| 1 |
// |
| 2 |
// ZQUATEV: Diagonalization of quaternionic matrices |
| 3 |
// File : blocked.cc |
| 4 |
// Copyright (c) 2016, Toru Shiozaki ([email protected]) |
| 5 |
// All rights reserved. |
| 6 |
// |
| 7 |
// Redistribution and use in source and binary forms, with or without |
| 8 |
// modification, are permitted provided that the following conditions are met: |
| 9 |
// |
| 10 |
// 1. Redistributions of source code must retain the above copyright notice, this |
| 11 |
// list of conditions and the following disclaimer. |
| 12 |
// 2. Redistributions in binary form must reproduce the above copyright notice, |
| 13 |
// this list of conditions and the following disclaimer in the documentation |
| 14 |
// and/or other materials provided with the distribution. |
| 15 |
// |
| 16 |
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND |
| 17 |
// ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED |
| 18 |
// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE |
| 19 |
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR |
| 20 |
// ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES |
| 21 |
// (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; |
| 22 |
// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND |
| 23 |
// ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
| 24 |
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
| 25 |
// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 26 |
// |
| 27 |
// The views and conclusions contained in the software and documentation are those |
| 28 |
// of the authors and should not be interpreted as representing official policies, |
| 29 |
// either expressed or implied, of the FreeBSD Project. |
| 30 |
// |
| 31 |
|
| 32 |
#include "zquatev.h" |
| 33 |
#include "f77.h" |
| 34 |
#include "supermat.h" |
| 35 |
#include <cassert> |
| 36 |
|
| 37 |
using namespace std; |
| 38 |
|
| 39 |
namespace ts { |
| 40 |
namespace impl { |
| 41 |
|
| 42 |
void panel_update(const int n, const int nb, |
| 43 |
complex<double>* const D0, complex<double>* const D1, |
| 44 |
complex<double>* const Q0, complex<double>* const Q1, |
| 45 |
const int ld, const int norig, complex<double>* const work_ptr) { |
| 46 |
|
| 47 |
assert(n > nb+1); // otherwise does not work |
| 48 |
|
| 49 |
complex<double>* cptr = work_ptr; |
| 50 |
auto work1_3n = cptr; cptr += (n-1)*3; |
| 51 |
auto work2_3nb = cptr; cptr += nb*3; |
| 52 |
auto work3_nb = cptr; cptr += nb; |
| 53 |
auto work4_nb = cptr; cptr += nb; |
| 54 |
SuperMatrix<3,3> T(cptr, nb, nb); cptr += 9*nb*nb; |
| 55 |
SuperMatrix<3,1> R(cptr, nb, nb); cptr += 3*nb*nb; |
| 56 |
SuperMatrix<1,3> S(cptr, nb, nb); cptr += 3*nb*nb; |
| 57 |
SuperMatrix<1,2> W(cptr, n-1, nb, n-1, 1); cptr += 2*(n-1)*nb; |
| 58 |
SuperMatrix<1,3> YD(cptr, n-1, nb, n-1, 1, true, true); cptr += 2*(n-1)*nb + n-1; |
| 59 |
SuperMatrix<1,3> ZD(cptr, n-1, nb, n-1, 1, true, true); cptr += 2*(n-1)*nb + n-1; |
| 60 |
SuperMatrix<1,3> YE(cptr, n-1, nb, n-1, 1, true, true); cptr += 2*(n-1)*nb + n-1; |
| 61 |
SuperMatrix<1,3> ZE(cptr, n-1, nb, n-1, 1, true, true); |
| 62 |
|
| 63 |
// TODO once the paper is out, add comments with equation numbers. |
| 64 |
for (int k = 0; k != nb; ++k) { |
| 65 |
// prepare the k-th column using compact WY-like representation |
| 66 |
if (k > 0) { |
| 67 |
SuperMatrix<1,1> d(D0+k*ld+1, n-1, 1, n-1, 1, false); |
| 68 |
SuperMatrix<1,1> e(D1+k*ld+1, n-1, 1, n-1, 1, false); |
| 69 |
|
| 70 |
SuperMatrix<3,1> x(work2_3nb, W.mptr(0), 1, W.mptr(0), 1); |
| 71 |
x.nptr(2) = 1; |
| 72 |
SuperMatrix<2,1> x2 = x.trunc<2>(); |
| 73 |
W.cut_row<0>(k-1, x2); |
| 74 |
x.data<2,0>(0,0) = 1.0; |
| 75 |
contract<_N>(1.0, YD, x, d); |
| 76 |
contract<_N>(1.0, YE, x, e); |
| 77 |
|
| 78 |
SuperMatrix<1,1> y(work1_3n, n-1, 1); |
| 79 |
contract<_N>(1.0, ZE, x, y); |
| 80 |
d.add_lastcolumn<0>(y); |
| 81 |
y.reset(); |
| 82 |
contract<_N>(-1.0, ZD, x, y); |
| 83 |
e.add_lastcolumn<0>(y); |
| 84 |
|
| 85 |
const int sf = k-1; |
| 86 |
auto ds = d.shift(sf); |
| 87 |
auto es = e.shift(sf); |
| 88 |
auto Ws = W.shift(sf); |
| 89 |
SuperMatrix<3,1> y1(work1_3n, k, 1, k, 1); |
| 90 |
y1.nptr(0) = W.mptr(0); |
| 91 |
y1.nptr(1) = W.mptr(1); |
| 92 |
auto y1s = y1.trunc<2>(); |
| 93 |
contract<_C>(1.0, W, d, y1s); |
| 94 |
zaxpy_(k, 1.0, d.block(0,0), 1, y1.ptr<2,0>(0,0), 1); |
| 95 |
SuperMatrix<3,1> y2(work2_3nb, k, 1); |
| 96 |
contract_tr<_C>(1.0, T, y1, y2, work4_nb); |
| 97 |
contract<_N>(1.0, Ws, y2.trunc<2>(), ds); |
| 98 |
ds.data<0,0>(0,0) += y2.data<2,0>(k-1,0); |
| 99 |
SuperMatrix<1,1> y3(work3_nb, k, 1); |
| 100 |
contract_tr<_C>(1.0, R, y1, y3, work4_nb); |
| 101 |
y3.conj(); |
| 102 |
y2.reset(); |
| 103 |
contract_tr<_C>(1.0, S, y3, y2, work4_nb); |
| 104 |
SuperMatrix<1,1> y4(work1_3n, n-sf, 1); |
| 105 |
contract<_N>(1.0, Ws, y2.trunc<2>(), y4); |
| 106 |
y4.data<0,0>(0,0) += y2.data<2,0>(k-1,0); |
| 107 |
y4.conj(); |
| 108 |
|
| 109 |
SuperMatrix<3,1> y5(work2_3nb, k, 1, k, 1); |
| 110 |
y5.nptr(0) = W.mptr(0); |
| 111 |
y5.nptr(1) = W.mptr(1); |
| 112 |
auto y5s = y5.trunc<2>(); |
| 113 |
contract<_T>(1.0, W, e, y5s); |
| 114 |
zaxpy_(k, 1.0, e.block(0,0), 1, y5.ptr<2,0>(0,0), 1); |
| 115 |
es.add_lastcolumn<0>(y4); |
| 116 |
|
| 117 |
SuperMatrix<3,1> y5x(work1_3n, y5); |
| 118 |
SuperMatrix<1,1> y6(work3_nb, k, 1); |
| 119 |
contract_tr<_T>(1.0, R, y5x, y6, work4_nb); |
| 120 |
SuperMatrix<3,1> y7(work2_3nb, k, 1); |
| 121 |
contract_tr<_C>(1.0, S, y6, y7, work4_nb); |
| 122 |
contract<_N>(-1.0, Ws, y7.trunc<2>(), ds); |
| 123 |
ds.data<0,0>(0,0) -= y7.data<2,0>(k-1,0); |
| 124 |
y7.reset(); |
| 125 |
contract_tr<_T>(1.0, T, y5x, y7, work4_nb); |
| 126 |
y7.conj(); |
| 127 |
SuperMatrix<1,1> y8(work1_3n, n-sf, 1); |
| 128 |
contract<_N>(1.0, Ws, y7.trunc<2>(), y8); |
| 129 |
y8.data<0,0>(0,0) += y7.data<2,0>(k-1,0); |
| 130 |
y8.conj(); |
| 131 |
es.add_lastcolumn<0>(y8); |
| 132 |
} |
| 133 |
if (k == n-1) break; |
| 134 |
|
| 135 |
const int len = n-k-1; |
| 136 |
complex<double> alpha = D1[k*ld+k+1]; |
| 137 |
if (len > 1) { |
| 138 |
complex<double>* dnow = D1+k*ld+k+1; |
| 139 |
complex<double> tau; |
| 140 |
dnow[0] = 1.0; |
| 141 |
zlarfg_(len, alpha, dnow+1, 1, tau); |
| 142 |
tau = conj(tau); |
| 143 |
conj_n(dnow, len); |
| 144 |
|
| 145 |
if (k == 0) { |
| 146 |
W.write_lastcolumn<0>(dnow, len); |
| 147 |
T.data<0,0>(0,0) = -tau; |
| 148 |
zgemv_("N", len, len, -tau, D0+ld+1, ld, dnow, 1, 0.0, YD.block(0,0), 1); |
| 149 |
zgemv_("N", len, len, -tau, D1+ld+1, ld, dnow, 1, 0.0, YE.block(0,0), 1); |
| 150 |
|
| 151 |
zaxpy_(len, -conj(tau)*zdotc_(len, dnow, 1, D0+1, 1), dnow, 1, D0+1, 1); |
| 152 |
} else { |
| 153 |
R.append_row<0>(); |
| 154 |
SuperMatrix<2,1> x(work2_3nb, k, 1); |
| 155 |
|
| 156 |
SuperMatrix<1,1> v(work1_3n, n-k-1, 1, n-k-1, 1); |
| 157 |
v.write_lastcolumn<0>(dnow, len); |
| 158 |
contract<_C>(-tau, W.shift(k), v, x); |
| 159 |
|
| 160 |
SuperMatrix<3,1> v2(work1_3n, k, 1); |
| 161 |
contract_tr<_N>(1.0, T.slice<0,2>(), x, v2, work4_nb); |
| 162 |
T.append_column<0>(v2); |
| 163 |
T.append_row<0,0>(k, -tau); |
| 164 |
|
| 165 |
SuperMatrix<1,1> v3(work3_nb, k, 1); |
| 166 |
contract_tr<_N>(1.0, S.slice<0,2>(), x, v3, work4_nb); |
| 167 |
S.append_column<0>(v3); |
| 168 |
W.append_column<0>(dnow, len, k); |
| 169 |
|
| 170 |
SuperMatrix<1,1> yx(work1_3n, n-1, 1); |
| 171 |
contract<_N>(1.0, YD.slice<0,2>(), x, yx); |
| 172 |
YD.append_column<0>(yx); |
| 173 |
zgemv_("N", n-1, len, -tau, D0+(k+1)*ld+1, ld, dnow, 1, 1.0, YD.ptr<0,0>(0,k), 1); |
| 174 |
yx.reset(); |
| 175 |
contract<_N>(1.0, YE.slice<0,2>(), x, yx); |
| 176 |
YE.append_column<0>(yx); |
| 177 |
zgemv_("N", n-1, len, -tau, D1+(k+1)*ld+1, ld, dnow, 1, 1.0, YE.ptr<0,0>(0,k), 1); |
| 178 |
yx.reset(); |
| 179 |
contract<_N>(1.0, ZD.slice<0,2>(), x, yx); |
| 180 |
ZD.append_column<0>(yx); |
| 181 |
yx.reset(); |
| 182 |
contract<_N>(1.0, ZE.slice<0,2>(), x, yx); |
| 183 |
ZE.append_column<0>(yx); |
| 184 |
|
| 185 |
zaxpy_(len, -conj(tau)*zdotc_(len, dnow, 1, D0+k*ld+k+1, 1), dnow, 1, D0+k*ld+k+1, 1); |
| 186 |
} |
| 187 |
} |
| 188 |
|
| 189 |
double c; |
| 190 |
complex<double> s, dum; |
| 191 |
zlartg_(D0[k+1+k*ld], alpha, c, s, dum); |
| 192 |
assert(abs(-conj(s)*D0[k+1+k*ld]+c*alpha) < 1.0e-10); |
| 193 |
D0[k+1+k*ld] = c*D0[k+1+k*ld] + s*alpha; |
| 194 |
|
| 195 |
const double cbar = c-1.0; |
| 196 |
const complex<double> sbar = conj(s); |
| 197 |
if (k == 0) { |
| 198 |
assert(abs(W.data<0,0>(0,0)-1.0)<1.0e-10); |
| 199 |
T.data<0,2>(0,0) = T.data<0,0>(0,0)*cbar; |
| 200 |
T.data<2,2>(0,0) = cbar; |
| 201 |
R.data<0,0>(0,0) = T.data<0,0>(0,0); |
| 202 |
R.data<2,0>(0,0) = 1.0; |
| 203 |
S.data<0,2>(0,0) = -sbar; |
| 204 |
|
| 205 |
auto YD0 = YD.slice<0>(); |
| 206 |
auto YE0 = YE.slice<0>(); |
| 207 |
YD.add_lastcolumn<2>(YD0, cbar); |
| 208 |
YE.add_lastcolumn<2>(YE0, cbar); |
| 209 |
zaxpy_(n-1, cbar, D0+ld+1, 1, YD.block(0,2), 1); |
| 210 |
zaxpy_(n-1, cbar, D1+ld+1, 1, YE.block(0,2), 1); |
| 211 |
SuperMatrix<1,1> yd0b(work1_3n, YD0); |
| 212 |
SuperMatrix<1,1> ye0b(work1_3n+n-1, YE0); |
| 213 |
zaxpy_(n-1, 1.0, D0+ld+1, 1, yd0b.block(0,0), 1); |
| 214 |
zaxpy_(n-1, 1.0, D1+ld+1, 1, ye0b.block(0,0), 1); |
| 215 |
yd0b.conj(); |
| 216 |
ye0b.conj(); |
| 217 |
ZD.add_lastcolumn<2>(yd0b, -sbar); |
| 218 |
ZE.add_lastcolumn<2>(ye0b, -sbar); |
| 219 |
} else { |
| 220 |
SuperMatrix<2,1> x(work2_3nb, k+1, 1); |
| 221 |
W.cut_row<0>(k, x); |
| 222 |
|
| 223 |
SuperMatrix<1,1> yx(work1_3n, n-1, 1); |
| 224 |
SuperMatrix<1,1> zx(work1_3n+n-1, n-1, 1); |
| 225 |
contract<_N>(1.0, YD.slice<0,2>(), x, yx); |
| 226 |
contract<_N>(1.0, ZD.slice<0,2>(), x, zx); |
| 227 |
fill_n(YD.block(0,2), n-1, 0.0); |
| 228 |
fill_n(ZD.block(0,2), n-1, 0.0); |
| 229 |
ZD.add_lastcolumn<2>(zx, cbar); |
| 230 |
zx.conj(); |
| 231 |
YD.add_lastcolumn<2>(zx, sbar); |
| 232 |
zaxpy_(n-1, 1.0, D0+(k+1)*ld+1, 1, yx.block(0,0), 1); |
| 233 |
YD.add_lastcolumn<2>(yx, cbar); |
| 234 |
yx.conj(); |
| 235 |
ZD.add_lastcolumn<2>(yx, -sbar); |
| 236 |
|
| 237 |
yx.reset(); |
| 238 |
zx.reset(); |
| 239 |
contract<_N>(1.0, YE.slice<0,2>(), x, yx); |
| 240 |
contract<_N>(1.0, ZE.slice<0,2>(), x, zx); |
| 241 |
fill_n(YE.block(0,2), n-1, 0.0); |
| 242 |
fill_n(ZE.block(0,2), n-1, 0.0); |
| 243 |
ZE.add_lastcolumn<2>(zx, cbar); |
| 244 |
zx.conj(); |
| 245 |
YE.add_lastcolumn<2>(zx, sbar); |
| 246 |
zaxpy_(n-1, 1.0, D1+(k+1)*ld+1, 1, yx.block(0,0), 1); |
| 247 |
YE.add_lastcolumn<2>(yx, cbar); |
| 248 |
yx.conj(); |
| 249 |
ZE.add_lastcolumn<2>(yx, -sbar); |
| 250 |
|
| 251 |
SuperMatrix<3,1> v(work1_3n, k+1, 1); |
| 252 |
contract_tr<_N>(1.0, T.slice<0,2>(), x, v, work4_nb); |
| 253 |
SuperMatrix<1,1> y(work3_nb, k+1, 1); |
| 254 |
contract_tr<_N>(1.0, S.slice<0,2>(), x, y, work4_nb); |
| 255 |
y.conj(); |
| 256 |
SuperMatrix<3,1> z(work2_3nb, k+1, 1); |
| 257 |
contract_tr<_N>(sbar, R, y, z, work4_nb); |
| 258 |
y.conj(); |
| 259 |
T.append_column<2>(z); |
| 260 |
T.add_lastcolumn<2>(v, cbar); |
| 261 |
T.append_row<2,2>(k, cbar); |
| 262 |
|
| 263 |
R.append_column<0>(v); |
| 264 |
R.append_row<2,0>(k,1.0); |
| 265 |
y.scale(cbar); |
| 266 |
S.append_column<2>(y); |
| 267 |
S.append_row<0,2>(k, -sbar); |
| 268 |
} |
| 269 |
|
| 270 |
if (len > 1) { |
| 271 |
complex<double>* dnow = D0+k*ld+k+1; |
| 272 |
complex<double> ctau; |
| 273 |
complex<double> alpha2 = dnow[0]; |
| 274 |
dnow[0] = 1.0; |
| 275 |
zlarfg_(len, alpha2, dnow+1, 1, ctau); |
| 276 |
|
| 277 |
if (k == 0) { |
| 278 |
W.write_lastcolumn<1>(dnow, len); |
| 279 |
const complex<double> zz = -ctau*zdotc_(len, W.block(0,0), 1, dnow, 1); |
| 280 |
T.data<0,1>(0,0) = zz*T.data<0,0>(0,0) -ctau*T.data<0,2>(0,0); |
| 281 |
T.data<2,1>(0,0) = -ctau*T.data<2,2>(0,0); |
| 282 |
T.data<1,1>(0,0) = -ctau; |
| 283 |
S.data<0,1>(0,0) = -ctau*S.data<0,2>(0,0); |
| 284 |
|
| 285 |
zgemv_("N", n-1, n-1, -ctau, D0+ld+1, ld, dnow, 1, 0.0, YD.block(0,1), 1); |
| 286 |
zgemv_("N", n-1, n-1, -ctau, D1+ld+1, ld, dnow, 1, 0.0, YE.block(0,1), 1); |
| 287 |
YD.add_lastcolumn<1>(YD.slice<0>(), zz); |
| 288 |
YE.add_lastcolumn<1>(YE.slice<0>(), zz); |
| 289 |
YD.add_lastcolumn<1>(YD.slice<2>(), -ctau); |
| 290 |
YE.add_lastcolumn<1>(YE.slice<2>(), -ctau); |
| 291 |
ZD.add_lastcolumn<1>(ZD.slice<2>(), -ctau); |
| 292 |
ZE.add_lastcolumn<1>(ZE.slice<2>(), -ctau); |
| 293 |
} else { |
| 294 |
R.append_row<1>(); |
| 295 |
SuperMatrix<3,1> x(work2_3nb, k+1, 1, k+1, 1); |
| 296 |
x.nptr(1) = k; |
| 297 |
|
| 298 |
SuperMatrix<1,1> v(work1_3n, n-k-1, 1, n-k-1, 1); |
| 299 |
v.write_lastcolumn<0>(dnow, len); |
| 300 |
SuperMatrix<2,1> x2 = x.trunc<2>(); |
| 301 |
contract<_C>(-ctau, W.shift(k), v, x2); |
| 302 |
x.data<2,0>(k,0) = -ctau; |
| 303 |
|
| 304 |
SuperMatrix<3,1> v2(work1_3n, k+1, 1); |
| 305 |
contract_tr<_N>(1.0, T, x, v2, work4_nb); |
| 306 |
T.append_column<1>(v2); |
| 307 |
T.append_row<1,1>(k, -ctau); |
| 308 |
|
| 309 |
SuperMatrix<1,1> v3(work3_nb, k+1, 1); |
| 310 |
contract_tr<_N>(1.0, S, x, v3, work4_nb); |
| 311 |
S.append_column<1>(v3); |
| 312 |
W.append_column<1>(dnow, len, k); |
| 313 |
|
| 314 |
x.data<2,0>(0,0) = -ctau; |
| 315 |
x.nptr(2) = 1; |
| 316 |
|
| 317 |
SuperMatrix<1,1> yx(work1_3n, n-1, 1); |
| 318 |
contract<_N>(1.0, YD, x, yx); |
| 319 |
YD.append_column<1>(yx); |
| 320 |
zgemv_("N", n-1, len, -ctau, D0+(k+1)*ld+1, ld, dnow, 1, 1.0, YD.ptr<0,1>(0,k), 1); |
| 321 |
yx.reset(); |
| 322 |
contract<_N>(1.0, YE, x, yx); |
| 323 |
YE.append_column<1>(yx); |
| 324 |
zgemv_("N", n-1, len, -ctau, D1+(k+1)*ld+1, ld, dnow, 1, 1.0, YE.ptr<0,1>(0,k), 1); |
| 325 |
yx.reset(); |
| 326 |
contract<_N>(1.0, ZD, x, yx); |
| 327 |
ZD.append_column<1>(yx); |
| 328 |
yx.reset(); |
| 329 |
contract<_N>(1.0, ZE, x, yx); |
| 330 |
ZE.append_column<1>(yx); |
| 331 |
} |
| 332 |
dnow[0] = alpha2; |
| 333 |
} |
| 334 |
} |
| 335 |
|
| 336 |
// finally update |
| 337 |
const int nrem = n-nb; |
| 338 |
if (nrem > 0) { |
| 339 |
// first update D and E. |
| 340 |
// Y^D + Z^E |
| 341 |
zaxpy_((n-1)*(nb*2+1), 1.0, ZE.block(0,0), 1, YD.block(0,0), 1); |
| 342 |
zaxpy_((n-1)*(nb*2+1), -1.0, ZD.block(0,0), 1, YE.block(0,0), 1); |
| 343 |
// Multiplied by W^+ |
| 344 |
zgemm3m_("N", "C", n-1, nrem, nb*2, 1.0, YD.block(0,0), n-1, W.block(0,0)+nb-1, n-1, 1.0, D0+nb*ld+1, ld); |
| 345 |
zgemm3m_("N", "C", n-1, nrem, nb*2, 1.0, YE.block(0,0), n-1, W.block(0,0)+nb-1, n-1, 1.0, D1+nb*ld+1, ld); |
| 346 |
assert(YD.nptr(0) == n-1 && YD.mptr(2) == 1); |
| 347 |
zaxpy_(n-1, 1.0, YD.block(0,2), 1, D0+nb*ld+1, 1); |
| 348 |
zaxpy_(n-1, 1.0, YE.block(0,2), 1, D1+nb*ld+1, 1); |
| 349 |
|
| 350 |
// now I can destroy Y and Z. |
| 351 |
complex<double>* const ptr = YD.block(0,0); |
| 352 |
|
| 353 |
// D^+W |
| 354 |
SuperMatrix<1,3> DW(ptr, nrem, nb, nrem, nb, /*init*/false); // 3 used |
| 355 |
zgemm3m_("C", "N", nrem, nb*2, n-1, 1.0, D0+nb*ld+1, ld, W.block(0,0), n-1, 0.0, DW.block(0,0), nrem); |
| 356 |
transpose_conj(nb, nrem, D0+nb*ld+1, ld, DW.block(0,2), nrem); |
| 357 |
// E^TW |
| 358 |
SuperMatrix<1,3> EW(ptr+nrem*nb*3, nrem, nb, nrem, nb, /*init*/false); // 6 used |
| 359 |
zgemm3m_("T", "N", nrem, nb*2, n-1, 1.0, D1+nb*ld+1, ld, W.block(0,0), n-1, 0.0, EW.block(0,0), nrem); |
| 360 |
transpose(nb, nrem, D1+nb*ld+1, ld, EW.block(0,2), nrem); |
| 361 |
|
| 362 |
// WT^+ |
| 363 |
SuperMatrix<1,3> WT(ptr+nrem*nb*6, nrem, nb, nrem, nb, /*init*/true); // 9 used |
| 364 |
contract<_N, _C>(1.0, W.shift(nb-1), T.slice<0,2>(), WT); |
| 365 |
for (int i = 0; i != nb; ++i) { |
| 366 |
WT.data<0,0>(0,i) += conj(T.data<0,2>(i,nb-1)); |
| 367 |
WT.data<0,1>(0,i) += conj(T.data<1,2>(i,nb-1)); |
| 368 |
WT.data<0,2>(0,i) += conj(T.data<2,2>(i,nb-1)); |
| 369 |
} |
| 370 |
|
| 371 |
// D <- WT^+W^+D |
| 372 |
zgemm3m_("N", "C", nrem, nrem, nb*3, 1.0, WT.block(0,0), nrem, DW.block(0,0), nrem, 1.0, D0+nb*ld+nb, ld); |
| 373 |
// E <- W^*T^T W^TE |
| 374 |
conj_n(WT.block(0,0), nrem*nb*3); |
| 375 |
zgemm3m_("N", "T", nrem, nrem, nb*3, 1.0, WT.block(0,0), nrem, EW.block(0,0), nrem, 1.0, D1+nb*ld+nb, ld); |
| 376 |
|
| 377 |
// next first form WS^+ |
| 378 |
SuperMatrix<1,1> WS(ptr+nrem*nb*6, nrem, nb, nrem, nb, /*init*/true); // 7 used |
| 379 |
contract<_N, _C>(1.0, W.shift(nb-1), S.slice<0,2>(), WS); |
| 380 |
for (int i = 0; i != nb; ++i) |
| 381 |
WS.data<0,0>(0,i) += conj(S.data<0,2>(i,nb-1)); |
| 382 |
|
| 383 |
// E^T WR |
| 384 |
SuperMatrix<1,1> EWR(ptr+nrem*nb*7, nrem, nb, nrem, nb, /*init*/true); // 8 used |
| 385 |
contract<_N, _N>(1.0, EW, R, EWR); |
| 386 |
// -WS^+ R^T W^T E = -WS^+ (E^T WR)^T |
| 387 |
zgemm3m_("N", "T", nrem, nrem, nb, -1.0, WS.block(0,0), nrem, EWR.block(0,0), nrem, 1.0, D0+nb*ld+nb, ld); |
| 388 |
|
| 389 |
// D^+WR |
| 390 |
SuperMatrix<1,1> DWR(ptr+nrem*nb*7, nrem, nb, nrem, nb, /*init*/true); // 8 used |
| 391 |
contract<_N, _N>(1.0, DW, R, DWR); |
| 392 |
// W^* S^T R^+ W^+D = W^* S^T(D^+ WR)^+ |
| 393 |
conj_n(WS.block(0,0), nrem*nb); |
| 394 |
zgemm3m_("N", "C", nrem, nrem, nb, 1.0, WS.block(0,0), nrem, DWR.block(0,0), nrem, 1.0, D1+nb*ld+nb, ld); |
| 395 |
} |
| 396 |
|
| 397 |
if (n != norig) { |
| 398 |
complex<double>* const ptr = YD.block(0,0); |
| 399 |
// Q0W |
| 400 |
SuperMatrix<1,3> Q0W(ptr, norig, nb, norig, nb, /*init*/false); // 3 used |
| 401 |
zgemm3m_("N", "N", norig, nb*2, n-1, 1.0, Q0+ld, ld, W.block(0,0), n-1, 0.0, Q0W.block(0,0), norig); |
| 402 |
fill_n(Q0W.block(0,2), norig*nb, 0.0); |
| 403 |
for (int i = 0; i != nb; ++i) |
| 404 |
zaxpy_(norig, 1.0, Q0+ld*(i+1), 1, Q0W.block(0,2)+norig*i, 1); |
| 405 |
// Q1W |
| 406 |
SuperMatrix<1,3> Q1W(ptr+norig*nb*3, norig, nb, norig, nb, /*init*/false); // 6 used |
| 407 |
zgemm3m_("N", "N", norig, nb*2, n-1, 1.0, Q1+ld, ld, W.block(0,0), n-1, 0.0, Q1W.block(0,0), norig); |
| 408 |
fill_n(Q1W.block(0,2), norig*nb, 0.0); |
| 409 |
for (int i = 0; i != nb; ++i) |
| 410 |
zaxpy_(norig, 1.0, Q1+ld*(i+1), 1, Q1W.block(0,2)+norig*i, 1); |
| 411 |
|
| 412 |
// WT^+ |
| 413 |
SuperMatrix<1,3> WT(ptr+norig*nb*6, n-1, nb, n-1, nb, /*init*/true); // 9 used |
| 414 |
transpose_conj(nb, nb, T.block(0,2), nb, WT.block(0,0), n-1); |
| 415 |
transpose_conj(nb, nb, T.block(1,2), nb, WT.block(0,1), n-1); |
| 416 |
transpose_conj(nb, nb, T.block(2,2), nb, WT.block(0,2), n-1); |
| 417 |
contract<_N, _C>(1.0, W, T.slice<0,2>(), WT); |
| 418 |
|
| 419 |
// Q0 <- Q0WTW^+ and Q1 <- Q1WTW^+ |
| 420 |
zgemm3m_("N", "C", norig, n-1, 3*nb, 1.0, Q0W.block(0,0), norig, WT.block(0,0), n-1, 1.0, Q0+ld, ld); |
| 421 |
zgemm3m_("N", "C", norig, n-1, 3*nb, 1.0, Q1W.block(0,0), norig, WT.block(0,0), n-1, 1.0, Q1+ld, ld); |
| 422 |
|
| 423 |
// next first form WS^+ |
| 424 |
SuperMatrix<1,1> WS(ptr+norig*nb*6, n-1, nb, n-1, nb, /*init*/true); // 7 used |
| 425 |
transpose_conj(nb, nb, S.block(0,2), nb, WS.block(0,0), n-1); |
| 426 |
contract<_N, _C>(1.0, W, S.slice<0,2>(), WS); |
| 427 |
|
| 428 |
// R^T W^T Q0^T |
| 429 |
SuperMatrix<1,1> RWQ0(ptr+norig*nb*6+(n-1)*nb, nb, norig, nb, norig, /*init*/true); // 8 used |
| 430 |
contract<_T, _T>(1.0, R, Q0W, RWQ0); |
| 431 |
// Q1 <- Q0^* W^* R^* SW^+ |
| 432 |
zgemm3m_("C", "C", norig, n-1, nb, -1.0, RWQ0.block(0,0), nb, WS.block(0,0), n-1, 1.0, Q1+ld, ld); |
| 433 |
|
| 434 |
// R^T W^T Q1^T |
| 435 |
SuperMatrix<1,1> RWQ1(ptr+norig*nb*6+(n-1)*nb, nb, norig, nb, norig, /*init*/true); // 8 used |
| 436 |
contract<_T, _T>(1.0, R, Q1W, RWQ1); |
| 437 |
// Q0 <- Q1^* W^* R^* SW^+ |
| 438 |
zgemm3m_("C", "C", norig, n-1, nb, 1.0, RWQ1.block(0,0), nb, WS.block(0,0), n-1, 1.0, Q0+ld, ld); |
| 439 |
} else { |
| 440 |
// in the first run this can be simplified, because Q0 = unit, Q1 = 0 on input. |
| 441 |
assert(abs(zdotc_(n*n, Q0, 1, Q0, 1)-static_cast<double>(n)) < 1.0e-10 |
| 442 |
&& abs(zdotc_(n*n, Q1, 1, Q1, 1)) < 1.0e-10); |
| 443 |
|
| 444 |
complex<double>* const ptr = YD.block(0,0); |
| 445 |
SuperMatrix<1,3> WT(ptr, n-1, nb, n-1, nb, /*init*/true); |
| 446 |
transpose_conj(nb, nb, T.block(0,2), nb, WT.block(0,0), n-1); |
| 447 |
transpose_conj(nb, nb, T.block(1,2), nb, WT.block(0,1), n-1); |
| 448 |
transpose_conj(nb, nb, T.block(2,2), nb, WT.block(0,2), n-1); |
| 449 |
contract<_N, _C>(1.0, W, T.slice<0,2>(), WT); |
| 450 |
|
| 451 |
transpose_conj(n-1, nb, WT.block(0,2), n-1, Q0+ld+1, ld); |
| 452 |
for (int i = 0; i != nb; ++i) |
| 453 |
Q0[(i+1)*ld+i+1] += 1.0; |
| 454 |
zgemm3m_("N", "C", n-1, n-1, 2*nb, 1.0, W.block(0,0), n-1, WT.block(0,0), n-1, 1.0, Q0+ld+1, ld); |
| 455 |
|
| 456 |
SuperMatrix<1,1> WS(ptr, n-1, nb, n-1, nb, /*init*/true); |
| 457 |
transpose_conj(nb, nb, S.block(0,2), nb, WS.block(0,0), n-1); |
| 458 |
contract<_N, _C>(1.0, W, S.slice<0,2>(), WS); |
| 459 |
|
| 460 |
SuperMatrix<1,1> RW(ptr+(n-1)*nb, nb, n-1, nb, n-1, /*init*/true); |
| 461 |
transpose(nb, nb, R.block(2,0), nb, RW.block(0,0), nb); |
| 462 |
contract<_T, _T>(1.0, R.trunc<2>(), W, RW); |
| 463 |
zgemm3m_("C", "C", n-1, n-1, nb, -1.0, RW.block(0,0), nb, WS.block(0,0), n-1, 1.0, Q1+ld+1, ld); |
| 464 |
} |
| 465 |
|
| 466 |
} |
| 467 |
|
| 468 |
} } |