Fix import performance regression and update time limits

Unfortunately, #3030 doubled the import
time for our performance test model, but the check there was too lax to catch it.

Sometimes, `smart_subs_dict` is applied to individual expressions,
sometimes to complete matrices. If it's only applied to individual expressions
or small matrices, avoid the overhead of flattening the substitutions first.

Update the time limits for the performance tests to catch such regressions earlier.

Fixes #3048.

Author: dweindl

python/sdist/amici/importers/utils.py

@@ -426,27 +426,54 @@ def smart_subs_dict(
     else:
         s = [(eid, expr[field]) for eid, expr in subs.items()]
 
-    if not reverse:
-        # counter-intuitive, but we need to reverse the order for reverse=False
-        s.reverse()
-
-    with sp.evaluate(False):
-        # The new expressions may themselves contain symbols to be substituted.
-        #  We flatten them out first, so that the substitutions in `sym` can be
-        #  performed simultaneously, which is usually more efficient than
-        #  repeatedly substituting into `sym`.
-        # TODO(performance): This could probably be made more efficient by
-        #   combining with toposort used to order `subs` in the first place.
-        #   Some substitutions could be combined, and some terms not present in
-        #   `sym` could be skipped.
-        for i in range(len(s) - 1):
-            for j in range(i + 1, len(s)):
-                if s[j][1].has(s[i][0]):
-                    s[j] = s[j][0], s[j][1].xreplace({s[i][0]: s[i][1]})
-
-        s = dict(s)
-        sym = sym.xreplace(s)
-    return sym
+    # We have the choice to flatten the replacement expressions first or to
+    #  substitute them one by one into `sym`. Flattening first is usually
+    #  more efficient if `sym` is large (e.g., a matrix with many elements)
+    #  and `subs` is cascading (i.e., substitutions depend on other
+    #  substitutions). Otherwise, substituting one by one is usually more
+    #  efficient, because flattening scales quadratically with the number of
+    #  substitutions.
+    #  The exact threshold is somewhat arbitrary and may need to be
+    #  adjusted in the future.
+    flatten_first = isinstance(sym, sp.MatrixBase) and sym.rows * sym.cols > 20
+
+    if flatten_first:
+        if not reverse:
+            # counter-intuitive, but on this branch, we need to reverse the
+            #  order for `reverse=False`
+            s.reverse()
+
+        with sp.evaluate(False):
+            # The new expressions may themselves contain symbols to be
+            #  substituted. We flatten them out first, so that the
+            #  substitutions in `sym` can be performed simultaneously,
+            #  which can be more efficient than repeatedly substituting into
+            #  `sym`.
+            # TODO(performance): This could probably be made more efficient by
+            #   combining with toposort used to order `subs` in the first
+            #   place.
+            #   Some substitutions could be combined, and some terms not
+            #   present in `sym` could be skipped.
+            for i in range(len(s) - 1):
+                for j in range(i + 1, len(s)):
+                    if s[j][1].has(s[i][0]):
+                        s[j] = s[j][0], s[j][1].xreplace({s[i][0]: s[i][1]})
+
+            s = dict(s)
+            sym = sym.xreplace(s)
+        return sym
+
+    else:
+        if reverse:
+            s.reverse()
+
+        with sp.evaluate(False):
+            for old, new in s:
+                # note that substitution may change free symbols,
+                #  so we have to do this recursively
+                if sym.has(old):
+                    sym = sym.xreplace({old: new})
+        return sym
 
 
 def smart_subs(element: sp.Expr, old: sp.Symbol, new: sp.Expr) -> sp.Expr:

tests/performance/reference.yml

@@ -1,11 +1,11 @@
 # Reference wall times (seconds) with some buffer
 create_sdist: 16
 install_sdist: 150
-petab_import: 2100
-install_model: 120
+petab_import: 720
+install_model: 60
 install_model_O0: 40
-install_model_O1: 90
-install_model_O2: 120
+install_model_O1: 45
+install_model_O2: 60
 forward_simulation: 2
 forward_sensitivities: 2
 adjoint_sensitivities: 2.5