09-25-2018 04:49 AM
I really think Intel Short Vector Math Library (SVML) is great.
Yet I find I'm missing few features which would be great if you could add them:
- Ability to Select Precision by Function Naming.
Currently it seems there are 3 levels of accuracy (See Intel® Math Kernel Library 2018 - Vector Mathematics (VM) Performance and Accuracy Data which basically is a wrapper around SVML) which are controlled by Floating Point Precision Mode (fp:precise for HA [1 ULP], fp:fast=1 for LA [4 ULP] and fp:fast=2 for EP [Half of bits are correct]). I wish we had a direct calls for each of the function flavor. This is useful in case we have code that needs different accuracy levels (Of the same function) in different calls.
- Better Documentation of Accuracy
Current documentation of SVML doesn't say anything about accuracy.
- Optimized Call Case
There are cases which can be optimized which are not. Simple cases is Trigonometric functions which contains multiplication of Pi. The power Function in cases the Power is Int or in cases it is the same for all elements, etc...
- Open Source
Well, this is a wishful thinking, but it would be great if you made SVML open source so everybody could enjoy it.
- Development Tools
- Intel® C++ Compiler
- Intel® Parallel Studio XE
- Intel® System Studio
- Parallel Computing