#ROBOMERGE-SOURCE: CL 15379156 in //UE5/Release-5.0-EarlyAccess/...
#ROBOMERGE-BOT: STARSHIP (Release-5.0-EarlyAccess -> Main) (v771-15082668)
[CL 15383280 by matt kuhlenschmidt in ue5-main branch]
#rb none
#jira FORT-256140, UE-89856
#ushell-cherrypick of 11894978 by Johan.Torp
#ROBOMERGE-SOURCE: CL 11894988 in //UE4/Release-4.25/... via CL 11894989
#ROBOMERGE-BOT: RELEASE (Release-4.25Plus -> Main) (v656-11643781)
[CL 11894990 by johan torp in Main branch]
#rnx
#rb none
#ROBOMERGE-SOURCE: CL 10869240 via CL 10869516 via CL 10869902
#ROBOMERGE-BOT: (v613-10869866)
[CL 10870584 by ryan durand in Main branch]
We've had several cases where the previously default alphabetical sorting caused performance issues when it wasn't needed. We now require you to explicitly choose either slow alphabetical sorting or fast non-alphabetical order.
The fast order does not decode, access or compare actual string data, but the order is not stable between process runs.
#rb steve.robb
[CL 5915265 by Johan Torp in Dev-Core branch]
Encoding improvements
* New hash table implementation
--- Move away from 16-bit hashes since we need more than 64k buckets to hold 2M entries efficiently
--- Change to CityHash64, which is faster and stronger
--- Remove hardcoded max limit
* NAME_INDEX changed from contigouos int to monotonically increasing int
--- Opens up for future deduplication schemes that are better than number suffix dedup.
--- Saves some memory since we don't need to maintain a contiguous array
--- Typed up to cause compile errors when used as integer directly
* Avoid touching data repeatedly, normal path only does single hash of data
* New constructor that allows supplying string length up front
* Avoid dynamic allocations in string conversions >= 128 characters
* Avoid extra copying when splitting numbers
* More efficient IsPureAnsi check
* Only do one global lazy initialization call instead of multiple
* Switch to faster RW locks
* Switch from single lock to sharded hash map with separate locks
* Memory optimizations that reduces per entry overhead: 4 bytes slots, 2 byte headers and don't store null terminator
Improved API & documentation
* Document that IsValid() rarely makes sense
* Hide global state such as GetNames()
* Reduce amount of implementation details visible in header
* NameTypes.h size down by ~1/3 while adding documentation, stronger type safety and new APIs
Future possibilities
* Memory savings: Removing public NAME_INDEX and global FName array allows using the 32-bit FNameEntryId for arbitrary deduplication schemes. This can save both actual stored strings memory by deduplication and half the size of FName instances from 8B to 4B in shipping / test configs by removing the number part.
* Implementation can be tweaked further, for instance could persist 32-bit slot index hash inside slot to increase encoding performance in development / debug at the cost of memory.
Perf & mem results for internal project:
* Editor startup wall time: 12% speedup, 22.0s -> 19.5s
--- Warm disk and asset registry cache
--- Some gains from nametable serialization improvements
--- Some of these gains might be from orthogonal asset discovery optimizations
* Win64 Test Client memory usage with ~400k names: 30.3Mb -> 22.2Mb
#rb steve.robb, pj.kack
#jira UE-59973
[CL 5774657 by Johan Torp in Dev-Core branch]
