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

+68 -20
View File
@@ -2569,10 +2569,10 @@ struct wuffs_lzw__decoder__struct {
uint32_t f_literal_width;
uint32_t f_flush_j;
uint8_t f_suffixes[4096];
uint8_t f_suffixes[4096][8];
uint16_t f_prefixes[4096];
uint16_t f_prefix_lengths[4096];
uint8_t f_output[8192];
uint8_t f_output[8199];
struct {
uint32_t coro_susp_point;
@@ -2586,8 +2586,12 @@ struct wuffs_lzw__decoder__struct {
uint32_t v_bits;
uint32_t v_n_bits;
uint32_t v_code;
uint16_t v_lm1_a;
uint32_t v_c;
uint32_t v_o;
uint32_t v_steps;
uint8_t v_first_byte;
uint16_t v_lm1_b;
} c_decode[1];
struct {
uint32_t coro_susp_point;
@@ -9507,8 +9511,12 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
uint32_t v_bits;
uint32_t v_n_bits;
uint32_t v_code;
uint16_t v_lm1_a;
uint32_t v_c;
uint32_t v_o;
uint32_t v_steps;
uint8_t v_first_byte;
uint16_t v_lm1_b;
uint8_t* iop_a_src = NULL;
uint8_t* io0_a_src = NULL;
@@ -9539,8 +9547,12 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
v_bits = self->private_impl.c_decode[0].v_bits;
v_n_bits = self->private_impl.c_decode[0].v_n_bits;
v_code = self->private_impl.c_decode[0].v_code;
v_lm1_a = self->private_impl.c_decode[0].v_lm1_a;
v_c = self->private_impl.c_decode[0].v_c;
v_o = self->private_impl.c_decode[0].v_o;
v_steps = self->private_impl.c_decode[0].v_steps;
v_first_byte = self->private_impl.c_decode[0].v_first_byte;
v_lm1_b = self->private_impl.c_decode[0].v_lm1_b;
} else {
}
switch (coro_susp_point) {
@@ -9553,11 +9565,12 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
v_clear_code = (((uint32_t)(1)) << v_literal_width);
v_end_code = (v_clear_code + 1);
v_save_code = v_end_code;
v_prev_code = 0;
v_prev_code = v_end_code;
v_width = (v_literal_width + 1);
v_j = 0;
while (v_j < v_clear_code) {
self->private_impl.f_prefix_lengths[v_j] = 0;
self->private_impl.f_suffixes[v_j][0] = ((uint8_t)(v_j));
v_j += 1;
}
v_j = 0;
@@ -9604,11 +9617,21 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
self->private_impl.f_output[v_j] = ((uint8_t)(v_code));
v_j = ((v_j + 1) & 8191);
if (v_save_code <= 4095) {
self->private_impl.f_suffixes[v_save_code] = ((uint8_t)(v_code));
self->private_impl.f_prefixes[v_save_code] =
((uint16_t)(v_prev_code));
self->private_impl.f_prefix_lengths[v_save_code] =
v_lm1_a =
((self->private_impl.f_prefix_lengths[v_prev_code] + 1) & 4095);
self->private_impl.f_prefix_lengths[v_save_code] = v_lm1_a;
if ((v_lm1_a % 8) != 0) {
self->private_impl.f_prefixes[v_save_code] =
self->private_impl.f_prefixes[v_prev_code];
memcpy(self->private_impl.f_suffixes[v_save_code],
self->private_impl.f_suffixes[v_prev_code], 8);
self->private_impl.f_suffixes[v_save_code][(v_lm1_a % 8)] =
((uint8_t)(v_code));
} else {
self->private_impl.f_prefixes[v_save_code] =
((uint16_t)(v_prev_code));
self->private_impl.f_suffixes[v_save_code][0] = ((uint8_t)(v_code));
}
v_save_code += 1;
if ((v_save_code == (((uint32_t)(1)) << v_width)) && (v_width < 12)) {
v_width += 1;
@@ -9620,32 +9643,53 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
goto label_1_break;
}
v_save_code = v_end_code;
v_prev_code = 0;
v_prev_code = v_end_code;
v_width = (v_literal_width + 1);
} else if (v_code <= v_save_code) {
v_c = v_code;
if (v_code == v_save_code) {
v_c = v_prev_code;
}
v_o = ((v_j + ((uint32_t)(self->private_impl.f_prefix_lengths[v_c]))) &
v_o = ((v_j + (((uint32_t)(self->private_impl.f_prefix_lengths[v_c])) &
4294967288)) &
8191);
v_j = ((v_o + 1) & 8191);
while (v_c >= v_clear_code) {
self->private_impl.f_output[v_o] = self->private_impl.f_suffixes[v_c];
v_o = ((v_o - 1) & 8191);
v_j = ((v_j + 1 +
((uint32_t)(self->private_impl.f_prefix_lengths[v_c]))) &
8191);
v_steps = (((uint32_t)(self->private_impl.f_prefix_lengths[v_c])) >> 3);
while (true) {
memcpy((self->private_impl.f_output) + (v_o),
(self->private_impl.f_suffixes[v_c]) + (0), 8);
if (v_steps <= 0) {
goto label_2_break;
}
v_steps -= 1;
v_o = ((v_o - 8) & 8191);
v_c = ((uint32_t)(self->private_impl.f_prefixes[v_c]));
}
self->private_impl.f_output[v_o] = ((uint8_t)(v_c));
label_2_break:;
v_first_byte = self->private_impl.f_suffixes[v_c][0];
if (v_code == v_save_code) {
self->private_impl.f_output[v_j] = ((uint8_t)(v_c));
self->private_impl.f_output[v_j] = v_first_byte;
v_j = ((v_j + 1) & 8191);
}
if (v_save_code <= 4095) {
self->private_impl.f_suffixes[v_save_code] = ((uint8_t)(v_c));
self->private_impl.f_prefixes[v_save_code] =
((uint16_t)(v_prev_code));
self->private_impl.f_prefix_lengths[v_save_code] =
v_lm1_b =
((self->private_impl.f_prefix_lengths[v_prev_code] + 1) & 4095);
self->private_impl.f_prefix_lengths[v_save_code] = v_lm1_b;
if ((v_lm1_b % 8) != 0) {
self->private_impl.f_prefixes[v_save_code] =
self->private_impl.f_prefixes[v_prev_code];
memcpy(self->private_impl.f_suffixes[v_save_code],
self->private_impl.f_suffixes[v_prev_code], 8);
self->private_impl.f_suffixes[v_save_code][(v_lm1_b % 8)] =
v_first_byte;
} else {
self->private_impl.f_prefixes[v_save_code] =
((uint16_t)(v_prev_code));
self->private_impl.f_suffixes[v_save_code][0] =
((uint8_t)(v_first_byte));
}
v_save_code += 1;
if ((v_save_code == (((uint32_t)(1)) << v_width)) && (v_width < 12)) {
v_width += 1;
@@ -9715,8 +9759,12 @@ suspend:
self->private_impl.c_decode[0].v_bits = v_bits;
self->private_impl.c_decode[0].v_n_bits = v_n_bits;
self->private_impl.c_decode[0].v_code = v_code;
self->private_impl.c_decode[0].v_lm1_a = v_lm1_a;
self->private_impl.c_decode[0].v_c = v_c;
self->private_impl.c_decode[0].v_o = v_o;
self->private_impl.c_decode[0].v_steps = v_steps;
self->private_impl.c_decode[0].v_first_byte = v_first_byte;
self->private_impl.c_decode[0].v_lm1_b = v_lm1_b;
goto exit;
exit:
@@ -9785,7 +9833,7 @@ wuffs_lzw__decoder__flush(wuffs_lzw__decoder* self,
v_s = wuffs_base__slice_u8__subslice_ij(
((wuffs_base__slice_u8){
.ptr = self->private_impl.f_output,
.len = 8192,
.len = 8199,
}),
v_i, v_j);
v_n = wuffs_base__io_writer__copy_from_slice(&iop_a_dst, io1_a_dst, v_s);