compress/flate: optimize history-copy decoding.
The forwardCopy function could be re-written in asm, and the copyHuff
method could probably be rolled into huffmanBlock and copyHist, but
I'm leaving those changes for future CLs.
compress/flate benchmarks:
benchmark old ns/op new ns/op delta
BenchmarkDecoderBestSpeed1K 385327 435140 +12.93%
BenchmarkDecoderBestSpeed10K
1245190 1062112 -14.70%
BenchmarkDecoderBestSpeed100K
8512365 5833680 -31.47%
BenchmarkDecoderDefaultCompression1K 382225 421301 +10.22%
BenchmarkDecoderDefaultCompression10K 867950 613890 -29.27%
BenchmarkDecoderDefaultCompression100K
5658240 2466726 -56.40%
BenchmarkDecoderBestCompression1K 383760 421634 +9.87%
BenchmarkDecoderBestCompression10K 867743 614671 -29.16%
BenchmarkDecoderBestCompression100K
5660160 2464996 -56.45%
image/png benchmarks:
benchmark old ns/op new ns/op delta
BenchmarkDecodeGray
2540834 2389624 -5.95%
BenchmarkDecodeNRGBAGradient
10052700 9534565 -5.15%
BenchmarkDecodeNRGBAOpaque
8704710 8163430 -6.22%
BenchmarkDecodePaletted
1458779 1325017 -9.17%
BenchmarkDecodeRGB
7183606 6794668 -5.41%
Wall time for Denis Cheremisov's PNG-decoding program given in
https://groups.google.com/group/golang-nuts/browse_thread/thread/
22aa8a05040fdd49
Before: 3.07s
After: 2.32s
Delta: -24%
Before profile:
Total: 304 samples
159 52.3% 52.3% 251 82.6% compress/flate.(*decompressor).huffmanBlock
58 19.1% 71.4% 76 25.0% compress/flate.(*decompressor).huffSym
32 10.5% 81.9% 32 10.5% hash/adler32.update
16 5.3% 87.2% 22 7.2% bufio.(*Reader).ReadByte
16 5.3% 92.4% 37 12.2% compress/flate.(*decompressor).moreBits
7 2.3% 94.7% 7 2.3% hash/crc32.update
7 2.3% 97.0% 7 2.3% runtime.memmove
5 1.6% 98.7% 5 1.6% scanblock
2 0.7% 99.3% 9 3.0% runtime.copy
1 0.3% 99.7% 1 0.3% compress/flate.(*huffmanDecoder).init
After profile:
Total: 230 samples
59 25.7% 25.7% 70 30.4% compress/flate.(*decompressor).huffSym
45 19.6% 45.2% 45 19.6% hash/adler32.update
35 15.2% 60.4% 35 15.2% compress/flate.forwardCopy
20 8.7% 69.1% 151 65.7% compress/flate.(*decompressor).huffmanBlock
16 7.0% 76.1% 24 10.4% compress/flate.(*decompressor).moreBits
15 6.5% 82.6% 15 6.5% runtime.memmove
11 4.8% 87.4% 50 21.7% compress/flate.(*decompressor).copyHist
7 3.0% 90.4% 7 3.0% hash/crc32.update
6 2.6% 93.0% 9 3.9% bufio.(*Reader).ReadByte
4 1.7% 94.8% 4 1.7% runtime.slicearray
R=rsc, rogpeppe, dave
CC=golang-dev, krasin
https://golang.org/cl/
6127064