The filling of the control vector with distributed mapping is very slow. The reason is that in the filling loop, the nbd variable, which is an array of size n, is completly overwrited at each iteration instead of assigning a single value. This bug has not affected simulations so far because the correct value is assigned, but in the case of distributed mapping over very large domains, the time taken to fill the control vector becomes excessive. O(n²) instead of O(n)
do i = 1, setup%nrrp
if (options%optimize%rr_parameters(i) .eq. 0) cycle
do col = 1, mesh%ncol
do row = 1, mesh%nrow
if (mesh%active_cell(row, col) .eq. 0) cycle
j = j + 1
parameters%control%x(j) = parameters%rr_parameters%values(row, col, i)
parameters%control%l(j) = options%optimize%l_rr_parameters(i)
parameters%control%u(j) = options%optimize%u_rr_parameters(i)
parameters%control%nbd = 2
write (name, '(a,i0,a,i0)') trim(parameters%rr_parameters%keys(i)), row, "-", col
parameters%control%name(j) = name
end do
end do
end do
The solution is to affect the value for the corresponding indice such that:
The filling of the control vector with distributed mapping is very slow. The reason is that in the filling loop, the
nbdvariable, which is an array of sizen, is completly overwrited at each iteration instead of assigning a single value. This bug has not affected simulations so far because the correct value is assigned, but in the case of distributed mapping over very large domains, the time taken to fill the control vector becomes excessive. O(n²) instead of O(n)The solution is to affect the value for the corresponding indice such that: