Intel® Software Network

Valery,
I guess you are trying to work on ia32 architecture, therefore you have to
1)setting  L"/opt/intel/mkl/10.2.1.017/lib/32
and
2) using another linking line:
-Wl,--start-group -lmkl_intel -lmkl_intel_thread -lmkl_core -Wl,--end-group -liomp5 -lpthread

As a result your linking line should be like:

icc stuff.c -I"opt/intel/mkl/10.2.1.017/include/" -L"/opt/intel/mkl/10.2.1.017/lib/32"  -Wl,--start-group -lmkl_intel -lmkl_intel_thread -lmkl_core -Wl,--end-group -liomp5 -lpthread

for more info how to link properly please refer to the KB article: http://software.intel.com/en-us/articles/intel-mkl-link-line-advisor/
--Gennady

Thank you. But with this new linking line don't helped with segmentation fault.

Documentation says that when all eigenvalues are computed either by "A" parameter of by "I" parameter with i = 1 and j = n (this lines don't work), the routine references array isuppz, but it seems valid to me.

Well. At  the first glance all input parameters are correct.
Valery, can you check the problem when all arrays will be allocated dynamically.
For the reproducing the problem - what is the CPU type you are working on?

My distro is Ubuntu 9.04, and processor Core 2 Quad Q9400.
I have done it with dynamical memory, but segfault remains. Also, it's intresting, when I uncomment bzero call, the program finishes normally.

#include <stdio.h>
#include <mkl_lapack.h>
#include <unistd.h>
#include <string.h>

int main(){
int n = 3;
// 1 2 3
// 2 4 5
// 3 5 7
double MM[1000] = {1, 2, 3, 2, 4, 5, 3, 5, 7};
double *M = (double *) malloc( sizeof(double) * 1000);
memcpy(M, MM, sizeof(double) * 9);
//bzero(M, sizeof(double) * 1000);
double a,b;
int i,j;
double accuracy = 1e-6;
int m;
double *eigVal = (double *) malloc( sizeof(double) * 1000);
double *eigVect = (double *) malloc( sizeof(double) * 1000);
int *isuppz = (int *) malloc( sizeof(int) * 1000);
double *work = (double *) malloc( sizeof(double) * 1000);
int *iwork = (int *) malloc( sizeof(int) * 1000);
int many = 1000;
int status = 0;

// First two work:
a = -100;
b = 100;
dsyevr("V", "V", "L", &n, M, &n, &a, &b, &i, &j, &accuracy, &m, eigVal, eigVect, &n, isuppz,
work, &many, iwork, &many, &status);

i = 1; j = 2;
dsyevr("V", "I", "L", &n, M, &n, &a, &b, &i, &j, &accuracy, &m, eigVal, eigVect, &n, isuppz,
work, &many, iwork, &many, &status);

// Second two make segfaults:
i = 1; j = 3;
dsyevr("V", "I", "L", &n, M, &n, &a, &b, &i, &j, &accuracy, &m, eigVal, eigVect, &n, isuppz,
work, &many, iwork, &many, &status);

dsyevr("V", "A", "L", &n, M, &n, &a, &b, &i, &j, &accuracy, &m, eigVal, eigVect, &n, isuppz,
work, &many, iwork, &many, &status);

int k;
for(k = 0; k < n; k ++){
printf("%f ", eigVal[k]);
}
printf("\n");

return 0;
}

btw, originally i checked the problem on win32 and got the following output:
-0.600823 0.695387 11.905435

Do you have the similar results on ubuntu?

No, I have got -0.247820, 0.338816, 11.908962. But I checked them with maxima and online solvers, they are correct.

For myself, I made a = -1e20 and b = 1e20 with "V" option, and make error in my subroutine if dsyevr doesn't find all values.

And I will try it in Ubuntu 8.10 this week.

Valery,
the problem you reported has been fixed in the latest MKL v.10.2 Update 2 available now.
You can download it from intel registratin center.
The version is shipped with Intel Compiler Proffesional Edition, version 11.1 for Linux ( build  056 ).

See the log i 've got running your test:

[root@XXXXX 555076]# ./fix.out

FIRST STEPS

SECOND STEPS

THIRD STEPS

4 STEPS

-0.600823 0.695387 11.905435

--Gennady