add some function

src/common/common.h

@@ -28,4 +28,4 @@ inline double time_diff(Time t1, Time t2) {
   return counter.count() / 1000.0;
 }
 
-}
+}  // namespace paddle_mobile

src/framework/cl/cl_image.h

@@ -118,7 +118,20 @@ class CLImage {
     cl_image_format cf = {.image_channel_order = CL_RGBA,
                           .image_channel_data_type = CL_HALF_FLOAT};
     // NCHW -> [W * (C+3)/4, H * N]
-    DLOG << tensor_dims_;
+    tensor_dims_ = dim;
+    if (tensor_data) {
+      tensor_data_ = tensor_data;
+    } else {
+      int numel = 1;
+      for (int i = 0; i < dim.size(); i++) {
+        numel *= dim[i];
+      }
+      tensor_data_ = static_cast<float *>(
+          paddle_mobile::memory::Alloc(sizeof(float) * numel));
+      for (int i = 0; i < numel; i++) {
+        tensor_data_[i] = 0;
+      }
+    }
     size_t N, C, H, W;
     if (tensor_dims_.size() == 4) {
       N = tensor_dims_[0];

src/framework/executor.cpp

@@ -936,7 +936,7 @@ void Executor<GPU_CL, Precision::FP32>::InitMemory() {
         cl_image->SetTensorData(tensorInput, ddim);
 
         delete origin_data;
-        paddle_mobile::memory::Free(tensorInput);
+        //        paddle_mobile::memory::Free(tensorInput);
       } else {
         if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
           auto cl_image = var->template GetMutable<framework::CLImage>();

src/framework/operator.cpp

@@ -73,12 +73,15 @@ void OperatorBase<Dtype>::Run() {
         } else {
           CLImage *cl_image = vari->template GetMutable<framework::CLImage>();
           //                        cl_command_queue commandQueue =
-          //              scope_->GetCLScpoe()->CommandQueue(); Tensor *tmp ;
+          //                        scope_->GetCLScpoe()->CommandQueue(); Tensor
+          //                        *tmp ;
           //                        CLImageToTensor(cl_image,tmp,commandQueue);
           //                        tmp->Resize(cl_image->dims());
+          const float *input = cl_image->data<float>();
           if (cl_image) {
-            //                  DLOG<<type_<<" input- "<<key<<"="<<*tmp;
             DLOG << type_ << " input- " << key << "=" << cl_image->dims();
+            //              if(input)
+            //              DLOG<<type_<<" input- "<<key<<"="<<*input;
           }
         }
 
@@ -95,15 +98,24 @@ void OperatorBase<Dtype>::Run() {
       auto vari = scope_->FindVar(var_vec_out[i]);
       if (vari->IsInitialized()) {
 #ifdef PADDLE_MOBILE_CL
+        if (type_ == "fetch") {
+          Tensor *tensor = vari->template GetMutable<framework::LoDTensor>();
+          if (tensor)
+            DLOG << type_ << " output- " << key << "=" << tensor->dims();
+        } else {
           CLImage *cl_image = vari->template GetMutable<framework::CLImage>();
           //          cl_command_queue commandQueue =
           //          scope_->GetCLScpoe()->CommandQueue(); Tensor *tmp ;
           //          CLImageToTensor(cl_image,tmp,commandQueue);
           //          tmp->Resize(cl_image->dims());
           if (cl_image) {
-          //              DLOG<<type_<<" output- "<<key<<"="<<*tmp;
+            const float *output = cl_image->data<float>();
             DLOG << type_ << " output- " << key << "=" << cl_image->dims();
+            //                  if(output)
+            //                  DLOG<<type_<<" output- "<<key<<"="<<*output;
           }
+        }
+
 #else
         Tensor *tensor = vari->template GetMutable<framework::LoDTensor>();
         if (tensor) DLOG << type_ << " output- " << key << "=" << *tensor;

src/operators/kernel/cl/feed_kernel.cpp

@@ -27,17 +27,17 @@ bool FeedKernel<GPU_CL, float>::Init(FeedParam<GPU_CL> *param) {
 
 template <>
 void FeedKernel<GPU_CL, float>::Compute(const FeedParam<GPU_CL> &param) {
-  DLOG << "feed_kernel";
   auto kernel = this->cl_helper_.KernelAt(0);
   cl_int status;
   auto output = param.Out();
-  auto input = param.InputX();
-  DLOG << " input: " << input;
+  const Tensor *input = param.InputX();
+  const float *input_data = nullptr;
+  input_data = input->data<float>();
 
-  const float *input_data = input->data<float>();
   cl_mem cl_image = output->GetCLImage();
   int height = output->dims()[2];
   int width = output->dims()[3];
+  DLOG << output->dims();
   status = clSetKernelArg(kernel, 0, sizeof(cl_mem), &input_data);
   status = clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_image);
   status = clSetKernelArg(kernel, 0, sizeof(cl_mem), &width);

test/net/test_googlenet.cpp

@@ -30,7 +30,8 @@ int main() {
   auto time1 = time();
   if (paddle_mobile.Load(g_googlenet, optimize)) {
     auto time2 = paddle_mobile::time();
-    std::cout << "load cost :" << paddle_mobile::time_diff(time1, time2) << "ms" << std::endl;
+    std::cout << "load cost :" << paddle_mobile::time_diff(time1, time2) << "ms"
+              << std::endl;
     std::vector<float> input;
     std::vector<int64_t> dims{1, 3, 224, 224};
     GetInput<float>(g_test_image_1x3x224x224, &input, dims);

test/net/test_mobilenet_GPU.cpp

@@ -26,7 +26,8 @@ int main() {
   auto isok = paddle_mobile.Load(g_mobilenet, false);
   if (isok) {
     auto time2 = paddle_mobile::time();
-    std::cout << "load cost :" << paddle_mobile::time_diff(time1, time1) << "ms" << std::endl;
+    std::cout << "load cost :" << paddle_mobile::time_diff(time1, time1) << "ms"
+              << std::endl;
 
     std::vector<float> input;
     std::vector<int64_t> dims{1, 3, 224, 224};
@@ -48,8 +49,8 @@ int main() {
     }
     DLOG << vec_result;
     auto time4 = paddle_mobile::time();
-    std::cout << "predict cost :" << paddle_mobile::time_diff(time3, time4) / 10 << "ms"
-              << std::endl;
+    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 "