void
GradDpElement :: computeStiffnessMatrix_ku(FloatMatrix &answer, MatResponseMode rMode, TimeStep *tStep)
{
double dV;
NLStructuralElement *elem = this->giveNLStructuralElement();
FloatArray Nk;
FloatMatrix B, DkuB, Dku;
StructuralCrossSection *cs = elem->giveStructuralCrossSection();
answer.clear();
int nlGeo = elem->giveGeometryMode();
for ( auto &gp: *elem->giveIntegrationRule(0) ) {
GradDpMaterialExtensionInterface *dpmat = dynamic_cast< GradDpMaterialExtensionInterface * >(
cs->giveMaterialInterface(GradDpMaterialExtensionInterfaceType, gp) );
if ( !dpmat ) {
OOFEM_ERROR("Material doesn't implement the required DpGrad interface!");
}
elem->computeBmatrixAt(gp, B);
if ( nlGeo == 1 ) {
if ( elem->domain->giveEngngModel()->giveFormulation() == AL ) {
elem->computeBmatrixAt(gp, B);
} else {
elem->computeBHmatrixAt(gp, B);
}
}
dpmat->givePDGradMatrix_ku(Dku, rMode, gp, tStep);
this->computeNkappaMatrixAt(gp, Nk);
dV = elem->computeVolumeAround(gp);
DkuB.beProductOf(Dku, B);
answer.plusProductUnsym(Nk, DkuB, -dV);
if ( dpmat->giveAveragingType() == 2 ) {
double dl1, dl2, dl3;
FloatArray Gk;
FloatMatrix D, DB, LDB;
FloatMatrix Bk, BktM22, BktM22Gk, BktM12, BktM12Gk, M22(2, 2), M12(2, 2);
FloatMatrix dL1(1, 3), dL2(1, 3), result1, result2, dLdS, n(2, 2);
this->computeBkappaMatrixAt(gp, Bk);
dpmat->givePDGradMatrix_uu(D, rMode, gp, tStep);
dpmat->givePDGradMatrix_LD(dLdS, rMode, gp, tStep);
this->computeNonlocalGradient(Gk, gp, tStep);
dl1 = dLdS.at(3, 3);
dl2 = dLdS.at(4, 4);
dl3 = dLdS.at(5, 5);
n.at(1, 1) = dLdS.at(1, 1);
n.at(1, 2) = dLdS.at(1, 2);
n.at(2, 1) = dLdS.at(2, 1);
n.at(2, 2) = dLdS.at(2, 2);
// first term Bk^T M22 G L1 D B
// M22 = n2 \otimes n2
M22.at(1, 1) = n.at(1, 2) * n.at(1, 2);
M22.at(1, 2) = n.at(1, 2) * n.at(2, 2);
M22.at(2, 1) = n.at(2, 2) * n.at(1, 2);
M22.at(2, 2) = n.at(2, 2) * n.at(2, 2);
// dL1
dL1.at(1, 1) = dl1 * n.at(1, 1) * n.at(1, 1) + dl2 *n.at(1, 2) * n.at(1, 2);
dL1.at(1, 2) = dl1 * n.at(2, 1) * n.at(2, 1) + dl2 *n.at(2, 2) * n.at(2, 2);
dL1.at(1, 3) = dl1 * n.at(1, 1) * n.at(2, 1) + dl2 *n.at(1, 2) * n.at(2, 2);
DB.beProductOf(D, B);
LDB.beProductOf(dL1, DB);
BktM22.beTProductOf(Bk, M22);
///@todo This can't possibly work if this is uncommented (!) / Mikael
//BktM22Gk.beProductOf(BktM22,Gk);
result1.beProductOf(BktM22Gk, LDB);
answer.add(dV, result1);
// This would be slightly shorter and faster;
//GkLDB.beProductOf(Gk, LDB);
//MGkLDB.beProductOf(M22, GkLDB);
//answer.plusProductUnsym(Bk, MGkLDB, dV);
// M12 + M21 = n1 \otimes n2 + n2 \otimes n1
M12.at(1, 1) = n.at(1, 1) * n.at(1, 2) + n.at(1, 2) * n.at(1, 1);
M12.at(1, 2) = n.at(1, 1) * n.at(2, 2) + n.at(1, 2) * n.at(2, 1);
M12.at(2, 1) = n.at(2, 1) * n.at(1, 2) + n.at(2, 2) * n.at(1, 1);
M12.at(2, 2) = n.at(2, 1) * n.at(2, 2) + n.at(2, 2) * n.at(2, 1);
//dL2
dL2.at(1, 1) = dl3 * ( n.at(1, 1) * n.at(1, 2) + n.at(1, 1) * n.at(1, 2) );
dL2.at(1, 2) = dl3 * ( n.at(2, 1) * n.at(2, 2) + n.at(2, 1) * n.at(2, 2) );
dL2.at(1, 3) = dl3 * ( n.at(1, 2) * n.at(2, 1) + n.at(1, 1) * n.at(2, 2) );
LDB.beProductOf(dL2, DB);
BktM12.beTProductOf(Bk, M12);
///@todo This can't possibly work if this is uncommented (!) / Mikael
//BktM12Gk.beProductOf(BktM12,Gk);
result2.beProductOf(BktM12Gk, LDB);
answer.add(dV, result2);
// This would be slightly shorter and faster;
//GkLDB.beProductOf(Gk, LDB);
//MGkLDB.beProductOf(M12, GkLDB);
//answer.plusProductUnsym(Bk, MGkLDB, dV);
}
}
}
extern "C" magma_int_t
magma_zgessm_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t k, magma_int_t ib,
magma_int_t *ipiv,
magmaDoubleComplex *dL1, magma_int_t lddl1,
magmaDoubleComplex *dL, magma_int_t lddl,
magmaDoubleComplex *dA, magma_int_t ldda,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
December 2013
Purpose
=======
ZGESSM applies the factors L computed by ZGETRF_INCPIV to
a complex M-by-N tile A.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
K (input) INTEGER
The number of columns of the matrix L. K >= 0.
IB (input) INTEGER
The inner-blocking size. IB >= 0.
IPIV (input) INTEGER array on the cpu.
The pivot indices array of size K as returned by
ZGETRF_INCPIV.
dL1 (input) DOUBLE COMPLEX array, dimension(LDDL1, N)
The IB-by-K matrix in which is stored L^(-1) as returned by GETRF_INCPIV
LDDL1 (input) INTEGER
The leading dimension of the array L1. LDDL1 >= max(1,2*IB).
dL (input) DOUBLE COMPLEX array, dimension(LDDL, N)
The M-by-K lower triangular tile on the gpu.
LDDL (input) INTEGER
The leading dimension of the array L. LDDL >= max(1,M).
dA (input/output) DOUBLE COMPLEX array, dimension (LDDA, N)
On entry, the M-by-N tile A on the gpu.
On exit, updated by the application of L on the gpu.
===================================================================== */
#define AT(i,j) (dAT + (i)*ldda + (j) )
#define L(i,j) (dL + (i) + (j)*lddl )
#define dL1(j) (dL1 + (j)*lddl1)
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
int i, s, sb;
magmaDoubleComplex *dAT;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (ldda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
if ( (storev == 'C') || (storev == 'c') ) {
magmablas_zgetmo_in( dA, dAT, ldda, m, n );
} else {
dAT = dA;
}
s = k / ib;
for(i = 0; i < k; i += ib) {
sb = min(ib, k-i);
magmablas_zlaswp( n, dAT, ldda, i+1, i+sb, ipiv, 1 );
#ifndef WITHOUTTRTRI
magma_ztrmm( MagmaRight, MagmaLower, MagmaTrans, MagmaUnit,
n, sb,
c_one, dL1(i), lddl1,
AT(i, 0), ldda);
#else
magma_ztrsm( MagmaRight, MagmaLower, MagmaTrans, MagmaUnit,
n, sb,
c_one, L( i, i), lddl,
AT(i, 0), ldda);
#endif
if ( (i+sb) < m) {
magma_zgemm( MagmaNoTrans, MagmaTrans,
n, m-(i+sb), sb,
c_neg_one, AT(i, 0), ldda,
L( i+sb, i), lddl,
c_one, AT(i+sb, 0), ldda );
}
}
if ( (storev == 'C') || (storev == 'c') ) {
magmablas_zgetmo_in( dA, dAT, ldda, m, n );
}
return *info;
/* End of MAGMA_ZGETRF_GPU */
}
void
GradDpElement :: computeStiffnessMatrix(FloatMatrix &answer, MatResponseMode rMode, TimeStep *tStep)
{
//set displacement and nonlocal location array
this->setDisplacementLocationArray();
this->setNonlocalLocationArray();
NLStructuralElement *elem = this->giveNLStructuralElement();
StructuralCrossSection *cs = elem->giveStructuralCrossSection();
FloatMatrix B, D, DB;
FloatMatrix DkuB, Dku;
FloatArray Nk;
FloatMatrix SNk, gPSigma;
FloatMatrix lStiff;
FloatMatrix Bk, LBk;
FloatMatrix answer_uu, answer_ku, answer_uk, answer_kk;
int nlGeo = elem->giveGeometryMode();
bool matStiffSymmFlag = elem->giveCrossSection()->isCharacteristicMtrxSymmetric(rMode);
for ( auto &gp : *elem->giveIntegrationRule(0) ) {
GradDpMaterialExtensionInterface *dpmat = dynamic_cast< GradDpMaterialExtensionInterface * >(
cs->giveMaterialInterface(GradDpMaterialExtensionInterfaceType, gp) );
if ( !dpmat ) {
OOFEM_ERROR("Material doesn't implement the required DpGrad interface!");
}
double dV = elem->computeVolumeAround(gp);
if ( nlGeo == 0 ) {
elem->computeBmatrixAt(gp, B);
} else if ( nlGeo == 1 ) {
if ( elem->domain->giveEngngModel()->giveFormulation() == AL ) {
elem->computeBmatrixAt(gp, B);
} else {
elem->computeBHmatrixAt(gp, B);
}
}
this->computeNkappaMatrixAt(gp, Nk);
this->computeBkappaMatrixAt(gp, Bk);
dpmat->givePDGradMatrix_uu(D, rMode, gp, tStep);
dpmat->givePDGradMatrix_ku(Dku, rMode, gp, tStep);
dpmat->givePDGradMatrix_uk(gPSigma, rMode, gp, tStep);
dpmat->givePDGradMatrix_kk(lStiff, rMode, gp, tStep);
/////////////////////////////////////////////////////////////////// uu:
DB.beProductOf(D, B);
if ( matStiffSymmFlag ) {
answer_uu.plusProductSymmUpper(B, DB, dV);
} else {
answer_uu.plusProductUnsym(B, DB, dV);
}
//////////////////////////////////////////////////////////////////////// ku:
DkuB.beProductOf(Dku, B);
answer_ku.plusProductUnsym(Nk, DkuB, -dV);
if ( dpmat->giveAveragingType() == 2 ) {
double dl1, dl2, dl3;
FloatMatrix LDB;
FloatMatrix GkLDB, MGkLDB;
FloatMatrix M22, M12;
FloatMatrix dL1(1, 3), dL2(1, 3), dLdS;
FloatArray Gk, n1, n2;
dpmat->givePDGradMatrix_LD(dLdS, rMode, gp, tStep);
this->computeNonlocalGradient(Gk, gp, tStep);
dl1 = dLdS.at(3, 3);
dl2 = dLdS.at(4, 4);
dl3 = dLdS.at(5, 5);
n1 = {dLdS.at(1, 1), dLdS.at(2, 1)};
n2 = {dLdS.at(1, 2), dLdS.at(2, 2)};
// first term Bk^T M22 G L1 D B
// M22 = n2 \otimes n2
M22.plusDyadUnsym(n2, n2, 1.);
// dL1
dL1.at(1, 1) = dl1 * n1.at(1) * n1.at(1) + dl2 * n2.at(1) * n2.at(1);
dL1.at(1, 2) = dl1 * n1.at(2) * n1.at(2) + dl2 * n2.at(2) * n2.at(2);
dL1.at(1, 3) = dl1 * n1.at(1) * n1.at(2) + dl2 * n2.at(1) * n2.at(2);
LDB.beProductOf(dL1, DB);
GkLDB.beProductOf(Gk, LDB);
MGkLDB.beProductOf(M22, GkLDB);
answer.plusProductUnsym(Bk, MGkLDB, dV);
// M12 + M21 = n1 \otimes n2 + n2 \otimes n1
M12.plusDyadUnsym(n1, n2, 1.);
M12.plusDyadUnsym(n2, n1, 1.);
//dL2
dL2.at(1, 1) = dl3 * ( n1.at(1) * n2.at(1) + n1.at(1) * n2.at(1) );
dL2.at(1, 2) = dl3 * ( n1.at(2) * n2.at(2) + n1.at(2) * n2.at(2) );
dL2.at(1, 3) = dl3 * ( n1.at(2) * n2.at(1) + n1.at(1) * n2.at(2) );
// Bk * ((M12 * L2 + M22 * L1) * DB)
LDB.beProductOf(dL2, DB);
GkLDB.beProductOf(Gk, LDB);
MGkLDB.beProductOf(M12, GkLDB);
answer.plusProductUnsym(Bk, MGkLDB, dV);
}
//////////////////////////////////////////////////////////////////////// uk:
SNk.beProductOf(gPSigma, Nk);
answer_uk.plusProductUnsym(B, SNk, -dV); // uk
/////////////////////////////////////////////////////////////////////// kk:
answer_kk.plusProductUnsym(Nk, Nk, dV);
if ( dpmat->giveAveragingType() == 0 || dpmat->giveAveragingType() == 1 ) {
double l = lStiff.at(1, 1);
answer_kk.plusProductUnsym(Bk, Bk, l * l * dV);
} else if ( dpmat->giveAveragingType() == 2 ) {
LBk.beProductOf(lStiff, Bk);
answer_kk.plusProductUnsym(Bk, LBk, dV);
}
}
if ( elem->domain->giveEngngModel()->giveFormulation() == AL ) {
FloatMatrix initialStressMatrix;
elem->computeInitialStressMatrix(initialStressMatrix, tStep);
answer_uu.add(initialStressMatrix);
}
if ( matStiffSymmFlag ) {
answer_uu.symmetrized();
}
answer.resize(totalSize, totalSize);
answer.zero();
answer.assemble(answer_uu, locU);
answer.assemble(answer_uk, locU, locK);
answer.assemble(answer_ku, locK, locU);
answer.assemble(answer_kk,locK);
}
/**
Purpose
-------
ZGESSM applies the factors L computed by ZGETRF_INCPIV to
a complex M-by-N tile A.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in]
k INTEGER
The number of columns of the matrix L. K >= 0.
@param[in]
ib INTEGER
The inner-blocking size. IB >= 0.
@param[in]
ipiv INTEGER array on the cpu.
The pivot indices array of size K as returned by
ZGETRF_INCPIV.
@param[in]
dL1 DOUBLE COMPLEX array, dimension(LDDL1, N)
The IB-by-K matrix in which is stored L^(-1) as returned by GETRF_INCPIV
@param[in]
lddl1 INTEGER
The leading dimension of the array L1. LDDL1 >= max(1,2*IB).
@param[in]
dL DOUBLE COMPLEX array, dimension(LDDL, N)
The M-by-K lower triangular tile on the gpu.
@param[in]
lddl INTEGER
The leading dimension of the array L. LDDL >= max(1,M).
@param[in,out]
dA DOUBLE COMPLEX array, dimension (LDDA, N)
On entry, the M-by-N tile A on the gpu.
On exit, updated by the application of L on the gpu.
@param[in]
ldda INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
@ingroup magma_zgesv_tile
********************************************************************/
extern "C" magma_int_t
magma_zgessm_gpu( magma_order_t order, magma_int_t m, magma_int_t n, magma_int_t k, magma_int_t ib,
magma_int_t *ipiv,
magmaDoubleComplex *dL1, magma_int_t lddl1,
magmaDoubleComplex *dL, magma_int_t lddl,
magmaDoubleComplex *dA, magma_int_t ldda,
magma_int_t *info)
{
#define AT(i,j) (dAT + (i)*ldda + (j) )
#define L(i,j) (dL + (i) + (j)*lddl )
#define dL1(j) (dL1 + (j)*lddl1)
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
int i, s, sb;
magmaDoubleComplex *dAT;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (ldda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
if ( order == MagmaColMajor ) {
magmablas_zgetmo_in( dA, dAT, ldda, m, n );
} else {
dAT = dA;
}
s = k / ib;
for (i = 0; i < k; i += ib) {
sb = min(ib, k-i);
magmablas_zlaswp( n, dAT, ldda, i+1, i+sb, ipiv, 1 );
#ifndef WITHOUTTRTRI
magma_ztrmm( MagmaRight, MagmaLower, MagmaTrans, MagmaUnit,
n, sb,
c_one, dL1(i), lddl1,
AT(i, 0), ldda);
#else
magma_ztrsm( MagmaRight, MagmaLower, MagmaTrans, MagmaUnit,
n, sb,
c_one, L( i, i), lddl,
AT(i, 0), ldda);
#endif
if ( (i+sb) < m) {
magma_zgemm( MagmaNoTrans, MagmaTrans,
n, m-(i+sb), sb,
c_neg_one, AT(i, 0), ldda,
L( i+sb, i), lddl,
c_one, AT(i+sb, 0), ldda );
}
}
if ( order == MagmaColMajor ) {
magmablas_zgetmo_in( dA, dAT, ldda, m, n );
}
return *info;
} /* magma_zgessm_gpu */
