The graphs below demonstrate the performance of double-precision exp and log functions in Yeppp! library compared to other mathematical libraries. The performance and accuracy are measured with Hysteria on [0.1, 10000.0] for log and [-707.0, 707.0] for exp function (the intervals are the chosen the same as in MKL VML performance and accuracy report to simplify comparison). CRLibM, Cephes, FDLibM, and SLEEF libraries, which are available in source code, were compiled with gcc 4.8 with -O3 and appropriate -march flags.

On Intel machines Yeppp! performs on the same level as Intel libraries, and on AMD machines it outperforms any alternative by a good margin. Yeppp! combines high performance with good accuracy: the maximum measured error is about 1.0 ULP for exp and 1.4 ULP for log. This can be interpreted as the last 1.0-1.5 bits being inaccurate, which is not a problem for most use-cases (after all, there are 51+ accurate bits!). Yeppp! also properly handles all special cases, such as infinite input, negative input for log, etc.

Not bad for a software written in Python, Yeppp?

Update on July 14, 2013: performance on Haswell

Update on July 18, 2013: the above numbers are measured for single-threaded execution. Haswell, Sandy Bridge, Bobcat, and Piledriver processors were tested with disabled DVFS (Intel Turbo Boost/AMD Turbo Core technologies).

---