From 7d2740db83f08683f710ec01caf46e2e478458ed Mon Sep 17 00:00:00 2001 From: Tao Luo Date: Tue, 19 Mar 2019 16:44:20 +0800 Subject: [PATCH] Revert "cache runtime_context" --- paddle/fluid/framework/ir/CMakeLists.txt | 1 - .../ir/runtime_context_cache_pass.cc | 39 ------------------- .../framework/ir/runtime_context_cache_pass.h | 32 --------------- paddle/fluid/framework/operator.cc | 28 ++++--------- paddle/fluid/framework/operator.h | 11 ------ paddle/fluid/inference/api/analysis_config.cc | 8 ---- .../inference/api/paddle_analysis_config.h | 26 ------------- .../tests/api/analyzer_pyramid_dnn_tester.cc | 1 - .../inference/tests/api/config_printer.h | 3 +- paddle/fluid/pybind/inference_api.cc | 4 -- 10 files changed, 8 insertions(+), 145 deletions(-) delete mode 100644 paddle/fluid/framework/ir/runtime_context_cache_pass.cc delete mode 100644 paddle/fluid/framework/ir/runtime_context_cache_pass.h diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index a79a53867..8d2cc5ade 100644 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/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 diff --git a/paddle/fluid/framework/ir/runtime_context_cache_pass.cc b/paddle/fluid/framework/ir/runtime_context_cache_pass.cc deleted file mode 100644 index 67b29512c..000000000 --- a/paddle/fluid/framework/ir/runtime_context_cache_pass.cc +++ /dev/null @@ -1,39 +0,0 @@ -/* Copyright (c) 2019 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 "paddle/fluid/framework/ir/runtime_context_cache_pass.h" -#include -#include "paddle/fluid/framework/operator.h" - -namespace paddle { -namespace framework { -namespace ir { - -std::unique_ptr RuntimeContextCachePass::ApplyImpl( - std::unique_ptr graph) const { - VLOG(3) << "Applies Runtime Context Cache strategy."; - for (const Node* n : graph->Nodes()) { - if (n->IsOp()) { - n->Op()->SetAttr(kEnableCacheRuntimeContext, true); - } - } - return graph; -} - -} // namespace ir -} // namespace framework -} // namespace paddle - -REGISTER_PASS(runtime_context_cache_pass, - paddle::framework::ir::RuntimeContextCachePass); diff --git a/paddle/fluid/framework/ir/runtime_context_cache_pass.h b/paddle/fluid/framework/ir/runtime_context_cache_pass.h deleted file mode 100644 index a6cf1a9ae..000000000 --- a/paddle/fluid/framework/ir/runtime_context_cache_pass.h +++ /dev/null @@ -1,32 +0,0 @@ -/* Copyright (c) 2019 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. */ - -#pragma once - -#include -#include "paddle/fluid/framework/ir/pass.h" - -namespace paddle { -namespace framework { -namespace ir { - -class RuntimeContextCachePass : public Pass { - protected: - std::unique_ptr ApplyImpl( - std::unique_ptr graph) const override; -}; - -} // namespace ir -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index ab96201b3..44821aadf 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -874,23 +874,9 @@ std::vector* 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 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. diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index ca5f0e27b..822bf5c9c 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.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 runtime_ctx_; - mutable const Scope* pre_scope_ = nullptr; }; extern bool OpSupportGPU(const std::string& op_type); diff --git a/paddle/fluid/inference/api/analysis_config.cc b/paddle/fluid/inference/api/analysis_config.cc index 4cad8a9df..92526f4e7 100644 --- a/paddle/fluid/inference/api/analysis_config.cc +++ b/paddle/fluid/inference/api/analysis_config.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(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(); } diff --git a/paddle/fluid/inference/api/paddle_analysis_config.h b/paddle/fluid/inference/api/paddle_analysis_config.h index 3b7faa540..9b05c3350 100644 --- a/paddle/fluid/inference/api/paddle_analysis_config.h +++ b/paddle/fluid/inference/api/paddle_analysis_config.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_; diff --git a/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc b/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc index e1787a717..5157bd280 100644 --- a/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc @@ -107,7 +107,6 @@ void SetConfig(AnalysisConfig *cfg) { cfg->DisableGpu(); cfg->SwitchSpecifyInputNames(); cfg->SwitchIrOptim(); - cfg->SwitchRuntimeContextCache(); if (FLAGS_zero_copy) { cfg->SwitchUseFeedFetchOps(false); } diff --git a/paddle/fluid/inference/tests/api/config_printer.h b/paddle/fluid/inference/tests/api/config_printer.h index b7b39d4dd..b0c23fbd5 100644 --- a/paddle/fluid/inference/tests/api/config_printer.h +++ b/paddle/fluid/inference/tests/api/config_printer.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"; diff --git a/paddle/fluid/pybind/inference_api.cc b/paddle/fluid/pybind/inference_api.cc index 11e9725ae..236afc77f 100644 --- a/paddle/fluid/pybind/inference_api.cc +++ b/paddle/fluid/pybind/inference_api.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); } -- GitLab