hi,
I'm facing a problem probably due to misalignment so I decided to use mkl_*alloc with Parallel Direct Sparse Solver.

The solver needs 64byte alignment that is succesfully obtained my mkl_malloc/mkl_calloc specifying the alignment parameter.
Unfortunately this is not possible with mkl_realloc: its default is a simple 16byte alignment and nothing can be done to bypass this default behaviour.
Any workaround?

Thanks,
Dario

void* mkl_realloc (void *ptr,

Gennady_F_Intel — Wed, 02 Sep 2015 12:46:12 GMT

void* mkl_realloc (void *ptr, size_t size); If ptr is NULL, the function works as mkl_malloc !

Sorry, maybe I didn't specify

DarioMangoni — Wed, 02 Sep 2015 13:07:53 GMT

Sorry, maybe I didn't specify that I WANT to reallocate that vector! If I didn't want to, I'd use mkl_malloc!
If I use mkl_realloc to reallocate, the address is no more 64byte aligned!

I can create another array with mkl_malloc and then copy the data, but the memory peak (mkl_peak_mem_usage()) is incredibly higher than using mkl_realloc.

Another question... does mkl_realloc try to "expand" the current array before moving it elsewhere?

thanks for the issue. mkl

Gennady_F_Intel — Thu, 03 Sep 2015 13:50:12 GMT

thanks for the issue. mkl_realloc doesn't remember the alignment of the "original" array but it must to keep this alignment. We will fix this into one of the next releases/updates.

the problem has been fixed in

Gennady_F_Intel — Fri, 10 Feb 2017 05:36:40 GMT

the problem has been fixed in MKL v.11.3.1

Please amend the

Eugene_S_1 — Wed, 20 Jun 2018 09:13:06 GMT

Please amend the documentation on mkl_realloc by explicitly stating that alignment is preserved. Now it says nothing about the alignment of a new buffer.