internal/xgbin: remove use of most of the calls to binary.Read

[sbinet]

This CL gets rid of most of the calls to binary.Read which makes heavy
use of reflection.

Also, drop bufio.Reader in favor of io.Reader.

Fixes #13.

[coveralls]

Pull Request Test Coverage Report for Build 52

• 31 of 52 (59.62%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.1%) to 60.07%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
internal/xgbin/xgbin_io.go	31	52	59.62%

Totals
Change from base Build 51:	-0.1%
Covered Lines:	689
Relevant Lines:	1147

---

💛 - Coveralls

[dmitryikh]

@sbinet , thanks for your help to the project.

I did some benchmark on loading speed:

func BenchmarkHiggsLoading(b *testing.B) {
	for i := 0; i < b.N; i++ {
		_, err := XGEnsembleFromFile(filepath.Join("testdata", "xghiggs.model"))
		if err != nil {
			b.Skip(err.Error())
		}
	}
}

[leaves] go test -bench Loading
goos: darwin
goarch: amd64
pkg: github.com/dmitryikh/leaves
BenchmarkHiggsLoading-8   	       5	 207094885 ns/op   <--- after removing binary.Read
PASS
ok  	github.com/dmitryikh/leaves	2.179s
[leaves] git rebase -i HEAD~2
Stopped at b724c93...  [+] xgboost model loading benchmark
...
[leaves] go test -bench Loading
goos: darwin
goarch: amd64
pkg: github.com/dmitryikh/leaves
BenchmarkHiggsLoading-8   	       5	 213136769 ns/op    <--- before removing binary.Read
PASS
ok  	github.com/dmitryikh/leaves	2.177s

As you can see the speedup is around ~4%

If you wish you can repeat the experiment taking xghiggs.model from:
https://github.com/dmitryikh/leaves/releases/tag/testdata_v2

Your PR seems nice to me. And if you don't want to add something, I will merge it.

[sbinet]

that's a bit disappointing... :)

there's probably some more perf to squeeze out of ReadStruct.
also, right now ReadFloat32Slice is only used in one place where we just discard the newly allocated []float32. Perhaps we could just discard it up-front instead of creating it and then discarding it...

[dmitryikh]

@sbinet, I did some investigations:

1. ReadFloat32Slices is called vary rarely (once per tree if SizeLeafVector > 0 which is false in the benchmark)
2. ReadInt32Slice is called once for model.
3. Most (almost all time) we seats in the ReadStruct function:

[leaves] go tool pprof  leaves.test cpu.out
File: leaves.test
Type: cpu
Time: Sep 21, 2018 at 7:21am (MSK)
Duration: 1.41s, Total samples = 1.28s (90.88%)
Entering interactive mode (type "help" for commands, "o" for options)
(pprof) list ReadStruct
Total: 1.28s
ROUTINE ======================== github.com/dmitryikh/leaves/internal/xgbin.ReadStruct in /Users/d.khominich/code/go/src/github.com/dmitryikh/leaves/internal/xgbin/xgbin_io.go
         0      1.13s (flat, cum) 88.28% of Total
         .          .    134:	NameGbm string
         .          .    135:}
         .          .    136:
         .          .    137:// ReadStruct - read arbitrary data structure from binary stream
         .          .    138:func ReadStruct(reader *bufio.Reader, dst interface{}) error {
         .      1.13s    139:	err := binary.Read(reader, binary.LittleEndian, dst)
         .          .    140:	if err != nil {
         .          .    141:		return err
         .          .    142:	}
         .          .    143:	return nil
         .          .    144:}

And memory allocations:

(pprof) list ReadStruct
Total: 943.14MB
ROUTINE ======================== github.com/dmitryikh/leaves/internal/xgbin.ReadStruct in /Users/d.khominich/code/go/src/github.com/dmitryikh/leaves/internal/xgbin/xgbin_io.go
         0   445.01MB (flat, cum) 47.18% of Total
         .          .    134:	NameGbm string
         .          .    135:}
         .          .    136:
         .          .    137:// ReadStruct - read arbitrary data structure from binary stream
         .          .    138:func ReadStruct(reader *bufio.Reader, dst interface{}) error {
         .   445.01MB    139:	err := binary.Read(reader, binary.LittleEndian, dst)
         .          .    140:	if err != nil {
         .          .    141:		return err
         .          .    142:	}
         .          .    143:	return nil
         .          .    144:}

Do you have any ideas how to speedup ReadStruct? Is it possible to read just sizeof(SomeStructType) raw bytes and then cast them to SomeStructType object?

[sbinet]

I am in the process of introducing a 'Decoder' type (like for json) that should deal with these pesky structs.
Stay tuned.

[dmitryikh]

out of date