diff --git a/.gitignore b/.gitignore
old mode 100755
new mode 100644
diff --git a/rs_amd64.go b/rs_amd64.go
index b3a486f..75301d4 100644
--- a/rs_amd64.go
+++ b/rs_amd64.go
@@ -3,6 +3,8 @@ package reedsolomon
 import (
 	"errors"
 	"sync"
+
+	"github.com/templexxx/cpufeat"
 )
 
 // SIMD Instruction Extensions
@@ -12,22 +14,25 @@ const (
 	ssse3
 )
 
-func getEXT() int {
-	if hasAVX2() {
-		return avx2
-	} else if hasSSSE3() {
-		return ssse3
+var extension = none
+
+func init() {
+	getEXT()
+}
+
+func getEXT() {
+	if cpufeat.X86.HasAVX2 {
+		extension = avx2
+		return
+	} else if cpufeat.X86.HasSSSE3 {
+		extension = ssse3
+		return
 	} else {
-		return none
+		extension = none
+		return
 	}
 }
 
-//go:noescape
-func hasAVX2() bool
-
-//go:noescape
-func hasSSSE3() bool
-
 //go:noescape
 func copy32B(dst, src []byte) // Need SSE2(introduced in 2001)
 
@@ -75,14 +80,13 @@ type (
 
 func newRS(d, p int, em matrix) (enc Encoder) {
 	g := em[d*d:]
-	ext := getEXT()
-	if ext == none {
+	if extension == none {
 		return &encBase{data: d, parity: p, encode: em, gen: g}
 	}
 	t := make([]byte, d*p*32)
 	initTbl(g, p, d, t)
 	ok := okCache(d, p)
-	if ext == avx2 {
+	if extension == avx2 {
 		e := &encAVX2{data: d, parity: p, encode: em, gen: g, tbl: t, enableCache: ok,
 			inverseCache: iCache{data: make(map[uint32][]byte)}}
 		return e
diff --git a/rs_amd64.s b/rs_amd64.s
index 7d7e7a1..b8f4885 100644
--- a/rs_amd64.s
+++ b/rs_amd64.s
@@ -323,16 +323,6 @@ one16b:
 	JNE      ymm
 	RET
 
-TEXT ·hasAVX2(SB), NOSPLIT, $0
-	XORQ AX, AX
-	XORQ CX, CX
-	ADDL $7, AX
-	CPUID
-	SHRQ $5, BX
-	ANDQ $1, BX
-	MOVB BX, ret+0(FP)
-	RET
-
 // func mulVectSSSE3(tbl, d, p []byte)
 TEXT ·mulVectSSSE3(SB), NOSPLIT, $0
     MOVQ    i+24(FP), in
@@ -399,15 +389,6 @@ loop:
 	JNZ    loop
 	RET
 
-TEXT ·hasSSSE3(SB), NOSPLIT, $0
-	XORQ AX, AX
-	INCL AX
-	CPUID
-	SHRQ $9, CX
-	ANDQ $1, CX
-	MOVB CX, ret+0(FP)
-	RET
-
 // func copy32B(dst, src []byte)
 TEXT ·copy32B(SB), NOSPLIT, $0
     MOVQ dst+0(FP), SI