libflame revision_anchor
Functions
FLA_Bidiag_UT_u_opt_var3.c File Reference

(r)

Functions

FLA_Error FLA_Bidiag_UT_u_opt_var3 (FLA_Obj A, FLA_Obj TU, FLA_Obj TV)
 
FLA_Error FLA_Bidiag_UT_u_step_opt_var3 (FLA_Obj A, FLA_Obj T, FLA_Obj S)
 
FLA_Error FLA_Bidiag_UT_u_step_ops_var3 (int m_A, int n_A, int m_TS, float *buff_A, int rs_A, int cs_A, float *buff_T, int rs_T, int cs_T, float *buff_S, int rs_S, int cs_S)
 
FLA_Error FLA_Bidiag_UT_u_step_opd_var3 (int m_A, int n_A, int m_TS, double *buff_A, int rs_A, int cs_A, double *buff_T, int rs_T, int cs_T, double *buff_S, int rs_S, int cs_S)
 
FLA_Error FLA_Bidiag_UT_u_step_opc_var3 (int m_A, int n_A, int m_TS, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_T, int rs_T, int cs_T, scomplex *buff_S, int rs_S, int cs_S)
 
FLA_Error FLA_Bidiag_UT_u_step_opz_var3 (int m_A, int n_A, int m_TS, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_T, int rs_T, int cs_T, dcomplex *buff_S, int rs_S, int cs_S)
 

Function Documentation

◆ FLA_Bidiag_UT_u_opt_var3()

FLA_Error FLA_Bidiag_UT_u_opt_var3 ( FLA_Obj A,
FLA_Obj TU,
FLA_Obj TV )
14{
15 return FLA_Bidiag_UT_u_step_opt_var3( A, TU, TV );
16}
FLA_Bidiag_UT_u_step_opt_var3
FLA_Error FLA_Bidiag_UT_u_step_opt_var3(FLA_Obj A, FLA_Obj T, FLA_Obj S)
Definition FLA_Bidiag_UT_u_opt_var3.c:18

References FLA_Bidiag_UT_u_step_opt_var3().

Referenced by FLA_Bidiag_UT_u().

◆ FLA_Bidiag_UT_u_step_opc_var3()

FLA_Error FLA_Bidiag_UT_u_step_opc_var3 ( int m_A,
int n_A,
int m_TS,
scomplex * buff_A,
int rs_A,
int cs_A,
scomplex * buff_T,
int rs_T,
int cs_T,
scomplex * buff_S,
int rs_S,
int cs_S )
956{
957 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
958 scomplex* buff_0 = FLA_COMPLEX_PTR( FLA_ZERO );
959 scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
960
961 scomplex alpha12;
962 scomplex minus_conj_alpha12;
963 scomplex psi11_minus_alpha12;
964 scomplex minus_inv_tau11;
965 scomplex minus_upsilon11;
966 scomplex minus_conj_nu11;
967 scomplex minus_conj_psi11;
968 scomplex minus_zeta11;
969 scomplex beta;
970 int i;
971
972 // b_alg = FLA_Obj_length( T );
973 int b_alg = m_TS;
974
975 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &w );
976 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &ap );
977 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &u );
978 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &up );
979 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &v );
980 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &y );
981 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &z );
982 scomplex* buff_w = ( scomplex* ) FLA_malloc( m_A * sizeof( *buff_A ) );
983 scomplex* buff_ap = ( scomplex* ) FLA_malloc( n_A * sizeof( *buff_A ) );
984 scomplex* buff_u = ( scomplex* ) FLA_malloc( m_A * sizeof( *buff_A ) );
985 scomplex* buff_up = ( scomplex* ) FLA_malloc( m_A * sizeof( *buff_A ) );
986 scomplex* buff_v = ( scomplex* ) FLA_malloc( n_A * sizeof( *buff_A ) );
987 scomplex* buff_y = ( scomplex* ) FLA_malloc( n_A * sizeof( *buff_A ) );
988 scomplex* buff_z = ( scomplex* ) FLA_malloc( m_A * sizeof( *buff_A ) );
989 int inc_w = 1;
990 int inc_ap = 1;
991 int inc_u = 1;
992 int inc_up = 1;
993 int inc_v = 1;
994 int inc_y = 1;
995 int inc_z = 1;
996
997 for ( i = 0; i < b_alg; ++i )
998 {
999 scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
1000 scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
1001 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
1002 scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
1003 scomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
1004 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
1005 scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
1006
1007 scomplex* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
1008 scomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
1009
1010 scomplex* s01 = buff_S + (i )*cs_S + (0 )*rs_S;
1011 scomplex* sigma11 = buff_S + (i )*cs_S + (i )*rs_S;
1012
1013 scomplex* w21 = buff_w + (i+1)*inc_w;
1014
1015 scomplex* a12p = buff_ap + (i+1)*inc_ap;
1016
1017 scomplex* upsilon11 = buff_u + (i )*inc_u;
1018 scomplex* u21 = buff_u + (i+1)*inc_u;
1019
1020 scomplex* u21p = buff_up + (i+1)*inc_up;
1021
1022 scomplex* nu11 = buff_v + (i )*inc_v;
1023 scomplex* v21 = buff_v + (i+1)*inc_v;
1024
1025 scomplex* psi11 = buff_y + (i )*inc_y;
1026 scomplex* y21 = buff_y + (i+1)*inc_y;
1027
1028 scomplex* zeta11 = buff_z + (i )*inc_z;
1029 scomplex* z21 = buff_z + (i+1)*inc_z;
1030
1031 scomplex* a12p_t = a12p + (0 )*inc_ap;
1032 scomplex* a12p_b = a12p + (1 )*inc_ap;
1033
1034 scomplex* v21_t = v21 + (0 )*inc_v;
1035 scomplex* v21_b = v21 + (1 )*inc_v;
1036
1037 scomplex* a12t_l = a12t + (0 )*cs_A + (0 )*rs_A;
1038 scomplex* a12t_r = a12t + (1 )*cs_A + (0 )*rs_A;
1039
1040 scomplex* A22_l = A22 + (0 )*cs_A + (0 )*rs_A;
1041
1042 int m_ahead = m_A - i - 1;
1043 int n_ahead = n_A - i - 1;
1044 int m_behind = i;
1045 int n_behind = i;
1046
1047 /*------------------------------------------------------------*/
1048
1049 if ( m_behind > 0 )
1050 {
1051 // FLA_Copy( upsilon11, minus_upsilon11 );
1052 // FLA_Scal( FLA_MINUS_ONE, minus_upsilon11 );
1053 bl1_cmult3( buff_m1, upsilon11, &minus_upsilon11 );
1054
1055 // FLA_Copy( zeta11, minus_zeta11 );
1056 // FLA_Scal( FLA_MINUS_ONE, minus_zeta11 );
1057 bl1_cmult3( buff_m1, zeta11, &minus_zeta11 );
1058
1059 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, psi11, minus_conj_psi11 );
1060 // FLA_Scal( FLA_MINUS_ONE, minus_conj_psi11 );
1061 bl1_ccopyconj( psi11, &minus_conj_psi11 );
1062 bl1_cscals( buff_m1, &minus_conj_psi11 );
1063
1064 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, nu11, minus_conj_nu11 );
1065 // FLA_Scal( FLA_MINUS_ONE, minus_conj_nu11 );
1066 bl1_ccopyconj( nu11, &minus_conj_nu11 );
1067 bl1_cscals( buff_m1, &minus_conj_nu11 );
1068
1069 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_psi11, upsilon11, alpha11 );
1070 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_nu11, zeta11, alpha11 );
1071 bl1_caxpyv( BLIS1_NO_CONJUGATE,
1072 1,
1073 &minus_conj_psi11,
1074 upsilon11, 1,
1075 alpha11, 1 );
1076 bl1_caxpyv( BLIS1_NO_CONJUGATE,
1077 1,
1078 &minus_conj_nu11,
1079 zeta11, 1,
1080 alpha11, 1 );
1081
1082 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_psi11, u21, a21 );
1083 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_nu11, z21, a21 );
1084 bl1_caxpyv( BLIS1_NO_CONJUGATE,
1085 m_ahead,
1086 &minus_conj_psi11,
1087 u21, inc_u,
1088 a21, rs_A );
1089 bl1_caxpyv( BLIS1_NO_CONJUGATE,
1090 m_ahead,
1091 &minus_conj_nu11,
1092 z21, inc_z,
1093 a21, rs_A );
1094
1095 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, minus_upsilon11, y21, a12t );
1096 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, minus_zeta11, v21, a12t );
1097 bl1_caxpyv( BLIS1_CONJUGATE,
1098 n_ahead,
1099 &minus_upsilon11,
1100 y21, inc_y,
1101 a12t, cs_A );
1102 bl1_caxpyv( BLIS1_CONJUGATE,
1103 n_ahead,
1104 &minus_zeta11,
1105 v21, inc_v,
1106 a12t, cs_A );
1107 }
1108
1109 // FLA_Househ2_UT( FLA_LEFT,
1110 // alpha11,
1111 // a21, tau11 );
1112 // FLA_Copy( a21, u21p );
1113 FLA_Househ2_UT_l_opc( m_ahead,
1114 alpha11,
1115 a21, rs_A,
1116 tau11 );
1117 bl1_ccopyv( BLIS1_NO_CONJUGATE,
1118 m_ahead,
1119 a21, rs_A,
1120 u21p, inc_up );
1121
1122 if ( n_ahead > 0 )
1123 {
1124 // FLA_Copy( FLA_MINUS_ONE, minus_inv_tau11 );
1125 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, minus_inv_tau11 );
1126 bl1_cdiv3( buff_m1, tau11, &minus_inv_tau11 );
1127
1128 // FLA_Copyt( FLA_TRANSPOSE, a12t, a12p );
1129 // FLA_Axpyt( FLA_TRANSPOSE, minus_inv_tau11, a12t, a12p );
1130 bl1_ccopyv( BLIS1_NO_CONJUGATE,
1131 n_ahead,
1132 a12t, cs_A,
1133 a12p, inc_ap );
1134 bl1_caxpyv( BLIS1_NO_CONJUGATE,
1135 n_ahead,
1136 &minus_inv_tau11,
1137 a12t, cs_A,
1138 a12p, inc_ap );
1139 }
1140
1141 if ( m_behind > 0 )
1142 {
1143 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, u21, y21, A22 );
1144 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, z21, v21, A22 );
1145 bl1_cger( BLIS1_NO_CONJUGATE,
1146 BLIS1_CONJUGATE,
1147 m_ahead,
1148 n_ahead,
1149 buff_m1,
1150 u21, inc_u,
1151 y21, inc_y,
1152 A22, rs_A, cs_A );
1153 bl1_cger( BLIS1_NO_CONJUGATE,
1154 BLIS1_CONJUGATE,
1155 m_ahead,
1156 n_ahead,
1157 buff_m1,
1158 z21, inc_z,
1159 v21, inc_v,
1160 A22, rs_A, cs_A );
1161 }
1162
1163 if ( n_ahead > 0 )
1164 {
1165 // FLA_Gemvc( FLA_CONJ_TRANSPOSE, FLA_NO_CONJUGATE, FLA_ONE, A22, u21p, FLA_ZERO, y21 );
1166 bl1_cgemv( BLIS1_CONJ_TRANSPOSE,
1167 BLIS1_NO_CONJUGATE,
1168 m_ahead,
1169 n_ahead,
1170 buff_1,
1171 A22, rs_A, cs_A,
1172 u21p, inc_up,
1173 buff_0,
1174 y21, inc_y );
1175
1176 // FLA_Axpyt( FLA_CONJ_NO_TRANSPOSE, minus_inv_tau11, y21, a12p );
1177 bl1_caxpyv( BLIS1_CONJUGATE,
1178 n_ahead,
1179 &minus_inv_tau11,
1180 y21, inc_y,
1181 a12p, inc_ap );
1182
1183 // FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A22, a12p, FLA_ZERO, w21 );
1184 bl1_cgemv( BLIS1_NO_TRANSPOSE,
1185 BLIS1_CONJUGATE,
1186 m_ahead,
1187 n_ahead,
1188 buff_1,
1189 A22, rs_A, cs_A,
1190 a12p, inc_ap,
1191 buff_0,
1192 w21, inc_w );
1193
1194 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, FLA_ONE, a12t, y21 );
1195 bl1_caxpyv( BLIS1_CONJUGATE,
1196 n_ahead,
1197 buff_1,
1198 a12t, cs_A,
1199 y21, inc_y );
1200
1201 // FLA_Househ2s_UT( FLA_RIGHT,
1202 // a12p_t,
1203 // a12p_b,
1204 // alpha12, psi11_minus_alpha12, sigma11 );
1205 FLA_Househ2s_UT_r_opc( n_ahead - 1,
1206 a12p_t,
1207 a12p_b, inc_ap,
1208 &alpha12,
1209 &psi11_minus_alpha12,
1210 sigma11 );
1211
1212 // FLA_Copy( a12p, v21 );
1213 // FLA_Mult_add( FLA_MINUS_ONE, alpha12, v21_t );
1214 // FLA_Inv_scalc( FLA_NO_CONJUGATE, psi11_minus_alpha12, v21 );
1215 // FLA_Conjugate( v21_b );
1216 bl1_ccopyv( BLIS1_NO_CONJUGATE,
1217 n_ahead,
1218 a12p, inc_ap,
1219 v21, inc_v );
1220 bl1_cmult4( buff_m1, &alpha12, v21_t, v21_t );
1221 bl1_cinvscalv( BLIS1_NO_CONJUGATE,
1222 n_ahead,
1223 &psi11_minus_alpha12,
1224 v21, inc_v );
1225 bl1_cconjv( n_ahead - 1,
1226 v21_b, inc_v );
1227
1228 // FLA_Copyt( FLA_NO_TRANSPOSE, alpha12, a12t_l );
1229 // FLA_Copyt( FLA_TRANSPOSE, v21_b, a12t_r );
1230 *a12t_l = alpha12;
1231 bl1_ccopyv( BLIS1_NO_CONJUGATE,
1232 n_ahead - 1,
1233 v21_b, inc_v,
1234 a12t_r, cs_A );
1235 }
1236
1237 // FLA_Copy( u21p, u21 );
1238 bl1_ccopyv( BLIS1_NO_CONJUGATE,
1239 m_ahead,
1240 u21p, inc_up,
1241 u21, inc_u );
1242
1243 if ( n_ahead > 0 )
1244 {
1245 // FLA_Dotc( FLA_CONJUGATE, y21, v21, beta );
1246 // FLA_Scal( FLA_MINUS_ONE, beta );
1247 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, beta );
1248 bl1_cdot( BLIS1_CONJUGATE,
1249 n_ahead,
1250 y21, inc_y,
1251 v21, inc_v,
1252 &beta );
1253 bl1_cscals( &minus_inv_tau11, &beta );
1254
1255 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha12, minus_conj_alpha12 );
1256 // FLA_Scal( FLA_MINUS_ONE, minus_conj_alpha12 );
1257 bl1_ccopyconj( &alpha12, &minus_conj_alpha12 );
1258 bl1_cneg1( &minus_conj_alpha12 );
1259
1260 // FLA_Copy( w21, z21 );
1261 // FLA_Axpy( minus_conj_alpha12, A22_l, z21 );
1262 // FLA_Inv_scalc( FLA_CONJUGATE, psi11_minus_alpha12, z21 );
1263 // FLA_Axpy( beta, u21, z21 );
1264 bl1_ccopyv( BLIS1_NO_CONJUGATE,
1265 m_ahead,
1266 w21, inc_w,
1267 z21, inc_z );
1268 bl1_caxpyv( BLIS1_NO_CONJUGATE,
1269 m_ahead,
1270 &minus_conj_alpha12,
1271 A22_l, rs_A,
1272 z21, inc_z );
1273 bl1_cinvscalv( BLIS1_CONJUGATE,
1274 m_ahead,
1275 &psi11_minus_alpha12,
1276 z21, inc_z );
1277 bl1_caxpyv( BLIS1_NO_CONJUGATE,
1278 m_ahead,
1279 &beta,
1280 u21, inc_u,
1281 z21, inc_z );
1282
1283 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, y21 );
1284 // FLA_Inv_scalc( FLA_NO_CONJUGATE, sigma11, z21 );
1285 bl1_cinvscalv( BLIS1_CONJUGATE,
1286 n_ahead,
1287 tau11,
1288 y21, inc_y );
1289 bl1_cinvscalv( BLIS1_NO_CONJUGATE,
1290 m_ahead,
1291 sigma11,
1292 z21, inc_z );
1293
1294 // FLA_Gemv( FLA_CONJ_NO_TRANSPOSE, FLA_ONE, A02, v21, FLA_ZERO, s01 );
1295 bl1_cgemv( BLIS1_CONJ_NO_TRANSPOSE,
1296 BLIS1_NO_CONJUGATE,
1297 m_behind,
1298 n_ahead,
1299 buff_1,
1300 A02, rs_A, cs_A,
1301 v21, inc_v,
1302 buff_0,
1303 s01, rs_S );
1304 }
1305
1306 // FLA_Copyt( FLA_CONJ_TRANSPOSE, a10t, t01 );
1307 // FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, u21, FLA_ONE, t01 );
1308 bl1_ccopyv( BLIS1_CONJUGATE,
1309 n_behind,
1310 a10t, cs_A,
1311 t01, rs_T );
1312 bl1_cgemv( BLIS1_CONJ_TRANSPOSE,
1313 BLIS1_NO_CONJUGATE,
1314 m_ahead,
1315 n_behind,
1316 buff_1,
1317 A20, rs_A, cs_A,
1318 u21, inc_u,
1319 buff_1,
1320 t01, rs_T );
1321
1322 if ( m_behind + 1 == b_alg && n_ahead > 0 )
1323 {
1324 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, u21, y21, A22 );
1325 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, z21, v21, A22 );
1326 bl1_cger( BLIS1_NO_CONJUGATE,
1327 BLIS1_CONJUGATE,
1328 m_ahead,
1329 n_ahead,
1330 buff_m1,
1331 u21, inc_u,
1332 y21, inc_y,
1333 A22, rs_A, cs_A );
1334 bl1_cger( BLIS1_NO_CONJUGATE,
1335 BLIS1_CONJUGATE,
1336 m_ahead,
1337 n_ahead,
1338 buff_m1,
1339 z21, inc_z,
1340 v21, inc_v,
1341 A22, rs_A, cs_A );
1342 }
1343
1344 /*------------------------------------------------------------*/
1345
1346 }
1347
1348 // FLA_Obj_free( &w );
1349 // FLA_Obj_free( &ap );
1350 // FLA_Obj_free( &u );
1351 // FLA_Obj_free( &up );
1352 // FLA_Obj_free( &v );
1353 // FLA_Obj_free( &y );
1354 // FLA_Obj_free( &z );
1355 FLA_free( buff_w );
1356 FLA_free( buff_ap );
1357 FLA_free( buff_u );
1358 FLA_free( buff_up );
1359 FLA_free( buff_v );
1360 FLA_free( buff_y );
1361 FLA_free( buff_z );
1362
1363 return FLA_SUCCESS;
1364}
FLA_MINUS_ONE
FLA_Obj FLA_MINUS_ONE
Definition FLA_Init.c:22
FLA_ZERO
FLA_Obj FLA_ZERO
Definition FLA_Init.c:20
FLA_ONE
FLA_Obj FLA_ONE
Definition FLA_Init.c:18
FLA_free
void FLA_free(void *ptr)
Definition FLA_Memory.c:247
FLA_malloc
void * FLA_malloc(size_t size)
Definition FLA_Memory.c:111
FLA_Househ2s_UT_r_opc
FLA_Error FLA_Househ2s_UT_r_opc(int m_x2, scomplex *chi_1, scomplex *x2, int inc_x2, scomplex *alpha, scomplex *chi_1_minus_alpha, scomplex *tau)
Definition FLA_Househ2s_UT.c:589
FLA_Househ2_UT_l_opc
FLA_Error FLA_Househ2_UT_l_opc(int m_x2, scomplex *chi_1, scomplex *x2, int inc_x2, scomplex *tau)
Definition FLA_Househ2_UT.c:390
i
int i
Definition bl1_axmyv2.c:145
bl1_caxpyv
void bl1_caxpyv(conj1_t conj, int n, scomplex *alpha, scomplex *x, int incx, scomplex *y, int incy)
Definition bl1_axpyv.c:29
bl1_cconjv
void bl1_cconjv(int m, scomplex *x, int incx)
Definition bl1_conjv.c:23
bl1_ccopyv
void bl1_ccopyv(conj1_t conj, int m, scomplex *x, int incx, scomplex *y, int incy)
Definition bl1_copyv.c:49
bl1_cdot
void bl1_cdot(conj1_t conj, int n, scomplex *x, int incx, scomplex *y, int incy, scomplex *rho)
Definition bl1_dot.c:39
bl1_cgemv
void bl1_cgemv(trans1_t transa, conj1_t conjx, int m, int n, scomplex *alpha, scomplex *a, int a_rs, int a_cs, scomplex *x, int incx, scomplex *beta, scomplex *y, int incy)
Definition bl1_gemv.c:125
bl1_cger
void bl1_cger(conj1_t conjx, conj1_t conjy, int m, int n, scomplex *alpha, scomplex *x, int incx, scomplex *y, int incy, scomplex *a, int a_rs, int a_cs)
Definition bl1_ger.c:111
bl1_cinvscalv
void bl1_cinvscalv(conj1_t conj, int n, scomplex *alpha, scomplex *x, int incx)
Definition bl1_invscalv.c:52
BLIS1_NO_TRANSPOSE
@ BLIS1_NO_TRANSPOSE
Definition blis_type_defs.h:54
BLIS1_CONJ_TRANSPOSE
@ BLIS1_CONJ_TRANSPOSE
Definition blis_type_defs.h:57
BLIS1_CONJ_NO_TRANSPOSE
@ BLIS1_CONJ_NO_TRANSPOSE
Definition blis_type_defs.h:56
BLIS1_CONJUGATE
@ BLIS1_CONJUGATE
Definition blis_type_defs.h:82
BLIS1_NO_CONJUGATE
@ BLIS1_NO_CONJUGATE
Definition blis_type_defs.h:81
scomplex
Definition blis_type_defs.h:133

References bl1_caxpyv(), bl1_cconjv(), bl1_ccopyv(), bl1_cdot(), bl1_cgemv(), bl1_cger(), bl1_cinvscalv(), BLIS1_CONJ_NO_TRANSPOSE, BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, FLA_free(), FLA_Househ2_UT_l_opc(), FLA_Househ2s_UT_r_opc(), FLA_malloc(), FLA_MINUS_ONE, FLA_ONE, FLA_ZERO, and i.

Referenced by FLA_Bidiag_UT_u_step_opt_var3().

◆ FLA_Bidiag_UT_u_step_opd_var3()

FLA_Error FLA_Bidiag_UT_u_step_opd_var3 ( int m_A,
int n_A,
int m_TS,
double * buff_A,
int rs_A,
int cs_A,
double * buff_T,
int rs_T,
int cs_T,
double * buff_S,
int rs_S,
int cs_S )
538{
539 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
540 double* buff_0 = FLA_DOUBLE_PTR( FLA_ZERO );
541 double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
542
543 double alpha12;
544 double minus_conj_alpha12;
545 double psi11_minus_alpha12;
546 double minus_inv_tau11;
547 double minus_upsilon11;
548 double minus_conj_nu11;
549 double minus_conj_psi11;
550 double minus_zeta11;
551 double beta;
552 int i;
553
554 // b_alg = FLA_Obj_length( T );
555 int b_alg = m_TS;
556
557 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &w );
558 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &ap );
559 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &u );
560 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &up );
561 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &v );
562 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &y );
563 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &z );
564 double* buff_w = ( double* ) FLA_malloc( m_A * sizeof( *buff_A ) );
565 double* buff_ap = ( double* ) FLA_malloc( n_A * sizeof( *buff_A ) );
566 double* buff_u = ( double* ) FLA_malloc( m_A * sizeof( *buff_A ) );
567 double* buff_up = ( double* ) FLA_malloc( m_A * sizeof( *buff_A ) );
568 double* buff_v = ( double* ) FLA_malloc( n_A * sizeof( *buff_A ) );
569 double* buff_y = ( double* ) FLA_malloc( n_A * sizeof( *buff_A ) );
570 double* buff_z = ( double* ) FLA_malloc( m_A * sizeof( *buff_A ) );
571 int inc_w = 1;
572 int inc_ap = 1;
573 int inc_u = 1;
574 int inc_up = 1;
575 int inc_v = 1;
576 int inc_y = 1;
577 int inc_z = 1;
578
579 for ( i = 0; i < b_alg; ++i )
580 {
581 double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
582 double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
583 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
584 double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
585 double* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
586 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
587 double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
588
589 double* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
590 double* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
591
592 double* s01 = buff_S + (i )*cs_S + (0 )*rs_S;
593 double* sigma11 = buff_S + (i )*cs_S + (i )*rs_S;
594
595 double* w21 = buff_w + (i+1)*inc_w;
596
597 double* a12p = buff_ap + (i+1)*inc_ap;
598
599 double* upsilon11 = buff_u + (i )*inc_u;
600 double* u21 = buff_u + (i+1)*inc_u;
601
602 double* u21p = buff_up + (i+1)*inc_up;
603
604 double* nu11 = buff_v + (i )*inc_v;
605 double* v21 = buff_v + (i+1)*inc_v;
606
607 double* psi11 = buff_y + (i )*inc_y;
608 double* y21 = buff_y + (i+1)*inc_y;
609
610 double* zeta11 = buff_z + (i )*inc_z;
611 double* z21 = buff_z + (i+1)*inc_z;
612
613 double* a12p_t = a12p + (0 )*inc_ap;
614 double* a12p_b = a12p + (1 )*inc_ap;
615
616 double* v21_t = v21 + (0 )*inc_v;
617 double* v21_b = v21 + (1 )*inc_v;
618
619 double* a12t_l = a12t + (0 )*cs_A + (0 )*rs_A;
620 double* a12t_r = a12t + (1 )*cs_A + (0 )*rs_A;
621
622 double* A22_l = A22 + (0 )*cs_A + (0 )*rs_A;
623
624 int m_ahead = m_A - i - 1;
625 int n_ahead = n_A - i - 1;
626 int m_behind = i;
627 int n_behind = i;
628
629 /*------------------------------------------------------------*/
630
631 if ( m_behind > 0 )
632 {
633 // FLA_Copy( upsilon11, minus_upsilon11 );
634 // FLA_Scal( FLA_MINUS_ONE, minus_upsilon11 );
635 bl1_dmult3( buff_m1, upsilon11, &minus_upsilon11 );
636
637 // FLA_Copy( zeta11, minus_zeta11 );
638 // FLA_Scal( FLA_MINUS_ONE, minus_zeta11 );
639 bl1_dmult3( buff_m1, zeta11, &minus_zeta11 );
640
641 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, psi11, minus_conj_psi11 );
642 // FLA_Scal( FLA_MINUS_ONE, minus_conj_psi11 );
643 bl1_dcopyconj( psi11, &minus_conj_psi11 );
644 bl1_dscals( buff_m1, &minus_conj_psi11 );
645
646 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, nu11, minus_conj_nu11 );
647 // FLA_Scal( FLA_MINUS_ONE, minus_conj_nu11 );
648 bl1_dcopyconj( nu11, &minus_conj_nu11 );
649 bl1_dscals( buff_m1, &minus_conj_nu11 );
650
651 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_psi11, upsilon11, alpha11 );
652 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_nu11, zeta11, alpha11 );
653 bl1_daxpyv( BLIS1_NO_CONJUGATE,
654 1,
655 &minus_conj_psi11,
656 upsilon11, 1,
657 alpha11, 1 );
658 bl1_daxpyv( BLIS1_NO_CONJUGATE,
659 1,
660 &minus_conj_nu11,
661 zeta11, 1,
662 alpha11, 1 );
663
664 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_psi11, u21, a21 );
665 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_nu11, z21, a21 );
666 bl1_daxpyv( BLIS1_NO_CONJUGATE,
667 m_ahead,
668 &minus_conj_psi11,
669 u21, inc_u,
670 a21, rs_A );
671 bl1_daxpyv( BLIS1_NO_CONJUGATE,
672 m_ahead,
673 &minus_conj_nu11,
674 z21, inc_z,
675 a21, rs_A );
676
677 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, minus_upsilon11, y21, a12t );
678 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, minus_zeta11, v21, a12t );
679 bl1_daxpyv( BLIS1_CONJUGATE,
680 n_ahead,
681 &minus_upsilon11,
682 y21, inc_y,
683 a12t, cs_A );
684 bl1_daxpyv( BLIS1_CONJUGATE,
685 n_ahead,
686 &minus_zeta11,
687 v21, inc_v,
688 a12t, cs_A );
689 }
690
691 // FLA_Househ2_UT( FLA_LEFT,
692 // alpha11,
693 // a21, tau11 );
694 // FLA_Copy( a21, u21p );
695 FLA_Househ2_UT_l_opd( m_ahead,
696 alpha11,
697 a21, rs_A,
698 tau11 );
699 bl1_dcopyv( BLIS1_NO_CONJUGATE,
700 m_ahead,
701 a21, rs_A,
702 u21p, inc_up );
703
704 if ( n_ahead > 0 )
705 {
706 // FLA_Copy( FLA_MINUS_ONE, minus_inv_tau11 );
707 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, minus_inv_tau11 );
708 bl1_ddiv3( buff_m1, tau11, &minus_inv_tau11 );
709
710 // FLA_Copyt( FLA_TRANSPOSE, a12t, a12p );
711 // FLA_Axpyt( FLA_TRANSPOSE, minus_inv_tau11, a12t, a12p );
712 bl1_dcopyv( BLIS1_NO_CONJUGATE,
713 n_ahead,
714 a12t, cs_A,
715 a12p, inc_ap );
716 bl1_daxpyv( BLIS1_NO_CONJUGATE,
717 n_ahead,
718 &minus_inv_tau11,
719 a12t, cs_A,
720 a12p, inc_ap );
721 }
722
723 if ( m_behind > 0 )
724 {
725 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, u21, y21, A22 );
726 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, z21, v21, A22 );
727 bl1_dger( BLIS1_NO_CONJUGATE,
728 BLIS1_CONJUGATE,
729 m_ahead,
730 n_ahead,
731 buff_m1,
732 u21, inc_u,
733 y21, inc_y,
734 A22, rs_A, cs_A );
735 bl1_dger( BLIS1_NO_CONJUGATE,
736 BLIS1_CONJUGATE,
737 m_ahead,
738 n_ahead,
739 buff_m1,
740 z21, inc_z,
741 v21, inc_v,
742 A22, rs_A, cs_A );
743 }
744
745 if ( n_ahead > 0 )
746 {
747 // FLA_Gemvc( FLA_CONJ_TRANSPOSE, FLA_NO_CONJUGATE, FLA_ONE, A22, u21p, FLA_ZERO, y21 );
748 bl1_dgemv( BLIS1_CONJ_TRANSPOSE,
749 BLIS1_NO_CONJUGATE,
750 m_ahead,
751 n_ahead,
752 buff_1,
753 A22, rs_A, cs_A,
754 u21p, inc_up,
755 buff_0,
756 y21, inc_y );
757
758 // FLA_Axpyt( FLA_CONJ_NO_TRANSPOSE, minus_inv_tau11, y21, a12p );
759 bl1_daxpyv( BLIS1_CONJUGATE,
760 n_ahead,
761 &minus_inv_tau11,
762 y21, inc_y,
763 a12p, inc_ap );
764
765 // FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A22, a12p, FLA_ZERO, w21 );
766 bl1_dgemv( BLIS1_NO_TRANSPOSE,
767 BLIS1_CONJUGATE,
768 m_ahead,
769 n_ahead,
770 buff_1,
771 A22, rs_A, cs_A,
772 a12p, inc_ap,
773 buff_0,
774 w21, inc_w );
775
776 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, FLA_ONE, a12t, y21 );
777 bl1_daxpyv( BLIS1_CONJUGATE,
778 n_ahead,
779 buff_1,
780 a12t, cs_A,
781 y21, inc_y );
782
783 // FLA_Househ2s_UT( FLA_RIGHT,
784 // a12p_t,
785 // a12p_b,
786 // alpha12, psi11_minus_alpha12, sigma11 );
787 FLA_Househ2s_UT_r_opd( n_ahead - 1,
788 a12p_t,
789 a12p_b, inc_ap,
790 &alpha12,
791 &psi11_minus_alpha12,
792 sigma11 );
793
794 // FLA_Copy( a12p, v21 );
795 // FLA_Mult_add( FLA_MINUS_ONE, alpha12, v21_t );
796 // FLA_Inv_scalc( FLA_NO_CONJUGATE, psi11_minus_alpha12, v21 );
797 // FLA_Conjugate( v21_b );
798 bl1_dcopyv( BLIS1_NO_CONJUGATE,
799 n_ahead,
800 a12p, inc_ap,
801 v21, inc_v );
802 bl1_dmult4( buff_m1, &alpha12, v21_t, v21_t );
803 bl1_dinvscalv( BLIS1_NO_CONJUGATE,
804 n_ahead,
805 &psi11_minus_alpha12,
806 v21, inc_v );
807 bl1_dconjv( n_ahead - 1,
808 v21_b, inc_v );
809
810 // FLA_Copyt( FLA_NO_TRANSPOSE, alpha12, a12t_l );
811 // FLA_Copyt( FLA_TRANSPOSE, v21_b, a12t_r );
812 *a12t_l = alpha12;
813 bl1_dcopyv( BLIS1_NO_CONJUGATE,
814 n_ahead - 1,
815 v21_b, inc_v,
816 a12t_r, cs_A );
817 }
818
819 // FLA_Copy( u21p, u21 );
820 bl1_dcopyv( BLIS1_NO_CONJUGATE,
821 m_ahead,
822 u21p, inc_up,
823 u21, inc_u );
824
825 if ( n_ahead > 0 )
826 {
827 // FLA_Dotc( FLA_CONJUGATE, y21, v21, beta );
828 // FLA_Scal( FLA_MINUS_ONE, beta );
829 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, beta );
830 bl1_ddot( BLIS1_CONJUGATE,
831 n_ahead,
832 y21, inc_y,
833 v21, inc_v,
834 &beta );
835 bl1_dscals( &minus_inv_tau11, &beta );
836
837 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha12, minus_conj_alpha12 );
838 // FLA_Scal( FLA_MINUS_ONE, minus_conj_alpha12 );
839 bl1_dcopyconj( &alpha12, &minus_conj_alpha12 );
840 bl1_dneg1( &minus_conj_alpha12 );
841
842 // FLA_Copy( w21, z21 );
843 // FLA_Axpy( minus_conj_alpha12, A22_l, z21 );
844 // FLA_Inv_scalc( FLA_CONJUGATE, psi11_minus_alpha12, z21 );
845 // FLA_Axpy( beta, u21, z21 );
846 bl1_dcopyv( BLIS1_NO_CONJUGATE,
847 m_ahead,
848 w21, inc_w,
849 z21, inc_z );
850 bl1_daxpyv( BLIS1_NO_CONJUGATE,
851 m_ahead,
852 &minus_conj_alpha12,
853 A22_l, rs_A,
854 z21, inc_z );
855 bl1_dinvscalv( BLIS1_CONJUGATE,
856 m_ahead,
857 &psi11_minus_alpha12,
858 z21, inc_z );
859 bl1_daxpyv( BLIS1_NO_CONJUGATE,
860 m_ahead,
861 &beta,
862 u21, inc_u,
863 z21, inc_z );
864
865 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, y21 );
866 // FLA_Inv_scalc( FLA_NO_CONJUGATE, sigma11, z21 );
867 bl1_dinvscalv( BLIS1_CONJUGATE,
868 n_ahead,
869 tau11,
870 y21, inc_y );
871 bl1_dinvscalv( BLIS1_NO_CONJUGATE,
872 m_ahead,
873 sigma11,
874 z21, inc_z );
875
876 // FLA_Gemv( FLA_CONJ_NO_TRANSPOSE, FLA_ONE, A02, v21, FLA_ZERO, s01 );
877 bl1_dgemv( BLIS1_CONJ_NO_TRANSPOSE,
878 BLIS1_NO_CONJUGATE,
879 m_behind,
880 n_ahead,
881 buff_1,
882 A02, rs_A, cs_A,
883 v21, inc_v,
884 buff_0,
885 s01, rs_S );
886 }
887
888 // FLA_Copyt( FLA_CONJ_TRANSPOSE, a10t, t01 );
889 // FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, u21, FLA_ONE, t01 );
890 bl1_dcopyv( BLIS1_CONJUGATE,
891 n_behind,
892 a10t, cs_A,
893 t01, rs_T );
894 bl1_dgemv( BLIS1_CONJ_TRANSPOSE,
895 BLIS1_NO_CONJUGATE,
896 m_ahead,
897 n_behind,
898 buff_1,
899 A20, rs_A, cs_A,
900 u21, inc_u,
901 buff_1,
902 t01, rs_T );
903
904 if ( m_behind + 1 == b_alg && n_ahead > 0 )
905 {
906 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, u21, y21, A22 );
907 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, z21, v21, A22 );
908 bl1_dger( BLIS1_NO_CONJUGATE,
909 BLIS1_CONJUGATE,
910 m_ahead,
911 n_ahead,
912 buff_m1,
913 u21, inc_u,
914 y21, inc_y,
915 A22, rs_A, cs_A );
916 bl1_dger( BLIS1_NO_CONJUGATE,
917 BLIS1_CONJUGATE,
918 m_ahead,
919 n_ahead,
920 buff_m1,
921 z21, inc_z,
922 v21, inc_v,
923 A22, rs_A, cs_A );
924 }
925
926 /*------------------------------------------------------------*/
927
928 }
929
930 // FLA_Obj_free( &w );
931 // FLA_Obj_free( &ap );
932 // FLA_Obj_free( &u );
933 // FLA_Obj_free( &up );
934 // FLA_Obj_free( &v );
935 // FLA_Obj_free( &y );
936 // FLA_Obj_free( &z );
937 FLA_free( buff_w );
938 FLA_free( buff_ap );
939 FLA_free( buff_u );
940 FLA_free( buff_up );
941 FLA_free( buff_v );
942 FLA_free( buff_y );
943 FLA_free( buff_z );
944
945 return FLA_SUCCESS;
946}
FLA_Househ2_UT_l_opd
FLA_Error FLA_Househ2_UT_l_opd(int m_x2, double *chi_1, double *x2, int inc_x2, double *tau)
Definition FLA_Househ2_UT.c:274
FLA_Househ2s_UT_r_opd
FLA_Error FLA_Househ2s_UT_r_opd(int m_x2, double *chi_1, double *x2, int inc_x2, double *alpha, double *chi_1_minus_alpha, double *tau)
Definition FLA_Househ2s_UT.c:572
bl1_daxpyv
void bl1_daxpyv(conj1_t conj, int n, double *alpha, double *x, int incx, double *y, int incy)
Definition bl1_axpyv.c:21
bl1_dconjv
void bl1_dconjv(int m, double *x, int incx)
Definition bl1_conjv.c:18
bl1_dcopyv
void bl1_dcopyv(conj1_t conj, int m, double *x, int incx, double *y, int incy)
Definition bl1_copyv.c:42
bl1_ddot
void bl1_ddot(conj1_t conj, int n, double *x, int incx, double *y, int incy, double *rho)
Definition bl1_dot.c:26
bl1_dgemv
void bl1_dgemv(trans1_t transa, conj1_t conjx, int m, int n, double *alpha, double *a, int a_rs, int a_cs, double *x, int incx, double *beta, double *y, int incy)
Definition bl1_gemv.c:69
bl1_dger
void bl1_dger(conj1_t conjx, conj1_t conjy, int m, int n, double *alpha, double *x, int incx, double *y, int incy, double *a, int a_rs, int a_cs)
Definition bl1_ger.c:62
bl1_dinvscalv
void bl1_dinvscalv(conj1_t conj, int n, double *alpha, double *x, int incx)
Definition bl1_invscalv.c:26

References bl1_daxpyv(), bl1_dconjv(), bl1_dcopyv(), bl1_ddot(), bl1_dgemv(), bl1_dger(), bl1_dinvscalv(), BLIS1_CONJ_NO_TRANSPOSE, BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, FLA_free(), FLA_Househ2_UT_l_opd(), FLA_Househ2s_UT_r_opd(), FLA_malloc(), FLA_MINUS_ONE, FLA_ONE, FLA_ZERO, and i.

Referenced by FLA_Bidiag_UT_u_step_opt_var3().

◆ FLA_Bidiag_UT_u_step_ops_var3()

FLA_Error FLA_Bidiag_UT_u_step_ops_var3 ( int m_A,
int n_A,
int m_TS,
float * buff_A,
int rs_A,
int cs_A,
float * buff_T,
int rs_T,
int cs_T,
float * buff_S,
int rs_S,
int cs_S )
120{
121 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
122 float* buff_0 = FLA_FLOAT_PTR( FLA_ZERO );
123 float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
124
125 float alpha12;
126 float minus_conj_alpha12;
127 float psi11_minus_alpha12;
128 float minus_inv_tau11;
129 float minus_upsilon11;
130 float minus_conj_nu11;
131 float minus_conj_psi11;
132 float minus_zeta11;
133 float beta;
134 int i;
135
136 // b_alg = FLA_Obj_length( T );
137 int b_alg = m_TS;
138
139 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &w );
140 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &ap );
141 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &u );
142 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &up );
143 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &v );
144 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &y );
145 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &z );
146 float* buff_w = ( float* ) FLA_malloc( m_A * sizeof( *buff_A ) );
147 float* buff_ap = ( float* ) FLA_malloc( n_A * sizeof( *buff_A ) );
148 float* buff_u = ( float* ) FLA_malloc( m_A * sizeof( *buff_A ) );
149 float* buff_up = ( float* ) FLA_malloc( m_A * sizeof( *buff_A ) );
150 float* buff_v = ( float* ) FLA_malloc( n_A * sizeof( *buff_A ) );
151 float* buff_y = ( float* ) FLA_malloc( n_A * sizeof( *buff_A ) );
152 float* buff_z = ( float* ) FLA_malloc( m_A * sizeof( *buff_A ) );
153 int inc_w = 1;
154 int inc_ap = 1;
155 int inc_u = 1;
156 int inc_up = 1;
157 int inc_v = 1;
158 int inc_y = 1;
159 int inc_z = 1;
160
161 for ( i = 0; i < b_alg; ++i )
162 {
163 float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
164 float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
165 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
166 float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
167 float* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
168 float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
169 float* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
170
171 float* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
172 float* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
173
174 float* s01 = buff_S + (i )*cs_S + (0 )*rs_S;
175 float* sigma11 = buff_S + (i )*cs_S + (i )*rs_S;
176
177 float* w21 = buff_w + (i+1)*inc_w;
178
179 float* a12p = buff_ap + (i+1)*inc_ap;
180
181 float* upsilon11 = buff_u + (i )*inc_u;
182 float* u21 = buff_u + (i+1)*inc_u;
183
184 float* u21p = buff_up + (i+1)*inc_up;
185
186 float* nu11 = buff_v + (i )*inc_v;
187 float* v21 = buff_v + (i+1)*inc_v;
188
189 float* psi11 = buff_y + (i )*inc_y;
190 float* y21 = buff_y + (i+1)*inc_y;
191
192 float* zeta11 = buff_z + (i )*inc_z;
193 float* z21 = buff_z + (i+1)*inc_z;
194
195 float* a12p_t = a12p + (0 )*inc_ap;
196 float* a12p_b = a12p + (1 )*inc_ap;
197
198 float* v21_t = v21 + (0 )*inc_v;
199 float* v21_b = v21 + (1 )*inc_v;
200
201 float* a12t_l = a12t + (0 )*cs_A + (0 )*rs_A;
202 float* a12t_r = a12t + (1 )*cs_A + (0 )*rs_A;
203
204 float* A22_l = A22 + (0 )*cs_A + (0 )*rs_A;
205
206 int m_ahead = m_A - i - 1;
207 int n_ahead = n_A - i - 1;
208 int m_behind = i;
209 int n_behind = i;
210
211 /*------------------------------------------------------------*/
212
213 if ( m_behind > 0 )
214 {
215 // FLA_Copy( upsilon11, minus_upsilon11 );
216 // FLA_Scal( FLA_MINUS_ONE, minus_upsilon11 );
217 bl1_smult3( buff_m1, upsilon11, &minus_upsilon11 );
218
219 // FLA_Copy( zeta11, minus_zeta11 );
220 // FLA_Scal( FLA_MINUS_ONE, minus_zeta11 );
221 bl1_smult3( buff_m1, zeta11, &minus_zeta11 );
222
223 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, psi11, minus_conj_psi11 );
224 // FLA_Scal( FLA_MINUS_ONE, minus_conj_psi11 );
225 bl1_scopyconj( psi11, &minus_conj_psi11 );
226 bl1_sscals( buff_m1, &minus_conj_psi11 );
227
228 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, nu11, minus_conj_nu11 );
229 // FLA_Scal( FLA_MINUS_ONE, minus_conj_nu11 );
230 bl1_scopyconj( nu11, &minus_conj_nu11 );
231 bl1_sscals( buff_m1, &minus_conj_nu11 );
232
233 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_psi11, upsilon11, alpha11 );
234 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_nu11, zeta11, alpha11 );
235 bl1_saxpyv( BLIS1_NO_CONJUGATE,
236 1,
237 &minus_conj_psi11,
238 upsilon11, 1,
239 alpha11, 1 );
240 bl1_saxpyv( BLIS1_NO_CONJUGATE,
241 1,
242 &minus_conj_nu11,
243 zeta11, 1,
244 alpha11, 1 );
245
246 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_psi11, u21, a21 );
247 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_nu11, z21, a21 );
248 bl1_saxpyv( BLIS1_NO_CONJUGATE,
249 m_ahead,
250 &minus_conj_psi11,
251 u21, inc_u,
252 a21, rs_A );
253 bl1_saxpyv( BLIS1_NO_CONJUGATE,
254 m_ahead,
255 &minus_conj_nu11,
256 z21, inc_z,
257 a21, rs_A );
258
259 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, minus_upsilon11, y21, a12t );
260 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, minus_zeta11, v21, a12t );
261 bl1_saxpyv( BLIS1_CONJUGATE,
262 n_ahead,
263 &minus_upsilon11,
264 y21, inc_y,
265 a12t, cs_A );
266 bl1_saxpyv( BLIS1_CONJUGATE,
267 n_ahead,
268 &minus_zeta11,
269 v21, inc_v,
270 a12t, cs_A );
271 }
272
273 // FLA_Househ2_UT( FLA_LEFT,
274 // alpha11,
275 // a21, tau11 );
276 // FLA_Copy( a21, u21p );
277 FLA_Househ2_UT_l_ops( m_ahead,
278 alpha11,
279 a21, rs_A,
280 tau11 );
281 bl1_scopyv( BLIS1_NO_CONJUGATE,
282 m_ahead,
283 a21, rs_A,
284 u21p, inc_up );
285
286 if ( n_ahead > 0 )
287 {
288 // FLA_Copy( FLA_MINUS_ONE, minus_inv_tau11 );
289 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, minus_inv_tau11 );
290 bl1_sdiv3( buff_m1, tau11, &minus_inv_tau11 );
291
292 // FLA_Copyt( FLA_TRANSPOSE, a12t, a12p );
293 // FLA_Axpyt( FLA_TRANSPOSE, minus_inv_tau11, a12t, a12p );
294 bl1_scopyv( BLIS1_NO_CONJUGATE,
295 n_ahead,
296 a12t, cs_A,
297 a12p, inc_ap );
298 bl1_saxpyv( BLIS1_NO_CONJUGATE,
299 n_ahead,
300 &minus_inv_tau11,
301 a12t, cs_A,
302 a12p, inc_ap );
303 }
304
305 if ( m_behind > 0 )
306 {
307 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, u21, y21, A22 );
308 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, z21, v21, A22 );
309 bl1_sger( BLIS1_NO_CONJUGATE,
310 BLIS1_CONJUGATE,
311 m_ahead,
312 n_ahead,
313 buff_m1,
314 u21, inc_u,
315 y21, inc_y,
316 A22, rs_A, cs_A );
317 bl1_sger( BLIS1_NO_CONJUGATE,
318 BLIS1_CONJUGATE,
319 m_ahead,
320 n_ahead,
321 buff_m1,
322 z21, inc_z,
323 v21, inc_v,
324 A22, rs_A, cs_A );
325 }
326
327 if ( n_ahead > 0 )
328 {
329 // FLA_Gemvc( FLA_CONJ_TRANSPOSE, FLA_NO_CONJUGATE, FLA_ONE, A22, u21p, FLA_ZERO, y21 );
330 bl1_sgemv( BLIS1_CONJ_TRANSPOSE,
331 BLIS1_NO_CONJUGATE,
332 m_ahead,
333 n_ahead,
334 buff_1,
335 A22, rs_A, cs_A,
336 u21p, inc_up,
337 buff_0,
338 y21, inc_y );
339
340 // FLA_Axpyt( FLA_CONJ_NO_TRANSPOSE, minus_inv_tau11, y21, a12p );
341 bl1_saxpyv( BLIS1_CONJUGATE,
342 n_ahead,
343 &minus_inv_tau11,
344 y21, inc_y,
345 a12p, inc_ap );
346
347 // FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A22, a12p, FLA_ZERO, w21 );
348 bl1_sgemv( BLIS1_NO_TRANSPOSE,
349 BLIS1_CONJUGATE,
350 m_ahead,
351 n_ahead,
352 buff_1,
353 A22, rs_A, cs_A,
354 a12p, inc_ap,
355 buff_0,
356 w21, inc_w );
357
358 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, FLA_ONE, a12t, y21 );
359 bl1_saxpyv( BLIS1_CONJUGATE,
360 n_ahead,
361 buff_1,
362 a12t, cs_A,
363 y21, inc_y );
364
365 // FLA_Househ2s_UT( FLA_RIGHT,
366 // a12p_t,
367 // a12p_b,
368 // alpha12, psi11_minus_alpha12, sigma11 );
369 FLA_Househ2s_UT_r_ops( n_ahead - 1,
370 a12p_t,
371 a12p_b, inc_ap,
372 &alpha12,
373 &psi11_minus_alpha12,
374 sigma11 );
375
376 // FLA_Copy( a12p, v21 );
377 // FLA_Mult_add( FLA_MINUS_ONE, alpha12, v21_t );
378 // FLA_Inv_scalc( FLA_NO_CONJUGATE, psi11_minus_alpha12, v21 );
379 // FLA_Conjugate( v21_b );
380 bl1_scopyv( BLIS1_NO_CONJUGATE,
381 n_ahead,
382 a12p, inc_ap,
383 v21, inc_v );
384 bl1_smult4( buff_m1, &alpha12, v21_t, v21_t );
385 bl1_sinvscalv( BLIS1_NO_CONJUGATE,
386 n_ahead,
387 &psi11_minus_alpha12,
388 v21, inc_v );
389 bl1_sconjv( n_ahead - 1,
390 v21_b, inc_v );
391
392 // FLA_Copyt( FLA_NO_TRANSPOSE, alpha12, a12t_l );
393 // FLA_Copyt( FLA_TRANSPOSE, v21_b, a12t_r );
394 *a12t_l = alpha12;
395 bl1_scopyv( BLIS1_NO_CONJUGATE,
396 n_ahead - 1,
397 v21_b, inc_v,
398 a12t_r, cs_A );
399 }
400
401 // FLA_Copy( u21p, u21 );
402 bl1_scopyv( BLIS1_NO_CONJUGATE,
403 m_ahead,
404 u21p, inc_up,
405 u21, inc_u );
406
407 if ( n_ahead > 0 )
408 {
409 // FLA_Dotc( FLA_CONJUGATE, y21, v21, beta );
410 // FLA_Scal( FLA_MINUS_ONE, beta );
411 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, beta );
412 bl1_sdot( BLIS1_CONJUGATE,
413 n_ahead,
414 y21, inc_y,
415 v21, inc_v,
416 &beta );
417 bl1_sscals( &minus_inv_tau11, &beta );
418
419 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha12, minus_conj_alpha12 );
420 // FLA_Scal( FLA_MINUS_ONE, minus_conj_alpha12 );
421 bl1_scopyconj( &alpha12, &minus_conj_alpha12 );
422 bl1_sneg1( &minus_conj_alpha12 );
423
424 // FLA_Copy( w21, z21 );
425 // FLA_Axpy( minus_conj_alpha12, A22_l, z21 );
426 // FLA_Inv_scalc( FLA_CONJUGATE, psi11_minus_alpha12, z21 );
427 // FLA_Axpy( beta, u21, z21 );
428 bl1_scopyv( BLIS1_NO_CONJUGATE,
429 m_ahead,
430 w21, inc_w,
431 z21, inc_z );
432 bl1_saxpyv( BLIS1_NO_CONJUGATE,
433 m_ahead,
434 &minus_conj_alpha12,
435 A22_l, rs_A,
436 z21, inc_z );
437 bl1_sinvscalv( BLIS1_CONJUGATE,
438 m_ahead,
439 &psi11_minus_alpha12,
440 z21, inc_z );
441 bl1_saxpyv( BLIS1_NO_CONJUGATE,
442 m_ahead,
443 &beta,
444 u21, inc_u,
445 z21, inc_z );
446
447 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, y21 );
448 // FLA_Inv_scalc( FLA_NO_CONJUGATE, sigma11, z21 );
449 bl1_sinvscalv( BLIS1_CONJUGATE,
450 n_ahead,
451 tau11,
452 y21, inc_y );
453 bl1_sinvscalv( BLIS1_NO_CONJUGATE,
454 m_ahead,
455 sigma11,
456 z21, inc_z );
457
458 // FLA_Gemv( FLA_CONJ_NO_TRANSPOSE, FLA_ONE, A02, v21, FLA_ZERO, s01 );
459 bl1_sgemv( BLIS1_CONJ_NO_TRANSPOSE,
460 BLIS1_NO_CONJUGATE,
461 m_behind,
462 n_ahead,
463 buff_1,
464 A02, rs_A, cs_A,
465 v21, inc_v,
466 buff_0,
467 s01, rs_S );
468 }
469
470 // FLA_Copyt( FLA_CONJ_TRANSPOSE, a10t, t01 );
471 // FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, u21, FLA_ONE, t01 );
472 bl1_scopyv( BLIS1_CONJUGATE,
473 n_behind,
474 a10t, cs_A,
475 t01, rs_T );
476 bl1_sgemv( BLIS1_CONJ_TRANSPOSE,
477 BLIS1_NO_CONJUGATE,
478 m_ahead,
479 n_behind,
480 buff_1,
481 A20, rs_A, cs_A,
482 u21, inc_u,
483 buff_1,
484 t01, rs_T );
485
486 if ( m_behind + 1 == b_alg && n_ahead > 0 )
487 {
488 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, u21, y21, A22 );
489 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, z21, v21, A22 );
490 bl1_sger( BLIS1_NO_CONJUGATE,
491 BLIS1_CONJUGATE,
492 m_ahead,
493 n_ahead,
494 buff_m1,
495 u21, inc_u,
496 y21, inc_y,
497 A22, rs_A, cs_A );
498 bl1_sger( BLIS1_NO_CONJUGATE,
499 BLIS1_CONJUGATE,
500 m_ahead,
501 n_ahead,
502 buff_m1,
503 z21, inc_z,
504 v21, inc_v,
505 A22, rs_A, cs_A );
506 }
507
508 /*------------------------------------------------------------*/
509
510 }
511
512 // FLA_Obj_free( &w );
513 // FLA_Obj_free( &ap );
514 // FLA_Obj_free( &u );
515 // FLA_Obj_free( &up );
516 // FLA_Obj_free( &v );
517 // FLA_Obj_free( &y );
518 // FLA_Obj_free( &z );
519 FLA_free( buff_w );
520 FLA_free( buff_ap );
521 FLA_free( buff_u );
522 FLA_free( buff_up );
523 FLA_free( buff_v );
524 FLA_free( buff_y );
525 FLA_free( buff_z );
526
527 return FLA_SUCCESS;
528}
FLA_Househ2_UT_l_ops
FLA_Error FLA_Househ2_UT_l_ops(int m_x2, float *chi_1, float *x2, int inc_x2, float *tau)
Definition FLA_Househ2_UT.c:160
FLA_Househ2s_UT_r_ops
FLA_Error FLA_Househ2s_UT_r_ops(int m_x2, float *chi_1, float *x2, int inc_x2, float *alpha, float *chi_1_minus_alpha, float *tau)
Definition FLA_Househ2s_UT.c:555
bl1_saxpyv
void bl1_saxpyv(conj1_t conj, int n, float *alpha, float *x, int incx, float *y, int incy)
Definition bl1_axpyv.c:13
bl1_sconjv
void bl1_sconjv(int m, float *x, int incx)
Definition bl1_conjv.c:13
bl1_scopyv
void bl1_scopyv(conj1_t conj, int m, float *x, int incx, float *y, int incy)
Definition bl1_copyv.c:35
bl1_sdot
void bl1_sdot(conj1_t conj, int n, float *x, int incx, float *y, int incy, float *rho)
Definition bl1_dot.c:13
bl1_sgemv
void bl1_sgemv(trans1_t transa, conj1_t conjx, int m, int n, float *alpha, float *a, int a_rs, int a_cs, float *x, int incx, float *beta, float *y, int incy)
Definition bl1_gemv.c:13
bl1_sger
void bl1_sger(conj1_t conjx, conj1_t conjy, int m, int n, float *alpha, float *x, int incx, float *y, int incy, float *a, int a_rs, int a_cs)
Definition bl1_ger.c:13
bl1_sinvscalv
void bl1_sinvscalv(conj1_t conj, int n, float *alpha, float *x, int incx)
Definition bl1_invscalv.c:13

References bl1_saxpyv(), bl1_sconjv(), bl1_scopyv(), bl1_sdot(), bl1_sgemv(), bl1_sger(), bl1_sinvscalv(), BLIS1_CONJ_NO_TRANSPOSE, BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, FLA_free(), FLA_Househ2_UT_l_ops(), FLA_Househ2s_UT_r_ops(), FLA_malloc(), FLA_MINUS_ONE, FLA_ONE, FLA_ZERO, and i.

Referenced by FLA_Bidiag_UT_u_step_opt_var3().

◆ FLA_Bidiag_UT_u_step_opt_var3()

FLA_Error FLA_Bidiag_UT_u_step_opt_var3 ( FLA_Obj A,
FLA_Obj T,
FLA_Obj S )
19{
20 FLA_Datatype datatype;
21 int m_A, n_A, m_TS;
22 int rs_A, cs_A;
23 int rs_T, cs_T;
24 int rs_S, cs_S;
25
26 datatype = FLA_Obj_datatype( A );
27
28 m_A = FLA_Obj_length( A );
29 n_A = FLA_Obj_width( A );
30 m_TS = FLA_Obj_length( T );
31
32 rs_A = FLA_Obj_row_stride( A );
33 cs_A = FLA_Obj_col_stride( A );
34
35 rs_T = FLA_Obj_row_stride( T );
36 cs_T = FLA_Obj_col_stride( T );
37
38 rs_S = FLA_Obj_row_stride( S );
39 cs_S = FLA_Obj_col_stride( S );
40
41
42 switch ( datatype )
43 {
44 case FLA_FLOAT:
45 {
46 float* buff_A = FLA_FLOAT_PTR( A );
47 float* buff_T = FLA_FLOAT_PTR( T );
48 float* buff_S = FLA_FLOAT_PTR( S );
49
50 FLA_Bidiag_UT_u_step_ops_var3( m_A,
51 n_A,
52 m_TS,
53 buff_A, rs_A, cs_A,
54 buff_T, rs_T, cs_T,
55 buff_S, rs_S, cs_S );
56
57 break;
58 }
59
60 case FLA_DOUBLE:
61 {
62 double* buff_A = FLA_DOUBLE_PTR( A );
63 double* buff_T = FLA_DOUBLE_PTR( T );
64 double* buff_S = FLA_DOUBLE_PTR( S );
65
66 FLA_Bidiag_UT_u_step_opd_var3( m_A,
67 n_A,
68 m_TS,
69 buff_A, rs_A, cs_A,
70 buff_T, rs_T, cs_T,
71 buff_S, rs_S, cs_S );
72
73 break;
74 }
75
76 case FLA_COMPLEX:
77 {
78 scomplex* buff_A = FLA_COMPLEX_PTR( A );
79 scomplex* buff_T = FLA_COMPLEX_PTR( T );
80 scomplex* buff_S = FLA_COMPLEX_PTR( S );
81
82 FLA_Bidiag_UT_u_step_opc_var3( m_A,
83 n_A,
84 m_TS,
85 buff_A, rs_A, cs_A,
86 buff_T, rs_T, cs_T,
87 buff_S, rs_S, cs_S );
88
89 break;
90 }
91
92 case FLA_DOUBLE_COMPLEX:
93 {
94 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
95 dcomplex* buff_T = FLA_DOUBLE_COMPLEX_PTR( T );
96 dcomplex* buff_S = FLA_DOUBLE_COMPLEX_PTR( S );
97
98 FLA_Bidiag_UT_u_step_opz_var3( m_A,
99 n_A,
100 m_TS,
101 buff_A, rs_A, cs_A,
102 buff_T, rs_T, cs_T,
103 buff_S, rs_S, cs_S );
104
105 break;
106 }
107 }
108
109 return FLA_SUCCESS;
110}
FLA_Bidiag_UT_u_step_opz_var3
FLA_Error FLA_Bidiag_UT_u_step_opz_var3(int m_A, int n_A, int m_TS, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_T, int rs_T, int cs_T, dcomplex *buff_S, int rs_S, int cs_S)
Definition FLA_Bidiag_UT_u_opt_var3.c:1368
FLA_Bidiag_UT_u_step_ops_var3
FLA_Error FLA_Bidiag_UT_u_step_ops_var3(int m_A, int n_A, int m_TS, float *buff_A, int rs_A, int cs_A, float *buff_T, int rs_T, int cs_T, float *buff_S, int rs_S, int cs_S)
Definition FLA_Bidiag_UT_u_opt_var3.c:114
FLA_Bidiag_UT_u_step_opd_var3
FLA_Error FLA_Bidiag_UT_u_step_opd_var3(int m_A, int n_A, int m_TS, double *buff_A, int rs_A, int cs_A, double *buff_T, int rs_T, int cs_T, double *buff_S, int rs_S, int cs_S)
Definition FLA_Bidiag_UT_u_opt_var3.c:532
FLA_Bidiag_UT_u_step_opc_var3
FLA_Error FLA_Bidiag_UT_u_step_opc_var3(int m_A, int n_A, int m_TS, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_T, int rs_T, int cs_T, scomplex *buff_S, int rs_S, int cs_S)
Definition FLA_Bidiag_UT_u_opt_var3.c:950
FLA_Obj_width
dim_t FLA_Obj_width(FLA_Obj obj)
Definition FLA_Query.c:123
FLA_Obj_row_stride
dim_t FLA_Obj_row_stride(FLA_Obj obj)
Definition FLA_Query.c:167
FLA_Obj_length
dim_t FLA_Obj_length(FLA_Obj obj)
Definition FLA_Query.c:116
FLA_Obj_col_stride
dim_t FLA_Obj_col_stride(FLA_Obj obj)
Definition FLA_Query.c:174
FLA_Obj_datatype
FLA_Datatype FLA_Obj_datatype(FLA_Obj obj)
Definition FLA_Query.c:13
FLA_Datatype
int FLA_Datatype
Definition FLA_type_defs.h:49
dcomplex
Definition blis_type_defs.h:138

References FLA_Bidiag_UT_u_step_opc_var3(), FLA_Bidiag_UT_u_step_opd_var3(), FLA_Bidiag_UT_u_step_ops_var3(), FLA_Bidiag_UT_u_step_opz_var3(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), and FLA_Obj_width().

Referenced by FLA_Bidiag_UT_u_blk_var3(), and FLA_Bidiag_UT_u_opt_var3().

◆ FLA_Bidiag_UT_u_step_opz_var3()

FLA_Error FLA_Bidiag_UT_u_step_opz_var3 ( int m_A,
int n_A,
int m_TS,
dcomplex * buff_A,
int rs_A,
int cs_A,
dcomplex * buff_T,
int rs_T,
int cs_T,
dcomplex * buff_S,
int rs_S,
int cs_S )
1374{
1375 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
1376 dcomplex* buff_0 = FLA_DOUBLE_COMPLEX_PTR( FLA_ZERO );
1377 dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
1378
1379 dcomplex alpha12;
1380 dcomplex minus_conj_alpha12;
1381 dcomplex psi11_minus_alpha12;
1382 dcomplex minus_inv_tau11;
1383 dcomplex minus_upsilon11;
1384 dcomplex minus_conj_nu11;
1385 dcomplex minus_conj_psi11;
1386 dcomplex minus_zeta11;
1387 dcomplex beta;
1388 int i;
1389
1390 // b_alg = FLA_Obj_length( T );
1391 int b_alg = m_TS;
1392
1393 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &w );
1394 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &ap );
1395 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &u );
1396 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &up );
1397 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &v );
1398 // FLA_Obj_create( datatype_A, n_A, 1, 0, 0, &y );
1399 // FLA_Obj_create( datatype_A, m_A, 1, 0, 0, &z );
1400 dcomplex* buff_w = ( dcomplex* ) FLA_malloc( m_A * sizeof( *buff_A ) );
1401 dcomplex* buff_ap = ( dcomplex* ) FLA_malloc( n_A * sizeof( *buff_A ) );
1402 dcomplex* buff_u = ( dcomplex* ) FLA_malloc( m_A * sizeof( *buff_A ) );
1403 dcomplex* buff_up = ( dcomplex* ) FLA_malloc( m_A * sizeof( *buff_A ) );
1404 dcomplex* buff_v = ( dcomplex* ) FLA_malloc( n_A * sizeof( *buff_A ) );
1405 dcomplex* buff_y = ( dcomplex* ) FLA_malloc( n_A * sizeof( *buff_A ) );
1406 dcomplex* buff_z = ( dcomplex* ) FLA_malloc( m_A * sizeof( *buff_A ) );
1407 int inc_w = 1;
1408 int inc_ap = 1;
1409 int inc_u = 1;
1410 int inc_up = 1;
1411 int inc_v = 1;
1412 int inc_y = 1;
1413 int inc_z = 1;
1414
1415 for ( i = 0; i < b_alg; ++i )
1416 {
1417 dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
1418 dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
1419 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
1420 dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
1421 dcomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
1422 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
1423 dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
1424
1425 dcomplex* t01 = buff_T + (i )*cs_T + (0 )*rs_T;
1426 dcomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
1427
1428 dcomplex* s01 = buff_S + (i )*cs_S + (0 )*rs_S;
1429 dcomplex* sigma11 = buff_S + (i )*cs_S + (i )*rs_S;
1430
1431 dcomplex* w21 = buff_w + (i+1)*inc_w;
1432
1433 dcomplex* a12p = buff_ap + (i+1)*inc_ap;
1434
1435 dcomplex* upsilon11 = buff_u + (i )*inc_u;
1436 dcomplex* u21 = buff_u + (i+1)*inc_u;
1437
1438 dcomplex* u21p = buff_up + (i+1)*inc_up;
1439
1440 dcomplex* nu11 = buff_v + (i )*inc_v;
1441 dcomplex* v21 = buff_v + (i+1)*inc_v;
1442
1443 dcomplex* psi11 = buff_y + (i )*inc_y;
1444 dcomplex* y21 = buff_y + (i+1)*inc_y;
1445
1446 dcomplex* zeta11 = buff_z + (i )*inc_z;
1447 dcomplex* z21 = buff_z + (i+1)*inc_z;
1448
1449 dcomplex* a12p_t = a12p + (0 )*inc_ap;
1450 dcomplex* a12p_b = a12p + (1 )*inc_ap;
1451
1452 dcomplex* v21_t = v21 + (0 )*inc_v;
1453 dcomplex* v21_b = v21 + (1 )*inc_v;
1454
1455 dcomplex* a12t_l = a12t + (0 )*cs_A + (0 )*rs_A;
1456 dcomplex* a12t_r = a12t + (1 )*cs_A + (0 )*rs_A;
1457
1458 dcomplex* A22_l = A22 + (0 )*cs_A + (0 )*rs_A;
1459
1460 int m_ahead = m_A - i - 1;
1461 int n_ahead = n_A - i - 1;
1462 int m_behind = i;
1463 int n_behind = i;
1464
1465 /*------------------------------------------------------------*/
1466
1467 if ( m_behind > 0 )
1468 {
1469 // FLA_Copy( upsilon11, minus_upsilon11 );
1470 // FLA_Scal( FLA_MINUS_ONE, minus_upsilon11 );
1471 bl1_zmult3( buff_m1, upsilon11, &minus_upsilon11 );
1472
1473 // FLA_Copy( zeta11, minus_zeta11 );
1474 // FLA_Scal( FLA_MINUS_ONE, minus_zeta11 );
1475 bl1_zmult3( buff_m1, zeta11, &minus_zeta11 );
1476
1477 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, psi11, minus_conj_psi11 );
1478 // FLA_Scal( FLA_MINUS_ONE, minus_conj_psi11 );
1479 bl1_zcopyconj( psi11, &minus_conj_psi11 );
1480 bl1_zscals( buff_m1, &minus_conj_psi11 );
1481
1482 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, nu11, minus_conj_nu11 );
1483 // FLA_Scal( FLA_MINUS_ONE, minus_conj_nu11 );
1484 bl1_zcopyconj( nu11, &minus_conj_nu11 );
1485 bl1_zscals( buff_m1, &minus_conj_nu11 );
1486
1487 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_psi11, upsilon11, alpha11 );
1488 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_nu11, zeta11, alpha11 );
1489 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
1490 1,
1491 &minus_conj_psi11,
1492 upsilon11, 1,
1493 alpha11, 1 );
1494 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
1495 1,
1496 &minus_conj_nu11,
1497 zeta11, 1,
1498 alpha11, 1 );
1499
1500 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_psi11, u21, a21 );
1501 // FLA_Axpyt( FLA_NO_TRANSPOSE, minus_conj_nu11, z21, a21 );
1502 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
1503 m_ahead,
1504 &minus_conj_psi11,
1505 u21, inc_u,
1506 a21, rs_A );
1507 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
1508 m_ahead,
1509 &minus_conj_nu11,
1510 z21, inc_z,
1511 a21, rs_A );
1512
1513 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, minus_upsilon11, y21, a12t );
1514 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, minus_zeta11, v21, a12t );
1515 bl1_zaxpyv( BLIS1_CONJUGATE,
1516 n_ahead,
1517 &minus_upsilon11,
1518 y21, inc_y,
1519 a12t, cs_A );
1520 bl1_zaxpyv( BLIS1_CONJUGATE,
1521 n_ahead,
1522 &minus_zeta11,
1523 v21, inc_v,
1524 a12t, cs_A );
1525 }
1526
1527 // FLA_Househ2_UT( FLA_LEFT,
1528 // alpha11,
1529 // a21, tau11 );
1530 // FLA_Copy( a21, u21p );
1531 FLA_Househ2_UT_l_opz( m_ahead,
1532 alpha11,
1533 a21, rs_A,
1534 tau11 );
1535 bl1_zcopyv( BLIS1_NO_CONJUGATE,
1536 m_ahead,
1537 a21, rs_A,
1538 u21p, inc_up );
1539
1540 if ( n_ahead > 0 )
1541 {
1542 // FLA_Copy( FLA_MINUS_ONE, minus_inv_tau11 );
1543 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, minus_inv_tau11 );
1544 bl1_zdiv3( buff_m1, tau11, &minus_inv_tau11 );
1545
1546 // FLA_Copyt( FLA_TRANSPOSE, a12t, a12p );
1547 // FLA_Axpyt( FLA_TRANSPOSE, minus_inv_tau11, a12t, a12p );
1548 bl1_zcopyv( BLIS1_NO_CONJUGATE,
1549 n_ahead,
1550 a12t, cs_A,
1551 a12p, inc_ap );
1552 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
1553 n_ahead,
1554 &minus_inv_tau11,
1555 a12t, cs_A,
1556 a12p, inc_ap );
1557 }
1558
1559 if ( m_behind > 0 )
1560 {
1561 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, u21, y21, A22 );
1562 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, z21, v21, A22 );
1563 bl1_zger( BLIS1_NO_CONJUGATE,
1564 BLIS1_CONJUGATE,
1565 m_ahead,
1566 n_ahead,
1567 buff_m1,
1568 u21, inc_u,
1569 y21, inc_y,
1570 A22, rs_A, cs_A );
1571 bl1_zger( BLIS1_NO_CONJUGATE,
1572 BLIS1_CONJUGATE,
1573 m_ahead,
1574 n_ahead,
1575 buff_m1,
1576 z21, inc_z,
1577 v21, inc_v,
1578 A22, rs_A, cs_A );
1579 }
1580
1581 if ( n_ahead > 0 )
1582 {
1583 // FLA_Gemvc( FLA_CONJ_TRANSPOSE, FLA_NO_CONJUGATE, FLA_ONE, A22, u21p, FLA_ZERO, y21 );
1584 bl1_zgemv( BLIS1_CONJ_TRANSPOSE,
1585 BLIS1_NO_CONJUGATE,
1586 m_ahead,
1587 n_ahead,
1588 buff_1,
1589 A22, rs_A, cs_A,
1590 u21p, inc_up,
1591 buff_0,
1592 y21, inc_y );
1593
1594 // FLA_Axpyt( FLA_CONJ_NO_TRANSPOSE, minus_inv_tau11, y21, a12p );
1595 bl1_zaxpyv( BLIS1_CONJUGATE,
1596 n_ahead,
1597 &minus_inv_tau11,
1598 y21, inc_y,
1599 a12p, inc_ap );
1600
1601 // FLA_Gemvc( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A22, a12p, FLA_ZERO, w21 );
1602 bl1_zgemv( BLIS1_NO_TRANSPOSE,
1603 BLIS1_CONJUGATE,
1604 m_ahead,
1605 n_ahead,
1606 buff_1,
1607 A22, rs_A, cs_A,
1608 a12p, inc_ap,
1609 buff_0,
1610 w21, inc_w );
1611
1612 // FLA_Axpyt( FLA_CONJ_TRANSPOSE, FLA_ONE, a12t, y21 );
1613 bl1_zaxpyv( BLIS1_CONJUGATE,
1614 n_ahead,
1615 buff_1,
1616 a12t, cs_A,
1617 y21, inc_y );
1618
1619 // FLA_Househ2s_UT( FLA_RIGHT,
1620 // a12p_t,
1621 // a12p_b,
1622 // alpha12, psi11_minus_alpha12, sigma11 );
1623 FLA_Househ2s_UT_r_opz( n_ahead - 1,
1624 a12p_t,
1625 a12p_b, inc_ap,
1626 &alpha12,
1627 &psi11_minus_alpha12,
1628 sigma11 );
1629
1630 // FLA_Copy( a12p, v21 );
1631 // FLA_Mult_add( FLA_MINUS_ONE, alpha12, v21_t );
1632 // FLA_Inv_scalc( FLA_NO_CONJUGATE, psi11_minus_alpha12, v21 );
1633 // FLA_Conjugate( v21_b );
1634 bl1_zcopyv( BLIS1_NO_CONJUGATE,
1635 n_ahead,
1636 a12p, inc_ap,
1637 v21, inc_v );
1638 bl1_zmult4( buff_m1, &alpha12, v21_t, v21_t );
1639 bl1_zinvscalv( BLIS1_NO_CONJUGATE,
1640 n_ahead,
1641 &psi11_minus_alpha12,
1642 v21, inc_v );
1643 bl1_zconjv( n_ahead - 1,
1644 v21_b, inc_v );
1645
1646 // FLA_Copyt( FLA_NO_TRANSPOSE, alpha12, a12t_l );
1647 // FLA_Copyt( FLA_TRANSPOSE, v21_b, a12t_r );
1648 *a12t_l = alpha12;
1649 bl1_zcopyv( BLIS1_NO_CONJUGATE,
1650 n_ahead - 1,
1651 v21_b, inc_v,
1652 a12t_r, cs_A );
1653 }
1654
1655 // FLA_Copy( u21p, u21 );
1656 bl1_zcopyv( BLIS1_NO_CONJUGATE,
1657 m_ahead,
1658 u21p, inc_up,
1659 u21, inc_u );
1660
1661 if ( n_ahead > 0 )
1662 {
1663 // FLA_Dotc( FLA_CONJUGATE, y21, v21, beta );
1664 // FLA_Scal( FLA_MINUS_ONE, beta );
1665 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, beta );
1666 bl1_zdot( BLIS1_CONJUGATE,
1667 n_ahead,
1668 y21, inc_y,
1669 v21, inc_v,
1670 &beta );
1671 bl1_zscals( &minus_inv_tau11, &beta );
1672
1673 // FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, alpha12, minus_conj_alpha12 );
1674 // FLA_Scal( FLA_MINUS_ONE, minus_conj_alpha12 );
1675 bl1_zcopyconj( &alpha12, &minus_conj_alpha12 );
1676 bl1_zneg1( &minus_conj_alpha12 );
1677
1678 // FLA_Copy( w21, z21 );
1679 // FLA_Axpy( minus_conj_alpha12, A22_l, z21 );
1680 // FLA_Inv_scalc( FLA_CONJUGATE, psi11_minus_alpha12, z21 );
1681 // FLA_Axpy( beta, u21, z21 );
1682 bl1_zcopyv( BLIS1_NO_CONJUGATE,
1683 m_ahead,
1684 w21, inc_w,
1685 z21, inc_z );
1686 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
1687 m_ahead,
1688 &minus_conj_alpha12,
1689 A22_l, rs_A,
1690 z21, inc_z );
1691 bl1_zinvscalv( BLIS1_CONJUGATE,
1692 m_ahead,
1693 &psi11_minus_alpha12,
1694 z21, inc_z );
1695 bl1_zaxpyv( BLIS1_NO_CONJUGATE,
1696 m_ahead,
1697 &beta,
1698 u21, inc_u,
1699 z21, inc_z );
1700
1701 // FLA_Inv_scalc( FLA_NO_CONJUGATE, tau11, y21 );
1702 // FLA_Inv_scalc( FLA_NO_CONJUGATE, sigma11, z21 );
1703 bl1_zinvscalv( BLIS1_CONJUGATE,
1704 n_ahead,
1705 tau11,
1706 y21, inc_y );
1707 bl1_zinvscalv( BLIS1_NO_CONJUGATE,
1708 m_ahead,
1709 sigma11,
1710 z21, inc_z );
1711
1712 // FLA_Gemv( FLA_CONJ_NO_TRANSPOSE, FLA_ONE, A02, v21, FLA_ZERO, s01 );
1713 bl1_zgemv( BLIS1_CONJ_NO_TRANSPOSE,
1714 BLIS1_NO_CONJUGATE,
1715 m_behind,
1716 n_ahead,
1717 buff_1,
1718 A02, rs_A, cs_A,
1719 v21, inc_v,
1720 buff_0,
1721 s01, rs_S );
1722 }
1723
1724 // FLA_Copyt( FLA_CONJ_TRANSPOSE, a10t, t01 );
1725 // FLA_Gemv( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, u21, FLA_ONE, t01 );
1726 bl1_zcopyv( BLIS1_CONJUGATE,
1727 n_behind,
1728 a10t, cs_A,
1729 t01, rs_T );
1730 bl1_zgemv( BLIS1_CONJ_TRANSPOSE,
1731 BLIS1_NO_CONJUGATE,
1732 m_ahead,
1733 n_behind,
1734 buff_1,
1735 A20, rs_A, cs_A,
1736 u21, inc_u,
1737 buff_1,
1738 t01, rs_T );
1739
1740 if ( m_behind + 1 == b_alg && n_ahead > 0 )
1741 {
1742 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, u21, y21, A22 );
1743 // FLA_Gerc( FLA_NO_CONJUGATE, FLA_CONJUGATE, FLA_MINUS_ONE, z21, v21, A22 );
1744 bl1_zger( BLIS1_NO_CONJUGATE,
1745 BLIS1_CONJUGATE,
1746 m_ahead,
1747 n_ahead,
1748 buff_m1,
1749 u21, inc_u,
1750 y21, inc_y,
1751 A22, rs_A, cs_A );
1752 bl1_zger( BLIS1_NO_CONJUGATE,
1753 BLIS1_CONJUGATE,
1754 m_ahead,
1755 n_ahead,
1756 buff_m1,
1757 z21, inc_z,
1758 v21, inc_v,
1759 A22, rs_A, cs_A );
1760 }
1761
1762 /*------------------------------------------------------------*/
1763
1764 }
1765
1766 // FLA_Obj_free( &w );
1767 // FLA_Obj_free( &ap );
1768 // FLA_Obj_free( &u );
1769 // FLA_Obj_free( &up );
1770 // FLA_Obj_free( &v );
1771 // FLA_Obj_free( &y );
1772 // FLA_Obj_free( &z );
1773 FLA_free( buff_w );
1774 FLA_free( buff_ap );
1775 FLA_free( buff_u );
1776 FLA_free( buff_up );
1777 FLA_free( buff_v );
1778 FLA_free( buff_y );
1779 FLA_free( buff_z );
1780
1781 return FLA_SUCCESS;
1782}
FLA_Househ2_UT_l_opz
FLA_Error FLA_Househ2_UT_l_opz(int m_x2, dcomplex *chi_1, dcomplex *x2, int inc_x2, dcomplex *tau)
Definition FLA_Househ2_UT.c:521
FLA_Househ2s_UT_r_opz
FLA_Error FLA_Househ2s_UT_r_opz(int m_x2, dcomplex *chi_1, dcomplex *x2, int inc_x2, dcomplex *alpha, dcomplex *chi_1_minus_alpha, dcomplex *tau)
Definition FLA_Househ2s_UT.c:610
bl1_zaxpyv
void bl1_zaxpyv(conj1_t conj, int n, dcomplex *alpha, dcomplex *x, int incx, dcomplex *y, int incy)
Definition bl1_axpyv.c:60
bl1_zconjv
void bl1_zconjv(int m, dcomplex *x, int incx)
Definition bl1_conjv.c:34
bl1_zcopyv
void bl1_zcopyv(conj1_t conj, int m, dcomplex *x, int incx, dcomplex *y, int incy)
Definition bl1_copyv.c:63
bl1_zdot
void bl1_zdot(conj1_t conj, int n, dcomplex *x, int incx, dcomplex *y, int incy, dcomplex *rho)
Definition bl1_dot.c:65
bl1_zscals
bl1_zscals(beta, rho_yz)
bl1_zgemv
void bl1_zgemv(trans1_t transa, conj1_t conjx, int m, int n, dcomplex *alpha, dcomplex *a, int a_rs, int a_cs, dcomplex *x, int incx, dcomplex *beta, dcomplex *y, int incy)
Definition bl1_gemv.c:255
bl1_zger
void bl1_zger(conj1_t conjx, conj1_t conjy, int m, int n, dcomplex *alpha, dcomplex *x, int incx, dcomplex *y, int incy, dcomplex *a, int a_rs, int a_cs)
Definition bl1_ger.c:194
bl1_zinvscalv
void bl1_zinvscalv(conj1_t conj, int n, dcomplex *alpha, dcomplex *x, int incx)
Definition bl1_invscalv.c:78

References bl1_zaxpyv(), bl1_zconjv(), bl1_zcopyv(), bl1_zdot(), bl1_zgemv(), bl1_zger(), bl1_zinvscalv(), bl1_zscals(), BLIS1_CONJ_NO_TRANSPOSE, BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, FLA_free(), FLA_Househ2_UT_l_opz(), FLA_Househ2s_UT_r_opz(), FLA_malloc(), FLA_MINUS_ONE, FLA_ONE, FLA_ZERO, and i.

Referenced by FLA_Bidiag_UT_u_step_opt_var3().

Generated on Thu Sep 23 2021 17:46:17 for libflame by doxygen 1.12.0