Remove some if's from deflate.decode_huffman_fast

name                             old speed     new speed     delta

wuffs_deflate_decode_1k/clang    145MB/s ± 0%  151MB/s ± 1%  +4.54%  (p=0.008 n=5+5)
wuffs_deflate_decode_10k/clang   234MB/s ± 0%  230MB/s ± 0%  -1.75%  (p=0.008 n=5+5)
wuffs_deflate_decode_100k/clang  263MB/s ± 0%  276MB/s ± 1%  +5.17%  (p=0.008 n=5+5)

wuffs_deflate_decode_1k/gcc      157MB/s ± 1%  157MB/s ± 0%    ~     (p=0.421 n=5+5)
wuffs_deflate_decode_10k/gcc     265MB/s ± 0%  259MB/s ± 0%  -2.16%  (p=0.008 n=5+5)
wuffs_deflate_decode_100k/gcc    313MB/s ± 0%  315MB/s ± 0%  +0.39%  (p=0.016 n=5+5)

wuffs_gzip_decode_10k/clang      216MB/s ± 0%  212MB/s ± 0%  -1.69%  (p=0.016 n=5+4)
wuffs_gzip_decode_100k/clang     241MB/s ± 0%  252MB/s ± 0%  +4.67%  (p=0.016 n=5+4)

wuffs_gzip_decode_10k/gcc        251MB/s ± 1%  248MB/s ± 1%  -1.05%  (p=0.008 n=5+5)
wuffs_gzip_decode_100k/gcc       285MB/s ± 1%  291MB/s ± 0%  +2.16%  (p=0.008 n=5+5)

wuffs_zlib_decode_10k/clang      213MB/s ± 0%  210MB/s ± 0%  -1.18%  (p=0.008 n=5+5)
wuffs_zlib_decode_100k/clang     239MB/s ± 0%  247MB/s ± 0%  +3.08%  (p=0.008 n=5+5)

wuffs_zlib_decode_10k/gcc        244MB/s ± 0%  247MB/s ± 0%  +1.15%  (p=0.008 n=5+5)
wuffs_zlib_decode_100k/gcc       282MB/s ± 0%  291MB/s ± 0%  +2.90%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2018-12-01 10:26:09 +11:00
1 parent 776e82d542
commit 3e70383c6c
6 files changed
+52 -55

No files matched your search

+16 -24
View File
@@ -4885,11 +4885,11 @@ static const uint8_t wuffs_deflate__reverse8[256] = {
};
static const uint32_t wuffs_deflate__lcode_magic_numbers[32] = {
1073742592, 1073742848, 1073743104, 1073743360, 1073743616, 1073743872,
1073744128, 1073744384, 1073744656, 1073745168, 1073745680, 1073746192,
1073746720, 1073747744, 1073748768, 1073749792, 1073750832, 1073752880,
1073754928, 1073756976, 1073759040, 1073763136, 1073767232, 1073771328,
1073775440, 1073783632, 1073791824, 1073800016, 1073807872, 134217728,
1073741824, 1073742080, 1073742336, 1073742592, 1073742848, 1073743104,
1073743360, 1073743616, 1073743888, 1073744400, 1073744912, 1073745424,
1073745952, 1073746976, 1073748000, 1073749024, 1073750064, 1073752112,
1073754160, 1073756208, 1073758272, 1073762368, 1073766464, 1073770560,
1073774672, 1073782864, 1073791056, 1073799248, 1073807104, 134217728,
134217728, 134217728,
};
@@ -6147,7 +6147,7 @@ label_0_continue:;
wuffs_deflate__error__internal_error_inconsistent_huffman_decoder_state;
goto exit;
}
v_length = ((v_table_entry >> 8) & 32767);
v_length = (((v_table_entry >> 8) & 255) + 3);
v_table_entry_n_bits = ((v_table_entry >> 4) & 15);
if (v_table_entry_n_bits > 0) {
if (v_n_bits < 15) {
@@ -6160,18 +6160,14 @@ label_0_continue:;
} else {
}
v_length =
((v_length +
((v_bits)&WUFFS_BASE__LOW_BITS_MASK__U32(v_table_entry_n_bits))) &
32767);
(((v_length + 253 +
((v_bits)&WUFFS_BASE__LOW_BITS_MASK__U32(v_table_entry_n_bits))) &
255) +
3);
v_bits >>= v_table_entry_n_bits;
v_n_bits -= v_table_entry_n_bits;
} else {
}
if (v_length > 258) {
status =
wuffs_deflate__error__internal_error_inconsistent_huffman_decoder_state;
goto exit;
}
if (v_n_bits < 15) {
v_bits |= (((uint32_t)(wuffs_base__load_u8be(iop_a_src))) << v_n_bits);
(iop_a_src += 1, wuffs_base__return_empty_struct());
@@ -6254,11 +6250,6 @@ label_0_continue:;
if (v_length > v_hdist) {
v_length -= v_hdist;
v_hlen = v_hdist;
if (v_length > 258) {
status =
wuffs_deflate__error__internal_error_inconsistent_huffman_decoder_state;
goto exit;
}
} else {
v_hlen = v_length;
v_length = 0;
@@ -6543,7 +6534,7 @@ wuffs_deflate__decoder__decode_huffman_slow(wuffs_deflate__decoder* self,
wuffs_deflate__error__internal_error_inconsistent_huffman_decoder_state;
goto exit;
}
v_length = ((v_table_entry >> 8) & 32767);
v_length = (((v_table_entry >> 8) & 255) + 3);
v_table_entry_n_bits = ((v_table_entry >> 4) & 15);
if (v_table_entry_n_bits > 0) {
while (v_n_bits < v_table_entry_n_bits) {
@@ -6559,10 +6550,11 @@ wuffs_deflate__decoder__decode_huffman_slow(wuffs_deflate__decoder* self,
v_bits |= (v_b2 << v_n_bits);
v_n_bits += 8;
}
v_length =
((v_length +
((v_bits)&WUFFS_BASE__LOW_BITS_MASK__U32(v_table_entry_n_bits))) &
32767);
v_length = (((v_length + 253 +
((v_bits)&WUFFS_BASE__LOW_BITS_MASK__U32(
v_table_entry_n_bits))) &
255) +
3);
v_bits >>= v_table_entry_n_bits;
v_n_bits -= v_table_entry_n_bits;
}
+10 -4
View File
@@ -19,8 +19,13 @@ package main
// print-deflate-magic-numbers.go prints the std/deflate lcode_magic_numbers
// and dcode_magic_numbers values based on the tables in RFC 1951 secion 3.2.5.
//
// The lcode base numbers are biased by -3 so that (base_number_minus_3 +
// extra_bits) fits in the range [0, 255]. This makes a bitwise-and with 0xFF a
// no-op, in terms of computed value, but proves to the compiler that the
// result is within a certain range.
//
// The dcode base numbers are biased by -1 so that (base_number_minus_1 +
// extra_bits) fits in the range [0, 32767]. This makes a bitwise and with
// extra_bits) fits in the range [0, 32767]. This makes a bitwise-and with
// 0x7FFF a no-op, in terms of computed value, but proves to the compiler that
// the result is within a certain range. Furthermore, proving that (d + 1) > 0
// is trivial, for d of type u32[..something], compared to proving that d > 0,
@@ -41,16 +46,17 @@ func main() {
}
func main1() error {
// See the "base numbers" comment above.
biases := [2]uint32{3, 1}
for i := 0; i < 2; i++ {
bias := uint32(0)
if i != 0 {
fmt.Println()
bias = 1 // See the "dcode base numbers" comment above.
}
for j := 0; j < 32; j++ {
x := uint32(0x08000000)
if bn := baseNumbers[i][j]; bn != bad {
x = 0x40000000 | ((bn - bias) << 8) | (extraBits[i][j] << 4)
x = 0x40000000 | ((bn - biases[i]) << 8) | (extraBits[i][j] << 4)
}
fmt.Printf("0x%08X,", x)
if j&7 == 7 {
@@ -244,8 +244,8 @@ outer_loop:
table_entry = self->private_impl.f_huffs[0][top + (bits & mask)];
}
// length = base_number + extra_bits.
uint32_t length = (table_entry >> 8) & 0xFFFF;
// length = base_number_minus_3 + 3 + extra_bits.
uint32_t length = ((table_entry >> 8) & 0xFF) + 3;
{
uint32_t n = (table_entry >> 4) & 0x0F;
if (n) {
@@ -290,8 +290,8 @@ outer_loop:
table_entry = self->private_impl.f_huffs[1][top + (bits & mask)];
}
// dist_minus_1 = base_number_minus_1 + extra_bits + 1.
// distance = dist_minus_1 + 1.
// dist_minus_1 = base_number_minus_1 + extra_bits.
// distance = dist_minus_1 + 1.
uint32_t dist_minus_1 = (table_entry >> 8) & 0xFFFF;
{
uint32_t n = (table_entry >> 4) & 0x0F;
+4 -4
View File
@@ -41,10 +41,10 @@ pri status "?internal error: inconsistent n_bits"
// Some trailing elements are 0x08000000. Bit 27 indicates an invalid value.
pri const lcode_magic_numbers array[32] base.u32 = [
0x40000300, 0x40000400, 0x40000500, 0x40000600, 0x40000700, 0x40000800, 0x40000900, 0x40000A00,
0x40000B10, 0x40000D10, 0x40000F10, 0x40001110, 0x40001320, 0x40001720, 0x40001B20, 0x40001F20,
0x40002330, 0x40002B30, 0x40003330, 0x40003B30, 0x40004340, 0x40005340, 0x40006340, 0x40007340,
0x40008350, 0x4000A350, 0x4000C350, 0x4000E350, 0x40010200, 0x08000000, 0x08000000, 0x08000000,
0x40000000, 0x40000100, 0x40000200, 0x40000300, 0x40000400, 0x40000500, 0x40000600, 0x40000700,
0x40000810, 0x40000A10, 0x40000C10, 0x40000E10, 0x40001020, 0x40001420, 0x40001820, 0x40001C20,
0x40002030, 0x40002830, 0x40003030, 0x40003830, 0x40004040, 0x40005040, 0x40006040, 0x40007040,
0x40008050, 0x4000A050, 0x4000C050, 0x4000E050, 0x4000FF00, 0x08000000, 0x08000000, 0x08000000,
]
pri const dcode_magic_numbers array[32] base.u32 = [
+9 -15
View File
@@ -153,8 +153,11 @@ pri func decoder.decode_huffman_fast!?(dst base.io_writer, src base.io_reader) {
return "?internal error: inconsistent Huffman decoder state"
}
// length = base_number + extra_bits.
var length base.u32[..0x7FFF] = (table_entry >> 8) & 0x7FFF
// length = base_number_minus_3 + 3 + extra_bits.
//
// The -3 is from the bias in script/print-deflate-magic-numbers.go.
// That bias makes the "& 0xFF" 1 and 15-ish lines below correct.
var length base.u32[..258] = ((table_entry >> 8) & 0xFF) + 3
table_entry_n_bits = (table_entry >> 4) & 0x0F
if table_entry_n_bits > 0 {
// Ensure that we have at least 15 bits of input.
@@ -174,16 +177,15 @@ pri func decoder.decode_huffman_fast!?(dst base.io_writer, src base.io_reader) {
assert args.src.available() >= 6
assert n_bits >= 15
length = (length + bits.low_bits(n:table_entry_n_bits)) & 0x7FFF
// The "+ 253" is the same as "- 3", after the "& 0xFF", but the
// plus form won't require an underflow check.
length = ((length + 253 + bits.low_bits(n:table_entry_n_bits)) & 0xFF) + 3
bits >>= table_entry_n_bits
n_bits -= table_entry_n_bits
} else {
assert args.src.available() >= 6
}
if length > 258 {
return "?internal error: inconsistent Huffman decoder state"
}
// Ensure that we have at least 15 bits of input.
if n_bits < 15 {
@@ -252,7 +254,7 @@ pri func decoder.decode_huffman_fast!?(dst base.io_writer, src base.io_reader) {
// distance = dist_minus_1 + 1.
//
// The -1 is from the bias in script/print-deflate-magic-numbers.go.
// That bias makes the "& 0x7FFF" 15-ish lines below correct and
// That bias makes the "& 0x7FFF" 2 and 15-ish lines below correct and
// undoing that bias makes proving (dist_minus_1 + 1) > 0 trivial.
var dist_minus_1 base.u32[..0x7FFF] = (table_entry >> 8) & 0x7FFF
table_entry_n_bits = (table_entry >> 4) & 0x0F
@@ -281,7 +283,6 @@ pri func decoder.decode_huffman_fast!?(dst base.io_writer, src base.io_reader) {
// its presence minimizes the diff between decode_huffman_fast and
// decode_huffman_slow.
while true,
pre length <= 258,
pre args.dst.available() >= 258,
{
// Copy from this.history.
@@ -297,14 +298,9 @@ pri func decoder.decode_huffman_fast!?(dst base.io_writer, src base.io_reader) {
assert hdist < 0x8000 via "a < b: a < c; c <= b"(c:length)
length -= hdist
hlen = hdist
// TODO: this if check should be redundant.
if length > 258 {
return "?internal error: inconsistent Huffman decoder state"
}
} else {
hlen = length
length = 0
assert length <= 258
}
if this.history_index < hdist {
return "?bad distance"
@@ -324,7 +320,6 @@ pri func decoder.decode_huffman_fast!?(dst base.io_writer, src base.io_reader) {
// "etc", but its presence minimizes the diff between
// decode_huffman_fast and decode_huffman_slow.
while true,
inv length <= 258,
inv args.dst.available() >= 258,
{
// TODO: copy_from_slice32 should probably update the
@@ -350,7 +345,6 @@ pri func decoder.decode_huffman_fast!?(dst base.io_writer, src base.io_reader) {
}
}
// Once again, redundant but explicit assertions.
assert length <= 258
assert ((dist_minus_1 + 1) as base.u64) <= args.dst.since_mark().length()
assert args.dst.available() >= 258
+9 -4
View File
@@ -101,8 +101,11 @@ pri func decoder.decode_huffman_slow!??(dst base.io_writer, src base.io_reader)
return "?internal error: inconsistent Huffman decoder state"
}
// length = base_number + extra_bits.
var length base.u32[..0x7FFF] = (table_entry >> 8) & 0x7FFF
// length = base_number_minus_3 + 3 + extra_bits.
//
// The -3 is from the bias in script/print-deflate-magic-numbers.go.
// That bias makes the "& 0xFF" 1 and 15-ish lines below correct.
var length base.u32[..258] = ((table_entry >> 8) & 0xFF) + 3
table_entry_n_bits = (table_entry >> 4) & 0x0F
if table_entry_n_bits > 0 {
while n_bits < table_entry_n_bits,
@@ -113,7 +116,9 @@ pri func decoder.decode_huffman_slow!??(dst base.io_writer, src base.io_reader)
bits |= b2 << n_bits
n_bits += 8
}
length = (length + bits.low_bits(n:table_entry_n_bits)) & 0x7FFF
// The "+ 253" is the same as "- 3", after the "& 0xFF", but the
// plus form won't require an underflow check.
length = ((length + 253 + bits.low_bits(n:table_entry_n_bits)) & 0xFF) + 3
bits >>= table_entry_n_bits
n_bits -= table_entry_n_bits
}
@@ -166,7 +171,7 @@ pri func decoder.decode_huffman_slow!??(dst base.io_writer, src base.io_reader)
// distance = dist_minus_1 + 1.
//
// The -1 is from the bias in script/print-deflate-magic-numbers.go.
// That bias makes the "& 0x7FFF" 15-ish lines below correct and
// That bias makes the "& 0x7FFF" 2 and 15-ish lines below correct and
// undoing that bias makes proving (dist_minus_1 + 1) > 0 trivial.
var dist_minus_1 base.u32[..0x7FFF] = (table_entry >> 8) & 0x7FFF
table_entry_n_bits = (table_entry >> 4) & 0x0F