com.twitter.algebird.AdaptiveVector.scala Maven / Gradle / Ivy
/*
Copyright 2012 Twitter, Inc.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package com.twitter.algebird
import scala.annotation.tailrec
/**
* Some functions to create or convert AdaptiveVectors
*/
object AdaptiveVector {
/**
* When density >= this value * size, we switch to dense vectors
*/
val THRESHOLD = 0.25
def fill[V](size: Int)(sparse: V): AdaptiveVector[V] = SparseVector(Map.empty[Int, V], sparse, size)
def fromVector[V](v: Vector[V], sparseVal: V): AdaptiveVector[V] = {
if (v.size == 0) {
fill[V](0)(sparseVal)
} else {
val denseCount = v.count { _ != sparseVal }
val sz = v.size
if (denseCount < sz * THRESHOLD)
SparseVector(toMap(v, sparseVal), sparseVal, sz)
else
DenseVector(v, sparseVal, denseCount)
}
}
def fromMap[V](m: Map[Int, V], sparseVal: V, sizeOfDense: Int): AdaptiveVector[V] = {
if (m.size == 0) {
fill[V](sizeOfDense)(sparseVal)
} else {
val maxIdx = m.keys.max
require(maxIdx < sizeOfDense, "Max key (" + maxIdx + ") exceeds valid for size (" + sizeOfDense + ")")
val denseCount = m.count { _._2 != sparseVal }
if (denseCount < sizeOfDense * THRESHOLD)
SparseVector(m, sparseVal, sizeOfDense)
else
DenseVector(toVector(m, sparseVal, sizeOfDense), sparseVal, denseCount)
}
}
def toMap[V](v: AdaptiveVector[V]): Map[Int, V] =
v match {
case DenseVector(is, sv, _) => toMap(is, sv)
case SparseVector(m, _, _) => m
}
def toMap[V](iseq: IndexedSeq[V], sparse: V): Map[Int, V] =
iseq.view.zipWithIndex.filter { _._1 != sparse }.map { _.swap }.toMap
def toVector[V](m: Map[Int, V], sparse: V, size: Int): Vector[V] = {
// Mutable local variable to optimize performance
import scala.collection.mutable.Buffer
val buf = Buffer.fill[V](size)(sparse)
m.foreach { case (idx, v) => buf(idx) = v }
Vector(buf: _*)
}
def toVector[V](v: AdaptiveVector[V]): Vector[V] =
v match {
case DenseVector(is, sv, _) => is
case SparseVector(m, v, sz) => toVector(m, v, sz)
}
private def withSparse[V](v: AdaptiveVector[V], sv: V): AdaptiveVector[V] =
if (v.sparseValue == sv) v
else fromVector(toVector(v), sv)
private class AVSemigroup[V: Semigroup] extends Semigroup[AdaptiveVector[V]] {
private def valueIsNonZero(v: V): Boolean = implicitly[Semigroup[V]] match {
case m: Monoid[_] => m.isNonZero(v)
case _ => true
}
def plus(left: AdaptiveVector[V], right: AdaptiveVector[V]) = {
if (left.sparseValue != right.sparseValue) {
if (left.denseCount > right.denseCount) plus(withSparse(left, right.sparseValue), right)
else plus(left, withSparse(right, left.sparseValue))
} else {
// they have the same sparse value
val maxSize = Ordering[Int].max(left.size, right.size)
(left, right) match {
case (DenseVector(lv, ls, ld), DenseVector(rv, rs, rd)) =>
val vec = Semigroup.plus[IndexedSeq[V]](lv, rv) match {
case v: Vector[_] => v.asInstanceOf[Vector[V]]
case notV => Vector(notV: _*)
}
fromVector(vec, ls)
case _ if valueIsNonZero(left.sparseValue) =>
fromVector(Vector(Semigroup.plus(toVector(left): IndexedSeq[V],
toVector(right): IndexedSeq[V]): _*),
left.sparseValue)
case _ => // sparse is zero:
fromMap(Semigroup.plus(toMap(left), toMap(right)),
left.sparseValue,
maxSize)
}
}
}
}
private class AVMonoid[V: Monoid] extends AVSemigroup[V] with Monoid[AdaptiveVector[V]] {
val zero = AdaptiveVector.fill[V](0)(Monoid.zero[V])
override def isNonZero(v: AdaptiveVector[V]) = !isZero(v)
def isZero(v: AdaptiveVector[V]) = (v.size == 0) || {
val sparseAreZero = if (Monoid.isNonZero(v.sparseValue)) (v.denseCount == v.size) else true
sparseAreZero &&
v.denseIterator.forall { idxv => !Monoid.isNonZero(idxv._2) }
}
}
private class AVGroup[V: Group] extends AVMonoid[V] with Group[AdaptiveVector[V]] {
override def negate(v: AdaptiveVector[V]) =
fromVector(toVector(v).map(Group.negate(_)), Group.negate(v.sparseValue))
}
implicit def semigroup[V: Semigroup]: Semigroup[AdaptiveVector[V]] = new AVSemigroup[V]
implicit def monoid[V: Monoid]: Monoid[AdaptiveVector[V]] = new AVMonoid[V]
implicit def group[V: Group]: Group[AdaptiveVector[V]] = new AVGroup[V]
/*
* Equality when considering only the dense values (so size doesn't matter)
*/
def denseEquiv[V: Equiv]: Equiv[AdaptiveVector[V]] = Equiv.fromFunction[AdaptiveVector[V]] { (l, r) =>
val (lit, rit) = (l.denseIterator, r.denseIterator)
@tailrec
def iteq: Boolean =
(lit.hasNext, rit.hasNext) match {
case (true, true) =>
val (lnext, rnext) = (lit.next, rit.next)
if (lnext._1 == rnext._1 && Equiv[V].equiv(lnext._2, rnext._2))
iteq
else
false
case (false, false) => true
case _ => false
}
Equiv[V].equiv(l.sparseValue, r.sparseValue) && iteq
}
implicit def equiv[V: Equiv]: Equiv[AdaptiveVector[V]] =
Equiv.fromFunction[AdaptiveVector[V]] { (l, r) =>
(l.size == r.size) && (denseEquiv[V].equiv(l, r) ||
toVector(l).view.zip(toVector(r)).forall { case (lv, rv) => Equiv[V].equiv(lv, rv) })
}
}
/**
* An IndexedSeq that automatically switches representation between dense and sparse depending on sparsity
* Should be an efficient representation for all sizes, and it should not be necessary to special case
* immutable algebras based on the sparsity of the vectors.
*/
sealed trait AdaptiveVector[V] extends IndexedSeq[V] {
def length = size
def sparseValue: V
/** How many items are not sparse */
def denseCount: Int
def size: Int
def apply(idx: Int): V
def updated(idx: Int, v: V): AdaptiveVector[V]
/** Grow by adding count sparse values to the end */
def extend(count: Int): AdaptiveVector[V]
/** Iterator of indices and values of all non-sparse values */
def denseIterator: Iterator[(Int, V)]
/*
* Note that IndexedSeq provides hashCode and equals that
* work correctly based on length and apply.
*/
}
case class DenseVector[V](iseq: Vector[V], override val sparseValue: V, override val denseCount: Int)
extends AdaptiveVector[V] {
override def size = iseq.size
def apply(idx: Int) = iseq(idx)
def updated(idx: Int, v: V): AdaptiveVector[V] = {
val oldIsSparse = if (iseq(idx) == sparseValue) 1 else 0
val newIsSparse = if (v == sparseValue) 1 else 0
val newCount = denseCount - newIsSparse + oldIsSparse
if (denseCount < size * AdaptiveVector.THRESHOLD) {
// Go sparse
SparseVector(AdaptiveVector.toMap(iseq, sparseValue), sparseValue, size)
} else {
DenseVector(iseq.updated(idx, v), sparseValue, newCount)
}
}
def extend(cnt: Int) = {
val newSize = size + cnt
if (denseCount < newSize * AdaptiveVector.THRESHOLD) {
// Go sparse
SparseVector(AdaptiveVector.toMap(iseq, sparseValue), sparseValue, newSize)
} else {
// Stay dense
DenseVector(iseq ++ Vector.fill(cnt)(sparseValue), sparseValue, denseCount)
}
}
def denseIterator =
iseq.view.zipWithIndex.filter { _._1 != sparseValue }.map { _.swap }.iterator
}
case class SparseVector[V](map: Map[Int, V], override val sparseValue: V, override val size: Int) extends AdaptiveVector[V] {
def denseCount: Int = map.size
def apply(idx: Int) = {
require(idx >= 0 && idx < size, "Index out of range")
map.getOrElse(idx, sparseValue)
}
def updated(idx: Int, v: V): AdaptiveVector[V] = {
if (v == sparseValue) {
SparseVector(map - idx, sparseValue, size)
} else {
val newM = map + (idx -> v)
if (newM.size >= size * AdaptiveVector.THRESHOLD) {
// Go dense:
DenseVector(AdaptiveVector.toVector(newM, sparseValue, size), sparseValue, newM.size)
} else {
SparseVector(newM, sparseValue, size)
}
}
}
def extend(cnt: Int) = SparseVector(map, sparseValue, size + cnt)
private lazy val sortedList = map.toList.sortBy { _._1 }
def denseIterator = sortedList.iterator
}
© 2015 - 2025 Weber Informatics LLC | Privacy Policy