Tweak std/bzip2's "inverse of T" description

This commit is contained in:
Nigel Tao committed 2022-05-17 00:10:46 +10:00
1 parent 84ef91dda8
commit 2422997ee5
2 files changed
+22 -25

No files matched your search

+19 -22
View File
@@ -241,13 +241,13 @@ value (relative to each cumulative count) to all the 'a' elements in *L*
(assigning 0, 1 and 2), all the 'b' elements (assigning 3), all the 'c' (assigning 0, 1 and 2), all the 'b' elements (assigning 3), all the 'c'
elements (assinging 4) and all the 'r' elements (assigning 5). elements (assinging 4) and all the 'r' elements (assigning 5).
Index L T Index L T
0 c 4 0 c 4
1 a 0 1 a 0
2 r 5 2 r 5
3 a 1 3 a 1
4 a 2 4 a 2
5 b 3 5 b 3
Starting with the *I* value given in the (g) bitslice, which was 1, calculate Starting with the *I* value given in the (g) bitslice, which was 1, calculate
*I*, *T[I]*, *T[T[I]]*, *T[T[T[I]]]*, etc to give [1, 0, 4, 2, 5, 3]. Indexing *I*, *T[I]*, *T[T[I]]*, *T[T[T[I]]]*, etc to give [1, 0, 4, 2, 5, 3]. Indexing
@@ -258,23 +258,20 @@ uncompressed text.
## Alternative BWT Inversion ## Alternative BWT Inversion
It's not in the BWT technical report, but instead of walking *I*, *T[I]*, It's not in the BWT technical report, but instead of walking *I*, *T[I]*,
*T[T[I]]*, etc, an alternative but equivalent algorithm starts with the *Index *T[T[I]]*, etc, an alternative but equivalent algorithm works on *U*, the
L T* table but then sorts the rows by *T*, modifying the *Index* and *T* inverse of the *T* vector. For example *T[4] == 2* and so *U[2] == 4*.
columns (to produce *Index2* and *T2*) but leaving the *L* column unchanged.
For example *Index == 4* matches *T == 2*, so *T2 == 2* matches *Index2 == 4*.
Index2 L T2 Index L T F U
1 c 0 0 c 4 a 1
3 a 1 1 a 0 a 3
4 r 2 2 r 5 a 4
5 a 3 3 a 1 b 5
0 a 4 4 a 2 c 0
2 b 5 5 b 3 r 2
Our original *I* value was 1, and when *T2* is 1, *Index2* is 3. Call this Calculate *U[I]*, *U[U[I]]*, *U[U[U[I]]]*, etc to give [3, 5, 2, 4, 0, 1].
value *I2*. Calculate *I2*, *Index2[I2]*, *Index2[Index2[I2]]*, etc to give [3, Indexing *L* with these values gives "abraca", the uncompressed text (without
5, 2, 4, 0, 1]. Indexing *L* with these values gives "abraca", the uncompressed being reversed).
text (without being reversed).
## Final Run Length Encoding ## Final Run Length Encoding
+3 -3
View File
@@ -67,7 +67,7 @@ pub struct decoder? implements base.io_transformer(
// bwt is the Burrows Wheeler Transform state. Per the README.md file, each // bwt is the Burrows Wheeler Transform state. Per the README.md file, each
// entry is a row with the low 8 bits holding the L column and the high 20 // entry is a row with the low 8 bits holding the L column and the high 20
// bits holding the Index2 column. The middle 4 bits are unused. // bits holding the U column. The middle 4 bits are unused.
// //
// The read_code_lengths and build_huffman_tree methods also re-purpose // The read_code_lengths and build_huffman_tree methods also re-purpose
// this buffer to temporarily hold up to 258 symbols' code lengths. // this buffer to temporarily hold up to 258 symbols' code lengths.
@@ -727,8 +727,8 @@ pri func decoder.invert_bwt!() {
i += 1 i += 1
} endwhile } endwhile
// Second data pass, but per the README.md file, calculate the Index2 // Second data pass, but per the README.md file, calculate the U column
// column instead of the T column calculated by SRC-RR-124.pdf. // instead of the BWT technical report's T column.
i = 0 i = 0
while i < this.block_size { while i < this.block_size {
assert i < 900000 via "a < b: a < c; c <= b"(c: this.block_size) assert i < 900000 via "a < b: a < c; c <= b"(c: this.block_size)