daiduo/TurboPFor-Integer-Compression
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

TurboPFor: Floating point

Browse Source
This commit is contained in:
x
2023-03-18 21:18:51 +01:00
parent d8f4385c31
commit 1280058904
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
fp.c
View File

@ -157,7 +157,7 @@ size_t T2(p4nzzdec128v,USIZE)(unsigned char *in, size_t n, uint_t *out, uint_t s
/*---------------- TurboFloat XOR: last value Predictor with TurboPFor ---------------------------------------------------------
Compress significantly (115% - 160%) better and faster than facebook's Gorilla algorithm for values
BEST results are obtained with LOSSY COMPRESSION (using fppad32/fppad64 in bitutil.c)
BEST results are obtained with LOSSY COMPRESSION (using fprazor32/fprazor64 in bitutil.c)
Algorithm:
1: XOR value with previous value. We have now leading (for common sign/exponent bits) + mantissa trailing zero bits
2: Eliminate the common block leading zeros of sign/exponent by shifting all values in the block to left