New executor dev (#34407)

* first test version

* add test exec;

* add data transfer; test=develop

* add new exec head;

* add memcpy; test=develop

* add python fetch

* add new test

* add graph node; test=develop

* remove useless new executor test; test=develop

* remove gperf dependency; test=develop

* fix compile bugs; test=develop

* remove useless code; test=develop

* remove useless code; test=develop

* add uni test; test=develop

* polish code; test=develop

* polish code; test=develop

* add interpreter cmakefile; test=develop

* remove useless code; test=develop

paddle/fluid/framework/CMakeLists.txt

@@ -427,6 +427,9 @@ else()
   cc_test(custom_tensor_test SRCS custom_tensor_test.cc DEPS custom_tensor glog)
 endif()
 
+#cc_binary(test_executor SRCS test_executor.cc DEPS executor op_registry ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} )
+#cc_binary(new_executor SRCS new_exec_test.cc DEPS operator op_registry executor ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} profiler)
+
 set(FLUID_FRAMEWORK_MODULES proto_desc memory lod_tensor executor data_feed_proto layer dynamic_loader custom_operator)
 
 cc_library(paddle_framework DEPS ${FLUID_FRAMEWORK_MODULES})

paddle/fluid/framework/new_exec_test.cc

+// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <iostream>
+#include <string>
+
+#include <chrono>
+#include <map>
+#include <memory>
+#include <unordered_map>
+#include <vector>
+
+#include "paddle/fluid/framework/executor_gc_helper.h"
+#include "paddle/fluid/framework/garbage_collector.h"
+#include "paddle/fluid/framework/op_info.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/framework/variable.h"
+#include "paddle/fluid/platform/device_context.h"
+
+#include "paddle/fluid/pybind/pybind.h"
+
+#include "gperftools/profiler.h"
+#include "paddle/fluid/framework/new_exec.h"
+#include "paddle/fluid/platform/init.h"
+
+int main() {
+  paddle::framework::InitDevices();
+  paddle::framework::VariableScope global_scope;
+  auto place = paddle::platform::CUDAPlace(0);
+  auto test_prog = paddle::framework::load_from_file("lm_startup_program");
+  {
+    paddle::framework::build_variable_scope(test_prog, &global_scope);
+
+    std::vector<paddle::framework::OpFuncNode> vec_func_list;
+    std::vector<paddle::framework::OperatorBase*> op_list;
+    paddle::framework::build_op_func_list(test_prog, op_list, vec_func_list,
+                                          &global_scope, place);
+
+    // paddle::framework::exec_op_func_list( vec_func_list, op_list,
+    // global_scope, place );
+  }
+
+  cerr << "run main" << endl;
+  auto main_prog = paddle::framework::load_from_file("lm_main_program");
+
+  paddle::framework::build_variable_scope(main_prog, &global_scope);
+
+  std::vector<paddle::framework::OpFuncNode> vec_main_func_list;
+  std::vector<paddle::framework::OperatorBase*> op_main_list;
+  paddle::framework::build_op_func_list(
+      main_prog, op_main_list, vec_main_func_list, &global_scope, place);
+  paddle::framework::Scope scope;
+  paddle::framework::InterpreterCore interp_core(place, main_prog, test_prog,
+                                                 &scope);
+  auto start = std::chrono::steady_clock::now();
+  ProfilerStart("new_executor.prof");
+  for (size_t i = 0; i < 2320; ++i) {
+    if (i % 200 == 0) {
+      cerr << i << endl;
+    }
+    // paddle::framework::exec_op_func_list( vec_main_func_list, op_main_list,
+    // global_scope, place );
+    std::vector<paddle::framework::Tensor> vec_out;
+    interp_core.run({}, {}, {}, vec_out);
+  }
+  ProfilerStop();
+  auto end = std::chrono::steady_clock::now();
+  std::chrono::duration<double> diff = end - start;
+
+  cerr << "time cost " << diff.count() << endl;
+
+  return 1;
+}

paddle/fluid/framework/new_exec_util.h

+// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+/*************************************************************************
+  > File Name: new_exec_util.h
+  > Author: guanshanshan@baidu.com
+  > Created Time: Fri 23 Jul 2021 06:19:19 AM UTC
+ ************************************************************************/
+
+#pragma once
+
+#include <chrono>
+#include <iostream>
+#include <string>
+
+#include <map>
+#include <memory>
+#include <unordered_map>
+#include <vector>
+
+#include "paddle/fluid/framework/executor_gc_helper.h"
+#include "paddle/fluid/framework/garbage_collector.h"
+#include "paddle/fluid/framework/op_info.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/framework/variable.h"
+#include "paddle/fluid/framework/variable_helper.h"
+#include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/platform/init.h"
+
+namespace paddle {
+namespace framework {
+
+class RuntimeInferShapeContext : public InferShapeContext {
+ public:
+  RuntimeInferShapeContext(const OperatorBase& op, const RuntimeContext& ctx)
+      : op_(op), ctx_(ctx) {}
+
+  bool HasInput(const std::string& name) const override {
+    // has only one input
+    const auto& ins = ctx_.inputs;
+    auto it = ins.find(name);
+    if (it == ins.end()) {
+      return false;
+    }
+    const auto& in = it->second;
+    if (in.size() == 0) return false;
+    PADDLE_ENFORCE_EQ(
+        in.size(), 1UL,
+        platform::errors::InvalidArgument(
+            "Input %s should not contain more than one inputs.", name));
+    return in[0] != nullptr;
+  }
+
+  bool HasOutput(const std::string& name) const override {
+    // has only one output
+    const auto& outs = ctx_.outputs;
+    auto it = outs.find(name);
+    if (it == outs.end()) {
+      return false;
+    }
+    const auto& out = it->second;
+    if (out.size() == 0) {
+      return false;
+    }
+    PADDLE_ENFORCE_EQ(
+        out.size(), 1UL,
+        platform::errors::InvalidArgument(
+            "Output %s should not contain more than one outputs.", name));
+    return out[0] != nullptr;
+  }
+
+  bool HasInputs(const std::string& name) const override {
+    const auto& ins = ctx_.inputs;
+    auto it = ins.find(name);
+    if (it == ins.end() || it->second.empty()) {
+      return false;
+    }
+    for (auto& input : it->second) {
+      if (input == nullptr) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  bool HasOutputs(const std::string& name) const override {
+    const auto& outs = ctx_.outputs;
+    auto it = outs.find(name);
+    if (it == outs.end() || it->second.empty()) {
+      return false;
+    }
+    for (auto& output : it->second) {
+      if (output == nullptr) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  AttrReader Attrs() const override { return AttrReader(op_.Attrs()); }
+
+  std::vector<std::string> Inputs(const std::string& name) const override {
+    return op_.Inputs(name);
+  }
+
+  std::vector<std::string> Outputs(const std::string& name) const override {
+    return op_.Outputs(name);
+  }
+
+  std::string GetInputNameByIdx(size_t idx) const override {
+    auto& op_proto =
+        paddle::framework::OpInfoMap::Instance().Get(op_.Type()).proto_;
+    PADDLE_ENFORCE_LT(idx, op_proto->inputs().size(),
+                      platform::errors::OutOfRange(
+                          "The index should be less than the size of inputs of "
+                          "operator %s, but got index is %d and size is %d",
+                          op_.Type(), idx, op_proto->inputs().size()));
+    return op_proto->inputs()[idx].name();
+  }
+
+  std::string GetOutputNameByIdx(size_t idx) const override {
+    auto& op_proto =
+        paddle::framework::OpInfoMap::Instance().Get(op_.Type()).proto_;
+    PADDLE_ENFORCE_LT(
+        idx, op_proto->outputs().size(),
+        platform::errors::OutOfRange(
+            "The index should be less than the size of outputs of "
+            "operator %s, but got index is %d and size is %d",
+            op_.Type(), idx, op_proto->outputs().size()));
+    return op_proto->outputs()[idx].name();
+  }
+
+  void ShareDim(const std::string& in, const std::string& out, size_t i = 0,
+                size_t j = 0) override {
+    auto in_it = ctx_.inputs.find(in);
+    auto out_it = ctx_.outputs.find(out);
+    PADDLE_ENFORCE_NE(
+        in_it, ctx_.inputs.end(),
+        platform::errors::NotFound("Input %s does not exist.", in));
+    PADDLE_ENFORCE_NE(
+        out_it, ctx_.outputs.end(),
+        platform::errors::NotFound("Output %s does not exist.", out));
+    PADDLE_ENFORCE_LT(i, in_it->second.size(),
+                      platform::errors::InvalidArgument(
+                          "The index of input dimension is out of range, "
+                          "excepted index less than %zu, but received %zu.",
+                          in_it->second.size(), i));
+    PADDLE_ENFORCE_LT(j, out_it->second.size(),
+                      platform::errors::InvalidArgument(
+                          "The index of output dimension is out of range, "
+                          "excepted index less than %zu, but received %zu.",
+                          out_it->second.size(), j));
+
+    Variable* in_var = in_it->second[i];
+    Variable* out_var = out_it->second[j];
+
+    PADDLE_ENFORCE_EQ(
+        in_var->Type(), out_var->Type(),
+        platform::errors::InvalidArgument(
+            "The type of input (%s) and output (%s) are inconsistent.", in,
+            out));
+
+    if (in_var->IsType<framework::SelectedRows>()) {
+      auto& in_sele_rows = in_var->Get<framework::SelectedRows>();
+      auto out_sele_rows = out_var->GetMutable<framework::SelectedRows>();
+      out_sele_rows->mutable_value()->Resize(in_sele_rows.value().dims());
+      out_sele_rows->set_rows(in_sele_rows.rows());
+      out_sele_rows->set_height(in_sele_rows.height());
+    } else if (in_var->IsType<framework::LoDTensor>()) {
+      auto& in_lod_tensor = in_var->Get<framework::LoDTensor>();
+      auto* out_lod_tensor = out_var->GetMutable<framework::LoDTensor>();
+      out_lod_tensor->Resize(in_lod_tensor.dims());
+    } else {
+      PADDLE_THROW(platform::errors::Unimplemented(
+          "Currently, the input type of ShareDim only can be LoDTensor "
+          "or SelectedRows."));
+    }
+  }
+
+  void ShareAllLoD(const std::string& in,
+                   const std::string& out) const override {
+    auto in_it = ctx_.inputs.find(in);
+    auto out_it = ctx_.outputs.find(out);
+    PADDLE_ENFORCE_NE(in_it, ctx_.inputs.end(),
+                      platform::errors::NotFound(
+                          "Input [%s] found error in Op [%s]", in, op_.Type()));
+    PADDLE_ENFORCE_NE(
+        out_it, ctx_.outputs.end(),
+        platform::errors::NotFound("Output [%s] found error in Op [%s]", out,
+                                   op_.Type()));
+
+    auto& in_var_list = in_it->second;
+    auto& out_var_list = out_it->second;
+
+    PADDLE_ENFORCE_EQ(
+        in_var_list.size(), out_var_list.size(),
+        platform::errors::PreconditionNotMet(
+            "Op [%s]: Input var size should be equal with output var size",
+            op_.Type()));
+
+    auto& out_var_names = op_.Outputs(out);
+
+    for (size_t i = 0; i < in_var_list.size(); ++i) {
+      if (out_var_names[i] == framework::kEmptyVarName) {
+        continue;
+      }
+
+      Variable* in_var = in_var_list[i];
+      if (!in_var->IsType<LoDTensor>()) return;
+      Variable* out_var = out_var_list[i];
+      PADDLE_ENFORCE_EQ(out_var->IsType<LoDTensor>(), true,
+                        platform::errors::PreconditionNotMet(
+                            "The %d-th output of Output(%s) must be LoDTensor.",
+                            i, out_var_names[i]));
+      auto& in_tensor = in_var->Get<LoDTensor>();
+      auto* out_tensor = out_var->GetMutable<LoDTensor>();
+      out_tensor->set_lod(in_tensor.lod());
+#ifdef PADDLE_WITH_MKLDNN
+      if (in_tensor.layout() != DataLayout::kMKLDNN)
+#endif
+        out_tensor->set_layout(in_tensor.layout());
+    }
+  }
+
+  void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
+                size_t j = 0) const override {
+    auto in_it = ctx_.inputs.find(in);
+    auto out_it = ctx_.outputs.find(out);
+    PADDLE_ENFORCE_NE(
+        in_it, ctx_.inputs.end(),
+        platform::errors::NotFound("Input %s does not exist.", in));
+    PADDLE_ENFORCE_NE(
+        out_it, ctx_.outputs.end(),
+        platform::errors::NotFound("Output %s does not exist.", out));
+    PADDLE_ENFORCE_LT(i, in_it->second.size(),
+                      platform::errors::InvalidArgument(
+                          "The index of input dimension is out of range, "
+                          "excepted index less than %zu, but received %zu.",
+                          in_it->second.size(), i));
+    PADDLE_ENFORCE_LT(j, out_it->second.size(),
+                      platform::errors::InvalidArgument(
+                          "The index of output dimension is out of range, "
+                          "excepted index less than %zu, but received %zu.",
+                          out_it->second.size(), j));
+
+    Variable* in_var = in_it->second.at(i);
+    if (!in_var->IsType<LoDTensor>()) return;
+    Variable* out_var = out_it->second.at(j);
+    PADDLE_ENFORCE_EQ(
+        out_var->IsType<LoDTensor>(), true,
+        platform::errors::InvalidArgument(
+            "The %zu-th output of Output(%s) must be LoDTensor.", j, out));
+    auto& in_tensor = in_var->Get<LoDTensor>();
+    auto* out_tensor = out_var->GetMutable<LoDTensor>();
+    out_tensor->set_lod(in_tensor.lod());
+
+// TODO(dzhwinter) : reuse ShareLoD in most operators.
+// Need to call ShareLayout explicitly in sequence related ops.
+// Shall we have a better method to shared info between in/out Tensor?
+#ifdef PADDLE_WITH_MKLDNN
+    // Fix me: ugly workaround below
+    // Correct solution:
+    //    set_layout() should NOT be called here (i.e. ShareLoD). Instead,
+    //    layout of output tensor should be set "manually" in Compute()
+    //    of each OPKernel. The reason layout should NOT be shared between
+    //    input and output "automatically" (now by InferShape()->ShareLoD())
+    //    is that layout transform may occur after InferShape().
+    // Workaround:
+    //    Skip set_layout() when input layout is kMKLDNN
+    //    This is to avoid kMKLDNN is populated wrongly into a non-MKLDNN
+    //    OPKernel. In all MKLDNN OPkernel, set_layout(kMKLDNN) should be called
+    //    in Compute()
+    if (in_tensor.layout() != DataLayout::kMKLDNN)
+#endif
+      out_tensor->set_layout(in_tensor.layout());
+  }
+
+  int32_t GetLoDLevel(const std::string& in, size_t i = 0) const override {
+    PADDLE_THROW(platform::errors::PreconditionNotMet(
+        "GetLoDLevel is only used in compile time. The calculation of "
+        "output's actual lod is different among operators so that should be "
+        "set in the runtime kernel."));
+  }
+
+  void SetLoDLevel(const std::string& out, int32_t lod_level,
+                   size_t j = 0) const override {
+    PADDLE_THROW(platform::errors::PreconditionNotMet(
+        "SetLoDLevel is only used in compile time. The calculation of "
+        "output's actual lod is different among operators so that should be "
+        "set in the runtime kernel."));
+  }
+
+  bool IsRuntime() const override { return true; }
+
+  // TODO(paddle-dev): Can this be template?
+  std::vector<InferShapeVarPtr> GetInputVarPtrs(
+      const std::string& name) override {
+    const std::vector<Variable*>& vars = InputVars(name);
+    std::vector<InferShapeVarPtr> res;
+    res.reserve(vars.size());
+    res.insert(res.begin(), vars.begin(), vars.end());
+    return res;
+  }
+
+  std::vector<InferShapeVarPtr> GetOutputVarPtrs(
+      const std::string& name) override {
+    const std::vector<Variable*>& vars = OutputVars(name);
+    std::vector<InferShapeVarPtr> res;
+    res.reserve(vars.size());
+    res.insert(res.begin(), vars.begin(), vars.end());
+    return res;
+  }
+
+  DDim GetInputDim(const std::string& name) const override {
+    const std::vector<Variable*>& vars = InputVars(name);
+    PADDLE_ENFORCE_EQ(
+        vars.size(), 1UL,
+        platform::errors::InvalidArgument(
+            "Input(%s) should hold one element, but now it holds %zu elements.",
+            name, vars.size()));
+    return this->GetDim(vars[0]);
+  }
+
+  std::vector<DDim> GetInputsDim(const std::string& name) const override {
+    const std::vector<Variable*>& vars = InputVars(name);
+    return GetDims(vars);
+  }
+
+  std::vector<proto::VarType::Type> GetInputsVarType(
+      const std::string& name) const override {
+    return GetVarTypes(InputVars(name));
+  }
+
+  std::vector<proto::VarType::Type> GetOutputsVarType(
+      const std::string& name) const override {
+    return GetVarTypes(OutputVars(name));
+  }
+
+  void SetOutputDim(const std::string& name, const DDim& dim) override {
+    auto& vars = OutputVars(name);
+    PADDLE_ENFORCE_EQ(
+        vars.size(), 1UL,
+        platform::errors::InvalidArgument("Output(%s) should hold one element, "
+                                          "but now it holds %zu elements.",
+                                          name, vars.size()));
+    SetDim(vars[0], dim);
+  }
+
+  void SetOutputsDim(const std::string& name,
+                     const std::vector<DDim>& dims) override {
+    auto& vars = OutputVars(name);
+    SetDims(vars, dims);
+  }
+
+ protected:
+  DDim GetDim(Variable* var) const {
+    PADDLE_ENFORCE_NOT_NULL(
+        var, platform::errors::InvalidArgument("Input variable is nullptr."));
+    if (var->IsType<LoDTensor>()) {
+      return var->Get<LoDTensor>().dims();
+    } else if (var->IsType<SelectedRows>()) {
+      return var->Get<SelectedRows>().GetCompleteDims();
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Only LoDTensor or SelectedRows support 'GetDim', but input "
+          "Variable's type is %s.",
+          ToTypeName(var->Type())));
+    }
+  }
+
+  std::vector<DDim> GetDims(const std::vector<Variable*>& vars) const {
+    std::vector<DDim> ret;
+    ret.reserve(vars.size());
+    std::transform(vars.begin(), vars.end(), std::back_inserter(ret),
+                   [this](Variable* var) { return this->GetDim(var); });
+    return ret;
+  }
+
+  std::vector<DDim> GetRepeatedDims(const std::string& name) const override {
+    PADDLE_THROW(platform::errors::PreconditionNotMet(
+        "GetRepeatedDims method only ban be used in compile time."));
+  }
+
+  void SetDim(Variable* var, const DDim& dim) {
+    if (var->IsType<LoDTensor>()) {
+      var->GetMutable<LoDTensor>()->Resize(dim);
+    } else if (var->IsType<SelectedRows>()) {
+      var->GetMutable<SelectedRows>()->set_height(dim[0]);
+    } else {
+      PADDLE_THROW(platform::errors::Unimplemented(
+          "Variable type error, expect LoDTensor or SelectedRows, but received "
+          "(%s).",
+          ToTypeName(var->Type())));
+    }
+  }
+
+  void SetDims(const std::vector<Variable*>& vars,
+               const std::vector<DDim>& dims) {
+    size_t length = vars.size();
+    PADDLE_ENFORCE_EQ(length, dims.size(),
+                      platform::errors::InvalidArgument(
+                          "The number of input variables do not match the "
+                          "number of input dimensions, the number of variables "
+                          "is %zu, the number of dimensions is %zu.",
+                          length, dims.size()));
+    for (size_t i = 0; i < length; ++i) {
+      if (vars[i] == nullptr) {
+        continue;
+      }
+      SetDim(vars[i], dims[i]);
+    }
+  }
+
+  void SetRepeatedDims(const std::string& name,
+                       const std::vector<DDim>& dims) override {
+    PADDLE_THROW(platform::errors::PreconditionNotMet(
+        "SetRepeatedDims method only can be used in compile time."));
+  }
+
+  std::vector<proto::VarType::Type> GetVarTypes(
+      const std::vector<Variable*>& vars) const {
+    std::vector<proto::VarType::Type> retv;
+    retv.resize(vars.size());
+    std::transform(vars.begin(), vars.end(), retv.begin(),
+                   std::bind(std::mem_fn(&RuntimeInferShapeContext::GetVarType),
+                             this, std::placeholders::_1));
+    return retv;
+  }
+
+  proto::VarType::Type GetVarType(Variable* var) const {
+    return ToVarType(var->Type());
+  }
+
+ private:
+  const std::vector<Variable*>& InputVars(const std::string& name) const {
+    auto it = ctx_.inputs.find(name);
+    PADDLE_ENFORCE_NE(
+        it, ctx_.inputs.end(),
+        platform::errors::NotFound(
+            "Operator (%s) does not have the input (%s).", op_.Type(), name));
+    return it->second;
+  }
+
+  const std::vector<Variable*>& OutputVars(const std::string& name) const {
+    auto it = ctx_.outputs.find(name);
+    PADDLE_ENFORCE_NE(
+        it, ctx_.outputs.end(),
+        platform::errors::NotFound(
+            "Operator (%s) does not have the outputs (%s).", op_.Type(), name));
+    return it->second;
+  }
+
+  const OperatorBase& op_;
+  const RuntimeContext& ctx_;
+};
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/operators/fill_constant_op.cc

@@ -36,7 +36,6 @@ class FillConstantOp : public framework::OperatorWithKernel {
                 i, shape[i], framework::make_ddim(shape)));
       }
     }
-
     if (shape.empty() && ctx->HasInput("ShapeTensor")) {
       auto shape_dims = ctx->GetInputDim("ShapeTensor");
       int num_ele = 1;

paddle/fluid/operators/reduce_ops/reduce_op.h

@@ -591,7 +591,6 @@ class ReduceGradOp : public framework::OperatorWithKernel {
         (in_dtype >= 0) ? static_cast<framework::proto::VarType::Type>(in_dtype)
                         : OperatorWithKernel::IndicateVarDataType(
                               ctx, framework::GradVarName("Out"));
-
 #ifdef PADDLE_WITH_MKLDNN
     auto CanMKLDNNReduceGradBeUsed = [&]() {
       auto dx_dims = ctx.Input<Tensor>("X")->dims();

paddle/fluid/operators/softmax_with_cross_entropy_op.h

@@ -111,15 +111,12 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
     const Tensor* labels = context.Input<Tensor>("Label");
     Tensor* logit_grad =
         context.Output<Tensor>(framework::GradVarName("Logits"));
-
     const Tensor* softmax = context.Input<Tensor>("Softmax");
     const bool use_softmax = context.Attr<bool>("use_softmax");
-
     if (logit_grad != softmax || !use_softmax) {
       framework::TensorCopy(*softmax, context.GetPlace(),
                             context.device_context(), logit_grad);
     }
-
     const bool soft_label = context.Attr<bool>("soft_label");
     auto ignore_index = context.Attr<int>("ignore_index");
 
@@ -133,7 +130,6 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
     logit_grad_2d.ShareDataWith(*logit_grad).Resize({n, d});
     labels_2d.ShareDataWith(*labels).Resize({n, labels->numel() / n});
     out_grad_2d.ShareDataWith(*out_grad).Resize({n, d / axis_dim});
-
     auto out_grad_mat = framework::EigenMatrix<T>::From(out_grad_2d);
     auto logit_grad_mat = framework::EigenMatrix<T>::From(logit_grad_2d);
     auto& place = *context.template device_context<platform::CPUDeviceContext>()
@@ -147,9 +143,8 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
         logit_grad_mat.device(place) =
             out_grad_mat.broadcast(Eigen::DSizes<int, 2>(1, axis_dim)) *
             logit_grad_mat;
-      }
-      // use_softmax step2
-      else {
+      } else {
+        // use_softmax step2
         const int64_t* label_data = labels->data<int64_t>();
         T* logit_grad_data = logit_grad->data<T>();
         const T* out_grad_data = out_grad->data<T>();
@@ -180,7 +175,6 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
       }
       return;
     }
-
     // for use_softmax=False, continue
 
     if (soft_label) {

paddle/fluid/pybind/pybind.cc

@@ -42,6 +42,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_rank_table.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
+#include "paddle/fluid/framework/new_exec.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/op_version_registry.h"
@@ -1935,6 +1936,34 @@ All parameter, weight, gradient are variables in Paddle.
                  fetch_vars);
       });
 
+  py::class_<framework::InterpreterCore>(m, "InterpreterCore")
+      .def(py::init<const platform::Place &, const ProgramDesc &,
+                    const ProgramDesc &, Scope *>())
+      .def("run",
+           [](InterpreterCore &self,
+              const std::unordered_map<std::string, py::array> &input_dict,
+              std::vector<std::string> vec_fetch_name) {
+             pybind11::gil_scoped_release release;
+             std::vector<framework::Tensor> vec_tensor;
+             std::vector<std::string> vec_name;
+
+             for (auto &item : input_dict) {
+               framework::LoDTensor t;
+               SetTensorFromPyArray<platform::CPUPlace>(
+                   &t, item.second, platform::CPUPlace(), false);
+               vec_name.push_back(item.first);
+               vec_tensor.push_back(t);
+             }
+
+             std::vector<framework::Tensor> vec_out;
+             self.run(vec_name, vec_tensor, vec_fetch_name, &vec_out);
+             std::vector<py::array> vec_ret;
+             for (size_t i = 0; i < vec_out.size(); ++i) {
+               vec_ret.push_back(TensorToPyArray(vec_out[i], true));
+             }
+             return vec_ret;
+           });
+
   m.def("init_gflags", framework::InitGflags);
   m.def("init_glog", framework::InitGLOG);
   m.def("load_op_meta_info_and_register_op",

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -686,6 +686,8 @@ add_subdirectory(asp)
 
 add_subdirectory(ir)
 
+add_subdirectory(interpreter)
+
 if (WITH_TESTING)
     set_property(TEST test_parallel_executor_mnist PROPERTY ENVIRONMENT GLOG_vmodule=all_reduce_deps_pass=10)
     set_property(TEST test_parallel_executor_fix_op_run_order PROPERTY ENVIRONMENT GLOG_vmodule=fix_op_run_order_pass=10)

python/paddle/fluid/tests/unittests/interpreter/CMakeLists.txt

+file(GLOB TEST_INTERP_CASES RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_INTERP_CASES "${TEST_INTERP_CASES}")
+
+foreach(target ${TEST_INTERP_CASES})
+  py_test_modules(${target} MODULES ${target})
+endforeach()

python/paddle/fluid/tests/unittests/interpreter/test_interpreter.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import paddle
+from paddle.fluid import core
+from paddle.fluid.core import InterpreterCore
+
+import numpy as np
+
+paddle.enable_static()
+
+
+class LinearTestCase(unittest.TestCase):
+    def setUp(self):
+        self.place = paddle.CUDAPlace(0) if core.is_compiled_with_cuda(
+        ) else paddle.CPUPlace()
+
+    def test_interp_base(self):
+        a = paddle.static.data(name="a", shape=[2, 2], dtype='float32')
+        b = paddle.ones([2, 2]) * 2
+        t = paddle.static.nn.fc(a, 2)
+        c = t + b
+
+        main_program = paddle.fluid.default_main_program()
+        startup_program = paddle.fluid.default_startup_program()
+        p = core.Place()
+        p.set_place(self.place)
+        inter_core = InterpreterCore(p, main_program.desc, startup_program.desc,
+                                     core.Scope())
+
+        out = inter_core.run({
+            "a": np.ones(
+                [2, 2], dtype="float32") * 2
+        }, [c.name])
+        for i in range(10):
+            out = inter_core.run({
+                "a": np.ones(
+                    [2, 2], dtype="float32") * i
+            }, [c.name])
+
+
+if __name__ == "__main__":
+    unittest.main()