topic Quad precision for MKL functions in Intel® oneAPI Math Kernel Library

Quad precision for MKL functions

S__MPay — Mon, 01 May 2017 07:10:59 GMT

Is it possible to use quad precision for MKL functions and maintain the accuracy?

Is your question answered by

TimP — Mon, 01 May 2017 08:07:58 GMT

Is your question answered by previous threads about MKL and quad precision or real*16 ?

Quote:Tim P. wrote:

S__MPay — Mon, 01 May 2017 10:12:25 GMT

Tim P. wrote:

Is your question answered by previous threads about MKL and quad precision or real*16 ?

According to a similar threat:

Tim P. wrote:

You would have to compile the public source code (netlib.org) to make the quad precision versions of the functions you want. You would be giving up on the "performance" aspect of MKL, as there is no support for vectorization in quad precision.

But my confusion is for some inaccuracies that are causing trouble are gone when I use quad precision even though all I did was to simply call double precision version of MKL.

MKL doesn't support quad

Gennady_F_Intel — Mon, 01 May 2017 13:25:31 GMT

MKL doesn't support quad precision. MKL's API supports single & double precision only.

It's certainly possible to

TimP — Mon, 01 May 2017 15:01:06 GMT

It's certainly possible to use higher precision in your source code while calling the MKL in standard double precision. It's not normally necessary. If you care to isolate where you may be losing accuracy, and that doesn't make the problem evident, you might show it on the relevant forum.

When I try to calculate the

S__MPay — Mon, 01 May 2017 16:27:33 GMT

When I try to calculate the determinant of a matrix there is error in double precision while the error of quad precision is exactly zero:

    print *, 'zero determinant double precision' , dprMklDet(reshape((/ 6.d0, 3.d0, 5.d0, 4.d0, 4.d0, 7.d0, 12.d0, 6.d0, 10.d0 /),(/3, 3/)))
    print *, 'zero determinant  quad  precision ' , qprMklDet(reshape((/ 6.q0, 3.q0, 5.q0, 4.q0, 4.q0, 7.q0, 12.q0, 6.q0, 10.q0 /),(/3, 3/)))

Output is:

zero determinant double precision 1.065814103640150E-014
zero determinant quad precision 0.000000000000000000000000000000000E+0000

I will post more detail in this topic https://software.intel.com/en-us/forums/intel-math-kernel-library/topic/733238