diff --git a/lib/superfox/ftrsd3.f90 b/lib/superfox/ftrsd3.f90 new file mode 100644 index 000000000..62267c04e --- /dev/null +++ b/lib/superfox/ftrsd3.f90 @@ -0,0 +1,188 @@ +subroutine ftrsd3(rxdat,rxprob,rxdat2,rxprob2,ntrials0,correct,param,ntry) + +! Soft-decision decoder for Reed-Solomon codes. + +! This decoding scheme is built around Phil Karn's Berlekamp-Massey +! errors and erasures decoder. The approach is inspired by a number of +! publications, including the stochastic Chase decoder described +! in "Stochastic Chase Decoding of Reed-Solomon Codes", by Leroux et al., +! IEEE Communications Letters, Vol. 14, No. 9, September 2010 and +! "Soft-Decision Decoding of Reed-Solomon Codes Using Successive Error- +! and-Erasure Decoding," by Soo-Woong Lee and B. V. K. Vijaya Kumar. + +! Steve Franke K9AN and Joe Taylor K1JT + + use sfox_mod + + integer, dimension(0:NN-1) :: rxdat,rxprob,rxdat2,rxprob2,workdat, & + correct,indexes + integer rxdat(0:NN-1) !Hard-decision symbol values + integer rxprob(0:NN-1) !Probabilities that rxdat values are correct + integer rxdat2(0:NN-1) !Second most probable symbol values + integer rxprob2(0:NN-1) !Probabilities that rxdat2 values are correct + integer workdat(0:NN-1) !Work array + integer correct(0:NN-1) !Corrected codeword + integer indexes(0:NN-1) !For sorting probabilities + integer probs(0:NN-1) !Temp array for sorting probabilities + integer thresh0(0:NN-1) !Temp array for thresholds + integer era_pos(0:NN-KK-1) !Index values for erasures + integer*8 nseed,ir !No unsigned int in Fortran + integer pass,tmp + + integer perr(0:7,0:7) + data perr/ 4, 9,11,13,14,14,15,15, & + 2,20,20,30,40,50,50,50, & + 7,24,27,40,50,50,50,50, & + 13,25,35,46,52,70,50,50, & + 17,30,42,54,55,64,71,70, & + 25,39,48,57,64,66,77,77, & + 32,45,54,63,66,75,78,83, & + 51,58,57,66,72,77,82,86/ + + ntrials=ntrials0 + nhard=0 + nhard_min=32768 + nsoft=0 + nsoft_min=32768 + ntotal=0 + ntotal_min=32768 + nera_best=0 + nsym=nn + + do i=0,NN-1 + indexes(i)=i + probs(i)=rxprob(i) + enddo + + do pass=1,nsym-1 + do k=0,nsym-pass-1 + if(probs(k).lt.probs(k+1)) then + probs(k)=probs(k+1) + probs(k+1)=tmp + tmp=indexes(k) + indexes(k)=indexes(k+1) + indexes(k+1)=tmp + enddo + enddo + enddo + + era_pos=0 + numera=0 + workdat=rxdat + call rs_decode_sf(workdat,era_pos,numera,nerr) !Call the decoder + + if(nerr.ge.0) then +! Hard-decision decoding succeeded. Save codeword and some parameters. + nhard=count(workdat.ne.rxdat) + correct=workdat + param(0)=0 + param(1)=nhard + param(2)=0 + param(3)=0 + param(4)=0 + param(5)=0 + param(7)=1000*1000 !??? + ntry=0 + return + endif + +! Hard-decision decoding failed. Try the FT soft-decision method. +! Generate random erasure-locator vectors and see if any of them +! decode. This will generate a list of "candidate" codewords. The +! soft distance between each candidate codeword and the received +! word is estimated by finding the largest (pp1) and second-largest +! (pp2) outputs from a synchronized filter-bank operating on the +! symbol spectra, and using these to decide which candidate +! codeword is "best". + + nseed=1 !Seed for random numbers + ncandidates=0 + nsum=0 + do i=0,NN-1 + nsum=nsum+rxprob(i) + j=indexes(NN-1-i) + ratio=(float)rxprob2(j)/((float)rxprob(j)+0.01) + ii=7.999*ratio + jj=(NN-1-i)/8 + thresh0(i)=1.3*perr(jj)(ii) + enddo + if(nsum.le.0) return + + pp1=0. + pp2=0. + do k=1,ntrials + era_pos=0 + workdat=rxdat + +! Mark a subset of the symbols as erasures. +! Run through the ranked symbols, starting with the worst, i=0. +! NB: j is the symbol-vector index of the symbol with rank i. + + numera=0 + do i=0,NN-1 + j=indexes(126-i) + thresh=thresh0(i) + +! Generate a random number ir, 0 <= ir <= 100 (see POSIX.1-2001 example). + nseed=nseed*1103515245 + 12345 + ir=mod(nseed/65536),32768) + ir=(100*ir)/32768 + nseed=iand(ir,4294967295) + + if((ir.lt.thresh ) .and. numera.lt.(NN-KK)) then + era_pos(numera)=j + numera=numera+1 + endif + enddo + +! nerr=decode_rs_int(rs,workdat,era_pos,numera,0); + call rs_decode_sf(workdat,era_pos,numera,nerr) !Call the decoder + + if( nerr.ge.0) then + ! We have a candidate codeword. Find its hard and soft distance from + ! the received word. Also find pp1 and pp2 from the full array + ! s3(64,127) of synchronized symbol spectra. + ncandidates=ncandidates+1 + nhard=0 + nsoft=0 + do i=0,NN-1 + if(workdat(i).ne. rxdat(i)) then + nhard=nhard+1; + if(workdat(i) .ne. rxdat2(i)) nsoft=nsoft+rxprob(i) + endif + enddo + nsoft=(NN-1)*nsoft/nsum + ntotal=nsoft+nhard + + pp=0. + call getpp3(s3,workdat,pp) + if(pp.gt.pp1) then + pp2=pp1 + pp1=pp + nsoft_min=nsoft + nhard_min=nhard + ntotal_min=ntotal + correct=workdat + nera_best=numera + ntry=k + else + if(pp.gt.pp2 .and. pp.ne.pp1) pp2=pp + endif + if(nhard_min <= 41 && ntotal_min <= 71) exit !### New values ### + enddo + if(k.eq.ntrials) ntry=k + enddo + + param(0)=ncandidates + param(1)=nhard_min + param(2)=nsoft_min + param(3)=nera_best +! param(4)= pp1 > 0 ? 1000.0*pp2/pp1 : 1000.0 + param(5)=ntotal_min + param(6)=ntry + param(7)=1000.0*pp2 + param(8)=1000.0*pp1 + if(param(0).eq.0) param(2)=-1 + + return +end subroutine ftrsd3 diff --git a/lib/superfox/sfox_mod.f90 b/lib/superfox/sfox_mod.f90 index 64cf1c5ec..724f8c188 100644 --- a/lib/superfox/sfox_mod.f90 +++ b/lib/superfox/sfox_mod.f90 @@ -13,7 +13,6 @@ contains 1280,1296,1344,1350,1372,1400,1440,1470,1500,1512, & 1536,1568,1600,1620,1680,1728,1920,1944,1960,2000/ - MM=mm0 !Bits per symbol NQ=2**MM !Q, number of MFSK tones NN=nn0 !Channel symbols, before puncture