/*
- * Copyright 2015 Facebook, Inc.
+ * Copyright 2016 Facebook, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
#ifndef FOLLY_EXPERIMENTAL_ELIAS_FANO_CODING_H
#define FOLLY_EXPERIMENTAL_ELIAS_FANO_CODING_H
-#ifndef __GNUC__
-#error EliasFanoCoding.h requires GCC
-#endif
-
-#if !FOLLY_X64
-#error EliasFanoCoding.h requires x86_64
-#endif
-
#include <cstdlib>
#include <limits>
#include <type_traits>
-#include <boost/noncopyable.hpp>
-#include <glog/logging.h>
#include <folly/Bits.h>
-#include <folly/CpuId.h>
#include <folly/Likely.h>
+#include <folly/Portability.h>
#include <folly/Range.h>
+#include <folly/experimental/Instructions.h>
#include <folly/experimental/Select64.h>
+#include <glog/logging.h>
+
+#ifndef __GNUC__
+#error EliasFanoCoding.h requires GCC
+#endif
+
+#if !FOLLY_X64
+#error EliasFanoCoding.h requires x86_64
+#endif
#if __BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__
#error EliasFanoCoding.h requires little endianness
namespace folly { namespace compression {
-struct EliasFanoCompressedList {
- EliasFanoCompressedList() { }
+template <class Pointer>
+struct EliasFanoCompressedListBase {
+ EliasFanoCompressedListBase() = default;
- void free() {
- ::free(const_cast<unsigned char*>(lower.data()));
- ::free(const_cast<unsigned char*>(upper.data()));
- ::free(const_cast<unsigned char*>(skipPointers.data()));
- ::free(const_cast<unsigned char*>(forwardPointers.data()));
+ template <class OtherPointer>
+ EliasFanoCompressedListBase(
+ const EliasFanoCompressedListBase<OtherPointer>& other)
+ : size(other.size),
+ numLowerBits(other.numLowerBits),
+ data(other.data),
+ skipPointers(reinterpret_cast<Pointer>(other.skipPointers)),
+ forwardPointers(reinterpret_cast<Pointer>(other.forwardPointers)),
+ lower(reinterpret_cast<Pointer>(other.lower)),
+ upper(reinterpret_cast<Pointer>(other.upper)) { }
+
+ template <class T = Pointer>
+ auto free() -> decltype(::free(T(nullptr))) {
+ return ::free(data.data());
}
+ size_t upperSize() const { return data.end() - upper; }
+
size_t size = 0;
uint8_t numLowerBits = 0;
- // WARNING: EliasFanoCompressedList has no ownership of
- // lower, upper, skipPointers and forwardPointers.
- // The 7 bytes following the last byte of lower and upper
- // sequences should be readable.
- folly::ByteRange lower;
- folly::ByteRange upper;
+ // WARNING: EliasFanoCompressedList has no ownership of data. The 7
+ // bytes following the last byte should be readable.
+ folly::Range<Pointer> data;
- folly::ByteRange skipPointers;
- folly::ByteRange forwardPointers;
+ Pointer skipPointers = nullptr;
+ Pointer forwardPointers = nullptr;
+ Pointer lower = nullptr;
+ Pointer upper = nullptr;
};
+typedef EliasFanoCompressedListBase<const uint8_t*> EliasFanoCompressedList;
+typedef EliasFanoCompressedListBase<uint8_t*> MutableEliasFanoCompressedList;
+
template <class Value,
class SkipValue = size_t,
size_t kSkipQuantum = 0, // 0 = disabled
size_t kForwardQuantum = 0> // 0 = disabled
struct EliasFanoEncoderV2 {
static_assert(std::is_integral<Value>::value &&
- std::is_unsigned<Value>::value,
+ std::is_unsigned<Value>::value,
"Value should be unsigned integral");
typedef EliasFanoCompressedList CompressedList;
+ typedef MutableEliasFanoCompressedList MutableCompressedList;
typedef Value ValueType;
typedef SkipValue SkipValueType;
+ struct Layout;
static constexpr size_t skipQuantum = kSkipQuantum;
static constexpr size_t forwardQuantum = kForwardQuantum;
// Requires: input range (begin, end) is sorted (encoding
// crashes if it's not).
- // WARNING: encode() mallocates lower, upper, skipPointers
- // and forwardPointers. As EliasFanoCompressedList has
- // no ownership of them, you need to call free() explicitly.
+ // WARNING: encode() mallocates EliasFanoCompressedList::data. As
+ // EliasFanoCompressedList has no ownership of it, you need to call
+ // free() explicitly.
template <class RandomAccessIterator>
- static EliasFanoCompressedList encode(RandomAccessIterator begin,
- RandomAccessIterator end) {
+ static MutableCompressedList encode(RandomAccessIterator begin,
+ RandomAccessIterator end) {
if (begin == end) {
- return EliasFanoCompressedList();
+ return MutableCompressedList();
}
EliasFanoEncoderV2 encoder(end - begin, *(end - 1));
for (; begin != end; ++begin) {
return encoder.finish();
}
- EliasFanoEncoderV2(size_t size, ValueType upperBound) {
- if (size == 0) {
- return;
- }
-
- uint8_t numLowerBits = defaultNumLowerBits(upperBound, size);
-
- // This is detail::writeBits56 limitation.
- numLowerBits = std::min<uint8_t>(numLowerBits, 56);
- CHECK_LT(numLowerBits, 8 * sizeof(Value)); // As we shift by numLowerBits.
-
- // WARNING: Current read/write logic assumes that the 7 bytes
- // following the last byte of lower and upper sequences are
- // readable (stored value doesn't matter and won't be changed),
- // so we allocate additional 7B, but do not include them in size
- // of returned value.
-
- // *** Lower bits.
- const size_t lowerSize = (numLowerBits * size + 7) / 8;
- if (lowerSize > 0) { // numLowerBits != 0
- lower_ = static_cast<unsigned char*>(calloc(lowerSize + 7, 1));
- }
-
- // *** Upper bits.
- // Upper bits are stored using unary delta encoding.
- // For example, (3 5 5 9) will be encoded as 1000011001000_2.
- const size_t upperSizeBits =
- (upperBound >> numLowerBits) + // Number of 0-bits to be stored.
- size; // 1-bits.
- const size_t upperSize = (upperSizeBits + 7) / 8;
- upper_ = static_cast<unsigned char*>(calloc(upperSize + 7, 1));
-
- // *** Skip pointers.
- // Store (1-indexed) position of every skipQuantum-th
- // 0-bit in upper bits sequence.
- size_t numSkipPointers = 0;
- /* static */ if (skipQuantum != 0) {
- CHECK_LT(size, std::numeric_limits<SkipValueType>::max());
-
- // 8 * upperSize is used here instead of upperSizeBits, as that is
- // more serialization-friendly way (upperSizeBits isn't known outside of
- // this function, unlike upperSize; thus numSkipPointers could easily be
- // deduced from upperSize).
- numSkipPointers = (8 * upperSize - size) / (skipQuantum ?: 1);
- skipPointers_ = static_cast<SkipValueType*>(
- numSkipPointers == 0
- ? nullptr
- : calloc(numSkipPointers, sizeof(SkipValueType)));
- }
-
- // *** Forward pointers.
- // Store (1-indexed) position of every forwardQuantum-th
- // 1-bit in upper bits sequence.
- size_t numForwardPointers = 0;
- /* static */ if (forwardQuantum != 0) {
- CHECK_LT(upperBound >> numLowerBits,
- std::numeric_limits<SkipValueType>::max());
-
- // '?: 1' is a workaround for false 'division by zero' compile-time error.
- numForwardPointers = size / (forwardQuantum ?: 1);
- forwardPointers_ = static_cast<SkipValueType*>(
- numForwardPointers == 0
- ? nullptr
- : malloc(numForwardPointers * sizeof(SkipValueType)));
- }
-
- // *** Result.
- result_.size = size;
- result_.numLowerBits = numLowerBits;
- result_.lower.reset(lower_, lowerSize);
- result_.upper.reset(upper_, upperSize);
- result_.skipPointers.reset(
- reinterpret_cast<unsigned char*>(skipPointers_),
- numSkipPointers * sizeof(SkipValueType));
- result_.forwardPointers.reset(
- reinterpret_cast<unsigned char*>(forwardPointers_),
- numForwardPointers * sizeof(SkipValueType));
+ explicit EliasFanoEncoderV2(const MutableCompressedList& result)
+ : lower_(result.lower),
+ upper_(result.upper),
+ skipPointers_(reinterpret_cast<SkipValueType*>(
+ result.skipPointers)),
+ forwardPointers_(reinterpret_cast<SkipValueType*>(
+ result.forwardPointers)),
+ result_(result) {
+ memset(result.data.data(), 0, result.data.size());
}
+ EliasFanoEncoderV2(size_t size, ValueType upperBound)
+ : EliasFanoEncoderV2(
+ Layout::fromUpperBoundAndSize(upperBound, size).allocList()) { }
+
void add(ValueType value) {
+ CHECK_LT(value, std::numeric_limits<ValueType>::max());
CHECK_GE(value, lastValue_);
const auto numLowerBits = result_.numLowerBits;
++size_;
}
- const EliasFanoCompressedList& finish() const {
+ const MutableCompressedList& finish() const {
CHECK_EQ(size_, result_.size);
return result_;
}
size_t size_ = 0;
size_t skipPointersSize_ = 0;
- EliasFanoCompressedList result_;
+ MutableCompressedList result_;
};
-// NOTE: It's recommended to compile EF coding with -msse4.2, starting
-// with Nehalem, Intel CPUs support POPCNT instruction and gcc will emit
-// it for __builtin_popcountll intrinsic.
-// But we provide an alternative way for the client code: it can switch to
-// the appropriate version of EliasFanoReader<> in realtime (client should
-// implement this switching logic itself) by specifying instruction set to
-// use explicitly.
-namespace instructions {
-
-struct Default {
- static bool supported(const folly::CpuId& cpuId = {}) {
- return true;
- }
- static inline uint64_t popcount(uint64_t value) {
- return __builtin_popcountll(value);
- }
- static inline int ctz(uint64_t value) {
- DCHECK_GT(value, 0);
- return __builtin_ctzll(value);
+template <class Value,
+ class SkipValue,
+ size_t kSkipQuantum,
+ size_t kForwardQuantum>
+struct EliasFanoEncoderV2<Value,
+ SkipValue,
+ kSkipQuantum,
+ kForwardQuantum>::Layout {
+ static Layout fromUpperBoundAndSize(size_t upperBound, size_t size) {
+ // numLowerBits can be at most 56 because of detail::writeBits56.
+ const uint8_t numLowerBits = std::min(defaultNumLowerBits(upperBound,
+ size),
+ uint8_t(56));
+ // *** Upper bits.
+ // Upper bits are stored using unary delta encoding.
+ // For example, (3 5 5 9) will be encoded as 1000011001000_2.
+ const size_t upperSizeBits =
+ (upperBound >> numLowerBits) + // Number of 0-bits to be stored.
+ size; // 1-bits.
+ const size_t upper = (upperSizeBits + 7) / 8;
+
+ // *** Validity checks.
+ // Shift by numLowerBits must be valid.
+ CHECK_LT(numLowerBits, 8 * sizeof(Value));
+ CHECK_LT(size, std::numeric_limits<SkipValueType>::max());
+ CHECK_LT(upperBound >> numLowerBits,
+ std::numeric_limits<SkipValueType>::max());
+
+ return fromInternalSizes(numLowerBits, upper, size);
}
- static inline uint64_t blsr(uint64_t value) {
- return value & (value - 1);
+
+ static Layout fromInternalSizes(uint8_t numLowerBits,
+ size_t upper,
+ size_t size) {
+ Layout layout;
+ layout.size = size;
+ layout.numLowerBits = numLowerBits;
+
+ layout.lower = (numLowerBits * size + 7) / 8;
+ layout.upper = upper;
+
+ // *** Skip pointers.
+ // Store (1-indexed) position of every skipQuantum-th
+ // 0-bit in upper bits sequence.
+ /* static */ if (skipQuantum != 0) {
+ // 8 * upper is used here instead of upperSizeBits, as that is
+ // more serialization-friendly way (upperSizeBits doesn't need
+ // to be known by this function, unlike upper).
+
+ // '?: 1' is a workaround for false 'division by zero'
+ // compile-time error.
+ size_t numSkipPointers = (8 * upper - size) / (skipQuantum ?: 1);
+ layout.skipPointers = numSkipPointers * sizeof(SkipValueType);
+ }
+
+ // *** Forward pointers.
+ // Store (1-indexed) position of every forwardQuantum-th
+ // 1-bit in upper bits sequence.
+ /* static */ if (forwardQuantum != 0) {
+ size_t numForwardPointers = size / (forwardQuantum ?: 1);
+ layout.forwardPointers = numForwardPointers * sizeof(SkipValueType);
+ }
+
+ return layout;
}
-};
-struct Nehalem : public Default {
- static bool supported(const folly::CpuId& cpuId = {}) {
- return cpuId.popcnt();
+ size_t bytes() const {
+ return lower + upper + skipPointers + forwardPointers;
}
- static inline uint64_t popcount(uint64_t value) {
- // POPCNT is supported starting with Intel Nehalem, AMD K10.
- uint64_t result;
- asm ("popcntq %1, %0" : "=r" (result) : "r" (value));
+
+ template <class Range>
+ EliasFanoCompressedListBase<typename Range::iterator>
+ openList(Range& buf) const {
+ EliasFanoCompressedListBase<typename Range::iterator> result;
+ result.size = size;
+ result.numLowerBits = numLowerBits;
+ result.data = buf.subpiece(0, bytes());
+
+ auto advance = [&] (size_t n) {
+ auto begin = buf.data();
+ buf.advance(n);
+ return begin;
+ };
+
+ result.skipPointers = advance(skipPointers);
+ result.forwardPointers = advance(forwardPointers);
+ result.lower = advance(lower);
+ result.upper = advance(upper);
+
return result;
}
-};
-struct Haswell : public Nehalem {
- static bool supported(const folly::CpuId& cpuId = {}) {
- return Nehalem::supported(cpuId) && cpuId.bmi1();
- }
- static inline uint64_t blsr(uint64_t value) {
- // BMI1 is supported starting with Intel Haswell, AMD Piledriver.
- // BLSR combines two instuctions into one and reduces register pressure.
- uint64_t result;
- asm ("blsrq %1, %0" : "=r" (result) : "r" (value));
- return result;
+ MutableCompressedList allocList() const {
+ uint8_t* buf = nullptr;
+ // WARNING: Current read/write logic assumes that the 7 bytes
+ // following the last byte of lower and upper sequences are
+ // readable (stored value doesn't matter and won't be changed), so
+ // we allocate additional 7 bytes, but do not include them in size
+ // of returned value.
+ if (size > 0) {
+ buf = static_cast<uint8_t*>(malloc(bytes() + 7));
+ }
+ folly::MutableByteRange bufRange(buf, bytes());
+ return openList(bufRange);
}
-};
-} // namespace instructions
+ size_t size = 0;
+ uint8_t numLowerBits = 0;
+
+ // Sizes in bytes.
+ size_t lower = 0;
+ size_t upper = 0;
+ size_t skipPointers = 0;
+ size_t forwardPointers = 0;
+};
namespace detail {
public:
typedef typename Encoder::ValueType ValueType;
- explicit UpperBitsReader(const EliasFanoCompressedList& list)
- : forwardPointers_(list.forwardPointers.data()),
- skipPointers_(list.skipPointers.data()),
- start_(list.upper.data()) {
+ explicit UpperBitsReader(const typename Encoder::CompressedList& list)
+ : forwardPointers_(list.forwardPointers),
+ skipPointers_(list.skipPointers),
+ start_(list.upper) {
reset();
}
return skipToNext(v);
}
+ ValueType previousValue() const {
+ DCHECK_NE(position(), -1);
+ DCHECK_GT(position(), 0);
+
+ size_t outer = outer_;
+ block_t block = folly::loadUnaligned<block_t>(start_ + outer);
+ block &= (block_t(1) << inner_) - 1;
+
+ while (UNLIKELY(block == 0)) {
+ DCHECK_GE(outer, sizeof(block_t));
+ outer -= sizeof(block_t);
+ block = folly::loadUnaligned<block_t>(start_ + outer);
+ }
+
+ auto inner = 8 * sizeof(block_t) - 1 - Instructions::clz(block);
+ return static_cast<ValueType>(8 * outer + inner - (position_ - 1));
+ }
+
+ void setDone(size_t endPos) {
+ position_ = endPos;
+ }
+
private:
ValueType setValue() {
value_ = static_cast<ValueType>(8 * outer_ + inner_ - position_);
block_ &= ~((block_t(1) << (dest % 8)) - 1);
}
- typedef unsigned long long block_t;
+ typedef uint64_t block_t;
const unsigned char* const forwardPointers_;
const unsigned char* const skipPointers_;
const unsigned char* const start_;
} // namespace detail
+// If kUnchecked = true the caller must guarantee that all the
+// operations return valid elements, i.e., they would never return
+// false if checked.
template <class Encoder,
- class Instructions = instructions::Default>
-class EliasFanoReader : private boost::noncopyable {
+ class Instructions = instructions::Default,
+ bool kUnchecked = false>
+class EliasFanoReader {
public:
typedef Encoder EncoderType;
typedef typename Encoder::ValueType ValueType;
- explicit EliasFanoReader(const EliasFanoCompressedList& list)
- : list_(list),
- lowerMask_((ValueType(1) << list_.numLowerBits) - 1),
- upper_(list_) {
+ explicit EliasFanoReader(const typename Encoder::CompressedList& list)
+ : size_(list.size),
+ lower_(list.lower),
+ upper_(list),
+ lowerMask_((ValueType(1) << list.numLowerBits) - 1),
+ numLowerBits_(list.numLowerBits) {
DCHECK(Instructions::supported());
// To avoid extra branching during skipTo() while reading
// upper sequence we need to know the last element.
- if (UNLIKELY(list_.size == 0)) {
+ // If kUnchecked == true, we do not check that skipTo() is called
+ // within the bounds, so we can avoid initializing lastValue_.
+ if (kUnchecked || UNLIKELY(list.size == 0)) {
lastValue_ = 0;
return;
}
- ValueType lastUpperValue = 8 * list_.upper.size() - list_.size;
- auto it = list_.upper.end() - 1;
+ ValueType lastUpperValue = 8 * list.upperSize() - size_;
+ auto it = list.upper + list.upperSize() - 1;
DCHECK_NE(*it, 0);
lastUpperValue -= 8 - folly::findLastSet(*it);
- lastValue_ = readLowerPart(list_.size - 1) |
- (lastUpperValue << list_.numLowerBits);
+ lastValue_ = readLowerPart(size_ - 1) | (lastUpperValue << numLowerBits_);
}
void reset() {
upper_.reset();
- progress_ = 0;
- value_ = 0;
+ value_ = kInvalidValue;
}
bool next() {
- if (UNLIKELY(progress_ >= list_.size)) {
+ if (!kUnchecked && UNLIKELY(position() + 1 >= size_)) {
return setDone();
}
- value_ = readLowerPart(progress_) |
- (upper_.next() << list_.numLowerBits);
- ++progress_;
+ upper_.next();
+ value_ = readLowerPart(upper_.position()) |
+ (upper_.value() << numLowerBits_);
return true;
}
bool skip(size_t n) {
CHECK_GT(n, 0);
- progress_ += n;
- if (LIKELY(progress_ <= list_.size)) {
+ if (kUnchecked || LIKELY(position() + n < size_)) {
if (LIKELY(n < kLinearScanThreshold)) {
for (size_t i = 0; i < n; ++i) upper_.next();
} else {
upper_.skip(n);
}
- value_ = readLowerPart(progress_ - 1) |
- (upper_.value() << list_.numLowerBits);
+ value_ = readLowerPart(upper_.position()) |
+ (upper_.value() << numLowerBits_);
return true;
}
}
bool skipTo(ValueType value) {
- DCHECK_GE(value, value_);
- if (value <= value_) {
- return true;
- } else if (value > lastValue_) {
+ // Also works when value_ == kInvalidValue.
+ if (value != kInvalidValue) { DCHECK_GE(value + 1, value_ + 1); }
+
+ if (!kUnchecked && value > lastValue_) {
return setDone();
+ } else if (value == value_) {
+ return true;
}
- size_t upperValue = (value >> list_.numLowerBits);
+ size_t upperValue = (value >> numLowerBits_);
size_t upperSkip = upperValue - upper_.value();
// The average density of ones in upper bits is 1/2.
// LIKELY here seems to make things worse, even for small skips.
}
bool jump(size_t n) {
- if (LIKELY(n - 1 < list_.size)) { // n > 0 && n <= list_.size
- progress_ = n;
- value_ = readLowerPart(n - 1) | (upper_.jump(n) << list_.numLowerBits);
- return true;
- } else if (n == 0) {
- reset();
+ if (LIKELY(n < size_)) { // Also checks that n != -1.
+ value_ = readLowerPart(n) | (upper_.jump(n + 1) << numLowerBits_);
return true;
}
return setDone();
}
bool jumpTo(ValueType value) {
- if (value <= 0) {
- reset();
- return true;
- } else if (value > lastValue_) {
+ if (!kUnchecked && value > lastValue_) {
return setDone();
}
- upper_.jumpToNext(value >> list_.numLowerBits);
+ upper_.jumpToNext(value >> numLowerBits_);
iterateTo(value);
return true;
}
- size_t size() const { return list_.size; }
+ ValueType previousValue() const {
+ DCHECK_GT(position(), 0);
+ DCHECK_LT(position(), size());
+ return readLowerPart(upper_.position() - 1) |
+ (upper_.previousValue() << numLowerBits_);
+ }
- size_t position() const { return progress_ - 1; }
- ValueType value() const { return value_; }
+ size_t size() const { return size_; }
+
+ bool valid() const {
+ return position() < size(); // Also checks that position() != -1.
+ }
+
+ size_t position() const { return upper_.position(); }
+ ValueType value() const {
+ DCHECK(valid());
+ return value_;
+ }
private:
+ constexpr static ValueType kInvalidValue =
+ std::numeric_limits<ValueType>::max(); // Must hold kInvalidValue + 1 == 0.
+
bool setDone() {
- value_ = std::numeric_limits<ValueType>::max();
- progress_ = list_.size + 1;
+ value_ = kInvalidValue;
+ upper_.setDone(size_);
return false;
}
ValueType readLowerPart(size_t i) const {
- DCHECK_LT(i, list_.size);
- const size_t pos = i * list_.numLowerBits;
- const unsigned char* ptr = list_.lower.data() + (pos / 8);
+ DCHECK_LT(i, size_);
+ const size_t pos = i * numLowerBits_;
+ const unsigned char* ptr = lower_ + (pos / 8);
const uint64_t ptrv = folly::loadUnaligned<uint64_t>(ptr);
return lowerMask_ & (ptrv >> (pos % 8));
}
void iterateTo(ValueType value) {
while (true) {
value_ = readLowerPart(upper_.position()) |
- (upper_.value() << list_.numLowerBits);
+ (upper_.value() << numLowerBits_);
if (LIKELY(value_ >= value)) break;
upper_.next();
}
- progress_ = upper_.position() + 1;
}
constexpr static size_t kLinearScanThreshold = 8;
- const EliasFanoCompressedList list_;
- const ValueType lowerMask_;
+ size_t size_;
+ const uint8_t* lower_;
detail::UpperBitsReader<Encoder, Instructions> upper_;
- size_t progress_ = 0;
- ValueType value_ = 0;
+ const ValueType lowerMask_;
+ ValueType value_ = kInvalidValue;
ValueType lastValue_;
+ uint8_t numLowerBits_;
};
}} // namespaces
projects / folly.git / blobdiff