How to reduce transcendental function time like exp and log

[chen__kafty]

Hi,

I am using VTune profile tool to analysis myprogram to see if there is any opportunity to improve the code efficiency. From the report, I found the transcendental function like _libm_sse2_exp and _libm_sse2_log costs about 30% of total running time. If I am right, INTEL has announced that some new technology can be used to reduce the transcendental funtion time such as compiled the program by INTEL compiler(my program is writen by C++ and compiled by Vistual Studio 2005). Anybody can help to clarify my guess?

Thanks

[Peter_W_Intel]

You can save current performance data, and use intrinsic functions from the compiler for SSE4 or AVX support - it depends on your platform, then you cananalyze it again to compare two results.

Using Intel C++ compiler, just use option like "/Qax [SSE3|SSE4.1|SSE4.2]" - for example. Please go C++ compiler forum, if you have more questions.

Thanks, Peter

[TimP]

Intel C++ offers auto-vectorization, with automatic substitution of Short Vector Math library calls, in case your source code is appropriately organized. Also, the MKL VML performance library offers vector math functions. VS2005 is no longer supported by the latest Intel compiler release.
If your application permits use of float rather than double versions of the math functions, those are available in both the Microsoft and Intel libraries.