Accuracy/performance of Intel Math Library

[Robert_E_1]

Can you please point me to the accuracy and performance tables for the Intel Math Library described at https://software.intel.com/en-us/node/522653 ?

[Jing_Xu]

Hi Robert,

The link you mentioned is the User and Reference Guide of Intel C++ compiler 15. I suppose what you are interested in may be Math Kernel Library (MKL), which is not a part of the compiler. The latest version of MKL is 2017(https://software.intel.com/en-us/intel-mkl). Also, you may find documentation of MKL here.

Regarding to the accuracy and performance table, is this (https://software.intel.com/en-us/intel-mkl/benchmarks) what you want?

[Robert_E_1]

Actually, it is the Intel Math Library described at  https://software.intel.com/en-us/node/522653 that I am interested in performance figures for.  In particular, for the long double functions.  (As far as I am aware, the MKL does not contain long double functions.)  I am interested in seeing tables for IML similar to the ones given for MKL at https://software.intel.com/sites/products/documentation/doclib/mkl/vm/functions/_performanceall.html .

[Marius_C_Intel · Accepted solution]

Hello Robert,

We do not have performance data published for the math library described at  https://software.intel.com/en-us/node/522653. The long double functions (I assume you mean those using the 80-bit double-extended floating-point format) will have lower performance than the 32-bit single precision or 64-bit double precision ones for example, mainly because the latter can take advantage of the existing SIMD instructions (from 128-bit SSE to 512-bit AVX-512, depending on the processor you are using). Performance should still be/is very good for the long double functions, as they are based on native double-extended floating-point operations available in the Intel(r) Architecture x87 FPU (Floating-Point Unit).

Thanks,

Marius

[Robert_E_1]

Thanks for the information, Marius.  (Yes, you are correct, I did mean 80-bit FP.)  Robert