package filter

cmd/gotta/main.go

@@ -23,7 +23,7 @@ func init() {
 }
 
 func main() {
-	fmt.Fprintf(os.Stderr, "\r\nTTA1 lossless audio encoder/decoder, version %s{%d}\n\n", tta.Version, tta.BinaryVersion())
+	fmt.Fprintf(os.Stderr, "\r\nTTA1 lossless audio encoder/decoder, version %s\n\n", tta.Version)
 	flag.Parse()
 	if help || flag.NArg() < 1 || (!decode && !encode) {
 		fmt.Fprintf(os.Stderr, "\rUsage of gotta: [encode|decode] [passwd PASSWORD] INPUT_FILE OUTPUT_FILE\n\n")

codec.go

@@ -1,5 +1,9 @@
 package tta
 
+import (
+	"github.com/zyxar/tta/filter"
+)
+
 type Info struct {
 	format  uint32 // audio format
 	nch     uint32 // number of channels
@@ -23,7 +27,7 @@ func (a *adapter) init(k0, k1 uint32) {
 }
 
 type codec struct {
-	filter  Filter
+	filter  filter.Filter
 	adapter adapter
 	prev    int32
 }

const.go

@@ -17,7 +17,7 @@ const (
 	formatEncrypted = 2
 
 	// Version of TTA lib
-	Version = "0.0.1"
+	Version = "0.0.2"
 )
 
 var (
@@ -34,14 +34,6 @@ var (
 	errPartialWritten = errors.New("partial written")
 )
 
-const ( // CPU_ARCH_TYPE
-	cpuArchUndefined = iota
-	cpuArchIx86Sse2
-	cpuArchIx86Sse3
-	cpuArchIx86Sse4_1
-	cpuArchIx86Sse4_2
-)
-
 var bitMask = []uint32{
 	0x00000000, 0x00000001, 0x00000003, 0x00000007,
 	0x0000000f, 0x0000001f, 0x0000003f, 0x0000007f,

decoder.go

@@ -4,6 +4,7 @@ import (
 	"io"
 	"os"
 
+	"github.com/zyxar/tta/filter"
 	"github.com/zyxar/tta/wave"
 )
 
@@ -226,7 +227,7 @@ func (d *Decoder) frameInit(frame uint32, seekNeeded bool) (err error) {
 		d.flen = d.flenStd
 	}
 	for i := 0; i < d.channels; i++ {
-		d.codecs[i].filter = NewCompatibleFilter(d.data, shift)
+		d.codecs[i].filter = filter.New(d.data, shift)
 		d.codecs[i].adapter.init(10, 10)
 		d.codecs[i].prev = 0
 	}

encoder.go

@@ -5,6 +5,7 @@ import (
 	"io"
 	"os"
 
+	"github.com/zyxar/tta/filter"
 	"github.com/zyxar/tta/wave"
 )
 
@@ -228,7 +229,7 @@ func (e *Encoder) frameInit(frame uint32) (err error) {
 	}
 	// init entropy encoder
 	for i := 0; i < e.channels; i++ {
-		e.codecs[i].filter = NewCompatibleFilter(e.data, shift)
+		e.codecs[i].filter = filter.New(e.data, shift)
 		e.codecs[i].adapter.init(10, 10)
 		e.codecs[i].prev = 0
 	}

filter/cpu.go

@@ -1,4 +1,4 @@
-package tta
+package filter
 
 import (
 	"github.com/klauspost/cpuid"

filter/filter.go

@@ -1,4 +1,4 @@
-package tta
+package filter
 
 // TODO: SSE4 optimization
 
@@ -7,20 +7,20 @@ type Filter interface {
 	Encode(*int32)
 }
 
-type filterCompat struct {
-	index int32
-	error int32
-	round int32
-	shift uint32
-	qm    [8]int32
-	dx    [24]int32
-	dl    [24]int32
+type flt struct {
+	index  int32
+	error  int32
+	round  int32
+	shift  uint32
+	qm     [8]int32
+	dx     [24]int32
+	dl     [24]int32
+	decode func(*int32)
+	encode func(*int32)
 }
 
-// type filterSse filterCompat
-
-func NewCompatibleFilter(data [8]byte, shift uint32) Filter {
-	t := filterCompat{}
+func New(data [8]byte, shift uint32) Filter {
+	t := flt{}
 	t.shift = shift
 	t.round = 1 << uint32(shift-1)
 	t.qm[0] = int32(int8(data[0]))
@@ -31,10 +31,20 @@ func NewCompatibleFilter(data [8]byte, shift uint32) Filter {
 	t.qm[5] = int32(int8(data[5]))
 	t.qm[6] = int32(int8(data[6]))
 	t.qm[7] = int32(int8(data[7]))
+	t.decode = t.DecodeCompat
+	t.encode = t.EncodeCompat
 	return &t
 }
 
-func (t *filterCompat) Decode(in *int32) {
+func (f *flt) Decode(in *int32) {
+	f.decode(in)
+}
+
+func (f *flt) Encode(in *int32) {
+	f.encode(in)
+}
+
+func (t *flt) DecodeCompat(in *int32) {
 	pa := t.dl[:]
 	pb := t.qm[:]
 	pm := t.dx[:]
@@ -84,7 +94,7 @@ func (t *filterCompat) Decode(in *int32) {
 	pa[4] += pa[5]
 }
 
-func (t *filterCompat) Encode(in *int32) {
+func (t *flt) EncodeCompat(in *int32) {
 	pa := t.dl[:]
 	pb := t.qm[:]
 	pm := t.dx[:]

util.go

@@ -4,13 +4,6 @@ import (
 	"encoding/binary"
 )
 
-func BinaryVersion() byte {
-	if sseEnabled {
-		return cpuArchIx86Sse4_1
-	}
-	return cpuArchUndefined
-}
-
 func computeKeyDigits(p []byte) [8]byte {
 	var crcLow, crcHigh uint32 = 0xFFFFFFFF, 0xFFFFFFFF
 	for i := 0; i < len(p); i++ {