[cpu] Remove custom shape inference factories (openvinotoolkit#27924)

### Details: - Remove custom shape inference factories CPU nodes. ### Related PR - openvinotoolkit#27770 ### Tickets: - CVS-118704 --------- Signed-off-by: Raasz, Pawel <[email protected]> Co-authored-by: Michal Lukaszewski <[email protected]> Co-authored-by: Maksim Kutakov <[email protected]>
MirceaDan99 · Jan 22, 2025 · 87235ee · 87235ee
1 parent f1dd30e
commit 87235ee
Show file tree

Hide file tree

Showing 7 changed files with 61 additions and 76 deletions.
diff --git a/src/frontends/tensorflow/src/frontend.cpp b/src/frontends/tensorflow/src/frontend.cpp
@@ -466,12 +466,11 @@ std::shared_ptr<ov::Model> FrontEnd::convert(const ov::frontend::InputModel::Ptr
 
         // recommend to use openvino-tokenizers if some unconverted operations from tokenizers are met
         if (unsupported_ops_from_tokenizers.size() > 0) {
-            exception_message
-                << "\nEncountered unconverted operation(s) for which openvino-tokenizers package "
-                   "provides conversion extension(s): "
-                << unsupported_ops_from_tokenizers
-                << ". Install OpenVINO Tokenizers, refer to the documentation: "
-                   "https://docs.openvino.ai/2024/openvino-workflow-generative/ov-tokenizers.html \n";
+            exception_message << "\nEncountered unconverted operation(s) for which openvino-tokenizers package "
+                                 "provides conversion extension(s): "
+                              << unsupported_ops_from_tokenizers
+                              << ". Install OpenVINO Tokenizers, refer to the documentation: "
+                                 "https://docs.openvino.ai/2024/openvino-workflow-generative/ov-tokenizers.html \n";
         }
     }
 

diff --git a/src/plugins/intel_cpu/src/nodes/deconv.cpp b/src/plugins/intel_cpu/src/nodes/deconv.cpp
@@ -125,16 +125,43 @@ bool DeconvKey::operator==(const DeconvKey& rhs) const {
  * input. Since in case it exists, plugin should pass the input data to the shape inference function.
  *
  */
-class DeconfolutionShapeInferFactory : public ShapeInferFactory {
+class DeconvolutionShapeInferFactory : public ShapeInferFactory {
 public:
-    DeconfolutionShapeInferFactory(std::shared_ptr<ov::Node> op) : m_op(std::move(op)) {}
+    DeconvolutionShapeInferFactory(std::shared_ptr<ov::Node> op) : m_op(std::move(op)) {}
 
     ShapeInferPtr makeShapeInfer() const override {
-        const auto port_mask = (m_op->get_input_size() > 2) ? PortMask(2) : EMPTY_PORT_MASK;
-        return make_shape_inference(m_op, port_mask);
+        return std::make_shared<DeconvolutionShapeInfer>(m_op);
     }
 
 private:
+    class DeconvolutionShapeInfer : public IShapeInfer {
+    public:
+        DeconvolutionShapeInfer(const std::shared_ptr<ov::Node>& op)
+            : m_shape_infer(make_shape_inference(op)),
+              m_port_mask((op->get_input_size() > 2) ? PortMask(2) : EMPTY_PORT_MASK) {}
+
+        Result infer(const std::vector<std::reference_wrapper<const VectorDims>>& input_shapes,
+                     const std::unordered_map<size_t, MemoryPtr>& data_dependency) override {
+            return m_shape_infer->infer(input_shapes, data_dependency);
+        }
+
+        const ov::CoordinateDiff& get_pads_begin() override {
+            return m_shape_infer->get_pads_begin();
+        }
+
+        const ov::CoordinateDiff& get_pads_end() override {
+            return m_shape_infer->get_pads_end();
+        }
+
+        port_mask_t get_port_mask() const override {
+            return m_port_mask;
+        };
+
+    private:
+        ShapeInferPtr m_shape_infer;
+        const port_mask_t m_port_mask;
+    };
+
     std::shared_ptr<ov::Node> m_op;
 };
 }  // namespace
@@ -165,7 +192,7 @@ bool Deconvolution::isSupportedOperation(const std::shared_ptr<const ov::Node>&
 }
 
 Deconvolution::Deconvolution(const std::shared_ptr<ov::Node>& op, const GraphContext::CPtr& context)
-    : Node(op, context, DeconfolutionShapeInferFactory(op)) {
+    : Node(op, context, DeconvolutionShapeInferFactory(op)) {
     std::string errorMessage;
     if (!isSupportedOperation(op, errorMessage))
         OPENVINO_THROW_NOT_IMPLEMENTED(errorMessage);

diff --git a/src/plugins/intel_cpu/src/nodes/eye.cpp b/src/plugins/intel_cpu/src/nodes/eye.cpp
@@ -29,22 +29,8 @@ bool Eye::isSupportedOperation(const std::shared_ptr<const ov::Node>& op, std::s
     return true;
 }
 
-namespace {
-class EyeShapeInferFactory : public ShapeInferFactory {
-public:
-    EyeShapeInferFactory(std::shared_ptr<ov::Node> op) : m_op(std::move(op)) {}
-    ShapeInferPtr makeShapeInfer() const override {
-        return (m_op->get_input_size() == 4) ? make_shape_inference(m_op)
-                                             : make_shape_inference(m_op, PortMask(Eye::ROWS_NUM, Eye::COLS_NUM));
-    }
-
-private:
-    std::shared_ptr<ov::Node> m_op;
-};
-}  // namespace
-
 Eye::Eye(const std::shared_ptr<ov::Node>& op, const GraphContext::CPtr& context)
-    : Node(op, context, EyeShapeInferFactory(op)) {
+    : Node(op, context, NgraphShapeInferFactory(op)) {
     std::string errorMessage;
     if (!isSupportedOperation(op, errorMessage)) {
         OPENVINO_THROW_NOT_IMPLEMENTED(errorMessage);

diff --git a/src/plugins/intel_cpu/src/nodes/reference.cpp b/src/plugins/intel_cpu/src/nodes/reference.cpp
@@ -12,22 +12,10 @@
 namespace ov {
 namespace intel_cpu {
 
-class ReferenceShapeInferFactory : public ShapeInferFactory {
-public:
-    ReferenceShapeInferFactory(std::shared_ptr<ov::Node> op) : m_op{std::move(op)} {}
-
-    ShapeInferPtr makeShapeInfer() const override {
-        return make_shape_inference(m_op, FULL_PORT_MASK);
-    }
-
-private:
-    std::shared_ptr<ov::Node> m_op;
-};
-
 namespace node {
 
 Reference::Reference(const std::shared_ptr<ov::Node>& op, const GraphContext::CPtr& context, std::string errorMessage)
-    : Node(op, context, ReferenceShapeInferFactory(op)),
+    : Node(op, context, NgraphShapeInferFactory(op)),
       ovCoreNode(op),
       additionalErrorMessage(std::move(errorMessage)) {
     if (!op->has_evaluate()) {
@@ -61,7 +49,9 @@ void Reference::initSupportedPrimitiveDescriptors() {
     addSupportedPrimDesc(inputConfigurators, outputConfigurators, impl_desc_type::ref);
 }
 
-void Reference::createPrimitive() {}
+void Reference::createPrimitive() {
+    hasOutputShapeDataDependency = isDynamicNode() && outputShapeDataDependency();
+}
 
 void Reference::execute(const dnnl::stream& strm) {
     auto inputs = prepareInputs();
@@ -72,6 +62,14 @@ void Reference::execute(const dnnl::stream& strm) {
 }
 
 void Reference::executeDynamicImpl(const dnnl::stream& strm) {
+    if (!hasOutputShapeDataDependency) {
+        // if there is no data dependency for the output shape, we can execute the operation as is, similar to the
+        // static case, since the shapes are already calculated
+        execute(strm);
+        return;
+    }
+
+    // if there is data dependency, we need to perform shape inference first
     auto inputs = prepareInputs();
     ov::TensorVector outputs;
     auto result = Node::shapeInfer();
@@ -125,7 +123,9 @@ bool Reference::created() const {
 }
 
 bool Reference::needShapeInfer() const {
-    return false;
+    // If there is data dependency for the output shape, let's assume the node has internal dynamism (in general case),
+    // so we postpone the shape inference until the actual execution
+    return !hasOutputShapeDataDependency && Node::needShapeInfer();
 }
 
 ov::TensorVector Reference::prepareInputs() const {

diff --git a/src/plugins/intel_cpu/src/nodes/reference.h b/src/plugins/intel_cpu/src/nodes/reference.h
@@ -36,6 +36,7 @@ class Reference : public Node {
 private:
     const std::shared_ptr<ov::Node> ovCoreNode;
     const std::string additionalErrorMessage;
+    bool hasOutputShapeDataDependency = false;  // flag to cache the output shape data dependency check result
 };
 
 }  // namespace node

diff --git a/src/plugins/intel_cpu/src/shape_inference/shape_inference.cpp b/src/plugins/intel_cpu/src/shape_inference/shape_inference.cpp
@@ -234,8 +234,7 @@ class ShapeInferFallback : public ShapeInferBase {
 
     ov::optional<std::vector<StaticShape>> infer(const std::vector<StaticShapeRef>& input_shapes,
                                                  const ov::ITensorAccessor& tensor_accessor) override {
-        auto op = m_node.get();
-        std::vector<StaticShape> output_shapes;
+        const auto op = m_node.get();
 
         std::shared_ptr<ov::Node> local_op;
         ov::OutputVector new_inputs;
@@ -252,7 +251,7 @@ class ShapeInferFallback : public ShapeInferBase {
         local_op = op->clone_with_new_inputs(new_inputs);
         local_op->validate_and_infer_types();
 
-        output_shapes.resize(local_op->get_output_size());
+        std::vector<StaticShape> output_shapes(local_op->get_output_size());
         for (size_t i = 0; i < output_shapes.size(); ++i) {
             const auto& partial_shape = local_op->get_output_partial_shape(i);
 
@@ -265,6 +264,11 @@ class ShapeInferFallback : public ShapeInferBase {
 
         return {std::move(output_shapes)};
     }
+
+    port_mask_t get_port_mask() const override {
+        // For fallback return full port mask to try get data for all node's inputs
+        return FULL_PORT_MASK;
+    }
 };
 
 template <class TOp, IShapeInfer::port_mask_t MASK>
@@ -610,34 +614,6 @@ const IStaticShapeInferFactory::TRegistry IStaticShapeInferFactory::registry{
 #undef _OV_OP_SHAPE_INFER_MASK_REG
 #undef _OV_OP_SHAPE_INFER_VA_REG
 
-class ShapeInferCustomMask : public IShapeInfer {
-public:
-    ShapeInferCustomMask(ShapeInferPtr shape_infer, port_mask_t port_mask)
-        : m_shape_infer{std::move(shape_infer)},
-          m_port_mask{port_mask} {}
-
-    Result infer(const std::vector<std::reference_wrapper<const VectorDims>>& input_shapes,
-                 const std::unordered_map<size_t, MemoryPtr>& data_dependency) override {
-        return m_shape_infer->infer(input_shapes, data_dependency);
-    }
-
-    const ov::CoordinateDiff& get_pads_begin() override {
-        return m_shape_infer->get_pads_begin();
-    }
-
-    const ov::CoordinateDiff& get_pads_end() override {
-        return m_shape_infer->get_pads_end();
-    }
-
-    port_mask_t get_port_mask() const override {
-        return m_port_mask;
-    }
-
-private:
-    const ShapeInferPtr m_shape_infer;
-    const port_mask_t m_port_mask;
-};
-
 std::shared_ptr<IStaticShapeInfer> make_shape_inference(std::shared_ptr<ov::Node> op) {
     if (auto shape_infer = IStaticShapeInferFactory::make(op->get_type_info(), op)) {
         return shape_infer;
@@ -652,8 +628,5 @@ std::shared_ptr<IStaticShapeInfer> make_shape_inference(std::shared_ptr<ov::Node
     }
 }
 
-ShapeInferPtr make_shape_inference(std::shared_ptr<ov::Node> op, IShapeInfer::port_mask_t port_mask) {
-    return std::make_shared<ShapeInferCustomMask>(make_shape_inference(std::move(op)), port_mask);
-}
 }  // namespace intel_cpu
 }  // namespace ov
diff --git a/src/plugins/intel_cpu/src/shape_inference/shape_inference.hpp b/src/plugins/intel_cpu/src/shape_inference/shape_inference.hpp
@@ -32,6 +32,5 @@ class IStaticShapeInfer : public IShapeInfer {
 };
 
 std::shared_ptr<IStaticShapeInfer> make_shape_inference(std::shared_ptr<ov::Node> op);
-ShapeInferPtr make_shape_inference(std::shared_ptr<ov::Node> op, IShapeInfer::port_mask_t port_mask);
 }  // namespace intel_cpu
 }  // namespace ov