diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index d1193bbc8f5c65ced89f002c60fcb4bbf31d4af9..f60622eda2b3eca423d54c5d435e1657d581ac8a 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -409,7 +409,7 @@ cc_library(custom_operator SRCS custom_operator.cc DEPS tensor attribute framewo cc_test(custom_tensor_test SRCS custom_tensor_test.cc DEPS custom_tensor glog) #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) +cc_binary(new_executor SRCS new_exec_test.cc DEPS operator op_registry executor ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} profiler place) set(FLUID_FRAMEWORK_MODULES proto_desc memory lod_tensor executor data_feed_proto layer dynamic_loader custom_operator) diff --git a/paddle/fluid/framework/new_exec.h b/paddle/fluid/framework/new_exec.h index 015d5b7bac7e8a6d999fcca23bc790bd26c4c495..1ce87ecc938816bbffc35501cc9e874557ddb5b4 100644 --- a/paddle/fluid/framework/new_exec.h +++ b/paddle/fluid/framework/new_exec.h @@ -1,50 +1,146 @@ +// 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. + +#pragma once + +#include +#include #include #include #include #include -#include #include #include #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/platform/device_context.h" #include "paddle/fluid/framework/variable.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/variable_helper.h" +#include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/init.h" -#include -#include +namespace paddle { +namespace framework { +class RuntimeContextV2 { + public: + RuntimeContextV2(std::vector>& in_values, // NOLINT + std::vector>& out_values, // NOLINT + const std::map& in_name_map, + const std::map& out_name_map) + : input_values(std::move(in_values)), + output_values(std::move(out_values)), + input_name_map(in_name_map), + output_name_map(out_name_map) {} + std::vector> input_values; + std::vector> output_values; + const std::map& input_name_map; + const std::map& output_name_map; +}; -//USE_OP(fill_constant); -//USE_OP(elementwise_add); +class ExecutionContextV2 : public ExecutionContext { + public: + ExecutionContextV2(const OperatorBase& op, const Scope& scope, + const platform::DeviceContext& device_context, + const RuntimeContextV2& ctx) + : ExecutionContext(op, scope, device_context, RuntimeContext({}, {})), + ctx_(ctx) {} + + const std::vector MultiInputVar(const std::string& name) const { + LogVarUsageIfUnusedVarCheckEnabled(name); + + auto it = ctx_.input_name_map.find(name); + if (it == ctx_.input_name_map.end()) { + return {}; + } + // return {it->second.begin(), it->second.end()}; + return ctx_.input_values[it->second]; + } -using namespace std; + std::vector MultiOutputVar(const std::string& name) const { + auto it = ctx_.output_name_map.find(name); + if (it == ctx_.output_name_map.end()) { + return {}; + } + // return it->second; + return ctx_.output_values[it->second]; + } -namespace paddle { -namespace framework { + std::vector InNameList() const { + std::vector vec_temp; + vec_temp.reserve(ctx_.output_name_map.size()); + + for (auto& input : ctx_.output_name_map) { + vec_temp.push_back(input.first); + } + + return vec_temp; + } + + const Variable* InputVar(const std::string& name) const { + LogVarUsageIfUnusedVarCheckEnabled(name); + + auto it = ctx_.input_name_map.find(name); + if (it == ctx_.input_name_map.end()) return nullptr; + + PADDLE_ENFORCE_LE( + ctx_.input_values[it->second].size(), 1UL, + platform::errors::InvalidArgument( + "Operator %s's input %s should contain only one variable.", + GetOp().Type(), name)); + return ctx_.input_values[it->second].empty() + ? nullptr + : ctx_.input_values[it->second][0]; + } + + Variable* OutputVar(const std::string& name) const { + auto it = ctx_.output_name_map.find(name); + if (it == ctx_.output_name_map.end()) return nullptr; + + PADDLE_ENFORCE_LE( + ctx_.output_values[it->second].size(), 1UL, + platform::errors::InvalidArgument( + "Operator %s's output %s should contain only one variable.", + GetOp().Type(), name)); + return ctx_.output_values[it->second].empty() + ? nullptr + : ctx_.output_values[it->second][0]; + } + + const RuntimeContextV2& ctx_; +}; class RuntimeInferShapeContext : public InferShapeContext { public: - RuntimeInferShapeContext(const OperatorBase& op, const RuntimeContext& ctx) + RuntimeInferShapeContext(const OperatorBase& op, const RuntimeContextV2& ctx) : op_(op), ctx_(ctx) {} bool HasInput(const std::string& name) const override { // has only one input - const auto& ins = ctx_.inputs; + const auto& ins = ctx_.input_name_map; auto it = ins.find(name); if (it == ins.end()) { return false; } - const auto& in = it->second; + const auto& in = ctx_.input_values[it->second]; if (in.size() == 0) return false; PADDLE_ENFORCE_EQ( in.size(), 1UL, @@ -55,12 +151,12 @@ class RuntimeInferShapeContext : public InferShapeContext { bool HasOutput(const std::string& name) const override { // has only one output - const auto& outs = ctx_.outputs; + const auto& outs = ctx_.output_name_map; auto it = outs.find(name); if (it == outs.end()) { return false; } - const auto& out = it->second; + const auto& out = ctx_.output_values[it->second]; if (out.size() == 0) { return false; } @@ -72,12 +168,12 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasInputs(const std::string& name) const override { - const auto& ins = ctx_.inputs; + const auto& ins = ctx_.input_name_map; auto it = ins.find(name); - if (it == ins.end() || it->second.empty()) { + if (it == ins.end() || ctx_.input_values[it->second].empty()) { return false; } - for (auto& input : it->second) { + for (auto& input : ctx_.input_values[it->second]) { if (input == nullptr) { return false; } @@ -86,12 +182,12 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasOutputs(const std::string& name) const override { - const auto& outs = ctx_.outputs; + const auto& outs = ctx_.output_name_map; auto it = outs.find(name); - if (it == outs.end() || it->second.empty()) { + if (it == outs.end() || ctx_.output_values[it->second].empty()) { return false; } - for (auto& output : it->second) { + for (auto& output : ctx_.output_values[it->second]) { if (output == nullptr) { return false; } @@ -134,27 +230,27 @@ class RuntimeInferShapeContext : public InferShapeContext { 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); + auto in_it = ctx_.input_name_map.find(in); + auto out_it = ctx_.output_name_map.find(out); PADDLE_ENFORCE_NE( - in_it, ctx_.inputs.end(), + in_it, ctx_.input_name_map.end(), platform::errors::NotFound("Input %s does not exist.", in)); PADDLE_ENFORCE_NE( - out_it, ctx_.outputs.end(), + out_it, ctx_.output_name_map.end(), platform::errors::NotFound("Output %s does not exist.", out)); - PADDLE_ENFORCE_LT(i, in_it->second.size(), + PADDLE_ENFORCE_LT(i, ctx_.input_values[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(), + ctx_.input_values[in_it->second].size(), i)); + PADDLE_ENFORCE_LT(j, ctx_.output_values[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)); + ctx_.output_values[out_it->second].size(), j)); - Variable* in_var = in_it->second[i]; - Variable* out_var = out_it->second[j]; + Variable* in_var = ctx_.input_values[in_it->second][i]; + Variable* out_var = ctx_.output_values[out_it->second][j]; PADDLE_ENFORCE_EQ( in_var->Type(), out_var->Type(), @@ -181,18 +277,18 @@ class RuntimeInferShapeContext : public InferShapeContext { 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(), + auto in_it = ctx_.input_name_map.find(in); + auto out_it = ctx_.output_name_map.find(out); + PADDLE_ENFORCE_NE(in_it, ctx_.input_name_map.end(), platform::errors::NotFound( "Input [%s] found error in Op [%s]", in, op_.Type())); PADDLE_ENFORCE_NE( - out_it, ctx_.outputs.end(), + out_it, ctx_.output_name_map.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; + auto& in_var_list = ctx_.input_values[in_it->second]; + auto& out_var_list = ctx_.output_values[out_it->second]; PADDLE_ENFORCE_EQ( in_var_list.size(), out_var_list.size(), @@ -226,28 +322,28 @@ class RuntimeInferShapeContext : public InferShapeContext { 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); + auto in_it = ctx_.input_name_map.find(in); + auto out_it = ctx_.output_name_map.find(out); PADDLE_ENFORCE_NE( - in_it, ctx_.inputs.end(), + in_it, ctx_.input_name_map.end(), platform::errors::NotFound("Input %s does not exist.", in)); PADDLE_ENFORCE_NE( - out_it, ctx_.outputs.end(), + out_it, ctx_.output_name_map.end(), platform::errors::NotFound("Output %s does not exist.", out)); - PADDLE_ENFORCE_LT(i, in_it->second.size(), + PADDLE_ENFORCE_LT(i, ctx_.input_values[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(), + ctx_.input_values[in_it->second].size(), i)); + PADDLE_ENFORCE_LT(j, ctx_.output_values[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)); + ctx_.output_values[out_it->second].size(), j)); - Variable* in_var = in_it->second.at(i); + Variable* in_var = ctx_.input_values[in_it->second].at(i); if (!in_var->IsType()) return; - Variable* out_var = out_it->second.at(j); + Variable* out_var = ctx_.output_values[out_it->second].at(j); PADDLE_ENFORCE_EQ( out_var->IsType(), true, platform::errors::InvalidArgument( @@ -339,7 +435,7 @@ class RuntimeInferShapeContext : public InferShapeContext { } void SetOutputDim(const std::string& name, const DDim& dim) override { - //cerr << "set out dim" << endl; + // std::cerr << "set out dim" << std::endl; auto& vars = OutputVars(name); PADDLE_ENFORCE_EQ( vars.size(), 1UL, @@ -385,9 +481,7 @@ class RuntimeInferShapeContext : public InferShapeContext { } void SetDim(Variable* var, const DDim& dim) { - if (var->IsType()) { - var->GetMutable()->Resize(dim); } else if (var->IsType()) { var->GetMutable()->set_height(dim[0]); @@ -438,484 +532,515 @@ class RuntimeInferShapeContext : public InferShapeContext { private: const std::vector& InputVars(const std::string& name) const { - auto it = ctx_.inputs.find(name); + auto it = ctx_.input_name_map.find(name); PADDLE_ENFORCE_NE( - it, ctx_.inputs.end(), + it, ctx_.input_name_map.end(), platform::errors::NotFound( "Operator (%s) does not have the input (%s).", op_.Type(), name)); - return it->second; + return ctx_.input_values[it->second]; } const std::vector& OutputVars(const std::string& name) const { - auto it = ctx_.outputs.find(name); + auto it = ctx_.output_name_map.find(name); PADDLE_ENFORCE_NE( - it, ctx_.outputs.end(), + it, ctx_.output_name_map.end(), platform::errors::NotFound( "Operator (%s) does not have the outputs (%s).", op_.Type(), name)); - return it->second; + return ctx_.output_values[it->second]; } const OperatorBase& op_; - const RuntimeContext& ctx_; + const RuntimeContextV2& ctx_; }; - - -framework::ProgramDesc load_from_file( const std::string& file_name ) -{ +framework::ProgramDesc load_from_file(const std::string& file_name) { std::ifstream fin(file_name, std::ios::in | std::ios::binary); + if (!fin.is_open()) { + std::cout << "open file " << file_name << " faild!" << std::endl; + } fin.seekg(0, std::ios::end); std::string buffer(fin.tellg(), ' '); fin.seekg(0, std::ios::beg); fin.read(&buffer[0], buffer.size()); fin.close(); - - ProgramDesc program_desc( buffer ); + ProgramDesc program_desc(buffer); return program_desc; } - -struct VariableScope -{ - std::vector< std::unique_ptr > var_list; - std::map name2id; +struct VariableScope { + std::vector> var_list; + std::map name2id; }; +struct OpFuncNode { + // int unsed; + // std::map< std::string, std::vector > input_index; + // std::map< std::string, std::vector > output_index; + std::vector> input_index; + std::vector> output_index; + std::map input_name_map; + std::map output_name_map; + + using OpKernelFunc = std::function; + OpKernelFunc kernel_func_; +}; +int convert(const platform::Place& place) { + if (is_cpu_place(place)) { + return 0; + } + if (is_gpu_place(place)) { + return 1; + } + return -1; +} -struct OpFuncNode{ - - //int unsed; - std::map< std::string, std::vector > input_index; - std::map< std::string, std::vector > output_index; - - using OpKernelFunc = std::function; - OpKernelFunc kernel_func_; -}; +void build_variable_scope(const framework::ProgramDesc& pdesc, + VariableScope* var_scope) { + auto& global_block = pdesc.Block(0); -int convert(const platform::Place& place ) -{ - if ( is_cpu_place(place )) - { - return 0; + for (auto& var : global_block.AllVars()) { + if (var->Name() == framework::kEmptyVarName) { + continue; } - if( is_gpu_place( place )) - { - return 1; + // std::cerr << "var name " << var->Name() << std::endl; + + if (var_scope->name2id.find(var->Name()) == var_scope->name2id.end()) { + var_scope->name2id[var->Name()] = var_scope->var_list.size(); } - return -1; + auto v = new Variable(); + // v->GetMutable(); + InitializeVariable(v, var->GetType()); + var_scope->var_list.push_back(std::unique_ptr(v)); + } } -void build_variable_scope( const framework::ProgramDesc& pdesc, VariableScope* var_scope ) -{ +void build_op_func_list(const framework::ProgramDesc& pdesc, + std::vector& op_list, // NOLINT + std::vector& vec_func_list, // NOLINT + VariableScope* var_scope, + const platform::Place& place) { auto& global_block = pdesc.Block(0); - - - for (auto& var : global_block.AllVars()) { - if (var->Name() == framework::kEmptyVarName) { - continue; - } - //cerr << "var name " << var->Name() << endl; - if ( var_scope->name2id.find( var->Name() ) == var_scope->name2id.end() ) - { - var_scope->name2id[ var->Name() ] = var_scope->var_list.size(); + for (auto& op : global_block.AllOps()) { + // std::cerr << op->Type() << std::endl; + // bool debug = op->Type() == "softmax_with_cross_entropy_grad"; + bool debug = false; + + // std::cerr << "create op" << std::endl; + // auto op_base_u = OpRegistry::CreateOp(*op); + auto& info = OpInfoMap::Instance().Get(op->Type()); + + VariableNameMap inputs_1 = op->Inputs(); + VariableNameMap outputs_1 = op->Outputs(); + AttributeMap attrs_1 = op->GetAttrMap(); + + if (info.Checker() != nullptr) { + info.Checker()->Check(&attrs_1); + } + auto op_base = info.Creator()(op->Type(), inputs_1, outputs_1, attrs_1); + + auto input_names = op->Inputs(); + auto output_names = op->Outputs(); + + OpFuncNode op_func_node; + + // VariableValueMap ins_map; + // std::map > ins_name2id; + std::vector> ins_value; + std::vector> ins_index; + std::map ins_name_map; + for (auto& var_name_item : input_names) { + std::vector input_vars; + std::vector vec_ids; + input_vars.reserve(var_name_item.second.size()); + for (auto& var_name : var_name_item.second) { + auto it = var_scope->name2id.find(var_name); + assert(it != var_scope->name2id.end()); + input_vars.push_back(var_scope->var_list[it->second].get()); + vec_ids.push_back(it->second); } - - auto v = new Variable(); - //v->GetMutable(); - InitializeVariable(v, var->GetType()); - var_scope->var_list.push_back(std::unique_ptr(v)); - } -} + ins_value.emplace_back(std::move(input_vars)); + ins_index.emplace_back(std::move(vec_ids)); + ins_name_map[var_name_item.first] = ins_index.size() - 1; + // ins_map[ var_name_item.first ] = input_vars; + // ins_name2id[ var_name_item.first ] = vec_ids; + } + if (debug) std::cerr << "1" << std::endl; + + // VariableValueMap outs_map; + // std::map > outs_name2id; + std::vector> outs_value; + std::vector> outs_index; + std::map outs_name_map; + for (auto& var_name_item : output_names) { + std::vector output_vars; + std::vector vec_ids; + output_vars.reserve(var_name_item.second.size()); + for (auto& var_name : var_name_item.second) { + auto it = var_scope->name2id.find(var_name); + assert(it != var_scope->name2id.end()); + // std::cerr << it->second << "\t" << var_scope.var_list.size() << + // std::endl; + output_vars.push_back(var_scope->var_list[it->second].get()); + vec_ids.push_back(it->second); + } + outs_value.emplace_back(std::move(output_vars)); + outs_index.emplace_back(std::move(vec_ids)); + outs_name_map[var_name_item.first] = outs_index.size() - 1; + // outs_map[ var_name_item.first ] = output_vars; + // //std::cerr << ToTypeName(output_vars[0]->Type() ) << std::endl; + // outs_name2id[ var_name_item.first ] = vec_ids; + } -void build_op_func_list( const framework::ProgramDesc& pdesc, std::vector& op_list, - std::vector& vec_func_list, VariableScope* var_scope, - const platform::Place& place ) -{ - auto &global_block = pdesc.Block( 0 ); - - for ( auto& op : global_block.AllOps() ) - { - //cerr << op->Type() << endl; - //bool debug = op->Type() == "softmax_with_cross_entropy_grad"; - bool debug = false; - - //cerr << "create op" << endl; - //auto op_base_u = OpRegistry::CreateOp(*op); - auto& info = OpInfoMap::Instance().Get( op->Type() ); - - VariableNameMap inputs_1 = op->Inputs(); - VariableNameMap outputs_1 = op->Outputs(); - AttributeMap attrs_1 = op->GetAttrMap(); - - if (info.Checker() != nullptr) { - info.Checker()->Check(&attrs_1); - } - auto op_base = info.Creator()( op->Type(), inputs_1, outputs_1, attrs_1); - - auto input_names = op->Inputs(); - auto output_names = op->Outputs(); - - OpFuncNode op_func_node; - - VariableValueMap ins_map; - std::map< std::string, std::vector > ins_name2id; - for( auto& var_name_item : input_names) - { - std::vector input_vars; - std::vector vec_ids; - input_vars.reserve(var_name_item.second.size()); - for (auto& var_name : var_name_item.second) { - auto it = var_scope->name2id.find( var_name ); - assert( it != var_scope->name2id.end() ); - input_vars.push_back( var_scope->var_list[ it->second].get()); - vec_ids.push_back( it->second ); - } - ins_map[ var_name_item.first ] = input_vars; - ins_name2id[ var_name_item.first ] = vec_ids; + // op_func_node.input_index = ins_name2id; + // op_func_node.output_index = outs_name2id; + op_func_node.input_index = ins_index; + op_func_node.input_name_map = ins_name_map; + op_func_node.output_index = outs_index; + op_func_node.output_name_map = outs_name_map; + RuntimeContextV2 runtime_context(ins_value, outs_value, ins_name_map, + outs_name_map); + // runtime_context.inputs.swap( ins_map ); + // runtime_context.outputs.swap( outs_map ); + // runtime_context.input_values.swap(ins_value); + // runtime_context.input_name_map = ins_name_map; + // runtime_context.output_values.swap(outs_value); + // runtime_context.output_name_map = outs_name_map; + // std::cerr << "create runtime context" << std::endl; + RuntimeInferShapeContext infer_shape_ctx(*op_base, runtime_context); + static_cast(op_base)->InferShape( + &infer_shape_ctx); + // std::cerr << "fin infer shape" << std::endl; + auto& all_op_kernels = OperatorWithKernel::AllOpKernels(); + auto kernels_iter = all_op_kernels.find(op->Type()); + PADDLE_ENFORCE_NE( + kernels_iter, all_op_kernels.end(), + platform::errors::Unavailable( + "There are no kernels which are registered in the %s operator.", + op->Type())); + // std::cerr << "create kernel" << std::endl; + using OpKernelFunc = std::function; + using OpKernelMap = + std::unordered_map; + if (debug) std::cerr << "2" << std::endl; + OpKernelMap& kernels = kernels_iter->second; + // auto place = platform::CPUPlace(); + // auto place = platform::CUDAPlace(0); + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto* dev_ctx = pool.Get(place); + Scope scope; + auto exec_ctx = + ExecutionContextV2(*op_base, scope, *dev_ctx, runtime_context); + if (debug) std::cerr << "21" << std::endl; + auto expected_kernel_key = + dynamic_cast(op_base) + ->GetExpectedKernelType(exec_ctx); + if (debug) std::cerr << "22" << std::endl; + // std::cerr << "22" << std::endl; + + // add transfer log + // std::cerr << "in map size " << ins_map.size() << std::endl; + // VariableValueMap& ins_map_temp = runtime_context.inputs; + auto ins_map_temp = runtime_context.input_name_map; + // std::cerr << "ins map siz" << ins_map_temp.size() << std::endl; + for (auto& var_name_item : ins_map_temp) { + // std::cerr << "in name " << var_name_item.first << std::endl; + // auto& vec_ids = ins_name2id[ var_name_item.first ]; + for (size_t i = 0; + i < runtime_context.input_values[var_name_item.second].size(); ++i) { + auto var = runtime_context.input_values[var_name_item.second][i]; + auto tensor_in = static_cast(&(var->Get())); + if (!tensor_in->IsInitialized()) { + continue; } - if (debug ) cerr << "1" << endl; - - - VariableValueMap outs_map; - std::map > outs_name2id; - for( auto& var_name_item : output_names ) - { - std::vector output_vars; - std::vector vec_ids; - output_vars.reserve(var_name_item.second.size()); - for (auto& var_name : var_name_item.second) { - auto it = var_scope->name2id.find( var_name ); - assert( it != var_scope->name2id.end() ); - //cerr << it->second << "\t" << var_scope.var_list.size() << endl; - output_vars.push_back( var_scope->var_list[ it->second].get() ); - vec_ids.push_back( it->second ); - } - outs_map[ var_name_item.first ] = output_vars; - //cerr << ToTypeName(output_vars[0]->Type() ) << endl; - outs_name2id[ var_name_item.first ] = vec_ids; + // std::cerr << "i " << i << "\t" << tensor_in->IsInitialized() << + // std::endl; + auto kernel_type_for_var = + static_cast(op_base) + ->GetKernelTypeForVar(var_name_item.first, *tensor_in, + expected_kernel_key); + if (debug) { + std::cerr << "var name " << var_name_item.first << std::endl; + std::cerr << expected_kernel_key.place_ << "\t" + << kernel_type_for_var.place_ << std::endl; } - - - op_func_node.input_index = ins_name2id; - op_func_node.output_index = outs_name2id; - RuntimeContext runtime_context( {}, {}); - runtime_context.inputs.swap( ins_map ); - runtime_context.outputs.swap( outs_map ); - //cerr << "create runtime context" << endl; - RuntimeInferShapeContext infer_shape_ctx(*op_base, runtime_context); - static_cast(op_base)->InferShape( &infer_shape_ctx ); - //cerr << "fin infer shape" << endl; - auto& all_op_kernels = OperatorWithKernel::AllOpKernels(); - auto kernels_iter = all_op_kernels.find(op->Type() ); - PADDLE_ENFORCE_NE( - kernels_iter, all_op_kernels.end(), - platform::errors::Unavailable( - "There are no kernels which are registered in the %s operator.", - op->Type() )); - - //cerr << "create kernel" << endl; - using OpKernelFunc = std::function; - using OpKernelMap = - std::unordered_map; - if (debug ) cerr << "2" << endl; - OpKernelMap& kernels = kernels_iter->second; - //auto place = platform::CPUPlace(); - //auto place = platform::CUDAPlace(0); - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto* dev_ctx = pool.Get(place); - Scope scope; - auto exec_ctx = ExecutionContext(*op_base, scope, *dev_ctx, runtime_context ); - if (debug ) cerr << "21" << endl; - auto expected_kernel_key = dynamic_cast(op_base)->GetExpectedKernelType( exec_ctx ); - if (debug ) cerr << "22" << endl; - //cerr << "22" << endl; - - // add transfer log - //cerr << "in map size " << ins_map.size() << endl; - VariableValueMap& ins_map_temp = runtime_context.inputs; - //cerr << "ins map siz" << ins_map_temp.size() << endl; - for( auto& var_name_item : ins_map_temp ) - { - - //auto& vec_ids = ins_name2id[ var_name_item.first ]; - for( size_t i = 0; i < var_name_item.second.size(); ++i ) - { - auto var = var_name_item.second[i]; - auto tensor_in = static_cast(&(var->Get())); - if( !tensor_in->IsInitialized() ) - { - continue; - } - //cerr << "i " << i << "\t" << tensor_in->IsInitialized() << endl; - auto kernel_type_for_var = static_cast(op_base)->GetKernelTypeForVar( - var_name_item.first, *tensor_in, expected_kernel_key); - if( debug) - { - cerr << "var name " << var_name_item.first << endl; - cerr << expected_kernel_key.place_ << "\t" << kernel_type_for_var.place_ << endl; - } - if ( !platform::is_same_place(kernel_type_for_var.place_, - expected_kernel_key.place_) ) - { - if(debug) cerr << "add data transfer" << endl; - // need trans place - // add var in scope - // add copy op - std::string new_var_name = "temp_1" + to_string( var_scope->var_list.size() + 1); - auto v = new Variable(); - v->GetMutable(); - var_scope->name2id[ new_var_name ] = var_scope->var_list.size(); - var_scope->var_list.push_back(std::unique_ptr(v)); - - VariableNameMap copy_in_map; - //cerr << "ints name is " << input_names[var_name_item.first][i] << endl; - copy_in_map["X"] = { input_names[var_name_item.first][i] }; - VariableNameMap copy_out_map; - copy_out_map["Out"] = { new_var_name }; - AttributeMap attr_map; - attr_map["dst_place_type"] = convert( place ); - - std::map< std::string, std::vector > copy_ins_name2id; - copy_ins_name2id["X"] = ins_name2id[ var_name_item.first ]; - std::map< std::string, std::vector > copy_out_name2id; - copy_out_name2id["Out"] = { var_scope->name2id[new_var_name]}; - - //vec_ids[i] = var_scope->name2id[new_var_name]; - // update out runtime_context - op_func_node.input_index[ var_name_item.first ][i] = var_scope->name2id[new_var_name]; - - VariableValueMap copy_ins_value_map; - copy_ins_value_map["X"] = { var }; - VariableValueMap copy_outs_value_map; - copy_outs_value_map["Out"] = { v }; - - - - auto& copy_info = OpInfoMap::Instance().Get( "memcpy" ); - auto copy_op = copy_info.Creator()( "memcpy", copy_in_map, copy_out_map, attr_map); - if(debug) cerr << "create memcpy" << endl; - OpFuncNode copy_op_func_node; - copy_op_func_node.input_index = copy_ins_name2id; - copy_op_func_node.output_index = copy_out_name2id; - - RuntimeContext copy_runtime_context( {}, {}); - copy_runtime_context.inputs.swap( copy_ins_value_map ); - copy_runtime_context.outputs.swap( copy_outs_value_map ); - //cerr << "create runtime context" << endl; - RuntimeInferShapeContext copy_infer_shape_ctx(*copy_op, copy_runtime_context); - if(debug) cerr << "before infer shape" << endl; - static_cast(copy_op)->InferShape( ©_infer_shape_ctx ); - if(debug) cerr << "infer shape" << endl; - //cerr << "fin infer shape" << endl; - auto& all_op_kernels = OperatorWithKernel::AllOpKernels(); - auto kernels_iter = all_op_kernels.find( "memcpy" ); - PADDLE_ENFORCE_NE( - kernels_iter, all_op_kernels.end(), - platform::errors::Unavailable("There are no kernels which are registered in the memcpy operator.") ); - - - //cerr << "create kernel" << endl; - using OpKernelFunc = std::function; - using OpKernelMap = - std::unordered_map; - - OpKernelMap& kernels = kernels_iter->second; - //auto place = platform::CPUPlace(); - //auto place = platform::CUDAPlace(0); - - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto* dev_ctx = pool.Get(place); - Scope scope; - auto copy_exec_ctx = ExecutionContext(*copy_op, scope, *dev_ctx, copy_runtime_context ); - if (debug ) cerr << "21" << endl; - auto expected_kernel_key = dynamic_cast(copy_op)->GetExpectedKernelType( copy_exec_ctx ); - if (debug ) cerr << "22" << endl; - //cerr << "22" << endl; - auto kernel_iter = kernels.find(expected_kernel_key); - copy_op_func_node.kernel_func_ = OpKernelFunc( kernel_iter->second ); - copy_op_func_node.kernel_func_( copy_exec_ctx ); - if(debug) cerr << "run exe ctx" << endl; - - op_list.push_back( copy_op ); - vec_func_list.push_back( copy_op_func_node); - - - var_name_item.second[i] = v; - } - } + if (!platform::is_same_place(kernel_type_for_var.place_, + expected_kernel_key.place_)) { + if (debug) std::cerr << "add data transfer" << std::endl; + // need trans place + // add var in scope + // add copy op + std::string new_var_name = + "temp_1" + std::to_string(var_scope->var_list.size() + 1); + auto v = new Variable(); + v->GetMutable(); + var_scope->name2id[new_var_name] = var_scope->var_list.size(); + var_scope->var_list.push_back(std::unique_ptr(v)); + + VariableNameMap copy_in_map; + // std::cerr << "ints name is " << input_names[var_name_item.first][i] + // << std::endl; + copy_in_map["X"] = {input_names[var_name_item.first][i]}; + VariableNameMap copy_out_map; + copy_out_map["Out"] = {new_var_name}; + AttributeMap attr_map; + attr_map["dst_place_type"] = convert(place); + + // std::map< std::string, std::vector > copy_ins_name2id; + // copy_ins_name2id["X"] = ins_name2id[ var_name_item.first ]; + // std::map< std::string, std::vector > copy_out_name2id; + // copy_out_name2id["Out"] = { var_scope->name2id[new_var_name]}; + + // vec_ids[i] = var_scope->name2id[new_var_name]; + // update out runtime_context + op_func_node + .input_index[op_func_node.input_name_map[var_name_item.first]] + [i] = var_scope->name2id[new_var_name]; + + // VariableValueMap copy_ins_value_map; + // copy_ins_value_map["X"] = { var }; + // VariableValueMap copy_outs_value_map; + // copy_outs_value_map["Out"] = { v }; + + auto& copy_info = OpInfoMap::Instance().Get("memcpy"); + auto copy_op = copy_info.Creator()("memcpy", copy_in_map, + copy_out_map, attr_map); + if (debug) std::cerr << "create memcpy" << std::endl; + OpFuncNode copy_op_func_node; + // copy_op_func_node.input_index = copy_ins_name2id; + // copy_op_func_node.output_index = copy_out_name2id; + copy_op_func_node.input_index.push_back( + ins_index[ins_name_map[var_name_item.first]]); + copy_op_func_node.input_name_map["X"] = 0; + copy_op_func_node.output_index.push_back( + {var_scope->name2id[new_var_name]}); + copy_op_func_node.output_name_map["Out"] = 0; + std::vector> in_values; + std::vector> out_values; + in_values.push_back({var}); + out_values.push_back({v}); + RuntimeContextV2 copy_runtime_context( + in_values, out_values, copy_op_func_node.input_name_map, + copy_op_func_node.output_name_map); + // copy_runtime_context.input_values.push_back({var}); + // copy_runtime_context.input_name_map["X"] = 0; + // copy_runtime_context.output_values.push_back({v}); + // copy_runtime_context.output_name_map["Out"] = 0; + // copy_runtime_context.inputs.swap( copy_ins_value_map ); + // copy_runtime_context.outputs.swap( copy_outs_value_map ); + // std::cerr << "create runtime context" << std::endl; + RuntimeInferShapeContext copy_infer_shape_ctx(*copy_op, + copy_runtime_context); + if (debug) std::cerr << "before infer shape" << std::endl; + static_cast(copy_op) + ->InferShape(©_infer_shape_ctx); + if (debug) std::cerr << "infer shape" << std::endl; + // std::cerr << "fin infer shape" << std::endl; + auto& all_op_kernels = OperatorWithKernel::AllOpKernels(); + auto kernels_iter = all_op_kernels.find("memcpy"); + PADDLE_ENFORCE_NE(kernels_iter, all_op_kernels.end(), + platform::errors::Unavailable( + "There are no kernels which are registered in " + "the memcpy operator.")); + + // std::cerr << "create kernel" << std::endl; + using OpKernelFunc = std::function; + using OpKernelMap = std::unordered_map; + + OpKernelMap& kernels = kernels_iter->second; + // auto place = platform::CPUPlace(); + // auto place = platform::CUDAPlace(0); + + platform::DeviceContextPool& pool = + platform::DeviceContextPool::Instance(); + auto* dev_ctx = pool.Get(place); + Scope scope; + auto copy_exec_ctx = ExecutionContextV2(*copy_op, scope, *dev_ctx, + copy_runtime_context); + if (debug) std::cerr << "21" << std::endl; + auto expected_kernel_key = + dynamic_cast(copy_op) + ->GetExpectedKernelType(copy_exec_ctx); + if (debug) std::cerr << "22" << std::endl; + // std::cerr << "22" << std::endl; + auto kernel_iter = kernels.find(expected_kernel_key); + copy_op_func_node.kernel_func_ = OpKernelFunc(kernel_iter->second); + copy_op_func_node.kernel_func_(copy_exec_ctx); + if (debug) std::cerr << "run exe ctx" << std::endl; + + op_list.push_back(copy_op); + vec_func_list.push_back(copy_op_func_node); + + runtime_context.input_values[var_name_item.second][i] = v; } - - op_list.push_back( op_base ); - - auto kernel_iter = kernels.find(expected_kernel_key); - - if (debug ) cerr << "3" << endl; - op_func_node.kernel_func_ = OpKernelFunc(kernel_iter->second); - if (debug ) cerr << "3-1" << endl; - op_func_node.kernel_func_( exec_ctx ); - vec_func_list.push_back( op_func_node ); - if (debug ) cerr << "5" << endl; + } } - -} + op_list.push_back(op_base); + auto kernel_iter = kernels.find(expected_kernel_key); -void exec_op_func_list( const std::vector& vec_func_list, - std::vector< OperatorBase* >& op_list, - const VariableScope& var_scope, - const platform::Place& place) -{ - for( size_t i = 0; i < vec_func_list.size(); ++i ) - { - auto& func_node = vec_func_list[i]; - auto op_base = op_list[i]; - // build runtime cost - VariableValueMap ins_map; - for( auto& var_name_item : func_node.input_index) - { - std::vector input_vars; - - input_vars.reserve(var_name_item.second.size()); - for (auto& id : var_name_item.second) { - //cerr << var_name_item.first << "\t " << id << endl; - input_vars.emplace_back( var_scope.var_list[ id ].get() ); - } - ins_map.emplace( var_name_item.first, std::move(input_vars) ); - } + if (debug) std::cerr << "3" << std::endl; + op_func_node.kernel_func_ = OpKernelFunc(kernel_iter->second); + if (debug) std::cerr << "3-1" << std::endl; + op_func_node.kernel_func_(exec_ctx); + vec_func_list.push_back(op_func_node); + if (debug) std::cerr << "5" << std::endl; + } +} - VariableValueMap outs_map; - for( auto& var_name_item : func_node.output_index) - { - std::vector out_vars; - - out_vars.reserve(var_name_item.second.size()); - for (auto& id : var_name_item.second) { - //cerr << var_name_item.first << "\t " << id << endl; - out_vars.emplace_back( var_scope.var_list[ id ].get()); - } - outs_map.emplace( var_name_item.first, std::move( out_vars ) ); - } +void exec_op_func_list(const std::vector& vec_func_list, + std::vector& op_list, // NOLINT + const VariableScope& var_scope, + const platform::Place& place) { + for (size_t i = 0; i < vec_func_list.size(); ++i) { + auto& func_node = vec_func_list[i]; + auto op_base = op_list[i]; + // build runtime cost + // VariableValueMap ins_map; + std::vector> ins_map; + for (auto& var_name_item : func_node.input_name_map) { + std::vector input_vars; + + input_vars.reserve(func_node.input_index[var_name_item.second].size()); + for (auto& id : func_node.input_index[var_name_item.second]) { + // std::cerr << var_name_item.first << "\t " << id << std::endl; + input_vars.emplace_back(var_scope.var_list[id].get()); + } + // ins_map.emplace( var_name_item.first, std::move(input_vars) ); + ins_map.emplace_back(std::move(input_vars)); + } + + // VariableValueMap outs_map; + std::vector> outs_map; + for (auto& var_name_item : func_node.output_name_map) { + std::vector out_vars; - RuntimeContext runtime_context( {}, {}); - runtime_context.inputs.swap( ins_map ); - runtime_context.outputs.swap( outs_map ); - - RuntimeInferShapeContext infer_shape_ctx( *op_base, runtime_context); - - //dynamic_cast(op_base)->InferShape( &infer_shape_ctx ); - //RuntimeInferShapeContext infer_shape_ctx(*op_base, runtime_context); - static_cast(op_base)->InferShape( &infer_shape_ctx ); - - - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - //auto place = platform::CPUPlace(); - //auto place = platform::CUDAPlace(0); - auto* dev_ctx = pool.Get(place); - Scope scope; - - - auto exec_context = ExecutionContext(*op_base, scope, *dev_ctx, runtime_context ); - - func_node.kernel_func_( exec_context ); - + out_vars.reserve(func_node.output_index[var_name_item.second].size()); + for (auto& id : func_node.output_index[var_name_item.second]) { + // std::cerr << var_name_item.first << "\t " << id << std::endl; + out_vars.emplace_back(var_scope.var_list[id].get()); + } + // outs_map.emplace( var_name_item.first, std::move( out_vars ) ); + outs_map.emplace_back(std::move(out_vars)); } + + RuntimeContextV2 runtime_context( + ins_map, outs_map, func_node.input_name_map, func_node.output_name_map); + // runtime_context.inputs.swap( ins_map ); + // runtime_context.outputs.swap( outs_map ); + // runtime_context.input_values.swap(ins_map); + // runtime_context.output_values.swap(outs_map); + // runtime_context.input_name_map = func_node.input_name_map; + // runtime_context.output_name_map = func_node.output_name_map; + + RuntimeInferShapeContext infer_shape_ctx(*op_base, runtime_context); + + // dynamic_cast(op_base)->InferShape( + // &infer_shape_ctx ); + // RuntimeInferShapeContext infer_shape_ctx(*op_base, runtime_context); + static_cast(op_base)->InferShape( + &infer_shape_ctx); + + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + // auto place = platform::CPUPlace(); + // auto place = platform::CUDAPlace(0); + auto* dev_ctx = pool.Get(place); + Scope scope; + + auto exec_context = + ExecutionContextV2(*op_base, scope, *dev_ctx, runtime_context); + + func_node.kernel_func_(exec_context); + } } -class InterpreterCore -{ -public: - InterpreterCore( const platform::Place& place, const ProgramDesc& prog, const ProgramDesc& startup_prog) : place_(place), prog_(prog) { +class InterpreterCore { + public: + InterpreterCore(const platform::Place& place, const ProgramDesc& prog, + const ProgramDesc& startup_prog) + : place_(place), prog_(prog) { paddle::framework::InitDevices(); is_build = false; - paddle::framework::build_variable_scope( startup_prog, &global_scope ); - + paddle::framework::build_variable_scope(startup_prog, &global_scope); std::vector vec_func_list; - std::vector< paddle::framework::OperatorBase* > op_list; - paddle::framework::build_op_func_list( startup_prog, op_list, vec_func_list, &global_scope, place_); - - } - void run( const std::vector vec_name, const std::vector& vec_tensor, const vector& vec_fetch_name, - std::vector& vec_out) - { - //cerr << "run" << endl; - // set static data - if( is_build == false ) - { - paddle::framework::build_variable_scope( prog_, &global_scope ); + std::vector op_list; + paddle::framework::build_op_func_list(startup_prog, op_list, vec_func_list, + &global_scope, place_); + } + void run(const std::vector vec_name, + const std::vector& vec_tensor, + const std::vector& vec_fetch_name, + std::vector& vec_out) { // NOLINT + // std::cerr << "run" << std::endl; + // set static data + if (is_build == false) { + paddle::framework::build_variable_scope(prog_, &global_scope); } - for ( size_t i = 0; i < vec_name.size(); ++i ) - { - auto it = global_scope.name2id.find( vec_name[i] ); - //cerr << "find " << ( it != global_scope.name2id.end() ) <second]->GetMutable(); - //cerr << " get tensor" << endl; - feed_tensor->ShareDataWith( vec_tensor[i] ); - //cerr << "share buffer with" << endl; + + for (size_t i = 0; i < vec_name.size(); ++i) { + auto it = global_scope.name2id.find(vec_name[i]); + // std::cerr << "find " << (it != global_scope.name2id.end()) << + // std::endl; + assert(it != global_scope.name2id.end()); + + auto feed_tensor = + global_scope.var_list[it->second]->GetMutable(); + // std::cerr << " get tensor" << std::endl; + feed_tensor->ShareDataWith(vec_tensor[i]); + // std::cerr << "share buffer with" << std::endl; } - - if( is_build == false ) - { - paddle::framework::build_op_func_list( prog_, op_list, vec_func_list, &global_scope, place_); + + if (is_build == false) { + paddle::framework::build_op_func_list(prog_, op_list, vec_func_list, + &global_scope, place_); is_build = true; + } else { + paddle::framework::exec_op_func_list(vec_func_list, op_list, global_scope, + place_); } - else - { - paddle::framework::exec_op_func_list( vec_func_list, op_list, global_scope, place_ ); - } - - for( size_t i = 0; i < vec_fetch_name.size(); ++i ) - { - auto it = global_scope.name2id.find( vec_fetch_name[i] ); - assert( it != global_scope.name2id.end() ); - - auto fetch_tensor = global_scope.var_list[ it->second]->GetMutable(); - - - //cerr << "out " << fetch_tensor->data()[0] << endl; - if ( platform::is_gpu_place(fetch_tensor->place() ) ) - { - //cerr << "fetch gpu" << endl; - Tensor out; - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto* dev_ctx = pool.Get(place_); - dev_ctx->Wait(); - TensorCopySync(*fetch_tensor, platform::CPUPlace(), &out); - dev_ctx->Wait(); - //cerr << "out " << out << endl; - //cout << out.data()[0] << endl; - vec_out.push_back( out ); - } - else - { - cerr << "out " << *fetch_tensor << endl; - } + for (size_t i = 0; i < vec_fetch_name.size(); ++i) { + auto it = global_scope.name2id.find(vec_fetch_name[i]); + assert(it != global_scope.name2id.end()); + + auto fetch_tensor = + global_scope.var_list[it->second]->GetMutable(); + + // std::cerr << "out " << fetch_tensor->data()[0] << std::endl; + if (platform::is_gpu_place(fetch_tensor->place())) { + // std::cerr << "fetch gpu" << std::endl; + Tensor out; + platform::DeviceContextPool& pool = + platform::DeviceContextPool::Instance(); + auto* dev_ctx = pool.Get(place_); + dev_ctx->Wait(); + TensorCopySync(*fetch_tensor, platform::CPUPlace(), &out); + dev_ctx->Wait(); + // std::cerr << "out " << out << std::endl; + vec_out.push_back(out); + } else { + // std::cerr << "out " << *fetch_tensor << std::endl; + } } } -private: + + private: const platform::Place& place_; const ProgramDesc& prog_; paddle::framework::VariableScope global_scope; std::vector vec_func_list; - std::vector< paddle::framework::OperatorBase* > op_list; + std::vector op_list; bool is_build; - }; - -} -} - - - +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/new_exec_test.cc b/paddle/fluid/framework/new_exec_test.cc index 5a352e00ed05d31d33b88bd072dab72574ede1b9..ccb4d0d1aba50e16951f0548d42baa747c7128dc 100644 --- a/paddle/fluid/framework/new_exec_test.cc +++ b/paddle/fluid/framework/new_exec_test.cc @@ -1,6 +1,20 @@ -#include -#include +// 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 +#include +#include #include #include #include @@ -9,69 +23,58 @@ #include "paddle/fluid/framework/executor_gc_helper.h" #include "paddle/fluid/framework/garbage_collector.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/operator.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/tensor.h" -#include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/framework/variable.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/operator.h" - -#include "paddle/fluid/pybind/pybind.h" - +#include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/init.h" -#include "paddle/fluid/framework/new_exec.h" - -#include -#include - - -int main() -{ - paddle::framework::InitDevices(); - paddle::framework::VariableScope global_scope; - - - { - auto test_prog = paddle::framework::load_from_file( "lm_startup_program"); - paddle::framework::build_variable_scope( test_prog, &global_scope ); - - - std::vector vec_func_list; - std::vector> op_list; - paddle::framework::build_op_func_list( test_prog, op_list, vec_func_list, global_scope); - - paddle::framework::exec_op_func_list( vec_func_list, op_list, global_scope ); - } - - 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 vec_main_func_list; - std::vector> op_main_list; - paddle::framework::build_op_func_list( main_prog, op_main_list, vec_main_func_list, global_scope); +#include "paddle/fluid/pybind/pybind.h" - 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 ); +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 vec_func_list; + std::vector 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); + } + + std::cerr << "run main" << std::endl; + auto main_prog = paddle::framework::load_from_file("lm_main_program"); + + paddle::framework::build_variable_scope(main_prog, &global_scope); + + std::vector vec_main_func_list; + std::vector op_main_list; + paddle::framework::build_op_func_list( + main_prog, op_main_list, vec_main_func_list, &global_scope, place); + + auto start = std::chrono::steady_clock::now(); + // ProfilerStart("new_executor.prof"); + for (size_t i = 0; i < 2320; ++i) { + if (i % 200 == 0) { + std::cerr << i << std::endl; } - ProfilerStop(); - auto end = std::chrono::steady_clock::now(); - std::chrono::duration diff = end-start; - - cerr << "time cost " << diff.count() << endl; - + paddle::framework::exec_op_func_list(vec_main_func_list, op_main_list, + global_scope, place); + 33 + } + // ProfilerStop(); + auto end = std::chrono::steady_clock::now(); + std::chrono::duration diff = end - start; - return 1; + std::cerr << "time cost " << diff.count() << std::endl; + return 1; } diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 7a4ad9bd1824e91fe69eee7ce629039d07f3db1d..145e5411a88f395b48416b83aeb1a2868b52d3bc 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -1,11 +1,8 @@ /* Copyright (c) 2016 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. @@ -586,7 +583,6 @@ 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; @@ -602,7 +598,6 @@ class RuntimeInferShapeContext : public InferShapeContext { "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; @@ -620,7 +615,6 @@ class RuntimeInferShapeContext : public InferShapeContext { "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); @@ -634,7 +628,6 @@ class RuntimeInferShapeContext : public InferShapeContext { } return true; } - bool HasOutputs(const std::string& name) const override { const auto& outs = ctx_.outputs; auto it = outs.find(name); @@ -648,17 +641,13 @@ class RuntimeInferShapeContext : public InferShapeContext { } return true; } - AttrReader Attrs() const override { return AttrReader(op_.Attrs()); } - std::vector Inputs(const std::string& name) const override { return op_.Inputs(name); } - std::vector 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_; @@ -669,7 +658,6 @@ class RuntimeInferShapeContext : public InferShapeContext { 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_; @@ -681,7 +669,6 @@ class RuntimeInferShapeContext : public InferShapeContext { 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); @@ -702,16 +689,13 @@ class RuntimeInferShapeContext : public InferShapeContext { "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()) { auto& in_sele_rows = in_var->Get(); auto out_sele_rows = out_var->GetMutable(); @@ -728,7 +712,6 @@ class RuntimeInferShapeContext : public InferShapeContext { "or SelectedRows.")); } } - void ShareAllLoD(const std::string& in, const std::string& out) const override { auto in_it = ctx_.inputs.find(in); @@ -740,23 +723,18 @@ class RuntimeInferShapeContext : public InferShapeContext { 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()) return; Variable* out_var = out_var_list[i]; @@ -773,7 +751,6 @@ class RuntimeInferShapeContext : public InferShapeContext { 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); @@ -794,7 +771,6 @@ class RuntimeInferShapeContext : public InferShapeContext { "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()) return; Variable* out_var = out_it->second.at(j); @@ -805,7 +781,6 @@ class RuntimeInferShapeContext : public InferShapeContext { auto& in_tensor = in_var->Get(); auto* out_tensor = out_var->GetMutable(); 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? @@ -826,14 +801,12 @@ class RuntimeInferShapeContext : public InferShapeContext { #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( @@ -841,9 +814,7 @@ class RuntimeInferShapeContext : public InferShapeContext { "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 GetInputVarPtrs( const std::string& name) override { @@ -853,7 +824,6 @@ class RuntimeInferShapeContext : public InferShapeContext { res.insert(res.begin(), vars.begin(), vars.end()); return res; } - std::vector GetOutputVarPtrs( const std::string& name) override { const std::vector& vars = OutputVars(name); @@ -862,7 +832,6 @@ class RuntimeInferShapeContext : public InferShapeContext { res.insert(res.begin(), vars.begin(), vars.end()); return res; } - DDim GetInputDim(const std::string& name) const override { const std::vector& vars = InputVars(name); PADDLE_ENFORCE_EQ( @@ -872,22 +841,18 @@ class RuntimeInferShapeContext : public InferShapeContext { name, vars.size())); return this->GetDim(vars[0]); } - std::vector GetInputsDim(const std::string& name) const override { const std::vector& vars = InputVars(name); return GetDims(vars); } - std::vector GetInputsVarType( const std::string& name) const override { return GetVarTypes(InputVars(name)); } - std::vector 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( @@ -897,13 +862,11 @@ class RuntimeInferShapeContext : public InferShapeContext { name, vars.size())); SetDim(vars[0], dim); } - void SetOutputsDim(const std::string& name, const std::vector& dims) override { auto& vars = OutputVars(name); SetDims(vars, dims); } - protected: DDim GetDim(Variable* var) const { PADDLE_ENFORCE_NOT_NULL( @@ -919,7 +882,6 @@ class RuntimeInferShapeContext : public InferShapeContext { ToTypeName(var->Type()))); } } - std::vector GetDims(const std::vector& vars) const { std::vector ret; ret.reserve(vars.size()); @@ -927,12 +889,10 @@ class RuntimeInferShapeContext : public InferShapeContext { [this](Variable* var) { return this->GetDim(var); }); return ret; } - std::vector 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()) { var->GetMutable()->Resize(dim); @@ -945,7 +905,6 @@ class RuntimeInferShapeContext : public InferShapeContext { ToTypeName(var->Type()))); } } - void SetDims(const std::vector& vars, const std::vector& dims) { size_t length = vars.size(); @@ -962,13 +921,11 @@ class RuntimeInferShapeContext : public InferShapeContext { SetDim(vars[i], dims[i]); } } - void SetRepeatedDims(const std::string& name, const std::vector& dims) override { PADDLE_THROW(platform::errors::PreconditionNotMet( "SetRepeatedDims method only can be used in compile time.")); } - std::vector GetVarTypes( const std::vector& vars) const { std::vector retv; @@ -978,11 +935,9 @@ class RuntimeInferShapeContext : public InferShapeContext { this, std::placeholders::_1)); return retv; } - proto::VarType::Type GetVarType(Variable* var) const { return ToVarType(var->Type()); } - private: const std::vector& InputVars(const std::string& name) const { auto it = ctx_.inputs.find(name); @@ -992,7 +947,6 @@ class RuntimeInferShapeContext : public InferShapeContext { "Operator (%s) does not have the input (%s).", op_.Type(), name)); return it->second; } - const std::vector& OutputVars(const std::string& name) const { auto it = ctx_.outputs.find(name); PADDLE_ENFORCE_NE( @@ -1001,11 +955,10 @@ class RuntimeInferShapeContext : public InferShapeContext { "Operator (%s) does not have the outputs (%s).", op_.Type(), name)); return it->second; } - const OperatorBase& op_; const RuntimeContext& ctx_; }; */ } // namespace framework -} // namespace paddle +} // namespace paddle \ No newline at end of file