fix custom operation in fork

docs/tutorials/gluon/customop.md

@@ -30,6 +30,7 @@ Custom operator in python is easy to develop and good for prototyping, but may h
 import numpy as np
 import mxnet as mx
 from mxnet import gluon, autograd
+import os
 ```
 
 ## Parameter-less operators
@@ -214,5 +215,24 @@ y = dense(x)
 print(y)
 ```
 
+## Using custom operators with fork
+In Linux systems, the default method in multiprocessing to create process is by using fork. If there are unfinished async custom operations when forking, the program will be blocked because of python GIL. Always use sync calls like `wait_to_read` or `waitall` before calling fork.
+
+```
+x = mx.nd.array([0, 1, 2, 3])
+y = mx.nd.Custom(x, op_type='sigmoid')
+# unfinished async sigmoid operation will cause blocking
+os.fork()
+```
+
+Correctly handling this will make mxnet depend upon libpython, so the workaround now is to ensure that all custom operations are executed before forking process.
+
+```
+x = mx.nd.array([0, 1, 2, 3])
+y = mx.nd.Custom(x, op_type='sigmoid')
+# force execution by reading y
+print(y.asnumpy())
+os.fork()
+```
 
 <!-- INSERT SOURCE DOWNLOAD BUTTONS -->

src/c_api/c_api.cc

@@ -111,6 +111,7 @@ int MXRandomSeedContext(int seed, int dev_type, int dev_id) {
 
 int MXNotifyShutdown() {
   API_BEGIN();
+  mxnet::op::custom::CustomOperator::Get()->Stop();
   Engine::Get()->NotifyShutdown();
   API_END();
 }

src/initialize.cc

@@ -26,6 +26,7 @@
 #include <dmlc/logging.h>
 #include <mxnet/engine.h>
 #include "./engine/openmp.h"
+#include "./operator/custom/custom-inl.h"
 #if MXNET_USE_OPENCV
 #include <opencv2/opencv.hpp>
 #endif  // MXNET_USE_OPENCV
@@ -53,12 +54,15 @@ class LibraryInitializer {
 
 // disable openmp for multithreaded workers
 #ifndef _WIN32
+    using op::custom::CustomOperator;
     pthread_atfork(
       []() {
+        CustomOperator::Get()->Stop();
         Engine::Get()->Stop();
       },
       []() {
         Engine::Get()->Start();
+        CustomOperator::Get()->Start();
       },
       []() {
         // Conservative thread management for multiprocess workers
@@ -71,6 +75,7 @@ class LibraryInitializer {
 #endif  // MXNET_USE_OPENCV
         engine::OpenMP::Get()->set_enabled(false);
         Engine::Get()->Start();
+        CustomOperator::Get()->Start();
       });
 #endif
   }

src/operator/custom/custom-inl.h

@@ -136,7 +136,21 @@ class CustomOperator {
     cv_.notify_all();
   }
 
-  ~CustomOperator() {
+  static CustomOperator* Get() {
+    static CustomOperator inst;
+    return &inst;
+  }
+
+  void Start() {
+    num_free_threads = 0;
+    destructing_ = false;
+    naive_engine_ = true;
+    if (std::string("NaiveEngine") != dmlc::GetEnv("MXNET_ENGINE_TYPE", std::string())) {
+      naive_engine_ = false;
+    }
+  }
+
+  void Stop() {
     if (naive_engine_) return;
     {
       std::unique_lock<std::mutex> lock(mutex_);
@@ -145,17 +159,12 @@ class CustomOperator {
     }
     for (auto &worker : workers_)
       worker.join();
+    workers_.clear();
   }
 
-  static CustomOperator* Get();
-
  private:
-  CustomOperator() : num_free_threads(0) {
-    destructing_ = false;
-    naive_engine_ = true;
-    if (std::string("NaiveEngine") != dmlc::GetEnv("MXNET_ENGINE_TYPE", std::string())) {
-      naive_engine_ = false;
-    }
+  CustomOperator() {
+    this->Start();
   }
   void ThreadTarget() {
     std::unique_lock<std::mutex> lock(mutex_);

src/operator/custom/custom.cc

@@ -34,11 +34,6 @@ namespace mxnet {
 namespace op {
 namespace custom {
 
-CustomOperator* CustomOperator::Get() {
-  static CustomOperator inst;
-  return &inst;
-}
-
 struct CustomParam {
   std::string op_type;
   size_t num_args, num_outs, num_auxs;

tests/python/unittest/test_operator.py

@@ -5198,6 +5198,45 @@ def create_operator(self, ctx, shapes, dtypes):
         x = mx.nd.Custom(length=10, depth=10, op_type="no_input_op")
     assert_almost_equal(x.asnumpy(), np.ones(shape=(10, 10), dtype=np.float32))
 
+    # test custom operator fork
+    # see https://github.com/apache/incubator-mxnet/issues/14396
+    if not sys.platform.startswith('win'):  # no fork in windows
+        class AdditionOP(mx.operator.CustomOp):
+            def __init__(self):
+                super(AdditionOP, self).__init__()
+            def forward(self, is_train, req, in_data, out_data, aux):
+                out_data[0][:] = in_data[0] + in_data[1]
+            def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
+                in_grad[0][:] = out_grad[0]
+                in_grad[1][:] = out_grad[0]
+
+        @mx.operator.register("AdditionOP")
+        class AdditionOPProp(mx.operator.CustomOpProp):
+            def __init__(self):
+                super(AdditionOPProp, self).__init__()
+            def list_arguments(self):
+                return ['a', 'b']
+            def list_outputs(self):
+                return ['output']
+            def infer_shape(self, in_shape):
+                return in_shape, [in_shape[0]]
+            def create_operator(self, ctx, shapes, dtypes):
+                return AdditionOP()
+
+        def custom_add():
+            a = mx.nd.array([1, 2, 3])
+            b = mx.nd.array([4, 5, 6])
+            c = mx.nd.Custom(a, b, op_type='AdditionOP')
+            assert_almost_equal((a + b).asnumpy(), c.asnumpy())
+
+        custom_add()
+        from multiprocessing import Process
+        p = Process(target=custom_add)
+        p.daemon = True
+        p.start()
+        p.join(5)
+        assert not p.is_alive(), "deadlock may exist in custom operator"
+
 @with_seed()
 def test_psroipooling():
     for num_rois in [1, 2]: