// Copyright 2019+ Klaus Post. All rights reserved. // License information can be found in the LICENSE file. // Based on work by Yann Collet, released under BSD License. package zstd import ( "errors" "fmt" "math" ) const ( // For encoding we only support up to maxEncTableLog = 8 maxEncTablesize = 1 << maxTableLog maxEncTableMask = (1 << maxTableLog) - 1 minEncTablelog = 5 maxEncSymbolValue = maxMatchLengthSymbol ) // Scratch provides temporary storage for compression and decompression. type fseEncoder struct { symbolLen uint16 // Length of active part of the symbol table. actualTableLog uint8 // Selected tablelog. ct cTable // Compression tables. maxCount int // count of the most probable symbol zeroBits bool // no bits has prob > 50%. clearCount bool // clear count useRLE bool // This encoder is for RLE preDefined bool // This encoder is predefined. reUsed bool // Set to know when the encoder has been reused. rleVal uint8 // RLE Symbol maxBits uint8 // Maximum output bits after transform. // TODO: Technically zstd should be fine with 64 bytes. count [256]uint32 norm [256]int16 } // cTable contains tables used for compression. type cTable struct { tableSymbol []byte stateTable []uint16 symbolTT []symbolTransform } // symbolTransform contains the state transform for a symbol. type symbolTransform struct { deltaNbBits uint32 deltaFindState int16 outBits uint8 } // String prints values as a human readable string. func (s symbolTransform) String() string { return fmt.Sprintf("{deltabits: %08x, findstate:%d outbits:%d}", s.deltaNbBits, s.deltaFindState, s.outBits) } // Histogram allows to populate the histogram and skip that step in the compression, // It otherwise allows to inspect the histogram when compression is done. // To indicate that you have populated the histogram call HistogramFinished // with the value of the highest populated symbol, as well as the number of entries // in the most populated entry. These are accepted at face value. func (s *fseEncoder) Histogram() *[256]uint32 { return &s.count } // HistogramFinished can be called to indicate that the histogram has been populated. // maxSymbol is the index of the highest set symbol of the next data segment. // maxCount is the number of entries in the most populated entry. // These are accepted at face value. func (s *fseEncoder) HistogramFinished(maxSymbol uint8, maxCount int) { s.maxCount = maxCount s.symbolLen = uint16(maxSymbol) + 1 s.clearCount = maxCount != 0 } // allocCtable will allocate tables needed for compression. // If existing tables a re big enough, they are simply re-used. func (s *fseEncoder) allocCtable() { tableSize := 1 << s.actualTableLog // get tableSymbol that is big enough. if cap(s.ct.tableSymbol) < tableSize { s.ct.tableSymbol = make([]byte, tableSize) } s.ct.tableSymbol = s.ct.tableSymbol[:tableSize] ctSize := tableSize if cap(s.ct.stateTable) < ctSize { s.ct.stateTable = make([]uint16, ctSize) } s.ct.stateTable = s.ct.stateTable[:ctSize] if cap(s.ct.symbolTT) < 256 { s.ct.symbolTT = make([]symbolTransform, 256) } s.ct.symbolTT = s.ct.symbolTT[:256] } // buildCTable will populate the compression table so it is ready to be used. func (s *fseEncoder) buildCTable() error { tableSize := uint32(1 << s.actualTableLog) highThreshold := tableSize - 1 var cumul [256]int16 s.allocCtable() tableSymbol := s.ct.tableSymbol[:tableSize] // symbol start positions { cumul[0] = 0 for ui, v := range s.norm[:s.symbolLen-1] { u := byte(ui) // one less than reference if v == -1 { // Low proba symbol cumul[u+1] = cumul[u] + 1 tableSymbol[highThreshold] = u highThreshold-- } else { cumul[u+1] = cumul[u] + v } } // Encode last symbol separately to avoid overflowing u u := int(s.symbolLen - 1) v := s.norm[s.symbolLen-1] if v == -1 { // Low proba symbol cumul[u+1] = cumul[u] + 1 tableSymbol[highThreshold] = byte(u) highThreshold-- } else { cumul[u+1] = cumul[u] + v } if uint32(cumul[s.symbolLen]) != tableSize { return fmt.Errorf("internal error: expected cumul[s.symbolLen] (%d) == tableSize (%d)", cumul[s.symbolLen], tableSize) } cumul[s.symbolLen] = int16(tableSize) + 1 } // Spread symbols s.zeroBits = false { step := tableStep(tableSize) tableMask := tableSize - 1 var position uint32 // if any symbol > largeLimit, we may have 0 bits output. largeLimit := int16(1 << (s.actualTableLog - 1)) for ui, v := range s.norm[:s.symbolLen] { symbol := byte(ui) if v > largeLimit { s.zeroBits = true } for nbOccurrences := int16(0); nbOccurrences < v; nbOccurrences++ { tableSymbol[position] = symbol position = (position + step) & tableMask for position > highThreshold { position = (position + step) & tableMask } /* Low proba area */ } } // Check if we have gone through all positions if position != 0 { return errors.New("position!=0") } } // Build table table := s.ct.stateTable { tsi := int(tableSize) for u, v := range tableSymbol { // TableU16 : sorted by symbol order; gives next state value table[cumul[v]] = uint16(tsi + u) cumul[v]++ } } // Build Symbol Transformation Table { total := int16(0) symbolTT := s.ct.symbolTT[:s.symbolLen] tableLog := s.actualTableLog tl := (uint32(tableLog) << 16) - (1 << tableLog) for i, v := range s.norm[:s.symbolLen] { switch v { case 0: case -1, 1: symbolTT[i].deltaNbBits = tl symbolTT[i].deltaFindState = total - 1 total++ default: maxBitsOut := uint32(tableLog) - highBit(uint32(v-1)) minStatePlus := uint32(v) << maxBitsOut symbolTT[i].deltaNbBits = (maxBitsOut << 16) - minStatePlus symbolTT[i].deltaFindState = total - v total += v } } if total != int16(tableSize) { return fmt.Errorf("total mismatch %d (got) != %d (want)", total, tableSize) } } return nil } var rtbTable = [...]uint32{0, 473195, 504333, 520860, 550000, 700000, 750000, 830000} func (s *fseEncoder) setRLE(val byte) { s.allocCtable() s.actualTableLog = 0 s.ct.stateTable = s.ct.stateTable[:1] s.ct.symbolTT[val] = symbolTransform{ deltaFindState: 0, deltaNbBits: 0, } if debugEncoder { println("setRLE: val", val, "symbolTT", s.ct.symbolTT[val]) } s.rleVal = val s.useRLE = true } // setBits will set output bits for the transform. // if nil is provided, the number of bits is equal to the index. func (s *fseEncoder) setBits(transform []byte) { if s.reUsed || s.preDefined { return } if s.useRLE { if transform == nil { s.ct.symbolTT[s.rleVal].outBits = s.rleVal s.maxBits = s.rleVal return } s.maxBits = transform[s.rleVal] s.ct.symbolTT[s.rleVal].outBits = s.maxBits return } if transform == nil { for i := range s.ct.symbolTT[:s.symbolLen] { s.ct.symbolTT[i].outBits = uint8(i) } s.maxBits = uint8(s.symbolLen - 1) return } s.maxBits = 0 for i, v := range transform[:s.symbolLen] { s.ct.symbolTT[i].outBits = v if v > s.maxBits { // We could assume bits always going up, but we play safe. s.maxBits = v } } } // normalizeCount will normalize the count of the symbols so // the total is equal to the table size. // If successful, compression tables will also be made ready. func (s *fseEncoder) normalizeCount(length int) error { if s.reUsed { return nil } s.optimalTableLog(length) var ( tableLog = s.actualTableLog scale = 62 - uint64(tableLog) step = (1 << 62) / uint64(length) vStep = uint64(1) << (scale - 20) stillToDistribute = int16(1 << tableLog) largest int largestP int16 lowThreshold = (uint32)(length >> tableLog) ) if s.maxCount == length { s.useRLE = true return nil } s.useRLE = false for i, cnt := range s.count[:s.symbolLen] { // already handled // if (count[s] == s.length) return 0; /* rle special case */ if cnt == 0 { s.norm[i] = 0 continue } if cnt <= lowThreshold { s.norm[i] = -1 stillToDistribute-- } else { proba := (int16)((uint64(cnt) * step) >> scale) if proba < 8 { restToBeat := vStep * uint64(rtbTable[proba]) v := uint64(cnt)*step - (uint64(proba) << scale) if v > restToBeat { proba++ } } if proba > largestP { largestP = proba largest = i } s.norm[i] = proba stillToDistribute -= proba } } if -stillToDistribute >= (s.norm[largest] >> 1) { // corner case, need another normalization method err := s.normalizeCount2(length) if err != nil { return err } if debugAsserts { err = s.validateNorm() if err != nil { return err } } return s.buildCTable() } s.norm[largest] += stillToDistribute if debugAsserts { err := s.validateNorm() if err != nil { return err } } return s.buildCTable() } // Secondary normalization method. // To be used when primary method fails. func (s *fseEncoder) normalizeCount2(length int) error { const notYetAssigned = -2 var ( distributed uint32 total = uint32(length) tableLog = s.actualTableLog lowThreshold = total >> tableLog lowOne = (total * 3) >> (tableLog + 1) ) for i, cnt := range s.count[:s.symbolLen] { if cnt == 0 { s.norm[i] = 0 continue } if cnt <= lowThreshold { s.norm[i] = -1 distributed++ total -= cnt continue } if cnt <= lowOne { s.norm[i] = 1 distributed++ total -= cnt continue } s.norm[i] = notYetAssigned } toDistribute := (1 << tableLog) - distributed if (total / toDistribute) > lowOne { // risk of rounding to zero lowOne = (total * 3) / (toDistribute * 2) for i, cnt := range s.count[:s.symbolLen] { if (s.norm[i] == notYetAssigned) && (cnt <= lowOne) { s.norm[i] = 1 distributed++ total -= cnt continue } } toDistribute = (1 << tableLog) - distributed } if distributed == uint32(s.symbolLen)+1 { // all values are pretty poor; // probably incompressible data (should have already been detected); // find max, then give all remaining points to max var maxV int var maxC uint32 for i, cnt := range s.count[:s.symbolLen] { if cnt > maxC { maxV = i maxC = cnt } } s.norm[maxV] += int16(toDistribute) return nil } if total == 0 { // all of the symbols were low enough for the lowOne or lowThreshold for i := uint32(0); toDistribute > 0; i = (i + 1) % (uint32(s.symbolLen)) { if s.norm[i] > 0 { toDistribute-- s.norm[i]++ } } return nil } var ( vStepLog = 62 - uint64(tableLog) mid = uint64((1 << (vStepLog - 1)) - 1) rStep = (((1 << vStepLog) * uint64(toDistribute)) + mid) / uint64(total) // scale on remaining tmpTotal = mid ) for i, cnt := range s.count[:s.symbolLen] { if s.norm[i] == notYetAssigned { var ( end = tmpTotal + uint64(cnt)*rStep sStart = uint32(tmpTotal >> vStepLog) sEnd = uint32(end >> vStepLog) weight = sEnd - sStart ) if weight < 1 { return errors.New("weight < 1") } s.norm[i] = int16(weight) tmpTotal = end } } return nil } // optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog func (s *fseEncoder) optimalTableLog(length int) { tableLog := uint8(maxEncTableLog) minBitsSrc := highBit(uint32(length)) + 1 minBitsSymbols := highBit(uint32(s.symbolLen-1)) + 2 minBits := uint8(minBitsSymbols) if minBitsSrc < minBitsSymbols { minBits = uint8(minBitsSrc) } maxBitsSrc := uint8(highBit(uint32(length-1))) - 2 if maxBitsSrc < tableLog { // Accuracy can be reduced tableLog = maxBitsSrc } if minBits > tableLog { tableLog = minBits } // Need a minimum to safely represent all symbol values if tableLog < minEncTablelog { tableLog = minEncTablelog } if tableLog > maxEncTableLog { tableLog = maxEncTableLog } s.actualTableLog = tableLog } // validateNorm validates the normalized histogram table. func (s *fseEncoder) validateNorm() (err error) { var total int for _, v := range s.norm[:s.symbolLen] { if v >= 0 { total += int(v) } else { total -= int(v) } } defer func() { if err == nil { return } fmt.Printf("selected TableLog: %d, Symbol length: %d\n", s.actualTableLog, s.symbolLen) for i, v := range s.norm[:s.symbolLen] { fmt.Printf("%3d: %5d -> %4d \n", i, s.count[i], v) } }() if total != (1 << s.actualTableLog) { return fmt.Errorf("warning: Total == %d != %d", total, 1<> 3) + 3 + 2 // Write Table Size bitStream = uint32(tableLog - minEncTablelog) bitCount = uint(4) remaining = int16(tableSize + 1) /* +1 for extra accuracy */ threshold = int16(tableSize) nbBits = uint(tableLog + 1) outP = len(out) ) if cap(out) < outP+maxHeaderSize { out = append(out, make([]byte, maxHeaderSize*3)...) out = out[:len(out)-maxHeaderSize*3] } out = out[:outP+maxHeaderSize] // stops at 1 for remaining > 1 { if previous0 { start := charnum for s.norm[charnum] == 0 { charnum++ } for charnum >= start+24 { start += 24 bitStream += uint32(0xFFFF) << bitCount out[outP] = byte(bitStream) out[outP+1] = byte(bitStream >> 8) outP += 2 bitStream >>= 16 } for charnum >= start+3 { start += 3 bitStream += 3 << bitCount bitCount += 2 } bitStream += uint32(charnum-start) << bitCount bitCount += 2 if bitCount > 16 { out[outP] = byte(bitStream) out[outP+1] = byte(bitStream >> 8) outP += 2 bitStream >>= 16 bitCount -= 16 } } count := s.norm[charnum] charnum++ max := (2*threshold - 1) - remaining if count < 0 { remaining += count } else { remaining -= count } count++ // +1 for extra accuracy if count >= threshold { count += max // [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ } bitStream += uint32(count) << bitCount bitCount += nbBits if count < max { bitCount-- } previous0 = count == 1 if remaining < 1 { return nil, errors.New("internal error: remaining < 1") } for remaining < threshold { nbBits-- threshold >>= 1 } if bitCount > 16 { out[outP] = byte(bitStream) out[outP+1] = byte(bitStream >> 8) outP += 2 bitStream >>= 16 bitCount -= 16 } } if outP+2 > len(out) { return nil, fmt.Errorf("internal error: %d > %d, maxheader: %d, sl: %d, tl: %d, normcount: %v", outP+2, len(out), maxHeaderSize, s.symbolLen, int(tableLog), s.norm[:s.symbolLen]) } out[outP] = byte(bitStream) out[outP+1] = byte(bitStream >> 8) outP += int((bitCount + 7) / 8) if charnum > s.symbolLen { return nil, errors.New("internal error: charnum > s.symbolLen") } return out[:outP], nil } // Approximate symbol cost, as fractional value, using fixed-point format (accuracyLog fractional bits) // note 1 : assume symbolValue is valid (<= maxSymbolValue) // note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits * func (s *fseEncoder) bitCost(symbolValue uint8, accuracyLog uint32) uint32 { minNbBits := s.ct.symbolTT[symbolValue].deltaNbBits >> 16 threshold := (minNbBits + 1) << 16 if debugAsserts { if !(s.actualTableLog < 16) { panic("!s.actualTableLog < 16") } // ensure enough room for renormalization double shift if !(uint8(accuracyLog) < 31-s.actualTableLog) { panic("!uint8(accuracyLog) < 31-s.actualTableLog") } } tableSize := uint32(1) << s.actualTableLog deltaFromThreshold := threshold - (s.ct.symbolTT[symbolValue].deltaNbBits + tableSize) // linear interpolation (very approximate) normalizedDeltaFromThreshold := (deltaFromThreshold << accuracyLog) >> s.actualTableLog bitMultiplier := uint32(1) << accuracyLog if debugAsserts { if s.ct.symbolTT[symbolValue].deltaNbBits+tableSize > threshold { panic("s.ct.symbolTT[symbolValue].deltaNbBits+tableSize > threshold") } if normalizedDeltaFromThreshold > bitMultiplier { panic("normalizedDeltaFromThreshold > bitMultiplier") } } return (minNbBits+1)*bitMultiplier - normalizedDeltaFromThreshold } // Returns the cost in bits of encoding the distribution in count using ctable. // Histogram should only be up to the last non-zero symbol. // Returns an -1 if ctable cannot represent all the symbols in count. func (s *fseEncoder) approxSize(hist []uint32) uint32 { if int(s.symbolLen) < len(hist) { // More symbols than we have. return math.MaxUint32 } if s.useRLE { // We will never reuse RLE encoders. return math.MaxUint32 } const kAccuracyLog = 8 badCost := (uint32(s.actualTableLog) + 1) << kAccuracyLog var cost uint32 for i, v := range hist { if v == 0 { continue } if s.norm[i] == 0 { return math.MaxUint32 } bitCost := s.bitCost(uint8(i), kAccuracyLog) if bitCost > badCost { return math.MaxUint32 } cost += v * bitCost } return cost >> kAccuracyLog } // maxHeaderSize returns the maximum header size in bits. // This is not exact size, but we want a penalty for new tables anyway. func (s *fseEncoder) maxHeaderSize() uint32 { if s.preDefined { return 0 } if s.useRLE { return 8 } return (((uint32(s.symbolLen) * uint32(s.actualTableLog)) >> 3) + 3) * 8 } // cState contains the compression state of a stream. type cState struct { bw *bitWriter stateTable []uint16 state uint16 } // init will initialize the compression state to the first symbol of the stream. func (c *cState) init(bw *bitWriter, ct *cTable, first symbolTransform) { c.bw = bw c.stateTable = ct.stateTable if len(c.stateTable) == 1 { // RLE c.stateTable[0] = uint16(0) c.state = 0 return } nbBitsOut := (first.deltaNbBits + (1 << 15)) >> 16 im := int32((nbBitsOut << 16) - first.deltaNbBits) lu := (im >> nbBitsOut) + int32(first.deltaFindState) c.state = c.stateTable[lu] } // flush will write the tablelog to the output and flush the remaining full bytes. func (c *cState) flush(tableLog uint8) { c.bw.flush32() c.bw.addBits16NC(c.state, tableLog) }