diff --git a/CMakeLists.txt b/CMakeLists.txt
index a24ddaeb..05609af2 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -12,7 +12,7 @@ include_directories(include)
 
 option(ENABLE_SM4_TBOX "Enable SM4 merged S-Box implementation" ON)
 option(ENABLE_SM4_AARCH64 "Enable SM4 AARCH64 assembly implementation" OFF)
-option(ENABLE_GMUL_AARCH64 "Enable GF(2^128) Multiplication AArch64 assembly" OFF)
+option(ENABLE_SM4_CTR_AESNI_AVX "Enable SM4 CTR AESNI+AVX assembly implementation" OFF)
 
 option(ENABLE_SM4_ECB "Enable SM4 ECB mode" OFF)
 option(ENABLE_SM4_OFB "Enable SM4 OFB mode" OFF)
@@ -20,6 +20,7 @@ option(ENABLE_SM4_CFB "Enable SM4 CFB mode" OFF)
 option(ENABLE_SM4_CBC_MAC "Enable SM4-CBC-MAC" OFF)
 option(ENABLE_SM4_CCM "Enable SM4 CCM mode" OFF)
 
+option(ENABLE_GMUL_AARCH64 "Enable GF(2^128) Multiplication AArch64 assembly" OFF)
 
 
 set(src
@@ -322,12 +323,11 @@ if (ENABLE_SM4_AARCH64)
 	enable_language(ASM)
 endif()
 
-option(ENABLE_SM4_AESNI_AVX "Enable SM4 AESNI+AVX assembly implementation" OFF)
-if (ENABLE_SM4_AESNI_AVX)
-	message(STATUS "ENABLE_SM4_AESNI_AVX is ON")
-	add_definitions(-DENABLE_SM4_AESNI_AVX)
-	list(APPEND src src/sm4_aesni_avx.c)
-	list(APPEND tests sm4_aesni_avx)
+if (ENABLE_SM4_CTR_AESNI_AVX)
+	message(STATUS "ENABLE_SM4_CTR_AESNI_AVX is ON")
+	list(FIND src src/sm4_ctr.c sm4_ctr_index)
+	list(REMOVE_AT src ${sm4_ctr_index})
+	list(INSERT src ${sm4_ctr_index} src/sm4_ctr_aesni_avx.c)
 	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=native")
 endif()
 
diff --git a/src/sm4_ctr.c b/src/sm4_ctr.c
index 1679481e..b2a91486 100644
--- a/src/sm4_ctr.c
+++ b/src/sm4_ctr.c
@@ -22,7 +22,6 @@ static void ctr_incr(uint8_t a[16])
 	}
 }
 
-#ifndef ENABLE_SM4_AESNI_AVX
 void sm4_ctr_encrypt(const SM4_KEY *key, uint8_t ctr[16], const uint8_t *in, size_t inlen, uint8_t *out)
 {
 	uint8_t block[16];
@@ -38,7 +37,6 @@ void sm4_ctr_encrypt(const SM4_KEY *key, uint8_t ctr[16], const uint8_t *in, siz
 		inlen -= len;
 	}
 }
-#endif
 
 int sm4_ctr_encrypt_init(SM4_CTR_CTX *ctx,
 	const uint8_t key[SM4_BLOCK_SIZE], const uint8_t ctr[SM4_BLOCK_SIZE])
diff --git a/src/sm4_aesni_avx.c b/src/sm4_ctr_aesni_avx.c
similarity index 75%
rename from src/sm4_aesni_avx.c
rename to src/sm4_ctr_aesni_avx.c
index ee75142b..1e2d9dd9 100644
--- a/src/sm4_aesni_avx.c
+++ b/src/sm4_ctr_aesni_avx.c
@@ -1,5 +1,5 @@
 /*
- *  Copyright 2014-2022 The GmSSL Project. All Rights Reserved.
+ *  Copyright 2014-2024 The GmSSL Project. All Rights Reserved.
  *
  *  Licensed under the Apache License, Version 2.0 (the License); you may
  *  not use this file except in compliance with the License.
@@ -32,16 +32,13 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 SOFTWARE.
 */
 
-#include <stdio.h>
-#include <string.h>
-#include <stdlib.h>
-#include <stdint.h>
+#include <x86intrin.h>
 #include <gmssl/mem.h>
 #include <gmssl/sm4.h>
-#include <x86intrin.h>
+#include <gmssl/error.h>
 
 
-void sm4_aesni_avx_encrypt(const uint32_t rk[32], const uint8_t in[16 * 4], uint8_t out[16 * 4])
+static void sm4_aesni_avx_encrypt(const uint32_t rk[32], const uint8_t in[16 * 4], uint8_t out[16 * 4])
 {
 	// nibble mask
 	const __m128i c0f __attribute__((aligned(0x10))) = {
@@ -188,3 +185,66 @@ void sm4_ctr_encrypt(const SM4_KEY *key, uint8_t ctr[16], const uint8_t *in, siz
 
 	memset(blocks, 0, sizeof(blocks));
 }
+
+int sm4_ctr_encrypt_init(SM4_CTR_CTX *ctx,
+	const uint8_t key[SM4_BLOCK_SIZE], const uint8_t ctr[SM4_BLOCK_SIZE])
+{
+	sm4_set_encrypt_key(&ctx->sm4_key, key);
+	memcpy(ctx->ctr, ctr, SM4_BLOCK_SIZE);
+	memset(ctx->block, 0, SM4_BLOCK_SIZE);
+	ctx->block_nbytes = 0;
+	return 1;
+}
+
+int sm4_ctr_encrypt_update(SM4_CTR_CTX *ctx,
+	const uint8_t *in, size_t inlen, uint8_t *out, size_t *outlen)
+{
+	size_t left;
+	size_t nblocks;
+	size_t len;
+
+	if (ctx->block_nbytes >= SM4_BLOCK_SIZE) {
+		error_print();
+		return -1;
+	}
+	*outlen = 0;
+	if (ctx->block_nbytes) {
+		left = SM4_BLOCK_SIZE - ctx->block_nbytes;
+		if (inlen < left) {
+			memcpy(ctx->block + ctx->block_nbytes, in, inlen);
+			ctx->block_nbytes += inlen;
+			return 1;
+		}
+		memcpy(ctx->block + ctx->block_nbytes, in, left);
+		sm4_ctr_encrypt(&ctx->sm4_key, ctx->ctr, ctx->block, SM4_BLOCK_SIZE, out);
+		in += left;
+		inlen -= left;
+		out += SM4_BLOCK_SIZE;
+		*outlen += SM4_BLOCK_SIZE;
+	}
+	if (inlen >= SM4_BLOCK_SIZE) {
+		nblocks = inlen / SM4_BLOCK_SIZE;
+		len = nblocks * SM4_BLOCK_SIZE;
+		sm4_ctr_encrypt(&ctx->sm4_key, ctx->ctr, in, len, out);
+		in += len;
+		inlen -= len;
+		out += len;
+		*outlen += len;
+	}
+	if (inlen) {
+		memcpy(ctx->block, in, inlen);
+	}
+	ctx->block_nbytes = inlen;
+	return 1;
+}
+
+int sm4_ctr_encrypt_finish(SM4_CTR_CTX *ctx, uint8_t *out, size_t *outlen)
+{
+	if (ctx->block_nbytes >= SM4_BLOCK_SIZE) {
+		error_print();
+		return -1;
+	}
+	sm4_ctr_encrypt(&ctx->sm4_key, ctx->ctr, ctx->block, ctx->block_nbytes, out);
+	*outlen = ctx->block_nbytes;
+	return 1;
+}