add debug code

src/framework/cl/cl_half.cpp

@@ -498,3 +498,15 @@ float half2float(half_t h) {
                exponenttable[h >> 10];
   return *reinterpret_cast<float*>(&v);
 }
+
+void FloatArray2HalfArray(float *f_array, half_t *h_array, int count) {
+  for (int i = 0; i < count; ++i) {
+    h_array[i] = float2half(f_array[i]);
+  }
+}
+
+void HalfArray2FloatArray(half_t *h_array, float *f_array, int count) {
+  for (int i = 0; i < count; ++i) {
+    f_array[i] = float2half(h_array[i]);
+  }
+}

src/framework/cl/cl_half.h

@@ -18,4 +18,9 @@ limitations under the License. */
 typedef uint16_t half_t;
 
 half_t float2half(float f);
+
 float half2float(half_t h);
+
+void FloatArray2HalfArray(float *f_array, half_t *h_array, int count);
+
+void HalfArray2FloatArray(half_t *h_array, float *f_array, int count);

src/framework/cl/cl_image.h

@@ -226,5 +226,9 @@ void TensorToCLImage(Tensor *tensor, CLImage *image);
 
 void CLImageToTensor(CLImage *image, Tensor *tensor);
 
+#ifdef PADDLE_MOBILE_DEBUG
+Print &operator<<(Print &printer, const CLImage &image);
+#endif
+
 }  // namespace framework
 }  // namespace paddle_mobile

src/framework/executor.cpp

@@ -414,7 +414,7 @@ std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
     }
   }
 #else
-  for (int i = 0; i < ops.size(); i++) {
+  for (int i = 0; i < 1; i++) {
 #ifdef PADDLE_MOBILE_PROFILE
     struct timespec ts;
     clock_gettime(CLOCK_MONOTONIC, &ts);
@@ -428,6 +428,9 @@ std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
 #endif
   }
 #endif
+
+  DLOG << " predict return nullptr";
+  return nullptr;
   auto last_op = ops.rbegin();
   auto output_map = (*last_op)->Outputs();
   std::vector<std::string> out_keys = (*last_op)->GetOutKeys();
@@ -647,13 +650,18 @@ std::vector<typename Executor<Dtype, P>::Ptype> Executor<Dtype, P>::Predict(
     const std::vector<Ptype> &input, const std::vector<int64_t> &dims) {
   framework::Tensor tensor(input, framework::make_ddim(dims));
   std::shared_ptr<framework::Tensor> output_tensor = Predict(tensor, 0);
-  Executor<Dtype, P>::Ptype *output_ptr =
-      output_tensor->data<typename Executor<Dtype, P>::Ptype>();
-  std::vector<typename Executor<Dtype, P>::Ptype> result_vector;
-  for (int j = 0; j < output_tensor->numel(); ++j) {
-    result_vector.push_back(output_ptr[j]);
+  if (output_tensor != nullptr) {
+    Executor<Dtype, P>::Ptype *output_ptr =
+            output_tensor->data<typename Executor<Dtype, P>::Ptype>();
+    std::vector<typename Executor<Dtype, P>::Ptype> result_vector;
+    for (int j = 0; j < output_tensor->numel(); ++j) {
+      result_vector.push_back(output_ptr[j]);
+    }
+    return result_vector;
+  } else {
+    DLOG << "return  empty vector";
+    return {};
   }
-  return result_vector;
 }
 
 #ifdef PADDLE_MOBILE_FPGA

src/framework/operator.cpp

@@ -57,7 +57,9 @@ void OperatorBase<Dtype>::CheckAllInputOutputSet() const {}
 
 template <typename Dtype>
 void OperatorBase<Dtype>::Run() {
+  DLOG << " begin run " << type_;
   RunImpl();
+  DLOG << " end run " << type_;
 #ifdef PADDLE_MOBILE_DEBUG
   DLOG << "-------------" << type_ << "----------------------------";
   vector<string> input_keys = GetInputKeys();
@@ -100,8 +102,9 @@ void OperatorBase<Dtype>::Run() {
 #ifdef PADDLE_MOBILE_CL
         if (type_ == "fetch") {
           Tensor *tensor = vari->template GetMutable<framework::LoDTensor>();
-          if (tensor)
+          if (tensor){
             DLOG << type_ << " output- " << key << "=" << tensor->dims();
+          }
         } else {
           CLImage *cl_image = vari->template GetMutable<framework::CLImage>();
           //          cl_command_queue commandQueue =

src/operators/op_param.h

@@ -948,6 +948,7 @@ class FetchParam : public OpParam {
     input_x_ = InputXFrom<GType>(inputs, scope);
     out_ = OutFrom(outputs, scope);
   }
+
   const RType *InputX() const { return input_x_; }
   Tensor *Out() const { return out_; }
 

test/net/test_mobilenet_GPU.cpp

@@ -34,23 +34,23 @@ int main() {
     GetInput<float>(g_test_image_1x3x224x224_banana, &input, dims);
 
     auto vec_result = paddle_mobile.Predict(input, dims);
-    std::vector<float>::iterator biggest =
-        std::max_element(std::begin(vec_result), std::end(vec_result));
-    std::cout << " Max element is " << *biggest << " at position "
-              << std::distance(std::begin(vec_result), biggest) << std::endl;
-
-    // 预热十次
-    for (int i = 0; i < 10; ++i) {
-      auto vec_result = paddle_mobile.Predict(input, dims);
-    }
-    auto time3 = paddle_mobile::time();
-    for (int i = 0; i < 10; ++i) {
-      auto vec_result = paddle_mobile.Predict(input, dims);
-    }
-    DLOG << vec_result;
-    auto time4 = paddle_mobile::time();
-    std::cout << "predict cost :" << paddle_mobile::time_diff(time3, time4) / 10
-              << "ms" << std::endl;
+//    std::vector<float>::iterator biggest =
+//        std::max_element(std::begin(vec_result), std::end(vec_result));
+//    std::cout << " Max element is " << *biggest << " at position "
+//              << std::distance(std::begin(vec_result), biggest) << std::endl;
+
+
+//    for (int i = 0; i < 10; ++i) {
+//      auto vec_result = paddle_mobile.Predict(input, dims);
+//    }
+//    auto time3 = paddle_mobile::time();
+//    for (int i = 0; i < 10; ++i) {
+//      auto vec_result = paddle_mobile.Predict(input, dims);
+//    }
+//    DLOG << vec_result;
+//    auto time4 = paddle_mobile::time();
+//    std::cout << "predict cost :" << paddle_mobile::time_diff(time3, time4) / 10 << "ms"
+//              << std::endl;
   }
 
   std::cout << "如果结果Nan请查看: test/images/g_test_image_1x3x224x224_banana "