diff --git a/paddle/fluid/operators/math/cpu_vec.h b/paddle/fluid/operators/math/cpu_vec.h
index a2e2b5a7fed03a480ce81fba03806560824f8ef3..52f072eb0e01ebb05f6a2c9846bdf54c5b4c641e 100644
--- a/paddle/fluid/operators/math/cpu_vec.h
+++ b/paddle/fluid/operators/math/cpu_vec.h
@@ -77,6 +77,122 @@ inline void vec_sigmoid(const int n, const T* x, T* y) {
   }
 }
 
+template <>
+inline void vec_sigmoid<float, platform::jit::avx>(const int n, const float* x,
+                                                   float* y) {
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block) {  // can use larger threshold if necessary
+    vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+    return;
+  }
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);
+  __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);
+  __m256 zeros = _mm256_setzero_ps();
+  __m256 tmp;
+#define MOVE_ONE_STEP              \
+  tmp = _mm256_loadu_ps(x + i);    \
+  tmp = _mm256_max_ps(tmp, min);   \
+  tmp = _mm256_min_ps(tmp, max);   \
+  tmp = _mm256_sub_ps(zeros, tmp); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+  if (rest != 0) {
+    i = n - block;
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+
+  vec_exp<float>(n, y, y);
+
+  __m256 ones = _mm256_set1_ps(1.0f);
+#define MOVE_ONE_STEP             \
+  tmp = _mm256_loadu_ps(y + i);   \
+  tmp = _mm256_add_ps(ones, tmp); \
+  tmp = _mm256_div_ps(ones, tmp); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest == 0) {
+    return;
+  }
+  // can not continue move step
+  for (i = n - rest; i < n; ++i) {
+    y[i] = 1.f / (1.f + y[i]);
+  }
+#else
+  vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+#endif
+}
+
+template <>
+inline void vec_sigmoid<float, platform::jit::avx2>(const int n, const float* x,
+                                                    float* y) {
+  vec_sigmoid<float, platform::jit::avx>(n, x, y);
+}
+
+template <>
+inline void vec_sigmoid<float, platform::jit::avx512_common>(const int n,
+                                                             const float* x,
+                                                             float* y) {
+#ifdef __AVX512F__
+  constexpr int block = AVX512_FLOAT_BLOCK;
+  if (n < block) {
+    vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+    return;
+  }
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m512 max = _mm512_set1_ps(SIGMOID_THRESHOLD_MAX);
+  __m512 min = _mm512_set1_ps(SIGMOID_THRESHOLD_MIN);
+  __m512 zeros = _mm512_setzero_ps();
+  __m512 tmp;
+#define MOVE_ONE_STEP              \
+  tmp = _mm512_loadu_ps(x + i);    \
+  tmp = _mm512_max_ps(tmp, min);   \
+  tmp = _mm512_min_ps(tmp, max);   \
+  tmp = _mm512_sub_ps(zeros, tmp); \
+  _mm512_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+  if (rest != 0) {
+    i = n - block;
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+
+  vec_exp<float>(n, y, y);
+
+  __m512 ones = _mm512_set1_ps(1.0f);
+#define MOVE_ONE_STEP             \
+  tmp = _mm512_loadu_ps(y + i);   \
+  tmp = _mm512_add_ps(ones, tmp); \
+  tmp = _mm512_div_ps(ones, tmp); \
+  _mm512_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest == 0) {
+    return;
+  }
+  for (i = n - rest; i < n; ++i) {
+    y[i] = 1.f / (1.f + y[i]);
+  }
+#else
+  vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+#endif
+}
+
 template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
 inline void vec_tanh(const int n, const T* x, T* y) {
   for (int i = 0; i < n; ++i) {
diff --git a/paddle/fluid/operators/math/cpu_vec_test.cc b/paddle/fluid/operators/math/cpu_vec_test.cc
index 0888e44fa655f1e8f9dc774bb71efcf3e234c142..8b0e9c086a5709e2d21885442d072e14b2f57dc1 100644
--- a/paddle/fluid/operators/math/cpu_vec_test.cc
+++ b/paddle/fluid/operators/math/cpu_vec_test.cc
@@ -104,7 +104,7 @@ void TestAndBench(const int n, std::function<void(const int, const T*, T*)> tgt,
 TEST(CpuVecTest, sigmoid) {
   namespace jit = paddle::platform::jit;
   using namespace paddle::operators::math;  // NOLINT
-  for (auto sz : {1, 2, 15, 16, 32, 128, 200, 512}) {
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestAndBench<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
     TestAndBench<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
     TestAndBench<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
@@ -117,7 +117,7 @@ TEST(CpuVecTest, sigmoid) {
 TEST(CpuVecTest, tanh) {
   namespace jit = paddle::platform::jit;
   using namespace paddle::operators::math;  // NOLINT
-  for (auto sz : {1, 2, 15, 16, 32, 128, 200, 512}) {
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestAndBench<float>(sz, vec_tanh<float>, ref_tanh<float>);
     TestAndBench<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
     TestAndBench<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
@@ -130,7 +130,7 @@ TEST(CpuVecTest, tanh) {
 TEST(CpuVecTest, relu) {
   namespace jit = paddle::platform::jit;
   using namespace paddle::operators::math;  // NOLINT
-  for (auto sz : {1, 2, 15, 16, 32, 128, 200, 512}) {
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestAndBench<float>(sz, vec_relu<float>, ref_relu<float>);
     TestAndBench<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
     TestAndBench<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);