Optimize single block encodes (#199)

* Optimize single block encodes
* Blocks without history cannot have invalid repeat codes.
* Remove bounds check.
* Don't rank empty part.
* Big speedup on huff0 encoding.

huff0/bitwriter.go

@@ -38,14 +38,25 @@ func (b *bitWriter) addBits16Clean(value uint16, bits uint8) {
 	b.nBits += bits
 }
 
-// addBits16Clean will add up to 16 bits. value may not contain more set bits than indicated.
+// encSymbol will add up to 16 bits. value may not contain more set bits than indicated.
 // It will not check if there is space for them, so the caller must ensure that it has flushed recently.
 func (b *bitWriter) encSymbol(ct cTable, symbol byte) {
 	enc := ct[symbol]
 	b.bitContainer |= uint64(enc.val) << (b.nBits & 63)
 	b.nBits += enc.nBits
 }
 
+// encTwoSymbols will add up to 32 bits. value may not contain more set bits than indicated.
+// It will not check if there is space for them, so the caller must ensure that it has flushed recently.
+func (b *bitWriter) encTwoSymbols(ct cTable, av, bv byte) {
+	encA := ct[av]
+	encB := ct[bv]
+	sh := b.nBits & 63
+	combined := uint64(encA.val) | (uint64(encB.val) << (encA.nBits & 63))
+	b.bitContainer |= combined << sh
+	b.nBits += encA.nBits + encB.nBits
+}
+
 // addBits16ZeroNC will add up to 16 bits.
 // It will not check if there is space for them,
 // so the caller must ensure that it has flushed recently.

huff0/compress.go

@@ -163,28 +163,23 @@ func (s *Scratch) compress1xDo(dst, src []byte) ([]byte, error) {
 	for i := len(src) & 3; i > 0; i-- {
 		bw.encSymbol(cTable, src[n+i-1])
 	}
+	n -= 4
 	if s.actualTableLog <= 8 {
-		n -= 4
 		for ; n >= 0; n -= 4 {
 			tmp := src[n : n+4]
 			// tmp should be len 4
 			bw.flush32()
-			bw.encSymbol(cTable, tmp[3])
-			bw.encSymbol(cTable, tmp[2])
-			bw.encSymbol(cTable, tmp[1])
-			bw.encSymbol(cTable, tmp[0])
+			bw.encTwoSymbols(cTable, tmp[3], tmp[2])
+			bw.encTwoSymbols(cTable, tmp[1], tmp[0])
 		}
 	} else {
-		n -= 4
 		for ; n >= 0; n -= 4 {
 			tmp := src[n : n+4]
 			// tmp should be len 4
 			bw.flush32()
-			bw.encSymbol(cTable, tmp[3])
-			bw.encSymbol(cTable, tmp[2])
+			bw.encTwoSymbols(cTable, tmp[3], tmp[2])
 			bw.flush32()
-			bw.encSymbol(cTable, tmp[1])
-			bw.encSymbol(cTable, tmp[0])
+			bw.encTwoSymbols(cTable, tmp[1], tmp[0])
 		}
 	}
 	err := bw.close()
@@ -439,7 +434,7 @@ func (s *Scratch) buildCTable() error {
 		return fmt.Errorf("internal error: maxNbBits (%d) > tableLogMax (%d)", maxNbBits, tableLogMax)
 	}
 	var nbPerRank [tableLogMax + 1]uint16
-	var valPerRank [tableLogMax + 1]uint16
+	var valPerRank [16]uint16
 	for _, v := range huffNode[:nonNullRank+1] {
 		nbPerRank[v.nbBits]++
 	}
@@ -455,16 +450,17 @@ func (s *Scratch) buildCTable() error {
 	}
 
 	// push nbBits per symbol, symbol order
-	// TODO: changed `s.symbolLen` -> `nonNullRank+1` (micro-opt)
 	for _, v := range huffNode[:nonNullRank+1] {
 		s.cTable[v.symbol].nBits = v.nbBits
 	}
 
 	// assign value within rank, symbol order
-	for n, val := range s.cTable[:s.symbolLen] {
-		v := valPerRank[val.nBits]
-		s.cTable[n].val = v
-		valPerRank[val.nBits] = v + 1
+	t := s.cTable[:s.symbolLen]
+	for n, val := range t {
+		nbits := val.nBits & 15
+		v := valPerRank[nbits]
+		t[n].val = v
+		valPerRank[nbits] = v + 1
 	}
 
 	return nil
@@ -488,10 +484,12 @@ func (s *Scratch) huffSort() {
 		r := highBit32(v+1) & 31
 		rank[r].base++
 	}
-	for n := 30; n > 0; n-- {
+	// maxBitLength is log2(BlockSizeMax) + 1
+	const maxBitLength = 18 + 1
+	for n := maxBitLength; n > 0; n-- {
 		rank[n-1].base += rank[n].base
 	}
-	for n := range rank[:] {
+	for n := range rank[:maxBitLength] {
 		rank[n].current = rank[n].base
 	}
 	for n, c := range s.count[:s.symbolLen] {
@@ -510,7 +508,7 @@ func (s *Scratch) huffSort() {
 }
 
 func (s *Scratch) setMaxHeight(lastNonNull int) uint8 {
-	maxNbBits := s.TableLog
+	maxNbBits := s.actualTableLog
 	huffNode := s.nodes[1 : huffNodesLen+1]
 	//huffNode = huffNode[: huffNodesLen]
 

zstd/blockenc.go

@@ -299,6 +299,20 @@ func (b *blockEnc) encodeRaw(a []byte) {
 	}
 }
 
+// encodeRaw can be used to set the output to a raw representation of supplied bytes.
+func (b *blockEnc) encodeRawTo(dst, src []byte) []byte {
+	var bh blockHeader
+	bh.setLast(b.last)
+	bh.setSize(uint32(len(src)))
+	bh.setType(blockTypeRaw)
+	dst = bh.appendTo(dst)
+	dst = append(dst, src...)
+	if debug {
+		println("Adding RAW block, length", len(src))
+	}
+	return dst
+}
+
 // encodeLits can be used if the block is only litLen.
 func (b *blockEnc) encodeLits(raw bool) error {
 	var bh blockHeader
@@ -437,7 +451,7 @@ func fuzzFseEncoder(data []byte) int {
 	return 1
 }
 
-// encode will encode the block and put the output in b.output.
+// encode will encode the block and append the output in b.output.
 func (b *blockEnc) encode(raw bool) error {
 	if len(b.sequences) == 0 {
 		return b.encodeLits(raw)
@@ -451,6 +465,8 @@ func (b *blockEnc) encode(raw bool) error {
 	var lh literalsHeader
 	bh.setLast(b.last)
 	bh.setType(blockTypeCompressed)
+	// Store offset of the block header. Needed when we know the size.
+	bhOffset := len(b.output)
 	b.output = bh.appendTo(b.output)
 
 	var (
@@ -468,6 +484,7 @@ func (b *blockEnc) encode(raw bool) error {
 	} else {
 		err = huff0.ErrIncompressible
 	}
+
 	switch err {
 	case huff0.ErrIncompressible:
 		lh.setType(literalsBlockRaw)
@@ -735,18 +752,18 @@ func (b *blockEnc) encode(raw bool) error {
 	}
 	b.output = wr.out
 
-	if len(b.output)-3 >= b.size {
+	if len(b.output)-3-bhOffset >= b.size {
 		// Maybe even add a bigger margin.
 		b.litEnc.Reuse = huff0.ReusePolicyNone
 		return errIncompressible
 	}
 
 	// Size is output minus block header.
-	bh.setSize(uint32(len(b.output)) - 3)
+	bh.setSize(uint32(len(b.output)-bhOffset) - 3)
 	if debug {
 		println("Rewriting block header", bh)
 	}
-	_ = bh.appendTo(b.output[:0])
+	_ = bh.appendTo(b.output[bhOffset:bhOffset])
 	b.coders.setPrev(llEnc, mlEnc, ofEnc)
 	return nil
 }