* Updates filler item names to match the actual item names
* Adds more descriptive error message in case this error comes back
* Properly raises exception instead of just text
* Replaces exception with assert
* TLOZ: Fix non-deterministic item pool generation
The way the item pool was constructed involved iterating unions of sets.
Sets are unordered, so the order of iteration of these combined sets
would be non-deterministic, resulting in the items in the item pool
being generated in a different order with the same seed.
Rather than creating unions of sets at all, the original code has been
replaced with using Counter objects. As a dict subclass, Counter
maintains insertion order, and its update() method makes it simple to
combine the separate item dictionaries into a single dictionary with the
total count of each item across each of the separate item dictionaries.
Fixes#3664 - After investigating more deeply, the only differences I
could find between generations of the same seed was the order of items
created by TLOZ, so this patch appears to fix the non-deterministic
generation issue. I did manage to reproduce the non-deterministic
behaviour with just TLOZ in the end, but it was very rare. I'm not
entirely sure why generating with SMZ3 specifically would cause the
non-deterministic behaviour in TLOZ to be frequently present, whereas
generating with other games or multiple TLOZ yamls would not.
* Change import order
---------
Co-authored-by: NewSoupVi <57900059+NewSoupVi@users.noreply.github.com>
* Core: dump all item placements for generation failures
* pass the multiworld from remaining fill
* change how the args get handled to fix formatting
---------
Co-authored-by: NewSoupVi <57900059+NewSoupVi@users.noreply.github.com>
* makes start inventory from pool warn and fixes the itempool to match when it can not find a matching item to remove
* calc the difference correctly
* save new filler and non-removed items differently so we don't remove existing items at random
Fixes randomly placed ingredients not being deterministic (depending on settings)
and in turn also fixes logic not being deterministic if they get replaced by fragments.
All the types being copied are built-in types with their own `copy()`
methods, so using the `copy` module was a bit overkill and also slower.
This patch replaces the use of the `copy` module in
`CollectionState.copy()` with using the built-in `.copy()` methods.
The copying of `reachable_regions` and `blocked_connections` was also
iterating the keys of each dictionary and then looking up the value in
the dictionary for that key. It is faster, and I think more readable, to
iterate the dictionary's `.items()` instead.
For me, when generating a multiworld including the template yaml of
every world with `python -O .\Generate.py --skip_output`, this patch
saves about 2.1s. The overall generation duration for these yamls varies
quite a lot, but averages around 160s for me, so on average this patch
reduced overall generation duration (excluding output duration) by
around 1.3%.
Timing comparisons were made by calling time.perf_counter() at the start
and end of `CollectionState.copy()`'s body, and summing the differences
between the starts and ends of the method body into a global variable
that was printed at the end of generation.
Additional timing comparisons were made, using the `timeit` module, of
the individual function calls or dictionary comprehensions used to
perform the copying.
The main performance cost was `copy.deepcopy()`, which gets slow as the
number of keys multiplied by the number of values within the
sets/Counters gets large, e.g., to deepcopy a `dict[int, Counter[str]]`
with 100 keys and where each Counter contains 100 keys was 30x slower
than most other tested copying methods. Increasing the number of dict
keys or Counter keys only makes it slower.
