Improved ov::Tensor behavior (#7683)

* Improved ov::Tensor behavior

* Fixed python test for setShape on preallocated

* Fixed clang-format

inference-engine/ie_bridges/python/tests/test_Blob.py

@@ -130,11 +130,14 @@ def test_set_shape():
     assert blob.tensor_desc.dims == [1, 4, 128, 128]
     assert blob.buffer.shape == (1, 4, 128, 128)
 
+
+def test_cannot_set_shape_preallocated_memory():
+    tensor_desc = TensorDesc("FP32", [1, 3, 127, 127], "NHWC")
     array = np.ones([1, 3, 127, 127], dtype=np.float32)
     blob = Blob(tensor_desc, array)
+    with pytest.raises(RuntimeError) as e:
         blob.set_shape([1, 4, 128, 128])
-    assert blob.tensor_desc.dims == [1, 4, 128, 128]
+    assert "Cannot call setShape for Blobs created on top of preallocated memory" in str(e.value)
-    assert blob.buffer.shape == (1, 4, 128, 128)
 
 
 @pytest.mark.ngraph_dependent_test

inference-engine/src/inference_engine/include/ie/ie_blob.h

@@ -192,22 +192,7 @@ public:
      *
      * @param dims new shape
      */
-    void setShape(const SizeVector& dims) {
+    void setShape(const SizeVector& dims);
-        if (properProduct(dims) > properProduct(getTensorDesc().getDims())) {
-            // New blob shape requires more memory than old one -- reallocate
-            if (!deallocate())
-                IE_THROW() << "Cannot deallocate blob while an attempt to enlarge blob area in setShape.";
-
-            // Old and new ranks should match as well as layouts
-            getTensorDesc().setDims(dims);
-
-            allocate();
-            // no way to detect if allocation is successful other than map/unmap that we wouldn't like to do here
-        } else {
-            // Don't shrink area when new size fit the existing area
-            getTensorDesc().setDims(dims);
-        }
-    }
 
     /**
      * @deprecated Cast to MemoryBlob and use new wlock/rwlock API instead.

inference-engine/src/inference_engine/include/ie/ie_layouts.h

@@ -84,7 +84,7 @@ public:
     /**
      * @brief Returns the vector of order
      *
-     * @return order
+     * @return order of dimensions
      */
     const SizeVector& getOrder() const {
         return order;
@@ -93,7 +93,7 @@ public:
     /**
      * @brief Returns the per-dimension offset vector
      *
-     * @return offsets
+     * @return offsets in elements
      */
     const SizeVector& getOffsetPaddingToData() const {
         return offsetPaddingToData;
@@ -102,7 +102,7 @@ public:
     /**
      * @brief Returns the offset to the current memory block
      *
-     * @return offset
+     * @return offset in elements
      */
     size_t getOffsetPadding() const {
         return offsetPadding;
@@ -111,7 +111,7 @@ public:
     /**
      * @brief Returns strides for each dimension
      *
-     * @return strides
+     * @return strides in elements
      */
     const SizeVector& getStrides() const {
         return strides;

inference-engine/src/inference_engine/src/ie_blob_common.cpp

@@ -7,8 +7,55 @@
 #include <vector>
 
 #include "ie_blob.h"
+#include "system_allocator.hpp"
 
 namespace InferenceEngine {
+
+void Blob::setShape(const SizeVector& dims) {
+    // we don't want to allow setShape for:
+    // 1. ROI cases
+    {
+        size_t denseStride = 1;
+        const auto& blockedDims = getTensorDesc().getBlockingDesc().getBlockDims();
+        const auto& strides = getTensorDesc().getBlockingDesc().getStrides();
+
+        for (size_t i = 1; i <= strides.size(); i++) {
+            if (denseStride != strides[strides.size() - i]) {
+                IE_THROW() << "Blob::setShape requires dense blob";
+            }
+            denseStride *= blockedDims[blockedDims.size() - i];
+        }
+    }
+
+    // 2. Blobs created on top of preallocated memory
+    if (std::dynamic_pointer_cast<InferenceEngine::details::PreAllocator>(getAllocator())) {
+        IE_THROW() << "Cannot call setShape for Blobs created on top of preallocated memory.";
+    }
+
+    if (properProduct(dims) > properProduct(getTensorDesc().getDims())) {
+        // New blob shape requires more memory than old one -- reallocate
+        if (!deallocate()) {
+            IE_THROW() << "Cannot deallocate blob while an attempt to enlarge blob area in setShape.";
+        }
+
+        // Old and new ranks should match as well as layouts
+        getTensorDesc().setDims(dims);
+
+        allocate();
+        // no way to detect if allocation is successful other than map/unmap
+        // that we wouldn't like to do here; but for cases when we use SystemMemoryAllocator
+        // we can do it
+        if (std::dynamic_pointer_cast<InferenceEngine::SystemMemoryAllocator>(getAllocator())) {
+            if (buffer() == nullptr) {
+                IE_THROW() << "Failed to allocate memory in Blob::setShape";
+            }
+        }
+    } else {
+        // Don't shrink area when new size fit the existing area
+        getTensorDesc().setDims(dims);
+    }
+}
+
 Blob::Ptr Blob::createROI(const ROI& roi) const {
     if (getTensorDesc().getLayout() == Layout::NCHW || getTensorDesc().getLayout() == Layout::NHWC) {
         return createROI({roi.id, 0, roi.posY, roi.posX},

inference-engine/src/inference_engine/src/ie_layouts.cpp

@@ -288,6 +288,21 @@ BlockingDesc::BlockingDesc(const SizeVector& blocked_dims,
     if (blocked_dims.size() != dimOffsets.size())
         IE_THROW() << "Offsets are not initialized for all dimensions.";
     this->offsetPaddingToData = dimOffsets;
+
+    // check that strides are valid
+    {
+        size_t denseStride = 1;
+
+        for (size_t i = 1; i <= strides.size(); i++) {
+            if (denseStride > strides[strides.size() - i]) {
+                IE_THROW() << "Stride in " << (strides.size() - i)
+                           << "-th dimension "
+                              "is not valid; actual "
+                           << strides[strides.size() - i] << ", should be >= " << denseStride << std::endl;
+            }
+            denseStride = std::max(strides[strides.size() - i], denseStride) * blocked_dims[blocked_dims.size() - i];
+        }
+    }
 }
 
 BlockingDesc::BlockingDesc(const SizeVector& dims, Layout layout) : offsetPadding(0) {

ngraph/core/CMakeLists.txt

@@ -18,6 +18,7 @@ set(IE_SHARED_SRCS
     "${IE_SRC_ROOT}/system_allocator.cpp"
     "${IE_SRC_ROOT}/blob_factory.cpp"
     "${IE_SRC_ROOT}/ie_blob_common.cpp"
+    "${IE_SRC_ROOT}/system_allocator.cpp"
     "${IE_SRC_ROOT}/ie_layouts.cpp")
 set(MIXED_SRC ${IE_SHARED_SRCS}
     "${CMAKE_CURRENT_SOURCE_DIR}/src/runtime/allocator.cpp"

ngraph/core/include/openvino/runtime/tensor.hpp

@@ -69,16 +69,10 @@ public:
      * @param type Tensor element type
      * @param shape Tensor shape
      * @param host_ptr Pointer to pre-allocated host memory
-     * @param size Optional size of allocated host memory in elements. If it is not set (default is `0`), the size of
-     * memory supposed to be not less then ov::shape_size(shape) * type.size() in bytes.
      * @param strides Optional strides parameters in elements. Strides are supposed to be equal to shape if they are not
      * set
      */
-    Tensor(const element::Type type,
+    Tensor(const element::Type type, const Shape& shape, void* host_ptr, const Strides& strides = {});
-           const Shape& shape,
-           void* host_ptr,
-           const size_t size = 0,
-           const Strides& strides = {});
 
     /**
      * @brief Constructs region of interest (ROI) tensor form another tensor.

ngraph/core/src/runtime/ov_tensor.cpp

@@ -38,11 +38,7 @@ Tensor::Tensor(const element::Type element_type, const Shape& shape, const Alloc
     _impl->allocate();
 }
 
-Tensor::Tensor(const element::Type element_type,
+Tensor::Tensor(const element::Type element_type, const Shape& shape, void* host_ptr, const Strides& strides) {
-               const Shape& shape,
-               void* host_ptr,
-               const size_t size,
-               const Strides& strides) {
     ie::SizeVector blk_order(shape.size());
     std::iota(blk_order.begin(), blk_order.end(), 0);
     ie::SizeVector dim_offset(shape.size(), 0);
@@ -50,12 +46,6 @@ Tensor::Tensor(const element::Type element_type,
     if (strides.empty()) {
         blk_strides = ov::row_major_strides(shape);
     } else {
-        OPENVINO_ASSERT(shape.size() == strides.size(),
-                        "shape.size() (",
-                        shape.size(),
-                        ") must be equal to strides.size() (",
-                        strides.size(),
-                        ")");
         blk_strides.assign(strides.begin(), strides.end());
     }
 
@@ -64,8 +54,7 @@ Tensor::Tensor(const element::Type element_type,
                                          ie::TensorDesc{ie::details::convertPrecision(element_type),
                                                         shape,
                                                         ie::BlockingDesc{shape, blk_order, 0, dim_offset, blk_strides}},
-                                         host_ptr,
+                                         host_ptr);
-                                         size);
     } catch (const std::exception& ex) {
         throw ov::Exception(ex.what());
     } catch (...) {

ngraph/test/ov_tensor_test.cpp

@@ -11,6 +11,7 @@
 #include <openvino/core/strides.hpp>
 #include <openvino/core/type/element_type.hpp>
 
+#include "openvino/core/except.hpp"
 #include "openvino/runtime/allocator.hpp"
 #include "openvino/runtime/tensor.hpp"
 
@@ -58,7 +59,7 @@ TEST_F(OVTensorTest, canCreateTensorUsingMockAllocator) {
 TEST_F(OVTensorTest, canAccessExternalData) {
     ov::Shape shape = {1, 1, 3};
     float data[] = {5.f, 6.f, 7.f};
-    ov::runtime::Tensor t{ov::element::f32, shape, data, 3};
+    ov::runtime::Tensor t{ov::element::f32, shape, data};
     {
         float* ptr = t.data<float>();
         ASSERT_EQ(ptr[2], 7);
@@ -74,7 +75,8 @@ TEST_F(OVTensorTest, canAccessExternalData) {
 TEST_F(OVTensorTest, canAccessExternalDataWithStrides) {
     ov::Shape shape = {2, 3};
     float data[] = {5.f, 6.f, 7.f, 0.f, 1.f, 42.f, 3.f, 0.f};
-    ov::runtime::Tensor t{ov::element::f32, shape, data, 8, {4, 1}};
+    ov::runtime::Tensor t{ov::element::f32, shape, data, {4, 1}};
+    ASSERT_EQ(ov::Strides({4, 1}), t.get_strides());
     {
         ASSERT_EQ((ov::Shape{2, 3}), t.get_shape());
         float* ptr = t.data<float>();
@@ -90,7 +92,17 @@ TEST_F(OVTensorTest, cannotCreateTensorWithExternalNullptr) {
 TEST_F(OVTensorTest, cannotCreateTensorWithWrongStrides) {
     ov::Shape shape = {2, 3};
     float data[] = {5.f, 6.f, 7.f, 0.f, 1.f, 42.f, 3.f, 0.f};
-    ASSERT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, 8, {4, 1, 2}), ov::Exception);
+    {
+        // strides.size() != shape.size()
+        EXPECT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {6, 3, 1}), ov::Exception);
+    }
+    {
+        // strides values are element-wise >= ov::row_major_strides(shape) values
+        EXPECT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {2, 1}), ov::Exception);
+        EXPECT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {3, 0}), ov::Exception);
+        EXPECT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {3, 2}), ov::Exception);
+        EXPECT_NO_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {6, 2}));
+    }
 }
 
 TEST_F(OVTensorTest, saveDimsAndSizeAfterMove) {
@@ -115,36 +127,65 @@ TEST_F(OVTensorTest, saveDimsAndSizeAfterMove) {
 
 // SetShape
 TEST_F(OVTensorTest, canSetShape) {
+    const ov::Shape origShape({1, 2, 3});
     ov::runtime::Tensor t{ov::element::f32, {1, 2, 3}};
     const ov::Shape newShape({4, 5, 6});
-    ASSERT_EQ(t.get_shape(), (ov::Shape{1, 2, 3}));
+
+    const void* orig_data = t.data();
+    ASSERT_EQ(t.get_shape(), origShape);
     ASSERT_NO_THROW(t.set_shape({4, 5, 6}));
     ASSERT_EQ(newShape, t.get_shape());
     ASSERT_EQ(ov::row_major_strides(newShape), t.get_strides());
+    ASSERT_NE(orig_data, t.data());
 
     // check that setShape for copy changes original Tensor
     {
         ov::runtime::Tensor t2 = t;
-        t2.set_shape(newShape);
+        ASSERT_NO_THROW(t2.set_shape(newShape));
         ASSERT_EQ(newShape, t.get_shape());
         ASSERT_EQ(t2.get_shape(), t.get_shape());
+        orig_data = t.data();
+    }
+
+    // set_shape for smaller memory - does not perform reallocation
+    {
+        t.set_shape(origShape);
+        ASSERT_EQ(origShape, t.get_shape());
+        ASSERT_EQ(orig_data, t.data());
     }
 }
 
+TEST_F(OVTensorTest, cannotSetShapeOnPreallocatedMemory) {
+    float data[4 * 5 * 6 * 2];
+    ov::runtime::Tensor t{ov::element::f32, {1, 2, 3}, data};
+    const ov::Shape newShape({4, 5, 6});
+
+    ASSERT_THROW(t.set_shape(newShape), ov::Exception);
+}
+
 TEST_F(OVTensorTest, makeRangeRoiTensor) {
-    ov::runtime::Tensor t{ov::element::i8, {1, 3, 6, 5}};  // RGBp picture of size (WxH) = 5x6
+    ov::runtime::Tensor t{ov::element::i32, {1, 3, 6, 5}};  // RGBp picture of size (WxH) = 5x6
     ov::runtime::Tensor roi_tensor{t, {0, 0, 1, 2}, {1, 3, 5, 4}};
     ov::Shape ref_shape = {1, 3, 4, 2};
-    ptrdiff_t ref_offset = 7;
+    ptrdiff_t ref_offset_elems = 7;
+    ptrdiff_t ref_offset_bytes = ref_offset_elems * ov::element::i32.size();
     ov::Strides ref_strides = {90, 30, 5, 1};
     ASSERT_EQ(roi_tensor.get_shape(), ref_shape);
-    ASSERT_EQ(roi_tensor.data<int8_t>() - t.data<int8_t>(), ref_offset);
+    ASSERT_EQ(roi_tensor.data<int32_t>() - t.data<int32_t>(), ref_offset_elems);
-    ASSERT_EQ(reinterpret_cast<uint8_t*>(roi_tensor.data()) - reinterpret_cast<uint8_t*>(t.data()), ref_offset);
+    ASSERT_EQ(reinterpret_cast<uint8_t*>(roi_tensor.data()) - reinterpret_cast<uint8_t*>(t.data()), ref_offset_bytes);
     ASSERT_EQ(roi_tensor.get_strides(), t.get_strides());
     ASSERT_EQ(ref_strides, roi_tensor.get_strides());
     ASSERT_EQ(roi_tensor.get_element_type(), t.get_element_type());
 }
 
+TEST_F(OVTensorTest, cannotSetShapeOnRoiTensor) {
+    ov::runtime::Tensor t{ov::element::i32, {1, 3, 6, 5}};  // RGBp picture of size (WxH) = 5x6
+    ov::runtime::Tensor roi_tensor{t, {0, 0, 1, 2}, {1, 3, 5, 4}};
+    const ov::Shape newShape({4, 5, 6});
+
+    ASSERT_THROW(roi_tensor.set_shape(newShape), ov::Exception);
+}
+
 TEST_F(OVTensorTest, makeRangeRoiTensorInt4) {
     ov::runtime::Tensor t{ov::element::i4, {1, 6, 5, 3}};  // RGB picture of size (WxH) = 5x6
     ov::runtime::Tensor roi_tensor{t, {0, 1, 2, 0}, {1, 5, 4, 3}};