Can unroll the loop, but would look more ugly :)
JIT will probably do it for us.
Fix <= to < for IVs
Inline temp var (start is unused after assigning to low)
Big thanks to Parzival from the RNG discord community for chiseling the LCRNG search space down to the best performing implementation possible.
50x? Down from O(2^8) -> O(2^3) is 32x, but we no longer need to access two heap arrays (262KB no longer needed!). Everything can be calculated tightly with the stack.
f641f3eab2/RNG/LCG_Reversal.py (L31)
Rainbow tables is the only faster implementation. However, nobody is gonna hog many GB of RAM for O(1) reversals. This is ~O(2^3), down from O(2^8). Much better than the days of O(2^16)!
In this pull request I've changed a ton of method signatures to reflect the more-narrow types of Species, Move# and Form; additionally, I've narrowed other large collections that stored lists of species / permitted values, and reworked them to be more performant with the latest API spaghetti that PKHeX provides. Roamer met locations, usually in a range of [max-min]<64, can be quickly checked using a bitflag operation on a UInt64. Other collections (like "Is this from Colosseum or XD") were eliminated -- shadow state is not transferred COLO<->XD, so having a Shadow ID or matching the met location from a gift/wild encounter is a sufficient check for "originated in XD".
`Moveset` struct stores 4 moves, and exposes methods to interact with a moveset.
`IndividualValueSet` stores a 6 IV template (signed).
Performance impact:
* Less allocating on the heap: Moves - (8 bytes member ptr, 20 bytes heap->8 bytes member)
* Less allocating on the heap: IVs - (8 bytes member ptr, 28 bytes heap->8 bytes member)
* No heap pointers, no need to jump to grab data.
* Easy to inline logic for checking if moves are present (no linq usage with temporary collections).
End result is faster ctor times, less memory used, faster program.
Rewrites a good amount of legality APIs pertaining to:
* Legal moves that can be learned
* Evolution chains & cross-generation paths
* Memory validation with forgotten moves
In generation 8, there are 3 separate contexts an entity can exist in: SW/SH, BD/SP, and LA. Not every entity can cross between them, and not every entity from generation 7 can exist in generation 8 (Gogoat, etc). By creating class models representing the restrictions to cross each boundary, we are able to better track and validate data.
The old implementation of validating moves was greedy: it would iterate for all generations and evolutions, and build a full list of every move that can be learned, storing it on the heap. Now, we check one game group at a time to see if the entity can learn a move that hasn't yet been validated. End result is an algorithm that requires 0 allocation, and a smaller/quicker search space.
The old implementation of storing move parses was inefficient; for each move that was parsed, a new object is created and adjusted depending on the parse. Now, move parse results are `struct` and store the move parse contiguously in memory. End result is faster parsing and 0 memory allocation.
* `PersonalTable` objects have been improved with new API methods to check if a species+form can exist in the game.
* `IEncounterTemplate` objects have been improved to indicate the `EntityContext` they originate in (similar to `Generation`).
* Some APIs have been extended to accept `Span<T>` instead of Array/IEnumerable
was checking stale value
make loop max adjustable by caller; knowingly requesting squares is 1:65,536, so a higher loop count than 50k might guarantee more successes.
Maybe in the future we'd have separate algorithms to pre-choose seeds by choosing a PID and unrolling -> rolling.
Co-Authored-By: Kermalis <29823718+Kermalis@users.noreply.github.com>
ShowdownSet: Lessen allocation
MoveTutor: Remove boxing by calling the generic method instead of object method
Xoro8b: Add more xmldoc, use positive constant instead of inverse negative for parity
StadiumUtil: Use built-in endianness reversal methods
Many years ago, PKX used to be a >4,000 line bloated file, which spun off multiple classes like CommonEdits and most of the early non-GUI PKM related logic. Now, it's just a stub to source the latest generation & personal table.
Separate files = more concise info, and more room to grow to do more advanced things.
Makes the IsPresent methods public (no longer internal).
Single underscore discards (one of the c# language revisions allowed reusing the single underscore discard).
Remove a temporary allocation in BDSP flag editor
Instantiating from template now follows group seed -> spawn 1 correlation, including alpha move.
Differentiates static encounters that don't follow the ow8a correlation, scrambles EC to disassociate.
Adds rand64 to get initial seeds
Set correct level range to match slotSeed; not respecting the slot roll being valid, but whatever.
Adds structures to read/write saved spawner data such as seeds, counts.
Adds generator and validator to emulate the FixInitSpec builder used by the game logic
Similar to SW/SH raids, validating these in-process is not feasible due to the number crunching required.
This does not handle the encounter slot call or the follow-up level range call. Just the inner FixInitSpec ctor & fill.
level is calc'd:
randFloat(sum) -> slot float
rand.Next() -> gen_seed (for all the details)
rand.NextInt(delta) +min -> level
Co-Authored-By: Lusamine <30205550+Lusamine@users.noreply.github.com>
Existing `get`/`set` logic is flawed in that it doesn't work on Big Endian operating systems, and it allocates heap objects when it doesn't need to.
`System.Buffers.Binary.BinaryPrimitives` in the `System.Memory` NuGet package provides both Little Endian and Big Endian methods to read and write data; all the `get`/`set` operations have been reworked to use this new API. This removes the need for PKHeX's manual `BigEndian` class, as all functions are already covered by the BinaryPrimitives API.
The `StringConverter` has now been rewritten to accept a Span to read from & write to, no longer requiring a temporary StringBuilder.
Other Fixes included:
- The Super Training UI for Gen6 has been reworked according to the latest block structure additions.
- Cloning a Stadium2 Save File now works correctly (opening from the Folder browser list).
- Checksum & Sanity properties removed from parent PKM class, and is now implemented via interface.
