Merge branch 'superresoltion' of https://github.com/codeWorm2015/paddle-mobile into superresoltion

src/framework/executor.cpp

@@ -238,7 +238,7 @@ void Executor<Device, T>::InitCombineMemory() {
 
 template <typename Device, typename T>
 void Executor<Device, T>::InitNoPersistableMemory(
-    const LoDTensor &input_tensor) {
+    const Tensor &input_tensor) {
   for (const auto &block : program_desc_->Blocks()) {
     for (const auto &var_desc : block->Vars()) {
       auto var = program_.scope->Var(var_desc->Name());
@@ -336,9 +336,9 @@ void Executor<Device, T>::SetInput(const Tensor &input,
   auto *target_tensor = target_var->template GetMutable<LoDTensor>();
 
   if (config_.load_when_predict) {
-    if (target_tensor->IsInitialized() &&
-        target_tensor->dims() != input.dims()) {
-      InitNoPersistableMemory(*target_tensor);
+    if (input_dim_last_ != input.dims()) {
+      InitNoPersistableMemory(input);
+      input_dim_last_ = input.dims();
     }
   }
 
@@ -355,9 +355,9 @@ void Executor<Device, T>::SetInput(const LoDTensor &input,
   auto *target_tensor = target_var->template GetMutable<LoDTensor>();
 
   if (config_.load_when_predict) {
-    if (target_tensor->IsInitialized() &&
-        target_tensor->dims() != input.dims()) {
+    if (input_dim_last_ != input.dims()) {
       InitNoPersistableMemory(*target_tensor);
+      input_dim_last_ = input.dims();
     }
   }
 

src/framework/executor.h

@@ -65,7 +65,7 @@ class Executor {
                       LoDTensor *tensor) const;
   void InitMemory();
   void InitCombineMemory();
-  void InitNoPersistableMemory(const LoDTensor &input_tensor);
+  void InitNoPersistableMemory(const Tensor &input_tensor);
   void LoadMemory(void **data, const std::shared_ptr<VarDesc> var_desc,
                   LoDTensor *tensor);
 #ifdef PADDLE_MOBILE_CL

src/operators/math/depthwise_conv3x3.cpp

@@ -253,34 +253,29 @@ void DepthwiseConv3x3s1p1(const framework::Tensor *input,
                           framework::Tensor *output, framework::Tensor *bias,
                           bool if_bias, bool if_relu) {
 #if __ARM_NEON
-  const float *input_data = input->data<float>();
-  const float *filter_data = filter->data<float>();
-  float *output_data = output->mutable_data<float>();
-  const float *bias_data;
-  if (if_bias) {
-    bias_data = bias->data<float>();
-  }
-
-  const int h = static_cast<int>(input->dims()[2]);
-  const int w = static_cast<int>(input->dims()[3]);
-  //  const int l = h;
+  const float *bias_data = bias->data<float>();
   const int batch_size = static_cast<int>(input->dims()[0]);
   const int c = static_cast<int>(input->dims()[1]);
+  const int h = static_cast<int>(input->dims()[2]);
+  const int w = static_cast<int>(input->dims()[3]);
   const int hxw = h * w;
-  float32x4_t vbias = vdupq_n_f32(0.0);
+  // const int l = h;
 
   // leftTop, rightTop, leftBottom, rightBottom
-  int lt = 0;
-  int rt = w - 1;
-  int lb = (h - 1) * w;
-  int rb = h * w - 1;
+  const int lt = 0;
+  const int rt = w - 1;
+  const int lb = (h - 1) * w;
+  const int rb = h * w - 1;
 
   float32x4_t zero = vdupq_n_f32(0.0);
 
   for (int b = 0; b < batch_size; ++b) {
-    const float *filter_data_tmp = filter_data;
-
+#pragma omp parallel for
     for (int j = 0; j < c; ++j) {
+      const float *filter_data_tmp = filter->data<float>() + j * 9;
+      const float *input_data = input->data<float>() + j * hxw;
+      float *output_data = output->mutable_data<float>() + j * hxw;
+      float32x4_t vbias;
       if (if_bias) {
         vbias = vdupq_n_f32(bias_data[j]);
       }
@@ -552,9 +547,6 @@ void DepthwiseConv3x3s1p1(const framework::Tensor *input,
           }
         }
       }
-      output_data += hxw;
-      input_data += hxw;
-      filter_data_tmp += 9;
     }
   }
 #endif