[GPU] Use ov::Tensor intead of ngraph::HostTensor (#18925)

src/core/shape_inference/include/tensor_data_accessor.hpp

@@ -18,6 +18,9 @@ public:
      * @return      Tensor to data at port.
      */
     virtual Tensor operator()(size_t port) const = 0;
+
+protected:
+    ~ITensorAccessor() = default;
 };
 
 /**

src/plugins/intel_gpu/include/intel_gpu/runtime/memory.hpp

@@ -9,8 +9,6 @@
 #include "event.hpp"
 #include "engine_configuration.hpp"
 
-#include "ngraph/runtime/host_tensor.hpp"
-
 #include <type_traits>
 
 #ifdef ENABLE_ONEDNN_FOR_GPU
@@ -247,16 +245,4 @@ inline std::vector<T> read_vector(cldnn::memory::ptr mem, const cldnn::stream& s
     return out_vecs;
 }
 
-inline std::shared_ptr<ngraph::runtime::HostTensor> make_host_tensor(layout l, void* memory_pointer) {
-    ov::element::Type et = data_type_to_element_type(l.data_type);
-
-    return std::make_shared<ngraph::runtime::HostTensor>(et, l.get_shape(), memory_pointer);
-}
-
-inline ov::Tensor make_tensor(layout l, void* memory_pointer) {
-    ov::element::Type et = data_type_to_element_type(l.data_type);
-
-    return ov::Tensor(et, l.get_shape(), memory_pointer);
-}
-
 }  // namespace cldnn

src/plugins/intel_gpu/include/intel_gpu/runtime/tensor_accessor.hpp

@@ -0,0 +1,89 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/runtime/tensor.hpp"
+#include "tensor_data_accessor.hpp"
+
+#include "memory.hpp"
+#include "layout.hpp"
+
+namespace cldnn {
+
+inline ov::Tensor make_tensor(const layout& l, void* memory_pointer) {
+    ov::element::Type et = data_type_to_element_type(l.data_type);
+
+    return ov::Tensor(et, l.get_shape(), memory_pointer);
+}
+
+struct TensorsContainer final {
+    using MemoryMap = std::unordered_map<size_t, cldnn::memory::ptr>;
+    using TensorsMap = std::unordered_map<size_t, ov::Tensor>;
+
+    TensorsContainer(const cldnn::stream* stream, const std::map<size_t, cldnn::memory::ptr>& deps_map = {})
+        : m_stream(stream)
+        , m_memories(deps_map.begin(), deps_map.end()) { }
+
+    ~TensorsContainer() {
+        for (auto& port : m_locked_memories) {
+            m_memories.at(port)->unlock(*m_stream);
+        }
+    }
+
+    void emplace(size_t port, cldnn::memory::ptr mem) {
+        m_memories.emplace(port, mem);
+    }
+
+    void emplace(size_t port, const ov::Tensor& tensor) {
+        auto res = m_tensors.emplace(port, tensor);
+        OPENVINO_ASSERT(res.first != m_tensors.end());
+    }
+
+    template<typename ElementType>
+    void emplace(size_t port, std::vector<ElementType>& vector, data_types dt = data_types::i64) {
+        ov::Shape shape{vector.size()};
+        auto tensor = make_tensor({shape, dt, format::bfyx}, static_cast<void*>(vector.data()));
+        m_tensors.emplace(port, tensor);
+    }
+
+    size_t size() const { return m_tensors.size(); }
+    ov::Tensor operator[](std::size_t port) const {
+        if (m_memories.count(port) > 0) {
+            m_locked_memories.insert(port);
+            auto mem = m_memories.at(port);
+            auto ptr = mem->lock(*m_stream, cldnn::mem_lock_type::read);
+            return make_tensor(mem->get_layout(), ptr);
+        } else if (m_tensors.count(port) > 0) {
+            return m_tensors.at(port);
+        } else {
+            OPENVINO_THROW("[GPU] Can't get tensor for ", port, " port!\n");
+        }
+    }
+
+private:
+    const cldnn::stream* m_stream;
+    MemoryMap m_memories;
+    TensorsMap m_tensors;
+
+    mutable std::set<size_t> m_locked_memories = {};
+};
+
+class TensorAccessor final : public ov::ITensorAccessor {
+public:
+    explicit TensorAccessor(const TensorsContainer& container) : m_container(container) { }
+
+    ov::Tensor operator()(size_t port) const override {
+        return m_container[port];
+    }
+
+private:
+    const TensorsContainer& m_container;
+};
+
+inline cldnn::TensorAccessor make_tensor_accessor(const TensorsContainer& c) {
+    return cldnn::TensorAccessor(c);
+}
+
+}  // namespace cldnn

src/plugins/intel_gpu/src/graph/CMakeLists.txt

@@ -33,10 +33,9 @@ target_include_directories(${TARGET_NAME} PUBLIC
 target_compile_options(${TARGET_NAME} PRIVATE
   $<$<CONFIG:Release>:$<IF:$<CXX_COMPILER_ID:MSVC>,/Os,-Os>>)
 
-target_link_libraries(${TARGET_NAME} PUBLIC OpenCL::OpenCL)
+target_link_libraries(${TARGET_NAME} PUBLIC OpenCL::OpenCL ov_shape_inference)
 target_link_libraries(${TARGET_NAME} PRIVATE openvino_intel_gpu_kernels
                                              openvino_intel_gpu_runtime
-                                             ov_shape_inference
                                              openvino::itt
                                              openvino::runtime::dev
                                              openvino::runtime)

src/plugins/intel_gpu/src/graph/arg_max_min.cpp

@@ -87,17 +87,17 @@ std::vector<layout> arg_max_min_inst::calc_output_layouts(arg_max_min_node const
 
     auto& constant_mem = impl_param.memory_deps;
     if (desc->top_k > 0) {
-        std::map<size_t, ngraph::HostTensorPtr> const_data;
+        std::unordered_map<size_t, ov::Tensor> const_data;
         auto topk = desc->top_k;
-        auto top_k_tensor = std::make_shared<ngraph::runtime::HostTensor>(ov::element::u32, ov::Shape{1}, static_cast<void*>(&topk));
+        auto top_k_tensor = ov::Tensor(ov::element::u32, ov::Shape{1}, static_cast<void*>(&topk));
         const_data = { {1, top_k_tensor} };
 
         output_shapes = ov::op::shape_infer(&op, input_shapes, ov::make_tensor_accessor(const_data));
     } else if (constant_mem.count(1)) {
-        std::map<size_t, ngraph::HostTensorPtr> const_data;
+        std::unordered_map<size_t, ov::Tensor> const_data;
         auto target_shape_mem = constant_mem.at(1);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> target_shape_lock(target_shape_mem, impl_param.get_stream());
-        const_data.emplace(1, make_host_tensor(target_shape_mem->get_layout(), target_shape_lock.data()));
+        const_data.emplace(1, make_tensor(target_shape_mem->get_layout(), target_shape_lock.data()));
 
         output_shapes = ov::op::shape_infer(&op, input_shapes, ov::make_tensor_accessor(const_data));
     } else {

src/plugins/intel_gpu/src/graph/batch_to_space.cpp

@@ -112,7 +112,7 @@ std::vector<layout> batch_to_space_inst::calc_output_layouts(batch_to_space_node
         end_shape
     };
 
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
     if (desc->shape_constant) {
         auto block_sizes = tensor_to_vec(block_data, input0_format);
         auto begin_sizes = tensor_to_vec(begin_data, input0_format);
@@ -122,9 +122,9 @@ std::vector<layout> batch_to_space_inst::calc_output_layouts(batch_to_space_node
         auto begin_values = static_cast<void*>(begin_sizes.data());
         auto end_values = static_cast<void*>(end_sizes.data());
 
-        auto block_tensor = make_host_tensor({ block_shape, data_types::i32, input0_format }, block_values);
+        auto block_tensor = make_tensor({ block_shape, data_types::i32, input0_format }, block_values);
-        auto begin_tensor = make_host_tensor({ begin_shape, data_types::i32, input0_format }, begin_values);
+        auto begin_tensor = make_tensor({ begin_shape, data_types::i32, input0_format }, begin_values);
-        auto end_tensor = make_host_tensor({ end_shape, data_types::i32, input0_format }, end_values);
+        auto end_tensor = make_tensor({ end_shape, data_types::i32, input0_format }, end_values);
 
         const_data.emplace(1, block_tensor);
         const_data.emplace(2, begin_tensor);
@@ -140,9 +140,9 @@ std::vector<layout> batch_to_space_inst::calc_output_layouts(batch_to_space_node
         cldnn::mem_lock<uint8_t, mem_lock_type::read> lock2(begin_mem, impl_param.get_stream());
         cldnn::mem_lock<uint8_t, mem_lock_type::read> lock3(end_mem, impl_param.get_stream());
 
-        auto block_tensor = make_host_tensor(block_mem->get_layout(), lock1.data());
+        auto block_tensor = make_tensor(block_mem->get_layout(), lock1.data());
-        auto begin_tensor = make_host_tensor(begin_mem->get_layout(), lock2.data());
+        auto begin_tensor = make_tensor(begin_mem->get_layout(), lock2.data());
-        auto end_tensor = make_host_tensor(end_mem->get_layout(), lock3.data());
+        auto end_tensor = make_tensor(end_mem->get_layout(), lock3.data());
 
         const_data.emplace(1, block_tensor);
         const_data.emplace(2, begin_tensor);

src/plugins/intel_gpu/src/graph/border.cpp

@@ -3,6 +3,7 @@
 //
 
 #include "border_inst.h"
+#include "intel_gpu/runtime/tensor_accessor.hpp"
 #include "pad_shape_inference.hpp"
 
 #include "intel_gpu/runtime/error_handler.hpp"
@@ -41,85 +42,55 @@ std::vector<layout> border_inst::calc_output_layouts(border_node const& /*node*/
         output_type = impl_param.get_fused_output_layout().data_type;
     }
 
+    size_t in_rank = input0_layout.get_partial_shape().size();
+
     ov::op::v1::Pad op;
     op.set_pad_mode(desc->pad_mode);
 
     const bool is_begin_mem = (desc->non_constant_input_mask & border::PAD_NON_CONST_INPUT::BEGIN);
     const bool is_end_mem = (desc->non_constant_input_mask & border::PAD_NON_CONST_INPUT::END);
 
-    layout pads_begin_layout, pads_end_layout;
+    const size_t begin_mem_idx = is_begin_mem ? 1 : 0;
-    if (is_begin_mem) {
+    const size_t end_mem_idx = is_begin_mem ? 2 : 1;
-        pads_begin_layout = impl_param.get_input_layout(1);
-    }
-    if (is_end_mem) {
-        pads_end_layout = is_begin_mem ? impl_param.get_input_layout(2) : impl_param.get_input_layout(1);
-    }
-
-    ShapeType pads_begin_shape = is_begin_mem ? pads_begin_layout.get<ShapeType>() : ov::Shape{ desc->pads_begin.size() };
-    ShapeType pads_end_shape = is_end_mem ? pads_end_layout.get<ShapeType>() : ov::Shape{ desc->pads_end.size() };
-    std::vector<ShapeType> output_shapes;
-    std::vector<ShapeType> input_shapes = {
-        input0_layout.get<ShapeType>(),
-        pads_begin_shape,
-        pads_end_shape,
-    };
 
     auto& memory_deps = impl_param.memory_deps;
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    if ((is_begin_mem && memory_deps.count(begin_mem_idx) == 0) ||
-    auto ta = ov::make_tensor_accessor(const_data);
+        (is_end_mem && memory_deps.count(end_mem_idx) == 0)) {
-
+        return {layout{ShapeType::dynamic(static_cast<int64_t>(in_rank)), input0_layout.data_type, input0_layout.format}};
-    if ((is_begin_mem && memory_deps.count(1)) && (is_end_mem && memory_deps.count(2))) {
-        auto pads_begin_mem = memory_deps.at(1);
-        cldnn::mem_lock<uint8_t, mem_lock_type::read> pads_begin_lock(pads_begin_mem, impl_param.get_stream());
-        const_data.emplace(1, make_host_tensor(pads_begin_mem->get_layout(), pads_begin_lock.data()));
-
-        auto pads_end_mem = memory_deps.at(2);
-        cldnn::mem_lock<uint8_t, mem_lock_type::read> pads_end_lock(pads_end_mem, impl_param.get_stream());
-        const_data.emplace(2, make_host_tensor(pads_end_mem->get_layout(), pads_end_lock.data()));
-
-        output_shapes = ov::op::shape_infer(&op, input_shapes, ta);
-    } else if ((is_begin_mem || is_end_mem) && memory_deps.count(1)) {
-        if (is_begin_mem) {
-            auto pads_begin_mem = memory_deps.at(1);
-            cldnn::mem_lock<uint8_t, mem_lock_type::read> pads_begin_lock(pads_begin_mem, impl_param.get_stream());
-            const_data.emplace(1, make_host_tensor(pads_begin_mem->get_layout(), pads_begin_lock.data()));
-
-            auto pads_end_data = desc->pads_end;
-            auto pads_end_tensor = make_host_tensor({pads_end_shape, data_types::i64, format::bfyx}, static_cast<void*>(pads_end_data.data()));
-            const_data.emplace(2, pads_end_tensor);
-
-            output_shapes = ov::op::shape_infer(&op, input_shapes, ta);
-        } else {
-            auto pads_begin_data = desc->pads_begin;
-            auto pads_begin_tensor = make_host_tensor({pads_begin_shape, data_types::i64, format::bfyx}, static_cast<void*>(pads_begin_data.data()));
-            const_data.emplace(1, pads_begin_tensor);
-
-            auto pads_end_mem = memory_deps.at(1);
-            cldnn::mem_lock<uint8_t, mem_lock_type::read> pads_end_lock(pads_end_mem, impl_param.get_stream());
-            const_data.emplace(2, make_host_tensor(pads_end_mem->get_layout(), pads_end_lock.data()));
-
-            output_shapes = ov::op::shape_infer(&op, input_shapes, ta);
-        }
-    } else {
-        std::ptrdiff_t val = desc->pad_value;
-
-        auto pads_begin_data = desc->pads_begin;
-        if (is_begin_mem && desc->pad_mode == ov::op::PadMode::CONSTANT) {
-            pads_begin_data = {val, val, val, val};
-        }
-        auto pads_begin_tensor = make_host_tensor({pads_begin_shape, data_types::i64, format::bfyx}, static_cast<void*>(pads_begin_data.data()));
-        const_data.emplace(1, pads_begin_tensor);
-
-        auto pads_end_data = desc->pads_end;
-        if (is_end_mem && desc->pad_mode == ov::op::PadMode::CONSTANT) {
-            pads_end_data = {val, val, val, val};
-        }
-        auto pads_end_tensor = make_host_tensor({pads_end_shape, data_types::i64, format::bfyx}, static_cast<void*>(pads_end_data.data()));
-        const_data.emplace(2, pads_end_tensor);
-
-        output_shapes = ov::op::shape_infer(&op, input_shapes, ta);
     }
 
+    int64_t begin_size = desc->pads_begin.size();
+    int64_t end_size = desc->pads_end.size();
+
+    layout pads_begin_layout = is_begin_mem ? impl_param.get_input_layout(begin_mem_idx) : layout({ begin_size }, data_types::i64, format::bfyx);
+    layout pads_end_layout = is_end_mem ? impl_param.get_input_layout(end_mem_idx) : layout({ end_size }, data_types::i64, format::bfyx);
+
+    std::vector<ShapeType> input_shapes = {
+        input0_layout.get<ShapeType>(),
+        pads_begin_layout.get<ShapeType>(),
+        pads_end_layout.get<ShapeType>(),
+    };
+
+    TensorsContainer const_data(&impl_param.get_stream());
+
+    auto pads_begin_data = desc->pads_begin;
+    auto pads_end_data = desc->pads_end;
+
+    if (is_begin_mem) {
+        const_data.emplace(1, memory_deps.at(begin_mem_idx));
+    } else {
+        const_data.emplace(1, make_tensor(pads_begin_layout, static_cast<void*>(pads_begin_data.data())));
+    }
+
+    if (is_end_mem) {
+        const_data.emplace(2, memory_deps.at(end_mem_idx));
+    } else {
+        const_data.emplace(2, make_tensor(pads_end_layout, static_cast<void*>(pads_end_data.data())));
+    }
+
+    auto ta = cldnn::make_tensor_accessor(const_data);
+    std::vector<ShapeType> output_shapes = ov::op::shape_infer(&op, input_shapes, ta);
+
     format output_format = format::adjust_to_rank(input0_layout.format, output_shapes[0].size());
 
     return { layout{output_shapes[0], output_type, output_format} };

src/plugins/intel_gpu/src/graph/broadcast.cpp

@@ -61,11 +61,11 @@ std::vector<layout> broadcast_inst::calc_output_layouts(broadcast_node const& /*
     auto axes_mapping = desc->axes_mapping.to_vector();
     ShapeType axes_mapping_shape = ov::Shape{axes_mapping.size()};
 
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
     if (third_input_needed) {
         input_shapes.emplace_back(axes_mapping_shape);
 
-        auto axes_mapping_tensor = make_host_tensor({axes_mapping_shape, data_types::i64, format::bfyx},
+        auto axes_mapping_tensor = make_tensor({axes_mapping_shape, data_types::i64, format::bfyx},
                                                     static_cast<void*>(axes_mapping.data()));
         const_data.emplace(2, axes_mapping_tensor);
     }
@@ -74,12 +74,11 @@ std::vector<layout> broadcast_inst::calc_output_layouts(broadcast_node const& /*
     if (constant_mem.count(1)) {
         auto target_shape_mem = constant_mem.at(1);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> target_shape_lock(target_shape_mem, impl_param.get_stream());
-        const_data.emplace(1, make_host_tensor(target_shape_mem->get_layout(), target_shape_lock.data()));
+        const_data.emplace(1, make_tensor(target_shape_mem->get_layout(), target_shape_lock.data()));
         output_shapes = ov::op::v3::shape_infer(&op, input_shapes, ov::make_tensor_accessor(const_data));
     } else if (impl_param.input_layouts.size() == 1) {
         // predefined pattern shape
-        auto target_shape_tensor = make_host_tensor({pattern_shape, data_types::i64, format::bfyx},
+        auto target_shape_tensor = make_tensor({pattern_shape, data_types::i64, format::bfyx}, static_cast<void*>(target_shape.data()));
-                                                     static_cast<void*>(target_shape.data()));
         const_data.emplace(1, target_shape_tensor);
         output_shapes = ov::op::v3::shape_infer(&op, input_shapes, ov::make_tensor_accessor(const_data));
     } else if (impl_param.input_layouts.size() >= 2) {

src/plugins/intel_gpu/src/graph/crop.cpp

@@ -56,17 +56,17 @@ std::vector<layout> crop_inst::calc_output_layouts(const crop_node& /*node*/, co
 
     // TODO: calling shape_infer for all cropped outpus is redundant... Need to optimize.
     if (desc->op_mode == cldnn::crop_ngraph_op_mode::variadic_split) {
-        std::map<size_t, ngraph::HostTensorPtr> const_data;
+        std::unordered_map<size_t, ov::Tensor> const_data;
 
         OPENVINO_ASSERT(impl_param.memory_deps.count(1) > 0, "[GPU] Can't find Crop(ngraph VariadicSplit op mode) axis values memory dependency");
         auto axis_values_mem = impl_param.memory_deps.at(1);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> axis_values_mem_lock(axis_values_mem, impl_param.get_stream());
-        const_data.emplace(1, make_host_tensor(axis_values_mem->get_layout(), axis_values_mem_lock.data()));
+        const_data.emplace(1, make_tensor(axis_values_mem->get_layout(), axis_values_mem_lock.data()));
 
         if (impl_param.memory_deps.count(2) > 0) {
             auto split_length_mem = impl_param.memory_deps.at(2);
             cldnn::mem_lock<uint8_t, mem_lock_type::read> split_length_mem_lock(split_length_mem, impl_param.get_stream());
-            const_data.emplace(2, make_host_tensor(split_length_mem->get_layout(), split_length_mem_lock.data()));
+            const_data.emplace(2, make_tensor(split_length_mem->get_layout(), split_length_mem_lock.data()));
 
             ov::op::v1::VariadicSplit op;
             output_shapes = shape_infer(&op, input_shapes, ov::make_tensor_accessor(const_data));
@@ -76,12 +76,12 @@ std::vector<layout> crop_inst::calc_output_layouts(const crop_node& /*node*/, co
             return { layout{out_shape, input0_layout.data_type, input0_layout.format } };
         }
     } else if (desc->op_mode == cldnn::crop_ngraph_op_mode::split) {
-        std::map<size_t, ngraph::HostTensorPtr> const_data;
+        std::unordered_map<size_t, ov::Tensor> const_data;
 
         OPENVINO_ASSERT(impl_param.memory_deps.count(1) > 0, "[GPU] Can't find Crop(ngraph Split op mode) axis values memory dependency");
         auto axis_values_mem = impl_param.memory_deps.at(1);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> axis_values_mem_lock(axis_values_mem, impl_param.get_stream());
-        const_data.emplace(1, make_host_tensor(axis_values_mem->get_layout(), axis_values_mem_lock.data()));
+        const_data.emplace(1, make_tensor(axis_values_mem->get_layout(), axis_values_mem_lock.data()));
 
         ov::op::v1::Split op;
         op.set_num_splits(desc->num_splits);

src/plugins/intel_gpu/src/graph/gather.cpp

@@ -88,8 +88,8 @@ std::vector<layout> gather_inst::calc_output_layouts(gather_node const& /*node*/
 
     int64_t axis = desc->axis;
 
-    auto axis_tensor = std::make_shared<ngraph::runtime::HostTensor>(ov::element::i64, ov::Shape{1}, static_cast<void*>(&axis));
+    auto axis_tensor = ov::Tensor(ov::element::i64, ov::Shape{1}, static_cast<void*>(&axis));
-    std::map<size_t, std::shared_ptr<ngraph::runtime::HostTensor>> const_data = {{2, axis_tensor}};
+    std::unordered_map<size_t, ov::Tensor> const_data = {{2, axis_tensor}};
     output_shapes = ov::op::shape_infer(&op, input_shapes, ov::make_tensor_accessor(const_data));
 
     format output_format = format::adjust_to_rank(input0_layout.format, output_shapes[0].size());

src/plugins/intel_gpu/src/graph/impls/cpu/range.cpp

@@ -46,8 +46,8 @@ struct range_impl : public typed_primitive_impl<range> {
         auto ev = stream.create_user_event(false);
         auto params = instance.get_impl_params();
 
-        ov::HostTensorVector input_host_tensors;
+        ov::TensorVector input_host_tensors;
-        ov::HostTensorVector output_host_tensors;
+        ov::TensorVector output_host_tensors;
 
         std::vector<memory::ptr> input_mem_ptrs;
         for (size_t i = 0; i < instance.dependencies().size(); i++)
@@ -58,9 +58,9 @@ struct range_impl : public typed_primitive_impl<range> {
         cldnn::mem_lock<uint8_t, mem_lock_type::read> output_lock(output_mem_ptr, stream);
 
         for (size_t i = 0; i < input_mem_ptrs.size(); i++)
-            input_host_tensors.push_back(make_host_tensor(params->input_layouts[i], input_mem_ptrs[i]->lock(stream, mem_lock_type::read)));
+            input_host_tensors.push_back(make_tensor(params->input_layouts[i], input_mem_ptrs[i]->lock(stream, mem_lock_type::read)));
 
-        output_host_tensors.push_back(make_host_tensor(params->output_layouts[0], output_lock.data()));
+        output_host_tensors.push_back(make_tensor(params->output_layouts[0], output_lock.data()));
 
         if (!op) {
             const auto output_dt = params->get_output_layout().data_type;

src/plugins/intel_gpu/src/graph/include/primitive_inst.h

@@ -8,6 +8,7 @@
 #include "intel_gpu/runtime/event.hpp"
 #include "intel_gpu/runtime/memory.hpp"
 #include "intel_gpu/runtime/lru_cache.hpp"
+#include "intel_gpu/runtime/tensor_accessor.hpp"
 #include "intel_gpu/graph/network.hpp"
 #include "intel_gpu/runtime/utils.hpp"
 #include "program_node.h"

src/plugins/intel_gpu/src/graph/non_max_suppression.cpp

@@ -38,12 +38,12 @@ std::vector<layout> non_max_suppression_inst::calc_output_layouts(non_max_suppre
     };
 
     auto& memory_deps = impl_param.memory_deps;
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
     if (memory_deps.count(2)) {
         auto max_output_boxes_per_class_mem = memory_deps.at(2);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> max_output_boxes_per_class_lock(max_output_boxes_per_class_mem,
                                                                                       impl_param.get_stream());
-        auto max_output_boxes_per_class_tensor = make_host_tensor(max_output_boxes_per_class_mem->get_layout(),
+        auto max_output_boxes_per_class_tensor = make_tensor(max_output_boxes_per_class_mem->get_layout(),
                                                                   max_output_boxes_per_class_lock.data());
         const_data.emplace(2, max_output_boxes_per_class_tensor);
 

src/plugins/intel_gpu/src/graph/one_hot.cpp

@@ -67,8 +67,8 @@ std::vector<layout> one_hot_inst::calc_output_layouts(const one_hot_node& /*node
 
     int64_t depth = desc->depth;
 
-    auto depth_tensor = std::make_shared<ngraph::runtime::HostTensor>(ov::element::i64, ov::Shape{1}, static_cast<void*>(&depth));
+    auto depth_tensor = ov::Tensor(ov::element::i64, ov::Shape{1}, static_cast<void*>(&depth));
-    std::map<size_t, std::shared_ptr<ngraph::runtime::HostTensor>> const_data = {
+    std::unordered_map<size_t, ov::Tensor> const_data = {
         {1, depth_tensor}
     };
     std::vector<ShapeType> output_shapes =

src/plugins/intel_gpu/src/graph/prior_box.cpp

@@ -440,7 +440,7 @@ std::vector<layout> prior_box_inst::calc_output_layouts(prior_box_node const& /*
         impl_param.get_input_layout(1).get<ShapeType>()
     };
     std::vector<ShapeType> output_shapes = {ShapeType()};
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
 
     auto& memory_deps = impl_param.memory_deps;
 
@@ -451,7 +451,7 @@ std::vector<layout> prior_box_inst::calc_output_layouts(prior_box_node const& /*
         cldnn::mem_lock<uint8_t, mem_lock_type::read> output_size_lock(output_size_mem, impl_param.get_stream());
         cldnn::mem_lock<uint8_t, mem_lock_type::read> img_size_lock(img_size_mem, impl_param.get_stream());
 
-        const_data.emplace(0, make_host_tensor(output_size_mem->get_layout(), output_size_lock.data()));
+        const_data.emplace(0, make_tensor(output_size_mem->get_layout(), output_size_lock.data()));
 
         auto p_param = const_cast<kernel_impl_params*>(&impl_param);
         if (output_size_mem->get_layout().data_type == cldnn::data_types::i64) {

src/plugins/intel_gpu/src/graph/random_uniform.cpp

@@ -33,18 +33,18 @@ std::vector<layout> random_uniform_inst::calc_output_layouts(random_uniform_node
                                             impl_param.get_input_layout(2).get_partial_shape() };
 
     auto& memory_deps = impl_param.memory_deps;
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
 
     auto run_shape_infer = [&]() {
         ov::op::v8::RandomUniform op;
         if (memory_deps.count(1) > 0 && memory_deps.count(2) > 0) {
             auto min_val = memory_deps.at(1);
             cldnn::mem_lock<uint8_t, mem_lock_type::read> min_val_lock(min_val, impl_param.get_stream());
-            const_data.emplace(1, make_host_tensor(min_val->get_layout(), min_val_lock.data()));
+            const_data.emplace(1, make_tensor(min_val->get_layout(), min_val_lock.data()));
 
             auto max_val = memory_deps.at(2);
             cldnn::mem_lock<uint8_t, mem_lock_type::read> max_val_lock(max_val, impl_param.get_stream());
-            const_data.emplace(2, make_host_tensor(max_val->get_layout(), max_val_lock.data()));
+            const_data.emplace(2, make_tensor(max_val->get_layout(), max_val_lock.data()));
 
             return ov::op::v8::shape_infer(&op, input_shapes, ov::make_tensor_accessor(const_data));
         } else {
@@ -55,7 +55,7 @@ std::vector<layout> random_uniform_inst::calc_output_layouts(random_uniform_node
     if (memory_deps.count(0) > 0) {
         auto output_shape = memory_deps.at(0);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> output_shape_lock(output_shape, impl_param.get_stream());
-        const_data.emplace(0, make_host_tensor(output_shape->get_layout(), output_shape_lock.data()));
+        const_data.emplace(0, make_tensor(output_shape->get_layout(), output_shape_lock.data()));
 
         output_shapes = run_shape_infer();
     } else {

src/plugins/intel_gpu/src/graph/range.cpp

@@ -34,21 +34,21 @@ std::vector<layout> range_inst::calc_output_layouts(range_node const& /*node*/,
     std::vector<ShapeType> output_shapes = {ShapeType::dynamic(1)};
     std::vector<ShapeType> input_shapes = {ov::Shape(), ov::Shape(), ov::Shape()};
 
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
     auto& memory_deps = impl_param.memory_deps;
 
     if (memory_deps.count(0) > 0 && memory_deps.count(1) > 0 && memory_deps.count(2) > 0) {
         auto start_mem = memory_deps.at(0);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> start_mem_lock(start_mem, impl_param.get_stream());
-        const_data.emplace(0, make_host_tensor(start_mem->get_layout(), start_mem_lock.data()));
+        const_data.emplace(0, make_tensor(start_mem->get_layout(), start_mem_lock.data()));
 
         auto stop_mem = memory_deps.at(1);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> stop_mem_lock(stop_mem, impl_param.get_stream());
-        const_data.emplace(1, make_host_tensor(stop_mem->get_layout(), stop_mem_lock.data()));
+        const_data.emplace(1, make_tensor(stop_mem->get_layout(), stop_mem_lock.data()));
 
         auto step_mem = memory_deps.at(2);
         cldnn::mem_lock<uint8_t, mem_lock_type::read> step_mem_lock(step_mem, impl_param.get_stream());
-        const_data.emplace(2, make_host_tensor(step_mem->get_layout(), step_mem_lock.data()));
+        const_data.emplace(2, make_tensor(step_mem->get_layout(), step_mem_lock.data()));
 
         output_shapes = shape_infer(&op, input_shapes, ov::make_tensor_accessor(const_data));
     }

src/plugins/intel_gpu/src/graph/reduce.cpp

@@ -107,8 +107,8 @@ std::vector<layout> reduce_inst::calc_output_layouts(reduce_node const& /*node*/
     std::vector<ShapeType> output_shapes = {ShapeType()};
 
     auto axes = desc->axes;
-    auto axes_tensor = std::make_shared<ngraph::runtime::HostTensor>(ov::element::i64, ov::Shape{axes.size()}, axes.data());
+    auto axes_tensor = ov::Tensor(ov::element::i64, ov::Shape{axes.size()}, axes.data());
-    std::map<size_t, std::shared_ptr<ngraph::runtime::HostTensor>> const_data = {{1, axes_tensor}};
+    std::unordered_map<size_t, ov::Tensor> const_data = {{1, axes_tensor}};
     auto ta = ov::make_tensor_accessor(const_data);
 
     // shape infer by mode

src/plugins/intel_gpu/src/graph/reshape.cpp

@@ -85,7 +85,7 @@ std::vector<layout> reshape_inst::calc_output_layouts(reshape_node const& /*node
         pattern_shape,
     };
 
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
     const auto ta = ov::make_tensor_accessor(const_data);
 
     auto run_shape_infer = [&](reshape::reshape_mode mode) {
@@ -120,13 +120,13 @@ std::vector<layout> reshape_inst::calc_output_layouts(reshape_node const& /*node
         cldnn::mem_lock<uint8_t, mem_lock_type::read> pattern_lock(pattern_mem, impl_param.get_stream());
 
         auto pattern_ptr = pattern_lock.data();
-        auto pattern_tensor = make_host_tensor(pattern_mem->get_layout(), pattern_ptr);
+        auto pattern_tensor = make_tensor(pattern_mem->get_layout(), pattern_ptr);
 
         const_data.emplace(1, pattern_tensor);
         run_shape_infer(prim->mode);
     } else {
         auto pattern_data = prim->output_pattern;
-        auto pattern_tensor = make_host_tensor({pattern_shape, data_types::i64, format::bfyx}, static_cast<void*>(pattern_data.data()));
+        auto pattern_tensor = make_tensor({pattern_shape, data_types::i64, format::bfyx}, static_cast<void*>(pattern_data.data()));
 
         const_data.emplace(1, pattern_tensor);
         run_shape_infer(prim->mode);

src/plugins/intel_gpu/src/graph/space_to_batch.cpp

@@ -101,7 +101,7 @@ std::vector<layout> space_to_batch_inst::calc_output_layouts(space_to_batch_node
         end_shape
     };
 
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
     if (desc->shape_constant) {
         auto block_sizes = tensor_to_vec(block_data, input0_format);
         auto begin_sizes = tensor_to_vec(begin_data, input0_format);
@@ -111,9 +111,9 @@ std::vector<layout> space_to_batch_inst::calc_output_layouts(space_to_batch_node
         auto begin_values = static_cast<void*>(begin_sizes.data());
         auto end_values = static_cast<void*>(end_sizes.data());
 
-        auto block_tensor = make_host_tensor({ block_shape, data_types::i32, input0_format }, block_values);
+        auto block_tensor = make_tensor({ block_shape, data_types::i32, input0_format }, block_values);
-        auto begin_tensor = make_host_tensor({ begin_shape, data_types::i32, input0_format }, begin_values);
+        auto begin_tensor = make_tensor({ begin_shape, data_types::i32, input0_format }, begin_values);
-        auto end_tensor = make_host_tensor({ end_shape, data_types::i32, input0_format }, end_values);
+        auto end_tensor = make_tensor({ end_shape, data_types::i32, input0_format }, end_values);
 
         const_data.emplace(1, block_tensor);
         const_data.emplace(2, begin_tensor);
@@ -129,9 +129,9 @@ std::vector<layout> space_to_batch_inst::calc_output_layouts(space_to_batch_node
         cldnn::mem_lock<uint8_t, mem_lock_type::read> lock2(begin_mem, impl_param.get_stream());
         cldnn::mem_lock<uint8_t, mem_lock_type::read> lock3(end_mem, impl_param.get_stream());
 
-        auto block_tensor = make_host_tensor(block_mem->get_layout(), lock1.data());
+        auto block_tensor = make_tensor(block_mem->get_layout(), lock1.data());
-        auto begin_tensor = make_host_tensor(begin_mem->get_layout(), lock2.data());
+        auto begin_tensor = make_tensor(begin_mem->get_layout(), lock2.data());
-        auto end_tensor = make_host_tensor(end_mem->get_layout(), lock3.data());
+        auto end_tensor = make_tensor(end_mem->get_layout(), lock3.data());
 
         const_data.emplace(1, block_tensor);
         const_data.emplace(2, begin_tensor);

src/plugins/intel_gpu/src/graph/strided_slice.cpp

@@ -67,12 +67,12 @@ std::vector<layout> strided_slice_inst::calc_output_layouts(strided_slice_node c
     op.set_shrink_axis_mask(desc->shrink_axis_mask);
     op.set_ellipsis_mask_mask(desc->ellipsis_mask);
 
-    std::map<size_t, ngraph::HostTensorPtr> const_data;
+    std::unordered_map<size_t, ov::Tensor> const_data;
     const auto ta = ov::make_tensor_accessor(const_data);
     if (!begin_data.empty() && !end_data.empty() && !strides_data.empty()) {
-        auto begin_tensor = make_host_tensor({ begin_shape, data_types::i64, format::bfyx }, static_cast<void*>(begin_data.data()));
+        auto begin_tensor = make_tensor({ begin_shape, data_types::i64, format::bfyx }, static_cast<void*>(begin_data.data()));
-        auto end_tensor = make_host_tensor({ end_shape, data_types::i64, format::bfyx }, static_cast<void*>(end_data.data()));
+        auto end_tensor = make_tensor({ end_shape, data_types::i64, format::bfyx }, static_cast<void*>(end_data.data()));
-        auto strides_tensor = make_host_tensor({ strides_shape, data_types::i64, format::bfyx }, static_cast<void*>(strides_data.data()));
+        auto strides_tensor = make_tensor({ strides_shape, data_types::i64, format::bfyx }, static_cast<void*>(strides_data.data()));
 
         const_data.emplace(1, begin_tensor);
         const_data.emplace(2, end_tensor);
@@ -88,9 +88,9 @@ std::vector<layout> strided_slice_inst::calc_output_layouts(strided_slice_node c
         cldnn::mem_lock<uint8_t, mem_lock_type::read> lock2(end_mem, impl_param.get_stream());
         cldnn::mem_lock<uint8_t, mem_lock_type::read> lock3(strides_mem, impl_param.get_stream());
 
-        auto begin_tensor = make_host_tensor(begin_mem->get_layout(), lock1.data());
+        auto begin_tensor = make_tensor(begin_mem->get_layout(), lock1.data());
-        auto end_tensor = make_host_tensor(end_mem->get_layout(), lock2.data());
+        auto end_tensor = make_tensor(end_mem->get_layout(), lock2.data());
-        auto strides_tensor = make_host_tensor(strides_mem->get_layout(), lock3.data());
+        auto strides_tensor = make_tensor(strides_mem->get_layout(), lock3.data());
 
         const_data.emplace(1, begin_tensor);
         const_data.emplace(2, end_tensor);

src/plugins/intel_gpu/tests/unit/shape_infer/pad_si_test.cpp

@@ -177,21 +177,28 @@ INSTANTIATE_TEST_SUITE_P(smoke, pad_test_non_constant_input_begin,
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {1, 0, 3, 7},
             ov::op::PadMode::CONSTANT, 1.f,
-            layout{ov::PartialShape{3, 4, 36, 48}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx}
         },
         {
             layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {1, 0, 3, 7},
             ov::op::PadMode::CONSTANT, 1.f,
-            layout{ov::PartialShape{{2, -1}, {1, -1}, {4, -1}, {8, -1}}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx}
+        },
+        {
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx},
+            layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
+            layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {1, 0, 3, 7},
+            ov::op::PadMode::EDGE, 1.f,
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx}
         },
         {
             layout{ov::PartialShape::dynamic(2), data_types::f32, format::bfyx},
             layout{ov::PartialShape{2}, data_types::i64, format::bfyx}, {},
             layout{ov::PartialShape{2}, data_types::i64, format::bfyx}, {1, 0},
             ov::op::PadMode::CONSTANT, 1.f,
-            layout{ov::PartialShape{{2, -1}, {1, -1}}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(2), data_types::f32, format::bfyx}
         }
     }));
 
@@ -234,21 +241,21 @@ INSTANTIATE_TEST_SUITE_P(smoke, pad_test_non_constant_input_end,
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {1, 0, 3, 7},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
             ov::op::PadMode::CONSTANT, 1.f,
-            layout{ov::PartialShape{3, 4, 36, 48}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx}
         },
         {
             layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {1, 0, 3, 7},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
             ov::op::PadMode::CONSTANT, 1.f,
-            layout{ov::PartialShape{{2, -1}, {1, -1}, {4, -1}, {8, -1}}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx}
         },
         {
             layout{ov::PartialShape::dynamic(2), data_types::f32, format::bfyx},
             layout{ov::PartialShape{2}, data_types::i64, format::bfyx}, {1, 0},
             layout{ov::PartialShape{2}, data_types::i64, format::bfyx}, {},
             ov::op::PadMode::CONSTANT, 1.f,
-            layout{ov::PartialShape{{2, -1}, {1, -1}}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(2), data_types::f32, format::bfyx}
         }
     }));
 
@@ -294,21 +301,92 @@ INSTANTIATE_TEST_SUITE_P(smoke, pad_test_non_constant_input_begin_end,
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
             ov::op::PadMode::CONSTANT, 1.f,
-            layout{ov::PartialShape{3, 5, 34, 42}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx}
         },
         {
             layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
             layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {},
-            ov::op::PadMode::CONSTANT, 1.f,
+            ov::op::PadMode::EDGE, 1.f,
-            layout{ov::PartialShape{{2, -1}, {2, -1}, {2, -1}, {2, -1}}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx}
         },
         {
             layout{ov::PartialShape::dynamic(2), data_types::f32, format::bfyx},
             layout{ov::PartialShape{2}, data_types::i64, format::bfyx}, {},
             layout{ov::PartialShape{2}, data_types::i64, format::bfyx}, {},
             ov::op::PadMode::CONSTANT, 1.f,
-            layout{ov::PartialShape{{2, -1}, {2, -1}}, data_types::f32, format::bfyx}
+            layout{ov::PartialShape::dynamic(2), data_types::f32, format::bfyx}
+        }
+    }));
+
+class pad_test_non_constant_input_begin_end_with_data : public testing::TestWithParam<pad_test_params> { };
+
+TEST_P(pad_test_non_constant_input_begin_end_with_data, shape_infer) {
+    auto p = GetParam();
+
+    auto& engine = get_test_engine();
+
+    auto input0_prim = std::make_shared<input_layout>("input0", p.in_layout);
+    auto input1_prim = std::make_shared<input_layout>("input1", p.pads_begin_layout);
+    auto input2_prim = std::make_shared<input_layout>("input2", p.pads_end_layout);
+
+    auto border_prim = std::make_shared<border>("output",
+                                                std::vector<input_info>({input_info("input0"), input_info("input1"), input_info("input2")}),
+                                                border::PAD_NON_CONST_INPUT::BEGIN | border::PAD_NON_CONST_INPUT::END,
+                                                p.pads_begin_data,
+                                                p.pads_end_data,
+                                                p.pad_mode,
+                                                p.pad_value);
+    cldnn::program prog(engine);
+
+    auto& input0_node = prog.get_or_create(input0_prim);
+    auto& input1_node = prog.get_or_create(input1_prim);
+    auto& input2_node = prog.get_or_create(input2_prim);
+    auto& border_node = prog.get_or_create(border_prim);
+
+    program_wrapper::add_connection(prog, input0_node, border_node);
+    program_wrapper::add_connection(prog, input1_node, border_node);
+    program_wrapper::add_connection(prog, input2_node, border_node);
+
+    auto begin_mem = engine.allocate_memory(p.pads_begin_layout);
+    auto end_mem = engine.allocate_memory(p.pads_end_layout);
+
+    set_values<int64_t>(begin_mem, p.pads_begin_data);
+    set_values<int64_t>(end_mem, p.pads_end_data);
+    auto impl_params = border_node.get_kernel_impl_params();
+    impl_params->memory_deps = {
+        {1, begin_mem},
+        {2, end_mem}
+    };
+
+    auto res = border_inst::calc_output_layouts<ov::PartialShape>(border_node, *impl_params);
+
+    ASSERT_EQ(res.size(), 1);
+    ASSERT_EQ(res[0], p.expected_layout);
+}
+
+INSTANTIATE_TEST_SUITE_P(smoke, pad_test_non_constant_input_begin_end_with_data,
+    testing::ValuesIn(std::vector<pad_test_params>{
+        {
+            layout{ov::PartialShape{1, 3, 32, 40}, data_types::f32, format::bfyx},
+            layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {1, 2, 3, 4},
+            layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {3, 2, 1, 0},
+            ov::op::PadMode::CONSTANT, 1.f,
+            layout{ov::PartialShape{5, 7, 36, 44}, data_types::f32, format::bfyx}
+        },
+        {
+            layout{ov::PartialShape::dynamic(4), data_types::f32, format::bfyx},
+            layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {1, 2, 3, 4},
+            layout{ov::PartialShape{4}, data_types::i64, format::bfyx}, {3, 2, 1, 0},
+            ov::op::PadMode::EDGE, 1.f,
+            layout{ov::PartialShape{{4, -1}, {4, -1}, {4, -1}, {4, -1}}, data_types::f32, format::bfyx}
+        },
+        {
+            layout{ov::PartialShape{10, 20}, data_types::f32, format::bfyx},
+            layout{ov::PartialShape{2}, data_types::i64, format::bfyx}, {1, 2},
+            layout{ov::PartialShape{2}, data_types::i64, format::bfyx}, {3, 4},
+            ov::op::PadMode::EDGE, 1.f,
+            layout{ov::PartialShape{14, 26}, data_types::f32, format::bfyx}
         }
     }));