<div class="gmail_quote">On Mon, Feb 22, 2010 at 1:39 PM, Siarhei Siamashka <<a href="mailto:siarhei.siamashka@gmail.com">siarhei.siamashka@gmail.com</a>> wrote: <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: > On 02/19/2010 12:57 PM, Siarhei Siamashka wrote: > > Adding small increments to the values at the end of loop iteration could > > be the biggest source of precision loss. Replacing this with explicit > > calculation like 'pdx = pdx0 + cx * n' should improve precision and maybe > > allow to use floats freely. And floats work better with SIMD on any > > platforms. > > And all the SIMD we are covering have a multiply-accumulate instruction > that would be in use, I'm a bit more concerned about the sqrt usage > though... </div>The usage of sqrt is probably not a fatal performance problem. ARM11 VFP has a separate DS pipeline which can calculate divides or square roots simultaneously with the other operations. So it's only a matter of hiding very high square root calculation latency. ARM Cortex-A8 has special SIMD instructions intended to help calculating reciprocals and reciprocal square roots using Newton-Raphson method. SSE has SIMD instructions for calculating square roots. </blockquote><div> Note that SSE functions for square root and reciprocal are unfit for this task as they have very little precision (12 bits). At least one Newton-Raphson iteration must be done to have usable values. </div></div>