test(Pe): add dry run tests for pe (#14254)

Dry run tests will skip `Op.Run` and just perform job scheduling. It helps to analysis dead lock in PE.

test=develop

paddle/fluid/framework/details/execution_strategy.h

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #pragma once
+#include <cstddef>  // for size_t
 
 namespace paddle {
 namespace framework {
@@ -26,6 +27,7 @@ struct ExecutionStrategy {
   bool allow_op_delay_{false};
   size_t num_iteration_per_drop_scope_{100};
   ExecutorType type_{kDefault};
+  bool dry_run_{false};
 };
 
 }  //  namespace details

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -128,7 +128,9 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
     size_t complete = 0;
     while (op_to_run != nullptr) {
       try {
-        op_to_run->Run(strategy_.use_cuda_);
+        if (LIKELY(!strategy_.dry_run_)) {
+          op_to_run->Run(strategy_.use_cuda_);
+        }
         ++complete;
       } catch (...) {
         exception_.Catch(std::current_exception());

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -211,7 +211,9 @@ void ThreadedSSAGraphExecutor::RunOp(
       if (VLOG_IS_ON(10)) {
         VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
       }
-      op->Run(strategy_.use_cuda_);
+      if (LIKELY(!strategy_.dry_run_)) {
+        op->Run(strategy_.use_cuda_);
+      }
       VLOG(10) << op << " " << op->Name() << " Done ";
       running_ops_--;
       ready_var_q->Extend(op->Outputs());

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -48,7 +48,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   // Use topological sort algorithm
   FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
 
-  ~ThreadedSSAGraphExecutor() {}
+  ~ThreadedSSAGraphExecutor() final = default;
 
  private:
   void RunOp(const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,

paddle/fluid/framework/parallel_executor.cc

@@ -38,9 +38,20 @@ class ParallelExecutorPrivate {
   explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
       : places_(places) {}
 
+  ~ParallelExecutorPrivate() {
+    if (own_local_scope_) {
+      for (size_t i = 1; i < local_scopes_.size(); ++i) {
+        // Skip the first scope, since it is the global scope.
+        Scope *local_scope = local_scopes_[i];
+        if (global_scope_->HasKid(local_scope)) {
+          global_scope_->DeleteScope(local_scope);
+        }
+      }
+    }
+  }
   std::vector<platform::Place> places_;
   std::vector<Scope *> local_scopes_;
-  Scope *global_scope_;
+  Scope *global_scope_;  // not owned
   std::unique_ptr<details::SSAGraphExecutor> executor_;
 
 #ifdef PADDLE_WITH_CUDA
@@ -306,16 +317,6 @@ ParallelExecutor::~ParallelExecutor() {
   for (auto &p : member_->places_) {
     platform::DeviceContextPool::Instance().Get(p)->Wait();
   }
-
-  if (member_->own_local_scope_) {
-    for (size_t i = 1; i < member_->local_scopes_.size(); ++i) {
-      Scope *local_scope = member_->local_scopes_[i];
-      if (member_->global_scope_->HasKid(local_scope)) {
-        member_->global_scope_->DeleteScope(local_scope);
-      }
-    }
-  }
-
   // member_ must be destructed before gcs_ since the destructor of
   // ReferenceCountOpHandle use raw pointers of gcs_ inside.
   member_.reset();

paddle/fluid/pybind/pybind.cc

@@ -742,7 +742,12 @@ All parameter, weight, gradient are variables in Paddle.
                        will clean up the temp variables at the end of the current iteration.
                     2. In some NLP model, it may cause the GPU memory is insufficient,
                        in this case, you should reduce `num_iteration_per_drop_scope`.
-              )DOC");
+              )DOC")
+      .def_property("_dry_run",
+                    [](const ExecutionStrategy &self) { return self.dry_run_; },
+                    [](ExecutionStrategy &self, bool dry_run) {
+                      self.dry_run_ = dry_run;
+                    });
 
   exec_strategy.def_property(
       "use_experimental_executor",

python/paddle/fluid/layers/io.py

@@ -60,7 +60,7 @@ def data(name,
             For example if shape=[1], the resulting shape is [-1, 1].
           2. If shape contains -1, such as shape=[1, -1],
             append_batch_size will be enforced to be be False (ineffective).
-       dtype(int|float): The type of data : float32, float_16, int etc
+       dtype(basestring): The type of data : float32, float_16, int etc
        type(VarType): The output type. By default it is LOD_TENSOR.
        lod_level(int): The LoD Level. 0 means the input data is not a sequence.
        stop_gradient(bool): A boolean that mentions whether gradient should flow.

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

+# 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.
+
+import paddle.fluid as fluid
+import unittest
+import logging
+import six
+
+
+class TestBase(unittest.TestCase):
+    def main(self,
+             network_func,
+             iter=100,
+             iter_per_pe=100,
+             use_gpu=True,
+             use_experimental_executor=False):
+        if use_gpu and not fluid.core.is_compiled_with_cuda():
+            logging.warning(
+                "Paddle is not compiled with CUDA, skip GPU unittests")
+            return
+
+        main_prog = fluid.Program()
+        startup_prog = fluid.Program()
+        scope = fluid.Scope()
+        with fluid.program_guard(main_prog, startup_prog):
+            with fluid.scope_guard(scope):
+                loss = network_func()
+                fluid.Executor(
+                    fluid.CUDAPlace(0)
+                    if use_gpu else fluid.CPUPlace()).run(startup_prog)
+
+        for _ in six.moves.xrange(iter):
+            exe_strategy = fluid.ExecutionStrategy()
+            exe_strategy._dry_run = True
+            exe_strategy.use_experimental_executor = use_experimental_executor
+            pe = fluid.ParallelExecutor(
+                use_cuda=True,
+                loss_name=loss.name,
+                main_program=main_prog,
+                exec_strategy=exe_strategy)
+            for _ in six.moves.xrange(iter_per_pe):
+                pe.run([])
+
+
+class TestMNISTDryRun(TestBase):
+    def test_mnist_dry_run(self):
+        for use_gpu in (False, True):
+            for use_experimental_executor in (False, True):
+                self.main(
+                    network_func=TestMNISTDryRun.network_func,
+                    use_gpu=use_gpu,
+                    use_experimental_executor=use_experimental_executor)
+
+    @staticmethod
+    def network_func():
+        img = fluid.layers.data(name='img', shape=[784], dtype='float32')
+        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        hidden = img
+        for _ in six.moves.xrange(10):
+            hidden = fluid.layers.fc(input=img, size=200, act='tanh')
+        prediction = fluid.layers.fc(input=hidden, size=10, act='softmax')
+        loss = fluid.layers.cross_entropy(input=prediction, label=label)
+        avg_loss = fluid.layers.mean(loss)
+        fluid.optimizer.Adam().minimize(avg_loss)
+        return avg_loss
+
+
+if __name__ == '__main__':
+    unittest.main()