topic multi-threaded matrix multiplication in Intel® oneAPI Math Kernel Library

multi-threaded matrix multiplication

Likun_T_ — Wed, 18 Mar 2015 23:33:54 GMT

Dear Intel MKL developers,

I am integrating the MKL subroutine mkl_zcsrmultcsr in my MPI code. I tested a case with 16 processors, and mkl_zcrsmultcsr is called in every processor in parallel. Once it is called, multi-threaded computing is automatically activated.

The problem I encountered is that 12 processors among all processors work fine while the other 4 processors give memory corruption errors, moreover, these 4 processors can vary during each test. I am not sure what the problem would be. Your advise is well appreciated.

best,

Likun

Dear all,

Likun_T_ — Thu, 19 Mar 2015 20:09:33 GMT

Dear all,

Some update about my previous post. If I set number of processors to be 4, the call of mkl_zcsrmultcsr is successful in each processor, however, if i tried 8 or 16 processors, I got memory corruption errors.

The structure of my code is like this:

subroutine test

call test1()

call mkl_zcrsmultcrs()

call test2()

end subroutine test

all the calculations in subroutines test1, test2 and mkl_zcrsmultcrs are independent among processors. I am not sure if anything else is needed for multi-threaded computing.

Thanks,

Likun

Hello Likun,

Gennady_F_Intel — Fri, 20 Mar 2015 03:28:13 GMT

Hello Likun,

What do you by "independent among processors"? Are you working on the SMP system with 16 threads? Do you use the same A,B and C arrays which you use in mkl_zcsrmultcsr?

--Gennady