different optimization level, different output data

[woshiwuxin]

Hi!

I've tried to compile a package with different optimization levels, e.g. -O0, -O1, -O2 and -O3. However, the printed numerical values are slightly different in the last several digits. It's not very significantly, but annoying. I just want to know which optimization level can generate morerobustand safer binary, if we don't consider the execution time of the binary.

Thanks!

[mecej4]

There is no magic combination of floating point options that guarantee the best compromise between speed and accuracy for all applications and users. Read about the several available FP related options in the User Guide and try a few.

o Have you established that higher optimization levels correlate with lower accuracy?

o Can you rule out bugs in your code?

o Is it possible that the conditioning of the algorithm you used is at odds with your expectations as to accuracy?