Hi,

I installed CentOS 5.1 on 2 machines, one is EM64T and the other is AMD64. They generate different results for some floating point operations and I suspect it is a issue of the kernel or system run-time library.

I diff'ed their kernel images and the 2 machines have exactly the same kernel image. But, when I diff'ed the /lib directories of the two machines, I found that they are different. And, the /lib64 directories are different too.

I used the same installation source for both machines, and the EMT64T and AMD64 ISA are compatible. Why are the /lib and /lib64 directories different across platforms?

Thank you very much. Any idea would be appreciated.

Because the kernel has some ASM and C headers (kernel headers) that are arch specific, of course the contents are going to be different.

I've had similar problems; do you have a simple program that can demonstrate the differences in the floating point results? Obviously if they differ they can't both be correct (but it is possible and not too unlikely that both are wrong). On my part I didn't see any problems until I ran some fairly complex code and I haven't had time to try to reproduce the problems with simple code.

I suspect the problem might be something along the lines that libc is using various features to speed up calculations. For example, instead of using the FPU it might use an SSE2 instruction (AMD) instead and the algorithm used is buggy. There may be equally buggy routines in the Intel-specific routines. Sadly, this sort of problem is pretty common - too few programmers know their maths and some think that a 'shortcut' (translation: WRONG implementation) is acceptable. 'Matlab' has notoriously buggy math routines but you don't see many users complaining about it.

You may be interested in the following 2 papers.

Evaluation of IEEE 754 floating-point arithmetic compliance across a wide range of heterogeneous computers

Ensuring Numerical Quality in Grid Computing



The simplest and the best way to investigate the problem is by running the IEEE 754 Compliance Checker.

IEEECC754
http://www.cant.ua.ac.be/ieeecc754.html

Thanks! That should help me sort out this problem I have.