Data Structures

Kmem is bit, array and fast memory utilities library for multiplatform Kotlin 1.3

Build Status Maven Version

Table of contents:

Using with gradle

Requires Gradle 4.7 (JVM 8~10) for building and Kotlin >=1.3.11 for running:

def kmemVersion = "1.0.0"

repositories {
    maven { url "" }

dependencies {
    // For multiplatform projects
    implementation "com.soywiz:kmem:$kmemVersion"
    // For JVM/Android only
    implementation "com.soywiz:kmem-jvm:$kmemVersion"
    // For JS only
    implementation "com.soywiz:kmem-js:$kmemVersion"

// settigs.gradle

arraycopy and arrayfill

Kotlin 1.3 introduces an Array.copyInto extension, but the signature is a bit confusing. KMem introduces a wrapper around it that makes the signature more familiar (like Java’s System.arraycopy). This is the kind of function that works better like a global or static function since there is no obvious receiver even with the “To” suffix since the receiver should be a couple of parameters or even three considering the length and making it a extension method breaks the symmetry unless the count would be the receiver.

This signature is easy to remember: SRC -> DST, HOW MUCH. Both, src and dst are formed from a couple of parameters (an array, and a position). So:

arraycopy(src: Array, srcPos: Int, dst: Array, dstPos: Int, count: Int)

As for Kotlin 1.3, no Array.fill is provided. For fill, there is an obvious receiver, so it is exposed as an extension method instead. To make it symmetric with arraycopy, it is exposed in two flavors:

arrayfill(array: Array<T>, value: T, start: Int = 0, end: Int = this.size)
Array<T>.fill(value: T, start: Int = 0, end: Int = this.size)

UByteArrayInt and FloatArrayFromIntArray

Sometimes we just want to use some arrays like ByeArray as unsigned or IntArray as if the elements were float. You can do it manually by converting the values for each array access.

Since Kotlin 1.3, there is an UByteArray, but the problem is that it returns UByte values that are inconvenient to use in some cases since there is no autoconversion for fitting values.

UByteArrayInt works like UByte but receives and returns Int values instead, and just considers the lower 8 bits.

FloatArrayFromIntArray works by reinterpreting backed Int as Float in each access.

val uba = ByteArray(16).asUByteArrayInt()
val ba = uba.asByteArray()

uba[0] = 255
assertEquals(-1, ba[0])
assertEquals(255, uba[0])

val fa = IntArray(16).asFloatArray()
val ia = fa.asIntArray()

fa[0] = 1f
assertEquals(0x3f800000, ia[0])

MemBuffer, DataBuffer, Int8Buffer, Int16Buffer, Int32Buffer, Float32Buffer, Float64Buffer

Analogous to JavaScript typed arrays. MemBuffer is like an ArrayBuffer. DataBuffer (DataBuffer) and *TypeBuffer, works like viewws of a single ArrayBuffer. And they do a single thing: provide a view of the data in a fast way in an immutable way. It doesn’t work like the JVM Buffers that do too much in a mutable way (providing the data and mutating a pointer).


FBuffer combines all the *Buffer classes in a single class to provide a DataBuffer like class but with faster aligned access in all the targets.

ByteArrayBuilder, ByteArrayBuilderLE, ByteArrayBuilderBE

Analogous to StringBuilder, this class allows to generate a ByteArray by appending data. It also provides a buildByteArray method and two variants for simpler Little Endian and BigEndian writting (buildByteArrayLE and buildByteArrayBE).

val byteArray = buildByteArray {
    append(byteArrayOf(3, 4, 5))
assertEquals(9, byteArray.size)
assertEquals(listOf(1, 2, 3, 4, 5, 6, 0, 0, 0), { it.toInt() })

It provides append methods for appending bytes and s* and f* methods to append several primitive packed types.

ByteArrayReader, ByteArrayReaderLE, ByteArrayReaderBE

ByteArrayReader allows to sequentially read a ByteArray. For convenience it provides ByteArray.reader and extension methods.

val byteArray = buildByteArray { f32BE(1f).f32LE(2f) } {
    assertEquals(1f, f32BE())
    assertEquals(2f, f32LE())
    assertEquals(0, remaining)

It provides s* and f* methods to read several primitive packed types., ByteArray.write

Similar to ByteArrayBuilder and ByteArrayReader, Kmem exposes several methods for reading values directly from a ByteArray by just providing the index of the array to read from without any kind of allocation or intermediary object.

It provides variants for reading and writing Byte, Short, Char, Int24, Int, Long, Float16, Float, Double, in signed and unsigned forms, in little and big endian.


Float16 is an inline class backed by an Int that represents a 16-bit floating point. This representation is not used for computations traditionally except in GPUs, but can be used to store some values in a floating point format in half the size with less precission.

val fp16 = Float16.fromBits(0x1F00)
val fp16AsInt = fp16.toBits()
val fp32 = fp16.toFloat()

Bit tools

Int,Float,Long,Double reinterpret*

Kotlin has Float.Companion.fromBits and Float.toBits. But sometimes is more obvious to use reinterpret* extension methods instead:

inline fun Float.reinterpretAsInt() = this.toBits()
inline fun Int.reinterpretAsFloat() = Float.fromBits(this)
inline fun Double.reinterpretAsLong() = this.toBits()
inline fun Long.reinterpretAsDouble() = Double.fromBits(this)


Kotlin defines shl, shr and ushr infix functions for primitive integral types. But no way to rotate bits without losing its bit count. These extension functions do exactly that:

fun Int.rotateLeft(bits: Int): Int = (this shl bits) or (this ushr (32 - bits))
fun Int.rotateRight(bits: Int): Int = (this shl (32 - bits)) or (this ushr bits)

reverseBytes, reverseBits

When working with several endians (either little or big), you might want to convert between endians. And that’s exactly what reverseBytes extension function does.

assertEquals(0x04030201, 0x01020304.reverseBytes())

reverseBits does a reversing too but instead of doing it in octets it does the reversing bit to bit.

countLeadingZeros, countTrailingZeros, countLeadingOnes, countTrailingOnes

Some algorithms might require to count the number of leading zeros of a number in a binary representation. For example to compute ilog2. countLeadingZeros is usually a processor intrinsic and it is defined in JS (Math.clz32) and JVM (java.lang.Integer.numberOfLeadingZeros). Kmem uses actual to provide the fastest implementation for each platform.


Sometimes we have an unsigned value, and we want to make it signed by extending the last bit in a complement-of-two representation:

fun Int.signExtend(bits: Int): Int = (this shl (32 - bits)) shr (32 - bits) // Int.SIZE_BITS
val signExtended = 0xFFFF.signExtend(16) // -1

mask, extract, extractSigned, insert, hasFlags

To work at the bit level we might want to extract part of the bits of a number, or construct a new number replacing some bits, or verifying if a number has all certain bits set.

Int.mask allow to construct a bit mask of a certain number:

assertEquals(0b0, 0.mask())
assertEquals(0b1, 1.mask())
assertEquals(0b11, 2.mask())
assertEquals(0b111, 3.mask())

Int.extract has several signatures to extract either a single bit, or several bits at once from an integer:

assertEquals(true,   0b00111010.extract(1))
assertEquals(0b1101, 0b00111010.extract(1, 4))

Int.insert allows to construct a new integer from one changing a set of bits:

assertEquals(0b11010010,  0b00000010.insert(0b1101, 4))

Int.hasFlags allows to check if several bits are set at once:

assertEquals(true,  0b00111010.hasFlags(0b00101010))
assertEquals(false, 0b00111010.hasFlags(0b10000010))

Numeric tools


Methods to convert a Double or Float value into an Int in one step using different methods for approximating for the integral value.

clamp, convertRange, convertRangeClamped

convertRange, convertRangeClamped extension methods allows to convert one value from one range of values to another interpolating the value. The clamped version ensures that the converted value is not outside the target bounds.

clamp extension methods allows to generate a value between a range of values. When the value is outside bounds, it returns the nearest upper or lower bound that is inside the range.

isAlmostZero, isNanOrInfinite

Extensions for Float and Double to determine if the number is very near to zero or is not a finite number.

isOdd, isEven

Integral extension properties to determine if an integer is odd or even.


umod works like the remaining % operator, but works for negative values too, wrapping around the maximum type generating an unsigned value.

nextAlignedTo, prevAlignedTo, isAlignedTo, nextPowerOfTwo, prevPowerOfTwo, isPowerOfTwo

When working with memory, it is pretty usual to need to compute aligned addresses or power of two values from near values to perform some kind of optimizations (like reading sectors), create power of two textures, or work with instructions that require to be 16-byte aligned (like some aligned SIMD instructions or plain instructions in some processors).

assertEquals(20, 13.nextAlignedTo(10))
assertEquals(10, 13.prevAlignedTo(10))
assertEquals(false, 13.isAlignedTo(10))

assertEquals(32, 20.nextPowerOfTwo)
assertEquals(16, 20.prevPowerOfTwo)
assertEquals(false, 20.isPowerOfTwo)