Merge pull request #14807 from baojun-nervana/ngraph_follow_up

Follow ups on nGraph PR comments

paddle/fluid/framework/CMakeLists.txt

@@ -129,11 +129,13 @@ cc_test(version_test SRCS version_test.cc DEPS version)
 
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
 
-if(NOT WIN32)
-cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
-cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
-  shape_inference data_transform lod_tensor profiler)
-endif(NOT WIN32)
+if(WITH_NGRAPH)
+  if(NOT WIN32)
+    cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
+    cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
+      shape_inference data_transform lod_tensor profiler ngraph)
+  endif(NOT WIN32)
+endif(WITH_NGRAPH)
 
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
 nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
@@ -169,11 +171,15 @@ if(WITH_DISTRIBUTE)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 else()
+  if(WITH_NGRAPH)
     if(NOT WIN32)
-    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator variable_helper)
+      cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph ngraph_operator variable_helper)
     else(NOT WIN32)
       cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
     endif(NOT WIN32)
+  else(WITH_NGRAPH)
+    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
+  endif(WITH_NGRAPH)
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 

paddle/fluid/framework/executor.cc

@@ -17,7 +17,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/framework/lod_rank_table.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
-#include "paddle/fluid/framework/ngraph_operator.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/transfer_scope_cache.h"
@@ -26,6 +25,10 @@ limitations under the License. */
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
 
+#ifdef PADDLE_WITH_NGRAPH
+#include "paddle/fluid/framework/ngraph_operator.h"
+#endif
+
 DECLARE_bool(benchmark);
 DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
 DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
@@ -88,11 +91,11 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
 static void EnableFusedOp(ExecutorPrepareContext* ctx) {
 #ifdef PADDLE_WITH_NGRAPH
   VLOG(3) << "use_ngraph=True";
-  auto intervals = FusedOperator::FusedOpIntervals(&ctx->ops_);
+  auto intervals = NgraphOperator::NgraphOpIntervals(&ctx->ops_);
   for (auto& interval : intervals) {
-    auto* fused_op = new FusedOperator(ctx->prog_, ctx->block_id_,
-                                       interval.at(0), interval.at(1));
-    *interval[0] = std::unique_ptr<OperatorBase>(fused_op);
+    auto* ng_op = new NgraphOperator(ctx->prog_, ctx->block_id_, interval.at(0),
+                                     interval.at(1));
+    *interval[0] = std::unique_ptr<OperatorBase>(ng_op);
   }
   for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) {
     ctx->ops_.erase(it->at(0) + 1, it->at(1));

paddle/fluid/framework/ngraph_bridge.cc

@@ -12,7 +12,6 @@ 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. */
 
-#ifdef PADDLE_WITH_NGRAPH
 #include <algorithm>
 #include <functional>
 #include <vector>
@@ -27,14 +26,15 @@ namespace paddle {
 namespace framework {
 
 static std::shared_ptr<ngraph::Node> GetNode(
-    const std::shared_ptr<OperatorBase>& op, const std::string prm,
+    const std::shared_ptr<OperatorBase>& op, const std::string name,
     const VariableNameMap& var_map,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  auto& var_names = var_map.at(prm);
+  auto& var_names = var_map.at(name);
   PADDLE_ENFORCE_EQ(var_names.size(), 1,
-                    "op %s prm %s expects one associated var", op->Type(), prm);
+                    "op %s name %s expects one associated var", op->Type(),
+                    name);
   if (ngb_node_map->find(var_names[0]) != ngb_node_map->end()) {
     return (*ngb_node_map)[var_names[0]];
   } else {
@@ -43,42 +43,42 @@ static std::shared_ptr<ngraph::Node> GetNode(
 }
 
 static std::shared_ptr<ngraph::Node> GetInputNode(
-    const std::shared_ptr<OperatorBase>& op, const std::string prm,
+    const std::shared_ptr<OperatorBase>& op, const std::string name,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  return GetNode(op, prm, op->Inputs(), ngb_node_map);
+  return GetNode(op, name, op->Inputs(), ngb_node_map);
 }
 
 static std::shared_ptr<ngraph::Node> GetOutputNode(
-    const std::shared_ptr<OperatorBase>& op, const std::string prm,
+    const std::shared_ptr<OperatorBase>& op, const std::string name,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  return GetNode(op, prm, op->Outputs(), ngb_node_map);
+  return GetNode(op, name, op->Outputs(), ngb_node_map);
 }
 
 static void SetOutputNode(
-    const std::shared_ptr<OperatorBase>& op, const std::string prm,
+    const std::shared_ptr<OperatorBase>& op, const std::string name,
     std::shared_ptr<ngraph::Node> node,
     std::shared_ptr<
         std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
         ngb_node_map) {
-  auto& var_names = op->Outputs().at(prm);
+  auto& var_names = op->Outputs().at(name);
   if (var_names.size() == 1) {
     (*ngb_node_map)[var_names[0]] = node;
   } else if (var_names.size() == 0) {
     (*ngb_node_map)[""] = node;
   } else {
-    PADDLE_THROW("prm %s has more than 1 var_names.", prm);
+    PADDLE_THROW("name %s has more than 1 var_names.", name);
   }
 }
 
 static bool HasOutput(const std::shared_ptr<OperatorBase>& op,
-                      const std::string prm) {
+                      const std::string name) {
   auto& outputs = op->Outputs();
-  if (outputs.find(prm) == outputs.end()) return false;
-  return outputs.at(prm).size() > 0;
+  if (outputs.find(name) == outputs.end()) return false;
+  return outputs.at(name).size() > 0;
 }
 
 template <typename T>
@@ -118,4 +118,3 @@ void NgraphBridge::BuildNgNode(const std::shared_ptr<OperatorBase>& op) {
 
 }  // namespace framework
 }  // namespace paddle
-#endif

paddle/fluid/framework/ngraph_bridge.h

@@ -14,8 +14,6 @@ limitations under the License. */
 
 #pragma once
 
-#ifdef PADDLE_WITH_NGRAPH
-
 #include <algorithm>
 #include <map>
 #include <string>
@@ -53,4 +51,3 @@ class NgraphBridge {
 
 }  // namespace framework
 }  // namespace paddle
-#endif

paddle/fluid/framework/ngraph_operator.cc

@@ -12,7 +12,6 @@ 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. */
 
-#ifdef PADDLE_WITH_NGRAPH
 #include <glog/logging.h>
 
 #include <algorithm>
@@ -58,9 +57,9 @@ typedef enum {                /* nGraph support state on ops          */
 } op_state;
 
 // perform graph build through bridge and execute computation
-class NgraphOperator {
+class NgraphEngine {
  public:
-  explicit NgraphOperator(const Scope& scope, const platform::Place& place,
+  explicit NgraphEngine(const Scope& scope, const platform::Place& place,
                         const std::vector<std::shared_ptr<OperatorBase>>& ops,
                         const std::unordered_map<
                             std::string, ngraph::element::Type>& var_type_map,
@@ -132,7 +131,7 @@ class NgraphOperator {
 };
 
 std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
-FusedOperator::FusedOpIntervals(
+NgraphOperator::NgraphOpIntervals(
     std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops) {
   std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
       intervals;
@@ -185,7 +184,7 @@ FusedOperator::FusedOpIntervals(
   return intervals;
 }
 
-FusedOperator::FusedOperator(
+NgraphOperator::NgraphOperator(
     const ProgramDesc& prog, size_t block_id,
     std::vector<std::unique_ptr<OperatorBase>>::iterator start,
     std::vector<std::unique_ptr<OperatorBase>>::iterator end,
@@ -215,7 +214,7 @@ FusedOperator::FusedOperator(
   Process();
 }
 
-void FusedOperator::Process() {
+void NgraphOperator::Process() {
   auto& bdesc = pdesc_.Block(block_);
   for (auto& var : bdesc.AllVars()) {
     if (!(var->GetType() == proto::VarType::SELECTED_ROWS ||
@@ -251,7 +250,7 @@ void FusedOperator::Process() {
   }
 }
 
-void FusedOperator::RunImpl(const Scope& scope,
+void NgraphOperator::RunImpl(const Scope& scope,
                              const platform::Place& place) const {
   op_state ng_op_state = PARTIAL_TEST;
   auto& bdesc = pdesc_.Block(block_);
@@ -266,19 +265,19 @@ void FusedOperator::RunImpl(const Scope& scope,
     ng_op_state = ng_op_state == PARTIAL_TEST ? FULL_TEST : FULL_TRAIN;
   }
 
-  NgraphOperator ngraph_op(scope, place, fused_ops_, var_type_map_,
+  NgraphEngine ngraph_engine(scope, place, fused_ops_, var_type_map_,
                              persistables_, fetches_, post_op_inputs_,
                              ng_op_state);
-  ngraph_op.Run(scope, place);
+  ngraph_engine.Run(scope, place);
 }
 
 std::unordered_map<std::string, std::shared_ptr<ngraph::Function>>
-    NgraphOperator::func_cache_ = {};
+    NgraphEngine::func_cache_ = {};
 
-std::shared_ptr<ngraph::runtime::Backend> NgraphOperator::backend_ =
+std::shared_ptr<ngraph::runtime::Backend> NgraphEngine::backend_ =
     ngraph::runtime::Backend::create("CPU");
 
-void NgraphOperator::GetNgInputShape(std::shared_ptr<OperatorBase> op) {
+void NgraphEngine::GetNgInputShape(std::shared_ptr<OperatorBase> op) {
   op->RuntimeInferShape(scope_, place_);
   for (auto& var_name_item : op->Inputs()) {
     for (auto& var_name : var_name_item.second) {
@@ -301,7 +300,7 @@ void NgraphOperator::GetNgInputShape(std::shared_ptr<OperatorBase> op) {
   }
 }
 
-void NgraphOperator::BuildNgNodes() {
+void NgraphEngine::BuildNgNodes() {
   for (auto& var_name : var_out_) {
     if (var_node_map_->find(var_name) == var_node_map_->end()) {
       auto* var = scope_.FindVar(var_name);
@@ -323,7 +322,7 @@ void NgraphOperator::BuildNgNodes() {
   }
 }
 
-void NgraphOperator::BuildNgIO() {
+void NgraphEngine::BuildNgIO() {
   std::unordered_set<std::string> inputs;
   std::unordered_set<std::string> outputs;
 
@@ -395,7 +394,7 @@ void NgraphOperator::BuildNgIO() {
   }
 }
 
-void NgraphOperator::BuildNgFunction() {
+void NgraphEngine::BuildNgFunction() {
   BuildNgNodes();
   ngraph_function_ = nullptr;
   ngraph::NodeVector func_outputs;
@@ -416,7 +415,7 @@ void NgraphOperator::BuildNgFunction() {
       std::make_shared<ngraph::Function>(func_outputs, func_inputs);
 }
 
-std::shared_ptr<std::string> NgraphOperator::GetCacheKey() {
+std::shared_ptr<std::string> NgraphEngine::GetCacheKey() {
   auto cache_key = std::make_shared<std::string>("");
   *cache_key += std::to_string(fused_ops_.size());
   for (auto& op : fused_ops_) {
@@ -444,7 +443,7 @@ std::shared_ptr<std::string> NgraphOperator::GetCacheKey() {
   return cache_key;
 }
 
-void NgraphOperator::GetNgFunction() {
+void NgraphEngine::GetNgFunction() {
   bool cache_on = true;
   if (cache_on) {
     std::string cache_key_val = *GetCacheKey();
@@ -459,8 +458,7 @@ void NgraphOperator::GetNgFunction() {
   }
 }
 
-void NgraphOperator::Run(const Scope& scope,
-                         const platform::Place& place) const {
+void NgraphEngine::Run(const Scope& scope, const platform::Place& place) const {
   std::vector<std::shared_ptr<ngraph::runtime::Tensor>> t_in;
   std::vector<std::shared_ptr<ngraph::runtime::Tensor>> t_out;
 
@@ -545,7 +543,6 @@ void NgraphOperator::Run(const Scope& scope,
   }
 
   backend_->call(ngraph_function_, t_out, t_in);
-}  // NgraphOperator::RunImpl
+}  // NgraphEngine::RunImpl
 }  // namespace framework
 }  // namespace paddle
-#endif

paddle/fluid/framework/ngraph_operator.h

@@ -14,8 +14,6 @@ limitations under the License. */
 
 #pragma once
 
-#ifdef PADDLE_WITH_NGRAPH
-
 #include <algorithm>
 #include <string>
 #include <unordered_map>
@@ -34,14 +32,14 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class FusedOperator : public OperatorBase {
+class NgraphOperator : public OperatorBase {
  public:
   static std::vector<
       std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
-  FusedOpIntervals(
+  NgraphOpIntervals(
       std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops);
 
-  explicit FusedOperator(
+  explicit NgraphOperator(
       const ProgramDesc& prog, size_t block_id,
       std::vector<std::unique_ptr<OperatorBase>>::iterator start,
       std::vector<std::unique_ptr<OperatorBase>>::iterator end,
@@ -64,4 +62,3 @@ class FusedOperator : public OperatorBase {
 };
 }  // namespace framework
 }  // namespace paddle
-#endif