00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00022
00023
00024
00025
00026 #ifndef EIGEN_JACOBI_H
00027 #define EIGEN_JACOBI_H
00028
00047 template<typename Scalar> class JacobiRotation
00048 {
00049 public:
00050 typedef typename NumTraits<Scalar>::Real RealScalar;
00051
00053 JacobiRotation() {}
00054
00056 JacobiRotation(const Scalar& c, const Scalar& s) : m_c(c), m_s(s) {}
00057
00058 Scalar& c() { return m_c; }
00059 Scalar c() const { return m_c; }
00060 Scalar& s() { return m_s; }
00061 Scalar s() const { return m_s; }
00062
00064 JacobiRotation operator*(const JacobiRotation& other)
00065 {
00066 return JacobiRotation(m_c * other.m_c - internal::conj(m_s) * other.m_s,
00067 internal::conj(m_c * internal::conj(other.m_s) + internal::conj(m_s) * internal::conj(other.m_c)));
00068 }
00069
00071 JacobiRotation transpose() const { return JacobiRotation(m_c, -internal::conj(m_s)); }
00072
00074 JacobiRotation adjoint() const { return JacobiRotation(internal::conj(m_c), -m_s); }
00075
00076 template<typename Derived>
00077 bool makeJacobi(const MatrixBase<Derived>&, typename Derived::Index p, typename Derived::Index q);
00078 bool makeJacobi(RealScalar x, Scalar y, RealScalar z);
00079
00080 void makeGivens(const Scalar& p, const Scalar& q, Scalar* z=0);
00081
00082 protected:
00083 void makeGivens(const Scalar& p, const Scalar& q, Scalar* z, internal::true_type);
00084 void makeGivens(const Scalar& p, const Scalar& q, Scalar* z, internal::false_type);
00085
00086 Scalar m_c, m_s;
00087 };
00088
00094 template<typename Scalar>
00095 bool JacobiRotation<Scalar>::makeJacobi(RealScalar x, Scalar y, RealScalar z)
00096 {
00097 typedef typename NumTraits<Scalar>::Real RealScalar;
00098 if(y == Scalar(0))
00099 {
00100 m_c = Scalar(1);
00101 m_s = Scalar(0);
00102 return false;
00103 }
00104 else
00105 {
00106 RealScalar tau = (x-z)/(RealScalar(2)*internal::abs(y));
00107 RealScalar w = internal::sqrt(internal::abs2(tau) + 1);
00108 RealScalar t;
00109 if(tau>0)
00110 {
00111 t = RealScalar(1) / (tau + w);
00112 }
00113 else
00114 {
00115 t = RealScalar(1) / (tau - w);
00116 }
00117 RealScalar sign_t = t > 0 ? 1 : -1;
00118 RealScalar n = RealScalar(1) / internal::sqrt(internal::abs2(t)+1);
00119 m_s = - sign_t * (internal::conj(y) / internal::abs(y)) * internal::abs(t) * n;
00120 m_c = n;
00121 return true;
00122 }
00123 }
00124
00134 template<typename Scalar>
00135 template<typename Derived>
00136 inline bool JacobiRotation<Scalar>::makeJacobi(const MatrixBase<Derived>& m, typename Derived::Index p, typename Derived::Index q)
00137 {
00138 return makeJacobi(internal::real(m.coeff(p,p)), m.coeff(p,q), internal::real(m.coeff(q,q)));
00139 }
00140
00157 template<typename Scalar>
00158 void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* z)
00159 {
00160 makeGivens(p, q, z, typename internal::conditional<NumTraits<Scalar>::IsComplex, internal::true_type, internal::false_type>::type());
00161 }
00162
00163
00164
00165 template<typename Scalar>
00166 void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* r, internal::true_type)
00167 {
00168 if(q==Scalar(0))
00169 {
00170 m_c = internal::real(p)<0 ? Scalar(-1) : Scalar(1);
00171 m_s = 0;
00172 if(r) *r = m_c * p;
00173 }
00174 else if(p==Scalar(0))
00175 {
00176 m_c = 0;
00177 m_s = -q/internal::abs(q);
00178 if(r) *r = internal::abs(q);
00179 }
00180 else
00181 {
00182 RealScalar p1 = internal::norm1(p);
00183 RealScalar q1 = internal::norm1(q);
00184 if(p1>=q1)
00185 {
00186 Scalar ps = p / p1;
00187 RealScalar p2 = internal::abs2(ps);
00188 Scalar qs = q / p1;
00189 RealScalar q2 = internal::abs2(qs);
00190
00191 RealScalar u = internal::sqrt(RealScalar(1) + q2/p2);
00192 if(internal::real(p)<RealScalar(0))
00193 u = -u;
00194
00195 m_c = Scalar(1)/u;
00196 m_s = -qs*internal::conj(ps)*(m_c/p2);
00197 if(r) *r = p * u;
00198 }
00199 else
00200 {
00201 Scalar ps = p / q1;
00202 RealScalar p2 = internal::abs2(ps);
00203 Scalar qs = q / q1;
00204 RealScalar q2 = internal::abs2(qs);
00205
00206 RealScalar u = q1 * internal::sqrt(p2 + q2);
00207 if(internal::real(p)<RealScalar(0))
00208 u = -u;
00209
00210 p1 = internal::abs(p);
00211 ps = p/p1;
00212 m_c = p1/u;
00213 m_s = -internal::conj(ps) * (q/u);
00214 if(r) *r = ps * u;
00215 }
00216 }
00217 }
00218
00219
00220 template<typename Scalar>
00221 void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* r, internal::false_type)
00222 {
00223
00224 if(q==0)
00225 {
00226 m_c = p<Scalar(0) ? Scalar(-1) : Scalar(1);
00227 m_s = 0;
00228 if(r) *r = internal::abs(p);
00229 }
00230 else if(p==0)
00231 {
00232 m_c = 0;
00233 m_s = q<Scalar(0) ? Scalar(1) : Scalar(-1);
00234 if(r) *r = internal::abs(q);
00235 }
00236 else if(internal::abs(p) > internal::abs(q))
00237 {
00238 Scalar t = q/p;
00239 Scalar u = internal::sqrt(Scalar(1) + internal::abs2(t));
00240 if(p<Scalar(0))
00241 u = -u;
00242 m_c = Scalar(1)/u;
00243 m_s = -t * m_c;
00244 if(r) *r = p * u;
00245 }
00246 else
00247 {
00248 Scalar t = p/q;
00249 Scalar u = internal::sqrt(Scalar(1) + internal::abs2(t));
00250 if(q<Scalar(0))
00251 u = -u;
00252 m_s = -Scalar(1)/u;
00253 m_c = -t * m_s;
00254 if(r) *r = q * u;
00255 }
00256
00257 }
00258
00259
00260
00261
00262
00269 namespace internal {
00270 template<typename VectorX, typename VectorY, typename OtherScalar>
00271 void apply_rotation_in_the_plane(VectorX& _x, VectorY& _y, const JacobiRotation<OtherScalar>& j);
00272 }
00273
00280 template<typename Derived>
00281 template<typename OtherScalar>
00282 inline void MatrixBase<Derived>::applyOnTheLeft(Index p, Index q, const JacobiRotation<OtherScalar>& j)
00283 {
00284 RowXpr x(this->row(p));
00285 RowXpr y(this->row(q));
00286 internal::apply_rotation_in_the_plane(x, y, j);
00287 }
00288
00295 template<typename Derived>
00296 template<typename OtherScalar>
00297 inline void MatrixBase<Derived>::applyOnTheRight(Index p, Index q, const JacobiRotation<OtherScalar>& j)
00298 {
00299 ColXpr x(this->col(p));
00300 ColXpr y(this->col(q));
00301 internal::apply_rotation_in_the_plane(x, y, j.transpose());
00302 }
00303
00304 namespace internal {
00305 template<typename VectorX, typename VectorY, typename OtherScalar>
00306 void apply_rotation_in_the_plane(VectorX& _x, VectorY& _y, const JacobiRotation<OtherScalar>& j)
00307 {
00308 typedef typename VectorX::Index Index;
00309 typedef typename VectorX::Scalar Scalar;
00310 enum { PacketSize = packet_traits<Scalar>::size };
00311 typedef typename packet_traits<Scalar>::type Packet;
00312 eigen_assert(_x.size() == _y.size());
00313 Index size = _x.size();
00314 Index incrx = _x.innerStride();
00315 Index incry = _y.innerStride();
00316
00317 Scalar* EIGEN_RESTRICT x = &_x.coeffRef(0);
00318 Scalar* EIGEN_RESTRICT y = &_y.coeffRef(0);
00319
00320
00321
00322 if(VectorX::SizeAtCompileTime == Dynamic &&
00323 (VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
00324 ((incrx==1 && incry==1) || PacketSize == 1))
00325 {
00326
00327 enum { Peeling = 2 };
00328
00329 Index alignedStart = first_aligned(y, size);
00330 Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize;
00331
00332 const Packet pc = pset1<Packet>(j.c());
00333 const Packet ps = pset1<Packet>(j.s());
00334 conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex,false> pcj;
00335
00336 for(Index i=0; i<alignedStart; ++i)
00337 {
00338 Scalar xi = x[i];
00339 Scalar yi = y[i];
00340 x[i] = j.c() * xi + conj(j.s()) * yi;
00341 y[i] = -j.s() * xi + conj(j.c()) * yi;
00342 }
00343
00344 Scalar* EIGEN_RESTRICT px = x + alignedStart;
00345 Scalar* EIGEN_RESTRICT py = y + alignedStart;
00346
00347 if(first_aligned(x, size)==alignedStart)
00348 {
00349 for(Index i=alignedStart; i<alignedEnd; i+=PacketSize)
00350 {
00351 Packet xi = pload<Packet>(px);
00352 Packet yi = pload<Packet>(py);
00353 pstore(px, padd(pmul(pc,xi),pcj.pmul(ps,yi)));
00354 pstore(py, psub(pcj.pmul(pc,yi),pmul(ps,xi)));
00355 px += PacketSize;
00356 py += PacketSize;
00357 }
00358 }
00359 else
00360 {
00361 Index peelingEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize);
00362 for(Index i=alignedStart; i<peelingEnd; i+=Peeling*PacketSize)
00363 {
00364 Packet xi = ploadu<Packet>(px);
00365 Packet xi1 = ploadu<Packet>(px+PacketSize);
00366 Packet yi = pload <Packet>(py);
00367 Packet yi1 = pload <Packet>(py+PacketSize);
00368 pstoreu(px, padd(pmul(pc,xi),pcj.pmul(ps,yi)));
00369 pstoreu(px+PacketSize, padd(pmul(pc,xi1),pcj.pmul(ps,yi1)));
00370 pstore (py, psub(pcj.pmul(pc,yi),pmul(ps,xi)));
00371 pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pmul(ps,xi1)));
00372 px += Peeling*PacketSize;
00373 py += Peeling*PacketSize;
00374 }
00375 if(alignedEnd!=peelingEnd)
00376 {
00377 Packet xi = ploadu<Packet>(x+peelingEnd);
00378 Packet yi = pload <Packet>(y+peelingEnd);
00379 pstoreu(x+peelingEnd, padd(pmul(pc,xi),pcj.pmul(ps,yi)));
00380 pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pmul(ps,xi)));
00381 }
00382 }
00383
00384 for(Index i=alignedEnd; i<size; ++i)
00385 {
00386 Scalar xi = x[i];
00387 Scalar yi = y[i];
00388 x[i] = j.c() * xi + conj(j.s()) * yi;
00389 y[i] = -j.s() * xi + conj(j.c()) * yi;
00390 }
00391 }
00392
00393
00394 else if(VectorX::SizeAtCompileTime != Dynamic &&
00395 (VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
00396 (VectorX::Flags & VectorY::Flags & AlignedBit))
00397 {
00398 const Packet pc = pset1<Packet>(j.c());
00399 const Packet ps = pset1<Packet>(j.s());
00400 conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex,false> pcj;
00401 Scalar* EIGEN_RESTRICT px = x;
00402 Scalar* EIGEN_RESTRICT py = y;
00403 for(Index i=0; i<size; i+=PacketSize)
00404 {
00405 Packet xi = pload<Packet>(px);
00406 Packet yi = pload<Packet>(py);
00407 pstore(px, padd(pmul(pc,xi),pcj.pmul(ps,yi)));
00408 pstore(py, psub(pcj.pmul(pc,yi),pmul(ps,xi)));
00409 px += PacketSize;
00410 py += PacketSize;
00411 }
00412 }
00413
00414
00415 else
00416 {
00417 for(Index i=0; i<size; ++i)
00418 {
00419 Scalar xi = *x;
00420 Scalar yi = *y;
00421 *x = j.c() * xi + conj(j.s()) * yi;
00422 *y = -j.s() * xi + conj(j.c()) * yi;
00423 x += incrx;
00424 y += incry;
00425 }
00426 }
00427 }
00428 }
00429
00430 #endif // EIGEN_JACOBI_H