2 * Copyright 2014 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 * @author Philip Pronin (philipp@fb.com)
20 * Based on the paper by Sebastiano Vigna,
21 * "Quasi-succinct indices" (arxiv:1206.4300).
24 #ifndef FOLLY_EXPERIMENTAL_ELIAS_FANO_CODING_H
25 #define FOLLY_EXPERIMENTAL_ELIAS_FANO_CODING_H
28 #error EliasFanoCoding.h requires GCC
31 #if !defined(__x86_64__)
32 #error EliasFanoCoding.h requires x86_64
38 #include <type_traits>
39 #include <boost/noncopyable.hpp>
40 #include <glog/logging.h>
41 #include "folly/Bits.h"
42 #include "folly/CpuId.h"
43 #include "folly/Likely.h"
44 #include "folly/Range.h"
46 #if __BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__
47 #error EliasFanoCoding.h requires little endianness
50 namespace folly { namespace compression {
52 template <class Value,
53 class SkipValue = size_t,
54 size_t kSkipQuantum = 0, // 0 = disabled
55 size_t kForwardQuantum = 0> // 0 = disabled
56 struct EliasFanoCompressedList {
57 static_assert(std::is_integral<Value>::value &&
58 std::is_unsigned<Value>::value,
59 "Value should be unsigned integral");
61 typedef Value ValueType;
62 typedef SkipValue SkipValueType;
64 EliasFanoCompressedList()
65 : size(0), numLowerBits(0) { }
67 static constexpr size_t skipQuantum = kSkipQuantum;
68 static constexpr size_t forwardQuantum = kForwardQuantum;
73 // WARNING: EliasFanoCompressedList has no ownership of
74 // lower, upper, skipPointers and forwardPointers.
75 // The 7 bytes following the last byte of lower and upper
76 // sequences should be readable.
77 folly::ByteRange lower;
78 folly::ByteRange upper;
80 folly::ByteRange skipPointers;
81 folly::ByteRange forwardPointers;
84 ::free(const_cast<unsigned char*>(lower.data()));
85 ::free(const_cast<unsigned char*>(upper.data()));
86 ::free(const_cast<unsigned char*>(skipPointers.data()));
87 ::free(const_cast<unsigned char*>(forwardPointers.data()));
90 static uint8_t defaultNumLowerBits(size_t upperBound, size_t size) {
91 if (size == 0 || upperBound < size) {
94 // floor(log(upperBound / size));
95 return folly::findLastSet(upperBound / size) - 1;
98 // WARNING: encode() mallocates lower, upper, skipPointers
99 // and forwardPointers. As EliasFanoCompressedList has
100 // no ownership of them, you need to call free() explicitly.
101 static void encode(const ValueType* list, size_t size,
102 EliasFanoCompressedList& result) {
103 encode(list, list + size, result);
106 // Range (begin, end) should be sorted.
107 template <class RandomAccessIterator>
108 static void encode(RandomAccessIterator begin,
109 RandomAccessIterator end,
110 EliasFanoCompressedList& result) {
111 CHECK(std::is_sorted(begin, end));
114 const size_t size = end - begin;
117 result = EliasFanoCompressedList();
121 const ValueType upperBound = list[size - 1];
122 uint8_t numLowerBits = defaultNumLowerBits(upperBound, size);
124 // This is detail::writeBits56 limitation.
125 numLowerBits = std::min<uint8_t>(numLowerBits, 56);
126 CHECK_LT(numLowerBits, 8 * sizeof(Value)); // As we shift by numLowerBits.
128 // WARNING: Current read/write logic assumes that the 7 bytes
129 // following the last byte of lower and upper sequences are
130 // readable (stored value doesn't matter and won't be changed),
131 // so we allocate additional 7B, but do not include them in size
132 // of returned value.
135 const size_t lowerSize = (numLowerBits * size + 7) / 8;
136 unsigned char* lower = nullptr;
137 if (lowerSize > 0) { // numLowerBits != 0
138 lower = static_cast<unsigned char*>(calloc(lowerSize + 7, 1));
139 const ValueType lowerMask = (ValueType(1) << numLowerBits) - 1;
140 for (size_t i = 0; i < size; ++i) {
141 const ValueType lowerBits = list[i] & lowerMask;
142 writeBits56(lower, i * numLowerBits, numLowerBits, lowerBits);
147 // Upper bits are stored using unary delta encoding.
148 // For example, (3 5 5 9) will be encoded as 1000011001000_2.
149 const size_t upperSizeBits =
150 (upperBound >> numLowerBits) + // Number of 0-bits to be stored.
152 const size_t upperSize = (upperSizeBits + 7) / 8;
153 unsigned char* const upper =
154 static_cast<unsigned char*>(calloc(upperSize + 7, 1));
155 for (size_t i = 0; i < size; ++i) {
156 const ValueType upperBits = list[i] >> numLowerBits;
157 const size_t pos = upperBits + i; // upperBits 0-bits and (i + 1) 1-bits.
158 upper[pos / 8] |= 1U << (pos % 8);
161 // *** Skip pointers.
162 // Store (1-indexed) position of every skipQuantum-th
163 // 0-bit in upper bits sequence.
164 SkipValueType* skipPointers = nullptr;
165 size_t numSkipPointers = 0;
166 /* static */ if (skipQuantum != 0) {
167 // Workaround to avoid 'division by zero' compile-time error.
168 constexpr size_t q = skipQuantum ?: 1;
169 CHECK_LT(upperSizeBits, std::numeric_limits<SkipValueType>::max());
170 // 8 * upperSize is used here instead of upperSizeBits, as that is
171 // more serialization-friendly way.
172 numSkipPointers = (8 * upperSize - size) / q;
173 skipPointers = static_cast<SkipValueType*>(
176 : calloc(numSkipPointers, sizeof(SkipValueType)));
178 for (size_t i = 0, pos = 0; i < size; ++i) {
179 const ValueType upperBits = list[i] >> numLowerBits;
180 for (; (pos + 1) * q <= upperBits; ++pos) {
181 skipPointers[pos] = i + (pos + 1) * q;
186 // *** Forward pointers.
187 // Store (1-indexed) position of every forwardQuantum-th
188 // 1-bit in upper bits sequence.
189 SkipValueType* forwardPointers = nullptr;
190 size_t numForwardPointers = 0;
191 /* static */ if (forwardQuantum != 0) {
192 // Workaround to avoid 'division by zero' compile-time error.
193 constexpr size_t q = forwardQuantum ?: 1;
194 CHECK_LT(upperSizeBits, std::numeric_limits<SkipValueType>::max());
196 numForwardPointers = size / q;
197 forwardPointers = static_cast<SkipValueType*>(
198 numForwardPointers == 0
200 : malloc(numForwardPointers * sizeof(SkipValueType)));
202 for (size_t i = q - 1, pos = 0; i < size; i += q, ++pos) {
203 const ValueType upperBits = list[i] >> numLowerBits;
204 forwardPointers[pos] = upperBits + i + 1;
210 result.numLowerBits = numLowerBits;
211 result.lower.reset(lower, lowerSize);
212 result.upper.reset(upper, upperSize);
213 result.skipPointers.reset(
214 reinterpret_cast<unsigned char*>(skipPointers),
215 numSkipPointers * sizeof(SkipValueType));
216 result.forwardPointers.reset(
217 reinterpret_cast<unsigned char*>(forwardPointers),
218 numForwardPointers * sizeof(SkipValueType));
222 // Writes value (with len up to 56 bits) to data starting at pos-th bit.
223 static void writeBits56(unsigned char* data, size_t pos,
224 uint8_t len, uint64_t value) {
225 DCHECK_LE(uint32_t(len), 56);
226 DCHECK_EQ(0, value & ~((uint64_t(1) << len) - 1));
227 unsigned char* const ptr = data + (pos / 8);
228 uint64_t ptrv = folly::loadUnaligned<uint64_t>(ptr);
229 ptrv |= value << (pos % 8);
230 folly::storeUnaligned<uint64_t>(ptr, ptrv);
234 // NOTE: It's recommended to compile EF coding with -msse4.2, starting
235 // with Nehalem, Intel CPUs support POPCNT instruction and gcc will emit
236 // it for __builtin_popcountll intrinsic.
237 // But we provide an alternative way for the client code: it can switch to
238 // the appropriate version of EliasFanoReader<> in realtime (client should
239 // implement this switching logic itself) by specifying instruction set to
241 namespace instructions {
244 static bool supported() {
247 static inline uint64_t popcount(uint64_t value) {
248 return __builtin_popcountll(value);
250 static inline int ctz(uint64_t value) {
252 return __builtin_ctzll(value);
256 struct Fast : public Default {
257 static bool supported() {
259 return cpuId.popcnt();
261 static inline uint64_t popcount(uint64_t value) {
263 asm ("popcntq %1, %0" : "=r" (result) : "r" (value));
268 } // namespace instructions
272 template <class CompressedList, class Instructions>
273 class UpperBitsReader {
274 typedef typename CompressedList::SkipValueType SkipValueType;
276 typedef typename CompressedList::ValueType ValueType;
278 explicit UpperBitsReader(const CompressedList& list)
279 : forwardPointers_(list.forwardPointers.data()),
280 skipPointers_(list.skipPointers.data()),
281 start_(list.upper.data()),
282 block_(start_ != nullptr ? folly::loadUnaligned<block_t>(start_) : 0),
283 outer_(0), // outer offset: number of consumed bytes in upper.
284 inner_(-1), // inner offset: (bit) position in current block.
285 position_(-1), // index of current value (= #reads - 1).
288 size_t position() const { return position_; }
289 ValueType value() const { return value_; }
292 // Skip to the first non-zero block.
293 while (block_ == 0) {
294 outer_ += sizeof(block_t);
295 block_ = folly::loadUnaligned<block_t>(start_ + outer_);
299 inner_ = Instructions::ctz(block_);
300 block_ &= block_ - 1;
305 ValueType skip(size_t n) {
308 position_ += n; // n 1-bits will be read.
310 // Use forward pointer.
311 if (CompressedList::forwardQuantum > 0 &&
312 n > CompressedList::forwardQuantum) {
313 // Workaround to avoid 'division by zero' compile-time error.
314 constexpr size_t q = CompressedList::forwardQuantum ?: 1;
316 const size_t steps = position_ / q;
318 folly::loadUnaligned<SkipValueType>(
319 forwardPointers_ + (steps - 1) * sizeof(SkipValueType));
322 n = position_ + 1 - steps * q; // n is > 0.
323 // correct inner_ will be set at the end.
327 // Find necessary block.
328 while ((cnt = Instructions::popcount(block_)) < n) {
330 outer_ += sizeof(block_t);
331 block_ = folly::loadUnaligned<block_t>(start_ + outer_);
334 // NOTE: Trying to skip half-block here didn't show any
335 // performance improvements.
339 // Kill n - 1 least significant 1-bits.
340 for (size_t i = 0; i < n - 1; ++i) {
341 block_ &= block_ - 1;
344 inner_ = Instructions::ctz(block_);
345 block_ &= block_ - 1;
350 // Skip to the first element that is >= v and located *after* the current
351 // one (so even if current value equals v, position will be increased by 1).
352 ValueType skipToNext(ValueType v) {
353 DCHECK_GE(v, value_);
356 if (CompressedList::skipQuantum > 0 &&
357 v >= value_ + CompressedList::skipQuantum) {
358 // Workaround to avoid 'division by zero' compile-time error.
359 constexpr size_t q = CompressedList::skipQuantum ?: 1;
361 const size_t steps = v / q;
363 folly::loadUnaligned<SkipValueType>(
364 skipPointers_ + (steps - 1) * sizeof(SkipValueType));
367 position_ = dest - q * steps - 1;
368 // Correct inner_ and value_ will be set during the next()
371 // NOTE: Corresponding block of lower bits sequence may be
372 // prefetched here (via __builtin_prefetch), but experiments
373 // didn't show any significant improvements.
378 size_t skip = v - (8 * outer_ - position_ - 1);
380 constexpr size_t kBitsPerBlock = 8 * sizeof(block_t);
381 while ((cnt = Instructions::popcount(~block_)) < skip) {
383 position_ += kBitsPerBlock - cnt;
384 outer_ += sizeof(block_t);
385 block_ = folly::loadUnaligned<block_t>(start_ + outer_);
388 // Try to skip half-block.
389 constexpr size_t kBitsPerHalfBlock = 4 * sizeof(block_t);
390 constexpr block_t halfBlockMask = (block_t(1) << kBitsPerHalfBlock) - 1;
391 if ((cnt = Instructions::popcount(~block_ & halfBlockMask)) < skip) {
392 position_ += kBitsPerHalfBlock - cnt;
393 block_ &= ~halfBlockMask;
396 // Just skip until we see expected value.
397 while (next() < v) { }
402 ValueType setValue() {
403 value_ = static_cast<ValueType>(8 * outer_ + inner_ - position_);
407 void reposition(size_t dest) {
409 block_ = folly::loadUnaligned<block_t>(start_ + outer_);
410 block_ &= ~((block_t(1) << (dest % 8)) - 1);
413 typedef unsigned long long block_t;
414 const unsigned char* const forwardPointers_;
415 const unsigned char* const skipPointers_;
416 const unsigned char* const start_;
424 } // namespace detail
426 template <class CompressedList,
427 class Instructions = instructions::Default>
428 class EliasFanoReader : private boost::noncopyable {
430 typedef typename CompressedList::ValueType ValueType;
432 explicit EliasFanoReader(const CompressedList& list)
434 lowerMask_((ValueType(1) << list_.numLowerBits) - 1),
438 DCHECK(Instructions::supported());
439 // To avoid extra branching during skipTo() while reading
440 // upper sequence we need to know the last element.
441 if (UNLIKELY(list_.size == 0)) {
445 ValueType lastUpperValue = 8 * list_.upper.size() - list_.size;
446 auto it = list_.upper.end() - 1;
448 lastUpperValue -= 8 - folly::findLastSet(*it);
449 lastValue_ = readLowerPart(list_.size - 1) |
450 (lastUpperValue << list_.numLowerBits);
453 size_t size() const { return list_.size; }
455 size_t position() const { return progress_ - 1; }
456 ValueType value() const { return value_; }
459 if (UNLIKELY(progress_ == list_.size)) {
460 value_ = std::numeric_limits<ValueType>::max();
463 value_ = readLowerPart(progress_) |
464 (upper_.next() << list_.numLowerBits);
469 bool skip(size_t n) {
473 if (LIKELY(progress_ < list_.size)) {
474 value_ = readLowerPart(progress_) |
475 (upper_.skip(n) << list_.numLowerBits);
480 progress_ = list_.size;
481 value_ = std::numeric_limits<ValueType>::max();
485 bool skipTo(ValueType value) {
486 DCHECK_GE(value, value_);
487 if (value <= value_) {
490 if (value > lastValue_) {
491 progress_ = list_.size;
492 value_ = std::numeric_limits<ValueType>::max();
496 upper_.skipToNext(value >> list_.numLowerBits);
497 progress_ = upper_.position();
498 value_ = readLowerPart(progress_) |
499 (upper_.value() << list_.numLowerBits);
501 while (value_ < value) {
502 value_ = readLowerPart(progress_) |
503 (upper_.next() << list_.numLowerBits);
511 ValueType readLowerPart(size_t i) const {
512 const size_t pos = i * list_.numLowerBits;
513 const unsigned char* ptr = list_.lower.data() + (pos / 8);
514 const uint64_t ptrv = folly::loadUnaligned<uint64_t>(ptr);
515 return lowerMask_ & (ptrv >> (pos % 8));
518 const CompressedList list_;
519 const ValueType lowerMask_;
520 detail::UpperBitsReader<CompressedList, Instructions> upper_;
523 ValueType lastValue_;
528 #endif // FOLLY_EXPERIMENTAL_ELIAS_FANO_CODING_H