Add FastExecutor

paddle/fluid/framework/CMakeLists.txt

@@ -100,7 +100,11 @@ else()
 endif()
 
 
-cc_library(parallel_executor SRCS parallel_executor.cc DEPS threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph graph_viz_pass multi_devices_graph_pass multi_devices_graph_print_pass multi_devices_graph_check_pass)
+cc_library(parallel_executor SRCS parallel_executor.cc DEPS
+        threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
+        graph graph_viz_pass multi_devices_graph_pass
+        multi_devices_graph_print_pass multi_devices_graph_check_pass
+        fast_threaded_ssa_graph_executor)
 
 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/execution_strategy.h

@@ -19,10 +19,13 @@ namespace framework {
 namespace details {
 
 struct ExecutionStrategy {
+  enum ExecutorType { kDefault = 0, kExperimental = 1 };
+
   size_t num_threads_{0};
   bool use_cuda_{true};
   bool allow_op_delay_{false};
   size_t num_iteration_per_drop_scope_{100};
+  ExecutorType type_{kDefault};
 };
 
 }  //  namespace details

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -105,7 +105,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
       for (int i = 0; i < remaining; ++i) {
         complete_q.Pop();
       }
-      LOG(FATAL) << "On exception thrown, not implemented";
+      exception_.ReThrow();
     }
     num_complete += num_comp;
   }
@@ -127,6 +127,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
         op_to_run->Run(strategy_.use_cuda_);
         ++complete;
       } catch (...) {
+        exception_.Catch(std::current_exception());
         --remaining_;
         complete_q->Push(-1UL);
         return;
@@ -161,6 +162,8 @@ void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() {
         std::unordered_map<OpHandleBase *, std::atomic<int>>>(op_deps);
   });
 }
+
+const ir::Graph &FastThreadedSSAGraphExecutor::Graph() const { return *graph_; }
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h

@@ -34,6 +34,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
                                const std::vector<platform::Place> &places,
                                std::unique_ptr<ir::Graph> &&graph);
   FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
+  const ir::Graph &Graph() const override;
 
  private:
   ExecutionStrategy strategy_;
@@ -56,6 +57,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::future<
       std::unique_ptr<std::unordered_map<OpHandleBase *, std::atomic<int>>>>
       atomic_op_deps_;
+  ExceptionHolder exception_;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/parallel_executor.cc

@@ -25,6 +25,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
+#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
 #include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
@@ -193,8 +194,14 @@ ParallelExecutor::ParallelExecutor(
       member_->local_scopes_, member_->use_cuda_, build_strategy);
 #endif
 
+  if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
     member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
         exec_strategy, member_->local_scopes_, places, std::move(graph)));
+  } else {
+    member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
+        exec_strategy, member_->local_scopes_, places, std::move(graph)));
+  }
+
   member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
       exec_strategy, member_->local_scopes_, std::move(var_infos),
       member_->places_, std::move(member_->executor_)));

paddle/fluid/pybind/pybind.cc

@@ -593,8 +593,8 @@ All parameter, weight, gradient are variables in Paddle.
 
   // -- python binds for parallel executor.
   py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
-  py::class_<ExecutionStrategy>(pe, "ExecutionStrategy")
-      .def(py::init())
+  py::class_<ExecutionStrategy> exec_strategy(pe, "ExecutionStrategy");
+  exec_strategy.def(py::init())
       .def_property(
           "num_threads",
           [](const ExecutionStrategy &self) { return self.num_threads_; },
@@ -621,6 +621,15 @@ All parameter, weight, gradient are variables in Paddle.
           [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) {
             self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope;
           });
+  py::enum_<ExecutionStrategy::ExecutorType>(exec_strategy, "ExecutorType")
+      .value("Default", ExecutionStrategy::kDefault)
+      .value("Experimental", ExecutionStrategy::kExperimental);
+  exec_strategy.def_property(
+      "type", [](const ExecutionStrategy &self) { return self.type_; },
+      [](ExecutionStrategy &self, ExecutionStrategy::ExecutorType type) {
+        self.type_ = type;
+      });
+
   py::class_<BuildStrategy> build_strategy(pe, "BuildStrategy");
 
   py::enum_<BuildStrategy::ReduceStrategy>(build_strategy, "ReduceStrategy")

python/paddle/fluid/tests/unittests/parallel_executor_test_base.py

@@ -36,7 +36,8 @@ class TestParallelExecutorBase(unittest.TestCase):
                                   seed=None,
                                   use_parallel_executor=True,
                                   use_reduce=False,
-                                  optimizer=fluid.optimizer.Adam):
+                                  optimizer=fluid.optimizer.Adam,
+                                  use_fast_executor=False):
         def run_executor(exe, feed, fetch_list, program=None):
             if isinstance(exe, fluid.ParallelExecutor):
                 res = exe.run(fetch_list=fetch_list, feed=feed)
@@ -69,6 +70,8 @@ class TestParallelExecutorBase(unittest.TestCase):
             startup_exe.run(startup)
             exec_strategy = fluid.ExecutionStrategy()
             exec_strategy.allow_op_delay = allow_op_delay
+            if use_fast_executor:
+                exec_strategy.type = fluid.ExecutionStrategy.ExecutorType.Experimental
 
             build_strategy = fluid.BuildStrategy()
             build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce \

python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py

@@ -181,7 +181,9 @@ class TestMNIST(TestParallelExecutorBase):
             use_parallel_executor=True)
 
         self.assertAlmostEquals(
-            np.mean(parallel_first_loss), single_first_loss, delta=1e-6)
+            np.mean(parallel_first_loss),
+            single_first_loss,
+            delta=1e-6, )
         self.assertAlmostEquals(
             np.mean(parallel_last_loss), single_last_loss, delta=1e-6)
 
@@ -189,7 +191,7 @@ class TestMNIST(TestParallelExecutorBase):
         self.check_simple_fc_parallel_accuracy(True)
         self.check_simple_fc_parallel_accuracy(False)
 
-    def check_batchnorm_fc_convergence(self, use_cuda):
+    def check_batchnorm_fc_convergence(self, use_cuda, use_fast_executor):
         if use_cuda and not core.is_compiled_with_cuda():
             return
 
@@ -201,11 +203,13 @@ class TestMNIST(TestParallelExecutorBase):
             fc_with_batchnorm,
             feed_dict={"image": img,
                        "label": label},
-            use_cuda=use_cuda)
+            use_cuda=use_cuda,
+            use_fast_executor=use_fast_executor)
 
     def test_batchnorm_fc(self):
-        self.check_batchnorm_fc_convergence(True)
-        self.check_batchnorm_fc_convergence(False)
+        for use_cuda in (False, True):
+            for use_fast_executor in (False, True):
+                self.check_batchnorm_fc_convergence(use_cuda, use_fast_executor)
 
     def test_batchnorm_fc_with_new_strategy(self):
         # FIXME(zcd): close this test temporally.