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

+59 -1
View File
@@ -22,7 +22,10 @@ import (
t "github.com/google/wuffs/lang/token"
)
var errNoSuchBuiltin = errors.New("cgen: internal error: no such built-in")
var (
errNoSuchBuiltin = errors.New("cgen: internal error: no such built-in")
errOptimizationNotApplicable = errors.New("cgen: internal error: optimization not applicable")
)
func (g *gen) writeBuiltinCall(b *buffer, n *a.Expr, rp replacementPolicy, depth uint32) error {
if n.Operator() != t.IDOpenParen {
@@ -323,6 +326,10 @@ func (g *gen) writeBuiltinNumType(b *buffer, recv *a.Expr, method t.ID, args []*
func (g *gen) writeBuiltinSlice(b *buffer, recv *a.Expr, method t.ID, args []*a.Node, rp replacementPolicy, depth uint32) error {
switch method {
case t.IDCopyFromSlice:
if err := g.writeBuiltinSliceCopyFromSlice8(b, recv, method, args, rp, depth); err != errOptimizationNotApplicable {
return err
}
// TODO: don't assume that the slice is a slice of base.u8.
b.writes("wuffs_base__slice_u8__copy_from_slice(")
if err := g.writeExpr(b, recv, rp, depth); err != nil {
@@ -351,6 +358,57 @@ func (g *gen) writeBuiltinSlice(b *buffer, recv *a.Expr, method t.ID, args []*a.
return errNoSuchBuiltin
}
// writeBuiltinSliceCopyFromSlice8 writes an optimized version of:
//
// foo[fIndex:fIndex + 8].copy_from_slice!(s:bar[bIndex:bIndex + 8])
func (g *gen) writeBuiltinSliceCopyFromSlice8(b *buffer, recv *a.Expr, method t.ID, args []*a.Node, rp replacementPolicy, depth uint32) error {
if method != t.IDCopyFromSlice || len(args) != 1 {
return errOptimizationNotApplicable
}
foo, fIndex := matchFooIndexIndexPlus8(recv)
bar, bIndex := matchFooIndexIndexPlus8(args[0].AsArg().Value())
if foo == nil || bar == nil {
return errOptimizationNotApplicable
}
b.writes("memcpy((")
if err := g.writeExpr(b, foo, rp, depth); err != nil {
return err
}
b.writes(")+(")
if err := g.writeExpr(b, fIndex, rp, depth); err != nil {
return err
}
b.writes("),(")
if err := g.writeExpr(b, bar, rp, depth); err != nil {
return err
}
b.writes(")+(")
if err := g.writeExpr(b, bIndex, rp, depth); err != nil {
return err
}
// TODO: don't assume that the slice is a slice of base.u8.
b.writes("), 8)")
return nil
}
// matchFooIndexIndexPlus8 matches n with "foo[index:index + 8]". It returns
// nil values if there isn't a match.
func matchFooIndexIndexPlus8(n *a.Expr) (foo *a.Expr, index *a.Expr) {
if n.Operator() != t.IDColon {
return nil, nil
}
foo = n.LHS().AsExpr()
index = n.MHS().AsExpr()
rhs := n.RHS().AsExpr()
if index == nil || rhs == nil || rhs.Operator() != t.IDXBinaryPlus || !rhs.LHS().AsExpr().Eq(index) {
return nil, nil
}
if cv := rhs.RHS().AsExpr().ConstValue(); cv == nil || cv.Cmp(eight) != 0 {
return nil, nil
}
return foo, index
}
func (g *gen) writeBuiltinTable(b *buffer, recv *a.Expr, method t.ID, args []*a.Node, rp replacementPolicy, depth uint32) error {
field := ""
+5 -2
View File
@@ -40,8 +40,11 @@ import (
)
var (
zero = big.NewInt(0)
one = big.NewInt(1)
zero = big.NewInt(0)
one = big.NewInt(1)
eight = big.NewInt(8)
mibi = big.NewInt(1 << 20)
)
// Prefixes are prepended to names to form a namespace and to avoid e.g.
+36 -21
View File
@@ -98,26 +98,43 @@ func (g *gen) writeStatementAssign(b *buffer, n *a.Assign, depth uint32) error {
if err := g.writeSuspendibles(b, n.RHS(), depth); err != nil {
return err
}
opName, tilde := "", false
opName, closer := "", ""
op := n.Operator()
switch op {
case t.IDTildeSatPlusEq, t.IDTildeSatMinusEq:
uBits := uintBits(n.LHS().MType().QID())
if uBits == 0 {
return fmt.Errorf("unsupported tilde-operator type %q", n.LHS().MType().Str(g.tm))
}
uOp := "add"
if op != t.IDTildeSatPlusEq {
uOp = "sub"
}
b.printf("wuffs_base__u%d__sat_%s_indirect(&", uBits, uOp)
opName, tilde = ",", true
if lTyp := n.LHS().MType(); lTyp.IsNumType() || lTyp.IsBool() || lTyp.IsSliceType() {
op := n.Operator()
switch op {
case t.IDTildeSatPlusEq, t.IDTildeSatMinusEq:
uBits := uintBits(n.LHS().MType().QID())
if uBits == 0 {
return fmt.Errorf("unsupported tilde-operator type %q", lTyp.Str(g.tm))
}
uOp := "add"
if op != t.IDTildeSatPlusEq {
uOp = "sub"
}
b.printf("wuffs_base__u%d__sat_%s_indirect(&", uBits, uOp)
opName, closer = ",", ")"
default:
opName = cOpName(op)
if opName == "" {
return fmt.Errorf("unrecognized operator %q", op.AmbiguousForm().Str(g.tm))
default:
opName = cOpName(op)
if opName == "" {
return fmt.Errorf("unrecognized operator %q", op.AmbiguousForm().Str(g.tm))
}
}
} else {
if lTyp.IsArrayType() {
if nElem := lTyp.ArrayLength().ConstValue(); nElem.Sign() > 0 && nElem.Cmp(mibi) <= 0 {
if elemTyp := lTyp.Inner(); elemTyp.IsNumType() {
if nBits := uintBits(elemTyp.QID()); nBits != 0 {
b.writes("memcpy(")
opName, closer = ",", fmt.Sprintf(", %d)", nElem.Int64()*int64(nBits/8))
}
}
}
}
if closer == "" {
return fmt.Errorf("unsupported assignment type %q", lTyp.Str(g.tm))
}
}
@@ -128,9 +145,7 @@ func (g *gen) writeStatementAssign(b *buffer, n *a.Assign, depth uint32) error {
if err := g.writeExpr(b, n.RHS(), replaceCallSuspendibles, depth); err != nil {
return err
}
if tilde {
b.writeb(')')
}
b.writes(closer)
b.writes(";\n")
return nil
}