Fix: [ATen] Add some missing moves (pytorch#88514)

Related to pytorch#88512 , but for ATen. This should reduce a number of copies and inefficient atomic smart pointer increments.

Pull Request resolved: pytorch#88514
Approved by: https://github.com/jgong5, https://github.com/ezyang

Author: Skylion007

aten/src/ATen/InferSize.h

@@ -80,7 +80,7 @@ inline at::SymDimVector infer_size_dv(
     c10::SymInt numel) {
   auto res = at::SymDimVector(shape);
   infer_size_impl<c10::SymIntArrayRef, c10::SymInt, at::SymDimVector>(
-      shape, numel, res);
+      shape, std::move(numel), res);
   return res;
 }
 

aten/src/ATen/core/Formatting.cpp

@@ -13,7 +13,7 @@ std::ostream& operator<<(std::ostream & out, Backend b) {
   return out << toString(b);
 }
 
-std::ostream& operator<<(std::ostream & out, Scalar s) {
+std::ostream& operator<<(std::ostream & out, const Scalar& s) {
   if (s.isFloatingPoint()) {
     return out << s.toDouble();
   }
@@ -35,7 +35,7 @@ std::ostream& operator<<(std::ostream & out, Scalar s) {
   throw std::logic_error("Unknown type in Scalar");
 }
 
-std::string toString(Scalar s) {
+std::string toString(const Scalar& s) {
   std::stringstream out;
   out << s;
   return out.str();

aten/src/ATen/core/Formatting.h

@@ -8,8 +8,8 @@
 
 namespace c10 {
 TORCH_API std::ostream& operator<<(std::ostream& out, Backend b);
-TORCH_API std::ostream& operator<<(std::ostream & out, Scalar s);
-TORCH_API std::string toString(Scalar s);
+TORCH_API std::ostream& operator<<(std::ostream & out, const Scalar& s);
+TORCH_API std::string toString(const Scalar& s);
 }
 namespace at {
 

aten/src/ATen/native/TensorShape.cpp

@@ -204,10 +204,11 @@
 #include <ATen/ops/zeros_native.h>
 #endif
 
+#include <c10/util/StringUtil.h>
 #include <algorithm>
 #include <cstdint>
+#include <utility>
 #include <vector>
-#include <c10/util/StringUtil.h>
 
 namespace at {
 namespace meta {
@@ -416,7 +417,7 @@ Tensor& set_storage_meta__symint(Tensor& result, Storage storage, c10::SymInt st
     const auto itemsize = result.dtype().itemsize();
     c10::SymInt size_bytes = at::detail::computeStorageNbytes(
         size, stride, itemsize, storage_offset);
-    storage.set_nbytes(size_bytes);
+    storage.set_nbytes(std::move(size_bytes));
   }
   return result;
 }

aten/src/ATen/native/quantized/cpu/qconv_prepack.cpp

@@ -1,4 +1,5 @@
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
+#include <utility>
 #include <vector>
 
 #include <ATen/core/Tensor.h>
@@ -444,7 +445,7 @@ c10::intrusive_ptr<ConvPackedParamsBase<kSpatialDim>> PackedConvWeightsOnednn<
   exp_wgt.init(w_desc);
   exp_wgt.set_scale(wgt_scales); // Also for feed_from()
   exp_wgt.feed_from(wgt, transpose); // expect wgt to be in [OC IC KH KW] format
-  ideep::tensor * packed_weight_p = new ideep::tensor(exp_wgt);
+  ideep::tensor * packed_weight_p = new ideep::tensor(std::move(exp_wgt));
   packed_weight_p->set_scale(wgt_scales);
   packed_weight_p->set_zero_point(wgt_zero_points);
   std::unique_ptr<ideep::tensor> weight_ptr(packed_weight_p);

aten/src/ATen/native/quantized/cpu/qlinear_prepack.cpp

@@ -23,6 +23,7 @@
 #include <c10/util/irange.h>
 
 #include <algorithm>
+#include <utility>
 #include <vector>
 
 int register_linear_params();
@@ -249,7 +250,7 @@ c10::intrusive_ptr<LinearPackedParamsBase> PackedLinearWeightsOnednn::prepack(
                                                              dnnl::memory::data_type::u8);
   ideep::tensor exp_wgt(w_desc);
   exp_wgt.feed_from(wgt);
-  ideep::tensor * packed_weight_p = new ideep::tensor(exp_wgt);
+  ideep::tensor * packed_weight_p = new ideep::tensor(std::move(exp_wgt));
   packed_weight_p->set_scale(wgt_scales);
   packed_weight_p->set_zero_point(wgt_zero_points);
   std::unique_ptr<ideep::tensor> weight_ptr(packed_weight_p);

c10/core/Storage.h

@@ -76,7 +76,7 @@ struct C10_API Storage {
   }
 
   void set_nbytes(c10::SymInt size_bytes) const {
-    storage_impl_.get()->set_nbytes(size_bytes);
+    storage_impl_.get()->set_nbytes(std::move(size_bytes));
   }
 
   bool resizable() const {

c10/core/StorageImpl.h

@@ -112,7 +112,7 @@ struct C10_API StorageImpl : public c10::intrusive_ptr_target {
   }
 
   void set_nbytes(c10::SymInt size_bytes) {
-    size_bytes_ = size_bytes;
+    size_bytes_ = std::move(size_bytes);
   }
 
   bool resizable() const {

c10/core/WrapDimMinimal.cpp

@@ -14,7 +14,8 @@ T maybe_wrap_dim_slow(T dim, T dim_post_expr, bool wrap_scalar) {
         "Dimension specified as ",
         dim,
         " but tensor has no dimensions");
-    return c10::maybe_wrap_dim(dim, /*dim_post_expr=*/1, /*wrap_scalar=*/false);
+    return c10::maybe_wrap_dim(
+        std::move(dim), /*dim_post_expr=*/1, /*wrap_scalar=*/false);
   }
 
   T min = dim_post_expr * -1;

c10/core/WrapDimMinimal.h

@@ -38,7 +38,7 @@ inline c10::SymInt maybe_wrap_dim(
     c10::SymInt dim,
     c10::SymInt dim_post_expr,
     bool wrap_scalar = true) {
-  return _maybe_wrap_dim(dim, dim_post_expr, wrap_scalar);
+  return _maybe_wrap_dim(std::move(dim), std::move(dim_post_expr), wrap_scalar);
 }
 
 } // namespace c10