mirror of https://github.com/google/snappy.git
Re-work fast path that emits copies in zippy compression.
The primary motivation for the change is that FindMatchLength is likely to discover a difference in the first 8 bytes it compares. If that occurs then we know the length of the match is less than 12, because FindMatchLength is invoked after a 4-byte match is found. When emitting a copy, it is useful to know that the length is less than 12 because the two-byte variant of an emitted copy requires that. This is a performance-tuning change that should not affect the library's behavior. With FDO on perflab/Haswell the geometric mean for ZFlat/* went from 47,290ns to 45,741ns, an improvement of 3.4%. SAMPLE (before) BM_ZFlat/0 102824 102650 40691 951.4MB/s html (22.31 %) BM_ZFlat/1 1293512 1290442 3225 518.9MB/s urls (47.78 %) BM_ZFlat/2 10373 10353 417959 11.1GB/s jpg (99.95 %) BM_ZFlat/3 268 268 15745324 712.4MB/s jpg_200 (73.00 %) BM_ZFlat/4 12137 12113 342462 7.9GB/s pdf (83.30 %) BM_ZFlat/5 430672 429720 9724 909.0MB/s html4 (22.52 %) BM_ZFlat/6 420541 419636 9833 345.6MB/s txt1 (57.88 %) BM_ZFlat/7 373829 373158 10000 319.9MB/s txt2 (61.91 %) BM_ZFlat/8 1119014 1116604 3755 364.5MB/s txt3 (54.99 %) BM_ZFlat/9 1544203 1540657 2748 298.3MB/s txt4 (66.26 %) BM_ZFlat/10 91041 90866 46002 1.2GB/s pb (19.68 %) BM_ZFlat/11 332766 331990 10000 529.5MB/s gaviota (37.72 %) BM_ZFlat/12 39960 39886 100000 588.3MB/s cp (48.12 %) BM_ZFlat/13 14493 14465 287181 735.1MB/s c (42.47 %) BM_ZFlat/14 4447 4440 947927 799.3MB/s lsp (48.37 %) BM_ZFlat/15 1316362 1313350 3196 747.7MB/s xls (41.23 %) BM_ZFlat/16 312 311 10000000 613.0MB/s xls_200 (78.00 %) BM_ZFlat/17 388471 387502 10000 1.2GB/s bin (18.11 %) BM_ZFlat/18 65 64 64838208 2.9GB/s bin_200 (7.50 %) BM_ZFlat/19 65900 65787 63099 554.3MB/s sum (48.96 %) BM_ZFlat/20 6188 6177 681951 652.6MB/s man (59.21 %) SAMPLE (after) Benchmark Time(ns) CPU(ns) Iterations -------------------------------------------- BM_ZFlat/0 99259 99044 42428 986.0MB/s html (22.31 %) BM_ZFlat/1 1257039 1255276 3341 533.4MB/s urls (47.78 %) BM_ZFlat/2 10044 10030 405781 11.4GB/s jpg (99.95 %) BM_ZFlat/3 268 267 15732282 713.3MB/s jpg_200 (73.00 %) BM_ZFlat/4 11675 11657 358629 8.2GB/s pdf (83.30 %) BM_ZFlat/5 420951 419818 9739 930.5MB/s html4 (22.52 %) BM_ZFlat/6 415460 414632 10000 349.8MB/s txt1 (57.88 %) BM_ZFlat/7 367191 366436 10000 325.8MB/s txt2 (61.91 %) BM_ZFlat/8 1098345 1096036 3819 371.3MB/s txt3 (54.99 %) BM_ZFlat/9 1508701 1505306 2758 305.3MB/s txt4 (66.26 %) BM_ZFlat/10 87195 87031 47289 1.3GB/s pb (19.68 %) BM_ZFlat/11 322338 321637 10000 546.5MB/s gaviota (37.72 %) BM_ZFlat/12 36739 36668 100000 639.9MB/s cp (48.12 %) BM_ZFlat/13 13646 13618 304009 780.9MB/s c (42.47 %) BM_ZFlat/14 4249 4240 992456 837.0MB/s lsp (48.37 %) BM_ZFlat/15 1262925 1260012 3314 779.4MB/s xls (41.23 %) BM_ZFlat/16 308 308 10000000 619.8MB/s xls_200 (78.00 %) BM_ZFlat/17 379750 378944 10000 1.3GB/s bin (18.11 %) BM_ZFlat/18 62 62 67443280 3.0GB/s bin_200 (7.50 %) BM_ZFlat/19 61706 61587 67645 592.1MB/s sum (48.96 %) BM_ZFlat/20 5968 5958 698974 676.6MB/s man (59.21 %)
This commit is contained in:
Geoff Pike
Alkis Evlogimenos
parent
094c67de88
commit
38a5ec5fca
|
|
@ -70,11 +70,12 @@ char* CompressFragment(const char* input,
|
|||
uint16* table,
|
||||
const int table_size);
|
||||
|
||||
// Return the largest n such that
|
||||
// Find the largest n such that
|
||||
//
|
||||
// s1[0,n-1] == s2[0,n-1]
|
||||
// and n <= (s2_limit - s2).
|
||||
//
|
||||
// Return make_pair(n, n < 8).
|
||||
// Does not read *s2_limit or beyond.
|
||||
// Does not read *(s1 + (s2_limit - s2)) or beyond.
|
||||
// Requires that s2_limit >= s2.
|
||||
|
|
@ -82,11 +83,27 @@ char* CompressFragment(const char* input,
|
|||
// Separate implementation for x86_64, for speed. Uses the fact that
|
||||
// x86_64 is little endian.
|
||||
#if defined(ARCH_K8)
|
||||
static inline int FindMatchLength(const char* s1,
|
||||
const char* s2,
|
||||
const char* s2_limit) {
|
||||
static inline pair<size_t, bool> FindMatchLength(const char* s1,
|
||||
const char* s2,
|
||||
const char* s2_limit) {
|
||||
assert(s2_limit >= s2);
|
||||
int matched = 0;
|
||||
size_t matched = 0;
|
||||
|
||||
// This block isn't necessary for correctness; we could just start looping
|
||||
// immediately. As an optimization though, it is useful. It creates some not
|
||||
// uncommon code paths that determine, without extra effort, whether the match
|
||||
// length is less than 8. In short, we are hoping to avoid a conditional
|
||||
// branch, and perhaps get better code layout from the C++ compiler.
|
||||
if (PREDICT_TRUE(s2 <= s2_limit - 8)) {
|
||||
uint64 a1 = UNALIGNED_LOAD64(s1);
|
||||
uint64 a2 = UNALIGNED_LOAD64(s2);
|
||||
if (a1 != a2) {
|
||||
return pair<size_t, bool>(Bits::FindLSBSetNonZero64(a1 ^ a2) >> 3, true);
|
||||
} else {
|
||||
matched = 8;
|
||||
s2 += 8;
|
||||
}
|
||||
}
|
||||
|
||||
// Find out how long the match is. We loop over the data 64 bits at a
|
||||
// time until we find a 64-bit block that doesn't match; then we find
|
||||
|
|
@ -97,14 +114,11 @@ static inline int FindMatchLength(const char* s1,
|
|||
s2 += 8;
|
||||
matched += 8;
|
||||
} else {
|
||||
// On current (mid-2008) Opteron models there is a 3% more
|
||||
// efficient code sequence to find the first non-matching byte.
|
||||
// However, what follows is ~10% better on Intel Core 2 and newer,
|
||||
// and we expect AMD's bsf instruction to improve.
|
||||
uint64 x = UNALIGNED_LOAD64(s2) ^ UNALIGNED_LOAD64(s1 + matched);
|
||||
int matching_bits = Bits::FindLSBSetNonZero64(x);
|
||||
matched += matching_bits >> 3;
|
||||
return matched;
|
||||
assert(matched >= 8);
|
||||
return pair<size_t, bool>(matched, false);
|
||||
}
|
||||
}
|
||||
while (PREDICT_TRUE(s2 < s2_limit)) {
|
||||
|
|
@ -112,15 +126,15 @@ static inline int FindMatchLength(const char* s1,
|
|||
++s2;
|
||||
++matched;
|
||||
} else {
|
||||
return matched;
|
||||
return pair<size_t, bool>(matched, matched < 8);
|
||||
}
|
||||
}
|
||||
return matched;
|
||||
return pair<size_t, bool>(matched, matched < 8);
|
||||
}
|
||||
#else
|
||||
static inline int FindMatchLength(const char* s1,
|
||||
const char* s2,
|
||||
const char* s2_limit) {
|
||||
static inline pair<size_t, bool> FindMatchLength(const char* s1,
|
||||
const char* s2,
|
||||
const char* s2_limit) {
|
||||
// Implementation based on the x86-64 version, above.
|
||||
assert(s2_limit >= s2);
|
||||
int matched = 0;
|
||||
|
|
@ -140,7 +154,7 @@ static inline int FindMatchLength(const char* s1,
|
|||
++matched;
|
||||
}
|
||||
}
|
||||
return matched;
|
||||
return pair<size_t, bool>(matched, matched < 8);
|
||||
}
|
||||
#endif
|
||||
|
||||
|
|
|
|||
72
snappy.cc
72
snappy.cc
|
|
@ -41,7 +41,6 @@ namespace snappy {
|
|||
|
||||
using internal::COPY_1_BYTE_OFFSET;
|
||||
using internal::COPY_2_BYTE_OFFSET;
|
||||
using internal::COPY_4_BYTE_OFFSET;
|
||||
using internal::LITERAL;
|
||||
using internal::char_table;
|
||||
using internal::kMaximumTagLength;
|
||||
|
|
@ -199,43 +198,55 @@ static inline char* EmitLiteral(char* op,
|
|||
return op + len;
|
||||
}
|
||||
|
||||
static inline char* EmitCopyLessThan64(char* op, size_t offset, int len) {
|
||||
static inline char* EmitCopyAtMost64(char* op, size_t offset, size_t len,
|
||||
bool len_less_than_12) {
|
||||
assert(len <= 64);
|
||||
assert(len >= 4);
|
||||
assert(offset < 65536);
|
||||
assert(len_less_than_12 == (len < 12));
|
||||
|
||||
if ((len < 12) && (offset < 2048)) {
|
||||
size_t len_minus_4 = len - 4;
|
||||
assert(len_minus_4 < 8); // Must fit in 3 bits
|
||||
*op++ = COPY_1_BYTE_OFFSET + ((len_minus_4) << 2) + ((offset >> 8) << 5);
|
||||
if (len_less_than_12 && PREDICT_TRUE(offset < 2048)) {
|
||||
// offset fits in 11 bits. The 3 highest go in the top of the first byte,
|
||||
// and the rest go in the second byte.
|
||||
*op++ = COPY_1_BYTE_OFFSET + ((len - 4) << 2) + ((offset >> 3) & 0xe0);
|
||||
*op++ = offset & 0xff;
|
||||
} else {
|
||||
*op++ = COPY_2_BYTE_OFFSET + ((len-1) << 2);
|
||||
LittleEndian::Store16(op, offset);
|
||||
op += 2;
|
||||
// Write 4 bytes, though we only care about 3 of them. The output buffer
|
||||
// is required to have some slack, so the extra byte won't overrun it.
|
||||
uint32 u = COPY_2_BYTE_OFFSET + ((len - 1) << 2) + (offset << 8);
|
||||
LittleEndian::Store32(op, u);
|
||||
op += 3;
|
||||
}
|
||||
return op;
|
||||
}
|
||||
|
||||
static inline char* EmitCopy(char* op, size_t offset, int len) {
|
||||
// Emit 64 byte copies but make sure to keep at least four bytes reserved
|
||||
while (PREDICT_FALSE(len >= 68)) {
|
||||
op = EmitCopyLessThan64(op, offset, 64);
|
||||
len -= 64;
|
||||
}
|
||||
static inline char* EmitCopy(char* op, size_t offset, size_t len,
|
||||
bool len_less_than_12) {
|
||||
assert(len_less_than_12 == (len < 12));
|
||||
if (len_less_than_12) {
|
||||
return EmitCopyAtMost64(op, offset, len, true);
|
||||
} else {
|
||||
// A special case for len <= 64 might help, but so far measurements suggest
|
||||
// it's in the noise.
|
||||
|
||||
// Emit an extra 60 byte copy if have too much data to fit in one copy
|
||||
if (len > 64) {
|
||||
op = EmitCopyLessThan64(op, offset, 60);
|
||||
len -= 60;
|
||||
}
|
||||
// Emit 64 byte copies but make sure to keep at least four bytes reserved.
|
||||
while (PREDICT_FALSE(len >= 68)) {
|
||||
op = EmitCopyAtMost64(op, offset, 64, false);
|
||||
len -= 64;
|
||||
}
|
||||
|
||||
// Emit remainder
|
||||
op = EmitCopyLessThan64(op, offset, len);
|
||||
return op;
|
||||
// One or two copies will now finish the job.
|
||||
if (len > 64) {
|
||||
op = EmitCopyAtMost64(op, offset, 60, false);
|
||||
len -= 60;
|
||||
}
|
||||
|
||||
// Emit remainder.
|
||||
op = EmitCopyAtMost64(op, offset, len, len < 12);
|
||||
return op;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
bool GetUncompressedLength(const char* start, size_t n, size_t* result) {
|
||||
uint32 v = 0;
|
||||
const char* limit = start + n;
|
||||
|
|
@ -422,19 +433,20 @@ char* CompressFragment(const char* input,
|
|||
// We have a 4-byte match at ip, and no need to emit any
|
||||
// "literal bytes" prior to ip.
|
||||
const char* base = ip;
|
||||
int matched = 4 + FindMatchLength(candidate + 4, ip + 4, ip_end);
|
||||
pair<size_t, bool> p = FindMatchLength(candidate + 4, ip + 4, ip_end);
|
||||
size_t matched = 4 + p.first;
|
||||
ip += matched;
|
||||
size_t offset = base - candidate;
|
||||
assert(0 == memcmp(base, candidate, matched));
|
||||
op = EmitCopy(op, offset, matched);
|
||||
// We could immediately start working at ip now, but to improve
|
||||
// compression we first update table[Hash(ip - 1, ...)].
|
||||
const char* insert_tail = ip - 1;
|
||||
op = EmitCopy(op, offset, matched, p.second);
|
||||
next_emit = ip;
|
||||
if (PREDICT_FALSE(ip >= ip_limit)) {
|
||||
goto emit_remainder;
|
||||
}
|
||||
input_bytes = GetEightBytesAt(insert_tail);
|
||||
// We are now looking for a 4-byte match again. We read
|
||||
// table[Hash(ip, shift)] for that. To improve compression,
|
||||
// we also update table[Hash(ip - 1, shift)] and table[Hash(ip, shift)].
|
||||
input_bytes = GetEightBytesAt(ip - 1);
|
||||
uint32 prev_hash = HashBytes(GetUint32AtOffset(input_bytes, 0), shift);
|
||||
table[prev_hash] = ip - base_ip - 1;
|
||||
uint32 cur_hash = HashBytes(GetUint32AtOffset(input_bytes, 1), shift);
|
||||
|
|
|
|||
|
|
@ -1054,7 +1054,9 @@ TEST(Snappy, ZeroOffsetCopyValidation) {
|
|||
namespace {
|
||||
|
||||
int TestFindMatchLength(const char* s1, const char *s2, unsigned length) {
|
||||
return snappy::internal::FindMatchLength(s1, s2, s2 + length);
|
||||
pair<size_t, bool> p = snappy::internal::FindMatchLength(s1, s2, s2 + length);
|
||||
CHECK_EQ(p.first < 8, p.second);
|
||||
return p.first;
|
||||
}
|
||||
|
||||
} // namespace
|
||||
|
|
@ -1164,8 +1166,7 @@ TEST(Snappy, FindMatchLengthRandom) {
|
|||
}
|
||||
DataEndingAtUnreadablePage u(s);
|
||||
DataEndingAtUnreadablePage v(t);
|
||||
int matched = snappy::internal::FindMatchLength(
|
||||
u.data(), v.data(), v.data() + t.size());
|
||||
int matched = TestFindMatchLength(u.data(), v.data(), t.size());
|
||||
if (matched == t.size()) {
|
||||
EXPECT_EQ(s, t);
|
||||
} else {
|
||||
|
|
|
|||
Loading…
Reference in New Issue