Merge pull request #16287 from PaddlePaddle/revert-16002-runtime_context

Revert "cache runtime_context"

Author: luotao1

paddle/fluid/framework/ir/CMakeLists.txt

@@ -70,7 +70,6 @@ pass_library(conv_affine_channel_fuse_pass inference)
 pass_library(transpose_flatten_concat_fuse_pass inference)
 pass_library(identity_scale_op_clean_pass base)
 pass_library(sync_batch_norm_pass base)
-pass_library(runtime_context_cache_pass base)
 
 # There may be many transpose-flatten structures in a model, and the output of
 # these structures will be used as inputs to the concat Op. This pattern will

paddle/fluid/framework/ir/runtime_context_cache_pass.cc

@@ -874,23 +874,9 @@ std::vector<KernelConfig>* OperatorWithKernel::GetKernelConfig(
   return kernel_configs;
 }
 
-RuntimeContext* OperatorWithKernel::GetRuntimeContext(
-    const Scope& scope) const {
-  if (!HasAttr(kEnableCacheRuntimeContext)) {
-    return new RuntimeContext(Inputs(), Outputs(), scope);
-  } else {
-    const Scope* cur_scope = &scope;
-    if (!runtime_ctx_ || pre_scope_ != cur_scope) {
-      runtime_ctx_.reset(new RuntimeContext(Inputs(), Outputs(), scope));
-      pre_scope_ = cur_scope;
-    }
-    return runtime_ctx_.get();
-  }
-}
-
 void OperatorWithKernel::RunImpl(const Scope& scope,
                                  const platform::Place& place) const {
-  auto runtime_ctx = GetRuntimeContext(scope);
+  RuntimeContext ctx(Inputs(), Outputs(), scope);
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto* dev_ctx = pool.Get(place);
 
@@ -905,7 +891,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   OpKernelMap& kernels = kernels_iter->second;
 
   auto expected_kernel_key = this->GetExpectedKernelType(
-      ExecutionContext(*this, scope, *dev_ctx, *runtime_ctx, nullptr));
+      ExecutionContext(*this, scope, *dev_ctx, ctx, nullptr));
   VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
 
   auto kernel_iter = kernels.find(expected_kernel_key);
@@ -929,8 +915,8 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
 
   // do data transformScope &transfer_scope;
   std::vector<std::string> transfered_inplace_vars;
-  auto* transfer_scope = PrepareData(scope, expected_kernel_key,
-                                     &transfered_inplace_vars, runtime_ctx);
+  auto* transfer_scope =
+      PrepareData(scope, expected_kernel_key, &transfered_inplace_vars, &ctx);
 
   // exec scope is the scope that kernel actually executed on.
   const Scope& exec_scope =
@@ -941,13 +927,13 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   }
 
   if (!HasAttr(kAllKernelsMustComputeRuntimeShape)) {
-    RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, *runtime_ctx);
+    RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, ctx);
     this->InferShape(&infer_shape_ctx);
   }
   // TODO(panyx0718): ExecutionContext should only depend on RuntimeContext
   // not Scope. Imperative mode only pass inputs and get outputs.
-  kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx,
-                                       *runtime_ctx, kernel_configs));
+  kernel_iter->second(
+      ExecutionContext(*this, exec_scope, *dev_ctx, ctx, kernel_configs));
 
   if (!transfered_inplace_vars.empty()) {
     // there is inplace variable has been transfered.

paddle/fluid/framework/ir/runtime_context_cache_pass.h

@@ -62,14 +62,6 @@ constexpr char kZeroVarSuffix[] = "@ZERO";
 /// Variables with this suffix are the new Gradient.
 constexpr char kNewGradSuffix[] = "@NEWGRAD@";
 
-/// RuntimeContext is used to relate input/output names of Operator with
-/// the corresponding variables in name scope.
-/// If an Op has attribute kEnableCacheRuntimeContext, it means that in a same
-/// name scope, since the input/output names of this Op do not change in the
-/// execution, RuntimeContext could be created only at the first iteration of
-/// this Op's execution to save the elapsed time.
-constexpr char kEnableCacheRuntimeContext[] = "@ENABLE_CACHE_RUNTIME_CONTEXT@";
-
 /// If an Op has this attribute, all its kernels should calculate output
 /// variable's shape in the corresponding Compute() function. And
 /// OperatorWithKernel::RunImpl() would skip call this Op's InferShape()
@@ -464,7 +456,6 @@ class OperatorWithKernel : public OperatorBase {
   // same.
   proto::VarType::Type IndicateDataType(const ExecutionContext& ctx) const;
   void RunImpl(const Scope& scope, const platform::Place& place) const final;
-  RuntimeContext* GetRuntimeContext(const Scope& scope) const;
 
   /**
    * Transfer data from scope to a transfered scope. If there is no data need to
@@ -483,8 +474,6 @@ class OperatorWithKernel : public OperatorBase {
 
  protected:
   mutable OpKernelConfigsMap kernel_configs_map_;
-  mutable std::unique_ptr<RuntimeContext> runtime_ctx_;
-  mutable const Scope* pre_scope_ = nullptr;
 };
 
 extern bool OpSupportGPU(const std::string& op_type);

paddle/fluid/framework/operator.cc

@@ -118,9 +118,6 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
 
   CP_MEMBER(serialized_info_cache_);
 
-  // framework related.
-  CP_MEMBER(enable_runtime_context_cache_);
-
   if (use_gpu_) {
     pass_builder_.reset(new GpuPassStrategy(
         *static_cast<GpuPassStrategy *>(other.pass_builder())));
@@ -235,10 +232,6 @@ void AnalysisConfig::Update() {
   if (ir_debug_) {
     pass_builder()->TurnOnDebug();
   }
-
-  if (enable_runtime_context_cache_) {
-    pass_builder()->AppendPass("runtime_context_cache_pass");
-  }
 }
 
 std::string AnalysisConfig::SerializeInfoCache() {
@@ -272,7 +265,6 @@ std::string AnalysisConfig::SerializeInfoCache() {
 
   ss << specify_input_name_;
   ss << cpu_math_library_num_threads_;
-  ss << enable_runtime_context_cache_;
 
   return ss.str();
 }

paddle/fluid/framework/operator.h

@@ -194,23 +194,6 @@ struct AnalysisConfig {
   /** Tell whether the memory optimization is activated. */
   bool enable_memory_optim() const;
 
-  // framework related
-  /** \brief Control whether to perform runtime context cache optimization.
-   *
-   * If turned off, in Op's every execution, RuntimeContext would be called to
-   * relate input/output names of this Op with the corresponding variables in
-   * Scope.
-   */
-  void SwitchRuntimeContextCache(int x = true) {
-    enable_runtime_context_cache_ = x;
-  }
-  /** A boolean state tell whether the runtime context cache optimization is
-   * actived.
-   */
-  bool runtime_context_cache_enabled() const {
-    return enable_runtime_context_cache_;
-  }
-
   friend class ::paddle::AnalysisPredictor;
 
   /** NOTE just for developer, not an official API, easily to be broken.
@@ -271,15 +254,6 @@ struct AnalysisConfig {
 
   int cpu_math_library_num_threads_{1};
 
-  // framework related
-  // RuntimeContext is used to relate input/output names of Operator with
-  // the corresponding variables in Scope.
-  // If enable_runtime_context_cache_ is true, it means that in a same Scope,
-  // since the input/output names of this Op do not change in the execution,
-  // RuntimeContext could be created only at the first iteration of this Op's
-  // execution to save the elapsed time.
-  bool enable_runtime_context_cache_{false};
-
   // A runtime cache, shouldn't be transferred to others.
   std::string serialized_info_cache_;
 

paddle/fluid/inference/api/analysis_config.cc

@@ -107,7 +107,6 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->DisableGpu();
   cfg->SwitchSpecifyInputNames();
   cfg->SwitchIrOptim();
-  cfg->SwitchRuntimeContextCache();
   if (FLAGS_zero_copy) {
     cfg->SwitchUseFeedFetchOps(false);
   }

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -72,8 +72,7 @@ std::ostream &operator<<(std::ostream &os, const AnalysisConfig &config) {
   }
   os << GenSpaces(num_spaces) << "enable_ir_optim: " << config.ir_optim()
      << "\n";
-  os << GenSpaces(num_spaces)
-     << "use_runtime_context_cache: " << config.runtime_context_cache_enabled()
+  os << GenSpaces(num_spaces) << "enable_ir_optim: " << config.ir_optim()
      << "\n";
   os << GenSpaces(num_spaces)
      << "use_feed_fetch_ops: " << config.use_feed_fetch_ops_enabled() << "\n";

paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc

@@ -242,10 +242,6 @@ void BindAnalysisConfig(py::module *m) {
       .def("set_mkldnn_op", &AnalysisConfig::SetMKLDNNOp)
       .def("set_model_buffer", &AnalysisConfig::SetModelBuffer)
       .def("model_from_memory", &AnalysisConfig::model_from_memory)
-      .def("runtime_context_cache_enabled",
-           &AnalysisConfig::runtime_context_cache_enabled)
-      .def("switch_runtime_context_cache",
-           &AnalysisConfig::SwitchRuntimeContextCache, py::arg("x") = true)
       .def("pass_builder", &AnalysisConfig::pass_builder,
            py::return_value_policy::reference);
 }