topic Do the MKL FFT Handler creation uses malloc?? in Intel® oneAPI Math Kernel Library

Do the MKL FFT Handler creation uses malloc??

shiva_rama_krishna_b — Wed, 12 Nov 2014 12:05:46 GMT

Hi,

When we want to use MKL FFT we create a handler for that. So do this handler creation involves malloc system call? If not which method is used to allocate memory.

Thanks

sivaramakrishna

MKL calls the i_malloc

Evgueni_P_Intel — Wed, 12 Nov 2014 12:18:20 GMT

MKL calls the i_malloc callback function which has the same interface as malloc and can be overridden by the user.

Thank you Evgueni Petrov.

shiva_rama_krishna_b — Thu, 13 Nov 2014 04:56:10 GMT

Thank you Evg ueni Petrov.

what happens If more than one thread calls creation of handler of its own. do they run sequncially or parallel?.

Thanks

sivaramakrishna

MKL does not serialize

Evgueni_P_Intel — Thu, 13 Nov 2014 05:53:35 GMT

MKL does not serialize execution of DftiCreateDescriptor in different application threads.

Hi Evgueni Petrov,

shiva_rama_krishna_b — Fri, 21 Nov 2014 07:19:59 GMT

Hi Evgueni Petrov,

I was trying call FFT by multiple threads at a time. code snippet looks like below. I noticed [MKLSERVICE]@lock function in vtune profiler. when i pulled the handler creation outside the loop, i did not find [MKLSERVICE]@lock in vtune profiler. This is actually making me to believe that MKL is doing some locking while allocating handler.

#pragma omp parallel num_threads(236)
for(i = 0; i < 236; i++)
{
                          .
                          .
                          .
                       some functions..
                          .
                          .

                      create handler one for each 'i'.
                      Call FFT using 'i'.

}

Thanks

sivaramakrishna