shrestha/upgrade ng-core to master (#356)

ugrade ngraph_version = "e543688981ea2083a260929a70420f5ba6984d95"
Nagy Mostafa [MLIR] Fused Ops dialect declaration (#3860)

Fixed Parallel Executor to take in the node name

PTS asserts out if depth>0 but no of inputs and outputs is 0. Behavior to be fixed in another PR. Unit tests that have NoOp make a cluster with no inputs and outputs. disabled support of NoOp for them.

All pipelines except --var go to Parallel Executor.

Author: Shrestha Malik

CMakeLists.txt

@@ -248,7 +248,7 @@ if (NOT USE_PRE_BUILT_NGRAPH)
     ExternalProject_Add(
         ext_ngraph
         GIT_REPOSITORY https://github.com/NervanaSystems/ngraph
-        GIT_TAG effcc47d39de1d5e662d6b78a8d42675ecf7815d
+        GIT_TAG e543688981ea2083a260929a70420f5ba6984d95
         CMAKE_ARGS 
             -DNGRAPH_DISTRIBUTED_ENABLE=${NGRAPH_DISTRIBUTED_ENABLE}
             -DNGRAPH_INSTALL_PREFIX=${NGRAPH_ARTIFACTS_DIR}

bazel/WORKSPACE

@@ -55,11 +55,11 @@ tf_workspace(path_prefix = "", tf_repo_name = "org_tensorflow")
 http_archive(
     name = "ngraph",
     build_file = "//:bazel/ngraph.BUILD",
-    sha256 = "1249b2cbbd89d83405ae2aedf058258fe1290045a7ed3a80ae690e5b2cfe4e0b",
-    strip_prefix = "ngraph-effcc47d39de1d5e662d6b78a8d42675ecf7815d",
+    sha256 = "b178f456dcfb212e0ab97fbb9b60efe35e749c8c361234bf2810c71083b8c455",
+    strip_prefix = "ngraph-e543688981ea2083a260929a70420f5ba6984d95",
     urls = [
-        "https://mirror.bazel.build/github.com/NervanaSystems/ngraph/archive/effcc47d39de1d5e662d6b78a8d42675ecf7815d.tar.gz",
-        "https://github.com/NervanaSystems/ngraph/archive/effcc47d39de1d5e662d6b78a8d42675ecf7815d.tar.gz"
+        "https://mirror.bazel.build/github.com/NervanaSystems/ngraph/archive/e543688981ea2083a260929a70420f5ba6984d95.tar.gz",
+        "https://github.com/NervanaSystems/ngraph/archive/e543688981ea2083a260929a70420f5ba6984d95.tar.gz"
     ],
 )
 

bazel/ngraph.BUILD

@@ -68,7 +68,7 @@ cc_library(
         "-fstack-protector-all",
         '-D SHARED_LIB_PREFIX=\\"lib\\"',
         '-D SHARED_LIB_SUFFIX=\\".so\\"',
-        '-D NGRAPH_VERSION=\\"v0.27.0-rc.0\\"',
+        '-D NGRAPH_VERSION=\\"e543688981ea2083a260929a70420f5ba6984d95\\"',
         "-D NGRAPH_DEX_ONLY",
         '-D PROJECT_ROOT_DIR=\\"\\"',
         '-D NGRAPH_STATIC_LIB_ENABLE',
@@ -111,7 +111,7 @@ cc_library(
         "-fstack-protector-all",
         '-D SHARED_LIB_PREFIX=\\"lib\\"',
         '-D SHARED_LIB_SUFFIX=\\".so\\"',
-        '-D NGRAPH_VERSION=\\"effcc47d39de1d5e662d6b78a8d42675ecf7815d\\"',
+        '-D NGRAPH_VERSION=\\"e543688981ea2083a260929a70420f5ba6984d95\\"',
         "-D NGRAPH_DEX_ONLY",
         '-D PROJECT_ROOT_DIR=\\"\\"',
     ] + CXX_ABI,
@@ -202,6 +202,7 @@ cc_library(
         "src/ngraph/runtime/cpu/builder/softmax.cpp",
         "src/ngraph/runtime/cpu/builder/get_output_element.cpp",
         "src/ngraph/runtime/cpu/builder/sum.cpp",
+        "src/ngraph/runtime/cpu/builder/cum_sum.cpp",
         "src/ngraph/runtime/cpu/builder/topk.cpp",
         "src/ngraph/runtime/cpu/builder/tile.cpp",
         "src/ngraph/runtime/cpu/builder/update_slice.cpp",

build_ngtf.py

@@ -53,7 +53,7 @@ def main():
     '''
 
     # Component versions
-    ngraph_version = "effcc47d39de1d5e662d6b78a8d42675ecf7815d"  # Scott Cyphers CropAndResize op
+    ngraph_version = "e543688981ea2083a260929a70420f5ba6984d95"  # Nagy Mostafa [MLIR] Fused Ops dialect declaration (#3860)
     tf_version = "v1.14.0"
 
     # Command line parser options

ngraph_bridge/ngraph_encapsulate_op.cc

@@ -157,9 +157,9 @@ void NGraphEncapsulateOp::CreateParallelExecutor(OpKernelConstruction* ctx,
   }
 
   // Create the Executor object
-  m_parallel_executor = move(unique_ptr<NGraphExecutor>(
-      new NGraphExecutor(s_instance_id, cluster_id, graph_id, encap_subgraph,
-                         backend_name, my_function_cache_depth_in_items)));
+  m_parallel_executor = move(unique_ptr<NGraphExecutor>(new NGraphExecutor(
+      s_instance_id, cluster_id, graph_id, encap_subgraph, backend_name, name(),
+      my_function_cache_depth_in_items)));
 
   auto tensor_manager = m_parallel_executor->GetTensorManager();
   OP_REQUIRES(ctx, tensor_manager->GetNumberOfInputs() == ctx->num_inputs(),
@@ -302,7 +302,11 @@ void NGraphEncapsulateOp::CreateLegacyExecutor(OpKernelConstruction* ctx,
   bool exec_can_create_tensor =
       BackendManager::GetBackend(ng_encap_impl_.GetOpBackend())
           ->executable_can_create_tensors();
-  ng_encap_impl_.SetExecCanCreateTensor(exec_can_create_tensor);
+  if (ng_encap_impl_.GetOpBackend() == "CPU") {
+    ng_encap_impl_.SetExecCanCreateTensor(false);
+  } else {
+    ng_encap_impl_.SetExecCanCreateTensor(exec_can_create_tensor);
+  }
   NGRAPH_VLOG(5) << "Executable can " << (exec_can_create_tensor ? "" : "not")
                  << " create tensors";
 

ngraph_bridge/ngraph_executor.cc

@@ -62,12 +62,13 @@ namespace ngraph_bridge {
 NGraphExecutor::NGraphExecutor(int instance_id, int cluster_id, int graph_id,
                                unique_ptr<tensorflow::Graph>& graph,
                                const string& backend_name,
-                               const int cache_depth)
+                               const string& node_name, const int cache_depth)
     : m_instance_id(instance_id),
       m_ngraph_cluster_id(cluster_id),
       m_graph_id(graph_id),
       m_graph(std::move(graph)),
       m_op_backend_name(backend_name),
+      m_node_name(node_name),
       m_ng_data_cache(cache_depth) {
   // Sanity checks
   if (m_graph == nullptr) {
@@ -85,7 +86,6 @@ NGraphExecutor::NGraphExecutor(int instance_id, int cluster_id, int graph_id,
     throw std::runtime_error(string("Requested backend: '") +
                              m_op_backend_name + string("' not available."));
   }
-
   // Initialize the "m_input_is_static" vector as follows:
   // (1) create m_input_is_static with n+1 elements, where n is the max arg
   //     index

ngraph_bridge/ngraph_executor.h

@@ -42,7 +42,8 @@ class NGraphExecutor {
   // Transforms, compiles and executes TesnorFlow computation graph using nGraph
   explicit NGraphExecutor(int instance_id, int cluster_id, int graph_id,
                           unique_ptr<tensorflow::Graph>& graph,
-                          const string& backend_name, const int cache_depth);
+                          const string& backend_name, const string& node_name,
+                          const int cache_depth);
 
   ~NGraphExecutor();
 

test/python/tensorflow/tf_unittest_ngraph.patch

@@ -1,12 +1,14 @@
 diff --git a/tensorflow/python/framework/test_util.py b/tensorflow/python/framework/test_util.py
-index aa8ecf8161..b8d2266132 100644
+index 544190e23d..9b652bd49e 100644
 --- a/tensorflow/python/framework/test_util.py
 +++ b/tensorflow/python/framework/test_util.py
-@@ -971,6 +971,7 @@ class TensorFlowTestCase(googletest.TestCase):
+@@ -1662,6 +1662,9 @@ class TensorFlowTestCase(googletest.TestCase):
      self._threads = []
      self._tempdir = None
      self._cached_session = None
 +    import ngraph_bridge
++
++    ngraph_bridge.set_disabled_ops('NoOp')
 
    def setUp(self):
-     self._ClearCachedSession()
+     self._ClearCachedSession()

test/python/test_variableops_static_input.py

@@ -115,6 +115,11 @@ def test_variable_static_input_variables_dont_share_buffer(self):
         env_var_map = self.store_env_variables([buffer_sharing_env])
         self.set_env_variable(buffer_sharing_env, "0")
 
+        # disable NoOp it forms a single cluster with no inputs or outputs
+        # Pipelined Tensor Store thinks it is depth 0 and errors out
+        # TODO: Revisit PTS to set valid PTS depth
+        ngraph_bridge.set_disabled_ops('NoOp')
+
         # Run on nGraph
         ng_var_init_val, ng_mean_values, ng_var_final = self.with_ngraph(
             self.__run_test)
@@ -143,6 +148,7 @@ def test_variable_static_input_variables_dont_share_buffer(self):
         assert np.allclose(ng_var_final, tf_var_final)
 
         # clean up
+        ngraph_bridge.set_disabled_ops('')
         self.unset_env_variable(buffer_sharing_env)
         self.restore_env_variables(env_var_map)
 
@@ -153,6 +159,11 @@ def test_variable_static_input_variables_share_buffer(self):
         env_var_map = self.store_env_variables([buffer_sharing_env])
         self.set_env_variable(buffer_sharing_env, "1")
 
+        # disable NoOp it forms a single cluster with no inputs or outputs
+        # Pipelined Tensor Store thinks it is depth 0 and errors out
+        # TODO: Revisit PTS to set valid PTS depth
+        ngraph_bridge.set_disabled_ops('NoOp')
+
         # Run on nGraph
         ng_var_init_val, ng_mean_values, ng_var_final = self.with_ngraph(
             self.__run_test)
@@ -181,3 +192,4 @@ def test_variable_static_input_variables_share_buffer(self):
         # clean up
         self.unset_env_variable(buffer_sharing_env)
         self.restore_env_variables(env_var_map)
+        ngraph_bridge.set_disabled_ops('')

test/test_parallel_executor.cpp

@@ -60,22 +60,23 @@ TEST(ParallelExecutor, Construction) {
 
   // First test with a backend not yet created
   unique_ptr<NGraphExecutor> executor;
-  ASSERT_THROW(executor = unique_ptr<NGraphExecutor>(
-                   new NGraphExecutor(100, 500, 600, input_graph, "bogus", 5)),
+  ASSERT_THROW(executor = unique_ptr<NGraphExecutor>(new NGraphExecutor(
+                   100, 500, 600, input_graph, "bogus", "xyz_500", 5)),
                std::runtime_error);
 
   // Next test with a null graph not yet created
-  ASSERT_THROW(executor = unique_ptr<NGraphExecutor>(
-                   new NGraphExecutor(100, 500, 600, input_graph, "bogus", 12)),
+  ASSERT_THROW(executor = unique_ptr<NGraphExecutor>(new NGraphExecutor(
+                   100, 500, 600, input_graph, "bogus", "xyz_500", 12)),
                std::runtime_error);
 
   // Now read the graph
   ASSERT_OK(LoadGraphFromPbTxt("test_axpy_launchop.pbtxt", input_graph));
 
   // Next test with a backend after creating
   tf::ngraph_bridge::BackendManager::CreateBackend("INTERPRETER");
-  ASSERT_NO_THROW(executor = unique_ptr<NGraphExecutor>(new NGraphExecutor(
-                      100, 500, 600, input_graph, "INTERPRETER", 16)));
+  ASSERT_NO_THROW(
+      executor = unique_ptr<NGraphExecutor>(new NGraphExecutor(
+          100, 500, 600, input_graph, "INTERPRETER", "xyz_500", 16)));
 
   // Now that the object has been cobstructed, test various internal parts
   // TODO: Create a Test Class and mark that as a friend of the Executor class
@@ -93,7 +94,8 @@ TEST(ParallelExecutor, CompilerTest) {
   ASSERT_OK(LoadGraphFromPbTxt("test_axpy_launchop.pbtxt", input_graph));
 
   tf::ngraph_bridge::BackendManager::CreateBackend("INTERPRETER");
-  NGraphExecutor executor(100, 500, 600, input_graph, "INTERPRETER", 10);
+  NGraphExecutor executor(100, 500, 600, input_graph, "INTERPRETER", "xyz_500",
+                          10);
 
   // Create the inputs for this graph
   Tensor x(DT_FLOAT, TensorShape({2, 3}));
@@ -132,7 +134,8 @@ TEST(ParallelExecutor, ExecuteOnSingleThread) {
   unique_ptr<tf::Graph> input_graph;
   ASSERT_OK(LoadGraphFromPbTxt("test_axpy_launchop.pbtxt", input_graph));
   tf::ngraph_bridge::BackendManager::CreateBackend("INTERPRETER");
-  NGraphExecutor executor(100, 500, 600, input_graph, "INTERPRETER", 12);
+  NGraphExecutor executor(100, 500, 600, input_graph, "INTERPRETER", "xyz_500",
+                          12);
 
   // Create the inputs for this graph
   Tensor x(DT_FLOAT, TensorShape({2, 3}));
@@ -214,7 +217,8 @@ TEST(ParallelExecutor, ExecuteOnSingleThread8Bit) {
   }
 
   tf::ngraph_bridge::BackendManager::CreateBackend(backend_name);
-  NGraphExecutor executor(100, 500, 600, input_graph, backend_name, 5);
+  NGraphExecutor executor(100, 500, 600, input_graph, backend_name, "xyz_500",
+                          5);
 
   // Create the inputs for this graph
   Tensor x(DT_INT8, TensorShape({2, 2}));
@@ -296,7 +300,8 @@ TEST(ParallelExecutor, ExecuteOnMultipleThreads8Bit) {
   }
 
   tf::ngraph_bridge::BackendManager::CreateBackend(backend_name);
-  NGraphExecutor executor(100, 500, 600, input_graph, backend_name, 16);
+  NGraphExecutor executor(100, 500, 600, input_graph, backend_name, "xyz_500",
+                          16);
 
   // Create the inputs for this graph
   Tensor x(DT_INT8, TensorShape({2, 2}));
@@ -383,7 +388,8 @@ TEST(ParallelExecutor, ExecuteOnMultipleThreads) {
   unique_ptr<tf::Graph> input_graph;
   ASSERT_OK(LoadGraphFromPbTxt("test_axpy_launchop.pbtxt", input_graph));
   tf::ngraph_bridge::BackendManager::CreateBackend("INTERPRETER");
-  NGraphExecutor executor(100, 500, 600, input_graph, "INTERPRETER", 16);
+  NGraphExecutor executor(100, 500, 600, input_graph, "INTERPRETER", "xyz_500",
+                          16);
 
   // Create the inputs for this graph
   Tensor x(DT_FLOAT, TensorShape({2, 3}));