| 1 |
// |
| 2 |
// SPDX-License-Identifier: BSD-3-Clause |
| 3 |
// Copyright Contributors to the OpenEXR Project. |
| 4 |
// |
| 5 |
|
| 6 |
/// |
| 7 |
/// @file ImathMatrixAlgo.cpp |
| 8 |
/// |
| 9 |
/// @brief Functions operating on Matrix22, Matrix33, and Matrix44 types. |
| 10 |
/// |
| 11 |
|
| 12 |
// clang-format off |
| 13 |
/// @cond Doxygen_Suppress |
| 14 |
|
| 15 |
#include "ImathMatrixAlgo.h" |
| 16 |
#include <algorithm> |
| 17 |
#include <cmath> |
| 18 |
|
| 19 |
#if defined(IMATH_DLL) |
| 20 |
# define EXPORT_CONST __declspec(dllexport) |
| 21 |
#else |
| 22 |
# define EXPORT_CONST const |
| 23 |
#endif |
| 24 |
|
| 25 |
IMATH_INTERNAL_NAMESPACE_SOURCE_ENTER |
| 26 |
|
| 27 |
EXPORT_CONST M22f identity22f ( 1, 0, |
| 28 |
0, 1); |
| 29 |
EXPORT_CONST M22d identity22d ( 1, 0, |
| 30 |
0, 1); |
| 31 |
EXPORT_CONST M33f identity33f ( 1, 0, 0, |
| 32 |
0, 1, 0, |
| 33 |
0, 0, 1); |
| 34 |
EXPORT_CONST M33d identity33d ( 1, 0, 0, |
| 35 |
0, 1, 0, |
| 36 |
0, 0, 1); |
| 37 |
EXPORT_CONST M44f identity44f ( 1, 0, 0, 0, |
| 38 |
0, 1, 0, 0, |
| 39 |
0, 0, 1, 0, |
| 40 |
0, 0, 0, 1); |
| 41 |
EXPORT_CONST M44d identity44d ( 1, 0, 0, 0, |
| 42 |
0, 1, 0, 0, |
| 43 |
0, 0, 1, 0, |
| 44 |
0, 0, 0, 1); |
| 45 |
|
| 46 |
// clang-format on |
| 47 |
|
| 48 |
namespace |
| 49 |
{ |
| 50 |
|
| 51 |
class KahanSum |
| 52 |
{ |
| 53 |
public: |
| 54 |
KahanSum() : _total (0), _correction (0) {} |
| 55 |
|
| 56 |
void operator+= (const double val) |
| 57 |
{ |
| 58 |
const double y = val - _correction; |
| 59 |
const double t = _total + y; |
| 60 |
_correction = (t - _total) - y; |
| 61 |
_total = t; |
| 62 |
} |
| 63 |
|
| 64 |
double get() const { return _total; } |
| 65 |
|
| 66 |
private: |
| 67 |
double _total; |
| 68 |
double _correction; |
| 69 |
}; |
| 70 |
|
| 71 |
} // namespace |
| 72 |
|
| 73 |
template <typename T> |
| 74 |
M44d |
| 75 |
procrustesRotationAndTranslation (const Vec3<T>* A, |
| 76 |
const Vec3<T>* B, |
| 77 |
const T* weights, |
| 78 |
const size_t numPoints, |
| 79 |
const bool doScale) |
| 80 |
{ |
| 81 |
if (numPoints == 0) |
| 82 |
return M44d(); |
| 83 |
|
| 84 |
// Always do the accumulation in double precision: |
| 85 |
V3d Acenter (0.0); |
| 86 |
V3d Bcenter (0.0); |
| 87 |
double weightsSum = 0.0; |
| 88 |
|
| 89 |
if (weights == 0) |
| 90 |
{ |
| 91 |
for (size_t i = 0; i < numPoints; ++i) |
| 92 |
{ |
| 93 |
Acenter += (V3d) A[i]; |
| 94 |
Bcenter += (V3d) B[i]; |
| 95 |
} |
| 96 |
weightsSum = (double) numPoints; |
| 97 |
} |
| 98 |
else |
| 99 |
{ |
| 100 |
for (size_t i = 0; i < numPoints; ++i) |
| 101 |
{ |
| 102 |
const double w = weights[i]; |
| 103 |
weightsSum += w; |
| 104 |
|
| 105 |
Acenter += w * (V3d) A[i]; |
| 106 |
Bcenter += w * (V3d) B[i]; |
| 107 |
} |
| 108 |
} |
| 109 |
|
| 110 |
if (weightsSum == 0) |
| 111 |
return M44d(); |
| 112 |
|
| 113 |
Acenter /= weightsSum; |
| 114 |
Bcenter /= weightsSum; |
| 115 |
|
| 116 |
// |
| 117 |
// Find Q such that |Q*A - B| (actually A-Acenter and B-Bcenter, weighted) |
| 118 |
// is minimized in the least squares sense. |
| 119 |
// From Golub/Van Loan, p.601 |
| 120 |
// |
| 121 |
// A,B are 3xn |
| 122 |
// Let C = B A^T (where A is 3xn and B^T is nx3, so C is 3x3) |
| 123 |
// Compute the SVD: C = U D V^T (U,V rotations, D diagonal). |
| 124 |
// Throw away the D part, and return Q = U V^T |
| 125 |
M33d C (0.0); |
| 126 |
if (weights == 0) |
| 127 |
{ |
| 128 |
for (size_t i = 0; i < numPoints; ++i) |
| 129 |
C += outerProduct ((V3d) B[i] - Bcenter, (V3d) A[i] - Acenter); |
| 130 |
} |
| 131 |
else |
| 132 |
{ |
| 133 |
for (size_t i = 0; i < numPoints; ++i) |
| 134 |
{ |
| 135 |
const double w = weights[i]; |
| 136 |
C += outerProduct (w * ((V3d) B[i] - Bcenter), (V3d) A[i] - Acenter); |
| 137 |
} |
| 138 |
} |
| 139 |
|
| 140 |
M33d U, V; |
| 141 |
V3d S; |
| 142 |
jacobiSVD (C, U, S, V, std::numeric_limits<double>::epsilon(), true); |
| 143 |
|
| 144 |
// We want Q.transposed() here since we are going to be using it in the |
| 145 |
// Imath style (multiplying vectors on the right, v' = v*A^T): |
| 146 |
const M33d Qt = V * U.transposed(); |
| 147 |
|
| 148 |
double s = 1.0; |
| 149 |
if (doScale && numPoints > 1) |
| 150 |
{ |
| 151 |
// Finding a uniform scale: let us assume the Q is completely fixed |
| 152 |
// at this point (solving for both simultaneously seems much harder). |
| 153 |
// We are trying to compute (again, per Golub and van Loan) |
| 154 |
// min || s*A*Q - B ||_F |
| 155 |
// Notice that we've jammed a uniform scale in front of the Q. |
| 156 |
// Now, the Frobenius norm (the least squares norm over matrices) |
| 157 |
// has the neat property that it is equivalent to minimizing the trace |
| 158 |
// of M^T*M (see your friendly neighborhood linear algebra text for a |
| 159 |
// derivation). Thus, we can expand this out as |
| 160 |
// min tr (s*A*Q - B)^T*(s*A*Q - B) |
| 161 |
// = min tr(Q^T*A^T*s*s*A*Q) + tr(B^T*B) - 2*tr(Q^T*A^T*s*B) by linearity of the trace |
| 162 |
// = min s^2 tr(A^T*A) + tr(B^T*B) - 2*s*tr(Q^T*A^T*B) using the fact that the trace is invariant |
| 163 |
// under similarity transforms Q*M*Q^T |
| 164 |
// If we differentiate w.r.t. s and set this to 0, we get |
| 165 |
// 0 = 2*s*tr(A^T*A) - 2*tr(Q^T*A^T*B) |
| 166 |
// so |
| 167 |
// 2*s*tr(A^T*A) = 2*s*tr(Q^T*A^T*B) |
| 168 |
// s = tr(Q^T*A^T*B) / tr(A^T*A) |
| 169 |
|
| 170 |
KahanSum traceATA; |
| 171 |
if (weights == 0) |
| 172 |
{ |
| 173 |
for (size_t i = 0; i < numPoints; ++i) |
| 174 |
traceATA += ((V3d) A[i] - Acenter).length2(); |
| 175 |
} |
| 176 |
else |
| 177 |
{ |
| 178 |
for (size_t i = 0; i < numPoints; ++i) |
| 179 |
traceATA += ((double) weights[i]) * ((V3d) A[i] - Acenter).length2(); |
| 180 |
} |
| 181 |
|
| 182 |
KahanSum traceBATQ; |
| 183 |
for (int i = 0; i < 3; ++i) |
| 184 |
for (int j = 0; j < 3; ++j) |
| 185 |
traceBATQ += Qt[j][i] * C[i][j]; |
| 186 |
|
| 187 |
s = traceBATQ.get() / traceATA.get(); |
| 188 |
} |
| 189 |
|
| 190 |
// Q is the rotation part of what we want to return. |
| 191 |
// The entire transform is: |
| 192 |
// (translate origin to Bcenter) * Q * (translate Acenter to origin) |
| 193 |
// last first |
| 194 |
// The effect of this on a point is: |
| 195 |
// (translate origin to Bcenter) * Q * (translate Acenter to origin) * point |
| 196 |
// = (translate origin to Bcenter) * Q * (-Acenter + point) |
| 197 |
// = (translate origin to Bcenter) * (-Q*Acenter + Q*point) |
| 198 |
// = (translate origin to Bcenter) * (translate Q*Acenter to origin) * Q*point |
| 199 |
// = (translate Q*Acenter to Bcenter) * Q*point |
| 200 |
// So what we want to return is: |
| 201 |
// (translate Q*Acenter to Bcenter) * Q |
| 202 |
// |
| 203 |
// In block form, this is: |
| 204 |
// [ 1 0 0 | ] [ 0 ] [ 1 0 0 | ] [ 1 0 0 | ] [ | ] [ ] |
| 205 |
// [ 0 1 0 tb ] [ s*Q 0 ] [ 0 1 0 -ta ] = [ 0 1 0 tb ] [ s*Q -s*Q*ta ] = [ Q tb-s*Q*ta ] |
| 206 |
// [ 0 0 1 | ] [ 0 ] [ 0 0 1 | ] [ 0 0 1 | ] [ | ] [ ] |
| 207 |
// [ 0 0 0 1 ] [ 0 0 0 1 ] [ 0 0 0 1 ] [ 0 0 0 1 ] [ 0 0 0 1 ] [ 0 0 0 1 ] |
| 208 |
// (ofc the whole thing is transposed for Imath). |
| 209 |
const V3d translate = Bcenter - s * Acenter * Qt; |
| 210 |
|
| 211 |
return M44d (s * Qt.x[0][0], |
| 212 |
s * Qt.x[0][1], |
| 213 |
s * Qt.x[0][2], |
| 214 |
T (0), |
| 215 |
s * Qt.x[1][0], |
| 216 |
s * Qt.x[1][1], |
| 217 |
s * Qt.x[1][2], |
| 218 |
T (0), |
| 219 |
s * Qt.x[2][0], |
| 220 |
s * Qt.x[2][1], |
| 221 |
s * Qt.x[2][2], |
| 222 |
T (0), |
| 223 |
translate.x, |
| 224 |
translate.y, |
| 225 |
translate.z, |
| 226 |
T (1)); |
| 227 |
} // procrustesRotationAndTranslation |
| 228 |
|
| 229 |
/// |
| 230 |
/// Return the procrustes transformation of a set of points: the |
| 231 |
/// rotation, translation, and optionally the scale that comes closest |
| 232 |
/// in a least squares sense to transforming the `A` points into |
| 233 |
/// `B`. |
| 234 |
/// |
| 235 |
|
| 236 |
template <typename T> |
| 237 |
M44d |
| 238 |
procrustesRotationAndTranslation (const Vec3<T>* A, |
| 239 |
const Vec3<T>* B, |
| 240 |
const size_t numPoints, |
| 241 |
const bool doScale) |
| 242 |
{ |
| 243 |
return procrustesRotationAndTranslation (A, B, (const T*) 0, numPoints, doScale); |
| 244 |
} // procrustesRotationAndTranslation |
| 245 |
|
| 246 |
/// TODO |
| 247 |
template IMATH_EXPORT M44d procrustesRotationAndTranslation (const V3d* from, |
| 248 |
const V3d* to, |
| 249 |
const size_t numPoints, |
| 250 |
const bool doScale); |
| 251 |
/// TODO |
| 252 |
template IMATH_EXPORT M44d procrustesRotationAndTranslation (const V3f* from, |
| 253 |
const V3f* to, |
| 254 |
const size_t numPoints, |
| 255 |
const bool doScale); |
| 256 |
/// TODO |
| 257 |
template IMATH_EXPORT M44d procrustesRotationAndTranslation (const V3d* from, |
| 258 |
const V3d* to, |
| 259 |
const double* weights, |
| 260 |
const size_t numPoints, |
| 261 |
const bool doScale); |
| 262 |
/// TODO |
| 263 |
template IMATH_EXPORT M44d procrustesRotationAndTranslation (const V3f* from, |
| 264 |
const V3f* to, |
| 265 |
const float* weights, |
| 266 |
const size_t numPoints, |
| 267 |
const bool doScale); |
| 268 |
|
| 269 |
namespace |
| 270 |
{ |
| 271 |
|
| 272 |
// Applies the 2x2 Jacobi rotation |
| 273 |
// [ c s 0 ] [ 1 0 0 ] [ c 0 s ] |
| 274 |
// [ -s c 0 ] or [ 0 c s ] or [ 0 1 0 ] |
| 275 |
// [ 0 0 1 ] [ 0 -s c ] [ -s 0 c ] |
| 276 |
// from the right; that is, computes |
| 277 |
// J * A |
| 278 |
// for the Jacobi rotation J and the matrix A. This is efficient because we |
| 279 |
// only need to touch exactly the 2 columns that are affected, so we never |
| 280 |
// need to explicitly construct the J matrix. |
| 281 |
template <typename T, int j, int k> |
| 282 |
void |
| 283 |
jacobiRotateRight (IMATH_INTERNAL_NAMESPACE::Matrix33<T>& A, const T c, const T s) |
| 284 |
{ |
| 285 |
for (int i = 0; i < 3; ++i) |
| 286 |
{ |
| 287 |
const T tau1 = A[i][j]; |
| 288 |
const T tau2 = A[i][k]; |
| 289 |
A[i][j] = c * tau1 - s * tau2; |
| 290 |
A[i][k] = s * tau1 + c * tau2; |
| 291 |
} |
| 292 |
} |
| 293 |
|
| 294 |
template <typename T> |
| 295 |
void |
| 296 |
jacobiRotateRight (IMATH_INTERNAL_NAMESPACE::Matrix44<T>& A, |
| 297 |
const int j, |
| 298 |
const int k, |
| 299 |
const T c, |
| 300 |
const T s) |
| 301 |
{ |
| 302 |
for (int i = 0; i < 4; ++i) |
| 303 |
{ |
| 304 |
const T tau1 = A[i][j]; |
| 305 |
const T tau2 = A[i][k]; |
| 306 |
A[i][j] = c * tau1 - s * tau2; |
| 307 |
A[i][k] = s * tau1 + c * tau2; |
| 308 |
} |
| 309 |
} |
| 310 |
|
| 311 |
// This routine solves the 2x2 SVD: |
| 312 |
// [ c1 s1 ] [ w x ] [ c2 s2 ] [ d1 0 ] |
| 313 |
// [ ] [ ] [ ] = [ ] |
| 314 |
// [ -s1 c1 ] [ y z ] [ -s2 c2 ] [ 0 d2 ] |
| 315 |
// where |
| 316 |
// [ w x ] |
| 317 |
// A = [ ] |
| 318 |
// [ y z ] |
| 319 |
// is the subset of A consisting of the [j,k] entries, A([j k], [j k]) in |
| 320 |
// Matlab parlance. The method is the 'USVD' algorithm described in the |
| 321 |
// following paper: |
| 322 |
// 'Computation of the Singular Value Decomposition using Mesh-Connected Processors' |
| 323 |
// by Richard P. Brent, Franklin T. Luk, and Charles Van Loan |
| 324 |
// It breaks the computation into two steps: the first symmetrizes the matrix, |
| 325 |
// and the second diagonalizes the symmetric matrix. |
| 326 |
template <typename T, int j, int k, int l> |
| 327 |
bool |
| 328 |
twoSidedJacobiRotation (IMATH_INTERNAL_NAMESPACE::Matrix33<T>& A, |
| 329 |
IMATH_INTERNAL_NAMESPACE::Matrix33<T>& U, |
| 330 |
IMATH_INTERNAL_NAMESPACE::Matrix33<T>& V, |
| 331 |
const T tol) |
| 332 |
{ |
| 333 |
// Load everything into local variables to make things easier on the |
| 334 |
// optimizer: |
| 335 |
const T w = A[j][j]; |
| 336 |
const T x = A[j][k]; |
| 337 |
const T y = A[k][j]; |
| 338 |
const T z = A[k][k]; |
| 339 |
|
| 340 |
// We will keep track of whether we're actually performing any rotations, |
| 341 |
// since if the matrix is already diagonal we'll end up with the identity |
| 342 |
// as our Jacobi rotation and we can short-circuit. |
| 343 |
bool changed = false; |
| 344 |
|
| 345 |
// The first step is to symmetrize the 2x2 matrix, |
| 346 |
// [ c s ]^T [ w x ] = [ p q ] |
| 347 |
// [ -s c ] [ y z ] [ q r ] |
| 348 |
T mu_1 = w + z; |
| 349 |
T mu_2 = x - y; |
| 350 |
|
| 351 |
T c, s; |
| 352 |
if (std::abs (mu_2) <= tol * std::abs (mu_1)) // Already symmetric (to tolerance) |
| 353 |
{ // Note that the <= is important here |
| 354 |
c = T (1); // because we want to bypass the computation |
| 355 |
s = T (0); // of rho if mu_1 = mu_2 = 0. |
| 356 |
|
| 357 |
const T p = w; |
| 358 |
const T r = z; |
| 359 |
mu_1 = r - p; |
| 360 |
mu_2 = x + y; |
| 361 |
} |
| 362 |
else |
| 363 |
{ |
| 364 |
// TODO is there a native inverse square root function? |
| 365 |
const T rho = mu_1 / mu_2; |
| 366 |
|
| 367 |
s = T (1) / std::sqrt (T (1) + rho * rho); |
| 368 |
if (rho < 0) |
| 369 |
s = -s; |
| 370 |
c = s * rho; |
| 371 |
|
| 372 |
mu_1 = s * (x + y) + c * (z - w); // = r - p |
| 373 |
mu_2 = T (2) * (c * x - s * z); // = 2*q |
| 374 |
|
| 375 |
changed = true; |
| 376 |
} |
| 377 |
|
| 378 |
// The second stage diagonalizes, |
| 379 |
// [ c2 s2 ]^T [ p q ] [ c2 s2 ] = [ d1 0 ] |
| 380 |
// [ -s2 c2 ] [ q r ] [ -s2 c2 ] [ 0 d2 ] |
| 381 |
T c_2, s_2; |
| 382 |
if (std::abs (mu_2) <= tol * std::abs (mu_1)) |
| 383 |
{ |
| 384 |
c_2 = T (1); |
| 385 |
s_2 = T (0); |
| 386 |
} |
| 387 |
else |
| 388 |
{ |
| 389 |
const T rho_2 = mu_1 / mu_2; |
| 390 |
T t_2 = T (1) / (std::abs (rho_2) + std::sqrt (1 + rho_2 * rho_2)); |
| 391 |
if (rho_2 < 0) |
| 392 |
t_2 = -t_2; |
| 393 |
c_2 = T (1) / std::sqrt (T (1) + t_2 * t_2); |
| 394 |
s_2 = c_2 * t_2; |
| 395 |
|
| 396 |
changed = true; |
| 397 |
} |
| 398 |
|
| 399 |
const T c_1 = c_2 * c - s_2 * s; |
| 400 |
const T s_1 = s_2 * c + c_2 * s; |
| 401 |
|
| 402 |
if (!changed) |
| 403 |
{ |
| 404 |
// We've decided that the off-diagonal entries are already small |
| 405 |
// enough, so we'll set them to zero. This actually appears to result |
| 406 |
// in smaller errors than leaving them be, possibly because it prevents |
| 407 |
// us from trying to do extra rotations later that we don't need. |
| 408 |
A[k][j] = 0; |
| 409 |
A[j][k] = 0; |
| 410 |
return false; |
| 411 |
} |
| 412 |
|
| 413 |
const T d_1 = c_1 * (w * c_2 - x * s_2) - s_1 * (y * c_2 - z * s_2); |
| 414 |
const T d_2 = s_1 * (w * s_2 + x * c_2) + c_1 * (y * s_2 + z * c_2); |
| 415 |
|
| 416 |
// For the entries we just zeroed out, we'll just set them to 0, since |
| 417 |
// they should be 0 up to machine precision. |
| 418 |
A[j][j] = d_1; |
| 419 |
A[k][k] = d_2; |
| 420 |
A[k][j] = 0; |
| 421 |
A[j][k] = 0; |
| 422 |
|
| 423 |
// Rotate the entries that _weren't_ involved in the 2x2 SVD: |
| 424 |
{ |
| 425 |
// Rotate on the left by |
| 426 |
// [ c1 s1 0 ]^T [ c1 0 s1 ]^T [ 1 0 0 ]^T |
| 427 |
// [ -s1 c1 0 ] or [ 0 1 0 ] or [ 0 c1 s1 ] |
| 428 |
// [ 0 0 1 ] [ -s1 0 c1 ] [ 0 -s1 c1 ] |
| 429 |
// This has the effect of adding the (weighted) ith and jth _rows_ to |
| 430 |
// each other. |
| 431 |
const T tau1 = A[j][l]; |
| 432 |
const T tau2 = A[k][l]; |
| 433 |
A[j][l] = c_1 * tau1 - s_1 * tau2; |
| 434 |
A[k][l] = s_1 * tau1 + c_1 * tau2; |
| 435 |
} |
| 436 |
|
| 437 |
{ |
| 438 |
// Rotate on the right by |
| 439 |
// [ c2 s2 0 ] [ c2 0 s2 ] [ 1 0 0 ] |
| 440 |
// [ -s2 c2 0 ] or [ 0 1 0 ] or [ 0 c2 s2 ] |
| 441 |
// [ 0 0 1 ] [ -s2 0 c2 ] [ 0 -s2 c2 ] |
| 442 |
// This has the effect of adding the (weighted) ith and jth _columns_ to |
| 443 |
// each other. |
| 444 |
const T tau1 = A[l][j]; |
| 445 |
const T tau2 = A[l][k]; |
| 446 |
A[l][j] = c_2 * tau1 - s_2 * tau2; |
| 447 |
A[l][k] = s_2 * tau1 + c_2 * tau2; |
| 448 |
} |
| 449 |
|
| 450 |
// Now apply the rotations to U and V: |
| 451 |
// Remember that we have |
| 452 |
// R1^T * A * R2 = D |
| 453 |
// This is in the 2x2 case, but after doing a bunch of these |
| 454 |
// we will get something like this for the 3x3 case: |
| 455 |
// ... R1b^T * R1a^T * A * R2a * R2b * ... = D |
| 456 |
// ----------------- --------------- |
| 457 |
// = U^T = V |
| 458 |
// So, |
| 459 |
// U = R1a * R1b * ... |
| 460 |
// V = R2a * R2b * ... |
| 461 |
jacobiRotateRight<T, j, k> (U, c_1, s_1); |
| 462 |
jacobiRotateRight<T, j, k> (V, c_2, s_2); |
| 463 |
|
| 464 |
return true; |
| 465 |
} |
| 466 |
|
| 467 |
template <typename T> |
| 468 |
bool |
| 469 |
twoSidedJacobiRotation (IMATH_INTERNAL_NAMESPACE::Matrix44<T>& A, |
| 470 |
int j, |
| 471 |
int k, |
| 472 |
IMATH_INTERNAL_NAMESPACE::Matrix44<T>& U, |
| 473 |
IMATH_INTERNAL_NAMESPACE::Matrix44<T>& V, |
| 474 |
const T tol) |
| 475 |
{ |
| 476 |
// Load everything into local variables to make things easier on the |
| 477 |
// optimizer: |
| 478 |
const T w = A[j][j]; |
| 479 |
const T x = A[j][k]; |
| 480 |
const T y = A[k][j]; |
| 481 |
const T z = A[k][k]; |
| 482 |
|
| 483 |
// We will keep track of whether we're actually performing any rotations, |
| 484 |
// since if the matrix is already diagonal we'll end up with the identity |
| 485 |
// as our Jacobi rotation and we can short-circuit. |
| 486 |
bool changed = false; |
| 487 |
|
| 488 |
// The first step is to symmetrize the 2x2 matrix, |
| 489 |
// [ c s ]^T [ w x ] = [ p q ] |
| 490 |
// [ -s c ] [ y z ] [ q r ] |
| 491 |
T mu_1 = w + z; |
| 492 |
T mu_2 = x - y; |
| 493 |
|
| 494 |
T c, s; |
| 495 |
if (std::abs (mu_2) <= tol * std::abs (mu_1)) // Already symmetric (to tolerance) |
| 496 |
{ // Note that the <= is important here |
| 497 |
c = T (1); // because we want to bypass the computation |
| 498 |
s = T (0); // of rho if mu_1 = mu_2 = 0. |
| 499 |
|
| 500 |
const T p = w; |
| 501 |
const T r = z; |
| 502 |
mu_1 = r - p; |
| 503 |
mu_2 = x + y; |
| 504 |
} |
| 505 |
else |
| 506 |
{ |
| 507 |
// TODO is there a native inverse square root function? |
| 508 |
const T rho = mu_1 / mu_2; |
| 509 |
|
| 510 |
s = T (1) / std::sqrt (T (1) + rho * rho); |
| 511 |
if (rho < 0) |
| 512 |
s = -s; |
| 513 |
c = s * rho; |
| 514 |
|
| 515 |
mu_1 = s * (x + y) + c * (z - w); // = r - p |
| 516 |
mu_2 = T (2) * (c * x - s * z); // = 2*q |
| 517 |
|
| 518 |
changed = true; |
| 519 |
} |
| 520 |
|
| 521 |
// The second stage diagonalizes, |
| 522 |
// [ c2 s2 ]^T [ p q ] [ c2 s2 ] = [ d1 0 ] |
| 523 |
// [ -s2 c2 ] [ q r ] [ -s2 c2 ] [ 0 d2 ] |
| 524 |
T c_2, s_2; |
| 525 |
if (std::abs (mu_2) <= tol * std::abs (mu_1)) |
| 526 |
{ |
| 527 |
c_2 = T (1); |
| 528 |
s_2 = T (0); |
| 529 |
} |
| 530 |
else |
| 531 |
{ |
| 532 |
const T rho_2 = mu_1 / mu_2; |
| 533 |
T t_2 = T (1) / (std::abs (rho_2) + std::sqrt (1 + rho_2 * rho_2)); |
| 534 |
if (rho_2 < 0) |
| 535 |
t_2 = -t_2; |
| 536 |
c_2 = T (1) / std::sqrt (T (1) + t_2 * t_2); |
| 537 |
s_2 = c_2 * t_2; |
| 538 |
|
| 539 |
changed = true; |
| 540 |
} |
| 541 |
|
| 542 |
const T c_1 = c_2 * c - s_2 * s; |
| 543 |
const T s_1 = s_2 * c + c_2 * s; |
| 544 |
|
| 545 |
if (!changed) |
| 546 |
{ |
| 547 |
// We've decided that the off-diagonal entries are already small |
| 548 |
// enough, so we'll set them to zero. This actually appears to result |
| 549 |
// in smaller errors than leaving them be, possibly because it prevents |
| 550 |
// us from trying to do extra rotations later that we don't need. |
| 551 |
A[k][j] = 0; |
| 552 |
A[j][k] = 0; |
| 553 |
return false; |
| 554 |
} |
| 555 |
|
| 556 |
const T d_1 = c_1 * (w * c_2 - x * s_2) - s_1 * (y * c_2 - z * s_2); |
| 557 |
const T d_2 = s_1 * (w * s_2 + x * c_2) + c_1 * (y * s_2 + z * c_2); |
| 558 |
|
| 559 |
// For the entries we just zeroed out, we'll just set them to 0, since |
| 560 |
// they should be 0 up to machine precision. |
| 561 |
A[j][j] = d_1; |
| 562 |
A[k][k] = d_2; |
| 563 |
A[k][j] = 0; |
| 564 |
A[j][k] = 0; |
| 565 |
|
| 566 |
// Rotate the entries that _weren't_ involved in the 2x2 SVD: |
| 567 |
for (int l = 0; l < 4; ++l) |
| 568 |
{ |
| 569 |
if (l == j || l == k) |
| 570 |
continue; |
| 571 |
|
| 572 |
// Rotate on the left by |
| 573 |
// [ 1 ] |
| 574 |
// [ . ] |
| 575 |
// [ c2 s2 ] j |
| 576 |
// [ 1 ] |
| 577 |
// [ -s2 c2 ] k |
| 578 |
// [ . ] |
| 579 |
// [ 1 ] |
| 580 |
// j k |
| 581 |
// |
| 582 |
// This has the effect of adding the (weighted) ith and jth _rows_ to |
| 583 |
// each other. |
| 584 |
const T tau1 = A[j][l]; |
| 585 |
const T tau2 = A[k][l]; |
| 586 |
A[j][l] = c_1 * tau1 - s_1 * tau2; |
| 587 |
A[k][l] = s_1 * tau1 + c_1 * tau2; |
| 588 |
} |
| 589 |
|
| 590 |
for (int l = 0; l < 4; ++l) |
| 591 |
{ |
| 592 |
// We set the A[j/k][j/k] entries already |
| 593 |
if (l == j || l == k) |
| 594 |
continue; |
| 595 |
|
| 596 |
// Rotate on the right by |
| 597 |
// [ 1 ] |
| 598 |
// [ . ] |
| 599 |
// [ c2 s2 ] j |
| 600 |
// [ 1 ] |
| 601 |
// [ -s2 c2 ] k |
| 602 |
// [ . ] |
| 603 |
// [ 1 ] |
| 604 |
// j k |
| 605 |
// |
| 606 |
// This has the effect of adding the (weighted) ith and jth _columns_ to |
| 607 |
// each other. |
| 608 |
const T tau1 = A[l][j]; |
| 609 |
const T tau2 = A[l][k]; |
| 610 |
A[l][j] = c_2 * tau1 - s_2 * tau2; |
| 611 |
A[l][k] = s_2 * tau1 + c_2 * tau2; |
| 612 |
} |
| 613 |
|
| 614 |
// Now apply the rotations to U and V: |
| 615 |
// Remember that we have |
| 616 |
// R1^T * A * R2 = D |
| 617 |
// This is in the 2x2 case, but after doing a bunch of these |
| 618 |
// we will get something like this for the 3x3 case: |
| 619 |
// ... R1b^T * R1a^T * A * R2a * R2b * ... = D |
| 620 |
// ----------------- --------------- |
| 621 |
// = U^T = V |
| 622 |
// So, |
| 623 |
// U = R1a * R1b * ... |
| 624 |
// V = R2a * R2b * ... |
| 625 |
jacobiRotateRight (U, j, k, c_1, s_1); |
| 626 |
jacobiRotateRight (V, j, k, c_2, s_2); |
| 627 |
|
| 628 |
return true; |
| 629 |
} |
| 630 |
|
| 631 |
template <typename T> |
| 632 |
void |
| 633 |
swapColumns (IMATH_INTERNAL_NAMESPACE::Matrix33<T>& A, int j, int k) |
| 634 |
{ |
| 635 |
for (int i = 0; i < 3; ++i) |
| 636 |
std::swap (A[i][j], A[i][k]); |
| 637 |
} |
| 638 |
|
| 639 |
template <typename T> |
| 640 |
IMATH_CONSTEXPR14 T |
| 641 |
maxOffDiag (const IMATH_INTERNAL_NAMESPACE::Matrix33<T>& A) |
| 642 |
{ |
| 643 |
T result = 0; |
| 644 |
result = std::max (result, std::abs (A[0][1])); |
| 645 |
result = std::max (result, std::abs (A[0][2])); |
| 646 |
result = std::max (result, std::abs (A[1][0])); |
| 647 |
result = std::max (result, std::abs (A[1][2])); |
| 648 |
result = std::max (result, std::abs (A[2][0])); |
| 649 |
result = std::max (result, std::abs (A[2][1])); |
| 650 |
return result; |
| 651 |
} |
| 652 |
|
| 653 |
template <typename T> |
| 654 |
IMATH_CONSTEXPR14 T |
| 655 |
maxOffDiag (const IMATH_INTERNAL_NAMESPACE::Matrix44<T>& A) |
| 656 |
{ |
| 657 |
T result = 0; |
| 658 |
for (int i = 0; i < 4; ++i) |
| 659 |
{ |
| 660 |
for (int j = 0; j < 4; ++j) |
| 661 |
{ |
| 662 |
if (i != j) |
| 663 |
result = std::max (result, std::abs (A[i][j])); |
| 664 |
} |
| 665 |
} |
| 666 |
|
| 667 |
return result; |
| 668 |
} |
| 669 |
|
| 670 |
template <typename T> |
| 671 |
void |
| 672 |
twoSidedJacobiSVD (IMATH_INTERNAL_NAMESPACE::Matrix33<T> A, |
| 673 |
IMATH_INTERNAL_NAMESPACE::Matrix33<T>& U, |
| 674 |
IMATH_INTERNAL_NAMESPACE::Vec3<T>& S, |
| 675 |
IMATH_INTERNAL_NAMESPACE::Matrix33<T>& V, |
| 676 |
const T tol, |
| 677 |
const bool forcePositiveDeterminant) |
| 678 |
{ |
| 679 |
// The two-sided Jacobi SVD works by repeatedly zeroing out |
| 680 |
// off-diagonal entries of the matrix, 2 at a time. Basically, |
| 681 |
// we can take our 3x3 matrix, |
| 682 |
// [* * *] |
| 683 |
// [* * *] |
| 684 |
// [* * *] |
| 685 |
// and use a pair of orthogonal transforms to zero out, say, the |
| 686 |
// pair of entries (0, 1) and (1, 0): |
| 687 |
// [ c1 s1 ] [* * *] [ c2 s2 ] [* *] |
| 688 |
// [-s1 c1 ] [* * *] [-s2 c2 ] = [ * *] |
| 689 |
// [ 1] [* * *] [ 1] [* * *] |
| 690 |
// When we go to zero out the next pair of entries (say, (0, 2) and (2, 0)) |
| 691 |
// then we don't expect those entries to stay 0: |
| 692 |
// [ c1 s1 ] [* *] [ c2 s2 ] [* * ] |
| 693 |
// [-s1 c1 ] [ * *] [-s2 c2 ] = [* * *] |
| 694 |
// [ 1] [* * *] [ 1] [ * *] |
| 695 |
// However, if we keep doing this, we'll find that the off-diagonal entries |
| 696 |
// converge to 0 fairly quickly (convergence should be roughly cubic). The |
| 697 |
// result is a diagonal A matrix and a bunch of orthogonal transforms: |
| 698 |
// [* * *] [* ] |
| 699 |
// L1 L2 ... Ln [* * *] Rn ... R2 R1 = [ * ] |
| 700 |
// [* * *] [ *] |
| 701 |
// ------------ ------- ------------ ------- |
| 702 |
// U^T A V S |
| 703 |
// This turns out to be highly accurate because (1) orthogonal transforms |
| 704 |
// are extremely stable to compute and apply (this is why QR factorization |
| 705 |
// works so well, FWIW) and because (2) by applying everything to the original |
| 706 |
// matrix A instead of computing (A^T * A) we avoid any precision loss that |
| 707 |
// would result from that. |
| 708 |
U.makeIdentity(); |
| 709 |
V.makeIdentity(); |
| 710 |
|
| 711 |
const int maxIter = 20; // In case we get really unlucky, prevents infinite loops |
| 712 |
const T absTol = tol * maxOffDiag (A); // Tolerance is in terms of the maximum |
| 713 |
if (absTol != 0) // _off-diagonal_ entry. |
| 714 |
{ |
| 715 |
int numIter = 0; |
| 716 |
do |
| 717 |
{ |
| 718 |
++numIter; |
| 719 |
bool changed = twoSidedJacobiRotation<T, 0, 1, 2> (A, U, V, tol); |
| 720 |
changed = twoSidedJacobiRotation<T, 0, 2, 1> (A, U, V, tol) || changed; |
| 721 |
changed = twoSidedJacobiRotation<T, 1, 2, 0> (A, U, V, tol) || changed; |
| 722 |
if (!changed) |
| 723 |
break; |
| 724 |
} while (maxOffDiag (A) > absTol && numIter < maxIter); |
| 725 |
} |
| 726 |
|
| 727 |
// The off-diagonal entries are (effectively) 0, so whatever's left on the |
| 728 |
// diagonal are the singular values: |
| 729 |
S.x = A[0][0]; |
| 730 |
S.y = A[1][1]; |
| 731 |
S.z = A[2][2]; |
| 732 |
|
| 733 |
// Nothing thus far has guaranteed that the singular values are positive, |
| 734 |
// so let's go back through and flip them if not (since by contract we are |
| 735 |
// supposed to return all positive SVs): |
| 736 |
for (int i = 0; i < 3; ++i) |
| 737 |
{ |
| 738 |
if (S[i] < 0) |
| 739 |
{ |
| 740 |
// If we flip S[i], we need to flip the corresponding column of U |
| 741 |
// (we could also pick V if we wanted; it doesn't really matter): |
| 742 |
S[i] = -S[i]; |
| 743 |
for (int j = 0; j < 3; ++j) |
| 744 |
U[j][i] = -U[j][i]; |
| 745 |
} |
| 746 |
} |
| 747 |
|
| 748 |
// Order the singular values from largest to smallest; this requires |
| 749 |
// exactly two passes through the data using bubble sort: |
| 750 |
for (int i = 0; i < 2; ++i) |
| 751 |
{ |
| 752 |
for (int j = 0; j < (2 - i); ++j) |
| 753 |
{ |
| 754 |
// No absolute values necessary since we already ensured that |
| 755 |
// they're positive: |
| 756 |
if (S[j] < S[j + 1]) |
| 757 |
{ |
| 758 |
// If we swap singular values we also have to swap |
| 759 |
// corresponding columns in U and V: |
| 760 |
std::swap (S[j], S[j + 1]); |
| 761 |
swapColumns (U, j, j + 1); |
| 762 |
swapColumns (V, j, j + 1); |
| 763 |
} |
| 764 |
} |
| 765 |
} |
| 766 |
|
| 767 |
if (forcePositiveDeterminant) |
| 768 |
{ |
| 769 |
// We want to guarantee that the returned matrices always have positive |
| 770 |
// determinant. We can do this by adding the appropriate number of |
| 771 |
// matrices of the form: |
| 772 |
// [ 1 ] |
| 773 |
// L = [ 1 ] |
| 774 |
// [ -1 ] |
| 775 |
// Note that L' = L and L*L = Identity. Thus we can add: |
| 776 |
// U*L*L*S*V = (U*L)*(L*S)*V |
| 777 |
// if U has a negative determinant, and |
| 778 |
// U*S*L*L*V = U*(S*L)*(L*V) |
| 779 |
// if V has a neg. determinant. |
| 780 |
if (U.determinant() < 0) |
| 781 |
{ |
| 782 |
for (int i = 0; i < 3; ++i) |
| 783 |
U[i][2] = -U[i][2]; |
| 784 |
S.z = -S.z; |
| 785 |
} |
| 786 |
|
| 787 |
if (V.determinant() < 0) |
| 788 |
{ |
| 789 |
for (int i = 0; i < 3; ++i) |
| 790 |
V[i][2] = -V[i][2]; |
| 791 |
S.z = -S.z; |
| 792 |
} |
| 793 |
} |
| 794 |
} |
| 795 |
|
| 796 |
template <typename T> |
| 797 |
void |
| 798 |
twoSidedJacobiSVD (IMATH_INTERNAL_NAMESPACE::Matrix44<T> A, |
| 799 |
IMATH_INTERNAL_NAMESPACE::Matrix44<T>& U, |
| 800 |
IMATH_INTERNAL_NAMESPACE::Vec4<T>& S, |
| 801 |
IMATH_INTERNAL_NAMESPACE::Matrix44<T>& V, |
| 802 |
const T tol, |
| 803 |
const bool forcePositiveDeterminant) |
| 804 |
{ |
| 805 |
// Please see the Matrix33 version for a detailed description of the algorithm. |
| 806 |
U.makeIdentity(); |
| 807 |
V.makeIdentity(); |
| 808 |
|
| 809 |
const int maxIter = 20; // In case we get really unlucky, prevents infinite loops |
| 810 |
const T absTol = tol * maxOffDiag (A); // Tolerance is in terms of the maximum |
| 811 |
if (absTol != 0) // _off-diagonal_ entry. |
| 812 |
{ |
| 813 |
int numIter = 0; |
| 814 |
do |
| 815 |
{ |
| 816 |
++numIter; |
| 817 |
bool changed = twoSidedJacobiRotation (A, 0, 1, U, V, tol); |
| 818 |
changed = twoSidedJacobiRotation (A, 0, 2, U, V, tol) || changed; |
| 819 |
changed = twoSidedJacobiRotation (A, 0, 3, U, V, tol) || changed; |
| 820 |
changed = twoSidedJacobiRotation (A, 1, 2, U, V, tol) || changed; |
| 821 |
changed = twoSidedJacobiRotation (A, 1, 3, U, V, tol) || changed; |
| 822 |
changed = twoSidedJacobiRotation (A, 2, 3, U, V, tol) || changed; |
| 823 |
if (!changed) |
| 824 |
break; |
| 825 |
} while (maxOffDiag (A) > absTol && numIter < maxIter); |
| 826 |
} |
| 827 |
|
| 828 |
// The off-diagonal entries are (effectively) 0, so whatever's left on the |
| 829 |
// diagonal are the singular values: |
| 830 |
S[0] = A[0][0]; |
| 831 |
S[1] = A[1][1]; |
| 832 |
S[2] = A[2][2]; |
| 833 |
S[3] = A[3][3]; |
| 834 |
|
| 835 |
// Nothing thus far has guaranteed that the singular values are positive, |
| 836 |
// so let's go back through and flip them if not (since by contract we are |
| 837 |
// supposed to return all positive SVs): |
| 838 |
for (int i = 0; i < 4; ++i) |
| 839 |
{ |
| 840 |
if (S[i] < 0) |
| 841 |
{ |
| 842 |
// If we flip S[i], we need to flip the corresponding column of U |
| 843 |
// (we could also pick V if we wanted; it doesn't really matter): |
| 844 |
S[i] = -S[i]; |
| 845 |
for (int j = 0; j < 4; ++j) |
| 846 |
U[j][i] = -U[j][i]; |
| 847 |
} |
| 848 |
} |
| 849 |
|
| 850 |
// Order the singular values from largest to smallest using insertion sort: |
| 851 |
for (int i = 1; i < 4; ++i) |
| 852 |
{ |
| 853 |
const IMATH_INTERNAL_NAMESPACE::Vec4<T> uCol (U[0][i], U[1][i], U[2][i], U[3][i]); |
| 854 |
const IMATH_INTERNAL_NAMESPACE::Vec4<T> vCol (V[0][i], V[1][i], V[2][i], V[3][i]); |
| 855 |
const T sVal = S[i]; |
| 856 |
|
| 857 |
int j = i - 1; |
| 858 |
while (std::abs (S[j]) < std::abs (sVal)) |
| 859 |
{ |
| 860 |
for (int k = 0; k < 4; ++k) |
| 861 |
U[k][j + 1] = U[k][j]; |
| 862 |
for (int k = 0; k < 4; ++k) |
| 863 |
V[k][j + 1] = V[k][j]; |
| 864 |
S[j + 1] = S[j]; |
| 865 |
|
| 866 |
--j; |
| 867 |
if (j < 0) |
| 868 |
break; |
| 869 |
} |
| 870 |
|
| 871 |
for (int k = 0; k < 4; ++k) |
| 872 |
U[k][j + 1] = uCol[k]; |
| 873 |
for (int k = 0; k < 4; ++k) |
| 874 |
V[k][j + 1] = vCol[k]; |
| 875 |
S[j + 1] = sVal; |
| 876 |
} |
| 877 |
|
| 878 |
if (forcePositiveDeterminant) |
| 879 |
{ |
| 880 |
// We want to guarantee that the returned matrices always have positive |
| 881 |
// determinant. We can do this by adding the appropriate number of |
| 882 |
// matrices of the form: |
| 883 |
// [ 1 ] |
| 884 |
// L = [ 1 ] |
| 885 |
// [ 1 ] |
| 886 |
// [ -1 ] |
| 887 |
// Note that L' = L and L*L = Identity. Thus we can add: |
| 888 |
// U*L*L*S*V = (U*L)*(L*S)*V |
| 889 |
// if U has a negative determinant, and |
| 890 |
// U*S*L*L*V = U*(S*L)*(L*V) |
| 891 |
// if V has a neg. determinant. |
| 892 |
if (U.determinant() < 0) |
| 893 |
{ |
| 894 |
for (int i = 0; i < 4; ++i) |
| 895 |
U[i][3] = -U[i][3]; |
| 896 |
S[3] = -S[3]; |
| 897 |
} |
| 898 |
|
| 899 |
if (V.determinant() < 0) |
| 900 |
{ |
| 901 |
for (int i = 0; i < 4; ++i) |
| 902 |
V[i][3] = -V[i][3]; |
| 903 |
S[3] = -S[3]; |
| 904 |
} |
| 905 |
} |
| 906 |
} |
| 907 |
|
| 908 |
} // namespace |
| 909 |
|
| 910 |
/// TODO |
| 911 |
template <typename T> |
| 912 |
void |
| 913 |
jacobiSVD (const IMATH_INTERNAL_NAMESPACE::Matrix33<T>& A, |
| 914 |
IMATH_INTERNAL_NAMESPACE::Matrix33<T>& U, |
| 915 |
IMATH_INTERNAL_NAMESPACE::Vec3<T>& S, |
| 916 |
IMATH_INTERNAL_NAMESPACE::Matrix33<T>& V, |
| 917 |
const T tol, |
| 918 |
const bool forcePositiveDeterminant) |
| 919 |
{ |
| 920 |
twoSidedJacobiSVD (A, U, S, V, tol, forcePositiveDeterminant); |
| 921 |
} |
| 922 |
|
| 923 |
/// TODO |
| 924 |
template <typename T> |
| 925 |
void |
| 926 |
jacobiSVD (const IMATH_INTERNAL_NAMESPACE::Matrix44<T>& A, |
| 927 |
IMATH_INTERNAL_NAMESPACE::Matrix44<T>& U, |
| 928 |
IMATH_INTERNAL_NAMESPACE::Vec4<T>& S, |
| 929 |
IMATH_INTERNAL_NAMESPACE::Matrix44<T>& V, |
| 930 |
const T tol, |
| 931 |
const bool forcePositiveDeterminant) |
| 932 |
{ |
| 933 |
twoSidedJacobiSVD (A, U, S, V, tol, forcePositiveDeterminant); |
| 934 |
} |
| 935 |
|
| 936 |
/// TODO |
| 937 |
template IMATH_EXPORT void jacobiSVD (const IMATH_INTERNAL_NAMESPACE::Matrix33<float>& A, |
| 938 |
IMATH_INTERNAL_NAMESPACE::Matrix33<float>& U, |
| 939 |
IMATH_INTERNAL_NAMESPACE::Vec3<float>& S, |
| 940 |
IMATH_INTERNAL_NAMESPACE::Matrix33<float>& V, |
| 941 |
const float tol, |
| 942 |
const bool forcePositiveDeterminant); |
| 943 |
/// TODO |
| 944 |
template IMATH_EXPORT void jacobiSVD (const IMATH_INTERNAL_NAMESPACE::Matrix33<double>& A, |
| 945 |
IMATH_INTERNAL_NAMESPACE::Matrix33<double>& U, |
| 946 |
IMATH_INTERNAL_NAMESPACE::Vec3<double>& S, |
| 947 |
IMATH_INTERNAL_NAMESPACE::Matrix33<double>& V, |
| 948 |
const double tol, |
| 949 |
const bool forcePositiveDeterminant); |
| 950 |
/// TODO |
| 951 |
template IMATH_EXPORT void jacobiSVD (const IMATH_INTERNAL_NAMESPACE::Matrix44<float>& A, |
| 952 |
IMATH_INTERNAL_NAMESPACE::Matrix44<float>& U, |
| 953 |
IMATH_INTERNAL_NAMESPACE::Vec4<float>& S, |
| 954 |
IMATH_INTERNAL_NAMESPACE::Matrix44<float>& V, |
| 955 |
const float tol, |
| 956 |
const bool forcePositiveDeterminant); |
| 957 |
/// TODO |
| 958 |
template IMATH_EXPORT void jacobiSVD (const IMATH_INTERNAL_NAMESPACE::Matrix44<double>& A, |
| 959 |
IMATH_INTERNAL_NAMESPACE::Matrix44<double>& U, |
| 960 |
IMATH_INTERNAL_NAMESPACE::Vec4<double>& S, |
| 961 |
IMATH_INTERNAL_NAMESPACE::Matrix44<double>& V, |
| 962 |
const double tol, |
| 963 |
const bool forcePositiveDeterminant); |
| 964 |
|
| 965 |
namespace |
| 966 |
{ |
| 967 |
|
| 968 |
template <int j, int k, typename TM> |
| 969 |
inline void |
| 970 |
jacobiRotateRight (TM& A, const typename TM::BaseType s, const typename TM::BaseType tau) |
| 971 |
{ |
| 972 |
typedef typename TM::BaseType T; |
| 973 |
|
| 974 |
for (unsigned int i = 0; i < TM::dimensions(); ++i) |
| 975 |
{ |
| 976 |
const T nu1 = A[i][j]; |
| 977 |
const T nu2 = A[i][k]; |
| 978 |
A[i][j] -= s * (nu2 + tau * nu1); |
| 979 |
A[i][k] += s * (nu1 - tau * nu2); |
| 980 |
} |
| 981 |
} |
| 982 |
|
| 983 |
template <int j, int k, int l, typename T> |
| 984 |
bool |
| 985 |
jacobiRotation (Matrix33<T>& A, Matrix33<T>& V, Vec3<T>& Z, const T tol) |
| 986 |
{ |
| 987 |
// Load everything into local variables to make things easier on the |
| 988 |
// optimizer: |
| 989 |
const T x = A[j][j]; |
| 990 |
const T y = A[j][k]; |
| 991 |
const T z = A[k][k]; |
| 992 |
|
| 993 |
// The first stage diagonalizes, |
| 994 |
// [ c s ]^T [ x y ] [ c -s ] = [ d1 0 ] |
| 995 |
// [ -s c ] [ y z ] [ s c ] [ 0 d2 ] |
| 996 |
const T mu1 = z - x; |
| 997 |
const T mu2 = 2 * y; |
| 998 |
|
| 999 |
if (std::abs (mu2) <= tol * std::abs (mu1)) |
| 1000 |
{ |
| 1001 |
// We've decided that the off-diagonal entries are already small |
| 1002 |
// enough, so we'll set them to zero. This actually appears to result |
| 1003 |
// in smaller errors than leaving them be, possibly because it prevents |
| 1004 |
// us from trying to do extra rotations later that we don't need. |
| 1005 |
A[j][k] = 0; |
| 1006 |
return false; |
| 1007 |
} |
| 1008 |
const T rho = mu1 / mu2; |
| 1009 |
const T t = (rho < 0 ? T (-1) : T (1)) / (std::abs (rho) + std::sqrt (1 + rho * rho)); |
| 1010 |
const T c = T (1) / std::sqrt (T (1) + t * t); |
| 1011 |
const T s = t * c; |
| 1012 |
const T tau = s / (T (1) + c); |
| 1013 |
const T h = t * y; |
| 1014 |
|
| 1015 |
// Update diagonal elements. |
| 1016 |
Z[j] -= h; |
| 1017 |
Z[k] += h; |
| 1018 |
A[j][j] -= h; |
| 1019 |
A[k][k] += h; |
| 1020 |
|
| 1021 |
// For the entries we just zeroed out, we'll just set them to 0, since |
| 1022 |
// they should be 0 up to machine precision. |
| 1023 |
A[j][k] = 0; |
| 1024 |
|
| 1025 |
// We only update upper triagnular elements of A, since |
| 1026 |
// A is supposed to be symmetric. |
| 1027 |
T& offd1 = l < j ? A[l][j] : A[j][l]; |
| 1028 |
T& offd2 = l < k ? A[l][k] : A[k][l]; |
| 1029 |
const T nu1 = offd1; |
| 1030 |
const T nu2 = offd2; |
| 1031 |
offd1 = nu1 - s * (nu2 + tau * nu1); |
| 1032 |
offd2 = nu2 + s * (nu1 - tau * nu2); |
| 1033 |
|
| 1034 |
// Apply rotation to V |
| 1035 |
jacobiRotateRight<j, k> (V, s, tau); |
| 1036 |
|
| 1037 |
return true; |
| 1038 |
} |
| 1039 |
|
| 1040 |
template <int j, int k, int l1, int l2, typename T> |
| 1041 |
bool |
| 1042 |
jacobiRotation (Matrix44<T>& A, Matrix44<T>& V, Vec4<T>& Z, const T tol) |
| 1043 |
{ |
| 1044 |
const T x = A[j][j]; |
| 1045 |
const T y = A[j][k]; |
| 1046 |
const T z = A[k][k]; |
| 1047 |
|
| 1048 |
const T mu1 = z - x; |
| 1049 |
const T mu2 = T (2) * y; |
| 1050 |
|
| 1051 |
// Let's see if rho^(-1) = mu2 / mu1 is less than tol |
| 1052 |
// This test also checks if rho^2 will overflow |
| 1053 |
// when tol^(-1) < sqrt(std::numeric_limits<T>::max()). |
| 1054 |
if (std::abs (mu2) <= tol * std::abs (mu1)) |
| 1055 |
{ |
| 1056 |
A[j][k] = 0; |
| 1057 |
return true; |
| 1058 |
} |
| 1059 |
|
| 1060 |
const T rho = mu1 / mu2; |
| 1061 |
const T t = (rho < 0 ? T (-1) : T (1)) / (std::abs (rho) + std::sqrt (1 + rho * rho)); |
| 1062 |
const T c = T (1) / std::sqrt (T (1) + t * t); |
| 1063 |
const T s = c * t; |
| 1064 |
const T tau = s / (T (1) + c); |
| 1065 |
const T h = t * y; |
| 1066 |
|
| 1067 |
Z[j] -= h; |
| 1068 |
Z[k] += h; |
| 1069 |
A[j][j] -= h; |
| 1070 |
A[k][k] += h; |
| 1071 |
A[j][k] = 0; |
| 1072 |
|
| 1073 |
{ |
| 1074 |
T& offd1 = l1 < j ? A[l1][j] : A[j][l1]; |
| 1075 |
T& offd2 = l1 < k ? A[l1][k] : A[k][l1]; |
| 1076 |
const T nu1 = offd1; |
| 1077 |
const T nu2 = offd2; |
| 1078 |
offd1 -= s * (nu2 + tau * nu1); |
| 1079 |
offd2 += s * (nu1 - tau * nu2); |
| 1080 |
} |
| 1081 |
|
| 1082 |
{ |
| 1083 |
T& offd1 = l2 < j ? A[l2][j] : A[j][l2]; |
| 1084 |
T& offd2 = l2 < k ? A[l2][k] : A[k][l2]; |
| 1085 |
const T nu1 = offd1; |
| 1086 |
const T nu2 = offd2; |
| 1087 |
offd1 -= s * (nu2 + tau * nu1); |
| 1088 |
offd2 += s * (nu1 - tau * nu2); |
| 1089 |
} |
| 1090 |
|
| 1091 |
jacobiRotateRight<j, k> (V, s, tau); |
| 1092 |
|
| 1093 |
return true; |
| 1094 |
} |
| 1095 |
|
| 1096 |
template <typename TM> |
| 1097 |
IMATH_CONSTEXPR14 inline typename TM::BaseType |
| 1098 |
maxOffDiagSymm (const TM& A) |
| 1099 |
{ |
| 1100 |
typedef typename TM::BaseType T; |
| 1101 |
T result = 0; |
| 1102 |
for (unsigned int i = 0; i < TM::dimensions(); ++i) |
| 1103 |
for (unsigned int j = i + 1; j < TM::dimensions(); ++j) |
| 1104 |
result = std::max (result, std::abs (A[i][j])); |
| 1105 |
|
| 1106 |
return result; |
| 1107 |
} |
| 1108 |
|
| 1109 |
} // namespace |
| 1110 |
|
| 1111 |
template <typename T> |
| 1112 |
void |
| 1113 |
jacobiEigenSolver (Matrix33<T>& A, Vec3<T>& S, Matrix33<T>& V, const T tol) |
| 1114 |
{ |
| 1115 |
V.makeIdentity(); |
| 1116 |
for (int i = 0; i < 3; ++i) |
| 1117 |
{ |
| 1118 |
S[i] = A[i][i]; |
| 1119 |
} |
| 1120 |
|
| 1121 |
const int maxIter = 20; // In case we get really unlucky, prevents infinite loops |
| 1122 |
const T absTol = tol * maxOffDiagSymm (A); // Tolerance is in terms of the maximum |
| 1123 |
if (absTol != 0) // _off-diagonal_ entry. |
| 1124 |
{ |
| 1125 |
int numIter = 0; |
| 1126 |
do |
| 1127 |
{ |
| 1128 |
// Z is for accumulating small changes (h) to diagonal entries |
| 1129 |
// of A for one sweep. Adding h's directly to A might cause |
| 1130 |
// a cancellation effect when h is relatively very small to |
| 1131 |
// the corresponding diagonal entry of A and |
| 1132 |
// this will increase numerical errors |
| 1133 |
Vec3<T> Z (0, 0, 0); |
| 1134 |
++numIter; |
| 1135 |
bool changed = jacobiRotation<0, 1, 2> (A, V, Z, tol); |
| 1136 |
changed = jacobiRotation<0, 2, 1> (A, V, Z, tol) || changed; |
| 1137 |
changed = jacobiRotation<1, 2, 0> (A, V, Z, tol) || changed; |
| 1138 |
// One sweep passed. Add accumulated changes (Z) to singular values (S) |
| 1139 |
// Update diagonal elements of A for better accuracy as well. |
| 1140 |
for (int i = 0; i < 3; ++i) |
| 1141 |
{ |
| 1142 |
A[i][i] = S[i] += Z[i]; |
| 1143 |
} |
| 1144 |
if (!changed) |
| 1145 |
break; |
| 1146 |
} while (maxOffDiagSymm (A) > absTol && numIter < maxIter); |
| 1147 |
} |
| 1148 |
} |
| 1149 |
|
| 1150 |
template <typename T> |
| 1151 |
void |
| 1152 |
jacobiEigenSolver (Matrix44<T>& A, Vec4<T>& S, Matrix44<T>& V, const T tol) |
| 1153 |
{ |
| 1154 |
V.makeIdentity(); |
| 1155 |
|
| 1156 |
for (int i = 0; i < 4; ++i) |
| 1157 |
{ |
| 1158 |
S[i] = A[i][i]; |
| 1159 |
} |
| 1160 |
|
| 1161 |
const int maxIter = 20; // In case we get really unlucky, prevents infinite loops |
| 1162 |
const T absTol = tol * maxOffDiagSymm (A); // Tolerance is in terms of the maximum |
| 1163 |
if (absTol != 0) // _off-diagonal_ entry. |
| 1164 |
{ |
| 1165 |
int numIter = 0; |
| 1166 |
do |
| 1167 |
{ |
| 1168 |
++numIter; |
| 1169 |
Vec4<T> Z (0, 0, 0, 0); |
| 1170 |
bool changed = jacobiRotation<0, 1, 2, 3> (A, V, Z, tol); |
| 1171 |
changed = jacobiRotation<0, 2, 1, 3> (A, V, Z, tol) || changed; |
| 1172 |
changed = jacobiRotation<0, 3, 1, 2> (A, V, Z, tol) || changed; |
| 1173 |
changed = jacobiRotation<1, 2, 0, 3> (A, V, Z, tol) || changed; |
| 1174 |
changed = jacobiRotation<1, 3, 0, 2> (A, V, Z, tol) || changed; |
| 1175 |
changed = jacobiRotation<2, 3, 0, 1> (A, V, Z, tol) || changed; |
| 1176 |
for (int i = 0; i < 4; ++i) |
| 1177 |
{ |
| 1178 |
A[i][i] = S[i] += Z[i]; |
| 1179 |
} |
| 1180 |
if (!changed) |
| 1181 |
break; |
| 1182 |
} while (maxOffDiagSymm (A) > absTol && numIter < maxIter); |
| 1183 |
} |
| 1184 |
} |
| 1185 |
|
| 1186 |
template <typename TM, typename TV> |
| 1187 |
void |
| 1188 |
maxEigenVector (TM& A, TV& V) |
| 1189 |
{ |
| 1190 |
TV S; |
| 1191 |
TM MV; |
| 1192 |
jacobiEigenSolver (A, S, MV); |
| 1193 |
|
| 1194 |
int maxIdx (0); |
| 1195 |
for (unsigned int i = 1; i < TV::dimensions(); ++i) |
| 1196 |
{ |
| 1197 |
if (std::abs (S[i]) > std::abs (S[maxIdx])) |
| 1198 |
maxIdx = i; |
| 1199 |
} |
| 1200 |
|
| 1201 |
for (unsigned int i = 0; i < TV::dimensions(); ++i) |
| 1202 |
V[i] = MV[i][maxIdx]; |
| 1203 |
} |
| 1204 |
|
| 1205 |
template <typename TM, typename TV> |
| 1206 |
void |
| 1207 |
minEigenVector (TM& A, TV& V) |
| 1208 |
{ |
| 1209 |
TV S; |
| 1210 |
TM MV; |
| 1211 |
jacobiEigenSolver (A, S, MV); |
| 1212 |
|
| 1213 |
int minIdx (0); |
| 1214 |
for (unsigned int i = 1; i < TV::dimensions(); ++i) |
| 1215 |
{ |
| 1216 |
if (std::abs (S[i]) < std::abs (S[minIdx])) |
| 1217 |
minIdx = i; |
| 1218 |
} |
| 1219 |
|
| 1220 |
for (unsigned int i = 0; i < TV::dimensions(); ++i) |
| 1221 |
V[i] = MV[i][minIdx]; |
| 1222 |
} |
| 1223 |
|
| 1224 |
template IMATH_EXPORT void |
| 1225 |
jacobiEigenSolver (Matrix33<float>& A, Vec3<float>& S, Matrix33<float>& V, const float tol); |
| 1226 |
template IMATH_EXPORT void |
| 1227 |
jacobiEigenSolver (Matrix33<double>& A, Vec3<double>& S, Matrix33<double>& V, const double tol); |
| 1228 |
template IMATH_EXPORT void |
| 1229 |
jacobiEigenSolver (Matrix44<float>& A, Vec4<float>& S, Matrix44<float>& V, const float tol); |
| 1230 |
template IMATH_EXPORT void |
| 1231 |
jacobiEigenSolver (Matrix44<double>& A, Vec4<double>& S, Matrix44<double>& V, const double tol); |
| 1232 |
|
| 1233 |
template IMATH_EXPORT void maxEigenVector (Matrix33<float>& A, Vec3<float>& S); |
| 1234 |
template IMATH_EXPORT void maxEigenVector (Matrix44<float>& A, Vec4<float>& S); |
| 1235 |
template IMATH_EXPORT void maxEigenVector (Matrix33<double>& A, Vec3<double>& S); |
| 1236 |
template IMATH_EXPORT void maxEigenVector (Matrix44<double>& A, Vec4<double>& S); |
| 1237 |
|
| 1238 |
template IMATH_EXPORT void minEigenVector (Matrix33<float>& A, Vec3<float>& S); |
| 1239 |
template IMATH_EXPORT void minEigenVector (Matrix44<float>& A, Vec4<float>& S); |
| 1240 |
template IMATH_EXPORT void minEigenVector (Matrix33<double>& A, Vec3<double>& S); |
| 1241 |
template IMATH_EXPORT void minEigenVector (Matrix44<double>& A, Vec4<double>& S); |
| 1242 |
|
| 1243 |
|
| 1244 |
IMATH_INTERNAL_NAMESPACE_SOURCE_EXIT |
| 1245 |
|
| 1246 |
/// @endcond |