Do not preallocate reshape, reuse input tensor instead.

mace/core/tensor.h

@@ -223,16 +223,19 @@ class Tensor {
     }
   }
 
-  inline void ResizeWithBuffer(const std::vector<index_t> &shape,
-                               BufferBase *buffer) {
-    MACE_CHECK(!has_opencl_image(), "Cannot resize image, use ResizeImage.");
-    shape_ = shape;
-    image_shape_.clear();
-    if (buffer_ != nullptr && is_buffer_owner_) {
+  // Make this tensor reuse other tensor's buffer.
+  // This tensor has the same dtype, shape and image_shape.
+  // It could be reshaped later (with image shape unchanged).
+  inline void ReuseTensorBuffer(const Tensor &other) {
+    if (is_buffer_owner_ && buffer_ != nullptr) {
       delete buffer_;
     }
-    buffer_ = buffer;
     is_buffer_owner_ = false;
+    buffer_ = other.buffer_;
+    allocator_ = other.allocator_;
+    dtype_ = other.dtype_;
+    shape_ = other.shape_;
+    image_shape_ = other.image_shape_;
   }
 
   inline MaceStatus ResizeImage(const std::vector<index_t> &shape,

mace/core/workspace.cc

@@ -176,7 +176,8 @@ MaceStatus Workspace::CreateOutputTensorBuffer(const NetDef &net_def,
     const int op_device =
         ProtoArgHelper::GetOptionalArg<OperatorDef, int>(
             op, "device", static_cast<int>(device_type));
-    if (op_device == device_type && !op.mem_id().empty()) {
+    if (op_device == device_type && !op.mem_id().empty()
+        && op.type() != "Reshape") {
       auto mem_ids = op.mem_id();
       int count = mem_ids.size();
       for (int i = 0; i < count; ++i) {

mace/kernels/reshape.h

@@ -36,7 +36,8 @@ struct ReshapeFunctor {
                   Tensor *output,
                   StatsFuture *future) {
     MACE_UNUSED(future);
-    output->ResizeWithBuffer(out_shape, input->UnderlyingBuffer());
+    output->ReuseTensorBuffer(*input);
+    output->Reshape(out_shape);
 
     return MACE_SUCCESS;
   }