From 43829fcbc2c042d8da29a97f784ab22ca44581ae Mon Sep 17 00:00:00 2001
From: Klaus Post <klauspost@gmail.com>
Date: Sat, 5 Feb 2022 04:36:01 -0800
Subject: [PATCH] flate: Improve huffman generation speed ~5-10% faster (#490)

Mainly gains for small blocks & fastest levels.

```
benchmark                               old ns/op     new ns/op     delta
BenchmarkEncodeDigitsConstant1e4-32     16080         15984         -0.60%
BenchmarkEncodeDigitsSpeed1e4-32        48105         45890         -4.60%
BenchmarkEncodeDigitsDefault1e4-32      114016        111234        -2.44%
BenchmarkEncodeDigitsCompress1e4-32     228570        225041        -1.54%
BenchmarkEncodeDigitsSL1e4-32           46695         45006         -3.62%
BenchmarkEncodeTwainConstant1e4-32      23557         22033         -6.47%
BenchmarkEncodeTwainSpeed1e4-32         68594         63177         -7.90%
BenchmarkEncodeTwainDefault1e4-32       112498        111798        -0.62%
BenchmarkEncodeTwainCompress1e4-32      282909        274804        -2.86%
BenchmarkEncodeTwainSL1e4-32            68406         62362         -8.84%

benchmark                               old MB/s     new MB/s     speedup
BenchmarkEncodeDigitsConstant1e4-32     621.89       625.64       1.01x
BenchmarkEncodeDigitsSpeed1e4-32        207.88       217.91       1.05x
BenchmarkEncodeDigitsDefault1e4-32      87.71        89.90        1.02x
BenchmarkEncodeDigitsCompress1e4-32     43.75        44.44        1.02x
BenchmarkEncodeDigitsSL1e4-32           214.16       222.19       1.04x
BenchmarkEncodeTwainConstant1e4-32      424.51       453.87       1.07x
BenchmarkEncodeTwainSpeed1e4-32         145.79       158.29       1.09x
BenchmarkEncodeTwainDefault1e4-32       88.89        89.45        1.01x
BenchmarkEncodeTwainCompress1e4-32      35.35        36.39        1.03x
BenchmarkEncodeTwainSL1e4-32            146.19       160.35       1.10x
```

(output is byte exact)
---
 flate/huffman_code.go | 29 ++++++++++++++++++++---------
 1 file changed, 20 insertions(+), 9 deletions(-)

diff --git a/flate/huffman_code.go b/flate/huffman_code.go
index 9989830b53..9ab497c275 100644
--- a/flate/huffman_code.go
+++ b/flate/huffman_code.go
@@ -188,14 +188,19 @@ func (h *huffmanEncoder) bitCounts(list []literalNode, maxBits int32) []int32 {
 	// of the level j ancestor.
 	var leafCounts [maxBitsLimit][maxBitsLimit]int32
 
+	// Descending to only have 1 bounds check.
+	l2f := int32(list[2].freq)
+	l1f := int32(list[1].freq)
+	l0f := int32(list[0].freq) + int32(list[1].freq)
+
 	for level := int32(1); level <= maxBits; level++ {
 		// For every level, the first two items are the first two characters.
 		// We initialize the levels as if we had already figured this out.
 		levels[level] = levelInfo{
 			level:        level,
-			lastFreq:     int32(list[1].freq),
-			nextCharFreq: int32(list[2].freq),
-			nextPairFreq: int32(list[0].freq) + int32(list[1].freq),
+			lastFreq:     l1f,
+			nextCharFreq: l2f,
+			nextPairFreq: l0f,
 		}
 		leafCounts[level][level] = 2
 		if level == 1 {
@@ -206,8 +211,8 @@ func (h *huffmanEncoder) bitCounts(list []literalNode, maxBits int32) []int32 {
 	// We need a total of 2*n - 2 items at top level and have already generated 2.
 	levels[maxBits].needed = 2*n - 4
 
-	level := maxBits
-	for {
+	level := uint32(maxBits)
+	for level < 16 {
 		l := &levels[level]
 		if l.nextPairFreq == math.MaxInt32 && l.nextCharFreq == math.MaxInt32 {
 			// We've run out of both leafs and pairs.
@@ -239,7 +244,13 @@ func (h *huffmanEncoder) bitCounts(list []literalNode, maxBits int32) []int32 {
 			// more values in the level below
 			l.lastFreq = l.nextPairFreq
 			// Take leaf counts from the lower level, except counts[level] remains the same.
-			copy(leafCounts[level][:level], leafCounts[level-1][:level])
+			if true {
+				save := leafCounts[level][level]
+				leafCounts[level] = leafCounts[level-1]
+				leafCounts[level][level] = save
+			} else {
+				copy(leafCounts[level][:level], leafCounts[level-1][:level])
+			}
 			levels[l.level-1].needed = 2
 		}
 
@@ -310,6 +321,7 @@ func (h *huffmanEncoder) assignEncodingAndSize(bitCount []int32, list []literalN
 // maxBits  The maximum number of bits to use for any literal.
 func (h *huffmanEncoder) generate(freq []uint16, maxBits int32) {
 	list := h.freqcache[:len(freq)+1]
+	codes := h.codes[:len(freq)]
 	// Number of non-zero literals
 	count := 0
 	// Set list to be the set of all non-zero literals and their frequencies
@@ -318,11 +330,10 @@ func (h *huffmanEncoder) generate(freq []uint16, maxBits int32) {
 			list[count] = literalNode{uint16(i), f}
 			count++
 		} else {
-			list[count] = literalNode{}
-			h.codes[i].len = 0
+			codes[i].len = 0
 		}
 	}
-	list[len(freq)] = literalNode{}
+	list[count] = literalNode{}
 
 	list = list[:count]
 	if count <= 2 {