Agner`s CPU blog

Choosing the most efficient function library can be a nightmare to a programmer. I have tried to calculate the cosine function with different libraries and compare the calculation time. The best version is 19 times faster than the worst!

AMD have now updated their math libraries and added CPU dispatching. There are two versions of code in AMD's LIBM library: One for the SSE2 instruction set and one for AVX and FMA4. Intel processors will run the inferior SSE2 branch because they don't have the FMA4 instruction set. The incompatibility between Intel's and AMD's FMA instructions is another scandal, which I have discussed in this blog post. The AMD library does not check the CPU brand name as Intel libraries do. It only checks for the FMA4 instructions which are not supported by Intel processors, although - quite ironically - they were designed by Intel. It will be possible to run the better branch on Intel processors if Intel decides to support the FMA4 instruction set in the future.

The following table shows the results of some tests I have made of different math libraries.

The Gnu function library (glibc) uses an outdated and inefficient code. The Microsoft library has decent performance in the 64-bit version, but of course it supports only the Windows platform. Intel's general math library is no better in my test case, but Intel's Short Vector Math Library (SVML) is very good. The SVML library supports vectors of 4 floats in an XMM register (SSE2) or a vector of 8 floats in a YMM register (AVX). It will choose the inferior generic path for non-Intel processors unless we replace Intel's CPU-dispatcher as described above. Intel's libraries are available for both Windows, Linux and Mac. AMD's LIBM library supports vectors of 4 floats. It is available for Windows and Linux, but only in 64-bit mode.

The sad conclusion is that we have no fully optimized math function library that supports all brands of x86 processors and all operating systems. If we want optimal performance on all processors, the best choice is to use Intel's SVML library and manipulate it into treating non-Intel processors better.

It would be nice if more people would work on improving glibc. This library supports all processors and platforms, but it is poorly optimized. Only a few memory and string functions in glibc have CPU dispatching, while the math functions have only old and poorly optimized versions. It would also be nice to have vector versions of the math functions in glibc because the Gnu compiler has support for such functions.