zstd: Minor performance tweaks

zstd/blockenc.go

@@ -51,7 +51,7 @@ func (b *blockEnc) init() {
 		if cap(b.literals) < maxCompressedBlockSize {
 			b.literals = make([]byte, 0, maxCompressedBlockSize)
 		}
-		const defSeqs = 200
+		const defSeqs = 2000
 		if cap(b.sequences) < defSeqs {
 			b.sequences = make([]seq, 0, defSeqs)
 		}
@@ -426,7 +426,7 @@ func fuzzFseEncoder(data []byte) int {
 		return 0
 	}
 	enc := fseEncoder{}
-	hist := enc.Histogram()[:256]
+	hist := enc.Histogram()
 	maxSym := uint8(0)
 	for i, v := range data {
 		v = v & 63
@@ -802,14 +802,13 @@ func (b *blockEnc) genCodes() {
 		// nothing to do
 		return
 	}
-
 	if len(b.sequences) > math.MaxUint16 {
 		panic("can only encode up to 64K sequences")
 	}
 	// No bounds checks after here:
-	llH := b.coders.llEnc.Histogram()[:256]
-	ofH := b.coders.ofEnc.Histogram()[:256]
-	mlH := b.coders.mlEnc.Histogram()[:256]
+	llH := b.coders.llEnc.Histogram()
+	ofH := b.coders.ofEnc.Histogram()
+	mlH := b.coders.mlEnc.Histogram()
 	for i := range llH {
 		llH[i] = 0
 	}

zstd/enc_base.go

@@ -108,11 +108,6 @@ func (e *fastBase) UseBlock(enc *blockEnc) {
 	e.blk = enc
 }
 
-func (e *fastBase) matchlenNoHist(s, t int32, src []byte) int32 {
-	// Extend the match to be as long as possible.
-	return int32(matchLen(src[s:], src[t:]))
-}
-
 func (e *fastBase) matchlen(s, t int32, src []byte) int32 {
 	if debugAsserts {
 		if s < 0 {
@@ -131,9 +126,24 @@ func (e *fastBase) matchlen(s, t int32, src []byte) int32 {
 			panic(fmt.Sprintf("len(src)-s (%d) > maxCompressedBlockSize (%d)", len(src)-int(s), maxCompressedBlockSize))
 		}
 	}
+	a := src[s:]
+	b := src[t:]
+	b = b[:len(a)]
+	end := int32((len(a) >> 3) << 3)
+	for i := int32(0); i < end; i += 8 {
+		if diff := load6432(a, i) ^ load6432(b, i); diff != 0 {
+			return i + int32(bits.TrailingZeros64(diff)>>3)
+		}
+	}
 
-	// Extend the match to be as long as possible.
-	return int32(matchLen(src[s:], src[t:]))
+	a = a[end:]
+	b = b[end:]
+	for i := range a {
+		if a[i] != b[i] {
+			return int32(i) + end
+		}
+	}
+	return int32(len(a)) + end
 }
 
 // Reset the encoding table.

zstd/enc_fast.go

@@ -6,8 +6,6 @@ package zstd
 
 import (
 	"fmt"
-	"math"
-	"math/bits"
 )
 
 const (
@@ -136,20 +134,7 @@ encodeLoop:
 				// Consider history as well.
 				var seq seq
 				var length int32
-				// length = 4 + e.matchlen(s+6, repIndex+4, src)
-				{
-					a := src[s+6:]
-					b := src[repIndex+4:]
-					endI := len(a) & (math.MaxInt32 - 7)
-					length = int32(endI) + 4
-					for i := 0; i < endI; i += 8 {
-						if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-							length = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-							break
-						}
-					}
-				}
-
+				length = 4 + e.matchlen(s+6, repIndex+4, src)
 				seq.matchLen = uint32(length - zstdMinMatch)
 
 				// We might be able to match backwards.
@@ -236,20 +221,7 @@ encodeLoop:
 		}
 
 		// Extend the 4-byte match as long as possible.
-		//l := e.matchlen(s+4, t+4, src) + 4
-		var l int32
-		{
-			a := src[s+4:]
-			b := src[t+4:]
-			endI := len(a) & (math.MaxInt32 - 7)
-			l = int32(endI) + 4
-			for i := 0; i < endI; i += 8 {
-				if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-					l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-					break
-				}
-			}
-		}
+		l := e.matchlen(s+4, t+4, src) + 4
 
 		// Extend backwards
 		tMin := s - e.maxMatchOff
@@ -286,20 +258,7 @@ encodeLoop:
 		if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) {
 			// We have at least 4 byte match.
 			// No need to check backwards. We come straight from a match
-			//l := 4 + e.matchlen(s+4, o2+4, src)
-			var l int32
-			{
-				a := src[s+4:]
-				b := src[o2+4:]
-				endI := len(a) & (math.MaxInt32 - 7)
-				l = int32(endI) + 4
-				for i := 0; i < endI; i += 8 {
-					if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-						l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-						break
-					}
-				}
-			}
+			l := 4 + e.matchlen(s+4, o2+4, src)
 
 			// Store this, since we have it.
 			nextHash := hashLen(cv, hashLog, tableFastHashLen)
@@ -418,21 +377,7 @@ encodeLoop:
 			if len(blk.sequences) > 2 && load3232(src, repIndex) == uint32(cv>>16) {
 				// Consider history as well.
 				var seq seq
-				// length := 4 + e.matchlen(s+6, repIndex+4, src)
-				// length := 4 + int32(matchLen(src[s+6:], src[repIndex+4:]))
-				var length int32
-				{
-					a := src[s+6:]
-					b := src[repIndex+4:]
-					endI := len(a) & (math.MaxInt32 - 7)
-					length = int32(endI) + 4
-					for i := 0; i < endI; i += 8 {
-						if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-							length = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-							break
-						}
-					}
-				}
+				length := 4 + e.matchlen(s+6, repIndex+4, src)
 
 				seq.matchLen = uint32(length - zstdMinMatch)
 
@@ -522,21 +467,7 @@ encodeLoop:
 			panic(fmt.Sprintf("t (%d) < 0 ", t))
 		}
 		// Extend the 4-byte match as long as possible.
-		//l := e.matchlenNoHist(s+4, t+4, src) + 4
-		// l := int32(matchLen(src[s+4:], src[t+4:])) + 4
-		var l int32
-		{
-			a := src[s+4:]
-			b := src[t+4:]
-			endI := len(a) & (math.MaxInt32 - 7)
-			l = int32(endI) + 4
-			for i := 0; i < endI; i += 8 {
-				if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-					l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-					break
-				}
-			}
-		}
+		l := e.matchlen(s+4, t+4, src) + 4
 
 		// Extend backwards
 		tMin := s - e.maxMatchOff
@@ -573,21 +504,7 @@ encodeLoop:
 		if o2 := s - offset2; len(blk.sequences) > 2 && load3232(src, o2) == uint32(cv) {
 			// We have at least 4 byte match.
 			// No need to check backwards. We come straight from a match
-			//l := 4 + e.matchlenNoHist(s+4, o2+4, src)
-			// l := 4 + int32(matchLen(src[s+4:], src[o2+4:]))
-			var l int32
-			{
-				a := src[s+4:]
-				b := src[o2+4:]
-				endI := len(a) & (math.MaxInt32 - 7)
-				l = int32(endI) + 4
-				for i := 0; i < endI; i += 8 {
-					if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-						l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-						break
-					}
-				}
-			}
+			l := 4 + e.matchlen(s+4, o2+4, src)
 
 			// Store this, since we have it.
 			nextHash := hashLen(cv, hashLog, tableFastHashLen)
@@ -731,19 +648,7 @@ encodeLoop:
 				// Consider history as well.
 				var seq seq
 				var length int32
-				// length = 4 + e.matchlen(s+6, repIndex+4, src)
-				{
-					a := src[s+6:]
-					b := src[repIndex+4:]
-					endI := len(a) & (math.MaxInt32 - 7)
-					length = int32(endI) + 4
-					for i := 0; i < endI; i += 8 {
-						if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-							length = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-							break
-						}
-					}
-				}
+				length = 4 + e.matchlen(s+6, repIndex+4, src)
 
 				seq.matchLen = uint32(length - zstdMinMatch)
 
@@ -831,20 +736,7 @@ encodeLoop:
 		}
 
 		// Extend the 4-byte match as long as possible.
-		//l := e.matchlen(s+4, t+4, src) + 4
-		var l int32
-		{
-			a := src[s+4:]
-			b := src[t+4:]
-			endI := len(a) & (math.MaxInt32 - 7)
-			l = int32(endI) + 4
-			for i := 0; i < endI; i += 8 {
-				if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-					l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-					break
-				}
-			}
-		}
+		l := e.matchlen(s+4, t+4, src) + 4
 
 		// Extend backwards
 		tMin := s - e.maxMatchOff
@@ -881,20 +773,7 @@ encodeLoop:
 		if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) {
 			// We have at least 4 byte match.
 			// No need to check backwards. We come straight from a match
-			//l := 4 + e.matchlen(s+4, o2+4, src)
-			var l int32
-			{
-				a := src[s+4:]
-				b := src[o2+4:]
-				endI := len(a) & (math.MaxInt32 - 7)
-				l = int32(endI) + 4
-				for i := 0; i < endI; i += 8 {
-					if diff := load64(a, i) ^ load64(b, i); diff != 0 {
-						l = int32(i+bits.TrailingZeros64(diff)>>3) + 4
-						break
-					}
-				}
-			}
+			l := 4 + e.matchlen(s+4, o2+4, src)
 
 			// Store this, since we have it.
 			nextHash := hashLen(cv, hashLog, tableFastHashLen)

zstd/encoder_test.go

@@ -51,7 +51,7 @@ func getEncOpts(cMax int) []testEncOpt {
 				addOpt("nolit", WithNoEntropyCompression(true))
 				addOpt("pad1k", WithEncoderPadding(1024))
 				addOpt("zerof", WithZeroFrames(true))
-				addOpt("singleseg", WithSingleSegment(true))
+				addOpt("1seg", WithSingleSegment(true))
 			}
 			if testing.Short() && conc == 2 {
 				break
@@ -904,10 +904,10 @@ func BenchmarkEncoder_EncodeAllXML(b *testing.B) {
 	}
 	dec.Close()
 
-	enc := Encoder{}
+	enc, _ := NewWriter(nil, WithEncoderConcurrency(1))
 	dst := enc.EncodeAll(in, nil)
 	wantSize := len(dst)
-	b.Log("Output size:", len(dst))
+	//b.Log("Output size:", len(dst))
 	b.ResetTimer()
 	b.ReportAllocs()
 	b.SetBytes(int64(len(in)))
@@ -994,7 +994,7 @@ func BenchmarkEncoder_EncodeAllHTML(b *testing.B) {
 		b.Fatal(err)
 	}
 
-	enc := Encoder{}
+	enc, _ := NewWriter(nil, WithEncoderConcurrency(1))
 	dst := enc.EncodeAll(in, nil)
 	wantSize := len(dst)
 	b.ResetTimer()
@@ -1018,7 +1018,7 @@ func BenchmarkEncoder_EncodeAllTwain(b *testing.B) {
 		b.Fatal(err)
 	}
 
-	enc := Encoder{}
+	enc, _ := NewWriter(nil, WithEncoderConcurrency(1))
 	dst := enc.EncodeAll(in, nil)
 	wantSize := len(dst)
 	b.ResetTimer()
@@ -1042,7 +1042,7 @@ func BenchmarkEncoder_EncodeAllPi(b *testing.B) {
 		b.Fatal(err)
 	}
 
-	enc := Encoder{}
+	enc, _ := NewWriter(nil, WithEncoderConcurrency(1))
 	dst := enc.EncodeAll(in, nil)
 	wantSize := len(dst)
 	b.ResetTimer()

zstd/fse_encoder.go

@@ -62,9 +62,8 @@ func (s symbolTransform) String() string {
 // 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.
-// The returned slice will always be length 256.
-func (s *fseEncoder) Histogram() []uint32 {
-	return s.count[:]
+func (s *fseEncoder) Histogram() *[256]uint32 {
+	return &s.count
 }
 
 // HistogramFinished can be called to indicate that the histogram has been populated.