From 5593e2b2ddcaeea40d28d7007369105a3bff7e66 Mon Sep 17 00:00:00 2001
From: Klaus Post <klauspost@gmail.com>
Date: Thu, 2 Dec 2021 07:39:56 -0800
Subject: [PATCH] Unroll pure go xor loop (#172)

* Unroll pure go xor loop

Testing with `go test -bench=x1x -tags=noasm -short`

```
before:
BenchmarkEncode2x1x1M-32           13658             87980 ns/op        35754.96 MB/s

after:
BenchmarkEncode2x1x1M-32           21633             55498 ns/op        56682.24 MB/s
```
---
 galois_amd64.go     |  2 +-
 galois_arm64.go     |  2 +-
 galois_noasm.go     | 23 ++++++++++++++++++-----
 reedsolomon_test.go |  1 +
 4 files changed, 21 insertions(+), 7 deletions(-)

diff --git a/galois_amd64.go b/galois_amd64.go
index 2db8621..03754d2 100644
--- a/galois_amd64.go
+++ b/galois_amd64.go
@@ -114,7 +114,7 @@ func galMulSliceXor(c byte, in, out []byte, o *options) {
 	}
 }
 
-// slice galois add
+// simple slice xor
 func sliceXor(in, out []byte, o *options) {
 	if o.useSSE2 {
 		if len(in) >= bigSwitchover {
diff --git a/galois_arm64.go b/galois_arm64.go
index 93acd74..df79a98 100644
--- a/galois_arm64.go
+++ b/galois_arm64.go
@@ -51,7 +51,7 @@ func galMulSliceXor(c byte, in, out []byte, o *options) {
 	}
 }
 
-// slice galois add
+// simple slice xor
 func sliceXor(in, out []byte, o *options) {
 
 	galXorNEON(in, out)
diff --git a/galois_noasm.go b/galois_noasm.go
index 6d3a5a7..7ef78f8 100644
--- a/galois_noasm.go
+++ b/galois_noasm.go
@@ -7,6 +7,8 @@
 
 package reedsolomon
 
+import "encoding/binary"
+
 func galMulSlice(c byte, in, out []byte, o *options) {
 	out = out[:len(in)]
 	if c == 1 {
@@ -22,9 +24,7 @@ func galMulSlice(c byte, in, out []byte, o *options) {
 func galMulSliceXor(c byte, in, out []byte, o *options) {
 	out = out[:len(in)]
 	if c == 1 {
-		for n, input := range in {
-			out[n] ^= input
-		}
+		sliceXor(in, out, o)
 		return
 	}
 	mt := mulTable[c][:256]
@@ -33,8 +33,21 @@ func galMulSliceXor(c byte, in, out []byte, o *options) {
 	}
 }
 
-// slice galois add
-func sliceXor(in, out []byte, o *options) {
+// simple slice xor
+func sliceXor(in, out []byte, _ *options) {
+	for len(out) >= 32 {
+		inS := in[:32]
+		v0 := binary.LittleEndian.Uint64(out[:]) ^ binary.LittleEndian.Uint64(inS[:])
+		v1 := binary.LittleEndian.Uint64(out[8:]) ^ binary.LittleEndian.Uint64(inS[8:])
+		v2 := binary.LittleEndian.Uint64(out[16:]) ^ binary.LittleEndian.Uint64(inS[16:])
+		v3 := binary.LittleEndian.Uint64(out[24:]) ^ binary.LittleEndian.Uint64(inS[24:])
+		binary.LittleEndian.PutUint64(out[:], v0)
+		binary.LittleEndian.PutUint64(out[8:], v1)
+		binary.LittleEndian.PutUint64(out[16:], v2)
+		binary.LittleEndian.PutUint64(out[24:], v3)
+		out = out[32:]
+		in = in[32:]
+	}
 	for n, input := range in {
 		out[n] ^= input
 	}
diff --git a/reedsolomon_test.go b/reedsolomon_test.go
index 411d55e..ce5d176 100644
--- a/reedsolomon_test.go
+++ b/reedsolomon_test.go
@@ -28,6 +28,7 @@ func TestMain(m *testing.M) {
 }
 
 func testOptions(o ...Option) []Option {
+	o = append(o, WithFastOneParityMatrix())
 	if *noSSSE3 {
 		o = append(o, withSSSE3(false))
 	}