[incubator-mxnet] branch master updated: fix engine crash in shutdown phase (#14382)

[wk...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

wkcn pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git


The following commit(s) were added to refs/heads/master by this push:
     new 4f5cba5  fix engine crash in shutdown phase (#14382)
4f5cba5 is described below

commit 4f5cba59dd44b91bfdd59025106e3bc116600a86
Author: Wang Jiajun <wa...@gmail.com>
AuthorDate: Mon Mar 11 11:56:53 2019 +0800

    fix engine crash in shutdown phase (#14382)
    
    * fix engine crash in shutdown phase
    
    * fix lint
    
    * Revert "Bypass ThreadedEngine in test_operator_gpu.py:test_convolution_multiple_streams. (#14338)"
    
    This reverts commit d6eafca2555b58746f51052fdce96a264d02a84a.
---
 src/engine/threaded_engine.h          |  9 +++++++++
 tests/python/gpu/test_operator_gpu.py | 12 +-----------
 2 files changed, 10 insertions(+), 11 deletions(-)

diff --git a/src/engine/threaded_engine.h b/src/engine/threaded_engine.h
index ab06ca1..640eac4 100644
--- a/src/engine/threaded_engine.h
+++ b/src/engine/threaded_engine.h
@@ -30,6 +30,7 @@
 #include <dmlc/base.h>
 #include <dmlc/logging.h>
 #include <dmlc/omp.h>
+#include <mxnet/storage.h>
 #include <vector>
 #include <functional>
 #include <condition_variable>
@@ -306,6 +307,8 @@ class ThreadedEngine : public Engine {
     objpool_varblk_ref_ = common::ObjectPool<VersionedVarBlock>::_GetSharedRef();
     objpool_var_ref_    = common::ObjectPool<ThreadedVar>::_GetSharedRef();
 
+    storage_ref_ = Storage::_GetSharedRef();
+
     // Get a ref to the profiler so that it doesn't get killed before us
     profiler::Profiler::Get(&profiler_);
   }
@@ -549,6 +552,12 @@ class ThreadedEngine : public Engine {
   std::shared_ptr<common::ObjectPool<VersionedVarBlock> > objpool_varblk_ref_;
   std::shared_ptr<common::ObjectPool<ThreadedVar> >       objpool_var_ref_;
 
+  /*!
+   * \brief Async destruction of some objects is relied on storage,
+   *  prevent it from being destructed too early
+   */
+  std::shared_ptr<Storage> storage_ref_;
+
 #if MXNET_USE_CUDA
   /*! \brief Number of GPU devices available */
   std::atomic<int> device_count_{-1};
diff --git a/tests/python/gpu/test_operator_gpu.py b/tests/python/gpu/test_operator_gpu.py
index 7d7c2ed..c12c94b 100644
--- a/tests/python/gpu/test_operator_gpu.py
+++ b/tests/python/gpu/test_operator_gpu.py
@@ -547,18 +547,8 @@ def _conv_with_num_streams(seed):
 
 @with_seed()
 def test_convolution_multiple_streams():
-    engines = ['NaiveEngine', 'ThreadedEngine', 'ThreadedEnginePerDevice']
-
-    if os.getenv('MXNET_ENGINE_TYPE') is not None:
-        engines = [os.getenv('MXNET_ENGINE_TYPE'),]
-        print("Only running against '%s'" % engines[0], file=sys.stderr, end='')
-    # Remove this else clause when the ThreadedEngine can handle this test
-    else:
-        engines.remove('ThreadedEngine')
-        print("SKIP: 'ThreadedEngine', only running against %s" % engines, file=sys.stderr, end='')
-
     for num_streams in [1, 2]:
-        for engine in engines:
+        for engine in ['NaiveEngine', 'ThreadedEngine', 'ThreadedEnginePerDevice']:
             print("Starting engine %s with %d streams." % (engine, num_streams), file=sys.stderr)
             run_in_spawned_process(_conv_with_num_streams,
                 {'MXNET_GPU_WORKER_NSTREAMS' : num_streams, 'MXNET_ENGINE_TYPE' : engine})