jscaux
/
ABACUS
2
0
Atdalīts 0
Kods Problēmas 1 Izmaiņu pieprasījumi Laidieni Vikivietne Aktivitāte

Accelerate Newton's method for LiebLin by moving allocations up

This commit is contained in:
J.-S. Caux 2018-09-07 17:35:20 +02:00
vecāks 565216ffdd
revīzija b5f3b0755c
2 mainīti faili ar 9 papildinājumiem un 7 dzēšanām
Rādīt statistiku Lejupielādēt ielāpa failu Lejupielādēt izmaiņu failu Izvērst visus failus Savērst visus failus

2
include/ABACUS_LiebLin.h
Parādīt failu

@ -82,7 +82,7 @@ namespace ABACUS {
void Set_Free_lambdaocs();
void Iterate_BAE(DP damping);
void Iterate_BAE_S(DP damping);
void Iterate_BAE_Newton(DP damping);
void Iterate_BAE_Newton (DP damping, Vect_DP& RHSBAE, Vect_DP& dlambdaoc, SQMat_DP& Gaudin, Vect_INT& indx);
void Compute_Energy ();
void Compute_Momentum ();
DP Kernel (int a, int b);

14
src/LIEBLIN/LiebLin_Bethe_State.cc
Parādīt failu

@ -275,8 +275,14 @@ namespace ABACUS {
DP damping = 1.0;
DP diffsq_prev = 1.0e+6;
Vect_DP RHSBAE (0.0, N); // contains RHS of BAEs
Vect_DP dlambdaoc (0.0, N); // contains delta lambdaoc computed from Newton's method
SQMat_DP Gaudin (0.0, N);
Vect_INT indx (N);
while (diffsq > prec && !is_nan(diffsq) && iter_Newton < 100) {
(*this).Iterate_BAE_Newton(damping);
(*this).Iterate_BAE_Newton(damping, RHSBAE, dlambdaoc, Gaudin, indx);
if (diffsq > diffsq_prev && damping > 0.5) damping /= 2.0;
else if (diffsq < diffsq_prev) damping = 1.0;
diffsq_prev = diffsq;
@ -469,12 +475,8 @@ namespace ABACUS {
/**
Performs one step of the matrix Newton method on the rapidities.
*/
void LiebLin_Bethe_State::Iterate_BAE_Newton (DP damping)
void LiebLin_Bethe_State::Iterate_BAE_Newton (DP damping, Vect_DP& RHSBAE, Vect_DP& dlambdaoc, SQMat_DP& Gaudin, Vect_INT& indx)
{
Vect_DP RHSBAE (0.0, N); // contains RHS of BAEs
Vect_DP dlambdaoc (0.0, N); // contains delta lambdaoc computed from Newton's method
SQMat_DP Gaudin (0.0, N);
Vect_INT indx (N);
DP sumtheta = 0.0;
// for large |lambda|, use atan (lambda) = sgn(lambda) pi/2 - atan(1/lambda)
int atanintshift = 0;