Fill the LZW bit buffer less often

name                             old speed     new speed     delta

wuffs_gif_decode_1k_bw/clang     292MB/s ± 1%  256MB/s ± 1%  -12.29%  (p=0.008 n=5+5)
wuffs_gif_decode_1k_color/clang  143MB/s ± 0%  159MB/s ± 0%  +10.95%  (p=0.008 n=5+5)
wuffs_gif_decode_10k/clang       150MB/s ± 1%  165MB/s ± 0%  +10.24%  (p=0.008 n=5+5)
wuffs_gif_decode_100k/clang      152MB/s ± 0%  150MB/s ± 0%   -1.19%  (p=0.016 n=4+5)
wuffs_gif_decode_1000k/clang     156MB/s ± 1%  153MB/s ± 1%   -1.60%  (p=0.008 n=5+5)

wuffs_gif_decode_1k_bw/gcc       250MB/s ± 1%  330MB/s ± 0%  +31.99%  (p=0.008 n=5+5)
wuffs_gif_decode_1k_color/gcc    172MB/s ± 1%  184MB/s ± 1%   +7.40%  (p=0.008 n=5+5)
wuffs_gif_decode_10k/gcc         189MB/s ± 0%  193MB/s ± 0%   +2.31%  (p=0.008 n=5+5)
wuffs_gif_decode_100k/gcc        166MB/s ± 0%  180MB/s ± 0%   +8.56%  (p=0.008 n=5+5)
wuffs_gif_decode_1000k/gcc       171MB/s ± 0%  186MB/s ± 0%   +8.19%  (p=0.008 n=5+5)

wuffs_lzw_decode_20k/clang       168MB/s ± 3%  168MB/s ± 0%     ~     (p=0.222 n=5+5)
wuffs_lzw_decode_100k/clang      252MB/s ± 1%  239MB/s ± 0%   -4.89%  (p=0.008 n=5+5)

wuffs_lzw_decode_20k/gcc         209MB/s ± 0%  217MB/s ± 0%   +4.28%  (p=0.008 n=5+5)
wuffs_lzw_decode_100k/gcc        307MB/s ± 0%  314MB/s ± 0%   +2.27%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2018-10-20 07:22:23 +11:00
1 parent 24e7ee0104
commit 6945bdc75c
2 files changed
+62 -48

No files matched your search

+32 -25
View File
@@ -9564,30 +9564,37 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
v_bits = 0;
v_n_bits = 0;
while (true) {
if (((uint64_t)(io1_a_src - iop_a_src)) >= 4) {
v_bits |= (((uint32_t)(wuffs_base__load_u32le(iop_a_src))) << v_n_bits);
(iop_a_src += ((31 - v_n_bits) >> 3),
wuffs_base__return_empty_struct());
v_n_bits |= 24;
} else {
while (v_n_bits < v_width) {
while (((uint64_t)(io1_a_src - iop_a_src)) <= 0) {
if (v_j > 0) {
self->private_impl.f_flush_j = v_j;
WUFFS_BASE__COROUTINE_SUSPENSION_POINT(1);
status = wuffs_lzw__decoder__flush(self, a_dst);
if (status) {
goto suspend;
}
v_j = 0;
}
status = wuffs_base__suspension__short_read;
WUFFS_BASE__COROUTINE_SUSPENSION_POINT_MAYBE_SUSPEND(2);
}
if (v_n_bits < v_width) {
if (((uint64_t)(io1_a_src - iop_a_src)) >= 4) {
v_bits |=
(((uint32_t)(wuffs_base__load_u8be(iop_a_src))) << v_n_bits);
(iop_a_src += 1, wuffs_base__return_empty_struct());
v_n_bits += 8;
(((uint32_t)(wuffs_base__load_u32le(iop_a_src))) << v_n_bits);
(iop_a_src += ((31 - v_n_bits) >> 3),
wuffs_base__return_empty_struct());
v_n_bits |= 24;
} else {
while (true) {
while (((uint64_t)(io1_a_src - iop_a_src)) <= 0) {
if (v_j > 0) {
self->private_impl.f_flush_j = v_j;
WUFFS_BASE__COROUTINE_SUSPENSION_POINT(1);
status = wuffs_lzw__decoder__flush(self, a_dst);
if (status) {
goto suspend;
}
v_j = 0;
}
status = wuffs_base__suspension__short_read;
WUFFS_BASE__COROUTINE_SUSPENSION_POINT_MAYBE_SUSPEND(2);
}
v_bits |=
(((uint32_t)(wuffs_base__load_u8be(iop_a_src))) << v_n_bits);
(iop_a_src += 1, wuffs_base__return_empty_struct());
v_n_bits += 8;
if (v_n_bits >= v_width) {
goto label_0_break;
}
}
label_0_break:;
}
}
v_code = ((v_bits) & ((1 << (v_width)) - 1));
@@ -9610,7 +9617,7 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
}
} else if (v_code <= v_end_code) {
if (v_code == v_end_code) {
goto label_0_break;
goto label_1_break;
}
v_save_code = v_end_code;
v_prev_code = 0;
@@ -9669,7 +9676,7 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
v_j = 0;
}
}
label_0_break:;
label_1_break:;
if (v_j > 0) {
while (v_n_bits >= 8) {
v_n_bits -= 8;
+30 -23
View File
@@ -63,33 +63,40 @@ pub func decoder.decode!??(dst base.io_writer, src base.io_reader) {
var n_bits base.u32[..31]
while true {
if args.src.available() >= 4 {
// Read 4 bytes, using the "Variant 4" technique of
// https://fgiesen.wordpress.com/2018/02/20/reading-bits-in-far-too-many-ways-part-2/
bits |= (args.src.peek_u32le() as base.u32) ~mod<< n_bits
args.src.skip_fast!(actual:(31 - n_bits) >> 3, worst_case:3)
n_bits |= 24
assert width <= n_bits via "a <= b: a <= c; c <= b"(c:12)
assert n_bits >= width via "a >= b: b <= a"()
} else {
while n_bits < width,
post n_bits >= width,
{
if n_bits < width {
if args.src.available() >= 4 {
// Read 4 bytes, using the "Variant 4" technique of
// https://fgiesen.wordpress.com/2018/02/20/reading-bits-in-far-too-many-ways-part-2/
bits |= (args.src.peek_u32le() as base.u32) ~mod<< n_bits
args.src.skip_fast!(actual:(31 - n_bits) >> 3, worst_case:3)
n_bits |= 24
assert width <= n_bits via "a <= b: a <= c; c <= b"(c:12)
assert n_bits >= width via "a >= b: b <= a"()
} else {
assert n_bits < 12 via "a < b: a < c; c <= b"(c:width)
while args.src.available() <= 0,
inv n_bits < 12,
post args.src.available() >= 1,
while true,
pre n_bits < 12,
post n_bits >= width,
{
if j > 0 {
this.flush_j = j
this.flush!??(dst:args.dst)
j = 0
while args.src.available() <= 0,
inv n_bits < 12,
post args.src.available() >= 1,
{
if j > 0 {
this.flush_j = j
this.flush!??(dst:args.dst)
j = 0
}
yield status "$short read"
}
yield status "$short read"
bits |= (args.src.peek_u8() as base.u32) << n_bits
args.src.skip_fast!(actual:1, worst_case:1)
n_bits += 8
if n_bits >= width {
break
}
assert n_bits < 12 via "a < b: a < c; c <= b"(c:width)
}
bits |= (args.src.peek_u8() as base.u32) << n_bits
args.src.skip_fast!(actual:1, worst_case:1)
n_bits += 8
}
}