Skip to content

Commit 63e3152

Browse files
sgbihumlukasze
sgbihu
and
mlukasze
authored
Fix access violation when Tensor is 0-size shape constant and scalar … (#34895)
### Details: - This case will make `element_count == 0`, `m_shape.size() == 1`, but `shape_size(m_shape) == 0`. - The `repeats` in `src\plugins\intel_cpu\src\nodes\broadcast.cpp` is `0`. But the `srcDims.size() = 1`. That make the program access `repeats[0]` and cause the access violation. Port the #34874 to 2026.1 ### Tickets: - [CVS-183409](https://jira.devtools.intel.com/browse/CVS-183409) ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).* --------- Co-authored-by: Michal Lukaszewski <michal.lukaszewski@intel.com>
1 parent f37a2e7 commit 63e3152

4 files changed

Lines changed: 173 additions & 106 deletions

File tree

‎src/frontends/onnx/frontend/src/core/tensor.cpp‎

Lines changed: 114 additions & 106 deletions
Original file line numberDiff line numberDiff line change
@@ -471,114 +471,122 @@ std::shared_ptr<ov::op::v0::Constant> Tensor::get_ov_constant() const {
471471
"The size of the external data file does not match the byte size of an initializer '" + get_name() +
472472
"' in the model");
473473
}
474-
} else if (element_count == shape_size(m_shape) && m_tensor_proto != nullptr) {
475-
switch (m_tensor_proto->data_type()) {
476-
case TensorProto_DataType::TensorProto_DataType_FLOAT:
477-
case TensorProto_DataType::TensorProto_DataType_DOUBLE:
478-
case TensorProto_DataType::TensorProto_DataType_INT32:
479-
case TensorProto_DataType::TensorProto_DataType_INT64:
480-
case TensorProto_DataType::TensorProto_DataType_UINT32:
481-
case TensorProto_DataType::TensorProto_DataType_UINT64:
482-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data_ptr());
483-
break;
484-
case TensorProto_DataType::TensorProto_DataType_INT4:
485-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int8_t>().data());
486-
break;
487-
case TensorProto_DataType::TensorProto_DataType_INT8:
488-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int8_t>().data());
489-
break;
490-
case TensorProto_DataType::TensorProto_DataType_INT16:
491-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int16_t>().data());
492-
break;
493-
case TensorProto_DataType::TensorProto_DataType_UINT4:
494-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint8_t>().data());
495-
break;
496-
case TensorProto_DataType::TensorProto_DataType_UINT8:
497-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint8_t>().data());
498-
break;
499-
case TensorProto_DataType::TensorProto_DataType_UINT16:
500-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint16_t>().data());
501-
break;
502-
case TensorProto_DataType::TensorProto_DataType_BOOL:
503-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<char>().data());
504-
break;
505-
case TensorProto_DataType::TensorProto_DataType_BFLOAT16:
506-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::bfloat16>().data());
507-
break;
508-
case TensorProto_DataType::TensorProto_DataType_FLOAT16:
509-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float16>().data());
510-
break;
511-
case TensorProto_DataType::TensorProto_DataType_FLOAT8E4M3FN:
512-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float8_e4m3>().data());
513-
break;
514-
case TensorProto_DataType::TensorProto_DataType_FLOAT8E5M2:
515-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float8_e5m2>().data());
516-
break;
517-
case TensorProto_DataType::TensorProto_DataType_STRING:
518-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<std::string>().data());
519-
break;
520-
default:
521-
ONNX_UNSUPPORTED_DATA_TYPE(
522-
m_tensor_proto->data_type(),
523-
"BOOL, BFLOAT16, FLOAT8E4M3FN, FLOAT8E5M2, FLOAT, FLOAT16, DOUBLE, INT4, INT8, INT16, INT32, INT64, "
524-
"UINT4, UINT8, UINT16, UINT32, UINT64, STRING");
474+
} else {
475+
const auto shape_elements = shape_size(m_shape);
476+
if (element_count != shape_elements && !(element_count == 0 && m_shape.empty())) {
477+
FRONT_END_THROW(
478+
"The number of elements implied by the data size does not match the shape of an initializer '" +
479+
get_name() + "' in the model");
525480
}
526-
} else if (element_count == shape_size(m_shape) && m_tensor_place != nullptr) {
527-
switch (m_tensor_place->get_element_type()) {
528-
case ov::element::f32:
529-
case ov::element::f64:
530-
case ov::element::i32:
531-
case ov::element::i64:
532-
case ov::element::u32:
533-
case ov::element::u64:
534-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data_ptr());
535-
break;
536-
case ov::element::i4:
537-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int8_t>().data());
538-
break;
539-
case ov::element::i8:
540-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int8_t>().data());
541-
break;
542-
case ov::element::i16:
543-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int16_t>().data());
544-
break;
545-
case ov::element::u4:
546-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint8_t>().data());
547-
break;
548-
case ov::element::u8:
549-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint8_t>().data());
550-
break;
551-
case ov::element::u16:
552-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint16_t>().data());
553-
break;
554-
case ov::element::boolean:
555-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<char>().data());
556-
break;
557-
case ov::element::bf16:
558-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::bfloat16>().data());
559-
break;
560-
case ov::element::f16:
561-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float16>().data());
562-
break;
563-
case ov::element::f8e4m3:
564-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float8_e4m3>().data());
565-
break;
566-
case ov::element::f8e5m2:
567-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float8_e5m2>().data());
568-
break;
569-
case ov::element::string:
570-
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<std::string>().data());
571-
break;
572-
default:
573-
ONNX_UNSUPPORTED_DATA_TYPE(
574-
m_tensor_proto->data_type(),
575-
"BOOL, BFLOAT16, FLOAT8E4M3FN, FLOAT8E5M2, FLOAT, FLOAT16, DOUBLE, INT4, INT8, INT16, INT32, INT64, "
576-
"UINT4, UINT8, UINT16, UINT32, UINT64, STRING");
481+
if (element_count == 0) {
482+
constant = common::make_failsafe_constant(ov_type);
483+
} else if (m_tensor_proto != nullptr) {
484+
switch (m_tensor_proto->data_type()) {
485+
case TensorProto_DataType::TensorProto_DataType_FLOAT:
486+
case TensorProto_DataType::TensorProto_DataType_DOUBLE:
487+
case TensorProto_DataType::TensorProto_DataType_INT32:
488+
case TensorProto_DataType::TensorProto_DataType_INT64:
489+
case TensorProto_DataType::TensorProto_DataType_UINT32:
490+
case TensorProto_DataType::TensorProto_DataType_UINT64:
491+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data_ptr());
492+
break;
493+
case TensorProto_DataType::TensorProto_DataType_INT4:
494+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int8_t>().data());
495+
break;
496+
case TensorProto_DataType::TensorProto_DataType_INT8:
497+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int8_t>().data());
498+
break;
499+
case TensorProto_DataType::TensorProto_DataType_INT16:
500+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int16_t>().data());
501+
break;
502+
case TensorProto_DataType::TensorProto_DataType_UINT4:
503+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint8_t>().data());
504+
break;
505+
case TensorProto_DataType::TensorProto_DataType_UINT8:
506+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint8_t>().data());
507+
break;
508+
case TensorProto_DataType::TensorProto_DataType_UINT16:
509+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint16_t>().data());
510+
break;
511+
case TensorProto_DataType::TensorProto_DataType_BOOL:
512+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<char>().data());
513+
break;
514+
case TensorProto_DataType::TensorProto_DataType_BFLOAT16:
515+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::bfloat16>().data());
516+
break;
517+
case TensorProto_DataType::TensorProto_DataType_FLOAT16:
518+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float16>().data());
519+
break;
520+
case TensorProto_DataType::TensorProto_DataType_FLOAT8E4M3FN:
521+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float8_e4m3>().data());
522+
break;
523+
case TensorProto_DataType::TensorProto_DataType_FLOAT8E5M2:
524+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float8_e5m2>().data());
525+
break;
526+
case TensorProto_DataType::TensorProto_DataType_STRING:
527+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<std::string>().data());
528+
break;
529+
default:
530+
ONNX_UNSUPPORTED_DATA_TYPE(m_tensor_proto->data_type(),
531+
"BOOL, BFLOAT16, FLOAT8E4M3FN, FLOAT8E5M2, FLOAT, FLOAT16, DOUBLE, INT4, "
532+
"INT8, INT16, INT32, INT64, "
533+
"UINT4, UINT8, UINT16, UINT32, UINT64, STRING");
534+
}
535+
} else if (m_tensor_place != nullptr) {
536+
switch (m_tensor_place->get_element_type()) {
537+
case ov::element::f32:
538+
case ov::element::f64:
539+
case ov::element::i32:
540+
case ov::element::i64:
541+
case ov::element::u32:
542+
case ov::element::u64:
543+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data_ptr());
544+
break;
545+
case ov::element::i4:
546+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int8_t>().data());
547+
break;
548+
case ov::element::i8:
549+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int8_t>().data());
550+
break;
551+
case ov::element::i16:
552+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<int16_t>().data());
553+
break;
554+
case ov::element::u4:
555+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint8_t>().data());
556+
break;
557+
case ov::element::u8:
558+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint8_t>().data());
559+
break;
560+
case ov::element::u16:
561+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<uint16_t>().data());
562+
break;
563+
case ov::element::boolean:
564+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<char>().data());
565+
break;
566+
case ov::element::bf16:
567+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::bfloat16>().data());
568+
break;
569+
case ov::element::f16:
570+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float16>().data());
571+
break;
572+
case ov::element::f8e4m3:
573+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float8_e4m3>().data());
574+
break;
575+
case ov::element::f8e5m2:
576+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<ov::float8_e5m2>().data());
577+
break;
578+
case ov::element::string:
579+
constant = std::make_shared<ov::op::v0::Constant>(ov_type, m_shape, get_data<std::string>().data());
580+
break;
581+
default:
582+
ONNX_UNSUPPORTED_DATA_TYPE(m_tensor_proto->data_type(),
583+
"BOOL, BFLOAT16, FLOAT8E4M3FN, FLOAT8E5M2, FLOAT, FLOAT16, DOUBLE, INT4, "
584+
"INT8, INT16, INT32, INT64, "
585+
"UINT4, UINT8, UINT16, UINT32, UINT64, STRING");
586+
}
587+
} else {
588+
FRONT_END_THROW("Tensor shape doesn't match data size");
577589
}
578-
} else if (element_count == 0 && m_shape.size() == 0) {
579-
constant = common::make_failsafe_constant(ov_type);
580-
} else {
581-
FRONT_END_THROW("Tensor shape doesn't match data size");
582590
}
583591

584592
if (m_tensor_proto != nullptr && m_tensor_proto->has_name()) {

‎src/frontends/onnx/frontend/src/op/reshape.cpp‎

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -6,6 +6,7 @@
66

77
#include "core/operator_set.hpp"
88
#include "exceptions.hpp"
9+
#include "utils/common.hpp"
910
#include "utils/reshape.hpp"
1011
using namespace ov::op;
1112

@@ -23,6 +24,12 @@ ov::OutputVector reshape(const ov::frontend::onnx::Node& node) {
2324
// Since opset 5 the target shape is provided as input
2425
if (ov_inputs.size() == 2) {
2526
pattern = ov_inputs.at(1);
27+
if (common::is_failsafe_node(pattern.get_node_shared_ptr())) {
28+
// in case the "shape" input is connected to a failsafe node created in place of an invalid initializer
29+
// the target shape should be ignored and this Expand operation should not modify its input tensor
30+
// the Broadcast created below should be eliminated later on by an appropriate optimization pass
31+
pattern = v0::Constant::create(ov::element::i64, {0}, {});
32+
}
2633
} else {
2734
// Added in onnx reshape version 14
2835
special_zero = !node.get_attribute_value<int64_t>("allowzero", 0);

‎src/frontends/onnx/tests/models/expand_scalar_failsafe_node.prototxt‎

Lines changed: 41 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,41 @@
1+
ir_version: 13
2+
producer_name: "OpenVINO ONNX Frontend"
3+
graph {
4+
name: ""
5+
node {
6+
input: "inputA"
7+
input: "shape"
8+
output: "expandOutput_0"
9+
name: "_0"
10+
op_type: "Expand"
11+
}
12+
initializer {
13+
dims: 0
14+
data_type: 7
15+
name: "shape"
16+
raw_data: ""
17+
}
18+
input {
19+
name: "inputA"
20+
type {
21+
tensor_type {
22+
elem_type: 1
23+
shape {
24+
}
25+
}
26+
}
27+
}
28+
output {
29+
name: "expandOutput_0"
30+
type {
31+
tensor_type {
32+
elem_type: 1
33+
shape {
34+
}
35+
}
36+
}
37+
}
38+
}
39+
opset_import {
40+
version: 25
41+
}

‎src/frontends/onnx/tests/onnx_import.in.cpp‎

Lines changed: 11 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -5517,6 +5517,17 @@ OPENVINO_TEST(${BACKEND_NAME}, onnx_model_expand_failsafe_node) {
55175517
test_case.run();
55185518
}
55195519

5520+
OPENVINO_TEST(${BACKEND_NAME}, onnx_model_expand_scalar_failsafe_node) {
5521+
const auto model = convert_model("expand_scalar_failsafe_node.onnx");
5522+
5523+
auto test_case = ov::test::TestCase(model, s_device);
5524+
const auto input_data = std::vector<float>{1.0f};
5525+
test_case.add_input<float>(input_data);
5526+
// the target shape is an empty constant so the Expand operation should not modify the input shape
5527+
test_case.add_expected_output<float>(input_data);
5528+
test_case.run();
5529+
}
5530+
55205531
OPENVINO_TEST(${BACKEND_NAME}, onnx_scan15_fib_like) {
55215532
const auto model = convert_model("scan15_fib_like.onnx");
55225533

0 commit comments

Comments
 (0)