Hi, I am looking for some good fixed point algorithms for logarithms and exponents. Any pointers would be helpful. Thanks and Regards, -Anoop Deoras
Fixed point algorithms
Started by ●August 20, 2004
Reply by ●August 20, 20042004-08-20
For hardware implementations, you might look at CORDIC. CORDIC wouldn't work well for general purpose or DSPs. In article <2f592dd3.0408200028.21978843@posting.google.com>, akd_ecc@yahoo.com (EC-AKD) wrote:>Hi, > > I am looking for some good fixed point algorithms for logarithms and >exponents. Any pointers would be helpful. > >Thanks and Regards, >-Anoop Deoras
Reply by ●August 24, 20042004-08-24
>>>>> "George" == George Bush <george.w.bush@whitehouse.com> writes:George> For hardware implementations, you might look at CORDIC. CORDIC wouldn't work George> well for general purpose or DSPs. Why not? Ray
Reply by ●August 24, 20042004-08-24
CORDIC uses bit operations which aren't effecient in general purpose operations. In article <sxdr7pw615c.fsf@edgedsp4.rtp.ericsson.se>, Raymond Toy <raymond.toy@ericsson.com> wrote:>>>>>> "George" == George Bush <george.w.bush@whitehouse.com> writes: > > George> For hardware implementations, you might look at CORDIC. CORDIC > wouldn't work > George> well for general purpose or DSPs. > >Why not? > >Ray
Reply by ●August 25, 20042004-08-25
>>>>> "George" == George Bush <george.w.bush@whitehouse.com> writes:George> CORDIC uses bit operations which aren't effecient in general purpose George> operations. Your CORDIC must be different from mine. All I had were compares, adds, subtracts, shifts, and table lookups. Sounds pretty general purpose to me. Ray
Reply by ●August 25, 20042004-08-25
Cordic for logarithms? Got a reference for that? Dirk george.w.bush@whitehouse.com (George Bush) wrote in message news:<%GnVc.8231$Bt5.4288@twister.socal.rr.com>...> For hardware implementations, you might look at CORDIC. CORDIC wouldn't work > well for general purpose or DSPs. > > In article <2f592dd3.0408200028.21978843@posting.google.com>, > akd_ecc@yahoo.com (EC-AKD) wrote: > >Hi, > > > > I am looking for some good fixed point algorithms for logarithms and > >exponents. Any pointers would be helpful. > > > >Thanks and Regards, > >-Anoop Deoras
Reply by ●August 25, 20042004-08-25
"Dirk Bell" <dirkman@erols.com> wrote in message news:6721a858.0408251532.725e6170@posting.google.com...> Cordic for logarithms? Got a reference for that? > > Dirk >http://www.magic-software.com/Documentation/CordicMethods.pdf Gives a set of formulas for sin(x), cos(x), atan(x/y), sinh(x), cosh(x) and atanh(x/y) http://www.xilinx.co.jp/products/software/sysgen/app_docs/user_guide_Chapter_10_Section_4_Subsection_25.htm points out the identity log (w) = 2 x atanh[(w-1) / (w+1)] Best wishes, --Phil Martel> george.w.bush@whitehouse.com (George Bush) wrote in messagenews:<%GnVc.8231$Bt5.4288@twister.socal.rr.com>...> > For hardware implementations, you might look at CORDIC. CORDIC wouldn'twork> > well for general purpose or DSPs. > > > > In article <2f592dd3.0408200028.21978843@posting.google.com>, > > akd_ecc@yahoo.com (EC-AKD) wrote: > > >Hi, > > > > > > I am looking for some good fixed point algorithms for logarithms and > > >exponents. Any pointers would be helpful. > > > > > >Thanks and Regards, > > >-Anoop Deoras
Reply by ●August 25, 20042004-08-25
Sounds pretty much the same except mine had some jumps. Do a cycle count and compare the number of cycles with the normal implementation. It's worse with a pipelined processor. In article <sxd4qmr402e.fsf@edgedsp4.rtp.ericsson.se>, Raymond Toy <raymond.toy@ericsson.com> wrote:>>>>>> "George" == George Bush <george.w.bush@whitehouse.com> writes: > > George> CORDIC uses bit operations which aren't effecient in general > purpose > George> operations. > >Your CORDIC must be different from mine. All I had were compares, >adds, subtracts, shifts, and table lookups. Sounds pretty general >purpose to me. > >Ray
Reply by ●August 26, 20042004-08-26
>>>>> "George" == George Bush <george.w.bush@whitehouse.com> writes:George> Sounds pretty much the same except mine had some jumps. Do a cycle count and George> compare the number of cycles with the normal implementation. It's worse with George> a pipelined processor. Well, yes of course. But you said bit operations. I don't consider a branch a bit operation. But, yes, branches hurt on pipelined processors, especially since you probably can't predict the branch in a CORDIC routine. Ray
Reply by ●August 28, 20042004-08-28
Where do you think binary weights come from? In article <sxdvff53nel.fsf@edgedsp4.rtp.ericsson.se>, Raymond Toy <raymond.toy@ericsson.com> wrote:>>>>>> "George" == George Bush <george.w.bush@whitehouse.com> writes: > > George> Sounds pretty much the same except mine had some jumps. Do a cycle > count and > George> compare the number of cycles with the normal implementation. It's > worse with > George> a pipelined processor. > >Well, yes of course. But you said bit operations. I don't consider a >branch a bit operation. > >But, yes, branches hurt on pipelined processors, especially since you >probably can't predict the branch in a CORDIC routine. > >Ray