Enable ConstantFolding inside MO Backend (#6916)

inference-engine/src/offline_transformations/include/disable_shapeof_constant_folding.hpp

@@ -0,0 +1,24 @@
+// Copyright (C) 2021 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <memory>
+
+#include <ngraph/pass/graph_rewrite.hpp>
+
+namespace ngraph {
+namespace pass {
+
+class DisableShapeOfConstantFolding;
+
+}  // namespace pass
+}  // namespace ngraph
+
+
+class ngraph::pass::DisableShapeOfConstantFolding: public ngraph::pass::MatcherPass {
+public:
+    NGRAPH_RTTI_DECLARATION;
+    DisableShapeOfConstantFolding();
+};

inference-engine/src/offline_transformations/src/disable_shapeof_constant_folding.cpp

@@ -0,0 +1,32 @@
+// Copyright (C) 2021 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include <memory>
+#include <ostream>
+#include <fstream>
+
+#include <ngraph/pattern/op/wrap_type.hpp>
+#include <ngraph/opsets/opset2.hpp>
+#include <ngraph/opsets/opset3.hpp>
+#include <ngraph/variant.hpp>
+#include <transformations/rt_info/disable_constant_folding.hpp>
+
+#include "disable_shapeof_constant_folding.hpp"
+
+NGRAPH_RTTI_DEFINITION(ngraph::pass::DisableShapeOfConstantFolding, "DisableShapeOfConstantFolding", 0);
+
+ngraph::pass::DisableShapeOfConstantFolding::DisableShapeOfConstantFolding() {
+    auto shape_of = pattern::wrap_type<opset2::ShapeOf, opset3::ShapeOf>([=](const Output<Node> & output) {
+        const auto & shape = output.get_partial_shape();
+        return shape.is_dynamic() || shape_size(shape.get_shape()) != 1;
+    });
+
+    ngraph::matcher_pass_callback callback = [=](pattern::Matcher& m) {
+        disable_constant_folding(m.get_match_root());
+        return true;
+    };
+
+    auto m = std::make_shared<ngraph::pattern::Matcher>(shape_of, "DisableShapeOfConstantFolding");
+    this->register_matcher(m, callback);
+}

inference-engine/src/offline_transformations/src/moc_transformations.cpp

@@ -5,8 +5,10 @@
 #include <memory>
 
 #include "moc_transformations.hpp"
+#include "disable_shapeof_constant_folding.hpp"
 
 #include <ngraph/pass/manager.hpp>
+#include <ngraph/pass/constant_folding.hpp>
 #include <transformations/init_node_info.hpp>
 #include <transformations/common_optimizations/gelu_fusion.hpp>
 #include <transformations/common_optimizations/softplus_fusion.hpp>
@@ -32,6 +34,7 @@
 #include <transformations/common_optimizations/lin_op_sequence_fusion.hpp>
 #include <transformations/common_optimizations/conv_mul_fusion.hpp>
 #include <transformations/common_optimizations/nop_elimination.hpp>
+#include <transformations/low_precision/disable_convert_constant_folding_on_const_path.hpp>
 
 NGRAPH_RTTI_DEFINITION(ngraph::pass::MOCTransformations, "MOCTransformations", 0);
 
@@ -48,6 +51,10 @@ bool ngraph::pass::MOCTransformations::run_on_function(std::shared_ptr<ngraph::F
     ngraph::pass::Manager manager(get_pass_config());
 
     manager.register_pass<ngraph::pass::InitNodeInfo>();
+    manager.register_pass<ngraph::pass::DisableConvertConstantFoldingOnConstPath>(
+            element::TypeVector{ ngraph::element::i8, ngraph::element::u8, ngraph::element::i4, ngraph::element::u4 });
+    manager.register_pass<ngraph::pass::DisableShapeOfConstantFolding>();
+    manager.register_pass<ngraph::pass::ConstantFolding>();
     manager.register_pass<ngraph::pass::RemoveFilteringBoxesBySize>();
     manager.register_pass<ngraph::pass::ConvertQuantizeDequantize>();
     manager.register_pass<ngraph::pass::SimplifyShapeOfSubGraph>();

inference-engine/src/transformations/include/transformations/low_precision/disable_convert_constant_folding_on_const_path.hpp

@@ -22,5 +22,5 @@ class ngraph::pass::DisableConvertConstantFoldingOnConstPath : public ngraph::pa
 public:
     NGRAPH_RTTI_DECLARATION;
     DisableConvertConstantFoldingOnConstPath(
-        const std::vector<ngraph::element::Type>& inputPrecisions = {});
+        const element::TypeVector & inputPrecisions = {});
 };

inference-engine/src/transformations/include/transformations/rt_info/disable_constant_folding.hpp

@@ -0,0 +1,44 @@
+// Copyright (C) 2021 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include <assert.h>
+#include <functional>
+#include <memory>
+#include <string>
+#include <set>
+
+#include <ngraph/node.hpp>
+#include <ngraph/variant.hpp>
+#include <transformations_visibility.hpp>
+
+
+namespace ngraph {
+
+/**
+ * @ingroup ie_runtime_attr_api
+ * @brief DisableConstantFolding disable ConstantFolding for given operation
+ */
+class TRANSFORMATIONS_API DisableConstantFolding {
+public:
+    DisableConstantFolding() = default;
+};
+
+extern template class TRANSFORMATIONS_API VariantImpl<DisableConstantFolding>;
+
+template<>
+class TRANSFORMATIONS_API VariantWrapper<DisableConstantFolding> : public VariantImpl<DisableConstantFolding> {
+public:
+    static constexpr VariantTypeInfo type_info{"DISABLED_CONSTANT_FOLDING", 0};
+
+    const VariantTypeInfo &get_type_info() const override {
+        return type_info;
+    }
+
+    VariantWrapper(const value_type &value) : VariantImpl<value_type>(value) {}
+
+    bool is_copyable() const override { return false; }
+};
+
+TRANSFORMATIONS_API void disable_constant_folding(const std::shared_ptr<Node>& node);
+}  // namespace ngraph

inference-engine/src/transformations/src/transformations/low_precision/disable_convert_constant_folding_on_const_path.cpp

@@ -20,7 +20,7 @@ using namespace ngraph;
 NGRAPH_RTTI_DEFINITION(ngraph::pass::DisableConvertConstantFoldingOnConstPath, "DisableConvertConstantFoldingOnConstPath", 0);
 
 ngraph::pass::DisableConvertConstantFoldingOnConstPath::DisableConvertConstantFoldingOnConstPath(
-    const std::vector<ngraph::element::Type>& inputPrecisions) {
+    const element::TypeVector & inputPrecisions) {
     auto matcherData = ngraph::pattern::any_input();
     auto matcherConvert = ngraph::pattern::wrap_type<opset3::Convert>({ matcherData }, pattern::consumers_count(1));
 

inference-engine/src/transformations/src/transformations/rt_info/disable_constant_folding.cpp

@@ -0,0 +1,14 @@
+// Copyright (C) 2021 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "transformations/rt_info/disable_constant_folding.hpp"
+
+template class ngraph::VariantImpl<ngraph::DisableConstantFolding>;
+
+constexpr ngraph::VariantTypeInfo ngraph::VariantWrapper<ngraph::DisableConstantFolding>::type_info;
+
+void ngraph::disable_constant_folding(const std::shared_ptr<Node>& node) {
+    auto & rt_info = node->get_rt_info();
+    rt_info[VariantWrapper<DisableConstantFolding>::type_info.name] = make_variant<DisableConstantFolding>({});
+}

inference-engine/tests/functional/inference_engine/transformations/disable_shapeof_constant_folding_tests.cpp

@@ -0,0 +1,78 @@
+// Copyright (C) 2021 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include <gtest/gtest.h>
+
+#include <string>
+#include <memory>
+#include <queue>
+
+#include <ngraph/function.hpp>
+#include <ngraph/opsets/opset6.hpp>
+#include <ngraph/pass/manager.hpp>
+#include <ngraph/pass/constant_folding.hpp>
+
+#include <disable_shapeof_constant_folding.hpp>
+#include <transformations/serialize.hpp>
+
+#include "common_test_utils/ngraph_test_utils.hpp"
+
+
+using namespace testing;
+using namespace ngraph;
+
+TEST(TransformationTests, DisableShapeOfConstantFolding) {
+    std::shared_ptr<Function> f, f_ref;
+    {
+        auto data = std::make_shared<opset6::Parameter>(element::f32, Shape{1, 4, 10, 10});
+        auto shape_of = std::make_shared<opset6::ShapeOf>(data);
+        auto abs = std::make_shared<opset6::Abs>(shape_of);
+        auto reshape = std::make_shared<opset6::Reshape>(data, abs, false);
+        f = std::make_shared<Function>(NodeVector{reshape}, ParameterVector{data});
+
+        pass::Manager m;
+        m.register_pass<pass::DisableShapeOfConstantFolding>();
+        m.register_pass<pass::ConstantFolding>();
+        m.run_passes(f);
+    }
+
+    {
+        auto data = std::make_shared<opset6::Parameter>(element::f32, Shape{1, 4, 10, 10});
+        auto shape_of = std::make_shared<opset6::ShapeOf>(data);
+        auto abs = std::make_shared<opset6::Abs>(shape_of);
+        auto reshape = std::make_shared<opset6::Reshape>(data, abs, false);
+        f_ref = std::make_shared<Function>(NodeVector{reshape}, ParameterVector{data});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ShapeOfShapeOfConstantFolding) {
+    std::shared_ptr<Function> f, f_ref;
+    {
+        auto data = std::make_shared<opset6::Parameter>(element::i64, Shape{1, 4, 10, 10});
+        auto shape_of = std::make_shared<opset6::ShapeOf>(data);
+        auto reshape = std::make_shared<opset6::Reshape>(data, shape_of, false);
+        auto rank = std::make_shared<opset6::ShapeOf>(shape_of);
+        auto mul = std::make_shared<opset6::Multiply>(reshape, rank);
+        f = std::make_shared<Function>(NodeVector{mul}, ParameterVector{data});
+
+        pass::Manager m;
+        m.register_pass<pass::DisableShapeOfConstantFolding>();
+        m.register_pass<pass::ConstantFolding>();
+        m.run_passes(f);
+    }
+
+    {
+        auto data = std::make_shared<opset6::Parameter>(element::i64, Shape{1, 4, 10, 10});
+        auto shape_of = std::make_shared<opset6::ShapeOf>(data);
+        auto reshape = std::make_shared<opset6::Reshape>(data, shape_of, false);
+        auto mul = std::make_shared<opset6::Multiply>(reshape, opset6::Constant::create(element::i64, Shape{1}, {4}));
+        f_ref = std::make_shared<Function>(NodeVector{mul}, ParameterVector{data});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}

ngraph/core/include/ngraph/validation_util.hpp

@@ -329,6 +329,8 @@ namespace ngraph
     /// that all the HostTensorPtrs are not equal to nullptr
     NGRAPH_API bool validate_host_tensor_vector(const HostTensorVector& v, const size_t& size);
 
+    NGRAPH_API bool could_propagate(const Output<Node>& output, std::vector<Node*>& order);
+
     namespace opset1
     {
         ///

ngraph/core/include/ngraph/variant.hpp

@@ -24,6 +24,7 @@ namespace ngraph
 
         virtual std::shared_ptr<ngraph::Variant> init(const std::shared_ptr<ngraph::Node>& node);
         virtual std::shared_ptr<ngraph::Variant> merge(const ngraph::NodeVector& nodes);
+        virtual bool is_copyable() const;
         virtual std::string to_string() { return ""; }
     };
 

ngraph/core/src/pass/constant_folding.cpp

@@ -6,6 +6,7 @@
 #include <ngraph/op/constant.hpp>
 #include "ngraph/op/util/sub_graph_base.hpp"
 #include "ngraph/rt_info.hpp"
+#include "ngraph/validation_util.hpp"
 
 using namespace std;
 using namespace ngraph;
@@ -101,7 +102,23 @@ bool ngraph::pass::ConstantFolding::pre_calculated_values_folding(
 
         for (auto& input_value : curr_node->input_values())
         {
-            if (input_value.get_tensor().has_and_set_bound())
+            // Check that ConstantFolding is not disabled on this path
+            std::vector<Node*> order;
+            auto status = could_propagate(input_value, order);
+            if (status)
+            {
+                for (const auto& node : order)
+                {
+                    const auto& rt_info = node->get_rt_info();
+                    if (rt_info.count("DISABLED_CONSTANT_FOLDING"))
+                    {
+                        status = false;
+                        break;
+                    }
+                }
+            }
+
+            if (status && input_value.get_tensor().has_and_set_bound())
             {
                 auto input_node = input_value.get_node_shared_ptr();
                 auto replacement =

ngraph/core/src/rt_info.cpp

@@ -8,46 +8,47 @@
 
 ngraph::Node::RTMap mergeRuntimeInfo(const ngraph::NodeVector& nodes)
 {
-    ngraph::Node::RTMap mergedInfo;
-    for (auto& node : nodes)
+    std::unordered_map<std::string, std::vector<std::shared_ptr<ngraph::Variant>>> attrs;
+    for (const auto& node : nodes)
     {
-        for (auto& item : node->get_rt_info())
+        for (const auto& item : node->get_rt_info())
         {
-            mergedInfo[item.first] = item.second;
+            if (item.second->is_copyable())
+            {
+                attrs[item.first].push_back(item.second);
+            }
         }
     }
 
-    ngraph::Node::RTMap newInfo;
-    for (auto& item : mergedInfo)
+    ngraph::Node::RTMap merged_attrs;
+    for (auto& item : attrs)
     {
-        size_t attributes_count = 0;
-        for (auto& node : nodes)
+        auto attr = *item.second.begin();
+        if (item.second.size() == 1)
         {
-            const auto& rt_info = node->get_rt_info();
-            if (rt_info.count(item.first))
+            merged_attrs[item.first] = attr;
+        }
+        else if (auto merge_attr = attr->merge(nodes))
         {
-                attributes_count++;
+            merged_attrs[item.first] = merge_attr;
         }
     }
 
-        if (attributes_count == 1)
-        {
-            newInfo[item.first] = item.second;
-        }
-        else if (auto merge_attr = item.second->merge(nodes))
-        {
-            newInfo[item.first] = merge_attr;
-        }
-    }
-
-    return newInfo;
+    return merged_attrs;
 }
 
 void ngraph::copy_runtime_info(std::shared_ptr<ngraph::Node> from, std::shared_ptr<ngraph::Node> to)
 {
-    auto& rtInfoFrom = from->get_rt_info();
-    auto& rtInfoTo = to->get_rt_info();
-    rtInfoTo = rtInfoFrom;
+    auto& attrs = to->get_rt_info();
+    attrs.clear();
+
+    for (const auto& item : from->get_rt_info())
+    {
+        if (item.second->is_copyable())
+        {
+            attrs[item.first] = item.second;
+        }
+    }
 }
 
 void ngraph::copy_runtime_info(std::shared_ptr<ngraph::Node> from, ngraph::NodeVector to)

ngraph/core/src/validation_util.cpp

@@ -1306,7 +1306,7 @@ void ngraph::evaluate_nodes(std::map<RawNodeOutput, HostTensorPtr>& value_map,
     }
 }
 
-bool could_propagate(const Output<Node>& output, std::vector<Node*>& order)
+bool ngraph::could_propagate(const Output<Node>& output, std::vector<Node*>& order)
 {
     bool status = true;
 
@@ -1367,7 +1367,7 @@ void propagate_rt_info(Node* node, const Output<Node>& final_port)
                 auto& rt_info = consumer->get_rt_info();
                 for (const auto& it : orig_rt_info)
                 {
-                    if (rt_info.find(it.first) == rt_info.end())
+                    if (rt_info.find(it.first) == rt_info.end() && it.second->is_copyable())
                         rt_info[it.first] = it.second;
                 }
             }

ngraph/core/src/variant.cpp

@@ -22,5 +22,10 @@ std::shared_ptr<ngraph::Variant> Variant::merge(const ngraph::NodeVector& nodes)
     return nullptr;
 }
 
+bool Variant::is_copyable() const
+{
+    return true;
+}
+
 template class ngraph::VariantImpl<std::string>;
 template class ngraph::VariantImpl<int64_t>;