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

+61
View File
@@ -35,6 +35,8 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryGreaterThan, xb, xa); err != nil {
return err
}
_ = xa
_ = xb
return nil
}},
@@ -55,6 +57,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessThan, xc, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -75,6 +80,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryEqEq, xc, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -95,6 +103,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessThan, xc, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -115,6 +126,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessEq, xc, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -135,6 +149,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessThan, xc, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -147,6 +164,8 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessThan, xb, xa); err != nil {
return err
}
_ = xa
_ = xb
return nil
}},
@@ -159,6 +178,8 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryGreaterEq, xb, xa); err != nil {
return err
}
_ = xa
_ = xb
return nil
}},
@@ -179,6 +200,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessEq, xc, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -199,6 +223,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryEqEq, xc, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -219,6 +246,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessEq, xc, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -231,6 +261,8 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessEq, xb, xa); err != nil {
return err
}
_ = xa
_ = xb
return nil
}},
@@ -251,6 +283,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessEq, zeroExpr, xb); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
@@ -284,6 +319,29 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessEq, xc0, xc); err != nil {
return err
}
_ = xa
_ = xb
_ = xb0
_ = xc
_ = xc0
return nil
}},
{`"a <= (a + b): 0 <= b"`, func(q *checker, n *a.Assert) error {
op, xa, t0 := parseBinaryOp(n.Condition())
if op != t.IDXBinaryLessEq {
return errFailed
}
op, xa, xb := parseBinaryOp(t0)
if op != t.IDXBinaryPlus {
return errFailed
}
// 0 <= b
if err := proveReasonRequirement(q, t.IDXBinaryLessEq, zeroExpr, xb); err != nil {
return err
}
_ = xa
_ = xb
return nil
}},
@@ -301,6 +359,9 @@ var reasons = [...]struct {
if err := proveReasonRequirement(q, t.IDXBinaryLessEq, xa, t1); err != nil {
return err
}
_ = xa
_ = xb
_ = xc
return nil
}},
}
+14
View File
@@ -26,6 +26,7 @@ import (
"go/format"
"io/ioutil"
"os"
"sort"
"strings"
)
@@ -49,6 +50,8 @@ var reasons = [...]string{
"a < (b + c): a < c; 0 <= b",
"a < (b + c): a < (b0 + c0); b0 <= b; c0 <= c",
"a <= (a + b): 0 <= b",
"(a + b) <= c: a <= (c - b)",
}
@@ -109,6 +112,17 @@ func gen(reason string) error {
}
}
{
namesAsList := []string(nil)
for name := range names {
namesAsList = append(namesAsList, name)
}
sort.Strings(namesAsList)
for _, name := range namesAsList {
fmt.Fprintf(&out, "_ = x%s\n", name)
}
}
fmt.Fprintf(&out, "return nil }},\n")
return nil
}