sharpetronics/zstd
1
0
Fork 0
mirror of https://github.com/facebook/zstd.git synced 2025-10-04 00:02:33 -04:00
Code Issues Projects Releases Wiki Activity

Updates ZSTD_RowFindBestMatch comment (#3947)

Browse Source 
Updates the comment on the head of `ZSTD_RowFindBestMatch` to make sure it's aligned with recent changes to the hash table.
This commit is contained in:
Yonatan Komornik 2024-03-12 15:10:07 -07:00 committed by GitHub
parent f65b9e27ce
commit b20703f273
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
lib/compress/zstd_lazy.c
View File

@ -1123,18 +1123,18 @@ ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 headGr
/* The high-level approach of the SIMD row based match finder is as follows:
* - Figure out where to insert the new entry:
* - Generate a hash from a byte along with an additional 1-byte "short hash". The additional byte is our "tag"
* - The hashTable is effectively split into groups or "rows" of 16 or 32 entries of U32, and the hash determines
* - Generate a hash for current input posistion and split it into a one byte of tag and `rowHashLog` bits of index.
* - The hash is salted by a value that changes on every contex reset, so when the same table is used
* we will avoid collisions that would otherwise slow us down by intorducing phantom matches.
* - The hashTable is effectively split into groups or "rows" of 15 or 31 entries of U32, and the index determines
* which row to insert into.
* - Determine the correct position within the row to insert the entry into. Each row of 16 or 32 can
* be considered as a circular buffer with a "head" index that resides in the tagTable.
* - Also insert the "tag" into the equivalent row and position in the tagTable.
* - Note: The tagTable has 17 or 33 1-byte entries per row, due to 16 or 32 tags, and 1 "head" entry.
* The 17 or 33 entry rows are spaced out to occur every 32 or 64 bytes, respectively,
* for alignment/performance reasons, leaving some bytes unused.
* - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte "short hash" and
* - Determine the correct position within the row to insert the entry into. Each row of 15 or 31 can
* be considered as a circular buffer with a "head" index that resides in the tagTable (overall 16 or 32 bytes
* per row).
* - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte tag calculated for the position and
* generate a bitfield that we can cycle through to check the collisions in the hash table.
* - Pick the longest match.
* - Insert the tag into the equivalent row and position in the tagTable.
*/
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR