/* === S Y N F I G ========================================================= */ /*! \file blineconvert.cpp ** \brief Template File ** ** $Id: blineconvert.cpp 465 2007-04-12 15:10:11Z dooglus $ ** ** \legal ** Copyright (c) 2002-2005 Robert B. Quattlebaum Jr., Adrian Bentley ** ** This package is free software; you can redistribute it and/or ** modify it under the terms of the GNU General Public License as ** published by the Free Software Foundation; either version 2 of ** the License, or (at your option) any later version. ** ** This package is distributed in the hope that it will be useful, ** but WITHOUT ANY WARRANTY; without even the implied warranty of ** MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU ** General Public License for more details. ** \endlegal */ /* ========================================================================= */ /* === H E A D E R S ======================================================= */ #ifdef USING_PCH # include "pch.h" #else #ifdef HAVE_CONFIG_H # include #endif #include "blineconvert.h" #include #include #include #include #include #include #include #include #endif /* === U S I N G =========================================================== */ using namespace std; using namespace etl; using namespace synfig; /* === M A C R O S ========================================================= */ #define EPSILON (1e-10) /* === G L O B A L S ======================================================= */ /* === P R O C E D U R E S ================================================= */ /* === M E T H O D S ======================================================= */ //Derivative Functions for numerical approximation //bias == 0 will get F' at f3, bias < 0 will get F' at f1, and bias > 0 will get F' at f5 template < class T > inline void FivePointdt(T &df, const T &f1, const T &f2, const T &f3, const T &f4, const T &f5, int bias) { if(bias == 0) { //middle df = (f1 - f2*8 + f4*8 - f5)*(1/12.0f); }else if(bias < 0) { //left df = (-f1*25 + f2*48 - f3*36 + f4*16 - f5*3)*(1/12.0f); }else { //right df = (f1*3 - f2*16 + f3*36 - f4*48 + f5*25)*(1/12.0f); } } template < class T > inline void ThreePointdt(T &df, const T &f1, const T &f2, const T &f3, int bias) { if(bias == 0) { //middle df = (-f1 + f3)*(1/2.0f); }else if(bias < 0) { //left df = (-f1*3 + f2*4 - f3)*(1/2.0f); }else { //right df = (f1 - f2*4 + f3*3)*(1/2.0f); } } template < class T > inline void ThreePointddt(T &df, const T &f1, const T &f2, const T &f3, int bias) { //a 3 point approximation pretends to have constant acceleration, so only one algorithm needed for left, middle, or right df = (f1 -f2*2 + f3)*(1/2.0f); } // WARNING -- totaly broken template < class T > inline void FivePointddt(T &df, const T &f1, const T &f2, const T &f3, int bias) { if(bias == 0) { assert(0); // !? //middle //df = (- f1 + f2*16 - f3*30 + f4*16 - f5)*(1/12.0f); }/*else if(bias < 0) { //left df = (f1*7 - f2*26*4 + f3*19*6 - f4*14*4 + f5*11)*(1/12.0f); }else { //right df = (f1*3 - f2*16 + f3*36 - f4*48 + f5*25)*(1/12.0f); }*/ //side ones don't work, use 3 point } //implement an arbitrary derivative //dumb algorithm template < class T > void DerivativeApprox(T &df, const T f[], const Real t[], int npoints, int indexval) { /* Lj(x) = PI_i!=j (x - xi) / PI_i!=j (xj - xi) so Lj'(x) = SUM_k PI_i!=j|k (x - xi) / PI_i!=j (xj - xi) */ unsigned int i,j,k,i0,i1; Real Lpj,mult,div,tj; Real tval = t[indexval]; //sum k for(j=0;j inline int sign(T f, T tol) { if(f < -tol) return -1; if(f > tol) return 1; return 0; } void GetFirstDerivatives(const std::vector &f, unsigned int left, unsigned int right, char *out, unsigned int dfstride) { unsigned int current = left; if(right - left < 2) return; else if(right - left < 3) { synfig::Vector v = f[left+1] - f[left]; //set both to the one we want *(synfig::Vector*)out = v; out += dfstride; *(synfig::Vector*)out = v; out += dfstride; } else if(right - left < 6/*5*/) //should use 3 point { //left then middle then right ThreePointdt(*(synfig::Vector*)out,f[left+0], f[left+1], f[left+2], -1); current += 1; out += dfstride; for(;current < right-1; current++, out += dfstride) { ThreePointdt(*(synfig::Vector*)out,f[current-1], f[current], f[current+1], 0); } ThreePointdt(*(synfig::Vector*)out,f[right-3], f[right-2], f[right-1], 1); current++; out += dfstride; }else //can use 5 point { //left 2 then middle bunch then right two //may want to use 3 point for inner edge ones FivePointdt(*(synfig::Vector*)out,f[left+0], f[left+1], f[left+2], f[left+3], f[left+4], -2); out += dfstride; FivePointdt(*(synfig::Vector*)out,f[left+1], f[left+2], f[left+3], f[left+4], f[left+5], -1); out += dfstride; current += 2; for(;current < right-2; current++, out += dfstride) { FivePointdt(*(synfig::Vector*)out,f[current-2], f[current-1], f[current], f[current+1], f[current+2], 0); } FivePointdt(*(synfig::Vector*)out,f[right-5], f[right-4], f[right-3], f[right-2], f[right-1], 1); out += dfstride; FivePointdt(*(synfig::Vector*)out,f[right-6], f[right-5], f[right-4], f[right-3], f[right-2], 2); out += dfstride; current += 2; } } void GetSimpleDerivatives(const std::vector &f, int left, int right, std::vector &df, int outleft, const std::vector &di) { int i1,i2,i; int offset = 2; //df = 1/2 (f[i+o]-f[i-o]) assert((int)df.size() >= right-left+outleft); //must be big enough for(i = left; i < right; ++i) { //right now indices (figure out distance later) i1 = std::max(left,i-offset); i2 = std::max(left,i+offset); df[outleft++] = (f[i2] - f[i1])*0.5f; } } //get the curve error from the double sample list of work points (hopefully that's enough) Real CurveError(const synfig::Point *pts, unsigned int n, std::vector &work, int left, int right) { if(right-left < 2) return -1; int i,j; //get distances to each point Real d,dtemp,dsum; //synfig::Vector v,vt; //synfig::Point p1,p2; synfig::Point pi; std::vector::const_iterator it;//,end = work.begin()+right; //unsigned int size = work.size(); //for each line, get distance d = 0; //starts at 0 for(i = 0; i < (int)n; ++i) { pi = pts[i]; dsum = FLT_MAX; it = work.begin()+left; //p2 = *it++; //put it at left+1 for(j = left/*+1*/; j < right; ++j,++it) { /*p1 = p2; p2 = *it; v = p2 - p1; vt = pi - p1; dtemp = v.mag_squared() > 1e-12 ? (vt*v)/v.mag_squared() : 0; //get the projected time value for the current line //get distance to line segment with the time value clamped 0-1 if(dtemp >= 1) //use p+v { vt += v; //makes it pp - (p+v) }else if(dtemp > 0) //use vt-proj { vt -= v*dtemp; // vt - proj_v(vt) //must normalize the projection vector to work } //else use p dtemp = vt.mag_squared();*/ dtemp = (pi - *it).mag_squared(); if(dtemp < dsum) dsum = dtemp; } //accumulate the points' min distance from the curve d += sqrt(dsum); } return d; } typedef synfigapp::BLineConverter::cpindex cpindex; //has the index data and the tangent scale data (relevant as it may be) int tesselate_curves(const std::vector &inds, const std::vector &f, const std::vector &df, std::vector &work) { if(inds.size() < 2) return 0; etl::hermite curve; int ntess = 0; std::vector::const_iterator j = inds.begin(),j2, end = inds.end(); unsigned int ibase = inds[0].curind; j2 = j++; for(; j != end; j2 = j++) { //if this curve has invalid error (in j) then retesselate its work points (requires reparametrization, etc.) if(j->error < 0) { //get the stepsize etc. for the number of points in here unsigned int n = j->curind - j2->curind + 1; //thats the number of points in the span unsigned int k, kend, i0, i3; //so reset the right chunk Real t, dt = 1/(Real)(n*2-2); //assuming that they own only n points //start at first intermediate t = 0; i0 = j2->curind; i3 = j->curind; k = (i0-ibase)*2; //start on first intermediary point (2x+1) kend = (i3-ibase)*2; //last point to set (not intermediary) //build hermite curve, it's easier curve.p1() = f[i0]; curve.p2() = f[i3]; curve.t1() = df[i0]*(df[i0].mag_squared() > 1e-4 ? j2->tangentscale/df[i0].mag() : j2->tangentscale); curve.t2() = df[i3]*(df[i3].mag_squared() > 1e-4 ? j->tangentscale/df[i3].mag() : j->tangentscale); curve.sync(); //MUST include the end point (since we are ignoring left one) for(; k < kend; ++k, t += dt) { work[k] = curve(t); } work[k] = curve(1); //k == kend, t == 1 -> c(t) == p2 ++ntess; } } return ntess; } synfigapp::BLineConverter::BLineConverter() { pixelwidth = 1; smoothness = 0.70f; width = 0; }; void synfigapp::BLineConverter::clear() { f.clear(); f_w.clear(); ftemp.clear(); df.clear(); cvt.clear(); brk.clear(); di.clear(); d_i.clear(); work.clear(); curind.clear(); } void synfigapp::BLineConverter::operator () (std::list &out, const std::list &in,const std::list &in_w) { //Profiling information /*etl::clock::value_type initialprocess=0, curveval=0, breakeval=0, disteval=0; etl::clock::value_type preproceval=0, tesseval=0, erroreval=0, spliteval=0; unsigned int numpre=0, numtess=0, numerror=0, numsplit=0; etl::clock_realtime timer,total;*/ //total.reset(); if(in.size()<=1) return; clear(); //removing digitization error harder than expected //intended to fix little pen errors caused by the way people draw (tiny juts in opposite direction) //Different solutions // Average at both end points (will probably eliminate many points at each end of the samples) // Average after the break points are found (weird points would still affect the curve) // Just always get rid of breaks at the beginning and end if they are a certain distance apart // This is will be current approach so all we do now is try to remove duplicate points //remove duplicate points - very bad for fitting //timer.reset(); { std::list::const_iterator i = in.begin(), end = in.end(); std::list::const_iterator iw = in_w.begin(); synfig::Point c; if(in.size() == in_w.size()) { for(;i != end; ++i,++iw) { //eliminate duplicate points if(*i != c) { f.push_back(c = *i); f_w.push_back(*iw); } } }else { for(;i != end; ++i) { //eliminate duplicate points if(*i != c) { f.push_back(c = *i); } } } } //initialprocess = timer(); if(f.size()<=6) return; //get curvature information //timer.reset(); { int i,i0,i1; synfig::Vector v1,v2; cvt.resize(f.size()); cvt.front() = 1; cvt.back() = 1; for(i = 1; i < (int)f.size()-1; ++i) { i0 = std::max(0,i - 2); i1 = std::min((int)(f.size()-1),i + 2); v1 = f[i] - f[i0]; v2 = f[i1] - f[i]; cvt[i] = (v1*v2)/(v1.mag()*v2.mag()); } } //curveval = timer(); //synfig::info("calculated curvature"); //find corner points and interpolate inside those //timer.reset(); { //break at sharp derivative points //TODO tolerance should be set based upon digitization resolution (length dependent index selection) Real tol = 0; //break tolerance, for the cosine of the change in angle (really high curvature or something) Real fixdistsq = 4*width*width; //the distance to ignore breaks at the end points (for fixing stuff) unsigned int i = 0; int maxi = -1, last=0; Real minc = 1; brk.push_back(0); for(i = 1; i < cvt.size()-1; ++i) { //insert if too sharp (we need to break the tangents to insert onto the break list) if(cvt[i] < tol) { if(cvt[i] < minc) { minc = cvt[i]; maxi = i; } }else if(maxi >= 0) { if(maxi >= last + 8) { //synfig::info("break: %d-%d",maxi+1,cvt.size()); brk.push_back(maxi); last = maxi; } maxi = -1; minc = 1; } } brk.push_back(i); //postprocess for breaks too close to eachother Real d = 0; Point p = f[brk.front()]; //first set for(i = 1; i < brk.size()-1; ++i) //do not want to include end point... { d = (f[brk[i]] - p).mag_squared(); if(d > fixdistsq) break; //don't want to group breaks if we ever get over the dist... } //want to erase all points before... if(i != 1) brk.erase(brk.begin(),brk.begin()+i-1); //end set p = f[brk.back()]; for(i = brk.size()-2; i > 0; --i) //start at one in from the end { d = (f[brk[i]] - p).mag_squared(); if(d > fixdistsq) break; //don't want to group breaks if we ever get over the dist } if(i != brk.size()-2) brk.erase(brk.begin()+i+2,brk.end()); //erase all points that we found... found none if i has not advanced //must not include the one we ended up on } //breakeval = timer(); //synfig::info("found break points: %d",brk.size()); //get the distance calculation of the entire curve (for tangent scaling) //timer.reset(); { synfig::Point p1,p2; p1=p2=f[0]; di.resize(f.size()); d_i.resize(f.size()); Real d = 0; for(unsigned int i = 0; i < f.size();) { d += (d_i[i] = (p2-p1).mag()); di[i] = d; p1=p2; p2=f[++i]; } } //disteval = timer(); //synfig::info("calculated distance"); //now break at every point - calculate new derivatives each time //TODO //must be sure that the break points are 3 or more apart //then must also store the breaks which are not smooth, etc. //and figure out tangents between there //for each pair of break points (as long as they are far enough apart) recursively subdivide stuff //ignore the detected intermediate points { unsigned int i0=0,i3=0,is=0; int i=0,j=0; bool done = false; Real errortol = smoothness*pixelwidth; //???? what the hell should this value be BLinePoint a; synfig::Vector v; //intemp = f; //don't want to smooth out the corners bool breaktan = false, setwidth; a.set_split_tangent_flag(false); //a.set_width(width); a.set_width(1.0f); setwidth = (f.size() == f_w.size()); for(j = 0; j < (int)brk.size() - 1; ++j) { //for b[j] to b[j+1] subdivide and stuff i0 = brk[j]; i3 = brk[j+1]; unsigned int size = i3-i0+1; //must include the end points //new derivatives //timer.reset(); ftemp.assign(f.begin()+i0, f.begin()+i3+1); for(i=0;i<20;++i) gaussian_blur_3(ftemp.begin(),ftemp.end(),false); df.resize(size); GetFirstDerivatives(ftemp,0,size,(char*)&df[0],sizeof(df[0])); //GetSimpleDerivatives(ftemp,0,size,df,0,di); //< don't have to worry about indexing stuff as it is all being taken car of right now //preproceval += timer(); //numpre++; work.resize(size*2-1); //guarantee that all points will be tesselated correctly (one point inbetween every 2 adjacent points) //if size of work is size*2-1, the step size should be 1/(size*2 - 2) //Real step = 1/(Real)(size*2 - 1); //start off with break points as indices curind.clear(); curind.push_back(cpindex(i0,di[i3]-di[i0],0)); //0 error because no curve on the left curind.push_back(cpindex(i3,di[i3]-di[i0],-1)); //error needs to be reevaluated done = false; //we want to loop unsigned int dcount = 0; //while there are still enough points between us, and the error is too high subdivide (and invalidate neighbors that share tangents) while(!done) { //tesselate all curves with invalid error values work[0] = f[i0]; //timer.reset(); /*numtess += */tesselate_curves(curind,f,df,work); //tesseval += timer(); //now get all error values //timer.reset(); for(i = 1; i < (int)curind.size(); ++i) { if(curind[i].error < 0) //must have been retesselated, so now recalculate error value { //evaluate error from points (starting at current index) int size = curind[i].curind - curind[i-1].curind + 1; curind[i].error = CurveError(&f[curind[i-1].curind], size, work,(curind[i-1].curind - i0)*2,(curind[i].curind - i0)*2+1); /*if(curind[i].error > 1.0e5) { synfig::info("Holy crap %d-%d error %f",curind[i-1].curind,curind[i].curind,curind[i].error); curind[i].error = -1; numtess += tesselate_curves(curind,f,df,work); curind[i].error = CurveError(&f[curind[i-1].curind], size, work,0,work.size());//(curind[i-1].curind - i0)*2,(curind[i].curind - i0)*2+1); }*/ //numerror++; } } //erroreval += timer(); //assume we're done done = true; //check each error to see if it's too big, if so, then subdivide etc. int indsize = (int)curind.size(); Real maxrelerror = 0; int maxi = -1;//, numpoints; //timer.reset(); //get the maximum error and split there for(i = 1; i < indsize; ++i) { //numpoints = curind[i].curind - curind[i-1].curind + 1; if(curind[i].error > maxrelerror && curind[i].curind - curind[i-1].curind > 2) //only accept if it's valid { maxrelerror = curind[i].error; maxi = i; } } //split if error is too great if(maxrelerror > errortol) { //add one to the left etc unsigned int ibase = curind[maxi-1].curind, itop = curind[maxi].curind, ibreak = (ibase + itop)/2; Real scale, scale2; assert(ibreak < f.size()); //synfig::info("Split %d -%d- %d, error: %f", ibase,ibreak,itop,maxrelerror); if(ibase != itop) { //invalidate current error of the changed tangents and add an extra segment //enforce minimum tangents property curind[maxi].error = -1; curind[maxi-1].error = -1; if(maxi+1 < indsize) curind[maxi+1].error = -1; //if there is a curve segment beyond this it will be effected as well scale = di[itop] - di[ibreak]; scale2 = maxi+1 < indsize ? di[curind[maxi+1].curind] - di[itop] : scale; //to the right valid? curind[maxi].tangentscale = std::min(scale, scale2); scale = di[ibreak] - di[ibase]; scale2 = maxi >= 2 ? di[ibase] - di[curind[maxi-2].curind] : scale; // to the left valid -2 ? curind[maxi-1].tangentscale = std::min(scale, scale2); scale = std::min(di[ibreak] - di[ibase], di[itop] - di[ibreak]); curind.insert(curind.begin()+maxi,cpindex(ibreak, scale, -1)); //curind.push_back(cpindex(ibreak, scale, -1)); //std::sort(curind.begin(), curind.end()); done = false; //numsplit++; } } //spliteval += timer(); dcount++; } //insert the last point too (just set tangent for now is = curind[0].curind; //first point inherits current tangent status v = df[is - i0]; if(v.mag_squared() > EPSILON) v *= (curind[0].tangentscale/v.mag()); if(!breaktan) a.set_tangent(v); else a.set_tangent2(v); a.set_vertex(f[is]); if(setwidth)a.set_width(f_w[is]); out.push_back(a); a.set_split_tangent_flag(false); //won't need to break anymore breaktan = false; for(i = 1; i < (int)curind.size()-1; ++i) { is = curind[i].curind; //first point inherits current tangent status v = df[is-i0]; if(v.mag_squared() > EPSILON) v *= (curind[i].tangentscale/v.mag()); a.set_tangent(v); // always inside, so guaranteed to be smooth a.set_vertex(f[is]); if(setwidth)a.set_width(f_w[is]); out.push_back(a); } //set the last point's data is = curind.back().curind; //should already be this v = df[is-i0]; if(v.mag_squared() > EPSILON) v *= (curind.back().tangentscale/v.mag()); a.set_tangent1(v); a.set_split_tangent_flag(true); breaktan = true; //will get the vertex and tangent 2 from next round } a.set_vertex(f[i3]); a.set_split_tangent_flag(false); if(setwidth) a.set_width(f_w[i3]); out.push_back(a); /*etl::clock::value_type totaltime = total(), misctime = totaltime - initialprocess - curveval - breakeval - disteval - preproceval - tesseval - erroreval - spliteval; synfig::info( "Curve Convert Profile:\n" "\tInitial Preprocess: %f\n" "\tCurvature Calculation: %f\n" "\tBreak Calculation: %f\n" "\tDistance Calculation: %f\n" " Algorithm: (numtimes,totaltime)\n" "\tPreprocess step: (%d,%f)\n" "\tTesselation step: (%d,%f)\n" "\tError step: (%d,%f)\n" "\tSplit step: (%d,%f)\n" " Num Input: %d, Num Output: %d\n" " Total time: %f, Misc time: %f\n", initialprocess, curveval,breakeval,disteval, numpre,preproceval,numtess,tesseval,numerror,erroreval,numsplit,spliteval, in.size(),out.size(), totaltime,misctime);*/ return; } } void synfigapp::BLineConverter::EnforceMinWidth(std::list &bline, synfig::Real min_pressure) { std::list::iterator i = bline.begin(), end = bline.end(); for(i = bline.begin(); i != end; ++i) { if(i->get_width() < min_pressure) { i->set_width(min_pressure); } } }