Merge pull request #12101 from panyx0718/ir

Initial IR change

Author: panyx0718

doc/fluid/design/ir/draft.md

@@ -0,0 +1,89 @@
+## Motivation
+
+There is a ```gap``` between the ```Program``` defined by
+user and the ```Executable``` that can be scheduled
+efficiently on heterogeneous hardware, either locally
+or distributedly.
+
+Usually, the ```gap``` is bridged by
+
+* A serious transformations with defined order.
+
+* These transformations usually involve
+```insert, delete, clustering, split, dependency analysis```.
+
+* Has a simple way to verify and debug each transformation.
+
+* Flexible to add, remove or customize transformations to fit
+the requirements of various algorithms (models) and hardware secenarios.
+
+Some other events also push us to a better unified pattern.
+
+* The deep learning framework is built around the concepts of graphs.
+To leverage tools such as compilation (e.g. TVM and nGraph) or
+cross-framework conversion (e.g. ONNX), we also need a intermediate
+representation that can be connected to the rest of the ecosystem.
+
+
+We need a unified pattern to naturally support the requirements
+described above. The pattern should fit both training, inference
+and other offline serielized model transformations.
+Learned from LLVM and other deep learning framework, we draft the
+design below.
+
+
+## Design
+
+### Major Concepts
+
+#### Node
+
+```Node``` represents an operation that performs some computation or
+a variable that is input or output of operation.
+
+```Node```s are connected to other ```Node```s via inputs and outputs.
+
+Other properties (maybe device placement information) can be added
+to ```Node``` in the future if it's a
+common requirement of many other ```Pass```es. Otherwise, it should live
+in a ```Node``` wrapper class that is private to some ```Pass``` or be
+a local member of a ```Pass```.
+
+#### Graph
+
+```Graph``` contains a list of ```Node```s, which are connected to
+each other via inputs and outputs.
+
+TODO: Better definitions for the graph.
+
+```Graph``` can also contain ```Attribute```s. ```Attribute```s
+can be ``any`` thing. For example, it can be a list of "wraper"
+nodes. The ```wrapper``` nodes compose ```Node```s and provide
+helper method for execution or transformation. ```Attribute```
+can also contain other things that describe some properties of
+the ```Graph``` or ```Graph``` nodes. ```Attribute``` can be passed
+across ```Pass```. However, it should be used with care.
+
+#### Pass
+
+```Pass``` represents a transformation of ```Graph```. Its input
+is a ```Graph``` and its output is also a ```Graph```. For example,
+a ```Pass``` can simply print out the ```Graph```. A ```Pass```
+can also fuse some ```Graph```'s ```Node```s.
+
+#### Optimize
+
+```Optimize``` contains a series of ```Pass``` with defined order.
+```Optimize``` transforms a ```Graph``` that only contains raw
+modeling logic to a ```Graph``` that can be run efficiently while
+maintaining the original modeling logic.
+
+
+### Optimize Process
+
+* Program is first converted to Graph.
+* Graph goes through a series of Pass
+* Graph is transformed from raw model logic to a
+form that is efficient to execute.
+
+Program->ProgramToGraph->Graph->Pass1->Graph->Pass2->Graph->Pass3->Graph->Executor

paddle/fluid/framework/CMakeLists.txt

@@ -1,4 +1,5 @@
 add_subdirectory(details)
+add_subdirectory(ir)
 # ddim lib
 proto_library(framework_proto SRCS framework.proto)
 
@@ -93,7 +94,7 @@ else()
 endif()
 
 
-cc_library(parallel_executor SRCS parallel_executor.cc DEPS ssa_graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor)
+cc_library(parallel_executor SRCS parallel_executor.cc DEPS ssa_graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph)
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)

paddle/fluid/framework/details/CMakeLists.txt

@@ -5,8 +5,7 @@ cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod
 cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
 cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
 
-cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base)
-cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph)
+cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS graph)
 cc_library(ssa_graph_printer SRCS ssa_graph_printer.cc DEPS ssa_graph_builder)
 cc_library(ssa_graph_checker SRCS ssa_graph_checker.cc DEPS ssa_graph_builder)
 
@@ -35,7 +34,7 @@ cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS
 
 cc_library(ssa_graph_builder_factory SRCS ssa_graph_builder_factory.cc DEPS multi_devices_graph_builder ssa_graph_printer ssa_graph_checker)
 
-cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framework_proto)
+cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
         simple_threadpool device_context)
 

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -23,20 +23,25 @@ namespace framework {
 namespace details {
 
 #ifdef PADDLE_WITH_CUDA
-AllReduceOpHandle::AllReduceOpHandle(const std::vector<Scope *> &local_scopes,
+AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
+                                     const std::vector<Scope *> &local_scopes,
                                      const std::vector<platform::Place> &places,
                                      const platform::NCCLContextMap *ctxs)
-    : local_scopes_(local_scopes), places_(places), nccl_ctxs_(ctxs) {
+    : OpHandleBase(node),
+      local_scopes_(local_scopes),
+      places_(places),
+      nccl_ctxs_(ctxs) {
   if (nccl_ctxs_) {
     for (auto &p : places_) {
       this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p);
     }
   }
 }
 #else
-AllReduceOpHandle::AllReduceOpHandle(const std::vector<Scope *> &local_scopes,
+AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
+                                     const std::vector<Scope *> &local_scopes,
                                      const std::vector<platform::Place> &places)
-    : local_scopes_(local_scopes), places_(places) {}
+    : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
 #endif
 
 void AllReduceOpHandle::RunImpl() {

paddle/fluid/framework/details/all_reduce_op_handle.h

@@ -30,11 +30,11 @@ namespace details {
 
 struct AllReduceOpHandle : public OpHandleBase {
 #ifdef PADDLE_WITH_CUDA
-  AllReduceOpHandle(const std::vector<Scope *> &local_scopes,
+  AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places,
                     const platform::NCCLContextMap *ctxs);
 #else
-  AllReduceOpHandle(const std::vector<Scope *> &local_scopes,
+  AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places);
 #endif
   std::string Name() const override;

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -35,20 +35,23 @@ namespace details {
 struct BroadcastOpHandle : public OpHandleBase {
  public:
 #ifdef PADDLE_WITH_CUDA
-  BroadcastOpHandle(const std::vector<Scope *> &local_scopes,
+  BroadcastOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places,
                     const platform::NCCLContextMap *nccl_ctxs)
-      : local_scopes_(local_scopes), places_(places), nccl_ctxs_(nccl_ctxs) {
+      : OpHandleBase(node),
+        local_scopes_(local_scopes),
+        places_(places),
+        nccl_ctxs_(nccl_ctxs) {
     if (nccl_ctxs_) {
       for (auto &p_ctx : nccl_ctxs_->contexts_) {
         dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get();
       }
     }
   }
 #else
-  BroadcastOpHandle(const std::vector<Scope *> &local_scopes,
+  BroadcastOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places)
-      : local_scopes_(local_scopes), places_(places) {}
+      : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
 #endif
 
   std::string Name() const override;

paddle/fluid/framework/details/broadcast_op_handle_test.cc

@@ -96,48 +96,61 @@ struct TestBroadcastOpHandle {
     }
     param_scopes_[input_scope_idx]->Var("input");
 
+    std::unique_ptr<ir::Node> n(
+        new ir::Node("node0", ir::Node::Type::kOperation));
     if (use_gpu_) {
 #ifdef PADDLE_WITH_CUDA
-      op_handle_.reset(
-          new BroadcastOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get()));
+      op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_,
+                                             nccl_ctxs_.get()));
 #else
       PADDLE_THROW("CUDA is not support.");
 #endif
     } else {
 #ifdef PADDLE_WITH_CUDA
-      op_handle_.reset(
-          new BroadcastOpHandle(local_scopes_, gpu_list_, nccl_ctxs_.get()));
+      op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_,
+                                             nccl_ctxs_.get()));
 #else
-      op_handle_.reset(new BroadcastOpHandle(local_scopes_, gpu_list_));
+      op_handle_.reset(
+          new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_));
 #endif
     }
 
-    auto* in_var_handle =
-        new VarHandle(1, input_scope_idx, "input", gpu_list_[input_scope_idx]);
+    std::unique_ptr<ir::Node> v(
+        new ir::Node("node1", ir::Node::Type::kVariable));
+    auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input",
+                                        gpu_list_[input_scope_idx]);
     vars_.emplace_back(in_var_handle);
     op_handle_->AddInput(in_var_handle);
 
     // add dummy var
-    vars_.emplace_back(new DummyVarHandle());
+
+    std::unique_ptr<ir::Node> v2(
+        new ir::Node("node2", ir::Node::Type::kVariable));
+    vars_.emplace_back(new DummyVarHandle(v2.get()));
     DummyVarHandle* dummy_var_handle =
         static_cast<DummyVarHandle*>(vars_.back().get());
-    dummy_var_handle->generated_op_ = nullptr;
+    dummy_var_handle->ClearGeneratedOp();
     op_handle_->AddInput(dummy_var_handle);
 
     for (size_t j = 0; j < gpu_list_.size(); ++j) {
       if (!use_gpu_) {
         op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get());
       }
-      VarHandle* out_var_handle = new VarHandle(2, j, "out", gpu_list_[j]);
+      std::unique_ptr<ir::Node> v3(
+          new ir::Node("node3", ir::Node::Type::kVariable));
+      VarHandle* out_var_handle =
+          new VarHandle(v3.get(), 2, j, "out", gpu_list_[j]);
       vars_.emplace_back(out_var_handle);
       op_handle_->AddOutput(out_var_handle);
     }
 
     // add dummy var
-    vars_.emplace_back(new DummyVarHandle());
+    std::unique_ptr<ir::Node> v4(
+        new ir::Node("node4", ir::Node::Type::kVariable));
+    vars_.emplace_back(new DummyVarHandle(v4.get()));
     DummyVarHandle* out_dummy_var_handle =
         static_cast<DummyVarHandle*>(vars_.back().get());
-    out_dummy_var_handle->generated_op_ = nullptr;
+    out_dummy_var_handle->ClearGeneratedOp();
     op_handle_->AddOutput(out_dummy_var_handle);
   }
 

paddle/fluid/framework/details/computation_op_handle.cc

@@ -19,9 +19,10 @@
 namespace paddle {
 namespace framework {
 namespace details {
-ComputationOpHandle::ComputationOpHandle(const OpDesc &op_desc, Scope *scope,
+ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
                                          platform::Place place)
-    : op_(framework::OpRegistry::CreateOp(op_desc)),
+    : OpHandleBase(node),
+      op_(framework::OpRegistry::CreateOp(*node->Op())),
       scope_(scope),
       place_(place) {}
 
@@ -35,8 +36,8 @@ void ComputationOpHandle::RunImpl() {
 
 bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) {
   bool need_wait =
-      in_var && in_var->generated_op_ &&
-      in_var->generated_op_->DeviceContext(place_) != dev_ctxes_[place_];
+      in_var && in_var->GeneratedOp() &&
+      in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_[place_];
   return need_wait;
 }
 

paddle/fluid/framework/details/computation_op_handle.h

@@ -28,8 +28,7 @@ namespace framework {
 namespace details {
 struct ComputationOpHandle : public OpHandleBase {
  public:
-  ComputationOpHandle(const OpDesc &op_desc, Scope *scope,
-                      platform::Place place);
+  ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place);
 
   std::string Name() const override;
 

paddle/fluid/framework/details/data_balance_op_handle.cc

@@ -22,10 +22,10 @@ namespace details {
 
 #ifdef PADDLE_WITH_CUDA
 DataBalanceOpHandle::DataBalanceOpHandle(
-    const std::vector<Scope *> &local_scopes,
+    ir::Node *node, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,
     const platform::NCCLContextMap *ctxs)
-    : local_scopes_(local_scopes), places_(places) {
+    : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {
   if (ctxs) {
     for (auto &p : places_) {
       this->dev_ctxes_[p] = ctxs->DevCtx(p);
@@ -34,9 +34,9 @@ DataBalanceOpHandle::DataBalanceOpHandle(
 }
 #else
 DataBalanceOpHandle::DataBalanceOpHandle(
-    const std::vector<Scope *> &local_scopes,
+    ir::Node *node, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places)
-    : local_scopes_(local_scopes), places_(places) {}
+    : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
 #endif
 
 std::string DataBalanceOpHandle::Name() const { return "data balance"; }