From f79a3a837cc162faee7c43c4cf616cbf8df706d3 Mon Sep 17 00:00:00 2001 From: sneaxiy Date: Fri, 22 Mar 2019 03:10:57 +0000 Subject: [PATCH] merge release/1.3 --- cmake/external/protobuf.cmake | 2 +- cmake/flags.cmake | 3 +- paddle/fluid/API.spec | 2 +- paddle/fluid/framework/details/CMakeLists.txt | 15 +- .../fluid/framework/details/build_strategy.cc | 4 +- .../framework/details/inplace_op_pass.cc | 35 +- .../fluid/framework/details/inplace_op_pass.h | 15 +- .../details/memory_optimize_helper.cc | 392 +++++++++- .../details/memory_optimize_helper.h | 119 +++- .../details/memory_optimize_helper_test.cc | 463 +++++++++++- .../framework/details/memory_optimize_pass.cc | 425 ++--------- .../framework/details/memory_optimize_pass.h | 57 +- .../details/memory_optimize_pass_test.cc | 417 ----------- .../details/sequential_execution_pass.cc | 1 + .../details/sequential_execution_pass.h | 2 - paddle/fluid/framework/inplace_op_inference.h | 2 +- .../framework/inplace_op_inference_test.cc | 33 +- .../ir/identity_scale_op_clean_pass.cc | 10 +- .../framework/ir/infer_clean_graph_pass.cc | 1 + paddle/fluid/framework/parallel_executor.cc | 11 +- .../analysis/ir_passes/subgraph_detector.cc | 71 -- .../analysis/ir_passes/subgraph_detector.h | 27 +- paddle/fluid/inference/api/paddle_api.h | 62 +- .../inference/api/paddle_pass_builder.cc | 46 ++ .../fluid/inference/api/paddle_pass_builder.h | 47 +- .../fluid/operators/controlflow/compare_op.cc | 10 +- .../detection/density_prior_box_op.h | 13 +- .../fluid/operators/detection/prior_box_op.h | 69 +- .../operators/detection/yolov3_loss_op.cc | 40 +- .../operators/elementwise/elementwise_op.h | 20 +- paddle/fluid/operators/lstm_op.h | 4 + paddle/fluid/operators/lstmp_op.h | 5 + paddle/fluid/operators/pool_op.cc | 144 ++-- paddle/fluid/operators/random_crop_op.h | 2 +- paddle/fluid/operators/slice_op.cu | 124 +++- .../softmax_with_cross_entropy_op.cu | 3 +- paddle/fluid/pybind/pybind.cc | 10 +- paddle/scripts/fast_install.sh | 669 ++++++++++++------ python/paddle/__init__.py | 1 + python/paddle/fluid/__init__.py | 1 - python/paddle/fluid/compiler.py | 5 +- .../fluid/contrib/int8_inference/README.md | 4 +- .../paddle/fluid/contrib/tests/CMakeLists.txt | 6 +- .../fluid/contrib/tests/test_calibration.py | 4 - python/paddle/fluid/executor.py | 68 +- python/paddle/fluid/framework.py | 3 +- python/paddle/fluid/io.py | 16 +- python/paddle/fluid/layers/control_flow.py | 20 +- python/paddle/fluid/layers/detection.py | 19 +- python/paddle/fluid/layers/io.py | 5 +- .../fluid/layers/layer_function_generator.py | 8 +- python/paddle/fluid/layers/nn.py | 101 ++- python/paddle/fluid/layers/ops.py | 4 +- python/paddle/fluid/optimizer.py | 6 +- python/paddle/fluid/parallel_executor.py | 5 +- .../fluid/tests/unittests/CMakeLists.txt | 9 + .../unittests/parallel_executor_test_base.py | 2 +- .../test_fuse_elewise_add_act_pass.py | 4 + .../fluid/tests/unittests/test_slice_op.py | 50 ++ .../fluid/transpiler/distribute_transpiler.py | 6 +- python/requirements.txt | 2 +- python/setup.py.in | 1 + tools/manylinux1/build_scripts/build.sh | 7 +- tools/run_mp.py | 129 ---- 64 files changed, 2149 insertions(+), 1712 deletions(-) delete mode 100644 paddle/fluid/framework/details/memory_optimize_pass_test.cc delete mode 100644 tools/run_mp.py diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake index e05b7694d..4c79d9d72 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -231,7 +231,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST) ) ENDFUNCTION() -SET(PROTOBUF_VERSION 3.1) +SET(PROTOBUF_VERSION 3.1.0) IF(NOT PROTOBUF_FOUND) build_protobuf(extern_protobuf FALSE) diff --git a/cmake/flags.cmake b/cmake/flags.cmake index 81e7868a6..36b533aa4 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -21,7 +21,7 @@ function(CheckCompilerCXX11Flag) if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 3.3) message(FATAL_ERROR "Unsupported Clang version. Clang >= 3.3 required.") endif() - endif() + endif() endif() endfunction() @@ -147,6 +147,7 @@ set(GPU_COMMON_FLAGS -Wno-error=unused-function # Warnings in Numpy Header. -Wno-error=array-bounds # Warnings in Eigen::array ) +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64") endif(NOT WIN32) if (APPLE) diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index 8a3c062db..230c96653 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -261,7 +261,7 @@ paddle.fluid.layers.increment ArgSpec(args=['x', 'value', 'in_place'], varargs=N paddle.fluid.layers.array_write ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.create_array ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.less_than ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords='ignored', defaults=(None, None)) -paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords='ignored', defaults=(None,)) +paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.array_read ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.array_length ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.IfElse.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,)) diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index 6621a59d3..dc308fd25 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -50,12 +50,15 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope) -cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper) +if(WITH_GPU) +cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info) +else() +cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info) +endif() + cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass) cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info) cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper) -cc_library(memory_early_delete_pass SRCS memory_early_delete_pass.cc DEPS memory_optimize_pass computation_op_handle scale_loss_grad_op_handle rpc_op_handle - all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass) cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle) cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper) cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass) @@ -67,13 +70,11 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle) -set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass inplace_op_pass) +set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass inplace_op_pass) if (WITH_GPU) list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass) endif() -cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph) -cc_test(memory_optimize_pass_test SRCS memory_optimize_pass_test.cc memory_optimize_pass.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry pass) - +cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry) cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS}) cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index 51ce97327..0c823b9ca 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -206,8 +206,6 @@ std::unique_ptr BuildStrategy::Apply( new std::vector(main_program.Block(0).AllOps()); graph->Set>(kAllOpDescs, all_op_descs); // take ownership - graph->Set(kGraphNodePool, - new GraphNodePool); // take ownership pass->Erase(kAllOpDescs); pass->SetNotOwned>(kAllOpDescs, all_op_descs); @@ -242,7 +240,9 @@ std::unique_ptr BuildStrategy::Apply( continue; } } + VLOG(3) << "Start Apply Pass " << pass->Type(); graph = pass->Apply(std::move(graph)); + VLOG(3) << "Finish Apply Pass " << pass->Type(); } return graph; } diff --git a/paddle/fluid/framework/details/inplace_op_pass.cc b/paddle/fluid/framework/details/inplace_op_pass.cc index 64368a5e8..c91fc81b2 100644 --- a/paddle/fluid/framework/details/inplace_op_pass.cc +++ b/paddle/fluid/framework/details/inplace_op_pass.cc @@ -49,7 +49,7 @@ DEFINE_bool( "If this option turns on, only these op in whitelist can be inplaced." "If it turns off, all of the running op can be candidate of inplaced op." "Such as scale, elementwise_add" - "By default, it's turned on"); + "By default, it's turned off"); DECLARE_string(memory_optimize_debug); @@ -171,16 +171,15 @@ void InplacePass::InplaceModifyDesc(const std::string& var, } } -const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var, - const std::string& cache_var, - const size_t& idx, - ir::Graph* graph) const { +const NodeSwapQueue InplacePass::TryInplaceModifyVar( + const std::string& var, const std::string& cache_var, const size_t& idx, + ir::Graph* graph) const { PADDLE_ENFORCE(var_nodes_[var].size() >= 1 && var_nodes_[var].at(0)->Var() != nullptr); std::unique_ptr var_desc(new VarDesc(*var_nodes_[var].at(0)->Var())); var_desc->SetName(cache_var); - SSANodePair swap_nodes; + NodeSwapQueue swap_nodes; for (size_t i = idx; i < view_.AllOps().size(); ++i) { auto* op = view_.AllOps()[i]; @@ -230,7 +229,7 @@ const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var, return swap_nodes; } -void InplacePass::CommitModify(const SSANodePair& swap_nodes, +void InplacePass::CommitModify(const NodeSwapQueue& swap_nodes, ir::Graph* graph) const { for (auto& pair : swap_nodes) { auto *node = pair.first, *cache_node = pair.second; @@ -245,7 +244,7 @@ void InplacePass::CommitModify(const SSANodePair& swap_nodes, } } -void InplacePass::WithdrawModify(const SSANodePair& nodes, +void InplacePass::WithdrawModify(const NodeSwapQueue& nodes, ir::Graph* graph) const { for (auto& pair : nodes) { auto *node = pair.first, *cache_node = pair.second; @@ -403,18 +402,20 @@ void GraphView::Build(ir::Graph* g) { // 2. track the nodes which used by parameter server. // these node can not be inplaced, otherwise trainer // pserver can not find each other name. - for (auto& node : g->Nodes()) { - if (!node->IsOp()) continue; - if (node->Name() == "send") { - for (auto& in : node->inputs) { - dup_nodes_.emplace(in->Name()); - } + auto update_skip_set = [&](ir::Node* node) { + for (auto& in : node->inputs) { + if (in->IsVar() && in->Var() != nullptr) dup_nodes_.emplace(in->Name()); } - if (node->Name() == "recv") { - for (auto& out : node->outputs) { + for (auto& out : node->outputs) { + if (out->IsVar() && out->Var() != nullptr) dup_nodes_.emplace(out->Name()); - } } + }; + for (auto& node : g->Nodes()) { + if (!node->IsOp()) continue; + if (node->Name() == "send") update_skip_set(node); + if (node->Name() == "recv") update_skip_set(node); + if (node->Name() == "prefetch") update_skip_set(node); } } diff --git a/paddle/fluid/framework/details/inplace_op_pass.h b/paddle/fluid/framework/details/inplace_op_pass.h index 1abcf1f27..7be7f3118 100644 --- a/paddle/fluid/framework/details/inplace_op_pass.h +++ b/paddle/fluid/framework/details/inplace_op_pass.h @@ -56,7 +56,8 @@ class GraphView { std::map> adj_list_; }; -typedef std::vector> SSANodePair; +// swap pairs in sequence +typedef std::vector> NodeSwapQueue; class InplacePass : public ir::Pass { public: InplacePass(); @@ -68,14 +69,14 @@ class InplacePass : public ir::Pass { void InitSSAGraphNodes() const; private: - const SSANodePair TryInplaceModifyVar(const std::string& var, - const std::string& cache_var, - const size_t& idx, - ir::Graph* graph) const; + const NodeSwapQueue TryInplaceModifyVar(const std::string& var, + const std::string& cache_var, + const size_t& idx, + ir::Graph* graph) const; - void CommitModify(const SSANodePair&, ir::Graph* graph) const; + void CommitModify(const NodeSwapQueue&, ir::Graph* graph) const; - void WithdrawModify(const SSANodePair& nodes, ir::Graph* graph) const; + void WithdrawModify(const NodeSwapQueue& nodes, ir::Graph* graph) const; void InplaceModifyDesc(const std::string& in_var, const std::string& out_var, const size_t& idx) const; diff --git a/paddle/fluid/framework/details/memory_optimize_helper.cc b/paddle/fluid/framework/details/memory_optimize_helper.cc index b56ef021e..6126c168c 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper.cc +++ b/paddle/fluid/framework/details/memory_optimize_helper.cc @@ -13,17 +13,114 @@ // limitations under the License. #include "paddle/fluid/framework/details/memory_optimize_helper.h" +#include +#include #include -#include +#include #include #include #include +#include "paddle/fluid/framework/var_desc.h" +#include "paddle/fluid/platform/cpu_info.h" + +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/gpu_info.h" +#endif // PADDLE_WITH_CUDA namespace paddle { namespace framework { namespace details { +using paddle::framework::VarDesc; + +std::vector SortOpLikeDescOrder(const ir::Graph& graph) { + PADDLE_ENFORCE(graph.Has(kAllOpDescs), + "Graph has no attribute of kAllOpDescs."); + // 1. get op desc order + auto& op_descs = graph.Get>(kAllOpDescs); + + // 2. topology sort order + auto nodes = graph.Nodes(); + std::deque ops; + FilterVariables(nodes, [&](ir::Node* op) { + if (op->IsOp() && op->Op() != nullptr) { + ops.emplace_back(op); + } + }); + std::unordered_map op_deps; + std::list ready_ops; + std::unordered_map> pending_ops; + + for (auto* op : ops) { + std::unordered_set preceding_op; + for (auto* in : op->inputs) { + if (in->inputs.empty()) continue; + PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp()); + preceding_op.emplace(in->inputs[0]); + pending_ops[in->inputs[0]].emplace(op); + } + op_deps[op] = preceding_op.size(); + if (preceding_op.empty()) { + ready_ops.emplace_back(op); + } + } + + // 3. generated op list based desc order and the topology order + std::vector ret; + std::list op_descs_list(op_descs.begin(), op_descs.end()); + + auto update_by_found_node = [&](ir::Node* found_node) { + for (auto* pending_op : pending_ops[found_node]) { + if (--op_deps[pending_op] == 0) { + ready_ops.emplace_back(pending_op); + } + } + ready_ops.remove(found_node); + ret.emplace_back(found_node); + }; + + while (!ready_ops.empty()) { + bool all_of_ready_op_unmatched = true; + for (auto it = op_descs_list.begin(); it != op_descs_list.end();) { + auto op_desc = *it; + ir::Node* found_node = nullptr; + for (auto* op : ready_ops) { + if (IsSameDesc(op->Op(), op_desc)) { + found_node = op; + break; + } + } + + // 3.1 op desc deleted by other pass + if (found_node == nullptr) { + ++it; + continue; + } else { + all_of_ready_op_unmatched = false; + it = op_descs_list.erase(it); + } + update_by_found_node(found_node); + } + + // 3.2 op descs are added by other pass + // preceding op non empty means some new op descs are + // created, but not contained in return node list. + // these new op desc may depend on each other. + std::list prev_ready_ops(ready_ops); + if (all_of_ready_op_unmatched) { + for (auto op : prev_ready_ops) { + update_by_found_node(op); + } + } + } + + PADDLE_ENFORCE(std::all_of( + op_deps.begin(), op_deps.end(), + [&](const std::pair& p) { return p.second == 0; })); + + return ret; +} -size_t NodeSizeInBytes(const VarDesc& node) { +size_t NodeSize(const VarDesc& node) { auto shape = node.GetShape(); int size = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies()); @@ -31,9 +128,9 @@ size_t NodeSizeInBytes(const VarDesc& node) { return type_size * std::abs(size); } -size_t NodeSizeInBytes(ir::Node* n) { +size_t NodeSize(ir::Node* n) { auto* desc = FindVarDescInBlock(n); - return NodeSizeInBytes(*desc); + return NodeSize(*desc); } std::string DebugStringImpl(VarDesc* var) { @@ -59,7 +156,6 @@ std::string DebugStringImpl(VarDesc* var) { std::string DebugString(ir::Node* var) { return DebugStringImpl(FindVarDescInBlock(var)); } -// return DebugString(var->Var()); } // NOTE(dzh): based ir node, if a large node has been reused // by a small size node, then next time it appear in pool, it will @@ -76,22 +172,26 @@ struct NodeComparator { bool operator()(ir::Node* lhs, ir::Node* rhs) const { auto* lhs_desc = FindVarDescInBlock(lhs); auto* rhs_desc = FindVarDescInBlock(rhs); + // match data type + if (lhs_desc->GetDataType() != rhs_desc->GetDataType()) { + return false; + } + // match shape auto lhs_shape = lhs_desc->GetShape(); auto rhs_shape = rhs_desc->GetShape(); if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) || (lhs_shape[0] != -1 && rhs_shape[0] != -1)) { - return NodeSizeInBytes(lhs) <= NodeSizeInBytes(rhs); + return NodeSize(lhs) <= NodeSize(rhs); } else { return false; } } }; -void OrderedNodeList::Insert(ir::Node* var, ir::Node* op) { +void OrderedSet::Insert(ir::Node* var) { PADDLE_ENFORCE(var->IsVar() && !var->IsCtrlVar()); - PADDLE_ENFORCE(op->IsOp()); if (mark_table_.count(var->Name()) != 0) { - mark_table_[var->Name()]->second.insert(op); + mark_table_[var->Name()]->emplace_back(var); return; } @@ -99,14 +199,15 @@ void OrderedNodeList::Insert(ir::Node* var, ir::Node* op) { auto var_shape = var_desc->GetShape(); int batch_size = static_cast(var_shape[0]); - NodeComparator compare_node; + NodeComparator functor; Iter it = nodes_.begin(); while (it != nodes_.end()) { - auto* cache_desc = FindVarDescInBlock(it->first); + auto& prev = it->front(); + auto* cache_desc = FindVarDescInBlock(prev); int cache_batch_size = cache_desc->GetShape()[0]; if ((cache_batch_size == -1 && batch_size == -1) || (cache_batch_size != -1 && batch_size != -1)) { - if (compare_node(it->first, var)) { + if (functor(prev, var)) { ++it; } else { break; @@ -118,62 +219,127 @@ void OrderedNodeList::Insert(ir::Node* var, ir::Node* op) { } } - it = - nodes_.insert(it, std::make_pair(var, std::unordered_set{op})); + it = nodes_.insert(it, {var}); mark_table_[var->Name()] = it; } -int OrderedNodeList::GetIndex(ir::Node* var) { +int OrderedSet::GetNodeIndexInPool(ir::Node* var) { return std::distance(nodes_.begin(), mark_table_[var->Name()]); } -ir::Node* OrderedNodeList::NodeMatch(ir::Node* var) const { +ir::Node* OrderedSet::FindBestFitNode(ir::Node* var) const { ir::Node* found_node = nullptr; - NodeComparator compare_node; + NodeComparator functor; for (auto it = nodes_.begin(); it != nodes_.end(); ++it) { - if (compare_node(var, it->first)) { - found_node = it->first; + auto& candidate = it->front(); + if (functor(var, candidate)) { + found_node = candidate; break; } } return found_node; } -void OrderedNodeList::Erase(ir::Node* var) { Erase(var->Name()); } +ir::Node* OrderedSet::FindNextBestFitNode(ir::Node* var, ir::Node* prev) const { + ir::Node* found_node = nullptr; + NodeComparator functor; + auto it = + std::find_if(nodes_.begin(), nodes_.end(), [&](const NodeVector& v) { + if (v.front() == prev) + return true; + else + return false; + }); + PADDLE_ENFORCE(it != nodes_.end(), "Not found previous in node list!"); + for (it = std::next(it); it != nodes_.end(); ++it) { + auto& candidate = it->front(); + if (functor(var, candidate)) { + found_node = candidate; + break; + } + } + return found_node; +} + +bool OrderedSet::Has(ir::Node* var) const { + if (mark_table_.count(var->Name())) { + auto& node_in_samename = mark_table_.at(var->Name()); + auto iter = + std::find_if(node_in_samename->begin(), node_in_samename->end(), + [&](ir::Node* n) { return n->Name() == var->Name(); }); + return iter != node_in_samename->end(); + } + return false; +} -void OrderedNodeList::Erase(const std::string& var) { +void OrderedSet::Erase(const std::string& var) { PADDLE_ENFORCE(mark_table_.count(var)); nodes_.erase(mark_table_[var]); mark_table_.erase(var); } -std::string OrderedNodeList::ToString() const { +void OrderedSet::Erase(ir::Node* var) { + PADDLE_ENFORCE(var != nullptr); + Erase(var->Name()); +} + +std::string OrderedSet::ToString() const { std::stringstream ss; for (auto it = nodes_.begin(); it != nodes_.end(); ++it) { - ss << DebugString(it->first) << " "; + for (auto& node : *it) { + ss << DebugString(node) << " "; + } } return ss.str(); } bool NodeCanReused(ir::Node* node) { + // valid the node is a var node if (node == nullptr || !node->IsVar() || node->IsCtrlVar()) return false; - // auto* desc = node->Var(); - bool flag = NodeCanReused(*node->Var()); + + bool flag = true; + // op output force generated in cpu, can not be reused. for (auto* op : node->inputs) { if (op->Op()->HasAttr("force_cpu")) { - // op output force generated in cpu, can not be reused. flag &= framework::AttrReader(op->Op()->GetAttrMap()) .Get("force_cpu") == 0; } } + // var desc validation. + flag &= NodeCanReused(*node->Var()); return flag; } +int MinChunkSize() { + int size{0}; +#ifdef PADDLE_WITH_CUDA + size = platform::GpuMinChunkSize(); +#else + size = platform::CpuMinChunkSize(); +#endif // PADDLE_WITH_CUDA + return size; +} + bool NodeCanReused(const VarDesc& node) { auto type = node.GetType(); - if (node.Persistable() || type != proto::VarType::LOD_TENSOR || - node.GetShape().empty()) { + // only these types holds bulk of gpu memory + if (!(type == proto::VarType::LOD_TENSOR || + type == proto::VarType::SELECTED_ROWS || + type == proto::VarType::LOD_TENSOR_ARRAY)) { + return false; + } + // persistable variable is parameter + if (node.Persistable()) { + return false; + } + // shape < min_chunk_size is meaningless. + // further more, fetched loss always has size = 1 + // which should not be reused. + auto shape = node.GetShape(); + int size = std::abs( + std::accumulate(shape.begin(), shape.end(), 1, std::multiplies())); + if (shape.empty() || size < MinChunkSize()) { return false; } // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad @@ -193,6 +359,176 @@ bool OpHasSubBlock(OpDesc* desc) { return false; } +ControlFlowGraph::ControlFlowGraph(const ir::Graph& graph) { + ops_ = SortOpLikeDescOrder(graph); + ConnectNodes(); +} + +void ControlFlowGraph::BuildCFGGraph() { + // FIXME(dzh): same effect with ConnectNodes, but use the control + // link to build dependency graph, it goes wrong in transformer. + for (ir::Node* op : ops_) { + for (auto& input_var : op->inputs) { + if (!input_var->inputs.empty()) { + PADDLE_ENFORCE( + input_var->inputs.size() == 1 && input_var->inputs[0]->IsOp(), + "Preceding Op Node of Var Node must be unique"); + auto* pred_op = input_var->inputs[0]; + if (pred_op->Op() != nullptr) { + predecessors_[op].insert(pred_op); + successors_[pred_op].insert(op); + } + } + if (input_var->IsVar() && !input_var->IsCtrlVar()) { + uses_[op].insert(input_var->Name()); + } + } + for (auto& output_var : op->outputs) { + // output var may be used by many op + for (auto* succ_op : output_var->outputs) { + if (succ_op->Op() != nullptr) { + successors_[op].insert(succ_op); + predecessors_[succ_op].insert(op); + } + } + if (output_var->IsVar() && !output_var->IsCtrlVar()) { + defs_[op].insert(output_var->Name()); + } + } + } +} + +void ControlFlowGraph::ConnectNodes() { + for (size_t i = 0; i < ops_.size(); ++i) { + auto& op = ops_[i]; + try { + auto& next_op = ops_.at(i + 1); + successors_[op].insert(next_op); + predecessors_[next_op].insert(op); + } catch (...) { + // do nothing + } + + FilterVariables(op->inputs, + [&](ir::Node* var) { uses_[op].emplace(var->Name()); }); + + FilterVariables(op->outputs, + [&](ir::Node* var) { defs_[op].emplace(var->Name()); }); + } +} + +void ControlFlowGraph::LiveVariableAnalysis() { + // NOTE(dzh): variable liveless analysis (a.k.a reversed_ops algorithm) + // compute the liveness of for each variable though reversed_ops algorithm. + // It iterates the operators from end to begin, compute the live in/live out + // variable set for each op, then the diff between in/out will be used for + // the variable reuse. For detail refer to + // http://www.cs.cornell.edu/courses/cs4120/2013fa/lectures/lec26-fa13.pdf + std::list work_list(ops_.rbegin(), ops_.rend()); + while (!work_list.empty()) { + ir::Node* op = work_list.front(); + work_list.pop_front(); + // get the live_in calculated before. Empty if first. + auto prev_live_in = std::move(live_in_[op]); + for (auto& s : successors_[op]) { + for (auto& var : live_in_[s]) { + live_out_[op].insert(var); + } + } + for (auto& var : uses_[op]) { + live_in_[op].insert(var); + } + for (auto& var : live_out_[op]) { + live_in_[op].insert(var); + } + for (auto& var : defs_[op]) { + live_in_[op].erase(var); + } + + // If the live_in is not changed, then the liveness analysis of + // predecessors is completed. + // + // Otherwise, recalculate the predecessors liveness + if (live_in_[op] != prev_live_in) { + for (auto& pre : predecessors_[op]) { + work_list.push_back(pre); + } + } + } +} + +void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node, + const std::string& new_node, + int begin_idx) { + // update graph from begin idx to the end + for (size_t i = begin_idx; i != ops_.size(); ++i) { + auto* op = ops_[i]; + if (uses_[op].find(old_node) != uses_[op].end()) { + uses_[op].erase(old_node); + uses_[op].insert(new_node); + } + if (defs_[op].find(old_node) != defs_[op].end()) { + defs_[op].erase(old_node); + defs_[op].insert(new_node); + } + if (live_in_[op].find(old_node) != live_in_[op].end()) { + live_in_[op].erase(old_node); + live_in_[op].insert(new_node); + } + if (live_out_[op].find(old_node) != live_out_[op].end()) { + live_out_[op].erase(old_node); + live_out_[op].insert(new_node); + } + } +} + +const std::set ControlFlowGraph::LiveIn(ir::Node* op) const { + auto it = live_in_.find(op); + PADDLE_ENFORCE( + it != live_in_.end(), + string::Sprintf("Expect %s in live_in, but Not Found.", op->Name())); + return it->second; +} + +const std::set ControlFlowGraph::LiveOut(ir::Node* op) const { + auto it = live_out_.find(op); + PADDLE_ENFORCE( + it != live_out_.end(), + string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); + return it->second; +} + +const std::set ControlFlowGraph::Use(ir::Node* op) const { + auto it = uses_.find(op); + PADDLE_ENFORCE( + it != uses_.end(), + string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); + return it->second; +} + +const std::vector ControlFlowGraph::Ops() const { return ops_; } + +std::vector& ControlFlowGraph::Ops() { return ops_; } + +ir::Node* ControlFlowGraph::GetNodeByName(const std::string& name, + ir::Node* op) const { + // in ssa-graph, different version nodes have same name, + // this function get the latest version var before target op + // It may return nullptr, such as data node. + ir::Node* found_node = nullptr; + for (auto* node : ops_) { + if (node == op) break; + for (auto& output : node->outputs) { + PADDLE_ENFORCE((output != nullptr && output->IsVar()), + "Output is empty!"); + if (output->Var() && output->Name() == name) { + found_node = output; + } + } + } + return found_node; +} + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/memory_optimize_helper.h b/paddle/fluid/framework/details/memory_optimize_helper.h index 064183d61..dba96309f 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper.h +++ b/paddle/fluid/framework/details/memory_optimize_helper.h @@ -17,6 +17,8 @@ #include #include #include +#include +#include #include #include #include @@ -27,41 +29,43 @@ namespace paddle { namespace framework { namespace details { -constexpr char kFetchedVars[] = "fetched_vars"; -constexpr char kGraphNodePool[] = "graph_node_pool"; +constexpr char kAllOpDescs[] = "all_op_descs"; -// NOTE(dzh): Variable and the operators use the var. -// for early delete pass. -// Because analysis var pass build base on ir::Node, which maybe released -// or modified between passes, so we use OpDesc* to mark ops. -using GraphNodePool = std::vector< - std::pair /* ops */>>; +std::vector SortOpLikeDescOrder(const ir::Graph& graph); -// NOTE(dzh): by default, it sort node in ascend order(by node bytes size). -// in fluid, -1 means the batch_size is determined in runtime. -// the node batch_size equal -1 always ranking in the front than the node not. +// NOTE(dzh): A ordered set for node reuse in memory optimize. +// the orderedset sort node in ascend order(by node bytes size). +// in fluid, -1 means the batch_size, which is determined in runtime. +// So the reuse happens between nodes who's batch_size both are -1 +// simultaneously or not. +// +// sort rule: +// rule 0 : smaller node ranking in front. +// rule 1 : batch_size equal -1 ranking in the front than the node not. +// // For example, // node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], .. -// O(1) insert, delete -class OrderedNodeList { - public: - using NodePair = std::pair>; - using Iter = typename std::list::iterator; - using ConstIter = typename std::list::const_iterator; - void Insert(ir::Node* var, ir::Node* op); +class OrderedSet { + public: + // nodes with same name exists in pool. + using NodeVector = std::vector; + using Iter = typename std::list::iterator; + using ConstIter = typename std::list::const_iterator; + void Insert(ir::Node* var); void Erase(ir::Node* var); - void Erase(const std::string& var); - - bool Has(ir::Node* var) { return mark_table_.count(var->Name()); } - - bool Has(const std::string& var) { return mark_table_.count(var); } - - ir::Node* NodeMatch(ir::Node* var) const; + bool Has(ir::Node* var) const; + void Clear() { + mark_table_.clear(); + nodes_.clear(); + } + // find the bestfit shape node block with var. + ir::Node* FindBestFitNode(ir::Node* var) const; + ir::Node* FindNextBestFitNode(ir::Node* var, ir::Node* prev) const; // map store non-const iterator, can not promise const - int GetIndex(ir::Node* var); + int GetNodeIndexInPool(ir::Node* var); // pool all node to string std::string ToString() const; @@ -69,18 +73,54 @@ class OrderedNodeList { Iter end() { return nodes_.end(); } ConstIter begin() const { return nodes_.begin(); } ConstIter end() const { return nodes_.end(); } - size_t size() const { return nodes_.size(); } - void Clear() { - mark_table_.clear(); - nodes_.clear(); - } + size_t size() const { return nodes_.size(); } private: // for searching. std::unordered_map mark_table_; - // node swap pairs. var -> ops dep var - std::list nodes_; + // node pool + std::list nodes_; +}; + +class ControlFlowGraph { + public: + ControlFlowGraph() = default; + // IR Graph + explicit ControlFlowGraph(const ir::Graph& graph); + + void LiveVariableAnalysis(); + + void RenameVarInCFGGraph(const std::string& old_node, + const std::string& new_node, int begin_idx); + + const std::set LiveIn(ir::Node* op) const; + const std::set LiveOut(ir::Node* op) const; + const std::set Use(ir::Node* op) const; + const std::vector Ops() const; + std::vector& Ops(); + + // for ssa-graph nodes + ir::Node* GetNodeByName(const std::string& name, ir::Node* op) const; + + private: + void BuildCFGGraph(); + void ConnectNodes(); + + using NodeListMap = std::unordered_map>; + using VarSetMap = std::map>; + // successors ops use the output variables. + NodeListMap successors_; + // predecessors ops generated input variables. + NodeListMap predecessors_; + // variables lived before run current op. + VarSetMap live_in_; + // variables lived after run current op. + VarSetMap live_out_; + VarSetMap uses_; // op inputs + VarSetMap defs_; // op outputs + + std::vector ops_; // op sequence by topology sort }; // valid a tensor can be reuse or not @@ -93,15 +133,24 @@ bool NodeCanReused(const VarDesc& node); bool OpHasSubBlock(OpDesc* desc); // node memory size in bytes -size_t NodeSizeInBytes(ir::Node* n); +size_t NodeSize(ir::Node* n); // node memory size in bytes -size_t NodeSizeInBytes(const VarDesc&); +size_t NodeSize(const VarDesc&); std::string DebugString(ir::Node* var); +// NOTE(dzhwinter) +// after node reuse, the replaced node shape is +// different with its VarDesc. So need to find the +// correct VarDesc in Block. VarDesc* FindVarDescInBlock(ir::Node* n); +static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) { + return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && + op1->Outputs() == op2->Outputs(); +} + template class FilterVariableImpl { public: diff --git a/paddle/fluid/framework/details/memory_optimize_helper_test.cc b/paddle/fluid/framework/details/memory_optimize_helper_test.cc index f2b9baf14..3cfe297a7 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper_test.cc +++ b/paddle/fluid/framework/details/memory_optimize_helper_test.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/details/memory_optimize_helper.h" #include #include +#include #include #include #include @@ -22,13 +23,19 @@ #include #include "glog/logging.h" #include "gtest/gtest.h" +#include "paddle/fluid/framework/details/graph_test_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/program_desc.h" namespace paddle { namespace framework { namespace details { -TEST(OrderedNodeList, Normal) { - OrderedNodeList pool; +TEST(OrderedSet, Normal) { + OrderedSet pool; std::vector> nodes; // clang-format off @@ -56,8 +63,15 @@ TEST(OrderedNodeList, Normal) { nodes.emplace_back(std::move(node)); } + // Insert for (auto& node : nodes) { - pool.Insert(node.get(), op.get()); + pool.Insert(node.get()); + } + + // Has/size + ASSERT_EQ(pool.size(), shapes.size()); + for (auto& node : nodes) { + ASSERT_TRUE(pool.Has(node.get())); } // assert its order and interface. @@ -66,14 +80,14 @@ TEST(OrderedNodeList, Normal) { std::cout << pool.ToString() << std::endl; ASSERT_EQ(pool.size(), static_cast(COUNT - 1)); - ASSERT_EQ(pool.GetIndex(nodes.back().get()), 0); + ASSERT_EQ(pool.GetNodeIndexInPool(nodes.back().get()), 0); { auto v1 = block_desc->Var("11"); v1->SetShape({-1, 256, 56, 56}); std::unique_ptr node1 = ir::CreateNodeForTest(v1); node1->inputs.emplace_back(op.get()); - auto* cache = pool.NodeMatch(node1.get()); + auto* cache = pool.FindBestFitNode(node1.get()); ASSERT_EQ(cache, nullptr); } { @@ -81,16 +95,447 @@ TEST(OrderedNodeList, Normal) { v2->SetShape({-1, 2, 5}); std::unique_ptr node1 = ir::CreateNodeForTest(v2); node1->inputs.emplace_back(op.get()); - auto* cache = pool.NodeMatch(node1.get()); - ASSERT_EQ(pool.GetIndex(cache), 2); // match 6:[-1,2,5] + auto* cache = pool.FindBestFitNode(node1.get()); + ASSERT_EQ(pool.GetNodeIndexInPool(cache), 2); // match 6:[-1,2,5] } { auto v3 = block_desc->Var("13"); v3->SetShape({2, 5}); std::unique_ptr node1 = ir::CreateNodeForTest(v3); node1->inputs.emplace_back(op.get()); - auto* cache = pool.NodeMatch(node1.get()); - ASSERT_EQ(pool.GetIndex(cache), 5); // match 4:[5,2] + auto* cache = pool.FindBestFitNode(node1.get()); + ASSERT_EQ(pool.GetNodeIndexInPool(cache), 5); // match 4:[5,2] + } +} + +TEST(OrderedSet, FindBestFitNode) { + OrderedSet pool; + std::vector> nodes; + ProgramDesc prog; + BlockDesc* block_desc = prog.MutableBlock(0); + auto* op_desc = block_desc->AppendOp(); + op_desc->SetType("dummy"); + std::unique_ptr op = ir::CreateNodeForTest(op_desc); + + { + auto desc = block_desc->Var("a"); + desc->SetShape({128, 128}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + { + auto desc = block_desc->Var("b"); + desc->SetShape({128, 129}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + { + auto desc = block_desc->Var("c"); + desc->SetShape({128, 128}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + + for (auto& node : nodes) { + pool.Insert(node.get()); + } + + // FindNextBestFitNode + auto* n = nodes[0].get(); + auto* cache = pool.FindBestFitNode(n); + PADDLE_ENFORCE(cache->Name() == "a"); + cache = pool.FindNextBestFitNode(n, cache); + PADDLE_ENFORCE(cache->Name() == "c"); + cache = pool.FindNextBestFitNode(n, cache); + PADDLE_ENFORCE(cache->Name() == "b"); +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_OPERATOR(sum, paddle::framework::DummyOp, + paddle::framework::SumOpMaker, + paddle::framework::DummyVarTypeInference); +REGISTER_OPERATOR(assign, paddle::framework::DummyOp, + paddle::framework::AssignOpMaker, + paddle::framework::DummyVarTypeInference); +REGISTER_OPERATOR(dummy, paddle::framework::DummyOp, + paddle::framework::SumOpMaker, + paddle::framework::DummyVarTypeInference); +/* + https://en.wikipedia.org/wiki/Live_variable_analysis + Create a customed classical dependency graph, left row is the instruction + number. + 1. a = 1 + 2. b = a + 3. c = a + 4. d = b + c + 5. e = d + + a--------+ + | | + b c + | | + d--------+ + | + e + Then analysis these variable's liveness range + */ + +namespace paddle { +namespace framework { +namespace details { + +inline static ProgramDesc FillProgramDesc() { + ProgramDesc prog; + prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("d")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("e")->SetType(proto::VarType::LOD_TENSOR); + { + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("assign"); + op->SetInput("X", {"a"}); + op->SetOutput("Out", {"b"}); + } + { + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("assign"); + op->SetInput("X", {"a"}); + op->SetOutput("Out", {"c"}); + } + { + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("sum"); + op->SetInput("X", {"b", "c"}); + op->SetOutput("Out", {"d"}); + } + { + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("assign"); + op->SetInput("X", {"d"}); + op->SetOutput("Out", {"e"}); + } + return prog; +} + +TEST(CFGGraph, IRGraph) { + // prepare ir graph + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + ControlFlowGraph cfg(graph); + cfg.LiveVariableAnalysis(); + + // test assign op + ASSERT_TRUE((std::set{"a"} == cfg.LiveIn(cfg.Ops()[0]))); + ASSERT_TRUE((std::set{"a", "b"} == cfg.LiveOut(cfg.Ops()[0]))); + + // test assign op + ASSERT_TRUE((std::set{"a", "b"} == cfg.LiveIn(cfg.Ops()[1]))); + ASSERT_TRUE((std::set{"b", "c"} == cfg.LiveOut(cfg.Ops()[1]))); + + // test sum op + ASSERT_TRUE((std::set{"b", "c"} == cfg.LiveIn(cfg.Ops()[2]))); + ASSERT_TRUE((std::set{"d"} == cfg.LiveOut(cfg.Ops()[2]))); + + // test assign op + ASSERT_TRUE((std::set{"d"} == cfg.LiveIn(cfg.Ops()[3]))); + ASSERT_TRUE((std::set{} == cfg.LiveOut(cfg.Ops()[3]))); +} + +// 1. normal test +TEST(SortOpLikeDescOrder, NormalTest) { + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + auto nodes = SortOpLikeDescOrder(graph); + auto op_descs = prog.Block(0).AllOps(); + for (size_t i = 0; i < nodes.size(); ++i) { + auto node = nodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +// 2. remove some op_desc +TEST(SortOpLikeDescOrder, RemoveOpDesc) { + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + auto nodes = graph.Nodes(); + auto op_descs = prog.Block(0).AllOps(); + ir::Node* found_node = nullptr; + for (auto node : nodes) { + if (node->IsOp() && node->outputs.back()->Name() == "e") { + found_node = node; + break; + } + } + PADDLE_ENFORCE(found_node != nullptr); + for (auto it = op_descs.begin(); it != op_descs.end();) { + if (IsSameDesc(*it, found_node->Op())) { + it = op_descs.erase(it); + } else { + ++it; + } + } + + auto find_node_in_graph = [&](std::string s) { + ir::Node* ret = nullptr; + for (auto n : graph.Nodes()) { + if (n->Name() == s) { + ret = n; + break; + } + } + PADDLE_ENFORCE(ret != nullptr); + return ret; + }; + + ir::Node* e = find_node_in_graph("e"); + ir::Node* d = find_node_in_graph("d"); + std::remove(d->outputs.begin(), d->outputs.end(), found_node); + graph.RemoveNode(found_node); + graph.RemoveNode(e); + + // other node keeps the same order + auto remain_nodes = SortOpLikeDescOrder(graph); + for (size_t i = 0; i < remain_nodes.size(); ++i) { + auto node = remain_nodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +// 3. add some op_desc +TEST(SortOpLikeDescOrder, AddOpDesc) { + auto prog = FillProgramDesc(); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + ir::Graph graph(prog); + + auto find_node_in_graph = [&](std::string s) { + ir::Node* ret = nullptr; + for (auto n : graph.Nodes()) { + if (n->Name() == s) { + ret = n; + break; + } + } + PADDLE_ENFORCE(ret != nullptr); + return ret; + }; + + // cached desc different with real one + // mimic the intermidiete pass modify the programdesc. + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + auto op_descs = prog.Block(0).AllOps(); + + auto op = prog.MutableBlock(0)->AppendOp(); + prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); + op->SetType("sum"); + op->SetInput("X", {"b", "c"}); + op->SetOutput("Out", {"d1"}); + ir::Node* node = graph.CreateOpNode(op); + ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); + ir::Node* b = find_node_in_graph("b"); + ir::Node* c = find_node_in_graph("c"); + node->outputs.emplace_back(d1); + node->inputs.emplace_back(b); + node->inputs.emplace_back(c); + d1->inputs.emplace_back(node); + b->outputs.emplace_back(node); + c->outputs.emplace_back(node); + op_descs.insert(op_descs.begin() + 4, op); + + auto nodes = SortOpLikeDescOrder(graph); + + for (size_t i = 0; i < nodes.size(); ++i) { + auto node = nodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +// 4. add and delete some op_desc +TEST(SortOpLikeDescOrder, AddAndDeleteOpDesc) { + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + auto find_node_in_graph = [&](std::string s) { + ir::Node* ret = nullptr; + for (auto n : graph.Nodes()) { + if (n->Name() == s) { + ret = n; + break; + } + } + PADDLE_ENFORCE(ret != nullptr); + return ret; + }; + + // remove sum node + auto op_descs = prog.Block(0).AllOps(); + ir::Node* found_node = nullptr; + auto nodes = graph.Nodes(); + for (auto node : nodes) { + if (node->Name() == "sum") { + found_node = node; + break; + } + } + PADDLE_ENFORCE(found_node != nullptr); + for (auto it = op_descs.begin(); it != op_descs.end();) { + if (IsSameDesc(*it, found_node->Op())) { + it = op_descs.erase(it); + } else { + ++it; + } + } + { + ir::Node* d = find_node_in_graph("d"); + ir::Node* c = find_node_in_graph("c"); + ir::Node* e = find_node_in_graph("e"); + std::remove(d->outputs.begin(), d->outputs.end(), found_node); + std::remove(c->outputs.begin(), c->outputs.end(), found_node); + ir::Node* pending_op = found_node->outputs[0]->outputs[0]; + graph.RemoveNode(e); + graph.RemoveNode(pending_op); + graph.RemoveNode(found_node); + } + + // add node + auto op = prog.MutableBlock(0)->AppendOp(); + prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); + op->SetType("sum"); + op->SetInput("X", {"b", "c"}); + op->SetOutput("Out", {"d1"}); + { + ir::Node* node = graph.CreateOpNode(op); + ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); + ir::Node* b = find_node_in_graph("b"); + ir::Node* c = find_node_in_graph("c"); + node->outputs.emplace_back(d1); + node->inputs.emplace_back(b); + node->inputs.emplace_back(c); + b->outputs.emplace_back(node); + c->outputs.emplace_back(node); + } + op_descs.insert(op_descs.begin() + 2, op); + + // check the order + auto mynodes = SortOpLikeDescOrder(graph); + for (size_t i = 0; i < mynodes.size(); ++i) { + auto node = mynodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +// 5. add and replace some op_desc inplace. +TEST(SortOpLikeDescOrder, AddAndReplaceOpDescInplace) { + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + auto find_node_in_graph = [&](std::string s) { + ir::Node* ret = nullptr; + for (auto n : graph.Nodes()) { + if (n->Name() == s) { + ret = n; + break; + } + } + PADDLE_ENFORCE(ret != nullptr); + return ret; + }; + + auto op_descs = prog.Block(0).AllOps(); + // add node + auto op = prog.MutableBlock(0)->AppendOp(); + prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); + op->SetType("sum"); + op->SetInput("X", {"b", "c"}); + op->SetOutput("Out", {"d1"}); + { + ir::Node* node = graph.CreateOpNode(op); + ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); + ir::Node* b = find_node_in_graph("b"); + ir::Node* c = find_node_in_graph("c"); + node->outputs.emplace_back(d1); + node->inputs.emplace_back(b); + node->inputs.emplace_back(c); + d1->inputs.emplace_back(node); + b->outputs.emplace_back(node); + c->outputs.emplace_back(node); + } + + op_descs.emplace_back(op); + + // replace op_desc inplace + auto nodes = graph.Nodes(); + ir::Node* found_node = nullptr; + for (auto node : nodes) { + if (node->IsOp() && node->Op() && node->Name() == "assign") { + if (node->outputs.size() == 1 && node->outputs[0]->Name() == "e") { + found_node = node; + break; + } + } + } + { + ir::Node* d = find_node_in_graph("d"); + ir::Node* e = find_node_in_graph("e"); + std::remove(d->outputs.begin(), d->outputs.end(), found_node); + std::remove(e->inputs.begin(), e->inputs.end(), found_node); + graph.RemoveNode(found_node); + } + op_descs.erase(op_descs.begin() + 3); + + auto replace_op = prog.MutableBlock(0)->AppendOp(); + replace_op->SetType("sum"); + replace_op->SetInput("X", {"d", "d1"}); + replace_op->SetOutput("Out", {"e"}); + { + ir::Node* sum2 = graph.CreateOpNode(replace_op); + ir::Node* e = find_node_in_graph("e"); + ir::Node* d = find_node_in_graph("d"); + ir::Node* d1 = find_node_in_graph("d1"); + sum2->inputs.emplace_back(d); + sum2->inputs.emplace_back(d1); + sum2->outputs.emplace_back(e); + e->inputs.emplace_back(sum2); + d->outputs.emplace_back(sum2); + d1->outputs.emplace_back(sum2); + } + + op_descs.emplace_back(replace_op); + // compare op order + auto graph_nodes = SortOpLikeDescOrder(graph); + for (size_t i = 0; i < graph_nodes.size(); ++i) { + auto node = graph_nodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); } } diff --git a/paddle/fluid/framework/details/memory_optimize_pass.cc b/paddle/fluid/framework/details/memory_optimize_pass.cc index 33ca45668..b35b967c7 100644 --- a/paddle/fluid/framework/details/memory_optimize_pass.cc +++ b/paddle/fluid/framework/details/memory_optimize_pass.cc @@ -43,16 +43,10 @@ namespace paddle { namespace framework { namespace details { -static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) { - return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && - op1->Outputs() == op2->Outputs(); -} - std::unique_ptr MemoryOptimizePass::ApplyImpl( std::unique_ptr graph) const { auto nodes = graph->Nodes(); - auto subblock_vars = GetSubBlockVars(nodes); - skip_set_.insert(subblock_vars.begin(), subblock_vars.end()); + CollectSkipVarsSet(nodes); cfg_.reset(new details::ControlFlowGraph(*graph)); cfg_->LiveVariableAnalysis(); @@ -75,82 +69,67 @@ std::unique_ptr MemoryOptimizePass::ApplyImpl( } for (auto& var : op->outputs) { - if (!NodeCanReused(var) || cfg_->Use(op).count(var->Name()) == 0 || - skip_set_.count(var->Name())) + if (var->IsVar() && !var->IsCtrlVar() && skip_set_.count(var->Name())) { + VLOG(3) << "Skip set contains variable of " << var->Name() + << "disable reuse on it. skipped"; continue; - ir::Node* cache = pool_.NodeMatch(var); - - if (var->Name() == FLAGS_memory_optimize_debug) { - VLOG(3) << "start match var " << DebugString(var) << " of op " - << op->Name(); - VLOG(3) << pool_.ToString(); - VLOG(3) << "matched in pool : " - << ((cache == nullptr) ? "False" : "True"); } + if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) { + ir::Node* cache = pool_.FindBestFitNode(var); + while (cache != nullptr && var->Name() == cache->Name()) { + VLOG(3) << "The same cache variable is cascade reused. " + << cache->Name() << " is re-filled to the pool after " + << "the reused op is finished. Current op can not " + << "replace it again. Skip this candidate."; + cache = pool_.FindNextBestFitNode(var, cache); + } + if (var->Name() == FLAGS_memory_optimize_debug) { + VLOG(3) << "start match var " << DebugString(var) << " of op " + << op->Name(); + VLOG(3) << pool_.ToString(); + VLOG(3) << "matched in pool : " + << ((cache == nullptr) ? "False" : "True"); + } - if (cache == nullptr) continue; - if (var->Name() == cache->Name()) { - VLOG(3) << "The same cache variable is cascade reused." << var->Name() - << " is re-filled to the pool after" - << "the reused op is finished. Current op can not " - << "replace it again. Skip this candidate."; - continue; - - int node_idx_in_pool = pool_.GetIndex(cache); - VLOG(3) << string::Sprintf( - "!!! %s, %s => %s, cache idx %d, pool size %d", - std::to_string(reuse_id++), DebugString(var), DebugString(cache), - node_idx_in_pool, static_cast(pool_.size())); - // update CFG Graph on the fly. - // reused var maybe re-fill into the pool - cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx); - // NOTE(dzhwinter): we need to both update the ProgramDesc - // and IR Graph. because op_desc/var_desc is used in CreateOp, - // CreateVar when running happens. But IR Graph - // define the dependence relationship between nodes. - RenameVarInGraphDesc(var->Name(), cache->Name(), idx); - RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get()); - - pool_.Erase(cache); - } - // fill the pool - std::unordered_set unlived_vars; - for (auto var : cfg_->LiveIn(op)) { - if (cfg_->LiveOut(op).count(var) == 0) { - unlived_vars.emplace(var); + if (cache != nullptr) { + int node_idx_in_pool = pool_.GetNodeIndexInPool(cache); + VLOG(3) << string::Sprintf( + "!!! %s, %s => %s, cache idx %d, pool size %d", + std::to_string(reuse_id++), DebugString(var), DebugString(cache), + node_idx_in_pool, static_cast(pool_.size())); + // NOTE(dzhwinter): update the ProgramDesc/IR Graph + // and the CFG Graph on the fly. + // + // IR Graph define the dependence relationship between nodes. + // + // ProgramDesc defines the input/output vars. Its used in + // CreateOp, CreateVar when running happens. + // + // CFG Graph store the liveness information, when reuse happens + // we also need to update the variable liveness. + const std::string var_name = var->Name(); + const std::string cache_name = cache->Name(); + + cfg_->RenameVarInCFGGraph(var_name, cache_name, idx); + RenameVarInGraphDesc(var_name, cache_name, idx); + RenameVarInGraphNode(var_name, cache_name, idx, graph.get()); + pool_.Erase(cache_name); } } - for (auto var : unlived_vars) { - ir::Node* var_node = cfg_->GetNodeFromVarName(var, op); + } + // fill the pool + for (auto var : cfg_->LiveIn(op)) { + if (cfg_->LiveOut(op).count(var) == 0) { + ir::Node* var_node = cfg_->GetNodeByName(var, op); + if (var_node == nullptr || var_node->IsCtrlVar()) continue; if (NodeCanReused(var_node) && !pool_.Has(var_node)) { - pool_.Insert(var_node, op); + pool_.Insert(var_node); } } } } graph->ResolveHazard(var_nodes_); - // For early delete pass. use GraphNodePool load the unlived vars. - // 1. find all deps op for each unlived var in memory pool. - for (auto& op : graph->Nodes()) { - for (auto& var : op->inputs) { - if (pool_.Has(var)) { - pool_.Insert(var, op); - } - } - } - // 2. convert ir node based memory pool to graph node - // because Node* maybe released bettwen passes. - auto& graph_pool = graph->Get(kGraphNodePool); - for (auto it = pool_.begin(); it != pool_.end(); ++it) { - std::unordered_set descs; - for (auto& op : it->second) { - PADDLE_ENFORCE(op->IsOp()); - descs.insert(op->Op()); - } - graph_pool.push_back(std::make_pair(it->first->Name(), descs)); - } - return graph; } @@ -199,12 +178,12 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const { PADDLE_ENFORCE(sub_op != nullptr); for (auto* var : sub_op->outputs) { if (NodeCanReused(var)) { - ir::Node* cache = pool_.NodeMatch(var); + ir::Node* cache = pool_.FindBestFitNode(var); if (cache != nullptr) { if (var->Var()->GetDataType() != cache->Var()->GetDataType()) { continue; } - int node_idx_in_pool = pool_.GetIndex(cache); + int node_idx_in_pool = pool_.GetNodeIndexInPool(cache); VLOG(3) << string::Sprintf( "!!! %s, %s => %s, cache idx %d, pool size %d", std::to_string(sub_reuse_id++), DebugString(var), @@ -224,20 +203,27 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const { } } -std::unordered_set MemoryOptimizePass::GetSubBlockVars( +void MemoryOptimizePass::CollectSkipVarsSet( const std::unordered_set& nodes) const { - std::unordered_set vars; + auto update_skip_set = [&](OpDesc* op_desc) { + auto inputs = op_desc->InputArgumentNames(); + auto outputs = op_desc->OutputArgumentNames(); + skip_set_.insert(inputs.begin(), inputs.end()); + skip_set_.insert(outputs.begin(), outputs.end()); + }; for (auto& op : nodes) { if (!op->IsOp() || op->Op() == nullptr) continue; auto* op_desc = op->Op(); - if (OpHasSubBlock(op_desc)) { - auto inputs = op_desc->InputArgumentNames(); - auto outputs = op_desc->OutputArgumentNames(); - vars.insert(inputs.begin(), inputs.end()); - vars.insert(outputs.begin(), outputs.end()); - } + // NOTE(dzhwinter): + // current block can not reuse next level block vars. + if (OpHasSubBlock(op_desc)) update_skip_set(op_desc); + // NOTE(dzhwinter): + // distributed ops input/output name need to + // keep same bettwen trainer/pserver + if (op_desc->Type() == "send") update_skip_set(op_desc); + if (op_desc->Type() == "recv") update_skip_set(op_desc); + if (op_desc->Type() == "prefetch") update_skip_set(op_desc); } - return vars; } void MemoryOptimizePass::RenameVarInGraphDesc(const std::string& var, @@ -291,8 +277,7 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, // redirect the input to the latest version of cache_var for (auto* node : op->inputs) { if (node->Name() == var) { - ir::Node* cache_node = graph->CreateVarNode(var_desc.get()); - var_nodes_[cache_var].emplace_back(cache_node); + ir::Node* cache_node = var_nodes_[cache_var].back(); // swap node to cache_node cache_node->outputs.insert(cache_node->outputs.end(), @@ -301,11 +286,15 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, auto* prev_op = node->inputs[0]; std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node, cache_node); - cache_node->inputs.emplace_back(prev_op); for (auto* next_op : node->outputs) { std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, cache_node); } + + // erase unused node + auto& nodes = var_nodes_.at(var); + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + graph->RemoveNode(node); } } @@ -325,271 +314,14 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, cache_node); } - } - } - } - - // release node of unused var in graph - for (auto* node : var_nodes_[var]) { - graph->RemoveNode(node); - } - var_nodes_.at(var).clear(); -} - -std::vector SortOpLikeDescOrder(const ir::Graph& graph) { - PADDLE_ENFORCE(graph.Has(kAllOpDescs), - "Graph has no attribute of kAllOpDescs."); - // 1. get op desc order - auto& op_descs = graph.Get>(kAllOpDescs); - - // 2. topology sort order - auto nodes = graph.Nodes(); - std::deque ops; - FilterVariables(nodes, [&](ir::Node* op) { - if (op->IsOp() && op->Op() != nullptr) { - ops.emplace_back(op); - } - }); - std::unordered_map op_deps; - std::list ready_ops; - std::unordered_map> pending_ops; - - for (auto* op : ops) { - std::unordered_set preceding_op; - for (auto* in : op->inputs) { - if (in->inputs.empty()) continue; - PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp()); - preceding_op.emplace(in->inputs[0]); - pending_ops[in->inputs[0]].emplace(op); - } - op_deps[op] = preceding_op.size(); - if (preceding_op.empty()) { - ready_ops.emplace_back(op); - } - } - - // 3. generated op list based desc order and the topology order - std::vector ret; - std::list op_descs_list(op_descs.begin(), op_descs.end()); - - auto update_by_found_node = [&](ir::Node* found_node) { - for (auto* pending_op : pending_ops[found_node]) { - if (--op_deps[pending_op] == 0) { - ready_ops.emplace_back(pending_op); - } - } - ready_ops.remove(found_node); - ret.emplace_back(found_node); - }; - - while (!ready_ops.empty()) { - bool all_of_ready_op_unmatched = true; - for (auto it = op_descs_list.begin(); it != op_descs_list.end();) { - auto op_desc = *it; - ir::Node* found_node = nullptr; - for (auto* op : ready_ops) { - if (IsSameDesc(op->Op(), op_desc)) { - found_node = op; - break; - } - } - - // 3.1 op desc deleted by other pass - if (found_node == nullptr) { - ++it; - continue; - } else { - all_of_ready_op_unmatched = false; - it = op_descs_list.erase(it); - } - update_by_found_node(found_node); - } - - // 3.2 op descs are added by other pass - // preceding op non empty means some new op descs are - // created, but not contained in return node list. - // these new op desc may depend on each other. - std::list prev_ready_ops(ready_ops); - if (all_of_ready_op_unmatched) { - for (auto op : prev_ready_ops) { - update_by_found_node(op); - } - } - } - - PADDLE_ENFORCE(std::all_of( - op_deps.begin(), op_deps.end(), - [&](const std::pair& p) { return p.second == 0; })); - - return ret; -} - -ControlFlowGraph::ControlFlowGraph(const ir::Graph& graph) { - ops_ = SortOpLikeDescOrder(graph); - ConnectNodes(); -} - -void ControlFlowGraph::BuildCFGGraph() { - // FIXME(dzh): same effect with ConnectNodes, but use the control - // link to build dependency graph, it goes wrong in transformer. - for (ir::Node* op : ops_) { - for (auto& input_var : op->inputs) { - if (!input_var->inputs.empty()) { - PADDLE_ENFORCE( - input_var->inputs.size() == 1 && input_var->inputs[0]->IsOp(), - "Preceding Op Node of Var Node must be unique"); - auto* pred_op = input_var->inputs[0]; - if (pred_op->Op() != nullptr) { - predecessors_[op].insert(pred_op); - successors_[pred_op].insert(op); - } - } - if (input_var->IsVar() && !input_var->IsCtrlVar()) { - uses_[op].insert(input_var->Name()); - } - } - for (auto& output_var : op->outputs) { - // output var may be used by many op - for (auto* succ_op : output_var->outputs) { - if (succ_op->Op() != nullptr) { - successors_[op].insert(succ_op); - predecessors_[succ_op].insert(op); - } - } - if (output_var->IsVar() && !output_var->IsCtrlVar()) { - defs_[op].insert(output_var->Name()); - } - } - } -} - -void ControlFlowGraph::ConnectNodes() { - for (size_t i = 0; i < ops_.size(); ++i) { - auto& op = ops_[i]; - try { - auto& next_op = ops_.at(i + 1); - successors_[op].insert(next_op); - predecessors_[next_op].insert(op); - } catch (...) { - // do nothing - } - - FilterVariables(op->inputs, - [&](ir::Node* var) { uses_[op].emplace(var->Name()); }); - - FilterVariables(op->outputs, - [&](ir::Node* var) { defs_[op].emplace(var->Name()); }); - } -} - -void ControlFlowGraph::LiveVariableAnalysis() { - // NOTE(dzh): variable liveless analysis (a.k.a reversed_ops algorithm) - // compute the liveness of for each variable though reversed_ops algorithm. - // It iterates the operators from end to begin, compute the live in/live out - // variable set for each op, then the diff between in/out will be used for - // the variable reuse. For detail refer to - // http://www.cs.cornell.edu/courses/cs4120/2013fa/lectures/lec26-fa13.pdf - std::list work_list(ops_.rbegin(), ops_.rend()); - while (!work_list.empty()) { - ir::Node* op = work_list.front(); - work_list.pop_front(); - // get the live_in calculated before. Empty if first. - auto prev_live_in = std::move(live_in_[op]); - for (auto& s : successors_[op]) { - for (auto& var : live_in_[s]) { - live_out_[op].insert(var); - } - } - for (auto& var : uses_[op]) { - live_in_[op].insert(var); - } - for (auto& var : live_out_[op]) { - live_in_[op].insert(var); - } - for (auto& var : defs_[op]) { - live_in_[op].erase(var); - } - - // If the live_in is not changed, then the liveness analysis of - // predecessors is completed. - // - // Otherwise, recalculate the predecessors liveness - if (live_in_[op] != prev_live_in) { - for (auto& pre : predecessors_[op]) { - work_list.push_back(pre); - } - } - } -} - -void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node, - const std::string& new_node, - int begin_idx) { - // update graph from begin idx to the end - for (size_t i = begin_idx; i != ops_.size(); ++i) { - auto* op = ops_[i]; - if (uses_[op].find(old_node) != uses_[op].end()) { - uses_[op].erase(old_node); - uses_[op].insert(new_node); - } - if (defs_[op].find(old_node) != defs_[op].end()) { - defs_[op].erase(old_node); - defs_[op].insert(new_node); - } - if (live_in_[op].find(old_node) != live_in_[op].end()) { - live_in_[op].erase(old_node); - live_in_[op].insert(new_node); - } - if (live_out_[op].find(old_node) != live_out_[op].end()) { - live_out_[op].erase(old_node); - live_out_[op].insert(new_node); - } - } -} - -const std::set ControlFlowGraph::LiveIn(ir::Node* op) const { - auto it = live_in_.find(op); - PADDLE_ENFORCE( - it != live_in_.end(), - string::Sprintf("Expect %s in live_in, but Not Found.", op->Name())); - return it->second; -} - -const std::set ControlFlowGraph::LiveOut(ir::Node* op) const { - auto it = live_out_.find(op); - PADDLE_ENFORCE( - it != live_out_.end(), - string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); - return it->second; -} - -const std::set ControlFlowGraph::Use(ir::Node* op) const { - auto it = uses_.find(op); - PADDLE_ENFORCE( - it != uses_.end(), - string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); - return it->second; -} - -const std::vector ControlFlowGraph::Ops() const { return ops_; } - -std::vector& ControlFlowGraph::Ops() { return ops_; } -ir::Node* ControlFlowGraph::GetNodeFromVarName(const std::string& name, - ir::Node* op) const { - // in ssa-graph, different version nodes have same name, - // this function get the latest version var before target op - // It may return nullptr, such as data node. - ir::Node* found_node = nullptr; - for (auto* node : ops_) { - if (node == op) break; - for (auto& output : node->outputs) { - if (output->Name() == name) { - found_node = output; + // erase unused node + auto& nodes = var_nodes_.at(var); + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + graph->RemoveNode(node); } } } - return found_node; } } // namespace details @@ -598,5 +330,4 @@ ir::Node* ControlFlowGraph::GetNodeFromVarName(const std::string& name, REGISTER_PASS(memory_optimize_pass, paddle::framework::details::MemoryOptimizePass) - .RequireGraphAttr(paddle::framework::details::kGraphNodePool) .RequireGraphAttr(paddle::framework::details::kAllOpDescs); diff --git a/paddle/fluid/framework/details/memory_optimize_pass.h b/paddle/fluid/framework/details/memory_optimize_pass.h index b3e026e0b..593ffc10f 100644 --- a/paddle/fluid/framework/details/memory_optimize_pass.h +++ b/paddle/fluid/framework/details/memory_optimize_pass.h @@ -32,20 +32,15 @@ namespace paddle { namespace framework { namespace details { -constexpr char kAllOpDescs[] = "all_op_descs"; - -std::vector SortOpLikeDescOrder(const ir::Graph& graph); - -class ControlFlowGraph; class MemoryOptimizePass : public ir::Pass { protected: std::unique_ptr ApplyImpl( std::unique_ptr graph) const override; - - private: // fill the variable map(var_nodes) by version. void InitSSAGraphNodes() const; + + private: // update program descs void RenameVarInGraphDesc(const std::string& var, const std::string& cache_var, size_t idx) const; @@ -55,13 +50,14 @@ class MemoryOptimizePass : public ir::Pass { ir::Graph* graph) const; void SubGraphOptimize(OpDesc* op_desc) const; - // scan subblock and collect the output/input variables. - std::unordered_set GetSubBlockVars( - const std::unordered_set&) const; + // 1. scan op with subblock and collect the output/input vars. + // while, while_grad, conditional_block + // 2. scan distributed ops and collect the output/input vars + void CollectSkipVarsSet(const std::unordered_set&) const; private: // Reuse Node Pool, Owned. - mutable OrderedNodeList pool_; + mutable OrderedSet pool_; // controlflow Graph mutable std::unique_ptr cfg_; // skip set @@ -70,45 +66,6 @@ class MemoryOptimizePass : public ir::Pass { mutable std::map> var_nodes_; }; -class ControlFlowGraph { - public: - ControlFlowGraph() = default; - // For IR Graph in parallelexecutor - explicit ControlFlowGraph(const ir::Graph& graph); - - void LiveVariableAnalysis(); - - void RenameVarInCFGGraph(const std::string& old_node, - const std::string& new_node, int begin_idx); - - const std::set LiveIn(ir::Node* op) const; - const std::set LiveOut(ir::Node* op) const; - const std::set Use(ir::Node* op) const; - const std::vector Ops() const; - std::vector& Ops(); - - // for ssa-graph nodes - ir::Node* GetNodeFromVarName(const std::string& name, ir::Node* op) const; - - private: - void BuildCFGGraph(); - void ConnectNodes(); - using NodeListMap = std::unordered_map>; - using VarSetMap = std::map>; - // successors ops use the output variables. - NodeListMap successors_; - // predecessors ops generated input variables. - NodeListMap predecessors_; - // variables lived before run current op. - VarSetMap live_in_; - // variables lived after run current op. - VarSetMap live_out_; - VarSetMap uses_; // op inputs - VarSetMap defs_; // op outputs - - std::vector ops_; // op sequence by topology sort -}; - } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/memory_optimize_pass_test.cc b/paddle/fluid/framework/details/memory_optimize_pass_test.cc deleted file mode 100644 index 3d3dfa935..000000000 --- a/paddle/fluid/framework/details/memory_optimize_pass_test.cc +++ /dev/null @@ -1,417 +0,0 @@ -// Copyright (c) 2018 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/details/memory_optimize_pass.h" -#include -#include -#include -#include "glog/logging.h" -#include "gtest/gtest.h" -#include "paddle/fluid/framework/details/graph_test_base.h" -#include "paddle/fluid/framework/ir/graph.h" -#include "paddle/fluid/framework/ir/graph_helper.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/operator.h" -#include "paddle/fluid/framework/program_desc.h" - -REGISTER_OPERATOR(sum, paddle::framework::DummyOp, - paddle::framework::SumOpMaker, - paddle::framework::DummyVarTypeInference); -REGISTER_OPERATOR(assign, paddle::framework::DummyOp, - paddle::framework::AssignOpMaker, - paddle::framework::DummyVarTypeInference); -REGISTER_OPERATOR(dummy, paddle::framework::DummyOp, - paddle::framework::SumOpMaker, - paddle::framework::DummyVarTypeInference); -/* - https://en.wikipedia.org/wiki/Live_variable_analysis - Create a customed classical dependency graph, left row is the instruction - number. - 1. a = 1 - 2. b = a - 3. c = a - 4. d = b + c - 5. e = d - - a--------+ - | | - b c - | | - d--------+ - | - e - Then analysis these variable's liveness range - */ - -namespace paddle { -namespace framework { -namespace details { - -static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) { - return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && - op1->Outputs() == op2->Outputs(); -} - -inline static ProgramDesc FillProgramDesc() { - ProgramDesc prog; - prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("d")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("e")->SetType(proto::VarType::LOD_TENSOR); - { - auto* op = prog.MutableBlock(0)->AppendOp(); - op->SetType("assign"); - op->SetInput("X", {"a"}); - op->SetOutput("Out", {"b"}); - } - { - auto* op = prog.MutableBlock(0)->AppendOp(); - op->SetType("assign"); - op->SetInput("X", {"a"}); - op->SetOutput("Out", {"c"}); - } - { - auto* op = prog.MutableBlock(0)->AppendOp(); - op->SetType("sum"); - op->SetInput("X", {"b", "c"}); - op->SetOutput("Out", {"d"}); - } - { - auto* op = prog.MutableBlock(0)->AppendOp(); - op->SetType("assign"); - op->SetInput("X", {"d"}); - op->SetOutput("Out", {"e"}); - } - return prog; -} - -TEST(CFGGraph, IRGraph) { - // prepare ir graph - auto prog = FillProgramDesc(); - ir::Graph graph(prog); - const std::vector* all_op_descs = - new std::vector(prog.Block(0).AllOps()); - graph.Set(details::kAllOpDescs, all_op_descs); // take ownership - - ControlFlowGraph cfg(graph); - cfg.LiveVariableAnalysis(); - - // test assign op - ASSERT_TRUE((std::set{"a"} == cfg.LiveIn(cfg.Ops()[0]))); - ASSERT_TRUE((std::set{"a", "b"} == cfg.LiveOut(cfg.Ops()[0]))); - - // test assign op - ASSERT_TRUE((std::set{"a", "b"} == cfg.LiveIn(cfg.Ops()[1]))); - ASSERT_TRUE((std::set{"b", "c"} == cfg.LiveOut(cfg.Ops()[1]))); - - // test sum op - ASSERT_TRUE((std::set{"b", "c"} == cfg.LiveIn(cfg.Ops()[2]))); - ASSERT_TRUE((std::set{"d"} == cfg.LiveOut(cfg.Ops()[2]))); - - // test assign op - ASSERT_TRUE((std::set{"d"} == cfg.LiveIn(cfg.Ops()[3]))); - ASSERT_TRUE((std::set{} == cfg.LiveOut(cfg.Ops()[3]))); -} - -// 1. normal test -TEST(SortOpLikeDescOrder, NormalTest) { - auto prog = FillProgramDesc(); - ir::Graph graph(prog); - const std::vector* all_op_descs = - new std::vector(prog.Block(0).AllOps()); - graph.Set(details::kAllOpDescs, all_op_descs); // take ownership - - auto nodes = SortOpLikeDescOrder(graph); - auto op_descs = prog.Block(0).AllOps(); - for (size_t i = 0; i < nodes.size(); ++i) { - auto node = nodes[i]; - auto op_desc = op_descs[i]; - ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); - } -} - -// 2. remove some op_desc -TEST(SortOpLikeDescOrder, RemoveOpDesc) { - auto prog = FillProgramDesc(); - ir::Graph graph(prog); - const std::vector* all_op_descs = - new std::vector(prog.Block(0).AllOps()); - graph.Set(details::kAllOpDescs, all_op_descs); // take ownership - auto nodes = graph.Nodes(); - auto op_descs = prog.Block(0).AllOps(); - ir::Node* found_node = nullptr; - for (auto node : nodes) { - if (node->IsOp() && node->outputs.back()->Name() == "e") { - found_node = node; - break; - } - } - PADDLE_ENFORCE(found_node != nullptr); - for (auto it = op_descs.begin(); it != op_descs.end();) { - if (IsSameDesc(*it, found_node->Op())) { - it = op_descs.erase(it); - } else { - ++it; - } - } - - auto find_node_in_graph = [&](std::string s) { - ir::Node* ret = nullptr; - for (auto n : graph.Nodes()) { - if (n->Name() == s) { - ret = n; - break; - } - } - PADDLE_ENFORCE(ret != nullptr); - return ret; - }; - - ir::Node* e = find_node_in_graph("e"); - ir::Node* d = find_node_in_graph("d"); - std::remove(d->outputs.begin(), d->outputs.end(), found_node); - graph.RemoveNode(found_node); - graph.RemoveNode(e); - - // other node keeps the same order - auto remain_nodes = SortOpLikeDescOrder(graph); - for (size_t i = 0; i < remain_nodes.size(); ++i) { - auto node = remain_nodes[i]; - auto op_desc = op_descs[i]; - ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); - } -} - -// 3. add some op_desc -TEST(SortOpLikeDescOrder, AddOpDesc) { - auto prog = FillProgramDesc(); - const std::vector* all_op_descs = - new std::vector(prog.Block(0).AllOps()); - ir::Graph graph(prog); - - auto find_node_in_graph = [&](std::string s) { - ir::Node* ret = nullptr; - for (auto n : graph.Nodes()) { - if (n->Name() == s) { - ret = n; - break; - } - } - PADDLE_ENFORCE(ret != nullptr); - return ret; - }; - - // cached desc different with real one - // mimic the intermidiete pass modify the programdesc. - graph.Set(details::kAllOpDescs, all_op_descs); // take ownership - - auto op_descs = prog.Block(0).AllOps(); - - auto op = prog.MutableBlock(0)->AppendOp(); - prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); - op->SetType("sum"); - op->SetInput("X", {"b", "c"}); - op->SetOutput("Out", {"d1"}); - ir::Node* node = graph.CreateOpNode(op); - ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); - ir::Node* b = find_node_in_graph("b"); - ir::Node* c = find_node_in_graph("c"); - node->outputs.emplace_back(d1); - node->inputs.emplace_back(b); - node->inputs.emplace_back(c); - d1->inputs.emplace_back(node); - b->outputs.emplace_back(node); - c->outputs.emplace_back(node); - op_descs.insert(op_descs.begin() + 4, op); - - auto nodes = SortOpLikeDescOrder(graph); - - for (size_t i = 0; i < nodes.size(); ++i) { - auto node = nodes[i]; - auto op_desc = op_descs[i]; - ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); - } -} - -// 4. add and delete some op_desc -TEST(SortOpLikeDescOrder, AddAndDeleteOpDesc) { - auto prog = FillProgramDesc(); - ir::Graph graph(prog); - const std::vector* all_op_descs = - new std::vector(prog.Block(0).AllOps()); - graph.Set(details::kAllOpDescs, all_op_descs); // take ownership - - auto find_node_in_graph = [&](std::string s) { - ir::Node* ret = nullptr; - for (auto n : graph.Nodes()) { - if (n->Name() == s) { - ret = n; - break; - } - } - PADDLE_ENFORCE(ret != nullptr); - return ret; - }; - - // remove sum node - auto op_descs = prog.Block(0).AllOps(); - ir::Node* found_node = nullptr; - auto nodes = graph.Nodes(); - for (auto node : nodes) { - if (node->Name() == "sum") { - found_node = node; - break; - } - } - PADDLE_ENFORCE(found_node != nullptr); - for (auto it = op_descs.begin(); it != op_descs.end();) { - if (IsSameDesc(*it, found_node->Op())) { - it = op_descs.erase(it); - } else { - ++it; - } - } - { - ir::Node* d = find_node_in_graph("d"); - ir::Node* c = find_node_in_graph("c"); - ir::Node* e = find_node_in_graph("e"); - std::remove(d->outputs.begin(), d->outputs.end(), found_node); - std::remove(c->outputs.begin(), c->outputs.end(), found_node); - ir::Node* pending_op = found_node->outputs[0]->outputs[0]; - graph.RemoveNode(e); - graph.RemoveNode(pending_op); - graph.RemoveNode(found_node); - } - - // add node - auto op = prog.MutableBlock(0)->AppendOp(); - prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); - op->SetType("sum"); - op->SetInput("X", {"b", "c"}); - op->SetOutput("Out", {"d1"}); - { - ir::Node* node = graph.CreateOpNode(op); - ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); - ir::Node* b = find_node_in_graph("b"); - ir::Node* c = find_node_in_graph("c"); - node->outputs.emplace_back(d1); - node->inputs.emplace_back(b); - node->inputs.emplace_back(c); - b->outputs.emplace_back(node); - c->outputs.emplace_back(node); - } - op_descs.insert(op_descs.begin() + 2, op); - - // check the order - auto mynodes = SortOpLikeDescOrder(graph); - for (size_t i = 0; i < mynodes.size(); ++i) { - auto node = mynodes[i]; - auto op_desc = op_descs[i]; - ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); - } -} - -// 5. add and replace some op_desc inplace. -TEST(SortOpLikeDescOrder, AddAndReplaceOpDescInplace) { - auto prog = FillProgramDesc(); - ir::Graph graph(prog); - const std::vector* all_op_descs = - new std::vector(prog.Block(0).AllOps()); - graph.Set(details::kAllOpDescs, all_op_descs); // take ownership - - auto find_node_in_graph = [&](std::string s) { - ir::Node* ret = nullptr; - for (auto n : graph.Nodes()) { - if (n->Name() == s) { - ret = n; - break; - } - } - PADDLE_ENFORCE(ret != nullptr); - return ret; - }; - - auto op_descs = prog.Block(0).AllOps(); - // add node - auto op = prog.MutableBlock(0)->AppendOp(); - prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); - op->SetType("sum"); - op->SetInput("X", {"b", "c"}); - op->SetOutput("Out", {"d1"}); - { - ir::Node* node = graph.CreateOpNode(op); - ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); - ir::Node* b = find_node_in_graph("b"); - ir::Node* c = find_node_in_graph("c"); - node->outputs.emplace_back(d1); - node->inputs.emplace_back(b); - node->inputs.emplace_back(c); - d1->inputs.emplace_back(node); - b->outputs.emplace_back(node); - c->outputs.emplace_back(node); - } - - op_descs.emplace_back(op); - - // replace op_desc inplace - auto nodes = graph.Nodes(); - ir::Node* found_node = nullptr; - for (auto node : nodes) { - if (node->IsOp() && node->Op() && node->Name() == "assign") { - if (node->outputs.size() == 1 && node->outputs[0]->Name() == "e") { - found_node = node; - break; - } - } - } - { - ir::Node* d = find_node_in_graph("d"); - ir::Node* e = find_node_in_graph("e"); - std::remove(d->outputs.begin(), d->outputs.end(), found_node); - std::remove(e->inputs.begin(), e->inputs.end(), found_node); - graph.RemoveNode(found_node); - } - op_descs.erase(op_descs.begin() + 3); - - auto replace_op = prog.MutableBlock(0)->AppendOp(); - replace_op->SetType("sum"); - replace_op->SetInput("X", {"d", "d1"}); - replace_op->SetOutput("Out", {"e"}); - { - ir::Node* sum2 = graph.CreateOpNode(replace_op); - ir::Node* e = find_node_in_graph("e"); - ir::Node* d = find_node_in_graph("d"); - ir::Node* d1 = find_node_in_graph("d1"); - sum2->inputs.emplace_back(d); - sum2->inputs.emplace_back(d1); - sum2->outputs.emplace_back(e); - e->inputs.emplace_back(sum2); - d->outputs.emplace_back(sum2); - d1->outputs.emplace_back(sum2); - } - - op_descs.emplace_back(replace_op); - // compare op order - auto graph_nodes = SortOpLikeDescOrder(graph); - for (size_t i = 0; i < graph_nodes.size(); ++i) { - auto node = graph_nodes[i]; - auto op_desc = op_descs[i]; - ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); - } -} - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/sequential_execution_pass.cc b/paddle/fluid/framework/details/sequential_execution_pass.cc index cc2c8bfef..879fb29d5 100644 --- a/paddle/fluid/framework/details/sequential_execution_pass.cc +++ b/paddle/fluid/framework/details/sequential_execution_pass.cc @@ -17,6 +17,7 @@ #include #include #include +#include "paddle/fluid/framework/details/memory_optimize_helper.h" #include "paddle/fluid/framework/op_proto_maker.h" namespace paddle { diff --git a/paddle/fluid/framework/details/sequential_execution_pass.h b/paddle/fluid/framework/details/sequential_execution_pass.h index a04c08bc2..ea3034877 100644 --- a/paddle/fluid/framework/details/sequential_execution_pass.h +++ b/paddle/fluid/framework/details/sequential_execution_pass.h @@ -21,8 +21,6 @@ namespace paddle { namespace framework { namespace details { -constexpr char kAllOpDescs[] = "all_op_descs"; - class SequentialExecutionPass : public ir::Pass { protected: std::unique_ptr ApplyImpl( diff --git a/paddle/fluid/framework/inplace_op_inference.h b/paddle/fluid/framework/inplace_op_inference.h index 03ab2a2b6..a3ccf677c 100644 --- a/paddle/fluid/framework/inplace_op_inference.h +++ b/paddle/fluid/framework/inplace_op_inference.h @@ -69,7 +69,7 @@ class InplaceInToOut : public InplaceOpInference { bool TryInplaceInputOutput(const VarDesc& in, const VarDesc& out) const { return in.Name() != out.Name() && details::NodeCanReused(in) && details::NodeCanReused(out) && - details::NodeSizeInBytes(out) <= details::NodeSizeInBytes(in); + details::NodeSize(out) <= details::NodeSize(in); } }; diff --git a/paddle/fluid/framework/inplace_op_inference_test.cc b/paddle/fluid/framework/inplace_op_inference_test.cc index 121f648a5..bf9d1dcd3 100644 --- a/paddle/fluid/framework/inplace_op_inference_test.cc +++ b/paddle/fluid/framework/inplace_op_inference_test.cc @@ -179,11 +179,11 @@ TEST(InferInplace, SingleOpInplaceInToOut) { op->SetOutput("Out", {"test2_out"}); prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64}); + prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64, 128, 128}); prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_out"); - prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 128, 128}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -201,11 +201,11 @@ TEST(InferInplace, SingleGradOpInplaceInToOut) { op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"}); prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16, 1024, 1024}); prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_out"); - prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -233,12 +233,12 @@ TEST(InferInplace, MultiOutInplaceInToOut) { prog.MutableBlock(0)->Var("o0"); prog.MutableBlock(0)->Var("y0"); prog.MutableBlock(0)->Var("z0"); - prog.MutableBlock(0)->Var("a0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("b0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("c0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("o0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("y0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("z0")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -267,15 +267,16 @@ TEST(InferInplace, MultiGradInplaceInToOut) { prog.MutableBlock(0)->Var("o0"); prog.MutableBlock(0)->Var("y0"); prog.MutableBlock(0)->Var("z0"); - prog.MutableBlock(0)->Var("a0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("b0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("c0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("o0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("y0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("z0")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); + EXPECT_EQ(in_to_outs.size(), 3ul); std::unordered_map expects = { {"o0", "a0"}, {"y0", "b0"}, {"z0", "c0"}, diff --git a/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc b/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc index 3b738aa15..5bdc0c5fa 100644 --- a/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc +++ b/paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc @@ -38,9 +38,13 @@ std::unique_ptr IdentityScaleOpCleanPass::ApplyImpl( ->assert_is_op("scale") ->assert_op_attr("scale", 1.) ->assert_op_attr("bias", 0.); - auto scale_out = detector.mutable_pattern() - ->NewNode("scale_out") - ->assert_is_op_output("scale"); + auto scale_out = + detector.mutable_pattern() + ->NewNode("scale_out") + ->assert_is_op_output("scale") + // scale's output var should has only one consumer, or it can't be + // removed. + ->assert_more([](Node* x) { return x->outputs.size() == 1UL; }); pre_op->LinksTo({scale_in}); scale_op->LinksFrom({scale_in}).LinksTo({scale_out}); diff --git a/paddle/fluid/framework/ir/infer_clean_graph_pass.cc b/paddle/fluid/framework/ir/infer_clean_graph_pass.cc index 7713ed1ea..6607c026a 100644 --- a/paddle/fluid/framework/ir/infer_clean_graph_pass.cc +++ b/paddle/fluid/framework/ir/infer_clean_graph_pass.cc @@ -37,6 +37,7 @@ class InferCleanGraphPass : public FusePassBase { std::unordered_set invalid_nodes; int valid_op = 0; for (auto* node : graph->Nodes()) { + PADDLE_ENFORCE_NOT_NULL(node); if (is_valid_node(node)) { invalid_nodes.insert(node); } else if (node->IsOp()) { diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index f61c9e3a9..ff7ef0cce 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -171,14 +171,6 @@ std::unique_ptr ParallelExecutorPrivate::PrepareGCAndRefCnts( eager_deletion_pass->SetNotOwned(details::kAllPlaces, &places_); graph = eager_deletion_pass->Apply(std::move(graph)); VLOG(10) << "EagerDeletionPass Applied"; - - if (build_strategy_.memory_early_delete_) { - auto early_delete_pass = - ir::PassRegistry::Instance().Get("memory_early_delete_pass"); - early_delete_pass->SetNotOwned(details::kGarbageCollector, &gcs_); - graph = early_delete_pass->Apply(std::move(graph)); - } - VLOG(10) << "MemoryEarlyDeletePass Applied."; } return graph; @@ -288,6 +280,8 @@ ParallelExecutor::ParallelExecutor( graphs.push_back(std::move(graph)); #endif auto max_memory_size = GetEagerDeletionThreshold(); + VLOG(10) << "Eager Deletion Threshold " + << static_cast(max_memory_size) / (1 << 30); if (max_memory_size >= 0) { for (size_t i = 0; i < graphs.size(); ++i) { graphs[i] = member_->PrepareGCAndRefCnts( @@ -506,6 +500,5 @@ ParallelExecutor::~ParallelExecutor() { } // namespace framework } // namespace paddle -USE_PASS(memory_early_delete_pass); USE_PASS(reference_count_pass); USE_PASS(eager_deletion_pass); diff --git a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc index a64f85ee9..96befe7f8 100644 --- a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc +++ b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc @@ -460,77 +460,6 @@ inline bool CheckNodeIndegreeEquals(const Node &node, size_t n) { return node.inputs.size() == n; } -NodesTSIterator::NodesTSIterator(const std::vector &source) { - PADDLE_ENFORCE(!source.empty(), - "Start points of topological sorting should not be empty!"); - // CHECK all the inputs' in-degree is 0 - for (auto *node : source) { - PADDLE_ENFORCE(CheckNodeIndegreeEquals(*node, 0)); - } - - std::unordered_set visited; - std::unordered_set to_visit{source.begin(), source.end()}; - - std::vector inlink_visited; - while (!to_visit.empty()) { - std::vector queue(to_visit.begin(), to_visit.end()); - for (auto *p : queue) { - if (Agent(p).deleted()) { - visited.insert(p); - to_visit.erase(p); - } - - inlink_visited.clear(); - - std::copy_if(p->inputs.begin(), p->inputs.end(), - std::back_inserter(inlink_visited), - [&](Node *x) -> bool { return visited.count(x) != 0; }); - - if (inlink_visited.size() == p->inputs.size()) { - sorted_.push_back(p); - for (auto *_ : p->outputs) { - if (!visited.count(_)) { - to_visit.insert(_); - } - } - - to_visit.erase(p); - visited.insert(p); - } - } - } -} - -NodesTSIterator::NodesTSIterator(const NodesTSIterator &other) - : sorted_(other.sorted_), cursor_(other.cursor_) {} - -Node &NodesTSIterator::operator*() { - PADDLE_ENFORCE_LT(cursor_, sorted_.size()); - return *sorted_[cursor_]; -} - -NodesTSIterator &NodesTSIterator::operator++() { - if (++cursor_ >= sorted_.size()) { - sorted_.clear(); - cursor_ = 0; - } - return *this; -} -NodesTSIterator &NodesTSIterator::operator=(const NodesTSIterator &other) { - cursor_ = other.cursor_; - sorted_ = other.sorted_; - return *this; -} - -bool NodesTSIterator::operator==(const NodesTSIterator &other) { - return sorted_ == other.sorted_ && cursor_ == other.cursor_; -} - -Node *NodesTSIterator::operator->() { - PADDLE_ENFORCE_LT(cursor_, sorted_.size()); - return sorted_[cursor_]; -} - } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h index ea88edd04..5d11c217b 100644 --- a/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h +++ b/paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h @@ -30,6 +30,7 @@ namespace inference { namespace analysis { using framework::ir::Graph; +using framework::ir::NodesTSIterator; const char kIsFunctionNode[] = "__is_function_node__"; const char kFunctionNodeSubGraph[] = "__function_node_sub_graph__"; @@ -132,32 +133,6 @@ struct Agent { framework::ir::Node *x_; }; -// Topological sorting iterator on nodes. -struct NodesTSIterator - : public std::iterator { - NodesTSIterator() = default; - explicit NodesTSIterator(const std::vector &source); - NodesTSIterator(NodesTSIterator &&other) - : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) { - other.cursor_ = 0; - } - NodesTSIterator(const NodesTSIterator &other); - - framework::ir::Node &operator*(); - NodesTSIterator &operator++(); - // TODO(Superjomn) current implementation just compare the first - // element, need to compare the graph and all the elements in the queue and - // set. - NodesTSIterator &operator=(const NodesTSIterator &other); - bool operator==(const NodesTSIterator &other); - bool operator!=(const NodesTSIterator &other) { return !(*this == other); } - framework::ir::Node *operator->(); - - private: - std::vector sorted_; - size_t cursor_{0}; -}; - // The nodes those have no input will be treated as start points. static std::vector ExtractStartPoints(const Graph &g) { std::vector result; diff --git a/paddle/fluid/inference/api/paddle_api.h b/paddle/fluid/inference/api/paddle_api.h index 8ac8bc529..f90a74b91 100644 --- a/paddle/fluid/inference/api/paddle_api.h +++ b/paddle/fluid/inference/api/paddle_api.h @@ -16,6 +16,12 @@ /*! \file paddle_api.h */ +/*! \mainpage Paddle Inference APIs + * \section intro_sec Introduction + * The Paddle inference library aims to offer an high performance inference SDK + * for Paddle users. + */ + #include #include #include @@ -34,26 +40,49 @@ enum PaddleDType { }; /** - *\brief Memory menager for PaddleTensor. + * \brief Memory manager for `PaddleTensor`. * - *The PaddleBuf holds a buffer for data input or output. The memory can be - *allocated by user or by PaddleBuf itself, but in any case, the PaddleBuf - *should be reused for better performance. + * The PaddleBuf holds a buffer for data input or output. The memory can be + * allocated by user or by PaddleBuf itself, but in any case, the PaddleBuf + * should be reused for better performance. * - *For user allocated memory, the following API can be used: - *- PaddleBuf(void* data, size_t length) to set an external memory by - *specifying - * the memory address and length. - *- Reset(void* data, size_t length) to reset the PaddleBuf with an external + * For user allocated memory, the following API can be used: + * - PaddleBuf(void* data, size_t length) to set an external memory by + * specifying the memory address and length. + * - Reset(void* data, size_t length) to reset the PaddleBuf with an external *memory. - *ATTENTION, for user allocated memory, deallocation should be done by users + * ATTENTION, for user allocated memory, deallocation should be done by users *externally after the program finished. The PaddleBuf won't do any allocation *or deallocation. * - *To have the PaddleBuf allocate and manage the memory: - *- PaddleBuf(size_t length) will allocate a memory of size `length`. - *- Resize(size_t length) resize the memory to no less than `length`, ATTENTION + * To have the PaddleBuf allocate and manage the memory: + * - PaddleBuf(size_t length) will allocate a memory of size `length`. + * - Resize(size_t length) resize the memory to no less than `length`, ATTENTION * if the allocated memory is larger than `length`, nothing will done. + * + * Usage: + * + * Let PaddleBuf manage the memory internally. + * \code{cpp} + * const int num_elements = 128; + * PaddleBuf buf(num_elements * sizeof(float)); + * \endcode + * + * Or + * \code{cpp} + * PaddleBuf buf; + * buf.Resize(num_elements * sizeof(float)); + * \endcode + * Works the exactly the same. + * + * One can also make the `PaddleBuf` use the external memory. + * \code{cpp} + * PaddleBuf buf; + * void* external_memory = new float[num_elements]; + * buf.Reset(external_memory, num_elements*sizeof(float)); + * ... + * delete[] external_memory; // manage the memory lifetime outside. + * \endcode */ class PaddleBuf { public: @@ -78,7 +107,7 @@ class PaddleBuf { /** Tell whether the buffer is empty. */ bool empty() const { return length_ == 0; } - /** Get the memory address. + /** Get the data's memory address. */ void* data() const { return data_; } /** Get the memory length. @@ -110,7 +139,8 @@ struct PaddleTensor { }; enum class PaddlePlace { kUNK = -1, kCPU, kGPU }; -/** Tensor without copy, currently only supports AnalysisPredictor. + +/** Tensor without copy, currently only supports `AnalysisPredictor`. */ class ZeroCopyTensor { public: @@ -269,9 +299,11 @@ struct NativeConfig : public PaddlePredictor::Config { * * Usage: * + * \code{.cpp} * NativeConfig config; * ... // change the configs. * auto native_predictor = CreatePaddlePredictor(config); + * \endcode * * FOR EXTENSION DEVELOPER: * Different predictors are designated by config type. Similar configs can be diff --git a/paddle/fluid/inference/api/paddle_pass_builder.cc b/paddle/fluid/inference/api/paddle_pass_builder.cc index 039389a4c..f9c13c2fa 100644 --- a/paddle/fluid/inference/api/paddle_pass_builder.cc +++ b/paddle/fluid/inference/api/paddle_pass_builder.cc @@ -66,8 +66,54 @@ void GpuPassStrategy::EnableMKLDNN() { LOG(ERROR) << "GPU not support MKLDNN yet"; } +GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) { + passes_.assign({ + "infer_clean_graph_pass", // + "identity_scale_op_clean_pass", // + "conv_affine_channel_fuse_pass", // + "conv_eltwiseadd_affine_channel_fuse_pass", // + "conv_bn_fuse_pass", // +#if CUDNN_VERSION >= 7100 // To run conv_fusion, the version of cudnn must be + // guaranteed at least v7 + "conv_elementwise_add_act_fuse_pass", // + "conv_elementwise_add2_act_fuse_pass", // + "conv_elementwise_add_fuse_pass", // +#endif + }); + + for (int i = 6; i >= 3; i--) { + passes_.push_back("transpose_flatten" + std::to_string(i) + + "_concat_fuse_pass"); + } + use_gpu_ = true; +} + void PaddlePassBuilder::AppendAnalysisPass(const std::string &pass) { analysis_passes_.push_back(pass); } +CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) { + // NOTE the large fusions should be located in the front, so that they will + // not be damaged by smaller ones. + passes_.assign({ + "infer_clean_graph_pass", // + "attention_lstm_fuse_pass", // + "seqpool_concat_fuse_pass", // + "seqconv_eltadd_relu_fuse_pass", // + // "embedding_fc_lstm_fuse_pass", // + "fc_lstm_fuse_pass", // + "mul_lstm_fuse_pass", // + "fc_gru_fuse_pass", // + "mul_gru_fuse_pass", // + "seq_concat_fc_fuse_pass", // + "fc_fuse_pass", // + "repeated_fc_relu_fuse_pass", // + "squared_mat_sub_fuse_pass", // + "conv_bn_fuse_pass", // + "conv_eltwiseadd_bn_fuse_pass", // + "is_test_pass", // + "identity_scale_op_clean_pass", // + }); + use_gpu_ = false; +} } // namespace paddle diff --git a/paddle/fluid/inference/api/paddle_pass_builder.h b/paddle/fluid/inference/api/paddle_pass_builder.h index aa353f12c..2524d89fc 100644 --- a/paddle/fluid/inference/api/paddle_pass_builder.h +++ b/paddle/fluid/inference/api/paddle_pass_builder.h @@ -97,30 +97,7 @@ class PassStrategy : public PaddlePassBuilder { */ class CpuPassStrategy : public PassStrategy { public: - CpuPassStrategy() : PassStrategy({}) { - // NOTE the large fusions should be located in the front, so that they will - // not be damaged by smaller ones. - passes_.assign({ - "infer_clean_graph_pass", // - "attention_lstm_fuse_pass", // - "seqpool_concat_fuse_pass", // - "seqconv_eltadd_relu_fuse_pass", // - // "embedding_fc_lstm_fuse_pass", // - "fc_lstm_fuse_pass", // - "mul_lstm_fuse_pass", // - "fc_gru_fuse_pass", // - "mul_gru_fuse_pass", // - "seq_concat_fc_fuse_pass", // - "fc_fuse_pass", // - "repeated_fc_relu_fuse_pass", // - "squared_mat_sub_fuse_pass", // - "conv_bn_fuse_pass", // - "conv_eltwiseadd_bn_fuse_pass", // - "is_test_pass", // - "identity_scale_op_clean_pass", // - }); - use_gpu_ = false; - } + CpuPassStrategy(); explicit CpuPassStrategy(const CpuPassStrategy &other) : PassStrategy(other.AllPasses()) {} @@ -153,27 +130,7 @@ class CpuPassStrategy : public PassStrategy { */ class GpuPassStrategy : public PassStrategy { public: - GpuPassStrategy() : PassStrategy({}) { - passes_.assign({ - "infer_clean_graph_pass", // - "identity_scale_op_clean_pass", // - "conv_affine_channel_fuse_pass", // - "conv_eltwiseadd_affine_channel_fuse_pass", // - "conv_bn_fuse_pass", // -#if CUDNN_VERSION >= 7100 // To run conv_fusion, the version of cudnn must be - // guaranteed at least v7 - "conv_elementwise_add_act_fuse_pass", // - "conv_elementwise_add2_act_fuse_pass", // - "conv_elementwise_add_fuse_pass", // -#endif - }); - - for (int i = 6; i >= 3; i--) { - passes_.push_back("transpose_flatten" + std::to_string(i) + - "_concat_fuse_pass"); - } - use_gpu_ = true; - } + GpuPassStrategy(); explicit GpuPassStrategy(const GpuPassStrategy &other) : PassStrategy(other.AllPasses()) { diff --git a/paddle/fluid/operators/controlflow/compare_op.cc b/paddle/fluid/operators/controlflow/compare_op.cc index 688457d4a..5d3f9b43f 100644 --- a/paddle/fluid/operators/controlflow/compare_op.cc +++ b/paddle/fluid/operators/controlflow/compare_op.cc @@ -51,6 +51,11 @@ class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { comment.type)); AddInput("Y", string::Sprintf("the right hand operand of %s operator", comment.type)); + AddAttr( + "axis", + "The start dimension index for broadcasting Y onto X. [default -1]") + .SetDefault(-1) + .EqualGreaterThan(-1); AddAttr("force_cpu", "Force fill output variable to cpu " "memory. Otherwise, fill output variable to the running " @@ -64,11 +69,6 @@ N-dim tensor. X and Y could be any type. The each element of the Out tensor is calculated by $%s$ )DOC", comment.equation)); - AddAttr( - "axis", - "The start dimension index for broadcasting Y onto X. [default -1]") - .SetDefault(-1) - .EqualGreaterThan(-1); } }; diff --git a/paddle/fluid/operators/detection/density_prior_box_op.h b/paddle/fluid/operators/detection/density_prior_box_op.h index 3591681fc..42137215e 100644 --- a/paddle/fluid/operators/detection/density_prior_box_op.h +++ b/paddle/fluid/operators/detection/density_prior_box_op.h @@ -72,7 +72,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { #ifdef PADDLE_WITH_MKLML #pragma omp parallel for #endif - for (int i = 0; i < fixed_ratios.size(); i++) { + for (size_t i = 0; i < fixed_ratios.size(); i++) { sqrt_fixed_ratios.push_back(sqrt(fixed_ratios[i])); } @@ -115,11 +115,10 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { } } if (clip) { - platform::Transform trans; - ClipFunctor clip_func; - trans(ctx.template device_context(), - boxes->data(), boxes->data() + boxes->numel(), - boxes->data(), clip_func); + T* dt = boxes->data(); + std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T { + return std::min(std::max(v, 0.), 1.); + }); } framework::Tensor var_t; var_t.mutable_data( @@ -141,7 +140,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { #pragma omp parallel for collapse(2) #endif for (int i = 0; i < box_num; ++i) { - for (int j = 0; j < variances.size(); ++j) { + for (size_t j = 0; j < variances.size(); ++j) { e_vars(i, j) = variances[j]; } } diff --git a/paddle/fluid/operators/detection/prior_box_op.h b/paddle/fluid/operators/detection/prior_box_op.h index 4e226abbb..f84405664 100644 --- a/paddle/fluid/operators/detection/prior_box_op.h +++ b/paddle/fluid/operators/detection/prior_box_op.h @@ -46,13 +46,6 @@ inline void ExpandAspectRatios(const std::vector& input_aspect_ratior, } } -template -struct ClipFunctor { - HOSTDEVICE inline T operator()(T in) const { - return std::min(std::max(in, 0.), 1.); - } -}; - template class PriorBoxOpKernel : public framework::OpKernel { public: @@ -101,31 +94,30 @@ class PriorBoxOpKernel : public framework::OpKernel { boxes->mutable_data(ctx.GetPlace()); vars->mutable_data(ctx.GetPlace()); - auto e_boxes = framework::EigenTensor::From(*boxes); + T* b_t = boxes->data(); for (int h = 0; h < feature_height; ++h) { for (int w = 0; w < feature_width; ++w) { T center_x = (w + offset) * step_width; T center_y = (h + offset) * step_height; T box_width, box_height; - int idx = 0; for (size_t s = 0; s < min_sizes.size(); ++s) { auto min_size = min_sizes[s]; if (min_max_aspect_ratios_order) { box_width = box_height = min_size / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // square prior with size sqrt(minSize * maxSize) box_width = box_height = sqrt(min_size * max_size) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } // priors with different aspect ratios for (size_t r = 0; r < aspect_ratios.size(); ++r) { @@ -135,11 +127,11 @@ class PriorBoxOpKernel : public framework::OpKernel { } box_width = min_size * sqrt(ar) / 2.; box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } } else { // priors with different aspect ratios @@ -147,21 +139,21 @@ class PriorBoxOpKernel : public framework::OpKernel { float ar = aspect_ratios[r]; box_width = min_size * sqrt(ar) / 2.; box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // square prior with size sqrt(minSize * maxSize) box_width = box_height = sqrt(min_size * max_size) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } } } @@ -169,11 +161,10 @@ class PriorBoxOpKernel : public framework::OpKernel { } if (clip) { - platform::Transform trans; - ClipFunctor clip_func; - trans(ctx.template device_context(), - boxes->data(), boxes->data() + boxes->numel(), - boxes->data(), clip_func); + T* dt = boxes->data(); + std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T { + return std::min(std::max(v, 0.), 1.); + }); } framework::Tensor var_t; diff --git a/paddle/fluid/operators/detection/yolov3_loss_op.cc b/paddle/fluid/operators/detection/yolov3_loss_op.cc index 2a69ad4b5..ab01bdf7c 100644 --- a/paddle/fluid/operators/detection/yolov3_loss_op.cc +++ b/paddle/fluid/operators/detection/yolov3_loss_op.cc @@ -144,34 +144,40 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker { "The ignore threshold to ignore confidence loss.") .SetDefault(0.7); AddComment(R"DOC( - This operator generate yolov3 loss by given predict result and ground + This operator generates yolov3 loss based on given predict result and ground truth boxes. The output of previous network is in shape [N, C, H, W], while H and W - should be the same, specify the grid size, each grid point predict given - number boxes, this given number is specified by anchors, it should be - half anchors length, which following will be represented as S. In the - second dimention(the channel dimention), C should be S * (class_num + 5), - class_num is the box categoriy number of source dataset(such as coco), - so in the second dimention, stores 4 box location coordinates x, y, w, h - and confidence score of the box and class one-hot key of each anchor box. + should be the same, H and W specify the grid size, each grid point predict + given number boxes, this given number, which following will be represented as S, + is specified by the number of anchors, In the second dimension(the channel + dimension), C should be equal to S * (class_num + 5), class_num is the object + category number of source dataset(such as 80 in coco dataset), so in the + second(channel) dimension, apart from 4 box location coordinates x, y, w, h, + also includes confidence score of the box and class one-hot key of each anchor box. - While the 4 location coordinates if $$tx, ty, tw, th$$, the box predictions - correspnd to: + Assume the 4 location coordinates are :math:`t_x, t_y, t_w, t_h`, the box predictions + should be as follows: $$ - b_x = \sigma(t_x) + c_x - b_y = \sigma(t_y) + c_y + b_x = \\sigma(t_x) + c_x + $$ + $$ + b_y = \\sigma(t_y) + c_y + $$ + $$ b_w = p_w e^{t_w} + $$ + $$ b_h = p_h e^{t_h} $$ - While $$c_x, c_y$$ is the left top corner of current grid and $$p_w, p_h$$ - is specified by anchors. + In the equation above, :math:`c_x, c_y` is the left top corner of current grid + and :math:`p_w, p_h` is specified by anchors. As for confidence score, it is the logistic regression value of IoU between anchor boxes and ground truth boxes, the score of the anchor box which has - the max IoU should be 1, and if the anchor box has IoU bigger then ignore + the max IoU should be 1, and if the anchor box has IoU bigger than ignore thresh, the confidence score loss of this anchor box will be ignored. Therefore, the yolov3 loss consist of three major parts, box location loss, @@ -186,13 +192,13 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker { In order to trade off box coordinate losses between big boxes and small boxes, box coordinate losses will be mutiplied by scale weight, which is - calculated as follow. + calculated as follows. $$ weight_{box} = 2.0 - t_w * t_h $$ - Final loss will be represented as follow. + Final loss will be represented as follows. $$ loss = (loss_{xy} + loss_{wh}) * weight_{box} diff --git a/paddle/fluid/operators/elementwise/elementwise_op.h b/paddle/fluid/operators/elementwise/elementwise_op.h index d04bb8f33..91e441526 100644 --- a/paddle/fluid/operators/elementwise/elementwise_op.h +++ b/paddle/fluid/operators/elementwise/elementwise_op.h @@ -264,6 +264,23 @@ class ElementwiseOpInplace : public framework::InplaceInToOut { } }; +class ElementwiseGradOpInplace : public framework::InplaceInToOut { + public: + using framework::InplaceInToOut::InplaceInToOut; + + protected: + std::unordered_map Apply( + const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + std::unordered_map ret; + if (block->HasVar(framework::GradVarName("X")) && + block->HasVar(framework::GradVarName("Out"))) { + ret[framework::GradVarName("Out")] = framework::GradVarName("X"); + } + return ret; + } +}; + } // namespace operators } // namespace paddle @@ -316,4 +333,5 @@ class ElementwiseOpInplace : public framework::InplaceInToOut { op_type##GradMaker, \ ::paddle::operators::ElementwiseOpInplace); \ REGISTER_OPERATOR(op_type##_grad, \ - ::paddle::operators::ElementwiseOpExplicitGrad) + ::paddle::operators::ElementwiseOpExplicitGrad, \ + ::paddle::operators::ElementwiseGradOpInplace) diff --git a/paddle/fluid/operators/lstm_op.h b/paddle/fluid/operators/lstm_op.h index 7d62d2d02..3f110024b 100644 --- a/paddle/fluid/operators/lstm_op.h +++ b/paddle/fluid/operators/lstm_op.h @@ -311,6 +311,10 @@ class LSTMGradKernel : public framework::OpKernel { lstm_grad.prev_state_grad = c0_g ? ordered_c0_g.data() : nullptr; } + // lstm_value.output_value not used in bp, set to nullptr + // lstm_grad.state_active_grad not used in bp, set to nullptr + lstm_value.output_value = nullptr; + lstm_grad.state_active_grad = nullptr; int cur_batch_size = bend - bstart; math::LstmUnitGradFunctor::compute( device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size, diff --git a/paddle/fluid/operators/lstmp_op.h b/paddle/fluid/operators/lstmp_op.h index 370dd04d1..1f11e57dc 100644 --- a/paddle/fluid/operators/lstmp_op.h +++ b/paddle/fluid/operators/lstmp_op.h @@ -405,6 +405,11 @@ class LSTMPGradKernel : public framework::OpKernel { } int cur_batch_size = bend - bstart; + // lstmp_value.output_value not used in bp, set to null + // lstmp_grad.state_active_grad not used in bp, set to null + lstmp_value.output_value = nullptr; + lstmp_grad.state_active_grad = nullptr; + math::LstmUnitGradFunctor::compute( device_ctx, lstmp_value, lstmp_grad, frame_size, cur_batch_size, gate_act, cell_act, cand_act); diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index 5399ae556..0a0ece162 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -168,9 +168,10 @@ void Pool2dOpMaker::Make() { "be ignored."); // TODO(Chengduo): Add checker. // (Currently, // TypedAttrChecker don't support vector type.) - AddAttr("global_pooling", - "(bool, default false) Whether to use the global pooling. " - "If global_pooling = true, ksize and paddings will be ignored.") + AddAttr( + "global_pooling", + "(bool, default false) Whether to use the global pooling. " + "If global_pooling = true, kernel size and paddings will be ignored.") .SetDefault(false); AddAttr>("strides", "(vector, default {1, 1}), strides(height, " @@ -182,7 +183,7 @@ void Pool2dOpMaker::Make() { "paddings", "(vector, default {0,0}), paddings(height, width) of pooling " "operator." - "If global_pooling = true, paddings and ksize will be ignored.") + "If global_pooling = true, paddings and kernel size will be ignored.") .SetDefault({0, 0}); AddAttr( "exclusive", @@ -204,7 +205,7 @@ void Pool2dOpMaker::Make() { .SetDefault(false); AddAttr( "ceil_mode", - "(bool, default false) Wether to use the ceil function to calculate " + "(bool, default false) Whether to use the ceil function to calculate " "output height and width. False is the default. If it is set to False, " "the floor function will be used.") .SetDefault(false); @@ -259,31 +260,40 @@ Example: W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1 $$ - For exclusive = true: + For exclusive = false: $$ hstart = i * strides[0] - paddings[0] + $$ + $$ hend = hstart + ksize[0] + $$ + $$ wstart = j * strides[1] - paddings[1] + $$ + $$ wend = wstart + ksize[1] + $$ + $$ Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{ksize[0] * ksize[1]} $$ - For exclusive = false: + + For exclusive = true: $$ hstart = max(0, i * strides[0] - paddings[0]) + $$ + $$ hend = min(H, hstart + ksize[0]) + $$ + $$ wstart = max(0, j * strides[1] - paddings[1]) + $$ + $$ wend = min(W, wstart + ksize[1]) + $$ + $$ Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} $$ - For adaptive = true: - $$ - hstart = floor(i * H_{in} / H_{out}) - hend = ceil((i + 1) * H_{in} / H_{out}) - wstart = floor(j * W_{in} / W_{out}) - wend = ceil((j + 1) * W_{in} / W_{out}) - Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} - $$ )DOC"); } @@ -324,7 +334,7 @@ void Pool3dOpMaker::Make() { AddAttr( "global_pooling", "(bool, default false) Whether to use the global pooling. " - "If global_pooling = true, ksize and paddings wille be ignored.") + "If global_pooling = true, kernel size and paddings will be ignored.") .SetDefault(false); AddAttr>( "strides", @@ -359,7 +369,7 @@ void Pool3dOpMaker::Make() { .SetDefault(false); AddAttr( "ceil_mode", - "(bool, default false) Wether to use the ceil function to calculate " + "(bool, default false) Whether to use the ceil function to calculate " "output height and width. False is the default. If it is set to False, " "the floor function will be used.") .SetDefault(false); @@ -392,48 +402,68 @@ Example: Output: Out shape: $(N, C, D_{out}, H_{out}, W_{out})$ For ceil_mode = false: - $$ - D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\ - H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\ - W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1 - $$ + $$ + D_{out} = \\frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 + $$ + $$ + H_{out} = \\frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[2]} + 1 + $$ + $$ + W_{out} = \\frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1 + $$ For ceil_mode = true: - $$ - D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0] + strides[0] -1)}{strides[0]} + 1 \\ - H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1] + strides[1] -1)}{strides[1]} + 1 \\ - W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2] + strides[2] -1)}{strides[2]} + 1 - $$ - For exclusive = true: - $$ - dstart = i * strides[0] - paddings[0] - dend = dstart + ksize[0] - hstart = j * strides[1] - paddings[1] - hend = hstart + ksize[1] - wstart = k * strides[2] - paddings[2] - wend = wstart + ksize[2] - Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{ksize[0] * ksize[1] * ksize[2]} - $$ + $$ + D_{out} = \\frac{(D_{in} - ksize[0] + 2 * paddings[0] + strides[0] -1)}{strides[0]} + 1 + $$ + $$ + H_{out} = \\frac{(H_{in} - ksize[1] + 2 * paddings[1] + strides[1] -1)}{strides[1]} + 1 + $$ + $$ + W_{out} = \\frac{(W_{in} - ksize[2] + 2 * paddings[2] + strides[2] -1)}{strides[2]} + 1 + $$ + For exclusive = false: - $$ - dstart = max(0, i * strides[0] - paddings[0]) - dend = min(D, dstart + ksize[0]) - hstart = max(0, j * strides[1] - paddings[1]) - hend = min(H, hstart + ksize[1]) - wstart = max(0, k * strides[2] - paddings[2]) - wend = min(W, wstart + ksize[2]) - Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)} - $$ - - For adaptive = true: - $$ - dstart = floor(i * D_{in} / D_{out}) - dend = ceil((i + 1) * D_{in} / D_{out}) - hstart = floor(j * H_{in} / H_{out}) - hend = ceil((j + 1) * H_{in} / H_{out}) - wstart = floor(k * W_{in} / W_{out}) - wend = ceil((k + 1) * W_{in} / W_{out}) - Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)} - $$ + $$ + dstart = i * strides[0] - paddings[0] + $$ + $$ + dend = dstart + ksize[0] + $$ + $$ + hstart = j * strides[1] - paddings[1] + $$ + $$ + hend = hstart + ksize[1] + $$ + $$ + wstart = k * strides[2] - paddings[2] + $$ + $$ + wend = wstart + ksize[2] + $$ + $$ + Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{ksize[0] * ksize[1] * ksize[2]} + $$ + + For exclusive = true: + $$ + dstart = max(0, i * strides[0] - paddings[0]) + $$ + $$ + dend = min(D, dstart + ksize[0]) + $$ + $$ + hend = min(H, hstart + ksize[1]) + $$ + $$ + wstart = max(0, k * strides[2] - paddings[2]) + $$ + $$ + wend = min(W, wstart + ksize[2]) + $$ + $$ + Output(i ,j, k) = \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)} + $$ )DOC"); } diff --git a/paddle/fluid/operators/random_crop_op.h b/paddle/fluid/operators/random_crop_op.h index d68ba9d66..ee034b270 100644 --- a/paddle/fluid/operators/random_crop_op.h +++ b/paddle/fluid/operators/random_crop_op.h @@ -121,7 +121,7 @@ struct RandomCropFunctor { HOSTDEVICE void operator()(size_t ins_idx) { typename Random::Engine engine(seed_); engine.discard(ins_idx * (rank_ - num_batchsize_dims_)); - size_t offsets[9]; + size_t offsets[9] = {}; for (int i = num_batchsize_dims_; i < rank_; ++i) { typename Random::template UniformIntDist dist( 0, x_dims_[i] - out_dims_[i]); diff --git a/paddle/fluid/operators/slice_op.cu b/paddle/fluid/operators/slice_op.cu index 5efecb78d..24a564f9e 100644 --- a/paddle/fluid/operators/slice_op.cu +++ b/paddle/fluid/operators/slice_op.cu @@ -12,18 +12,138 @@ 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 "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/slice_op.h" +#include "paddle/fluid/platform/cuda_device_function.h" +#include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/float16.h" + +namespace paddle { +namespace operators { + +using platform::PADDLE_CUDA_NUM_THREADS; + +template +__global__ void Padding(const paddle::platform::float16* d_out, + const int* out_dims, const int* in_dims, + const int* offsets, int64_t n, + paddle::platform::float16* d_in) { + int64_t out_idx = threadIdx.x + blockDim.x * blockIdx.x; + if (out_idx < n) { + int64_t out_idx_tmp = out_idx; + int coords[D] = {0}; + for (int i = D - 1; i >= 0; --i) { + coords[i] = out_idx_tmp % out_dims[i]; + out_idx_tmp /= out_dims[i]; + coords[i] += offsets[i]; + } + + int64_t in_idx = 0; + for (int i = 0; i < D; ++i) { + in_idx = in_idx * in_dims[i] + coords[i]; + } + + d_in[in_idx] = d_out[out_idx]; + } +} + +template <> +class SliceGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* d_out = ctx.Input(framework::GradVarName("Out")); + auto* d_in = ctx.Output(framework::GradVarName("Input")); + d_in->mutable_data(ctx.GetPlace()); + + auto out_dims = d_out->dims(); + auto in_dims = d_in->dims(); + int rank = out_dims.size(); + std::vector offsets(rank, 0); + auto axes = ctx.Attr>("axes"); + auto starts = ctx.Attr>("starts"); + + for (size_t i = 0; i < starts.size(); ++i) { + if (starts[i] < 0) { + starts[i] += in_dims[axes[i]]; + } + offsets[axes[i]] = std::max(starts[i], 0); + } + + math::SetConstant + set_zero; + auto& dev_ctx = + ctx.template device_context(); + set_zero(dev_ctx, d_in, static_cast(0)); + + int64_t numel = d_out->numel(); + dim3 blocks((numel - 1) / PADDLE_CUDA_NUM_THREADS + 1); + dim3 threads(PADDLE_CUDA_NUM_THREADS); + auto stream = ctx.cuda_device_context().stream(); + + auto out_shape = framework::vectorize2int(out_dims); + thrust::device_vector out_dims_vec(out_shape.begin(), out_shape.end()); + auto in_shape = framework::vectorize2int(in_dims); + thrust::device_vector in_dims_vec(in_shape.begin(), in_shape.end()); + thrust::device_vector offsets_vec(offsets.begin(), offsets.end()); + const int* out_dims_ptr = thrust::raw_pointer_cast(out_dims_vec.data()); + const int* in_dims_ptr = thrust::raw_pointer_cast(in_dims_vec.data()); + const int* offsets_ptr = thrust::raw_pointer_cast(offsets_vec.data()); + + switch (rank) { + case 1: + Padding<1><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 2: + Padding<2><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 3: + Padding<3><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 4: + Padding<4><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 5: + Padding<5><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + case 6: + Padding<6><<>>( + d_out->data(), out_dims_ptr, in_dims_ptr, + offsets_ptr, numel, d_in->data()); + break; + } + } +}; + +} // namespace operators +} // namespace paddle namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( slice, ops::SliceKernel, ops::SliceKernel, ops::SliceKernel, - ops::SliceKernel); + ops::SliceKernel, + ops::SliceKernel); REGISTER_OP_CUDA_KERNEL( slice_grad, ops::SliceGradKernel, ops::SliceGradKernel, ops::SliceGradKernel, - ops::SliceGradKernel); + ops::SliceGradKernel, + ops::SliceGradKernel); diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu index 52b8dcc68..89aaac4cb 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu @@ -439,7 +439,8 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel { context.Input(framework::GradVarName("Loss"))->data(); Tensor* logit_grad = context.Output(framework::GradVarName("Logits")); - logit_grad->ShareDataWith(*context.Input("Softmax")); + framework::TensorCopy(*context.Input("Softmax"), context.GetPlace(), + context.device_context(), logit_grad); T* logit_grad_data = logit_grad->data(); const int batch_size = logit_grad->dims()[0]; diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 9eec8bfb5..93c492f6b 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -398,9 +398,9 @@ PYBIND11_MODULE(core, m) { py::arg("recursive_sequence_lengths"), R"DOC( Set LoD of the LoDTensor according to recursive sequence length. - For example, if recursive_sequence_lengths=[2, 3], meaning that + For example, if recursive_sequence_lengths=[[2, 3]], meaning that there are two sequences with length 2 and 3 respectively, the - corresponding lod would be [0, 2, 2+3], i.e, [0, 2, 5]. + corresponding lod would be [[0, 2, 2+3]], i.e, [[0, 2, 5]]. Args: recursive_sequence_lengths (List[List[int]]): sequence lengths. @@ -803,9 +803,11 @@ All parameter, weight, gradient are variables in Paddle. .def(py::init()) .def("close", &Executor::Close) .def("run", [](Executor &self, const ProgramDesc &prog, Scope *scope, - int block_id, bool create_local_scope, bool create_vars) { + int block_id, bool create_local_scope, bool create_vars, + const std::vector &fetch_vars) { pybind11::gil_scoped_release release; - self.Run(prog, scope, block_id, create_local_scope, create_vars); + self.Run(prog, scope, block_id, create_local_scope, create_vars, + fetch_vars); }); m.def("init_gflags", framework::InitGflags); diff --git a/paddle/scripts/fast_install.sh b/paddle/scripts/fast_install.sh index b960d0f00..0461944ca 100644 --- a/paddle/scripts/fast_install.sh +++ b/paddle/scripts/fast_install.sh @@ -1,5 +1,37 @@ #!/bin/bash +## purple to echo +function purple(){ + echo -e "\033[35m$1\033[0m" +} + + +## green to echo +function green(){ + echo -e "\033[32m$1\033[0m" +} + +## Error to warning with blink +function bred(){ + echo -e "\033[31m\033[01m\033[05m$1\033[0m" +} + +## Error to warning with blink +function byellow(){ + echo -e "\033[33m\033[01m\033[05m$1\033[0m" +} + + +## Error +function red(){ + echo -e "\033[31m\033[01m$1\033[0m" +} + +## warning +function yellow(){ + echo -e "\033[33m\033[01m$1\033[0m" +} + path='http://paddlepaddle.org/download?url=' #release_version=`curl -s https://pypi.org/project/paddlepaddle/|grep -E "/project/paddlepaddle/"|grep "release"|awk -F '/' '{print $(NF-1)}'|head -1` release_version=1.2.0 @@ -228,36 +260,128 @@ function checkLinuxPaddleVersion(){ done } -function checkLinuxPip(){ +function checkPythonVirtualenv(){ while true do - echo "请输入您要使用的pip目录（您可以另起终端，并使用which pip来查看）：" - read -p "" pip_path - if [ "$pip_path" == "" -o ! -f "$pip_path" ];then - echo "检测结果：pip不存在,请重新输入" - continue - fi - python_version=`$pip_path --version|awk -F "[ |)]" '{print $6}'|sed 's#\.##g'` - if [ "$python_version" == "27" ];then - uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27mu"` - if [[ "$uncode" == "" ]];then - uncode= - else - uncode=u - fi - fi - if [ "$python_version" == "" ];then - echo "检测结果：pip不存在,请重新输入" - else - version_list=`echo "${python_list[@]}" | grep "$python_version" ` - if [ "$version_list" != "" ];then - echo "检测结果：找到python${python_version}版本" - break - else - echo "检测结果：找不到可用的 pip, 我们只支持Python27/35/36/37及其对应的pip, 请重新输入， 或使用ctrl + c退出 " - fi - fi + read -p " + 是否使用python virtualenv虚环境安装(y/n)": check_virtualenv + case $check_virtualenv in + y) + echo "为您使用python虚环境安装" + ;; + n) + break + ;; + *) + continue + ;; + esac + + virtualenv_path=`which virtualenv 2>&1` + if [ "$virtualenv_path" == "" ];then + $python_path -m pip install virtualenv + if [ "$?" != '0' ];then + echo "安装虚拟环境失败,请检查本地环境" + fi + fi + + while true + do + read -p "请输入虚拟环境名字：" virtualenv_name + if [ "$virtualenv_name" == "" ];then + echo "不能为空" + continue + fi + break + done + + virtualenv -p $python_path ${virtualenv_name} + if [ "$?" != 0 ];then + echo "创建虚环境失败,请检查环境" + exit 2 + fi + cd ${virtualenv_name} + source ./bin/activate + + if [ "$?" == 0 ];then + use_virtualenv= + python_path=`which python` + break + else + echo "创建虚环境失败,请检查环境" + exit 2 + fi + done +} + +function checkLinuxPython(){ + python_path=`which python 2>/dev/null` + while true + do + if [ "$python_path" == '' ];then + while true + do + read -p "没有找到默认的python版本,请输入要安装的python路径:" python_path + python_path=`$python_path -V` + if [ "$python_path" != "" ];then + break + else + echo "输入路径有误,未找到pyrhon" + fi done + fi + + python_version=`$python_path -V 2>&1|awk -F '[ .]' '{print $2$3}'` + pip_version=`$python_path -m pip -V|awk -F '[ .]' '{print $2}'` + while true + do + read -p " + 找到python版本$python_version,使用请输入y,选择其他版本请输n(y/n):" check_python + case $check_python in + n) + read -p "请指定您的python路径:" new_python_path + python_V=`$new_python_path -V 2>/dev/null` + if [ "$python_V" != "" ];then + python_path=$new_python_path + python_version=`$python_path -V 2>&1|awk -F '[ .]' '{print $2$3}'` + pip_version=`python -m pip -V|awk -F '[ .]' '{print $2}'` + echo "您的python版本为${python_version}" + break + else + echo 输入有误,未找到python路径 + fi + ;; + y) + break + ;; + *) + echo "输入有误，请重新输入." + continue + ;; + esac + done + + if [ "$pip_version" -lt 9 ];then + echo "您的pip版本小于9.0.1 请升级pip (pip install --upgrade pip)" + exit 0 + fi + + if [ "$python_version" == "27" ];then + uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27mu"` + if [[ "$uncode" == "" ]];then + uncode= + else + uncode=u + fi + fi + + version_list=`echo "${python_list[@]}" | grep "$python_version" ` + if [ "$version_list" == "" ];then + echo "找不到可用的 pip, 我们只支持Python27/35/36/37及其对应的pip, 请重新输入， 或使用ctrl + c退出 " + else + break + fi + done } function checkLinuxAVX(){ @@ -287,25 +411,36 @@ function PipLinuxInstall(){ wheel_cpu_develop="http://paddle-wheel.bj.bcebos.com/latest-cpu-${AVX}-${math}/paddlepaddle-latest-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" wheel_gpu_develop="http://paddle-wheel.bj.bcebos.com/latest-gpu-cuda${CUDA}-cudnn${CUDNN}-${AVX}-${math}/paddlepaddle_gpu-latest-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" - if [[ "$paddle_version" == "2" ]];then if [[ "$GPU" == "gpu" ]];then if [[ ${AVX} == "avx" ]];then rm -rf `echo $wheel_gpu_release|awk -F '/' '{print $NF}'` wget -q $wheel_gpu_release if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi else rm -rf `echo $wheel_gpu_release_novax|awk -F '/' '{print $NF}'` wget -q $wheel_gpu_release_novax if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release_noavx + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release_noavx + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi fi @@ -313,9 +448,15 @@ function PipLinuxInstall(){ rm -rf `echo $wheel_cpu_release|awk -F '/' '{print $NF}'` wget -q $wheel_cpu_release if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_release + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_release + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi fi @@ -324,18 +465,30 @@ function PipLinuxInstall(){ rm -rf `echo $wheel_gpu_develop|awk -F '/' '{print $NF}'` wget -q $wheel_gpu_develop if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_develop + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_develop + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi else rm -rf `echo $wheel_cpu_develop|awk -F '/' '{print $NF}'` wget -q $wheel_cpu_develop if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_develop + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_develop + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi fi @@ -575,95 +728,122 @@ gpu_list=( echo echo "Step 5. 检测pip版本" echo - checkLinuxPip + checkLinuxPython echo checkLinuxAVX + echo + echo "Step 6.是否使用Python的虚拟环境" + use_virtualenv="--user" + checkPythonVirtualenv echo "*********************2. 开始安装*****************************" PipLinuxInstall + if [ "$check_virtualenv" == 'y' ];then + echo "虚环境创建成功，请cd 进入${virtualenv_name}, 执行 source bin/activate　进入虚环境。退出虚环境执行 deactivate命令。 + 更多虚环境使用方法请参考virtualenv官网:https://virtualenv.pypa.io/en/latest/" + fi +} + +function clearMacPythonEnv(){ + python_version="" + python_brief_version="" + python_root="" } function checkMacPython2(){ while true do - read -p " - => 未能在常规路径下找到Python2，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python2（注意Python版本不能低于2.7.15） - 如希望自定义Python路径，请输入路径：" python_root - echo python_version=`$python_root --version 2>&1 1>&1` - if [ $? == "0" ];then - : + if [[ $? == "0" ]];then + if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ];then + clearMacPythonEnv + else + check_python=`echo $python_version | grep "Python 2"` + if [[ -n "$check_python" ]];then + while true + do + echo -e " => 在您的环境中找到 \033[32m[ $python_version ]\033[0m, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " + read -p "" use_python + echo + use_python=`echo $use_python | tr 'A-Z' 'a-z'` + if [[ "$use_python" == "y" ]]||[[ "$use_python" == "" ]];then + use_python="y" + break + elif [[ "$use_python" == "n" ]];then + clearMacPythonEnv + break + else + red " 输入错误，请重新输入(y/n)" + fi + done + if [[ "$use_python" == "y" ]];then + return 0 + fi + else + red " 您输入Python的不是Python2" + clearMacPythonEnv + fi + fi else - python_version="" + clearMacPythonEnv + red " => 未能在常规路径下找到可用的Python2，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python2（注意Python版本不能低于2.7.15）" + read -p " 如希望自定义Python路径，请输入路径 + 如果希望重新选择Python版本，请回车：" python_root + echo + if [[ "$python_root" == "" ]];then + python_V="" + clearMacPythonEnv + return 1 + fi fi - check_python=`echo $python_version | grep "Python 2"` - if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then - python_version="" - elif [ -n "$check_python" ];then - while true - do - read -p " - => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " use_python - echo - use_python=`echo $use_python | tr 'A-Z' 'a-z'` - if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then - use_python="y" - break - elif [ "$use_python" == "n" ];then - python_root="" - break - else - echo "输入错误，请重新输入(y/n)" - fi - done - if [ "$use_python" == "y" ];then - break - fi - else - echo "您输入Python的不是Python2" - python_version="" - fi done } function checkMacPython3(){ while true do - read -p " - => 未能在常规路径下找到Python3，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载Python3 - 如希望自定义Python路径，请输入路径：" python_root - python_version=`$python_root --version 2>&1 1>&1` - if [ $? == "0" ];then - : + python_version=`$python_root --version 2>&1 1>&1` + if [[ $? == "0" ]];then + if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then + clearMacPythonEnv + else + check_python=`echo $python_version | grep "Python 3"` + if [[ -n "$check_python" ]];then + while true + do + echo -e " => 在您的环境中找到 \033[32m[ $python_version ]\033[0m, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " + read -p "" use_python + echo + use_python=`echo $use_python | tr 'A-Z' 'a-z'` + if [[ "$use_python" == "y" ]]||[[ "$use_python" == "" ]];then + use_python="y" + break + elif [[ "$use_python" == "n" ]];then + clearMacPythonEnv + break + else + red " 输入错误，请重新输入(y/n)" + fi + done + if [[ "$use_python" == "y" ]];then + return 0 + fi + else + red " 您输入Python的不是Python3" + clearMacPythonEnv + fi + fi else - python_version="" + clearMacPythonEnv + red " => 未能在常规路径下找到可用的Python3，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python3（注意Python版本不能低于3.5.x)" + read -p " 如希望自定义Python路径，请输入路径 + 如果希望重新选择Python版本，请回车：" python_root + echo + if [[ "$python_root" == "" ]];then + python_V="" + clearMacPythonEnv + return 1 + fi fi - check_python=`echo $python_version | grep "Python 3"` - if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then - python_version="" - elif [ -n "$check_python" ] ;then - while true - do - read -p " - => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " use_python - echo - use_python=`echo $use_python | tr 'A-Z' 'a-z'` - if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then - use_python="y" - break - elif [ "$use_python" == "n" ];then - python_root="" - break - else - echo "输入错误，请重新输入(y/n)" - fi - done - if [ "$use_python" == "y" ];then - break - fi - else - echo "您输入Python的不是Python3" - python_version="" - fi done } @@ -672,145 +852,160 @@ function checkMacPaddleVersion(){ do read -n1 -p "Step 2. 选择PaddlePaddle的版本，请按回车键继续..." echo - read -p " - 1. 开发版：对应Github上develop分支，如您需要开发、或希望使用PaddlePaddle最新功能，请选用此版本 - 2. 稳定版（推荐）：如您无特殊开发需求，建议使用此版本，目前最新的版本号为 ${release_version} - - => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. 稳定版 】 。请在这里输入并回车：" paddle_version - if [ "$paddle_version" == "1" ]||[ "$paddle_version" == "2" ];then + yellow " 1. 开发版：对应Github上develop分支，如您需要开发、或希望使用PaddlePaddle最新功能，请选用此版本" + yellow " 2. 稳定版（推荐）：如您无特殊开发需求，建议使用此版本，目前最新的版本号为 ${release_version}" + read -p " => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. 稳定版 】 。请在这里输入并回车：" paddle_version + if [[ "$paddle_version" == "1" ]]||[[ "$paddle_version" == "2" ]];then echo - echo "您选择了数字【"$paddle_version" 】" + yellow " 您选择了数字【"$paddle_version" 】" echo break else paddle_version="2" echo - echo "您选择了数字【2】" + yellow " 您选择了数字【2】" echo break fi done } +function initCheckMacPython2(){ + echo + yellow " 您选择了Python "$python_V"，正在寻找符合要求的Python 2版本" + echo + python_root=`which python2.7` + if [[ "$python_root" == "" ]];then + python_root=`which python` + fi + checkMacPython2 + if [[ "$?" == "1" ]];then + return 1 + else + return 0 + fi +} -function checkMacPythonVersion(){ - while true - do - read -n1 -p "Step 3. 选择Python版本，请按回车键继续..." - read -p " - 2. 使用python 2.x - 3. 使用python 3.x +function initCheckMacPython3(){ + echo + yellow " 您选择了Python "$python_V"，正在寻找符合您要求的Python 2版本" + echo + python_root=`which python3` + checkMacPython3 + if [[ "$?" == "1" ]];then + return 1 + else + return 0 + fi +} - => 请输入数字2或3。如输入其他字符或直接回车，将会默认使用【Python 2 】。请在这里输入并回车：" python_V - echo - if [ "$python_V" == "" ];then - python_V="2" +function checkMacPip(){ + if [[ "$python_V" == "2" ]]||[[ "$python_V" == "3" ]];then + + python_brief_version=`$python_root -m pip -V |awk -F "[ |)]" '{print $6}'|sed 's#\.##g'` + if [[ ${python_brief_version} == "" ]];then + red "您输入的python：${python_root} 对应的pip不可用，请检查此pip或重新选择其他python" + echo + return 1 fi - echo "您选择了数字【"$python_V"】，正在寻找符合您要求的Python版本，请按回车键继续..." - echo - if [ "$python_V" == "2" ];then - python_root=`which python2.7` - if [ "$python_root" == "" ];then - python_root=`which python` - fi - python_version=`$python_root --version 2>&1 1>&1` - if [ $? == "0" ];then - : - else - python_version="" - fi - if [ "$python_root" == "" ]||[ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ]||[ "$python_root" == "/usr/bin/python2.7" -a "$python_version" == "Python 2.7.10" ];then - checkMacPython2 - fi - while true - do - read -p " - => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车：" use_python - echo - use_python=`echo $use_python | tr 'A-Z' 'a-z'` - if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then - break - elif [ "$use_python" == "n" ];then - python_root="" - checkMacPython2 - break + pip_version=`$python_root -m pip -V |awk -F '[ .]' '{print $2}'` + if [[ 9 -le ${pip_version} ]];then + : + else + red "您的pip版本过低，请安装pip 9.0.1及以上的版本" + echo + return 1 + fi + if [[ "$python_brief_version" == "" ]];then + clearMacPythonEnv + red "您的 $python_root 对应的pip存在问题，请按ctrl + c退出后重新安装pip，或切换其他python版本" + echo + return 1 + else + if [[ $python_brief_version == "27" ]];then + uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27"` + if [[ $uncode == "" ]];then + uncode="mu" else - echo "输入错误，请重新输入(y/n)" + uncode="m" fi - done - - elif [ "$python_V" == "3" ];then - python_root=`which python3` - python_version=`$python_root --version 2>&1 1>&1` - if [ $? == "0" ];then - : - else - python_version="" - fi - if [ "$python_root" == "" ]||[ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ];then - checkMacPython3 - fi - while true - do - read -p " - => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车：" use_python + fi + version_list=`echo "${python_list[@]}" | grep "$python_brief_version" ` + if [[ "$version_list" != "" ]];then + return 0 + else + red "未找到可用的pip或pip3。PaddlePaddle目前支持：Python2.7/3.5/3.6/3.7及其对应的pip, 请重新输入，或使用ctrl + c退出" echo - use_python=`echo $use_python | tr 'A-Z' 'a-z'` - if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then - break - elif [ "$use_python" == "n" ];then - checkMacPython3 - break - else - echo "输入错误，请重新输入(y/n)" - fi - done - else - : - fi + clearMacPythonEnv + return 1 + fi + fi + fi +} - if [ "$python_V" == "2" ]||[ "$python_V" == "3" ];then - python_brief_version=`$python_root -m pip -V |awk -F "[ |)]" '{print $6}'|sed 's#\.##g'` - if [[ $python_brief_version == "27" ]];then - uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27"` - if [[ $uncode == "" ]];then - uncode="mu" - else - uncode="m" - fi - fi - version_list=`echo "${python_list[@]}" | grep "$python_brief_version" ` - if [ "$version_list" != "" ];then - break +function checkMacPythonVersion(){ + while true + do + read -n1 -p "Step 3. 选择Python版本，请按回车键继续..." + echo + yellow " 2. 使用python 2.x" + yellow " 3. 使用python 3.x" + read -p " => 请输入数字2或3。如输入其他字符或直接回车，将会默认使用【Python 2 】。请在这里输入并回车：" python_V + if [[ "$python_V" == "" ]];then + python_V="2" + fi + if [[ "$python_V" == "2" ]];then + initCheckMacPython2 + if [[ "$?" == "0" ]];then + checkMacPip + if [[ "$?" == "0" ]];then + return 0 + else + : + fi else - echo "未找到可用的pip或pip3。PaddlePaddle目前支持：Python2.7/3.5/3.6/3.7及其对应的pip, 请重新输入，或使用ctrl + c退出" - fi - else - echo "输入错误，请重新输入" - fi + : + fi + elif [[ "$python_V" == "3" ]];then + initCheckMacPython3 + if [[ "$?" == "0" ]];then + checkMacPip + if [[ "$?" == "0" ]];then + return 0 + else + : + fi + else + : + fi + else + red "输入错误，请重新输入" + fi done } function checkMacAVX(){ read -n1 -p "Step 4. 检测您的Mac是否支持AVX指令集，请按回车键继续..." - echo if [[ $AVX != "" ]];then AVX="avx" - echo "检测结果：支持" + echo "" + green " 检测结果：支持" + echo "" + return 0 else - read -n1 -p "检测结果：不支持。非常抱歉，PaddlePaddle在Mac系统暂不提供no_avx类型的安装包，您可以选择在Linux系统中安装no_avx版的PaddlePaddle, 请按回车键退出..." - exit + red " 检测结果：不支持。非常抱歉，PaddlePaddle在Mac系统暂不提供no_avx类型的安装包，您可以选择在Linux系统中安装no_avx版的PaddlePaddle, 请按回车键退出..." + echo + return 1 fi - echo } function checkMacGPU(){ read -n1 -p "Step 5. 选择CPU/GPU版本，请按回车键继续..." echo if [[ $GPU != "" ]];then - echo "MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" + yellow " MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" else - echo "MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" + yellow " MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" GPU=cpu fi echo @@ -822,38 +1017,44 @@ function macos() { while true do + checkMacPaddleVersion + checkMacPythonVersion + checkMacAVX + checkMacGPU - echo "*********************2. 开始安装*****************************" + green "*********************2. 开始安装*****************************" echo - read -n1 -p "即将为您下载并安装PaddlePaddle，请按回车键继续..." + yellow "即将为您下载并安装PaddlePaddle，请按回车键继续..." + read -n1 -p "" echo if [[ $paddle_version == "2" ]];then $python_root -m pip install paddlepaddle - if [ $? == "0" ];then - echo "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + if [[ $? == "0" ]];then + green "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" break else rm $whl_cpu_release - echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" echo"" echo "==========================================================================================" echo"" exit 1 fi else - if [ -f $whl_cpu_develop ];then + if [[ -f $whl_cpu_develop ]];then $python_root -m pip install $whl_cpu_develop - if [ $? == "0" ];then + if [[ $? == "0" ]];then rm -rf $whl_cpu_develop - echo "安装成功！小提示：可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + # TODO add install success check here + green "安装成功！小提示：可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" break else - echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" echo"" echo "==========================================================================================" echo"" @@ -861,15 +1062,15 @@ function macos() { fi else wget ${path}$whl_cpu_develop -O $whl_cpu_develop - if [ $? == "0" ];then + if [[ $? == "0" ]];then $python_root -m pip install $whl_cpu_develop - if [ $? == "0" ];then + if [[ $? == "0" ]];then rm $wheel_cpu_develop - echo "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + green "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" break else rm $whl_cpu_release - echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" echo"" echo "==========================================================================================" echo"" @@ -877,7 +1078,7 @@ function macos() { fi else rm $whl_cpu_develop - echo "未能正常安装PaddlePaddle，请检查您的网络 或者确认您是否安装有 wget，或者ctrl + c退出后反馈至https://github.com/PaddlePaddle/Paddle/issues" + red "未能正常安装PaddlePaddle，请检查您的网络 或者确认您是否安装有 wget，或者ctrl + c退出后反馈至https://github.com/PaddlePaddle/Paddle/issues" echo"" echo "==========================================================================================" echo"" @@ -890,33 +1091,35 @@ function macos() { function main() { echo "*********************************" - echo "欢迎使用PaddlePaddle快速安装脚本" + green "欢迎使用PaddlePaddle快速安装脚本" echo "*********************************" echo - echo "如果您在安装过程中遇到任何问题，请在https://github.com/PaddlePaddle/Paddle/issues反馈，我们的工作人员将会帮您答疑解惑" + yellow "如果您在安装过程中遇到任何问题，请在https://github.com/PaddlePaddle/Paddle/issues反馈，我们的工作人员将会帮您答疑解惑" echo - echo "本安装包将帮助您在Linux或Mac系统下安装PaddlePaddle，包括 1）安装前的准备和 2）开始安装 两部分" + echo "本安装包将帮助您在Linux或Mac系统下安装PaddlePaddle,包括" + yellow "1）安装前的准备" + yellow "2）开始安装" echo read -n1 -p "请按回车键进行下一步..." echo echo - echo "*********************1. 安装前的准备*****************************" + green "*********************1. 安装前的准备*****************************" echo echo "Step 1. 正在检测您的操作系统信息..." echo SYSTEM=`uname -s` - if [ "$SYSTEM" == "Darwin" ];then - echo "您的系统为：MAC OSX" + if [[ "$SYSTEM" == "Darwin" ]];then + yellow " 您的系统为：MAC OSX" echo macos else - echo "您的系统为：Linux" + yellow " 您的系统为：Linux" echo OS=`cat /etc/issue|awk 'NR==1 {print $1}'` - if [ $OS == "\S" ] || [ "$OS" == "CentOS" ] || [ $OS == "Ubuntu" ];then + if [[ $OS == "\S" ]] || [[ "$OS" == "CentOS" ]] || [[ $OS == "Ubuntu" ]];then linux else - echo "您的系统不在本安装包的支持范围，如您需要在windows环境下安装PaddlePaddle，请您参考PaddlePaddle官网的windows安装文档" + red "您的系统不在本安装包的支持范围，如您需要在windows环境下安装PaddlePaddle，请您参考PaddlePaddle官网的windows安装文档" fi fi } diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py index 53746afdb..fe2ae67ec 100644 --- a/python/paddle/__init__.py +++ b/python/paddle/__init__.py @@ -25,4 +25,5 @@ import paddle.reader import paddle.dataset import paddle.batch import paddle.compat +import paddle.distributed batch = batch.batch diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index 396f36e18..aa1f85734 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -161,7 +161,6 @@ def __bootstrap__(): 'times_excess_than_required_tmp_allocation', 'enable_inplace_whitelist' ] - core.init_gflags([sys.argv[0]] + ["--tryfromenv=" + ",".join(read_env_flags)]) core.init_glog(sys.argv[0]) diff --git a/python/paddle/fluid/compiler.py b/python/paddle/fluid/compiler.py index ef0242942..a258ad76c 100644 --- a/python/paddle/fluid/compiler.py +++ b/python/paddle/fluid/compiler.py @@ -177,7 +177,10 @@ class CompiledProgram(object): # FIXME(dzhwinter): enable_inplace should be after memory_optimize # if turn on python memory optimize, turn off the inplace_pass. - self._build_strategy.enable_inplace = False if self._program._is_mem_optimized else True + if self._build_strategy.memory_optimize is True: + self._build_strategy.memory_optimize = False if self._program._is_mem_optimized else True + if self._build_strategy.enable_inplace is True: + self._build_strategy.enable_inplace = False if self._program._is_mem_optimized else True if self._build_strategy.num_trainers > 1 and trainers_endpoints: assert self._build_strategy.num_trainers == len( diff --git a/python/paddle/fluid/contrib/int8_inference/README.md b/python/paddle/fluid/contrib/int8_inference/README.md index a9691dad4..460ae393f 100644 --- a/python/paddle/fluid/contrib/int8_inference/README.md +++ b/python/paddle/fluid/contrib/int8_inference/README.md @@ -63,10 +63,10 @@ Notes: ## 4. How to reproduce the results * Small dataset ```bash -python python/paddle/fluid/contrib/tests/test_calibration.py +FLAGS_use_mkldnn=true python python/paddle/fluid/contrib/tests/test_calibration.py ``` * Full dataset ```bash -DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py +FLAGS_use_mkldnn=true DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py ``` diff --git a/python/paddle/fluid/contrib/tests/CMakeLists.txt b/python/paddle/fluid/contrib/tests/CMakeLists.txt index 81aee1233..a2c594164 100644 --- a/python/paddle/fluid/contrib/tests/CMakeLists.txt +++ b/python/paddle/fluid/contrib/tests/CMakeLists.txt @@ -6,5 +6,9 @@ if(APPLE OR WIN32 OR NOT WITH_MKL) endif() foreach(src ${TEST_OPS}) - py_test(${src} SRCS ${src}.py) + if(src MATCHES "test_calibration") + py_test(${src} SRCS ${src}.py ENVS FLAGS_use_mkldnn=true) + else() + py_test(${src} SRCS ${src}.py) + endif() endforeach() diff --git a/python/paddle/fluid/contrib/tests/test_calibration.py b/python/paddle/fluid/contrib/tests/test_calibration.py index 424ea245a..b9f938beb 100644 --- a/python/paddle/fluid/contrib/tests/test_calibration.py +++ b/python/paddle/fluid/contrib/tests/test_calibration.py @@ -199,7 +199,6 @@ class TestCalibrationForResnet50(unittest.TestCase): def run_program(self, model_path, generate_int8=False, algo='direct'): image_shape = [3, 224, 224] - os.environ['FLAGS_use_mkldnn'] = 'True' fluid.memory_optimize(fluid.default_main_program()) @@ -241,9 +240,6 @@ class TestCalibrationForResnet50(unittest.TestCase): label = label.reshape([-1, 1]) running_program = calibrator.sampling_program.clone( ) if generate_int8 else infer_program.clone() - for op in running_program.current_block().ops: - if op.has_attr("use_mkldnn"): - op._set_attr("use_mkldnn", True) t1 = time.time() _, acc1, _ = exe.run( diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index 8815911ea..fc3b24fa8 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -261,45 +261,42 @@ def _as_lodtensor(data, place): class Executor(object): """ - An Executor in Python, only support the single-GPU running. For multi-cards, please refer to - ParallelExecutor. - Python executor takes a program, add feed operators and fetch operators to this program according + An Executor in Python, supports single/multiple-GPU running, and single/multiple-CPU running. + Python executor takes a program, adds feed operators and fetch operators to this program according to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides - the variables(or names) that user want to get after program run. Note: the executor will run all + the variables(or names) that user wants to get after program runs. Note: the executor will run all operators in the program but not only the operators dependent by the fetch_list. - It store the global variables into the global scope, and create a local scope for the temporary - variables. The local scope contents will be discarded after every minibatch forward/backward finished. - But the global scope variables will be persistent through different runs. - All of ops in program will be running in sequence. + It stores the global variables into the global scope, and creates a local scope for the temporary + variables. The contents in local scope may be discarded after every minibatch forward/backward + finished. But the global scope variables will be persistent through different runs. Example: - .. code-block:: python - # First create the Executor. - place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - exe = fluid.Executor(place) - - # Run the startup program once and only once. - # Not need to optimize/compile the startup program. - exe.run(fluid.default_startup_program()) - - # Run the main program directly without compile. - loss, = exe.run(fluid.default_main_program(), - feed=feed_dict, - fetch_list=[loss.name]) - # Or, compiled the program and run. See `CompiledProgram` for more detail. - compiled_prog = compiler.CompiledProgram( - fluid.default_main_program()).with_data_parallel( - loss_name=loss.name) - loss, = exe.run(compiled_prog, - feed=feed_dict, - fetch_list=[loss.name]) + + .. code-block:: python + + # First create the Executor. + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + + # Run the startup program once and only once. + # Not need to optimize/compile the startup program. + exe.run(fluid.default_startup_program()) + + # Run the main program directly without compile. + loss, = exe.run(fluid.default_main_program(), + feed=feed_dict, + fetch_list=[loss.name]) + # Or, compiled the program and run. See `CompiledProgram` for more detail. + compiled_prog = compiler.CompiledProgram( + fluid.default_main_program()).with_data_parallel( + loss_name=loss.name) + loss, = exe.run(compiled_prog, + feed=feed_dict, + fetch_list=[loss.name]) Args: place(core.CPUPlace|core.CUDAPlace(n)): indicate the executor run on which device - - Note: For debugging complicated network in parallel-GPUs, you can test it on the executor. - They has the exactly same arguments, and expected the same results. """ def __init__(self, place): @@ -382,6 +379,12 @@ class Executor(object): ] return outs + ''' + TODO(typhoonzero): Define "no longer use" meaning? Can user create + a new Executor for the same program and run? + TODO(panyx0718): Why ParallelExecutor doesn't have close? + ''' + def close(self): """ Close this executor. @@ -389,9 +392,6 @@ class Executor(object): You can no longer use this executor after calling this method. For the distributed training, this method would free the resource on PServers related to the current Trainer. - TODO(typhoonzero): Define "no longer use" meaning? Can user create - a new Executor for the same program and run? - TODO(panyx0718): Why ParallelExecutor doesn't have close? Example: >>> cpu = core.CPUPlace() diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index 832c97c7d..8d27b080a 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -555,7 +555,8 @@ class OpProtoHolder(object): return { core.op_proto_and_checker_maker.kOpRoleAttrName(), core.op_proto_and_checker_maker.kOpRoleVarAttrName(), - core.op_proto_and_checker_maker.kOpNameScopeAttrName() + core.op_proto_and_checker_maker.kOpNameScopeAttrName(), + core.op_proto_and_checker_maker.kOpCreationCallstackAttrName() } diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index a2abbf36c..74950c1f6 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -766,7 +766,10 @@ def _load_distributed_persistables(executor, dirname, main_program=None): dtype=slice_var.dtype, persistable=True) - dim1_flatten = reduce(lambda x, y: x * y, slice.shape[1:]) + dim1_flatten = 1 + if len(slice.shape) >= 2: + dim1_flatten = reduce(lambda x, y: x * y, slice.shape[1:]) + start = int(offset / dim1_flatten) end = int(offset / dim1_flatten + slice.shape[0]) @@ -892,7 +895,7 @@ def save_inference_model(dirname, True is supported. Returns: - None + target_var_name_list(list): The fetch variables' name list Raises: ValueError: If `feed_var_names` is not a list of basestring. @@ -945,11 +948,13 @@ def save_inference_model(dirname, # TODO(Superjomn) add an IR pass to remove 1-scale op. with program_guard(main_program): uniq_target_vars = [] - for var in target_vars: + for i, var in enumerate(target_vars): if isinstance(var, Variable): - var1 = layers.scale(var, 1.) - uniq_target_vars.append(var1) + var = layers.scale( + var, 1., name="save_infer_model/scale_{}".format(i)) + uniq_target_vars.append(var) target_vars = uniq_target_vars + target_var_name_list = [var.name for var in target_vars] # when a pserver and a trainer running on the same machine, mkdir may conflict try: @@ -1006,6 +1011,7 @@ def save_inference_model(dirname, params_filename = os.path.basename(params_filename) save_persistables(executor, dirname, main_program, params_filename) + return target_var_name_list def load_inference_model(dirname, diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 3a6753b01..539c9675b 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -506,9 +506,9 @@ class While(object): while loop control flow. Args: - cond (Variable): condition used to compare. + cond(Variable): condition used to compare. is_test(bool): A flag indicating whether execution is in test phase. - name (str): The name of this layer. + name(str): The name of this layer. Examples: .. code-block:: python @@ -589,7 +589,8 @@ class While(object): def lod_rank_table(x, level=0): - """LoD Rank Table Operator. Given an input variable **x** and a level number + """ + LoD Rank Table Operator. Given an input variable **x** and a level number of LoD, this layer creates a LodRankTable object. A LoDRankTable object contains a list of bi-element tuples. Each tuple consists of an index and a length, both of which are int type. Refering to specified level of LoD, @@ -883,10 +884,8 @@ def less_than(x, y, force_cpu=None, cond=None, **ignored): return cond -def equal(x, y, cond=None, **ignored): +def equal(x, y, cond=None): """ - **equal** - This layer returns the truth value of :math:`x == y` elementwise. Args: @@ -1458,7 +1457,6 @@ class DynamicRNN(object): Returns: The current timestep in the input sequence. - """ self._assert_in_rnn_block_("step_input") if not isinstance(x, Variable): @@ -1535,8 +1533,7 @@ class DynamicRNN(object): @signature_safe_contextmanager def block(self): """ - The block for user to define operators in RNN. See the class docstring - for more details. + The block for user to define operators in RNN. """ if self.status != DynamicRNN.BEFORE_RNN: raise ValueError("rnn.block() can only be invoke once") @@ -1640,8 +1637,7 @@ class DynamicRNN(object): dtype(str|numpy.dtype): The data type of the initialized memory. Returns: - the memory variable. - + The memory variable. """ self._assert_in_rnn_block_('memory') self._init_zero_idx_() @@ -1740,7 +1736,7 @@ class DynamicRNN(object): def output(self, *outputs): """ - mark the RNN output variables. + Mark the RNN output variables. Args: outputs: The output variables. diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index c983e2a44..0695bccd2 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -545,15 +545,16 @@ def yolov3_loss(x, TypeError: Attr ignore_thresh of yolov3_loss must be a float number Examples: - .. code-block:: python - - x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32') - gtbox = fluid.layers.data(name='gtbox', shape=[6, 5], dtype='float32') - gtlabel = fluid.layers.data(name='gtlabel', shape=[6, 1], dtype='int32') - anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326] - anchors = [0, 1, 2] - loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, class_num=80, anchors=anchors, - ignore_thresh=0.5, downsample_ratio=32) + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32') + gtbox = fluid.layers.data(name='gtbox', shape=[6, 5], dtype='float32') + gtlabel = fluid.layers.data(name='gtlabel', shape=[6, 1], dtype='int32') + anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326] + anchor_mask = [0, 1, 2] + loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, gtlabel=gtlabel, anchors=anchors, + anchor_mask=anchor_mask, class_num=80, + ignore_thresh=0.7, downsample_ratio=32) """ helper = LayerHelper('yolov3_loss', **locals()) diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index b88be6690..a9b391fd5 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -56,7 +56,10 @@ def data(name, Args: name(str): The name/alias of the function - shape(list): Tuple declaring the shape. + shape(list): Tuple declaring the shape. If :code:`append_batch_size` is + True and there is no -1 inside :code:`shape`, it should be + considered as the shape of the each sample. Otherwise, it + should be considered as the shape of the batched data. append_batch_size(bool): 1. If true, it prepends -1 to the shape. For example if shape=[1], the resulting shape is [-1, 1]. diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 09b1b3021..da6c24100 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -24,7 +24,7 @@ from ..framework import OpProtoHolder, Variable, core, convert_np_dtype_to_dtype from ..layer_helper import LayerHelper __all__ = [ - 'deprecated', 'generate_layer_fn', 'generate_layer_fn_noattr', 'autodoc', + 'deprecated', 'generate_layer_fn', 'generate_activation_fn', 'autodoc', 'templatedoc' ] @@ -89,6 +89,9 @@ def _generate_doc_string_(op_proto, additional_args_lines=None): buf.write('\n') skip_attrs = OpProtoHolder.generated_op_attr_names() + # attr use_mkldnn and is_test also should not be visible to users. + skip_attrs.add("use_mkldnn") + skip_attrs.add("is_test") for each_attr in op_proto.attrs: if each_attr.name in skip_attrs: @@ -226,7 +229,7 @@ def generate_layer_fn(op_type): return func -def generate_layer_fn_noattr(op_type): +def generate_activation_fn(op_type): """Register the Python layer for an Operator without Attribute. Args: @@ -246,6 +249,7 @@ def generate_layer_fn_noattr(op_type): func.__name__ = op_type func.__doc__ = _generate_doc_string_(op_proto) + return func diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 0e4b5aadc..cdd64e89b 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -2441,7 +2441,7 @@ def pool2d(input, data = fluid.layers.data( name='data', shape=[3, 32, 32], dtype='float32') - conv2d = fluid.layers.pool2d( + pool2d = fluid.layers.pool2d( input=data, pool_size=2, pool_type='max', @@ -2490,6 +2490,7 @@ def pool2d(input, return pool_out +@templatedoc() def pool3d(input, pool_size=-1, pool_type="max", @@ -2501,13 +2502,19 @@ def pool3d(input, name=None, exclusive=True): """ - This function adds the operator for pooling in 3-dimensions, using the - pooling configurations mentioned in input parameters. + ${comment} Args: - input (Variable): ${input_comment} - pool_size (int): ${ksize_comment} - pool_type (str): ${pooling_type_comment} + input (Variable): The input tensor of pooling operator. The format of + input tensor is NCDHW, where N is batch size, C is + the number of channels, D is the depth of the feature, + H is the height of the feature, and W is the width + of the feature. + pool_size (int|list|tuple): The pool kernel size. If pool kernel size + is a tuple or list, it must contain three integers, + (pool_size_Depth, pool_size_Height, pool_size_Width). + Otherwise, the pool kernel size will be the cube of an int. + pool_type (string): ${pooling_type_comment} pool_stride (int): stride of the pooling layer. pool_padding (int): padding size. global_pooling (bool): ${global_pooling_comment} @@ -2520,6 +2527,19 @@ def pool3d(input, Returns: Variable: output of pool3d layer. + + Examples: + + .. code-block:: python + + data = fluid.layers.data( + name='data', shape=[3, 32, 32, 32], dtype='float32') + pool3d = fluid.layers.pool3d( + input=data, + pool_size=2, + pool_type='max', + pool_stride=1, + global_pooling=False) """ if pool_type not in ["max", "avg"]: raise ValueError( @@ -2569,7 +2589,27 @@ def adaptive_pool2d(input, require_index=False, name=None): """ - ${comment} + **Adaptive Pool2d Operator** + The adaptive_pool2d operation calculates the output based on the input, pool_size, + pool_type parameters. Input(X) and output(Out) are in NCHW format, where N is batch + size, C is the number of channels, H is the height of the feature, and W is + the width of the feature. Parameters(pool_size) should contain two elements which + represent height and width, respectively. Also the H and W dimensions of output(Out) + is same as Parameter(pool_size). + + For average adaptive pool2d: + + .. math:: + + hstart &= floor(i * H_{in} / H_{out}) + + hend &= ceil((i + 1) * H_{in} / H_{out}) + + wstart &= floor(j * W_{in} / W_{out}) + + wend &= ceil((j + 1) * W_{in} / W_{out}) + + Output(i ,j) &= \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} Args: input (Variable): The input tensor of pooling operator. The format of @@ -2579,8 +2619,8 @@ def adaptive_pool2d(input, pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list, it must contain two integers, (pool_size_Height, pool_size_Width). pool_type: ${pooling_type_comment} - require_index (bool): If true, the index of max pooling point along with outputs. - it cannot be set in average pooling type. + require_index (bool): If true, the index of max pooling point will be returned along + with outputs. It cannot be set in average pooling type. name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. @@ -2661,18 +2701,42 @@ def adaptive_pool3d(input, require_index=False, name=None): """ - ${comment} + **Adaptive Pool3d Operator** + The adaptive_pool3d operation calculates the output based on the input, pool_size, + pool_type parameters. Input(X) and output(Out) are in NCDHW format, where N is batch + size, C is the number of channels, D is the depth of the feature, H is the height of + the feature, and W is the width of the feature. Parameters(pool_size) should contain + three elements which represent height and width, respectively. Also the D, H and W + dimensions of output(Out) is same as Parameter(pool_size). + + For average adaptive pool3d: + + .. math:: + + dstart &= floor(i * D_{in} / D_{out}) + + dend &= ceil((i + 1) * D_{in} / D_{out}) + + hstart &= floor(j * H_{in} / H_{out}) + + hend &= ceil((j + 1) * H_{in} / H_{out}) + + wstart &= floor(k * W_{in} / W_{out}) + + wend &= ceil((k + 1) * W_{in} / W_{out}) + + Output(i ,j, k) &= \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)} Args: input (Variable): The input tensor of pooling operator. The format of - input tensor is NCHW, where N is batch size, C is - the number of channels, H is the height of the - feature, and W is the width of the feature. + input tensor is NCDHW, where N is batch size, C is + the number of channels, D is the depth of the feature, + H is the height of the feature, and W is the width of the feature. pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list, - it must contain two integers, (Depth, Height, Width). + it must contain three integers, (Depth, Height, Width). pool_type: ${pooling_type_comment} - require_index (bool): If true, the index of max pooling point along with outputs. - it cannot be set in average pooling type. + require_index (bool): If true, the index of max pooling point will be returned along + with outputs. It cannot be set in average pooling type. name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. @@ -2709,7 +2773,7 @@ def adaptive_pool3d(input, name='data', shape=[3, 32, 32], dtype='float32') pool_out, mask = fluid.layers.adaptive_pool3d( input=data, - pool_size=[3, 3], + pool_size=[3, 3, 3], pool_type='avg') """ if pool_type not in ["max", "avg"]: @@ -2930,6 +2994,7 @@ def batch_norm(input, "momentum": momentum, "epsilon": epsilon, "is_test": is_test, + "data_layout": data_layout, "use_mkldnn": False, "fuse_with_relu": fuse_with_relu, "use_global_stats": use_global_stats @@ -3235,7 +3300,7 @@ def group_norm(input, # create output mean_out = helper.create_variable(dtype=dtype, stop_gradient=True) variance_out = helper.create_variable(dtype=dtype, stop_gradient=True) - group_norm_out = helper.create_variable(dtype) + group_norm_out = helper.create_variable(dtype=dtype) helper.append_op( type="group_norm", diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 3dcf9dc06..6b4dc4ac8 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -14,7 +14,7 @@ from __future__ import print_function import os -from .layer_function_generator import generate_layer_fn, generate_layer_fn_noattr +from .layer_function_generator import generate_layer_fn, generate_activation_fn from .. import core from ..framework import convert_np_dtype_to_dtype_ @@ -53,7 +53,7 @@ globals()['_elementwise_div'] = generate_layer_fn('elementwise_div') __all__ += __activations_noattr__ for _OP in set(__activations_noattr__): - globals()[_OP] = generate_layer_fn_noattr(_OP) + globals()[_OP] = generate_activation_fn(_OP) __all__ += ["uniform_random"] diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index fbd04f1eb..fe2b3fbbd 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -1368,9 +1368,9 @@ class FtrlOptimizer(Optimizer): Args: learning_rate (float|Variable): global learning rate. - l1 (float): - l2 (float): - lr_power (float): + l1 (float): L1 regularization strength. + l2 (float): L2 regularization strength. + lr_power (float): Learning Rate Power. regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer. name: A optional name prefix. diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 52b260efd..29b7f6aba 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -148,7 +148,10 @@ class ParallelExecutor(object): else framework.default_main_program() # FIXME(dzhwinter): enable_inplace should be after memory_optimize # if turn on python memory optimize, turn off the inplace_pass. - build_strategy.enable_inplace = False if main._is_mem_optimized else True + if build_strategy.memory_optimize is True: + build_strategy.memory_optimize = False if main._is_mem_optimized else True + if build_strategy.enable_inplace is True: + build_strategy.enable_inplace = False if main._is_mem_optimized else True scope = scope if scope is not None else executor.global_scope() if share_vars_from and not isinstance(share_vars_from, diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index 699181d01..289a48aac 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -77,6 +77,7 @@ list(REMOVE_ITEM TEST_OPS test_bilinear_interp_op) list(REMOVE_ITEM TEST_OPS test_nearest_interp_op) list(REMOVE_ITEM TEST_OPS test_imperative_resnet) list(REMOVE_ITEM TEST_OPS test_imperative_optimizer) +list(REMOVE_ITEM TEST_OPS test_ir_memory_optimize_transformer) foreach(TEST_OP ${TEST_OPS}) py_test_modules(${TEST_OP} MODULES ${TEST_OP}) endforeach(TEST_OP) @@ -107,10 +108,18 @@ py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SE py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL) set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 450) py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL) +if(NOT WIN32) +py_test_modules(test_ir_memory_optimize_transformer MODULES test_ir_memory_optimize_transformer SERIAL) +endif() if(NOT APPLE) py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL) + if(CMAKE_BUILD_TYPE STREQUAL "Debug") + # change the timeout from 600 to 1200, because in debug mode, this test need more time. + set_tests_properties(test_image_classification_resnet PROPERTIES TIMEOUT 1200) + endif() endif() + if (WITH_NGRAPH) add_subdirectory(ngraph) endif() diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py index c429c8af7..a94487e67 100644 --- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -79,7 +79,7 @@ class TestParallelExecutorBase(unittest.TestCase): if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops build_strategy.fuse_relu_depthwise_conv = fuse_relu_depthwise_conv - build_strategy.memory_optimize = use_ir_memory_optimize + build_strategy.memory_optimize = False if memory_opt else use_ir_memory_optimize # python memory optimization is conflict with inplace pass. # Use ir graph memory optimization after inplace pass is the correct way. build_strategy.enable_inplace = False if memory_opt else enable_inplace diff --git a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py index 03471a443..c1fb53ecf 100644 --- a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py +++ b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py @@ -121,6 +121,8 @@ class TestMNIST(TestParallelExecutorBase): regularization=fluid.regularizer.L2Decay(1e-6)) return optimizer + # NOTE(dzh): + # need to make it compatible with elewise fuse act not_fuse_op_first_loss, not_fuse_op_last_loss = self.check_network_convergence( model, feed_dict={"image": img, @@ -128,6 +130,7 @@ class TestMNIST(TestParallelExecutorBase): use_cuda=use_cuda, fuse_elewise_add_act_ops=False, memory_opt=False, + use_ir_memory_optimize=False, optimizer=_optimizer) fuse_op_first_loss, fuse_op_last_loss = self.check_network_convergence( model, @@ -136,6 +139,7 @@ class TestMNIST(TestParallelExecutorBase): use_cuda=use_cuda, fuse_elewise_add_act_ops=True, memory_opt=False, + use_ir_memory_optimize=False, optimizer=_optimizer) for loss in zip(not_fuse_op_first_loss, fuse_op_first_loss): diff --git a/python/paddle/fluid/tests/unittests/test_slice_op.py b/python/paddle/fluid/tests/unittests/test_slice_op.py index 4e6ed3a74..aefd8cb6d 100644 --- a/python/paddle/fluid/tests/unittests/test_slice_op.py +++ b/python/paddle/fluid/tests/unittests/test_slice_op.py @@ -16,6 +16,7 @@ from __future__ import print_function import unittest import numpy as np +import paddle.fluid.core as core from op_test import OpTest @@ -63,5 +64,54 @@ class TestCase2(TestSliceOp): self.out = self.input[-3:3, 0:100, :, 2:-1] +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestFP16(TestSliceOp): + def config(self): + self.dtype = "float16" + self.input = np.random.random([3, 4, 5, 6]).astype(self.dtype) + self.starts = [-3, 0, 2] + self.ends = [3, 100, -1] + self.axes = [0, 1, 3] + self.out = self.input[-3:3, 0:100, :, 2:-1] + + def test_check_output(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-5) + + def test_check_grad_normal(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['Input'], 'Out', max_relative_error=0.006) + + +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestFP16_2(TestSliceOp): + def config(self): + self.dtype = "float16" + self.input = np.random.random([3, 4, 5]).astype(self.dtype) + self.starts = [0] + self.ends = [1] + self.axes = [1] + self.out = self.input[:, 0:1, :] + + def test_check_output(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-5) + + def test_check_grad_normal(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['Input'], + 'Out', + max_relative_error=0.006, + numeric_grad_delta=0.5) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index a3293afbb..eb5406865 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -1020,7 +1020,11 @@ class DistributeTranspiler(object): skip_dim0 = 0 slice_vars = self.param_var_mapping[orig_var_name] - orig_dim1_flatten = reduce(lambda x, y: x * y, slice_vars[0].shape[1:]) + orig_dim1_flatten = 1 + + if len(slice_vars[0].shape) >= 2: + orig_dim1_flatten = reduce(lambda x, y: x * y, + slice_vars[0].shape[1:]) for slice_var in slice_vars[:block_idx]: skip_dim0 += slice_var.shape[0] diff --git a/python/requirements.txt b/python/requirements.txt index 5a70f1aa3..36bd5d426 100644 --- a/python/requirements.txt +++ b/python/requirements.txt @@ -1,6 +1,6 @@ requests==2.9.2 numpy>=1.12 -protobuf==3.1 +protobuf>=3.1.0 recordio>=0.1.0 matplotlib==2.2.3 # TODO: let python3 paddlepaddle package use latest matplotlib rarfile diff --git a/python/setup.py.in b/python/setup.py.in index f93f0cd13..a7c1e91f9 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -100,6 +100,7 @@ packages=['paddle', 'paddle.utils', 'paddle.dataset', 'paddle.reader', + 'paddle.distributed', 'paddle.fluid', 'paddle.fluid.imperative', 'paddle.fluid.proto', diff --git a/tools/manylinux1/build_scripts/build.sh b/tools/manylinux1/build_scripts/build.sh index 6c551eceb..9f0452097 100644 --- a/tools/manylinux1/build_scripts/build.sh +++ b/tools/manylinux1/build_scripts/build.sh @@ -17,7 +17,6 @@ OPENSSL_ROOT=openssl-1.1.0i OPENSSL_HASH=ebbfc844a8c8cc0ea5dc10b86c9ce97f401837f3fa08c17b2cdadc118253cf99 EPEL_RPM_HASH=e5ed9ecf22d0c4279e92075a64c757ad2b38049bcf5c16c4f2b75d5f6860dc0d DEVTOOLS_HASH=a8ebeb4bed624700f727179e6ef771dafe47651131a00a78b342251415646acc -PATCHELF_HASH=d9afdff4baeacfbc64861454f368b7f2c15c44d245293f7587bbf726bfe722fb CURL_ROOT=curl-7.49.1 CURL_HASH=eb63cec4bef692eab9db459033f409533e6d10e20942f4b060b32819e81885f1 AUTOCONF_ROOT=autoconf-2.69 @@ -107,11 +106,7 @@ curl-config --features rm -rf /usr/local/ssl # Install patchelf (latest with unreleased bug fixes) -curl -sLO http://nipy.bic.berkeley.edu/manylinux/patchelf-0.9njs2.tar.gz -check_sha256sum patchelf-0.9njs2.tar.gz $PATCHELF_HASH -tar -xzf patchelf-0.9njs2.tar.gz -(cd patchelf-0.9njs2 && ./configure && make && make install) -rm -rf patchelf-0.9njs2.tar.gz patchelf-0.9njs2 +yum install -y patchelf # Install latest pypi release of auditwheel LD_LIBRARY_PATH="${ORIGINAL_LD_LIBRARY_PATH}:$(dirname ${PY35_BIN})/lib" $PY35_BIN/pip install auditwheel diff --git a/tools/run_mp.py b/tools/run_mp.py deleted file mode 100644 index 2485400ab..000000000 --- a/tools/run_mp.py +++ /dev/null @@ -1,129 +0,0 @@ -# Copyright (c) 2018 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. - -from __future__ import print_function - -import subprocess -import os -import sys -import time -import argparse - -default_envs = { - "PADDLE_TRAINER_ENDPOINTS": - "127.0.0.1:6170,127.0.0.1:6171,127.0.0.1:6172,127.0.0.1:6173,127.0.0.1:6174,127.0.0.1:6175,127.0.0.1:6176,127.0.0.1:6177", - "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH", ""), - "PATH": os.getenv("PATH"), - "LD_PRELOAD": os.getenv("LD_PRELOAD", ""), - "PADDLE_TRAINERS_NUM": "8", - "NCCL_DEBUG": "INFO", - "GLOG_v": "0", - "NCCL_SOCKET_IFNAME": "eth0", - "NCCL_IB_GID_INDEX": "3", - "NCCL_IB_RETRY_CNT": "0", -} - -GPUS = 8 - - -def start_procs(gpus, cmd, log_dir): - procs = [] - log_fns = [] - os.system("mkdir -p %s" % log_dir) - # ======== update parent envs ======= - for k, v in os.environ.items(): - if k.startswith("FLAGS_") or k.startswith("NCCL_") or \ - k.startswith("GLOG_"): - default_envs[k] = v - - # ======== for dist training ======= - node_trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) - current_ip = os.getenv("POD_IP", "127.0.0.1") - trainer_ips = os.getenv("PADDLE_TRAINERS", current_ip).split(",") - num_nodes = len(trainer_ips) - all_nodes_devices_endpoints = "" - for n in trainer_ips: - for i in range(gpus): - if all_nodes_devices_endpoints: - all_nodes_devices_endpoints += "," - all_nodes_devices_endpoints += "%s:617%d" % (n, i) - nranks = num_nodes * gpus - # ======== for dist training ======= - - for i in range(gpus): - curr_env = {} - curr_env.update(default_envs) - curr_env.update({ - "FLAGS_selected_gpus": "%d" % i, - "PADDLE_TRAINER_ID": "%d" % (node_trainer_id * gpus + i), - "PADDLE_CURRENT_ENDPOINT": "%s:617%d" % (current_ip, i), - # nranks - "PADDLE_TRAINERS_NUM": "%d" % nranks, - "PADDLE_TRAINER_ENDPOINTS": all_nodes_devices_endpoints - }) - - print("starting process ", i, cmd, curr_env) - fn = open("%s/workerlog.%d" % (log_dir, i), "w") - log_fns.append(fn) - procs.append( - subprocess.Popen( - cmd.strip().split(" "), stdout=fn, stderr=fn, env=curr_env)) - - for i in range(gpus): - try: - procs[i].communicate() - procs[i].terminate() - log_fns[i].close() - except: - pass - - -def main(): - parser = argparse.ArgumentParser( - description='''start paddle training using multi-process mode. -NOTE: your train program ***must*** run as distributed nccl2 mode, -see: http://www.paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html#permalink-8--nccl2- -And your train program must read environment variables below in order to let different -process init properly: -FLAGS_selected_gpus -PADDLE_TRAINER_ID -PADDLE_CURRENT_ENDPOINT -PADDLE_TRAINERS_NUM -PADDLE_TRAINER_ENDPOINTS -POD_IP (current node ip address, not needed for local training) -''') - parser.add_argument( - '--gpus', - type=int, - default=8, - help='start number of processes for every gpu') - parser.add_argument( - '--cmd', - type=str, - default="", - help='command to run for each process, e.g. python train.py --lr 0.1') - parser.add_argument( - '--log_dir', - type=str, - default="mylog", - help='directory to put logs per process.') - args = parser.parse_args() - if args.cmd == "": - parser.print_help() - exit(0) - start_procs(args.gpus, args.cmd, args.log_dir) - - -if __name__ == "__main__": - main() -- GitLab