<html><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space; "><br><div><div>On 22-Feb-10, at 10:35 AM, Rodrigo Kumpera wrote:</div><br class="Apple-interchange-newline"><blockquote type="cite"><br><br><div class="gmail_quote">On Mon, Feb 22, 2010 at 1:39 PM, Siarhei Siamashka <span dir="ltr"><<a href="mailto:siarhei.siamashka@gmail.com">siarhei.siamashka@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"> <div class="im">On Friday 19 February 2010, Luca Barbato wrote:<br> > On 02/19/2010 12:57 PM, Siarhei Siamashka wrote:<br> > > Adding small increments to the values at the end of loop iteration could<br> > > be the biggest source of precision loss. Replacing this with explicit<br> > > calculation like 'pdx = pdx0 + cx * n' should improve precision and maybe<br> > > allow to use floats freely. And floats work better with SIMD on any<br> > > platforms.<br> ><br> > And all the SIMD we are covering have a multiply-accumulate instruction<br> > that would be in use, I'm a bit more concerned about the sqrt usage<br> > though...<br> <br> </div>The usage of sqrt is probably not a fatal performance problem.<br> <br> ARM11 VFP has a separate DS pipeline which can calculate divides or square<br> roots simultaneously with the other operations. So it's only a matter of<br> hiding very high square root calculation latency.<br> <br> ARM Cortex-A8 has special SIMD instructions intended to help calculating<br> reciprocals and reciprocal square roots using Newton-Raphson method.<br> <br> SSE has SIMD instructions for calculating square roots.<br> <br></blockquote><div><br>Note that SSE functions for square root and reciprocal are unfit for this task as they have<br>very little precision (12 bits). At least one Newton-Raphson iteration must be done to have<br>usable values.</div></div></blockquote><br></div><div>rsqrt is an approximation, however the regular square root instructions are not and are fine replacements for the sqrt() function. In fact, gcc on OS X already does this substitution.</div><div><br></div><div>-Jeff</div><div><br></div><div><br></div></body></html>