iterative refinement with extended preicision

[xilin1]

Hi,
In the reference manual for pardiso, it says

Note that if iparm(8)< 0, the accumulation of the residuum is using extended precision real and complex data types.

I am solving using iparm(28)=0, double precision. When I try to set iparm(8) to a negative integer for extended precision, the result is random values. If I set iparm(8) to a positive integer, I get slightly better accuracy in the solution.

Is there anything special I need to do to go from double precision refinement to quad precision refinement?

Xi Lin

[xilin1]

I forgot to mention, the refinement using quad precision works for unsymmetric matrices mtype=13 but fails for mtype=6. The settings of other parameters iparm(11),iparm(13),iparm(21) don't make any difference.

[Gennady_F_Intel]

Hi Xi Lin,

It would be nice if you provided for us a self-contained example that will helps us to reproduce the problem.

(Feel free to reply in Private if needed.)

What MKL version you are using?

Whats your linking line?

Is this ia32 or Intel64 architecture?

Regards, Gennady

[xilin1]

Quoting - Gennady Fedorov (Intel)

Hi Xi Lin,

It would be nice if you provided for us a self-contained example that will helps us to reproduce the problem.

(Feel free to reply in Private if needed.)

What MKL version you are using?

Whats your linking line?

Is this ia32 or Intel64 architecture?

Regards, Gennady

Hi ,
I have determined what the problem was. I am using the pgi 8.0 compiler which does not support extended precision (though the intel compiler does). Do you know of a workaround to use the iterative refinement with extended precision with pgi compiler?

My MKL is 10.2 , this is em64t libraries.

Xi Lin

[Sergey_Solovev__Inte]

Hi,
can you try to use Intel compiler for linking your example? Is the result correct with Intel compiler? Could you provide us with your test?
Sergey Solovev