[hackers] [flate] todo update || nsz

changeset: 118:f9e371f5cd64
tag: tip
user: nsz <nszabolcs_AT_gmail.com>
date: Sun Aug 23 20:58:30 2009 +0200
files: TODO
description:
todo update

diff -r 3b5e4a1000de -r f9e371f5cd64 TODO
--- a/TODO Sun Aug 23 20:06:20 2009 +0200
+++ b/TODO Sun Aug 23 20:58:30 2009 +0200
@@ -3,21 +3,17 @@
 fill output entirely
 man
 globals
-error message ?
+error message
 inflate assumes Flate* < 0
 _init _reset _free ?
-_zlib _gzip _zip
-deflate, inflate example -> flate_example
-test, benchmark:
- empty block
- all zero (rle)
- uncompressed block
- test with small nin/nout (1,2..9 bytes)
+_zlib _gzip _zip interface
+test, benchmark
 
 
 inflate
 -------
-callback interface: reading past the end of compressed data: unreachable data
+-callback interface
+return extra input
 (rev lookup vs revinc)
 (test/optimize uncompressed block)
 read less than 7 bits in clen decode
@@ -26,20 +22,14 @@
 
 deflate
 -------
-space opt:
+block split heuristics with lfreq, dfreq + cost
+special case: last block (better optimization for short input)
+check ushort vs int
+lzbuf + costbuf
+minor mods:
         overlap rbuf/dstwin
         hufflen: overlap arrays (parent+heap only)
-time opt:
- match loop unroll
+ match loop + shiftwin loop unroll
         bisect len/distbase -> lookup table
         rolling hash
-better compression:
- block split heuristics with lfreq, dfreq (optimal block size?)
- (zlib huffcode trick: if same freq then shorter code goes to the one with smaller subtree)
- last block can be allowed to be larger
-code cleanups:
- better organization (configurable blocksize)
- bounds on the compressend size
- input from in+nin instead of src+srcend
- setting s->state..
- ushort vs int
+ zlib huffcode trick: if same freq then shorter code goes to the one with smaller subtree
Received on Sun Aug 23 2009 - 19:00:34 UTC