| 1 |
// $Id$ |
| 2 |
/* |
| 3 |
************************************************************ |
| 4 |
* Copyright 2006 by ACcESS MNRF * |
| 5 |
* * |
| 6 |
* http://www.access.edu.au * |
| 7 |
* Primary Business: Queensland, Australia * |
| 8 |
* Licensed under the Open Software License version 3.0 * |
| 9 |
* http://www.opensource.org/licenses/osl-3.0.php * |
| 10 |
* * |
| 11 |
************************************************************ |
| 12 |
*/ |
| 13 |
|
| 14 |
#if !defined escript_LocalOps_H |
| 15 |
#define escript_LocalOps_H |
| 16 |
#ifdef __INTEL_COMPILER |
| 17 |
#include <mathimf.h> |
| 18 |
# define M_PI 3.14159265358979323846 /* pi */ |
| 19 |
#else |
| 20 |
#include <math.h> |
| 21 |
#endif |
| 22 |
|
| 23 |
namespace escript { |
| 24 |
|
| 25 |
|
| 26 |
/** |
| 27 |
\brief |
| 28 |
solves a 1x1 eigenvalue A*V=ev*V problem |
| 29 |
|
| 30 |
\param A00 Input - A_00 |
| 31 |
\param ev0 Output - eigenvalue |
| 32 |
*/ |
| 33 |
inline |
| 34 |
void eigenvalues1(const double A00,double* ev0) { |
| 35 |
|
| 36 |
*ev0=A00; |
| 37 |
|
| 38 |
} |
| 39 |
/** |
| 40 |
\brief |
| 41 |
solves a 2x2 eigenvalue A*V=ev*V problem for symmetric A |
| 42 |
|
| 43 |
\param A00 Input - A_00 |
| 44 |
\param A01 Input - A_01 |
| 45 |
\param A11 Input - A_11 |
| 46 |
\param ev0 Output - smallest eigenvalue |
| 47 |
\param ev1 Output - largest eigenvalue |
| 48 |
*/ |
| 49 |
inline |
| 50 |
void eigenvalues2(const double A00,const double A01,const double A11, |
| 51 |
double* ev0, double* ev1) { |
| 52 |
const register double trA=(A00+A11)/2.; |
| 53 |
const register double A_00=A00-trA; |
| 54 |
const register double A_11=A11-trA; |
| 55 |
const register double s=sqrt(A01*A01-A_00*A_11); |
| 56 |
*ev0=trA-s; |
| 57 |
*ev1=trA+s; |
| 58 |
} |
| 59 |
/** |
| 60 |
\brief |
| 61 |
solves a 3x3 eigenvalue A*V=ev*V problem for symmetric A |
| 62 |
|
| 63 |
\param A00 Input - A_00 |
| 64 |
\param A01 Input - A_01 |
| 65 |
\param A02 Input - A_02 |
| 66 |
\param A11 Input - A_11 |
| 67 |
\param A12 Input - A_12 |
| 68 |
\param A22 Input - A_22 |
| 69 |
\param ev0 Output - smallest eigenvalue |
| 70 |
\param ev1 Output - eigenvalue |
| 71 |
\param ev2 Output - largest eigenvalue |
| 72 |
*/ |
| 73 |
inline |
| 74 |
void eigenvalues3(const double A00, const double A01, const double A02, |
| 75 |
const double A11, const double A12, |
| 76 |
const double A22, |
| 77 |
double* ev0, double* ev1,double* ev2) { |
| 78 |
|
| 79 |
const register double trA=(A00+A11+A22)/3.; |
| 80 |
const register double A_00=A00-trA; |
| 81 |
const register double A_11=A11-trA; |
| 82 |
const register double A_22=A22-trA; |
| 83 |
const register double A01_2=A01*A01; |
| 84 |
const register double A02_2=A02*A02; |
| 85 |
const register double A12_2=A12*A12; |
| 86 |
const register double p=A02_2+A12_2+A01_2+(A_00*A_00+A_11*A_11+A_22*A_22)/2.; |
| 87 |
if (p<=0.) { |
| 88 |
*ev2=trA; |
| 89 |
*ev1=trA; |
| 90 |
*ev0=trA; |
| 91 |
|
| 92 |
} else { |
| 93 |
const register double q=(A02_2*A_11+A12_2*A_00+A01_2*A_22)-(A_00*A_11*A_22+2*A01*A12*A02); |
| 94 |
const register double sq_p=sqrt(p/3.); |
| 95 |
register double z=-q/(2*pow(sq_p,3)); |
| 96 |
if (z<-1.) { |
| 97 |
z=-1.; |
| 98 |
} else if (z>1.) { |
| 99 |
z=1.; |
| 100 |
} |
| 101 |
const register double alpha_3=acos(z)/3.; |
| 102 |
*ev2=trA+2.*sq_p*cos(alpha_3); |
| 103 |
*ev1=trA-2.*sq_p*cos(alpha_3+M_PI/3.); |
| 104 |
*ev0=trA-2.*sq_p*cos(alpha_3-M_PI/3.); |
| 105 |
} |
| 106 |
} |
| 107 |
/** |
| 108 |
\brief |
| 109 |
solves a 1x1 eigenvalue A*V=ev*V problem for symmetric A |
| 110 |
|
| 111 |
\param A00 Input - A_00 |
| 112 |
\param ev0 Output - eigenvalue |
| 113 |
\param V00 Output - eigenvector |
| 114 |
\param tol Input - tolerance to identify to eigenvalues |
| 115 |
*/ |
| 116 |
inline |
| 117 |
void eigenvalues_and_eigenvectors1(const double A00,double* ev0,double* V00,const double tol) |
| 118 |
{ |
| 119 |
eigenvalues1(A00,ev0); |
| 120 |
*V00=1.; |
| 121 |
return; |
| 122 |
} |
| 123 |
/** |
| 124 |
\brief |
| 125 |
returns a non-zero vector in the kernel of [[A00,A01],[A01,A11]] assuming that the kernel dimension is at least 1. |
| 126 |
|
| 127 |
\param A00 Input - matrix component |
| 128 |
\param A10 Input - matrix component |
| 129 |
\param A01 Input - matrix component |
| 130 |
\param A11 Input - matrix component |
| 131 |
\param V0 Output - vector component |
| 132 |
\param V1 Output - vector component |
| 133 |
*/ |
| 134 |
inline |
| 135 |
void vectorInKernel2(const double A00,const double A10,const double A01,const double A11, |
| 136 |
double* V0, double*V1) |
| 137 |
{ |
| 138 |
register double absA00=fabs(A00); |
| 139 |
register double absA10=fabs(A10); |
| 140 |
register double absA01=fabs(A01); |
| 141 |
register double absA11=fabs(A11); |
| 142 |
register double m=absA11>absA10 ? absA11 : absA10; |
| 143 |
if (absA00>m || absA01>m) { |
| 144 |
*V0=-A01; |
| 145 |
*V1=A00; |
| 146 |
} else { |
| 147 |
if (m<=0) { |
| 148 |
*V0=1.; |
| 149 |
*V1=0.; |
| 150 |
} else { |
| 151 |
*V0=A11; |
| 152 |
*V1=-A10; |
| 153 |
} |
| 154 |
} |
| 155 |
} |
| 156 |
/** |
| 157 |
\brief |
| 158 |
returns a non-zero vector in the kernel of [[A00,A01,A02],[A10,A11,A12],[A20,A21,A22]] |
| 159 |
assuming that the kernel dimension is at least 1 and A00 is non zero. |
| 160 |
|
| 161 |
\param A00 Input - matrix component |
| 162 |
\param A10 Input - matrix component |
| 163 |
\param A20 Input - matrix component |
| 164 |
\param A01 Input - matrix component |
| 165 |
\param A11 Input - matrix component |
| 166 |
\param A21 Input - matrix component |
| 167 |
\param A02 Input - matrix component |
| 168 |
\param A12 Input - matrix component |
| 169 |
\param A22 Input - matrix component |
| 170 |
\param V0 Output - vector component |
| 171 |
\param V1 Output - vector component |
| 172 |
\param V2 Output - vector component |
| 173 |
*/ |
| 174 |
inline |
| 175 |
void vectorInKernel3__nonZeroA00(const double A00,const double A10,const double A20, |
| 176 |
const double A01,const double A11,const double A21, |
| 177 |
const double A02,const double A12,const double A22, |
| 178 |
double* V0,double* V1,double* V2) |
| 179 |
{ |
| 180 |
double TEMP0,TEMP1; |
| 181 |
register const double I00=1./A00; |
| 182 |
register const double IA10=I00*A10; |
| 183 |
register const double IA20=I00*A20; |
| 184 |
vectorInKernel2(A11-IA10*A01,A12-IA10*A02, |
| 185 |
A21-IA20*A01,A22-IA20*A02,&TEMP0,&TEMP1); |
| 186 |
*V0=-(A10*TEMP0+A20*TEMP1); |
| 187 |
*V1=A00*TEMP0; |
| 188 |
*V2=A00*TEMP1; |
| 189 |
} |
| 190 |
|
| 191 |
/** |
| 192 |
\brief |
| 193 |
solves a 2x2 eigenvalue A*V=ev*V problem for symmetric A. Eigenvectors are |
| 194 |
ordered by increasing value and eigen vectors are normalizeVector3d such that |
| 195 |
length is zero and first non-zero component is positive. |
| 196 |
|
| 197 |
\param A00 Input - A_00 |
| 198 |
\param A01 Input - A_01 |
| 199 |
\param A11 Input - A_11 |
| 200 |
\param ev0 Output - smallest eigenvalue |
| 201 |
\param ev1 Output - eigenvalue |
| 202 |
\param V00 Output - eigenvector componenent coresponding to ev0 |
| 203 |
\param V10 Output - eigenvector componenent coresponding to ev0 |
| 204 |
\param V01 Output - eigenvector componenent coresponding to ev1 |
| 205 |
\param V11 Output - eigenvector componenent coresponding to ev1 |
| 206 |
\param tol Input - tolerance to identify to eigenvalues |
| 207 |
*/ |
| 208 |
inline |
| 209 |
void eigenvalues_and_eigenvectors2(const double A00,const double A01,const double A11, |
| 210 |
double* ev0, double* ev1, |
| 211 |
double* V00, double* V10, double* V01, double* V11, |
| 212 |
const double tol) |
| 213 |
{ |
| 214 |
double TEMP0,TEMP1; |
| 215 |
eigenvalues2(A00,A01,A11,ev0,ev1); |
| 216 |
const register double absev0=fabs(*ev0); |
| 217 |
const register double absev1=fabs(*ev1); |
| 218 |
register double max_ev=absev0>absev1 ? absev0 : absev1; |
| 219 |
if (fabs((*ev0)-(*ev1))<tol*max_ev) { |
| 220 |
*V00=1.; |
| 221 |
*V10=0.; |
| 222 |
*V01=0.; |
| 223 |
*V11=1.; |
| 224 |
} else { |
| 225 |
vectorInKernel2(A00-(*ev0),A01,A01,A11-(*ev0),&TEMP0,&TEMP1); |
| 226 |
const register double scale=1./sqrt(TEMP0*TEMP0+TEMP1*TEMP1); |
| 227 |
if (TEMP0<0.) { |
| 228 |
*V00=-TEMP0*scale; |
| 229 |
*V10=-TEMP1*scale; |
| 230 |
if (TEMP1<0.) { |
| 231 |
*V01= *V10; |
| 232 |
*V11=-(*V00); |
| 233 |
} else { |
| 234 |
*V01=-(*V10); |
| 235 |
*V11= (*V10); |
| 236 |
} |
| 237 |
} else if (TEMP0>0.) { |
| 238 |
*V00=TEMP0*scale; |
| 239 |
*V10=TEMP1*scale; |
| 240 |
if (TEMP1<0.) { |
| 241 |
*V01=-(*V10); |
| 242 |
*V11= (*V00); |
| 243 |
} else { |
| 244 |
*V01= (*V10); |
| 245 |
*V11=-(*V00); |
| 246 |
} |
| 247 |
} else { |
| 248 |
*V00=0.; |
| 249 |
*V10=1; |
| 250 |
*V11=0.; |
| 251 |
*V01=1.; |
| 252 |
} |
| 253 |
} |
| 254 |
} |
| 255 |
/** |
| 256 |
\brief |
| 257 |
nomalizes a 3-d vector such that length is one and first non-zero component is positive. |
| 258 |
|
| 259 |
\param V0 - vector componenent |
| 260 |
\param V1 - vector componenent |
| 261 |
\param V2 - vector componenent |
| 262 |
*/ |
| 263 |
inline |
| 264 |
void normalizeVector3(double* V0,double* V1,double* V2) |
| 265 |
{ |
| 266 |
register double s; |
| 267 |
if (*V0>0) { |
| 268 |
s=1./sqrt((*V0)*(*V0)+(*V1)*(*V1)+(*V2)*(*V2)); |
| 269 |
*V0*=s; |
| 270 |
*V1*=s; |
| 271 |
*V2*=s; |
| 272 |
} else if (*V0<0) { |
| 273 |
s=-1./sqrt((*V0)*(*V0)+(*V1)*(*V1)+(*V2)*(*V2)); |
| 274 |
*V0*=s; |
| 275 |
*V1*=s; |
| 276 |
*V2*=s; |
| 277 |
} else { |
| 278 |
if (*V1>0) { |
| 279 |
s=1./sqrt((*V1)*(*V1)+(*V2)*(*V2)); |
| 280 |
*V1*=s; |
| 281 |
*V2*=s; |
| 282 |
} else if (*V1<0) { |
| 283 |
s=-1./sqrt((*V1)*(*V1)+(*V2)*(*V2)); |
| 284 |
*V1*=s; |
| 285 |
*V2*=s; |
| 286 |
} else { |
| 287 |
*V2=1.; |
| 288 |
} |
| 289 |
} |
| 290 |
} |
| 291 |
/** |
| 292 |
\brief |
| 293 |
solves a 2x2 eigenvalue A*V=ev*V problem for symmetric A. Eigenvectors are |
| 294 |
ordered by increasing value and eigen vectors are normalizeVector3d such that |
| 295 |
length is zero and first non-zero component is positive. |
| 296 |
|
| 297 |
\param A00 Input - A_00 |
| 298 |
\param A01 Input - A_01 |
| 299 |
\param A11 Input - A_11 |
| 300 |
\param ev0 Output - smallest eigenvalue |
| 301 |
\param ev1 Output - eigenvalue |
| 302 |
\param V00 Output - eigenvector componenent coresponding to ev0 |
| 303 |
\param V10 Output - eigenvector componenent coresponding to ev0 |
| 304 |
\param V01 Output - eigenvector componenent coresponding to ev1 |
| 305 |
\param V11 Output - eigenvector componenent coresponding to ev1 |
| 306 |
\param tol Input - tolerance to identify to eigenvalues |
| 307 |
*/ |
| 308 |
inline |
| 309 |
void eigenvalues_and_eigenvectors3(const double A00, const double A01, const double A02, |
| 310 |
const double A11, const double A12, const double A22, |
| 311 |
double* ev0, double* ev1, double* ev2, |
| 312 |
double* V00, double* V10, double* V20, |
| 313 |
double* V01, double* V11, double* V21, |
| 314 |
double* V02, double* V12, double* V22, |
| 315 |
const double tol) |
| 316 |
{ |
| 317 |
register const double absA01=fabs(A01); |
| 318 |
register const double absA02=fabs(A02); |
| 319 |
register const double m=absA01>absA02 ? absA01 : absA02; |
| 320 |
if (m<=0) { |
| 321 |
double TEMP_V00,TEMP_V10,TEMP_V01,TEMP_V11,TEMP_EV0,TEMP_EV1; |
| 322 |
eigenvalues_and_eigenvectors2(A11,A12,A22, |
| 323 |
&TEMP_EV0,&TEMP_EV1, |
| 324 |
&TEMP_V00,&TEMP_V10,&TEMP_V01,&TEMP_V11,tol); |
| 325 |
if (A00<=TEMP_EV0) { |
| 326 |
*V00=1.; |
| 327 |
*V10=0.; |
| 328 |
*V20=0.; |
| 329 |
*V01=0.; |
| 330 |
*V11=TEMP_V00; |
| 331 |
*V21=TEMP_V10; |
| 332 |
*V02=0.; |
| 333 |
*V12=TEMP_V01; |
| 334 |
*V22=TEMP_V11; |
| 335 |
*ev0=A00; |
| 336 |
*ev1=TEMP_EV0; |
| 337 |
*ev2=TEMP_EV1; |
| 338 |
} else if (A00>TEMP_EV1) { |
| 339 |
*V02=1.; |
| 340 |
*V12=0.; |
| 341 |
*V22=0.; |
| 342 |
*V00=0.; |
| 343 |
*V10=TEMP_V00; |
| 344 |
*V20=TEMP_V10; |
| 345 |
*V01=0.; |
| 346 |
*V11=TEMP_V01; |
| 347 |
*V21=TEMP_V11; |
| 348 |
*ev0=TEMP_EV0; |
| 349 |
*ev1=TEMP_EV1; |
| 350 |
*ev2=A00; |
| 351 |
} else { |
| 352 |
*V01=1.; |
| 353 |
*V11=0.; |
| 354 |
*V21=0.; |
| 355 |
*V00=0.; |
| 356 |
*V10=TEMP_V00; |
| 357 |
*V20=TEMP_V10; |
| 358 |
*V02=0.; |
| 359 |
*V12=TEMP_V01; |
| 360 |
*V22=TEMP_V11; |
| 361 |
*ev0=TEMP_EV0; |
| 362 |
*ev1=A00; |
| 363 |
*ev2=TEMP_EV1; |
| 364 |
} |
| 365 |
} else { |
| 366 |
eigenvalues3(A00,A01,A02,A11,A12,A22,ev0,ev1,ev2); |
| 367 |
const register double absev0=fabs(*ev0); |
| 368 |
const register double absev1=fabs(*ev1); |
| 369 |
const register double absev2=fabs(*ev2); |
| 370 |
register double max_ev=absev0>absev1 ? absev0 : absev1; |
| 371 |
max_ev=max_ev>absev2 ? max_ev : absev2; |
| 372 |
const register double d_01=fabs((*ev0)-(*ev1)); |
| 373 |
const register double d_12=fabs((*ev1)-(*ev2)); |
| 374 |
const register double max_d=d_01>d_12 ? d_01 : d_12; |
| 375 |
if (max_d<=tol*max_ev) { |
| 376 |
*V00=1.; |
| 377 |
*V10=0; |
| 378 |
*V20=0; |
| 379 |
*V01=0; |
| 380 |
*V11=1.; |
| 381 |
*V21=0; |
| 382 |
*V02=0; |
| 383 |
*V12=0; |
| 384 |
*V22=1.; |
| 385 |
} else { |
| 386 |
const register double S00=A00-(*ev0); |
| 387 |
const register double absS00=fabs(S00); |
| 388 |
if (fabs(S00)>m) { |
| 389 |
vectorInKernel3__nonZeroA00(S00,A01,A02,A01,A11-(*ev0),A12,A02,A12,A22-(*ev0),V00,V10,V20); |
| 390 |
} else if (absA02<m) { |
| 391 |
vectorInKernel3__nonZeroA00(A01,A11-(*ev0),A12,S00,A01,A02,A02,A12,A22-(*ev0),V00,V10,V20); |
| 392 |
} else { |
| 393 |
vectorInKernel3__nonZeroA00(A02,A12,A22-(*ev0),S00,A01,A02,A01,A11-(*ev0),A12,V00,V10,V20); |
| 394 |
} |
| 395 |
normalizeVector3(V00,V10,V20);; |
| 396 |
const register double T00=A00-(*ev2); |
| 397 |
const register double absT00=fabs(T00); |
| 398 |
if (fabs(T00)>m) { |
| 399 |
vectorInKernel3__nonZeroA00(T00,A01,A02,A01,A11-(*ev2),A12,A02,A12,A22-(*ev2),V02,V12,V22); |
| 400 |
} else if (absA02<m) { |
| 401 |
vectorInKernel3__nonZeroA00(A01,A11-(*ev2),A12,T00,A01,A02,A02,A12,A22-(*ev2),V02,V12,V22); |
| 402 |
} else { |
| 403 |
vectorInKernel3__nonZeroA00(A02,A12,A22-(*ev2),T00,A01,A02,A01,A11-(*ev2),A12,V02,V12,V22); |
| 404 |
} |
| 405 |
const register double dot=(*V02)*(*V00)+(*V12)*(*V10)+(*V22)*(*V20); |
| 406 |
*V02-=dot*(*V00); |
| 407 |
*V12-=dot*(*V10); |
| 408 |
*V22-=dot*(*V20); |
| 409 |
normalizeVector3(V02,V12,V22); |
| 410 |
*V01=(*V10)*(*V22)-(*V12)*(*V20); |
| 411 |
*V11=(*V20)*(*V02)-(*V00)*(*V22); |
| 412 |
*V21=(*V00)*(*V12)-(*V02)*(*V10); |
| 413 |
normalizeVector3(V01,V11,V21); |
| 414 |
} |
| 415 |
} |
| 416 |
} |
| 417 |
|
| 418 |
// General tensor product: arg_2(SL x SR) = arg_0(SL x SM) * arg_1(SM x SR) |
| 419 |
// SM is the product of the last axis_offset entries in arg_0.getShape(). |
| 420 |
inline |
| 421 |
void matrix_matrix_product(const int SL, const int SM, const int SR, const double* A, const double* B, double* C, int transpose) |
| 422 |
{ |
| 423 |
if (transpose == 0) { |
| 424 |
for (int i=0; i<SL; i++) { |
| 425 |
for (int j=0; j<SR; j++) { |
| 426 |
double sum = 0.0; |
| 427 |
for (int l=0; l<SM; l++) { |
| 428 |
sum += A[i+SL*l] * B[l+SM*j]; |
| 429 |
} |
| 430 |
C[i+SL*j] = sum; |
| 431 |
} |
| 432 |
} |
| 433 |
} |
| 434 |
else if (transpose == 1) { |
| 435 |
for (int i=0; i<SL; i++) { |
| 436 |
for (int j=0; j<SR; j++) { |
| 437 |
double sum = 0.0; |
| 438 |
for (int l=0; l<SM; l++) { |
| 439 |
sum += A[i*SM+l] * B[l+SM*j]; |
| 440 |
} |
| 441 |
C[i+SL*j] = sum; |
| 442 |
} |
| 443 |
} |
| 444 |
} |
| 445 |
else if (transpose == 2) { |
| 446 |
for (int i=0; i<SL; i++) { |
| 447 |
for (int j=0; j<SR; j++) { |
| 448 |
double sum = 0.0; |
| 449 |
for (int l=0; l<SM; l++) { |
| 450 |
sum += A[i+SL*l] * B[l*SR+j]; |
| 451 |
} |
| 452 |
C[i+SL*j] = sum; |
| 453 |
} |
| 454 |
} |
| 455 |
} |
| 456 |
} |
| 457 |
|
| 458 |
} // end of namespace |
| 459 |
#endif |
Parent Directory
| 
Revision Log