From: Leon H. <leo...@ds...> - 2005-05-21 15:41:41
|
----- Original Message ----- From: "Brad Midgley" <bmi...@xm...> To: <gum...@li...> Sent: Saturday, May 21, 2005 4:01 PM Subject: [Gumstix-users] floating point replacements > Hi > > Are there any convenient libraries or routines you guys have come across > for replacing heavyweight floating point (add/mult/div) with short floats > or rationals? > > The bluetooth stereo-audio codec is rife with floating point and shows > pretty dismal performance on xscale (using uclibc fp emulation). > > I started investigating the idea of using rationals but there may need to > be a lot of hand-tuning to normalize the ratio in just the right places... Where I used to work, I replaced the standard IEEE floating point calculations used by the C compiler by ADI's old two-word FP format in a 16-bit DSP application and got it to run much faster. It was a lot of work, though. Leon |