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

+152 -66
View File
@@ -2568,9 +2568,11 @@ struct wuffs_lzw__decoder__struct {
uint32_t magic;
uint32_t f_literal_width;
uint8_t f_stack[4096];
uint32_t f_flush_j;
uint8_t f_suffixes[4096];
uint16_t f_prefixes[4096];
uint16_t f_prefix_lengths[4096];
uint8_t f_output[8192];
struct {
uint32_t coro_susp_point;
@@ -2580,13 +2582,19 @@ struct wuffs_lzw__decoder__struct {
uint32_t v_save_code;
uint32_t v_prev_code;
uint32_t v_width;
uint32_t v_j;
uint32_t v_bits;
uint32_t v_n_bits;
uint32_t v_code;
uint32_t v_s;
uint32_t v_c;
uint64_t v_n_copied;
uint32_t v_o;
} c_decode[1];
struct {
uint32_t coro_susp_point;
uint32_t v_i;
uint32_t v_j;
uint64_t v_n;
} c_flush[1];
} private_impl;
#ifdef __cplusplus
@@ -9412,6 +9420,8 @@ exit:
// ---------------- Status Codes Implementations
const char* wuffs_lzw__error__bad_code = "?lzw: bad code";
const char* wuffs_lzw__error__internal_error_inconsistent_i_j =
"?lzw: internal error: inconsistent i:j";
// ---------------- Private Consts
@@ -9419,6 +9429,10 @@ const char* wuffs_lzw__error__bad_code = "?lzw: bad code";
// ---------------- Private Function Prototypes
static wuffs_base__status //
wuffs_lzw__decoder__flush(wuffs_lzw__decoder* self,
wuffs_base__io_writer a_dst);
// ---------------- Initializer Implementations
wuffs_base__status WUFFS_BASE__WARN_UNUSED_RESULT //
@@ -9489,33 +9503,13 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
uint32_t v_save_code;
uint32_t v_prev_code;
uint32_t v_width;
uint32_t v_j;
uint32_t v_bits;
uint32_t v_n_bits;
uint32_t v_code;
uint32_t v_s;
uint32_t v_c;
wuffs_base__slice_u8 v_expansion;
uint64_t v_n_copied;
uint32_t v_o;
uint8_t* iop_a_dst = NULL;
uint8_t* io0_a_dst = NULL;
uint8_t* io1_a_dst = NULL;
WUFFS_BASE__IGNORE_POTENTIALLY_UNUSED_VARIABLE(io0_a_dst);
WUFFS_BASE__IGNORE_POTENTIALLY_UNUSED_VARIABLE(io1_a_dst);
if (a_dst.private_impl.buf) {
iop_a_dst =
a_dst.private_impl.buf->data.ptr + a_dst.private_impl.buf->meta.wi;
if (!a_dst.private_impl.mark) {
a_dst.private_impl.mark = iop_a_dst;
a_dst.private_impl.limit =
a_dst.private_impl.buf->data.ptr + a_dst.private_impl.buf->data.len;
}
if (a_dst.private_impl.buf->meta.closed) {
a_dst.private_impl.limit = iop_a_dst;
}
io0_a_dst = a_dst.private_impl.mark;
io1_a_dst = a_dst.private_impl.limit;
}
uint8_t* iop_a_src = NULL;
uint8_t* io0_a_src = NULL;
uint8_t* io1_a_src = NULL;
@@ -9541,15 +9535,13 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
v_save_code = self->private_impl.c_decode[0].v_save_code;
v_prev_code = self->private_impl.c_decode[0].v_prev_code;
v_width = self->private_impl.c_decode[0].v_width;
v_j = self->private_impl.c_decode[0].v_j;
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_s = self->private_impl.c_decode[0].v_s;
v_c = self->private_impl.c_decode[0].v_c;
v_expansion = ((wuffs_base__slice_u8){});
v_n_copied = self->private_impl.c_decode[0].v_n_copied;
v_o = self->private_impl.c_decode[0].v_o;
} else {
v_expansion = ((wuffs_base__slice_u8){});
}
switch (coro_susp_point) {
WUFFS_BASE__COROUTINE_SUSPENSION_POINT_0;
@@ -9563,6 +9555,12 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
v_save_code = v_end_code;
v_prev_code = 0;
v_width = (v_literal_width + 1);
v_j = 0;
while (v_j < v_clear_code) {
self->private_impl.f_prefix_lengths[v_j] = 0;
v_j += 1;
}
v_j = 0;
v_bits = 0;
v_n_bits = 0;
while (true) {
@@ -9582,16 +9580,14 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
v_bits >>= v_width;
v_n_bits -= v_width;
if (v_code < v_clear_code) {
WUFFS_BASE__COROUTINE_SUSPENSION_POINT(2);
if (iop_a_dst == io1_a_dst) {
status = wuffs_base__suspension__short_write;
goto suspend;
}
*iop_a_dst++ = ((uint8_t)(v_code));
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] =
((self->private_impl.f_prefix_lengths[v_prev_code] + 1) & 4095);
v_save_code += 1;
if ((v_save_code == (((uint32_t)(1)) << v_width)) && (v_width < 12)) {
v_width += 1;
@@ -9600,49 +9596,35 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
}
} else if (v_code <= v_end_code) {
if (v_code == v_end_code) {
status = NULL;
goto ok;
goto label_0_break;
}
v_save_code = v_end_code;
v_prev_code = 0;
v_width = (v_literal_width + 1);
} else if (v_code <= v_save_code) {
v_s = 4095;
v_c = v_code;
if (v_code == v_save_code) {
v_s -= 1;
v_c = v_prev_code;
}
v_o = ((v_j + ((uint32_t)(self->private_impl.f_prefix_lengths[v_c]))) &
8191);
v_j = ((v_o + 1) & 8191);
while (v_c >= v_clear_code) {
self->private_impl.f_stack[v_s] = self->private_impl.f_suffixes[v_c];
v_s = ((v_s - 1) & 4095);
self->private_impl.f_output[v_o] = self->private_impl.f_suffixes[v_c];
v_o = ((v_o - 1) & 8191);
v_c = ((uint32_t)(self->private_impl.f_prefixes[v_c]));
}
self->private_impl.f_stack[v_s] = ((uint8_t)(v_c));
self->private_impl.f_output[v_o] = ((uint8_t)(v_c));
if (v_code == v_save_code) {
self->private_impl.f_stack[4095] = ((uint8_t)(v_c));
self->private_impl.f_output[v_j] = ((uint8_t)(v_c));
v_j = ((v_j + 1) & 8191);
}
while (true) {
v_expansion = wuffs_base__slice_u8__subslice_i(
((wuffs_base__slice_u8){
.ptr = self->private_impl.f_stack,
.len = 4096,
}),
v_s);
v_n_copied = wuffs_base__io_writer__copy_from_slice(
&iop_a_dst, io1_a_dst, v_expansion);
if (v_n_copied == ((uint64_t)(v_expansion.len))) {
goto label_0_break;
}
v_s = ((v_s + ((uint32_t)((v_n_copied & 4095)))) & 4095);
status = wuffs_base__suspension__short_write;
WUFFS_BASE__COROUTINE_SUSPENSION_POINT_MAYBE_SUSPEND(3);
}
label_0_break:;
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] =
((self->private_impl.f_prefix_lengths[v_prev_code] + 1) & 4095);
v_save_code += 1;
if ((v_save_code == (((uint32_t)(1)) << v_width)) && (v_width < 12)) {
v_width += 1;
@@ -9653,6 +9635,24 @@ wuffs_lzw__decoder__decode(wuffs_lzw__decoder* self,
status = wuffs_lzw__error__bad_code;
goto exit;
}
if (v_j > 4095) {
self->private_impl.f_flush_j = v_j;
WUFFS_BASE__COROUTINE_SUSPENSION_POINT(2);
status = wuffs_lzw__decoder__flush(self, a_dst);
if (status) {
goto suspend;
}
v_j = 0;
}
}
label_0_break:;
if (v_j > 0) {
self->private_impl.f_flush_j = v_j;
WUFFS_BASE__COROUTINE_SUSPENSION_POINT(3);
status = wuffs_lzw__decoder__flush(self, a_dst);
if (status) {
goto suspend;
}
}
goto ok;
@@ -9670,19 +9670,15 @@ suspend:
self->private_impl.c_decode[0].v_save_code = v_save_code;
self->private_impl.c_decode[0].v_prev_code = v_prev_code;
self->private_impl.c_decode[0].v_width = v_width;
self->private_impl.c_decode[0].v_j = v_j;
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_s = v_s;
self->private_impl.c_decode[0].v_c = v_c;
self->private_impl.c_decode[0].v_n_copied = v_n_copied;
self->private_impl.c_decode[0].v_o = v_o;
goto exit;
exit:
if (a_dst.private_impl.buf) {
a_dst.private_impl.buf->meta.wi =
iop_a_dst - a_dst.private_impl.buf->data.ptr;
}
if (a_src.private_impl.buf) {
a_src.private_impl.buf->meta.ri =
iop_a_src - a_src.private_impl.buf->data.ptr;
@@ -9694,6 +9690,96 @@ exit:
return status;
}
// -------- func lzw.decoder.flush
static wuffs_base__status //
wuffs_lzw__decoder__flush(wuffs_lzw__decoder* self,
wuffs_base__io_writer a_dst) {
wuffs_base__status status = NULL;
uint32_t v_i;
uint32_t v_j;
wuffs_base__slice_u8 v_s;
uint64_t v_n;
uint8_t* iop_a_dst = NULL;
uint8_t* io0_a_dst = NULL;
uint8_t* io1_a_dst = NULL;
WUFFS_BASE__IGNORE_POTENTIALLY_UNUSED_VARIABLE(io0_a_dst);
WUFFS_BASE__IGNORE_POTENTIALLY_UNUSED_VARIABLE(io1_a_dst);
if (a_dst.private_impl.buf) {
iop_a_dst =
a_dst.private_impl.buf->data.ptr + a_dst.private_impl.buf->meta.wi;
if (!a_dst.private_impl.mark) {
a_dst.private_impl.mark = iop_a_dst;
a_dst.private_impl.limit =
a_dst.private_impl.buf->data.ptr + a_dst.private_impl.buf->data.len;
}
if (a_dst.private_impl.buf->meta.closed) {
a_dst.private_impl.limit = iop_a_dst;
}
io0_a_dst = a_dst.private_impl.mark;
io1_a_dst = a_dst.private_impl.limit;
}
uint32_t coro_susp_point = self->private_impl.c_flush[0].coro_susp_point;
if (coro_susp_point) {
v_i = self->private_impl.c_flush[0].v_i;
v_j = self->private_impl.c_flush[0].v_j;
v_s = ((wuffs_base__slice_u8){});
v_n = self->private_impl.c_flush[0].v_n;
} else {
v_s = ((wuffs_base__slice_u8){});
}
switch (coro_susp_point) {
WUFFS_BASE__COROUTINE_SUSPENSION_POINT_0;
v_i = 0;
v_j = self->private_impl.f_flush_j;
while (true) {
if (v_i > v_j) {
status = wuffs_lzw__error__internal_error_inconsistent_i_j;
goto exit;
}
v_s = wuffs_base__slice_u8__subslice_ij(
((wuffs_base__slice_u8){
.ptr = self->private_impl.f_output,
.len = 8192,
}),
v_i, v_j);
v_n = wuffs_base__io_writer__copy_from_slice(&iop_a_dst, io1_a_dst, v_s);
if (v_n == ((uint64_t)(v_s.len))) {
status = NULL;
goto ok;
}
v_i = ((v_i + ((uint32_t)((v_n & 8191)))) & 8191);
status = wuffs_base__suspension__short_write;
WUFFS_BASE__COROUTINE_SUSPENSION_POINT_MAYBE_SUSPEND(1);
}
goto ok;
ok:
self->private_impl.c_flush[0].coro_susp_point = 0;
goto exit;
}
goto suspend;
suspend:
self->private_impl.c_flush[0].coro_susp_point = coro_susp_point;
self->private_impl.c_flush[0].v_i = v_i;
self->private_impl.c_flush[0].v_j = v_j;
self->private_impl.c_flush[0].v_n = v_n;
goto exit;
exit:
if (a_dst.private_impl.buf) {
a_dst.private_impl.buf->meta.wi =
iop_a_dst - a_dst.private_impl.buf->data.ptr;
}
return status;
}
#endif // !defined(WUFFS_CONFIG__MODULES) || defined(WUFFS_CONFIG__MODULE__LZW)
#if !defined(WUFFS_CONFIG__MODULES) || defined(WUFFS_CONFIG__MODULE__ZLIB)
+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"
}
}