Remove unnecessary type dispatches from Variable::Impl ctor (pytorch#…

…13630)

Summary:
This should improve the performance of wrapping a tensor in a Variable
Pull Request resolved: pytorch#13630

Reviewed By: ezyang

Differential Revision: D12944960

Pulled By: zou3519

fbshipit-source-id: 89fa78a563e46a747d851a90ffd1b5cf3cd2d0d7

torch/csrc/autograd/variable.cpp

@@ -22,7 +22,7 @@
 namespace torch {
 namespace autograd {
 Variable::Impl::Impl(at::Tensor data, bool requires_grad, Edge gradient_edge)
-    : TensorImpl(data.type().type_id(), data.type().typeMeta(), data.type().allocator(), /* is variable */ true),
+    : TensorImpl(data.type_id(), data.dtype(), /*allocator=*/nullptr, /* is variable */ true),
       data_(std::move(data)),
       grad_fn_(std::move(gradient_edge.function)),
       requires_grad_(false),

torch/csrc/autograd/variable.h

@@ -319,7 +319,7 @@ struct TORCH_API Variable::Impl : public at::TensorImpl {
   /// variables.
   void set_requires_grad(bool requires_grad) override {
     AT_CHECK(
-        !requires_grad || at::isFloatingType(type().scalarType()),
+        !requires_grad || at::isFloatingType(at::typeMetaToScalarType(dtype())),
         "Only Tensors of floating point dtype can require gradients");
     requires_grad_ = requires_grad;
   }