diff --git a/zstd/_generate/gen.go b/zstd/_generate/gen.go
index 1914b0ec7c..b61a637020 100644
--- a/zstd/_generate/gen.go
+++ b/zstd/_generate/gen.go
@@ -561,7 +561,7 @@ type executeSimple struct{}
 //
 // See copyMemory()
 func (e executeSimple) copySize() int {
-	return 16
+	return 32
 }
 
 func (e executeSimple) generateProcedure(name string) {
@@ -716,14 +716,19 @@ func (e executeSimple) generateProcedure(name string) {
 func (e executeSimple) copyMemory(suffix string, src, dst, length reg.GPVirtual) {
 	label := "copy_" + suffix
 	ofs := GP64()
-	s := Mem{Base: src, Index: ofs, Scale: 1}
-	d := Mem{Base: dst, Index: ofs, Scale: 1}
+	s0 := Mem{Base: src, Index: ofs, Scale: 1}
+	s1 := Mem{Base: src, Index: ofs, Scale: 1, Disp: 16}
+	d0 := Mem{Base: dst, Index: ofs, Scale: 1}
+	d1 := Mem{Base: dst, Index: ofs, Scale: 1, Disp: 16}
 
 	XORQ(ofs, ofs)
 	Label(label)
-	t := XMM()
-	MOVUPS(s, t)
-	MOVUPS(t, d)
+	t0 := XMM()
+	t1 := XMM()
+	MOVUPS(s0, t0)
+	MOVUPS(s1, t1)
+	MOVUPS(t0, d0)
+	MOVUPS(t1, d1)
 	ADDQ(U8(e.copySize()), ofs)
 	CMPQ(ofs, length)
 	JB(LabelRef(label))
diff --git a/zstd/seqdec_amd64.s b/zstd/seqdec_amd64.s
index bee3fd8a04..699a19723d 100644
--- a/zstd/seqdec_amd64.s
+++ b/zstd/seqdec_amd64.s
@@ -619,7 +619,7 @@ main_loop:
 
 	// Check if we won't overflow ctx.out while fast copying
 	LEAQ (R10)(R11*1), R12
-	LEAQ 16(R8)(R12*1), R13
+	LEAQ 32(R8)(R12*1), R13
 	CMPQ R13, SI
 	JA   slow_path
 
@@ -634,8 +634,10 @@ main_loop:
 
 copy_1:
 	MOVUPS (DI)(R12*1), X0
+	MOVUPS 16(DI)(R12*1), X1
 	MOVUPS X0, (BX)(R12*1)
-	ADDQ   $0x10, R12
+	MOVUPS X1, 16(BX)(R12*1)
+	ADDQ   $0x20, R12
 	CMPQ   R12, R11
 	JB     copy_1
 	ADDQ   R11, DI
@@ -658,8 +660,10 @@ copy_match:
 
 copy_2:
 	MOVUPS (R12)(R11*1), X0
+	MOVUPS 16(R12)(R11*1), X1
 	MOVUPS X0, (BX)(R11*1)
-	ADDQ   $0x10, R11
+	MOVUPS X1, 16(BX)(R11*1)
+	ADDQ   $0x20, R11
 	CMPQ   R11, R10
 	JB     copy_2
 	ADDQ   R10, BX