Write LZW decoder output in batches

name                             old speed     new speed     delta

wuffs_gif_decode_1k_bw/clang     309MB/s ± 0%  306MB/s ± 0%   -0.93%  (p=0.008 n=5+5)
wuffs_gif_decode_1k_color/clang  160MB/s ± 0%  156MB/s ± 0%   -2.26%  (p=0.008 n=5+5)
wuffs_gif_decode_10k/clang       145MB/s ± 0%  153MB/s ± 0%   +5.21%  (p=0.008 n=5+5)
wuffs_gif_decode_100k/clang      131MB/s ± 1%  147MB/s ± 0%  +12.15%  (p=0.008 n=5+5)
wuffs_gif_decode_1000k/clang     134MB/s ± 0%  151MB/s ± 1%  +12.83%  (p=0.008 n=5+5)

wuffs_gif_decode_1k_bw/gcc       339MB/s ± 0%  319MB/s ± 0%   -5.99%  (p=0.008 n=5+5)
wuffs_gif_decode_1k_color/gcc    186MB/s ± 0%  165MB/s ± 0%  -10.93%  (p=0.008 n=5+5)
wuffs_gif_decode_10k/gcc         164MB/s ± 0%  171MB/s ± 1%   +4.19%  (p=0.008 n=5+5)
wuffs_gif_decode_100k/gcc        143MB/s ± 0%  161MB/s ± 1%  +12.32%  (p=0.008 n=5+5)
wuffs_gif_decode_1000k/gcc       146MB/s ± 0%  166MB/s ± 0%  +13.61%  (p=0.008 n=5+5)

wuffs_lzw_decode_20k/clang       142MB/s ± 0%  167MB/s ± 0%  +17.51%  (p=0.008 n=5+5)
wuffs_lzw_decode_100k/clang      186MB/s ± 1%  239MB/s ± 0%  +28.68%  (p=0.008 n=5+5)

wuffs_lzw_decode_20k/gcc         162MB/s ± 0%  192MB/s ± 0%  +18.63%  (p=0.008 n=5+5)
wuffs_lzw_decode_100k/gcc        235MB/s ± 1%  295MB/s ± 0%  +25.59%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2018-10-14 17:53:20 +11:00
1 parent 9573b88a02
commit 4e27b1280e
2 files changed
+224 -92

No files matched your search

+72 -26
View File
@@ -14,11 +14,15 @@
pub status "?bad code"
pri status "?internal error: inconsistent i:j"
pub struct decoder?(
literal_width base.u32[..8],
stack array[4096] base.u8,
flush_j base.u32[..8191],
suffixes array[4096] base.u8,
prefixes array[4096] base.u16[..4095],
prefix_lengths array[4096] base.u16[..4095],
output array[8192] base.u8,
)
pub func decoder.set_literal_width!(lw base.u32[2..8]) {
@@ -43,7 +47,18 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
var prev_code base.u32[..4095]
var width base.u32[..12] = literal_width + 1
// These variables yield src's bits in Least Significant Bits order.
// First, use the j variable to iterate over prefix_lengths.
var j base.u32[..8191]
while j < clear_code {
assert j < 256 via "a < b: a < c; c <= b"(c:clear_code)
this.prefix_lengths[j] = 0
j += 1
}
// For the rest of this function, j is such that output[:j] accumulates the
// decoded output to write (in batches) to args.dst.
j = 0
// These variables produce src's bits in Least Significant Bits order.
var bits base.u32
var n_bits base.u32
@@ -65,10 +80,12 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
if code < clear_code {
assert code < 256 via "a < b: a < c; c <= b"(c:clear_code)
args.dst.write_u8!??(x:code as base.u8)
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
save_code += 1
if (save_code == ((1 as base.u32) << width)) and (width < 12) {
width += 1
@@ -78,53 +95,50 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
} else if code <= end_code {
if code == end_code {
return
break
}
save_code = end_code
prev_code = 0
width = literal_width + 1
} else if code <= save_code {
var s base.u32[..4095] = 4095
var c base.u32[..4095] = code
if code == save_code {
s -= 1
c = prev_code
}
// 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].
//
// 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
while c >= clear_code,
inv n_bits < 8,
post c < 256 via "a < b: a < c; c <= b"(c:clear_code),
{
this.stack[s] = this.suffixes[c]
// The bitwise-and with 4095 is algorithmically a no-op, but is
// necessary for the overflow checker.
s = (s ~mod- 1) & 4095
this.output[o] = this.suffixes[c]
// This line is essentially "o -= 1". The "& 8191" is a no-op
// in practice, but is necessary for the overflow checker.
o = (o ~mod- 1) & 8191
c = this.prefixes[c] as base.u32
}
this.stack[s] = c as base.u8
this.output[o] = c as base.u8
if code == save_code {
this.stack[4095] = c as base.u8
}
while true,
inv n_bits < 8,
inv c < 256,
{
var expansion slice base.u8 = this.stack[s:]
var n_copied base.u64 = args.dst.copy_from_slice!(s:expansion)
if n_copied == expansion.length() {
break
}
s = (s + ((n_copied & 4095) as base.u32)) & 4095
yield status "$short write"
this.output[j] = c as base.u8
j = (j + 1) & 8191
}
if save_code <= 4095 {
this.suffixes[save_code] = c as base.u8
this.prefixes[save_code] = prev_code as base.u16
this.prefix_lengths[save_code] = (this.prefix_lengths[prev_code] + 1) & 4095
save_code += 1
if (save_code == ((1 as base.u32) << width)) and (width < 12) {
width += 1
@@ -135,5 +149,37 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
} else {
return status "?bad code"
}
// Flush the output if it could be too full to contain the entire
// decoding of the next code. The longest possible decoding is slightly
// less than 4096 and output's length is 8192, so a conservative
// threshold is ensuring that j <= 4095.
if j > 4095 {
this.flush_j = j
this.flush!??(dst:args.dst)
j = 0
}
}
if j > 0 {
this.flush_j = j
this.flush!??(dst:args.dst)
}
}
pri func decoder.flush!??(dst base.io_writer) {
var i base.u32[..8191]
var j base.u32[..8191] = this.flush_j
while true {
if i > j {
return status "?internal error: inconsistent i:j"
}
var s slice base.u8 = this.output[i:j]
var n base.u64 = args.dst.copy_from_slice!(s:s)
if n == s.length() {
return
}
i = (i + ((n & 8191) as base.u32)) & 8191
yield status "$short write"
}
}