@gmod/cram
Version:
read CRAM files with pure Javascript
210 lines (187 loc) • 6.22 kB
text/typescript
import CramCodec from './_base.ts'
import { CramMalformedError } from '../../errors.ts'
import type { Cursor, Cursors } from './_base.ts'
import type { HuffmanEncoding } from '../encoding.ts'
import type { CramFileBlock } from '../file.ts'
import type CramSlice from '../slice/index.ts'
/**
* Inlined getBits for huffman decoding - avoids function call overhead
*/
function getBitsInline(
data: Uint8Array,
cursor: Cursor,
numBits: number,
): number {
let { bytePosition, bitPosition } = cursor
// Fast path for single bit (common in huffman)
if (numBits === 1) {
const val = (data[bytePosition]! >> bitPosition) & 1
bitPosition -= 1
if (bitPosition < 0) {
bytePosition += 1
bitPosition = 7
}
cursor.bytePosition = bytePosition
cursor.bitPosition = bitPosition as Cursor['bitPosition']
return val
}
// General case
let val = 0
for (let i = 0; i < numBits; i++) {
val <<= 1
val |= (data[bytePosition]! >> bitPosition) & 1
bitPosition -= 1
if (bitPosition < 0) {
bytePosition += 1
bitPosition = 7
}
}
cursor.bytePosition = bytePosition
cursor.bitPosition = bitPosition
return val
}
function numberOfSetBits(ii: number) {
let i = (ii - (ii >> 1)) & 0x55555555
i = (i & 0x33333333) + ((i >> 2) & 0x33333333)
return (((i + (i >> 4)) & 0x0f0f0f0f) * 0x01010101) >> 24
}
interface Code {
bitLength: number
value: number
bitCode: number
}
export default class HuffmanIntCodec extends CramCodec<
'byte' | 'int',
HuffmanEncoding['parameters']
> {
private codes: Record<number, Code> = {}
private codeBook: Record<number, number[]> = {}
private sortedCodes: Code[] = []
// parallel flat arrays indexed by position in sortedCodes; flat arrays let
// V8 use monomorphic inline caches in the decode hot path
private valuesByIndex: number[] = []
private bitLengthsByIndex: number[] = []
// sparse lookup: bitCode -> index in sortedCodes, or -1 if absent
private bitCodeToIndex: number[] = []
constructor(
parameters: HuffmanEncoding['parameters'],
dataType: 'byte' | 'int',
) {
super(parameters, dataType)
if (!['byte', 'int'].includes(this.dataType)) {
throw new TypeError(
`${this.dataType} decoding not yet implemented by HUFFMAN_INT codec`,
)
}
this.buildCodeBook()
this.buildCodes()
this.buildCaches()
// degenerate zero-length huffman code: special-case the decoding.
// empty codeBook (no codes at all) is also valid for unused data series;
// decode() will throw 'Huffman symbol not found' if such a codec is used.
if (this.sortedCodes.length > 0 && this.sortedCodes[0]!.bitLength === 0) {
this._decode = this._decodeZeroLengthCode
}
}
buildCodeBook() {
// parse the parameters together into a `codes` data structure
const codes = new Array<{ symbol: number; bitLength: number }>(
this.parameters.numCodes,
)
for (let i = 0; i < this.parameters.numCodes; i++) {
codes[i] = {
symbol: this.parameters.symbols[i]!,
bitLength: this.parameters.bitLengths[i]!,
}
}
// sort the codes by bit length and symbol value
codes.sort((a, b) => a.bitLength - b.bitLength || a.symbol - b.symbol)
this.codeBook = {}
codes.forEach(code => {
if (!this.codeBook[code.bitLength]) {
this.codeBook[code.bitLength] = []
}
this.codeBook[code.bitLength]!.push(code.symbol)
})
}
buildCodes() {
this.codes = {} /* new TreeMap<Integer, HuffmanBitCode>(); */
let codeLength = 0
let codeValue = -1
Object.entries(this.codeBook).forEach(([bitLength, symbols]) => {
const bitLengthInt = Number.parseInt(bitLength, 10)
symbols.forEach(symbol => {
const code = {
bitLength: bitLengthInt,
value: symbol,
bitCode: 0,
}
codeValue = codeValue + 1
const delta = bitLengthInt - codeLength // new length?
codeValue = codeValue << delta // pad with 0's
code.bitCode = codeValue // calculated: huffman code
codeLength = codeLength + delta // adjust current code length
if (numberOfSetBits(codeValue) > bitLengthInt) {
throw new CramMalformedError('Symbol out of range')
}
this.codes[symbol] = code
})
})
}
buildCaches() {
this.sortedCodes = Object.values(this.codes).sort(
(a, b) => a.bitLength - b.bitLength || a.bitCode - b.bitCode,
)
this.valuesByIndex = this.sortedCodes.map(c => c.value)
this.bitLengthsByIndex = this.sortedCodes.map(c => c.bitLength)
if (this.sortedCodes.length > 0) {
let maxBitCode = 0
for (const code of this.sortedCodes) {
if (code.bitCode > maxBitCode) {
maxBitCode = code.bitCode
}
}
this.bitCodeToIndex = new Array(maxBitCode + 1).fill(-1)
for (let i = 0; i < this.sortedCodes.length; i++) {
this.bitCodeToIndex[this.sortedCodes[i]!.bitCode] = i
}
}
}
decode(
slice: CramSlice,
coreDataBlock: CramFileBlock,
_blocksByContentId: Record<number, CramFileBlock>,
cursors: Cursors,
) {
return this._decode(slice, coreDataBlock, cursors.coreBlock)
}
// the special case for zero-length codes
_decodeZeroLengthCode() {
return this.sortedCodes[0]!.value
}
_decode(_slice: CramSlice, coreDataBlock: CramFileBlock, coreCursor: Cursor) {
const input = coreDataBlock.content
let prevLen = 0
let bits = 0
for (let i = 0; i < this.sortedCodes.length; i += 1) {
const length = this.sortedCodes[i]!.bitLength
const bitsToRead = length - prevLen
if (bitsToRead > 0) {
bits <<= bitsToRead
bits |= getBitsInline(input, coreCursor, bitsToRead)
}
prevLen = length
const index = this.bitCodeToIndex[bits] ?? -1
if (index > -1 && this.bitLengthsByIndex[index] === length) {
return this.valuesByIndex[index]!
}
while (
i + 1 < this.sortedCodes.length &&
this.sortedCodes[i + 1]!.bitLength === length
) {
i += 1
}
}
throw new CramMalformedError('Huffman symbol not found.')
}
}