Set dims for constant with multiple values (#11116)

* Also fix issue with data transfer not handling Tensor<std::string> correctly.

onnxruntime/core/framework/data_transfer.cc

@@ -62,7 +62,14 @@ common::Status CPUDataTransfer::CopyTensor(const Tensor& src, Tensor& dst, int /
 #endif
     // Copying only happens between two same size tensors.
     ORT_ENFORCE(src.SizeInBytes() == dst.SizeInBytes());
-    memcpy(dst_data, src_data, src.SizeInBytes());
+    if (!src.IsDataTypeString()) {
+      memcpy(dst_data, src_data, src.SizeInBytes());
+    } else {
+      const auto* src_strings = src.Data<std::string>();
+      auto* dst_strings = dst.MutableData<std::string>();
+      std::copy(src_strings, src_strings + src.Shape().Size(), dst_strings);
+    }
+
     return Status::OK();
 #ifdef ENABLE_TRAINING
   }

onnxruntime/core/framework/tensorprotoutils.cc

@@ -228,7 +228,7 @@ Status UnpackTensorWithExternalData(const ONNX_NAMESPACE::TensorProto& /*tensor*
                                     /*out*/ std::string* /*p_data*/) {
   return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "External data type cannot be STRING.");
 }
-#endif  //!defined(ORT_MINIMAL_BUILD)
+#endif  //! defined(ORT_MINIMAL_BUILD)
 
 // implementation of type specific unpack of data contained within the TensorProto
 template <typename T>
@@ -637,10 +637,10 @@ Status TensorProtoToTensor(const Env& env, const ORTCHAR_T* model_path,
     raw_data = const_cast<char*>(tensor_proto.raw_data().data());
     // TODO The line above has const-correctness issues. Below is a possible fix which copies the tensor_proto data
     //      into a writeable buffer. However, it requires extra memory which may exceed the limit for certain tests.
-    //auto buffer = std::make_unique<char[]>(tensor_proto.raw_data().size());
-    //std::memcpy(buffer.get(), tensor_proto.raw_data().data(), tensor_proto.raw_data().size());
-    //deleter_for_file_data.d = OrtCallback{DeleteCharArray, buffer.get()};
-    //raw_data = buffer.release();
+    // auto buffer = std::make_unique<char[]>(tensor_proto.raw_data().size());
+    // std::memcpy(buffer.get(), tensor_proto.raw_data().data(), tensor_proto.raw_data().size());
+    // deleter_for_file_data.d = OrtCallback{DeleteCharArray, buffer.get()};
+    // raw_data = buffer.release();
     raw_data_len = tensor_proto.raw_data().size();
   }
 
@@ -807,6 +807,7 @@ common::Status ConstantNodeProtoToTensorProto(const ONNX_NAMESPACE::NodeProto& n
     case AttributeProto_AttributeType_FLOATS:
       tensor.set_data_type(TensorProto_DataType_FLOAT);
       *tensor.mutable_float_data() = constant_attribute.floats();
+      tensor.add_dims(constant_attribute.floats().size());
       break;
     case AttributeProto_AttributeType_INT:
       tensor.set_data_type(TensorProto_DataType_INT64);
@@ -815,6 +816,7 @@ common::Status ConstantNodeProtoToTensorProto(const ONNX_NAMESPACE::NodeProto& n
     case AttributeProto_AttributeType_INTS:
       tensor.set_data_type(TensorProto_DataType_INT64);
       *tensor.mutable_int64_data() = constant_attribute.ints();
+      tensor.add_dims(constant_attribute.ints().size());
       break;
     case AttributeProto_AttributeType_STRING:
       tensor.set_data_type(TensorProto_DataType_STRING);
@@ -823,6 +825,7 @@ common::Status ConstantNodeProtoToTensorProto(const ONNX_NAMESPACE::NodeProto& n
     case AttributeProto_AttributeType_STRINGS: {
       tensor.set_data_type(TensorProto_DataType_STRING);
       *tensor.mutable_string_data() = constant_attribute.strings();
+      tensor.add_dims(constant_attribute.strings().size());
       break;
     }
 #if !defined(DISABLE_SPARSE_TENSORS)

onnxruntime/test/providers/cpu/tensor/constant_test.cc

@@ -0,0 +1,44 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#include "core/session/inference_session.h"
+#include "core/util/math.h"
+#include "test/providers/provider_test_utils.h"
+#include "gtest/gtest.h"
+
+namespace onnxruntime {
+namespace test {
+
+// test that a multi value constant has the correct shape
+TEST(ConstantOpTest, MultiValueConstant_floats) {
+  OpTester test("Constant", 14, kOnnxDomain);
+  test.AddAttribute("value_floats", std::vector<float>{0.f, 1.f, 2.f, 3.f});
+  test.AddOutput<float>("Y", {4}, {0.f, 1.f, 2.f, 3.f});
+  test.Run();
+}
+
+TEST(ConstantOpTest, MultiValueConstant_ints) {
+  OpTester test("Constant", 14, kOnnxDomain);
+  test.AddAttribute("value_ints", std::vector<int64_t>{0, 1, 2, 3});
+  test.AddOutput<int64_t>("Y", {4}, {0, 1, 2, 3});
+  test.Run();
+}
+
+TEST(ConstantOpTest, MultiValueConstant_strings) {
+  OpTester test("Constant", 14, kOnnxDomain);
+  test.AddAttribute("value_strings", std::vector<std::string>{"zero", "one", "two", "three"});
+  test.AddOutput<std::string>("Y", {4}, {"zero", "one", "two", "three"});
+  test.Run();
+}
+
+// regression test - https://github.com/microsoft/onnxruntime/issues/11091
+TEST(ConstantOpTest, GH11091) {
+  auto model_uri = ORT_TSTR("testdata/constant_floats.onnx");
+  SessionOptions so;
+  so.session_logid = "ConstantOpTest.GH11091";
+  InferenceSession session_object{so, GetEnvironment()};
+  ASSERT_STATUS_OK(session_object.Load(model_uri));
+  ASSERT_STATUS_OK(session_object.Initialize());
+}
+}  // namespace test
+}  // namespace onnxruntime

onnxruntime/test/providers/provider_test_utils.cc

@@ -330,7 +330,7 @@ struct TensorCheck<BFloat16> {
     /// XXX: May need to adjust threshold as BFloat is coarse
     float threshold = 0.001f;
 #if defined(USE_TENSORRT) || defined(ENABLE_TRAINING) || defined(USE_CUDA) || defined(USE_ROCM)
-    threshold = 0.05f; // expect at least 95% close
+    threshold = 0.05f;  // expect at least 95% close
 #endif
     for (int i = 0; i < size; ++i) {
       if (std::isnan(f_expected[i])) {
@@ -653,7 +653,7 @@ void OpTester::AddSparseCsrTensorData(std::vector<Data>& data,
 
 void OpTester::AddSparseCsrTensorStrings(std::vector<Data>& data,
                                          const char* name,
-                                         gsl::span<const  int64_t> dims,
+                                         gsl::span<const int64_t> dims,
                                          gsl::span<const std::string> values,
                                          gsl::span<const int64_t> inner_indices,
                                          gsl::span<const int64_t> outer_indices,