1*> \brief \b DORHR_COL 2* 3* =========== DOCUMENTATION =========== 4* 5* Online html documentation available at 6* http://www.netlib.org/lapack/explore-html/ 7* 8*> \htmlonly 9*> Download DORHR_COL + dependencies 10*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dorhr_col.f"> 11*> [TGZ]</a> 12*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dorhr_col.f"> 13*> [ZIP]</a> 14*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dorhr_col.f"> 15*> [TXT]</a> 16*> \endhtmlonly 17* 18* Definition: 19* =========== 20* 21* SUBROUTINE DORHR_COL( M, N, NB, A, LDA, T, LDT, D, INFO ) 22* 23* .. Scalar Arguments .. 24* INTEGER INFO, LDA, LDT, M, N, NB 25* .. 26* .. Array Arguments .. 27* DOUBLE PRECISION A( LDA, * ), D( * ), T( LDT, * ) 28* .. 29* 30*> \par Purpose: 31* ============= 32*> 33*> \verbatim 34*> 35*> DORHR_COL takes an M-by-N real matrix Q_in with orthonormal columns 36*> as input, stored in A, and performs Householder Reconstruction (HR), 37*> i.e. reconstructs Householder vectors V(i) implicitly representing 38*> another M-by-N matrix Q_out, with the property that Q_in = Q_out*S, 39*> where S is an N-by-N diagonal matrix with diagonal entries 40*> equal to +1 or -1. The Householder vectors (columns V(i) of V) are 41*> stored in A on output, and the diagonal entries of S are stored in D. 42*> Block reflectors are also returned in T 43*> (same output format as DGEQRT). 44*> \endverbatim 45* 46* Arguments: 47* ========== 48* 49*> \param[in] M 50*> \verbatim 51*> M is INTEGER 52*> The number of rows of the matrix A. M >= 0. 53*> \endverbatim 54*> 55*> \param[in] N 56*> \verbatim 57*> N is INTEGER 58*> The number of columns of the matrix A. M >= N >= 0. 59*> \endverbatim 60*> 61*> \param[in] NB 62*> \verbatim 63*> NB is INTEGER 64*> The column block size to be used in the reconstruction 65*> of Householder column vector blocks in the array A and 66*> corresponding block reflectors in the array T. NB >= 1. 67*> (Note that if NB > N, then N is used instead of NB 68*> as the column block size.) 69*> \endverbatim 70*> 71*> \param[in,out] A 72*> \verbatim 73*> A is DOUBLE PRECISION array, dimension (LDA,N) 74*> 75*> On entry: 76*> 77*> The array A contains an M-by-N orthonormal matrix Q_in, 78*> i.e the columns of A are orthogonal unit vectors. 79*> 80*> On exit: 81*> 82*> The elements below the diagonal of A represent the unit 83*> lower-trapezoidal matrix V of Householder column vectors 84*> V(i). The unit diagonal entries of V are not stored 85*> (same format as the output below the diagonal in A from 86*> DGEQRT). The matrix T and the matrix V stored on output 87*> in A implicitly define Q_out. 88*> 89*> The elements above the diagonal contain the factor U 90*> of the "modified" LU-decomposition: 91*> Q_in - ( S ) = V * U 92*> ( 0 ) 93*> where 0 is a (M-N)-by-(M-N) zero matrix. 94*> \endverbatim 95*> 96*> \param[in] LDA 97*> \verbatim 98*> LDA is INTEGER 99*> The leading dimension of the array A. LDA >= max(1,M). 100*> \endverbatim 101*> 102*> \param[out] T 103*> \verbatim 104*> T is DOUBLE PRECISION array, 105*> dimension (LDT, N) 106*> 107*> Let NOCB = Number_of_output_col_blocks 108*> = CEIL(N/NB) 109*> 110*> On exit, T(1:NB, 1:N) contains NOCB upper-triangular 111*> block reflectors used to define Q_out stored in compact 112*> form as a sequence of upper-triangular NB-by-NB column 113*> blocks (same format as the output T in DGEQRT). 114*> The matrix T and the matrix V stored on output in A 115*> implicitly define Q_out. NOTE: The lower triangles 116*> below the upper-triangular blocks will be filled with 117*> zeros. See Further Details. 118*> \endverbatim 119*> 120*> \param[in] LDT 121*> \verbatim 122*> LDT is INTEGER 123*> The leading dimension of the array T. 124*> LDT >= max(1,min(NB,N)). 125*> \endverbatim 126*> 127*> \param[out] D 128*> \verbatim 129*> D is DOUBLE PRECISION array, dimension min(M,N). 130*> The elements can be only plus or minus one. 131*> 132*> D(i) is constructed as D(i) = -SIGN(Q_in_i(i,i)), where 133*> 1 <= i <= min(M,N), and Q_in_i is Q_in after performing 134*> i-1 steps of “modified” Gaussian elimination. 135*> See Further Details. 136*> \endverbatim 137*> 138*> \param[out] INFO 139*> \verbatim 140*> INFO is INTEGER 141*> = 0: successful exit 142*> < 0: if INFO = -i, the i-th argument had an illegal value 143*> \endverbatim 144*> 145*> \par Further Details: 146* ===================== 147*> 148*> \verbatim 149*> 150*> The computed M-by-M orthogonal factor Q_out is defined implicitly as 151*> a product of orthogonal matrices Q_out(i). Each Q_out(i) is stored in 152*> the compact WY-representation format in the corresponding blocks of 153*> matrices V (stored in A) and T. 154*> 155*> The M-by-N unit lower-trapezoidal matrix V stored in the M-by-N 156*> matrix A contains the column vectors V(i) in NB-size column 157*> blocks VB(j). For example, VB(1) contains the columns 158*> V(1), V(2), ... V(NB). NOTE: The unit entries on 159*> the diagonal of Y are not stored in A. 160*> 161*> The number of column blocks is 162*> 163*> NOCB = Number_of_output_col_blocks = CEIL(N/NB) 164*> 165*> where each block is of order NB except for the last block, which 166*> is of order LAST_NB = N - (NOCB-1)*NB. 167*> 168*> For example, if M=6, N=5 and NB=2, the matrix V is 169*> 170*> 171*> V = ( VB(1), VB(2), VB(3) ) = 172*> 173*> = ( 1 ) 174*> ( v21 1 ) 175*> ( v31 v32 1 ) 176*> ( v41 v42 v43 1 ) 177*> ( v51 v52 v53 v54 1 ) 178*> ( v61 v62 v63 v54 v65 ) 179*> 180*> 181*> For each of the column blocks VB(i), an upper-triangular block 182*> reflector TB(i) is computed. These blocks are stored as 183*> a sequence of upper-triangular column blocks in the NB-by-N 184*> matrix T. The size of each TB(i) block is NB-by-NB, except 185*> for the last block, whose size is LAST_NB-by-LAST_NB. 186*> 187*> For example, if M=6, N=5 and NB=2, the matrix T is 188*> 189*> T = ( TB(1), TB(2), TB(3) ) = 190*> 191*> = ( t11 t12 t13 t14 t15 ) 192*> ( t22 t24 ) 193*> 194*> 195*> The M-by-M factor Q_out is given as a product of NOCB 196*> orthogonal M-by-M matrices Q_out(i). 197*> 198*> Q_out = Q_out(1) * Q_out(2) * ... * Q_out(NOCB), 199*> 200*> where each matrix Q_out(i) is given by the WY-representation 201*> using corresponding blocks from the matrices V and T: 202*> 203*> Q_out(i) = I - VB(i) * TB(i) * (VB(i))**T, 204*> 205*> where I is the identity matrix. Here is the formula with matrix 206*> dimensions: 207*> 208*> Q(i){M-by-M} = I{M-by-M} - 209*> VB(i){M-by-INB} * TB(i){INB-by-INB} * (VB(i))**T {INB-by-M}, 210*> 211*> where INB = NB, except for the last block NOCB 212*> for which INB=LAST_NB. 213*> 214*> ===== 215*> NOTE: 216*> ===== 217*> 218*> If Q_in is the result of doing a QR factorization 219*> B = Q_in * R_in, then: 220*> 221*> B = (Q_out*S) * R_in = Q_out * (S * R_in) = Q_out * R_out. 222*> 223*> So if one wants to interpret Q_out as the result 224*> of the QR factorization of B, then the corresponding R_out 225*> should be equal to R_out = S * R_in, i.e. some rows of R_in 226*> should be multiplied by -1. 227*> 228*> For the details of the algorithm, see [1]. 229*> 230*> [1] "Reconstructing Householder vectors from tall-skinny QR", 231*> G. Ballard, J. Demmel, L. Grigori, M. Jacquelin, H.D. Nguyen, 232*> E. Solomonik, J. Parallel Distrib. Comput., 233*> vol. 85, pp. 3-31, 2015. 234*> \endverbatim 235*> 236* Authors: 237* ======== 238* 239*> \author Univ. of Tennessee 240*> \author Univ. of California Berkeley 241*> \author Univ. of Colorado Denver 242*> \author NAG Ltd. 243* 244*> \ingroup doubleOTHERcomputational 245* 246*> \par Contributors: 247* ================== 248*> 249*> \verbatim 250*> 251*> November 2019, Igor Kozachenko, 252*> Computer Science Division, 253*> University of California, Berkeley 254*> 255*> \endverbatim 256* 257* ===================================================================== 258 SUBROUTINE DORHR_COL( M, N, NB, A, LDA, T, LDT, D, INFO ) 259 IMPLICIT NONE 260* 261* -- LAPACK computational routine -- 262* -- LAPACK is a software package provided by Univ. of Tennessee, -- 263* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- 264* 265* .. Scalar Arguments .. 266 INTEGER INFO, LDA, LDT, M, N, NB 267* .. 268* .. Array Arguments .. 269 DOUBLE PRECISION A( LDA, * ), D( * ), T( LDT, * ) 270* .. 271* 272* ===================================================================== 273* 274* .. Parameters .. 275 DOUBLE PRECISION ONE, ZERO 276 PARAMETER ( ONE = 1.0D+0, ZERO = 0.0D+0 ) 277* .. 278* .. Local Scalars .. 279 INTEGER I, IINFO, J, JB, JBTEMP1, JBTEMP2, JNB, 280 $ NPLUSONE 281* .. 282* .. External Subroutines .. 283 EXTERNAL DCOPY, DLAORHR_COL_GETRFNP, DSCAL, DTRSM, 284 $ XERBLA 285* .. 286* .. Intrinsic Functions .. 287 INTRINSIC MAX, MIN 288* .. 289* .. Executable Statements .. 290* 291* Test the input parameters 292* 293 INFO = 0 294 IF( M.LT.0 ) THEN 295 INFO = -1 296 ELSE IF( N.LT.0 .OR. N.GT.M ) THEN 297 INFO = -2 298 ELSE IF( NB.LT.1 ) THEN 299 INFO = -3 300 ELSE IF( LDA.LT.MAX( 1, M ) ) THEN 301 INFO = -5 302 ELSE IF( LDT.LT.MAX( 1, MIN( NB, N ) ) ) THEN 303 INFO = -7 304 END IF 305* 306* Handle error in the input parameters. 307* 308 IF( INFO.NE.0 ) THEN 309 CALL XERBLA( 'DORHR_COL', -INFO ) 310 RETURN 311 END IF 312* 313* Quick return if possible 314* 315 IF( MIN( M, N ).EQ.0 ) THEN 316 RETURN 317 END IF 318* 319* On input, the M-by-N matrix A contains the orthogonal 320* M-by-N matrix Q_in. 321* 322* (1) Compute the unit lower-trapezoidal V (ones on the diagonal 323* are not stored) by performing the "modified" LU-decomposition. 324* 325* Q_in - ( S ) = V * U = ( V1 ) * U, 326* ( 0 ) ( V2 ) 327* 328* where 0 is an (M-N)-by-N zero matrix. 329* 330* (1-1) Factor V1 and U. 331 332 CALL DLAORHR_COL_GETRFNP( N, N, A, LDA, D, IINFO ) 333* 334* (1-2) Solve for V2. 335* 336 IF( M.GT.N ) THEN 337 CALL DTRSM( 'R', 'U', 'N', 'N', M-N, N, ONE, A, LDA, 338 $ A( N+1, 1 ), LDA ) 339 END IF 340* 341* (2) Reconstruct the block reflector T stored in T(1:NB, 1:N) 342* as a sequence of upper-triangular blocks with NB-size column 343* blocking. 344* 345* Loop over the column blocks of size NB of the array A(1:M,1:N) 346* and the array T(1:NB,1:N), JB is the column index of a column 347* block, JNB is the column block size at each step JB. 348* 349 NPLUSONE = N + 1 350 DO JB = 1, N, NB 351* 352* (2-0) Determine the column block size JNB. 353* 354 JNB = MIN( NPLUSONE-JB, NB ) 355* 356* (2-1) Copy the upper-triangular part of the current JNB-by-JNB 357* diagonal block U(JB) (of the N-by-N matrix U) stored 358* in A(JB:JB+JNB-1,JB:JB+JNB-1) into the upper-triangular part 359* of the current JNB-by-JNB block T(1:JNB,JB:JB+JNB-1) 360* column-by-column, total JNB*(JNB+1)/2 elements. 361* 362 JBTEMP1 = JB - 1 363 DO J = JB, JB+JNB-1 364 CALL DCOPY( J-JBTEMP1, A( JB, J ), 1, T( 1, J ), 1 ) 365 END DO 366* 367* (2-2) Perform on the upper-triangular part of the current 368* JNB-by-JNB diagonal block U(JB) (of the N-by-N matrix U) stored 369* in T(1:JNB,JB:JB+JNB-1) the following operation in place: 370* (-1)*U(JB)*S(JB), i.e the result will be stored in the upper- 371* triangular part of T(1:JNB,JB:JB+JNB-1). This multiplication 372* of the JNB-by-JNB diagonal block U(JB) by the JNB-by-JNB 373* diagonal block S(JB) of the N-by-N sign matrix S from the 374* right means changing the sign of each J-th column of the block 375* U(JB) according to the sign of the diagonal element of the block 376* S(JB), i.e. S(J,J) that is stored in the array element D(J). 377* 378 DO J = JB, JB+JNB-1 379 IF( D( J ).EQ.ONE ) THEN 380 CALL DSCAL( J-JBTEMP1, -ONE, T( 1, J ), 1 ) 381 END IF 382 END DO 383* 384* (2-3) Perform the triangular solve for the current block 385* matrix X(JB): 386* 387* X(JB) * (A(JB)**T) = B(JB), where: 388* 389* A(JB)**T is a JNB-by-JNB unit upper-triangular 390* coefficient block, and A(JB)=V1(JB), which 391* is a JNB-by-JNB unit lower-triangular block 392* stored in A(JB:JB+JNB-1,JB:JB+JNB-1). 393* The N-by-N matrix V1 is the upper part 394* of the M-by-N lower-trapezoidal matrix V 395* stored in A(1:M,1:N); 396* 397* B(JB) is a JNB-by-JNB upper-triangular right-hand 398* side block, B(JB) = (-1)*U(JB)*S(JB), and 399* B(JB) is stored in T(1:JNB,JB:JB+JNB-1); 400* 401* X(JB) is a JNB-by-JNB upper-triangular solution 402* block, X(JB) is the upper-triangular block 403* reflector T(JB), and X(JB) is stored 404* in T(1:JNB,JB:JB+JNB-1). 405* 406* In other words, we perform the triangular solve for the 407* upper-triangular block T(JB): 408* 409* T(JB) * (V1(JB)**T) = (-1)*U(JB)*S(JB). 410* 411* Even though the blocks X(JB) and B(JB) are upper- 412* triangular, the routine DTRSM will access all JNB**2 413* elements of the square T(1:JNB,JB:JB+JNB-1). Therefore, 414* we need to set to zero the elements of the block 415* T(1:JNB,JB:JB+JNB-1) below the diagonal before the call 416* to DTRSM. 417* 418* (2-3a) Set the elements to zero. 419* 420 JBTEMP2 = JB - 2 421 DO J = JB, JB+JNB-2 422 DO I = J-JBTEMP2, NB 423 T( I, J ) = ZERO 424 END DO 425 END DO 426* 427* (2-3b) Perform the triangular solve. 428* 429 CALL DTRSM( 'R', 'L', 'T', 'U', JNB, JNB, ONE, 430 $ A( JB, JB ), LDA, T( 1, JB ), LDT ) 431* 432 END DO 433* 434 RETURN 435* 436* End of DORHR_COL 437* 438 END 439