On Apr 21, 2006, at 9:13 AM, Décio Luiz Gazzoni Filho wrote:
On Apr 21, 2006, at 6:51 AM, Rustam Muginov wrote:
[snip]
And another question about optimization. I would like to compare how worse code would behave if compiler wont use the SSE/SSE2 instruction for scalar floating point, and try to compile code to use only x387 math. Which additional compiler flags I should pass to gcc to prohibit usage of SSE/SSE2?
I believe -mno-sse and -mno-sse2.
Use -mfpmath=387 for 387 floating point. This is normally set to 'sse' for gcc on Intel Macs. I would highly recommend doing test with each setting. I have some floating point fractal code (FFFF project) that performs better using the 387 floating point. I believe that the Intel compiler defaults to 387 floating point code. Don't assume that SSE floating point is faster.
Thanks, Steve |
