diff --git a/src/framework/load_ops.h b/src/framework/load_ops.h
index 2b76b0158fe06e8678208f6f98fcdb71f8d91e51..4c6842572e49daa283efa2d92bd43e4687d92e26 100644
--- a/src/framework/load_ops.h
+++ b/src/framework/load_ops.h
@@ -221,5 +221,9 @@ LOAD_FUSION_MATCHER(fusion_conv_bn);
 #ifdef ELEMENTWISESUB_OP
 LOAD_OP1(elementwise_sub, CPU)
 #endif
+#ifdef QUANT_OP
 LOAD_OP1(quantize, CPU);
+#endif
+#ifdef DEQUANT_OP
 LOAD_OP1(dequantize, CPU);
+#endif
diff --git a/src/operators/kernel/arm/quantize_kernel.cpp b/src/operators/kernel/arm/quantize_kernel.cpp
index e7552d2602b31f9a5c10e3d81122babae8fcf1a8..11a1f0a53d4886e1a07d258b76b3827671471dca 100644
--- a/src/operators/kernel/arm/quantize_kernel.cpp
+++ b/src/operators/kernel/arm/quantize_kernel.cpp
@@ -135,11 +135,15 @@ static void quantize_round_to_even(const Tensor *input, const float scale,
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = size >> 4;
   size_t remain = size & 0xF;
+
+  #pragma omp parallel for
   for (size_t i = 0; i < loop; ++i) {
-    float32x4_t r0 = vld1q_f32(x);
-    float32x4_t r1 = vld1q_f32(x + 4);
-    float32x4_t r2 = vld1q_f32(x + 8);
-    float32x4_t r3 = vld1q_f32(x + 12);
+    const float *local_x = x + (i << 4);
+    int8_t *local_y = y + (i << 4);
+    float32x4_t r0 = vld1q_f32(local_x);
+    float32x4_t r1 = vld1q_f32(local_x + 4);
+    float32x4_t r2 = vld1q_f32(local_x + 8);
+    float32x4_t r3 = vld1q_f32(local_x + 12);
     r0 = vmulq_n_f32(r0, scale);
     r1 = vmulq_n_f32(r1, scale);
     r2 = vmulq_n_f32(r2, scale);
@@ -156,12 +160,12 @@ static void quantize_round_to_even(const Tensor *input, const float scale,
     int16x8_t q6 = vcombine_s16(d2, d3);
     int8x8_t d5 = vmovn_s16(q5);
     int8x8_t d6 = vmovn_s16(q6);
-    vst1_s8(y, d5);
-    vst1_s8(y + 8, d6);
-    x += 16;
-    y += 16;
+    vst1_s8(local_y, d5);
+    vst1_s8(local_y + 8, d6);
   }
   size = remain;
+  x += (loop << 4);
+  y += (loop << 4);
 #endif
   for (size_t i = 0; i < size; ++i) {
     float value = x[i] * scale;
@@ -187,11 +191,15 @@ static void quantize_round_to_zero(const Tensor *input, const float scale,
 #ifdef defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = size >> 4;
   size_t remain = size & 0xF;
+
+  #pragma omp parallel for
   for (size_t i = 0; i < loop; ++i) {
-    float32x4_t r0 = vld1q_f32(x);
-    float32x4_t r1 = vld1q_f32(x + 4);
-    float32x4_t r2 = vld1q_f32(x + 8);
-    float32x4_t r3 = vld1q_f32(x + 12);
+    const float *local_x = x + (i << 4);
+    int8_t *local_y = y + (i << 4);
+    float32x4_t r0 = vld1q_f32(local_x);
+    float32x4_t r1 = vld1q_f32(local_x + 4);
+    float32x4_t r2 = vld1q_f32(local_x + 8);
+    float32x4_t r3 = vld1q_f32(local_x + 12);
     r0 = vmulq_n_f32(r0, scale);
     r1 = vmulq_n_f32(r1, scale);
     r2 = vmulq_n_f32(r2, scale);
@@ -208,12 +216,12 @@ static void quantize_round_to_zero(const Tensor *input, const float scale,
     int16x8_t q6 = vcombine_s16(d2, d3);
     int8x8_t d5 = vmovn_s16(q5);
     int8x8_t d6 = vmovn_s16(q6);
-    vst1_s8(y, d5);
-    vst1_s8(y + 8, d6);
-    x += 16;
-    y += 16;
+    vst1_s8(local_y, d5);
+    vst1_s8(local_y + 8, d6);
   }
   size = remain;
+  x += (loop << 4);
+  y += (loop << 4);
 #endif
   for (size_t i = 0; i < size; ++i) {
     y[i] = trunc(x[i] * scale);
@@ -228,11 +236,15 @@ static void quantize_round_to_nearest(const Tensor *input, const float scale,
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = size >> 4;
   size_t remain = size & 0xF;
+
+  #pragma omp parallel for
   for (size_t i = 0; i < loop; ++i) {
-    float32x4_t r0 = vld1q_f32(x);
-    float32x4_t r1 = vld1q_f32(x + 4);
-    float32x4_t r2 = vld1q_f32(x + 8);
-    float32x4_t r3 = vld1q_f32(x + 12);
+    const float *local_x = x + (i << 4);
+    int8_t *local_y = y + (i << 4);
+    float32x4_t r0 = vld1q_f32(local_x);
+    float32x4_t r1 = vld1q_f32(local_x + 4);
+    float32x4_t r2 = vld1q_f32(local_x + 8);
+    float32x4_t r3 = vld1q_f32(local_x + 12);
     r0 = vmulq_n_f32(r0, scale);
     r1 = vmulq_n_f32(r1, scale);
     r2 = vmulq_n_f32(r2, scale);
@@ -249,12 +261,12 @@ static void quantize_round_to_nearest(const Tensor *input, const float scale,
     int16x8_t q6 = vcombine_s16(d2, d3);
     int8x8_t d5 = vmovn_s16(q5);
     int8x8_t d6 = vmovn_s16(q6);
-    vst1_s8(y, d5);
-    vst1_s8(y + 8, d6);
-    x += 16;
-    y += 16;
+    vst1_s8(local_y, d5);
+    vst1_s8(local_y + 8, d6);
   }
   size = remain;
+  x += (loop << 4);
+  y += (loop << 4);
 #endif
   for (size_t i = 0; i < size; ++i) {
     y[i] = round(x[i] * scale);
diff --git a/src/operators/kernel/central-arm-func/elementwise_add_arm_func.h b/src/operators/kernel/central-arm-func/elementwise_add_arm_func.h
index 0c01ef0072444479d2d2e2f7676b842d89e432ec..b6288380a04c71b3d6467f7f6648db046ae9acc9 100644
--- a/src/operators/kernel/central-arm-func/elementwise_add_arm_func.h
+++ b/src/operators/kernel/central-arm-func/elementwise_add_arm_func.h
@@ -58,6 +58,7 @@ void ElementwiseAddCompute(const ElementwiseAddParam<CPU> &param) {
   const float *input_data = input_x->data<float>();
   float *output_data = Out->mutable_data<float>();
   for (int i = 0; i < batch; ++i) {
+    #pragma omp parallel for
     for (int j = 0; j < channels; ++j) {
       size_t offset = (i * channels + j) * elementwise_num;
       const float *input = input_data + offset;
diff --git a/test/net/test_googlenet.cpp b/test/net/test_googlenet.cpp
index b489f33bab364bebdc81c87a7b6c200082f3e079..f7d29942224b51734cf62988ba8f271f1fa05bc3 100644
--- a/test/net/test_googlenet.cpp
+++ b/test/net/test_googlenet.cpp
@@ -25,7 +25,7 @@ int main() {
   paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
 #endif
 
-  paddle_mobile.SetThreadNum(1);
+  paddle_mobile.SetThreadNum(4);
   bool optimize = true;
   auto time1 = time();
   if (paddle_mobile.Load(g_googlenet, optimize)) {