Temporary terms need to be computed per equation (as was done previously), and
not simply per block.
It’s necessary to track temporary terms per equation, because some equations
are evaluated instead of solved, and an equation E1 may depend on the value of
an endogenous Y computed by a previously evaluated equation E2; in this case,
if some temporary term TT of equation E2 contains Y, then TT needs to be
computed after E1, but before E2.
In particular, in dynamic models, temporary terms are now computed for
derivatives w.r.t. exogenous, and also w.r.t. endogenous variables that do not
belong to the block.
Also remove a message about elements below the cutoff that was no longer
correct (elements below the cutoff have no impact on the incidence matrix
outside of normalization).
By the way, this probably fixes a bug in the presence of external
functions (so-called TEF terms were not properly repeated in each per-block
static file).
Also remove debugging output in the M-file.
By the way, remove the BlockType stuff which was purely informative (and it’s
not worth carrying over prologue and epilogue information just for that).
Previously, the cutoff option would also impact the block decomposition itself,
since it would had an influence on the incidence matrix used for computing the
blocks and their derivatives.
The problem is that, in the general case, it’s quite possible that an element
of the numerical Jacobian be zero at the evaluation point, while being quite
different from zero along the simulation path. A typical example is an
expression of the type x*y, where y is an endogenous and x is an exogenous not
present in the initval block (and hence initialized to zero).
It has been superseded by ModelTree::blocks_derivatives.
By the way, fix the initial number of non-zero elements in sparse Jacobian.
Also avoid computing suboptimal temporary terms.
— use a std::map for storing block derivatives
— remove redundant ModelTree::first_chain_rule_derivatives structure
— remove unused codepaths in StaticModel
— DynamicModel: simplify code that determines the type of derivatives in a
block. We now use a slightly different categorization.
— by the way, fix the max lead/lag information for blocks that are obtained via
merging. A workaround was previously implemented in
DynamicModel::get_Derivative(), but it is no longer needed with this fix.
This was only adding unneeded complexity, for no clear reason (we’re very far
from reaching 2³¹ equations, and if we wanted to support models that large, it
would be better to use long integers to avoid being limited to 2³²).
— return output arguments on the left-hand side
— do not pass class members as input/output arguments
By the way, fix a (benign) vector allocation bug in
{Static,Dynamic}Model::computeChainRuleJacobian().
Backward incompatible change: the temporary terms for 3rd order are now stored
in "temporary_terms_third_derivative" (without the final "s"; same for external
functions), for consistency with the name of the slot for the derivatives
themselves ("third_derivative").
Ref dynare#217
Does not work for Julia mode, neither with block and bytecode.
Note: in DLL mode, the number of temporary terms in no longer given in the
‘ntt’ symbol; it must be computed using ‘M_.dynamic_tmp_nbr’.
Ref dynare#217
The preprocessor now writes all the symmetric elements in the "hp"
matrix (derivatives of the hessian w.r.t. parameters), for consistency with all
other derivatives output.
Previously it was only writing one of the two symmetric elements, when indices
of endogenous were different.
Also, no longer compute two times symmetric elements in derivation w.r.t.
parameters at order 2, for consistency with derivation w.r.t. endogenous.
It is therefore now necessary to duplicate them in the output to keep behavior
unchanged.
When creating the sparse matrix (in MATLAB or C mode), since storage is in
column-major order, output the first column, then the second, then the third.
This gives a significant performance boost in use_dll mode (at both compilation
and runtime), because it facilitates memory accesses and expression reusage.
New options "mexext" and "matlabroot" are introduced, so that the preprocessor
knows where to find MATLAB and which architecture to compile for.
Only recent gcc is now supported. A set of optimization flags is used so that
compilation goes reasonably fast on large models.
Consequently, options "msvc", "mingw" and "cygwin" have been removed.
In particular, it is necessary to turn back DataTree::AddVariable() into a
non-virtual method, because it is called from DataTree's constructor. Enforcing
the absence of leads/lags is now done using a new boolean DataTree::is_static.
Take advantage of the new copy constructor for handling
PlannerObjectiveStatement more elegantly.
Unfortunately it is not possible to implement *move* constructor / assigment
operators, because the reference ExprNode::datatree is not mutable.
This mimicks the structure of M-functions (though the logic for filling the
temporary terms vector is a bit different).
This change implied a modification in the way we compute the checksum in case
of block decomposition (the temporary terms for the C output are not correctly
computed in that case).
More precisely, remove those variants where temporary_terms can be specified
without temporary_terms_idxs, in order to make clear that the latter is
expected. For situations where the tt_idxs are not needed (C, block MATLAB), an
empty map has to be explicitly given.
In the presence of MLVs, the temporary terms indexing was corrupted. The code
was using the implicit assumption that the ExprNodeLess ordering was giving the
same ordering as the temporary terms indexes ordering. But MLVs can be higher
in ExprNodeLess ordering than some other temporary terms, while they have the
lowest temporary terms index, hence the bug.
Fix this by no longer relying on the ExprNodeLess ordering, and rather use a
full map<ExprNode *, int> for ModelTree::temporary_terms_idxs. By the way,
simplify the code by removing a few useless data structures (e.g.
ModelTree::temporary_terms_idxs_*).