[PHI] Remove reseting dtype, layout and allocation by arg_def for outputs in executor (#39781)

* remove SetAllocationForOutputTenosr

* add place param for copy kernel

* recover SetAllocationForOutputTenosr

* polish code

* fix empty_dev api bug

* remove reseting dtype and layout for output in executor

* fix merge bug

* [Phi] Add ClearHolder when re-alloc on new place in DeviceContext

* fix hostAlloc

* remove setting output allocation

* remove full_kernel_impl.h

* fix bug of xpu full_like
Co-authored-by: NAurelius84 <zhangliujie@baidu.com>

paddle/fluid/framework/operator.cc

@@ -2108,12 +2108,6 @@ void OperatorWithKernel::BuildPhiKernelContext(
             "Unsupported output `%s` type when call pt kernel.",
             framework::ToTypeName(var->Type())));
       }
-
-      experimental::ResetTensorDtypeAndLayoutByArgDef(tensor_out,
-                                                      output_defs.at(i));
-      SetAllocationForOutputTenosr(
-          tensor_out, phi::TransToPhiPlace(output_defs.at(i).backend));
-
       pt_kernel_context->EmplaceBackOutputWithoutSetRange(tensor_out);
     }
 

paddle/fluid/framework/phi_utils.cc

@@ -233,26 +233,5 @@ static void SetAllocationForUninitializedDenseTensor(
   dense_tensor->ResetHolder(shared_allocation);
 }
 
-void SetAllocationForOutputTenosr(phi::TensorBase* tensor,
-                                  const platform::Place& place) {
-  if (phi::DenseTensor::classof(tensor)) {
-    auto* dense_tensor = static_cast<phi::DenseTensor*>(tensor);
-    if (!dense_tensor->IsInitialized() || !(dense_tensor->place() == place)) {
-      SetAllocationForUninitializedDenseTensor(dense_tensor, place);
-    }
-  } else if (phi::SelectedRows::classof(tensor)) {
-    auto* selected_rows = static_cast<phi::SelectedRows*>(tensor);
-    if (!selected_rows->value().IsInitialized() ||
-        !(selected_rows->place() == place)) {
-      SetAllocationForUninitializedDenseTensor(selected_rows->mutable_value(),
-                                               place);
-    }
-  } else {
-    PADDLE_THROW(platform::errors::Unimplemented(
-        "Unsupported tensor type is received when setting allocation for "
-        "output tensor."));
-  }
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/phi_utils.h

@@ -62,9 +62,6 @@ class KernelArgsNameMaker {
 
 void InitDefaultKernelSignatureMap();
 
-void SetAllocationForOutputTenosr(phi::TensorBase* tensor,
-                                  const platform::Place& place);
-
 // TODO(Wilber): support others device context.
 template <typename T>
 struct ConvertToPhiContext {

paddle/fluid/imperative/prepared_operator.h

@@ -323,12 +323,6 @@ void BuildDygraphPhiKernelContext(
             "Unsupported output `%s` type when call pt kernel.",
             framework::ToTypeName(var->Type())));
       }
-
-      experimental::ResetTensorDtypeAndLayoutByArgDef(tensor_out,
-                                                      output_defs.at(i));
-      framework::SetAllocationForOutputTenosr(
-          tensor_out, phi::TransToPhiPlace(output_defs.at(i).backend));
-
       kernel_ctx->EmplaceBackOutputWithoutSetRange(tensor_out);
     }
     kernel_ctx->AssignOutputRange(std::make_pair(start_idx, end_idx), i);

paddle/phi/api/lib/utils/tensor_utils.cc

@@ -136,26 +136,5 @@ phi::ScalarArray MakePhiScalarArrayFromVarList(
   return result;
 }
 
-void ResetTensorDtypeAndLayoutByArgDef(phi::TensorBase* dst,
-                                       const phi::TensorArgDef& arg_def) {
-  VLOG(5) << "ResetTensor by TensorArgDef.";
-  if (phi::DenseTensor::classof(dst)) {
-    auto* dense_t = static_cast<phi::DenseTensor*>(dst);
-    auto* meta = phi::DenseTensorUtils::GetMutableMeta(dense_t);
-    meta->dtype = arg_def.dtype;
-    meta->layout = arg_def.layout;
-  } else if (phi::SelectedRows::classof(dst)) {
-    auto* selected_rows = static_cast<phi::SelectedRows*>(dst);
-    auto* meta =
-        phi::DenseTensorUtils::GetMutableMeta(selected_rows->mutable_value());
-    meta->dtype = arg_def.dtype;
-    meta->layout = arg_def.layout;
-  } else {
-    PADDLE_THROW(phi::errors::Unimplemented(
-        "Unsupported tensor type is received when reseting tensor dtype and "
-        "layout by argument definition."));
-  }
-}
-
 }  // namespace experimental
 }  // namespace paddle

paddle/phi/api/lib/utils/tensor_utils.h

@@ -42,8 +42,5 @@ phi::ScalarArray MakePhiScalarArrayFromVar(const framework::Variable& variable);
 phi::ScalarArray MakePhiScalarArrayFromVarList(
     const std::vector<framework::Variable*>& variable_list);
 
-void ResetTensorDtypeAndLayoutByArgDef(phi::TensorBase* dst,
-                                       const phi::TensorArgDef& arg_def);
-
 }  // namespace experimental
 }  // namespace paddle

paddle/phi/core/CMakeLists.txt

@@ -22,8 +22,8 @@ cc_library(sparse_csr_tensor SRCS sparse_csr_tensor.cc DEPS dense_tensor tensor_
 
 cc_library(meta_tensor SRCS meta_tensor.cc DEPS tensor_base tensor_meta dense_tensor)
 cc_library(infermeta_utils SRCS infermeta_utils.cc DEPS meta_tensor)
-cc_library(phi_device_context SRCS device_context.cc DEPS dense_tensor selected_rows)
 cc_library(selected_rows SRCS selected_rows_impl.cc DEPS dense_tensor phi_enforce ddim memcpy)
+cc_library(phi_device_context SRCS device_context.cc DEPS dense_tensor selected_rows)
 
 cc_library(phi_custom_kernel SRCS custom_kernel.cc DEPS kernel_factory convert_utils)
 

paddle/phi/core/dense_tensor_impl.cc

@@ -73,11 +73,6 @@ void DenseTensor::set_layout(const paddle::framework::DataLayout layout) {
 // Note: When you reset holder, you need to ensure the offset is correct
 void DenseTensor::ResetHolder(const std::shared_ptr<phi::Allocation>& holder) {
   if (holder_) {
-    // TODO(zyfncg): The change of static_cast<> in check will recover back
-    // when SetAllocationForOutputTenosr is deleted.
-    // Now the numel() may return -1, and will cast to a very large number when
-    // compare with a data with unsigned long type, this will make checking
-    // failed, so it's a temporary solution to deal with this problem.
     PADDLE_ENFORCE_LE(
         numel() * static_cast<int64_t>(SizeOf(dtype())) +
             static_cast<int64_t>(meta_.offset),

paddle/phi/kernels/xpu/full_kernel.cc

@@ -59,7 +59,7 @@ void FullKernel(const Context& dev_ctx,
                 const Scalar& val,
                 DataType dtype,
                 DenseTensor* out) {
-  out->ResizeAndAllocate(phi::make_ddim(shape.GetData()));
+  out->Resize(phi::make_ddim(shape.GetData()));
   FullValueXPU<T>(dev_ctx, out, val.to<T>());
 }
 
@@ -69,6 +69,7 @@ void FullLikeKernel(const Context& dev_ctx,
                     const Scalar& val,
                     DataType dtype,
                     DenseTensor* out) {
+  dev_ctx.template Alloc<T>(out);
   auto value = val.to<float>();
   using XPUInTDType = typename XPUTypeTrait<T>::Type;
   using CommonType = typename std::common_type<