Write LZW output 8 bytes at a time, not 1

The benchmarks are mixed, but I think it's still a net win.

----
name                                    old speed     new speed     delta

wuffs_gif_decode_1k_bw/clang            258MB/s ± 0%  253MB/s ± 1%    -2.04%  (p=0.008 n=5+5)
wuffs_gif_decode_1k_color/clang         159MB/s ± 0%  116MB/s ± 0%   -26.76%  (p=0.008 n=5+5)
wuffs_gif_decode_10k/clang              167MB/s ± 0%  138MB/s ± 0%   -17.42%  (p=0.008 n=5+5)
wuffs_gif_decode_100k_artificial/clang  205MB/s ± 0%  439MB/s ± 0%  +114.53%  (p=0.008 n=5+5)
wuffs_gif_decode_100k_realistic/clang   151MB/s ± 0%  164MB/s ± 0%    +8.41%  (p=0.008 n=5+5)
wuffs_gif_decode_1000k/clang            155MB/s ± 0%  165MB/s ± 0%    +6.50%  (p=0.008 n=5+5)
wuffs_gif_decode_anim_screencap/clang   255MB/s ± 0%  921MB/s ± 0%  +261.50%  (p=0.008 n=5+5)

wuffs_gif_decode_1k_bw/gcc              316MB/s ± 2%  267MB/s ± 1%   -15.78%  (p=0.008 n=5+5)
wuffs_gif_decode_1k_color/gcc           184MB/s ± 1%  134MB/s ± 0%   -27.57%  (p=0.008 n=5+5)
wuffs_gif_decode_10k/gcc                196MB/s ± 0%  161MB/s ± 0%   -17.50%  (p=0.008 n=5+5)
wuffs_gif_decode_100k_artificial/gcc    287MB/s ± 0%  449MB/s ± 1%   +56.22%  (p=0.008 n=5+5)
wuffs_gif_decode_100k_realistic/gcc     182MB/s ± 0%  182MB/s ± 1%      ~     (p=0.690 n=5+5)
wuffs_gif_decode_1000k/gcc              187MB/s ± 1%  185MB/s ± 1%      ~     (p=0.056 n=5+5)
wuffs_gif_decode_anim_screencap/gcc     397MB/s ± 0%  973MB/s ± 0%  +145.42%  (p=0.008 n=5+5)

wuffs_lzw_decode_20k/clang              170MB/s ± 0%  191MB/s ± 1%   +12.52%  (p=0.008 n=5+5)
wuffs_lzw_decode_100k/clang             242MB/s ± 0%  312MB/s ± 2%   +29.00%  (p=0.008 n=5+5)

wuffs_lzw_decode_20k/gcc                220MB/s ± 0%  227MB/s ± 0%    +3.31%  (p=0.008 n=5+5)
wuffs_lzw_decode_100k/gcc               317MB/s ± 0%  369MB/s ± 1%   +16.53%  (p=0.008 n=5+5)
----

LZW is the compression method used by GIF images. Roughly speaking, LZW
is a sequence of codes, and each code emits either 1 byte (a literal
code) or multiple bytes (a copy code). The speed gain in this commit
comes from, when emitting an N-byte copy code, this now requires N/8
iterations instead of N iterations, as each iteration now writes 8 bytes
instead of 1. Literal codes, both before and after this commit, requires
a single iteration.

Decoding should therefore get faster or slower depending on whether it's
copy-code-heavy or literal-code-heavy respectively. For what it's worth,
copy-code-heavy GIF images also usually compress better.

Running script/extract-giflzw.go (or examining the cached output at
test/data/analysis/giflzw-emission-lengths.txt) on the anim_screencap
(which got faster) and 1k_color (which got slower) cases (also known as
gifplayer-muybridge.gif and hippopotamus.regular.gif) shows that the
proportion of literal codes agrees with that explanation.

----
gifplayer-muybridge.gif (all frames)
 Percent   Count  Emission_Length_Bucket
   12.78   34658           1 (Literal Code)
   14.05   38106           2
   13.25   35937           3
   10.59   28707           4
    7.85   21286           5
    6.08   16481           6
    4.85   13156           7
    3.60    9764           8
    9.34   25318    9 ..  16
    4.89   13257   17 ..  32
    4.78   12954   33 ..  64
    4.72   12806   65 .. 128
    3.02    8190  127 .. 256
    0.20     550  256+

hippopotamus.regular.gif (all frames)
 Percent   Count  Emission_Length_Bucket
   82.66     672           1 (Literal Code)
   13.04     106           2
    2.83      23           3
    0.86       7           4
    0.37       3           5
    0.25       2           6
    0.00       0           7
    0.00       0           8
    0.00       0    9 ..  16
    0.00       0   17 ..  32
    0.00       0   33 ..  64
    0.00       0   65 .. 128
    0.00       0  127 .. 256
    0.00       0  256+
----

The new algorithm also adds more CPU and RAM overhead, regardless of
whether the input GIF image's LZW data is copy-code-heavy or
literal-code-heavy. For example, the suffix table grows by 8x, from 4KiB
to 32 KiB.

The gifplayer-muybridge.gif image is also larger (472 x 298) and the
hippopotamus.regular.gif image is also smaller (36 x 28). Higher fixed
costs may have larger (negative) relative effects on decoding smaller
images.

To repeat, I think it's still a net win.

An upcoming commit will update the std/lzw/README.md file to provide
more detail on the algorithm.

Thanks to Tushar Khandelwal for the idea.
This commit is contained in:
Nigel Tao committed 2018-10-21 14:48:40 +11:00
1 parent 8d7bd36179
commit 3338d6d1bc
7 files changed
+292 -67

No files matched your search

+49 -23
View File
@@ -19,10 +19,10 @@ pri status "?internal error: inconsistent I/O"
pub struct decoder?(
literal_width base.u32[..8],
flush_j base.u32[..8191],
suffixes array[4096] base.u8,
suffixes array[4096] array[8] base.u8,
prefixes array[4096] base.u16[..4095],
prefix_lengths array[4096] base.u16[..4095],
output array[8192] base.u8,
prefix_lengths array[4096] base.u16[..4095], // TODO: rename this field.
output array[8192 + 7] base.u8,
)
pub func decoder.set_literal_width!(lw base.u32[2..8]) {
@@ -44,7 +44,7 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
// next back-reference expands to. The README.md file also calls this value
// `max`. 4096 means do not save.
var save_code base.u32[..4096] = end_code
var prev_code base.u32[..4095]
var prev_code base.u32[..4095] = end_code
var width base.u32[..12] = literal_width + 1
// First, use the j variable to iterate over prefix_lengths.
@@ -52,6 +52,7 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
while j < clear_code {
assert j < 256 via "a < b: a < c; c <= b"(c:clear_code)
this.prefix_lengths[j] = 0
this.suffixes[j][0] = j as base.u8
j += 1
}
// For the rest of this function, j is such that output[:j] accumulates the
@@ -109,9 +110,18 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
this.output[j] = code as base.u8
j = (j + 1) & 8191
if save_code <= 4095 {
this.suffixes[save_code] = code as base.u8
this.prefixes[save_code] = prev_code as base.u16
this.prefix_lengths[save_code] = (this.prefix_lengths[prev_code] + 1) & 4095
var lm1_a base.u16[..4095] = (this.prefix_lengths[prev_code] + 1) & 4095
this.prefix_lengths[save_code] = lm1_a
if (lm1_a % 8) != 0 {
this.prefixes[save_code] = this.prefixes[prev_code]
this.suffixes[save_code] = this.suffixes[prev_code]
this.suffixes[save_code][lm1_a % 8] = code as base.u8
} else {
this.prefixes[save_code] = prev_code as base.u16
this.suffixes[save_code][0] = code as base.u8
}
save_code += 1
if (save_code == ((1 as base.u32) << width)) and (width < 12) {
width += 1
@@ -124,8 +134,9 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
break
}
save_code = end_code
prev_code = 0
prev_code = end_code
width = literal_width + 1
} else if code <= save_code {
var c base.u32[..4095] = code
if code == save_code {
@@ -134,35 +145,50 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
// Letting old_j and new_j denote the values of j before and after
// these two lines of code, the decoded bytes will be written to
// output[old_j:new_j]. They will be written back-to-front,
// starting by writing output[o], also known as output[new_j - 1].
// output[old_j:new_j]. They will be written back-to-front, 8 bytes
// at a time, starting by writing output[o:o + 8], which will
// contain output[new_j - 1].
//
// In the special case that code == save_code, the decoded bytes
// contain an extra copy (at the end) of the first byte, and will
// be written to output[old_j:new_j + 1].
var o base.u32[..8191] = (j + (this.prefix_lengths[c] as base.u32)) & 8191
j = (o + 1) & 8191
var o base.u32[..8191] = (j + ((this.prefix_lengths[c] as base.u32) & 0xFFFFFFF8)) & 8191
j = (j + 1 + (this.prefix_lengths[c] as base.u32)) & 8191
while c >= clear_code,
post c < 256 via "a < b: a < c; c <= b"(c:clear_code),
{
this.output[o] = this.suffixes[c]
// This line is essentially "o -= 1". The "& 8191" is a no-op
var steps base.u32 = (this.prefix_lengths[c] as base.u32) >> 3
while true {
assert o <= (o + 8) via "a <= (a + b): 0 <= b"(b:8)
this.output[o:o + 8].copy_from_slice!(s:this.suffixes[c][0:0 + 8])
if steps <= 0 {
break
}
steps -= 1
// This line is essentially "o -= 8". The "& 8191" is a no-op
// in practice, but is necessary for the overflow checker.
o = (o ~mod- 1) & 8191
o = (o ~mod- 8) & 8191
c = this.prefixes[c] as base.u32
}
this.output[o] = c as base.u8
var first_byte base.u8 = this.suffixes[c][0]
if code == save_code {
this.output[j] = c as base.u8
this.output[j] = first_byte
j = (j + 1) & 8191
}
if save_code <= 4095 {
this.suffixes[save_code] = c as base.u8
this.prefixes[save_code] = prev_code as base.u16
this.prefix_lengths[save_code] = (this.prefix_lengths[prev_code] + 1) & 4095
var lm1_b base.u16[..4095] = (this.prefix_lengths[prev_code] + 1) & 4095
this.prefix_lengths[save_code] = lm1_b
if (lm1_b % 8) != 0 {
this.prefixes[save_code] = this.prefixes[prev_code]
this.suffixes[save_code] = this.suffixes[prev_code]
this.suffixes[save_code][lm1_b % 8] = first_byte
} else {
this.prefixes[save_code] = prev_code as base.u16
this.suffixes[save_code][0] = first_byte as base.u8
}
save_code += 1
if (save_code == ((1 as base.u32) << width)) and (width < 12) {
width += 1