update kcp-go package

vendor/github.com/klauspost/reedsolomon/galois_amd64.go

@@ -0,0 +1,92 @@
+//+build !noasm
+//+build !appengine
+//+build !gccgo
+
+// Copyright 2015, Klaus Post, see LICENSE for details.
+
+package reedsolomon
+
+//go:noescape
+func galMulSSSE3(low, high, in, out []byte)
+
+//go:noescape
+func galMulSSSE3Xor(low, high, in, out []byte)
+
+//go:noescape
+func galMulAVX2Xor(low, high, in, out []byte)
+
+//go:noescape
+func galMulAVX2(low, high, in, out []byte)
+
+//go:noescape
+func sSE2XorSlice(in, out []byte)
+
+// This is what the assembler routines do in blocks of 16 bytes:
+/*
+func galMulSSSE3(low, high, in, out []byte) {
+	for n, input := range in {
+		l := input & 0xf
+		h := input >> 4
+		out[n] = low[l] ^ high[h]
+	}
+}
+
+func galMulSSSE3Xor(low, high, in, out []byte) {
+	for n, input := range in {
+		l := input & 0xf
+		h := input >> 4
+		out[n] ^= low[l] ^ high[h]
+	}
+}
+*/
+
+func galMulSlice(c byte, in, out []byte, o *options) {
+	var done int
+	if o.useAVX2 {
+		galMulAVX2(mulTableLow[c][:], mulTableHigh[c][:], in, out)
+		done = (len(in) >> 5) << 5
+	} else if o.useSSSE3 {
+		galMulSSSE3(mulTableLow[c][:], mulTableHigh[c][:], in, out)
+		done = (len(in) >> 4) << 4
+	}
+	remain := len(in) - done
+	if remain > 0 {
+		mt := mulTable[c][:256]
+		for i := done; i < len(in); i++ {
+			out[i] = mt[in[i]]
+		}
+	}
+}
+
+func galMulSliceXor(c byte, in, out []byte, o *options) {
+	var done int
+	if o.useAVX2 {
+		galMulAVX2Xor(mulTableLow[c][:], mulTableHigh[c][:], in, out)
+		done = (len(in) >> 5) << 5
+	} else if o.useSSSE3 {
+		galMulSSSE3Xor(mulTableLow[c][:], mulTableHigh[c][:], in, out)
+		done = (len(in) >> 4) << 4
+	}
+	remain := len(in) - done
+	if remain > 0 {
+		mt := mulTable[c][:256]
+		for i := done; i < len(in); i++ {
+			out[i] ^= mt[in[i]]
+		}
+	}
+}
+
+// slice galois add
+func sliceXor(in, out []byte, sse2 bool) {
+	var done int
+	if sse2 {
+		sSE2XorSlice(in, out)
+		done = (len(in) >> 4) << 4
+	}
+	remain := len(in) - done
+	if remain > 0 {
+		for i := done; i < len(in); i++ {
+			out[i] ^= in[i]
+		}
+	}
+}