topic One more data race threading issue in zgetrf (10.3 update 5) in Intel® oneAPI Math Kernel Library

One more data race threading issue in zgetrf (10.3 update 5)

AndrewC — Tue, 30 Aug 2011 16:24:38 GMT

MKL 10.3 update 5 has made many strides over previous versions in fixing data races in multi-thread applications. I run tests on my code under Intel Inspector XE , and the data races inside MKL are almost gone
... EXCEPT for this one that keeps cropping up. I do not believe it is a "harmless" data race. They never are...

This needs to run with OpenMP on, obviously.
ID Problem Sources Modules State
P1 Data race testzgetrf.cpp testzgetrf.exe New

The actual data race (write) is at this location.

mkl_core.dll!mkl_serv_deallocate - mkl_core.dll:4937
testzgetrf.exe!zgetrf - testzgetrf.exe:6101
testzgetrf.exe!test - testzgetrf.cpp:27

[cpp]#include "mkl_lapack.h"
#include 
void test()
{
#define NSIZE 20
		#pragma omp parallel for
		for( int iLoop=0;iLoop<20;iLoop++){
			MKL_Complex16 a[NSIZE][NSIZE];
			for( int i=0;i.real=1;
						a.imag=0;
					}else{
						a.real=0;
						a.imag=0;
					}
				}
			}
			int M = NSIZE;
			int N = NSIZE;
			int info;
			int ipvt[NSIZE];
			zgetrf(&M, &N, (MKL_Complex16 *)a, &M, ipvt, &info);
			//std::cout << info << std::endl;
		}
}

int main(int argc, char * argv[])
{
	test();
	return 0;
}

[/cpp]

One more data race threading issue in zgetrf (10.3 update 5)

yuriisig — Wed, 31 Aug 2011 10:03:49 GMT

MKL_Complex16a[NSIZE][NSIZE];??????????

One more data race threading issue in zgetrf (10.3 update 5)

AndrewC — Wed, 31 Aug 2011 13:46:41 GMT

And your point is what?

One more data race threading issue in zgetrf (10.3 update 5)

yuriisig — Wed, 31 Aug 2011 14:45:11 GMT

It is an one-dimensional array.

One more data race threading issue in zgetrf (10.3 update 5)

AndrewC — Wed, 31 Aug 2011 15:54:57 GMT

zgetrf expects a NxM matrix of complex double.
It declares a block of memory that can be though of as a one-d array [NSIZE*NSIZE] or a 2D array [NSIZE][NSIZE], lapack does not care how you declare your memory as long as the layout is column major, and since my array is square that is not an issue.

One more data race threading issue in zgetrf (10.3 update 5)

yuriisig — Wed, 31 Aug 2011 19:33:05 GMT

Very interesting. I did not know it. Tell explicitly as storage for a two-dimensional array is selected.
This selection of storage depends on the translator?

One more data race threading issue in zgetrf (10.3 update 5)

AndrewC — Wed, 31 Aug 2011 19:36:57 GMT

Is this a Turing test? And way off-topic

One more data race threading issue in zgetrf (10.3 update 5)

yuriisig — Wed, 31 Aug 2011 23:03:59 GMT

I hinted that you are mistaken. Consider an example of the fragmented storage.