GitLab

Also remove the det_cond_forecast interface, that was related to it.

Ref. dynare!1626
Closes: #39

Ref. dynare#1755

Incidentally, this makes it possible to use model-local variables with either
the “block” or “bytecode” option. (Ref: dynare#1243)

TODO: Check that we can do the same in the matlab version…

– Fix order of items in this structure. Previously, items were ordered
  according to the declaration order of parameters. Now, items are order
  according to lag order (first lag appears first)
– Gracefully handle the case where there is no autoregressive part
  (Closes: #52)

Ref. dynare#1726

Ref. dynare#1726

Closes: #53

This greatly facilitates the job of the compiler during the optimization pass,
since we promise that the various pointers do not overlap each other. It may
now be possible to reenable some of the optimization flags that were disabled
without sacrificing compilation time, but this needs more investigation.

For the gory details, see:
https://en.cppreference.com/w/c/language/restrict
https://cellperformance.beyond3d.com/articles/2006/05/demystifying-the-restrict-keyword.html

Sébastien Villemot authored Jul 24, 2020 and Stéphane Adjemian committed Jul 24, 2020

Also be more strict on the form of the target (must now be X(-1) or log(X(-1))
where X is *not* an auxiliary variable).

By the way, improve some comments in SymbolTable.

Sébastien Villemot authored Jul 23, 2020 and Stéphane Adjemian committed Jul 23, 2020

Closes: #54

This makes the output more useful when there are multiple PAC models.

Sébastien Villemot authored Jul 10, 2020 and Stéphane Adjemian committed Jul 13, 2020

The detection of the target EC variable to be used when constructing the
forward-looking expectation variable is rather fragile.

When the PAC model is written with an (non-)optimizing share of agents,
restrict the identification of the target variable to the optimizing
expression, to minimize the risk of wrong identification.

By the way, add a few comments, and a small simplification.

It is useless in the backward looking version, and it cannot be
evaluated in the MCE case (because of the forward looking recursion).

Sébastien Villemot authored Jul 08, 2020 and Stéphane Adjemian committed Jul 08, 2020

PAC MCE expectations: add corresponding auxiliary variables in dynamic_set_auxiliary_series.m and set_auxiliary_variables.m

By default, the preprocessor is supposed to only do the “unary ops”
transformation in the equations of VAR/PAC/trend component models.

However, the implementation was slightly different so far. It would detect
candidates to this transformation in the chosen equations, but it would then
perform the substitution in *all* equations.

This could lead for crashes, for example if the chosen equation contains
log(X(-1)), but another (non-chosen) equation has log(X(-2)). Then this latter
expression, even though it belongs to the same lag-equivalence class, is not
properly handled, causing a segfault.

Also do a few related cosmetic changes.

Previously they were returned as 3-column matrices. But this was inconsistent
with the M-file mode.

This is a preparatory step to allow use_dll with block decomposition.

Use 3-column format before calling sparse().

Ensure that the 3-column matrix is constructed in column-major order: data
locality will greatly improve performance once we implement use_dll.

This is much more explicit.

Block decomposition: for “solve two-boundaries” blocks, move the iteration loop outside the dynamic file

Block decomposition: for “solve two boundaries” block, no longer compute “b” (-residuals+g1*y) in the dynamic function