[C#][TVM EP] Fix issues related to using TVM EP in C# front-end (#12958)

Changes in this PR:
* Update building of Nuget package for TVM EP
* Update of documentation  for using TVM EP in C#

docs/TVM_EP.md

@@ -146,7 +146,7 @@ To verify the CUDA installation use `nvcc --version` in cmd.
 <br>
 <br>
 
-Build ONNX Runtime with TVM Execution Provider from source:
+#### **Build ONNX Runtime with TVM Execution Provider from source (Python):**
 - Use command line and clone sources from github:
 ```cmd
 git clone --recursive https://github.com/Microsoft/onnxruntime
@@ -183,8 +183,24 @@ print(onnxruntime.get_available_providers())
 pip uninstall onnxruntime-tvm
 ```
 
+#### **Build ONNX Runtime with TVM Execution Provider from source (C#):**
+- Use command line and clone sources from github:
+```cmd
+git clone --recursive https://github.com/Microsoft/onnxruntime
+cd onnxruntime
+```
+- CPU build:
+
+Make sure you download [nuget.exe](https://docs.microsoft.com/en-us/nuget/install-nuget-client-tools#nugetexe-cli) and add path to it into `PATH` env.
+```
+build.bat --config Release --build_nuget --skip_tests --parallel --use_tvm --skip_onnx_tests --cmake_generator "Visual Studio 17 2022" --llvm_config llvm-config.exe
+```
+- Install C# nuget package for TVM EP. Default path to the package is `<path_to_onnxruntime_root>\build\Windows\Release\Release`.
+
+
 ## Configuration options
 TVM Executor Provider can be configured with the following provider options:
+1. Python
 ```python
 po = [dict(executor=tvm_executor_type,
            so_folder=folder_with_pretuned_files,
@@ -201,6 +217,31 @@ po = [dict(executor=tvm_executor_type,
            input_shapes = input_shapes_str)]
 tvm_session = onnxruntime.InferenceSession(model_path, providers=["TvmExecutionProvider"], provider_options=po)
 ```
+
+2. C#
+
+Currently, only precompiled models are supported in C# (see the related section below).
+
+```CSharp
+SessionOptions session_options = new SessionOptions{};
+string tvm_ep_options = 
+  $"executor: {tvm_executor_type}, " +
+  $"so_folder: {folder_with_pretuned_files}, " +
+  $"check_hash: {check_hash}, " +
+  $"hash_file_path: {hash_file_path}, " +
+  $"target: {client_target}, " +
+  $"target_host: {client_target_host}, " +
+  $"opt_level: {client_opt_level}, " +
+  $"freeze_weights: {freeze}, " +
+  $"to_nhwc: {layout_transform}, " +
+  $"tuning_type: {tvm_optimizer_type}, " +
+  $"tuning_file_path: {client_tuning_logfile}, " +
+  $"input_names: {input_names_str}, " +
+  $"input_shapes: {input_shapes_str}";
+
+session_options.AppendExecutionProvider_Tvm(tvm_ep_options);
+using var tvm_session = new InferenceSession(modelFilePath, session_options);
+```
 <br>
 
 - `executor` is executor type used by TVM. There is choice between two types: GraphExecutor and VirtualMachine which are corresponded to "graph" and "vm" tags. VirtualMachine is used by default.

onnxruntime/core/providers/tvm/tvm_ep_options.cc

@@ -31,6 +31,9 @@ constexpr const char* kInputShapes = "input_shapes";
 
 static const std::unordered_set<std::string> valid_keys {
   std::string{kExecutor},
+  std::string{kSoFolder},
+  std::string{kCheckHash},
+  std::string{kHashFilePath},
   std::string{kTarget},
   std::string{kTargetHost},
   std::string{kOptLevel},

tools/ci_build/build.py

@@ -2116,6 +2116,7 @@ def build_nuget_package(
     elif use_cuda:
         package_name = '/p:OrtPackageId="Microsoft.ML.OnnxRuntime.Gpu"'
     elif use_tvm:
+        execution_provider = '/p:ExecutionProvider="tvm"'
         package_name = '/p:OrtPackageId="Microsoft.ML.OnnxRuntime.Tvm"'
     else:
         # use the solution file that includes Xamarin mobile targets
@@ -2157,7 +2158,7 @@ def build_nuget_package(
             run_subprocess(cmd_args, cwd=csharp_build_dir)
 
         if is_windows():
-            if use_openvino:
+            if use_openvino or use_tvm:
                 # user needs to make sure nuget is installed and added to the path variable
                 nuget_exe = "nuget.exe"
             else:

tools/nuget/generate_nuspec_for_native_nuget.py

@@ -119,7 +119,7 @@ def parse_arguments():
         required=False,
         default="None",
         type=str,
-        choices=["cuda", "dnnl", "openvino", "tensorrt", "snpe", "None"],
+        choices=["cuda", "dnnl", "openvino", "tensorrt", "snpe", "tvm", "None"],
         help="The selected execution provider for this build.",
     )
     parser.add_argument("--dependency_id", required=False, default="None", type=str, help="ependency id.")
@@ -335,6 +335,7 @@ def generate_files(list, args):
             "tensorrt_ep_shared_lib": "onnxruntime_providers_tensorrt.dll",
             "openvino_ep_shared_lib": "onnxruntime_providers_openvino.dll",
             "cuda_ep_shared_lib": "onnxruntime_providers_cuda.dll",
+            "tvm_ep_shared_lib": "onnxruntime_providers_tvm.lib",
             "onnxruntime_perf_test": "onnxruntime_perf_test.exe",
             "onnx_test_runner": "onnx_test_runner.exe",
         }
@@ -387,6 +388,14 @@ def generate_files(list, args):
         + '" target="build\\native\\include" />'
     )
 
+    if args.execution_provider == "tvm":
+        files_list.append(
+            "<file src="
+            + '"'
+            + os.path.join(args.sources_path, "include\\onnxruntime\\core\\providers\\tvm\\tvm_provider_factory.h")
+            + '" target="build\\native\\include" />'
+        )
+
     if args.execution_provider == "openvino":
         files_list.append(
             "<file src="
@@ -609,6 +618,38 @@ def generate_files(list, args):
             + '\\native" />'
         )
 
+    if args.execution_provider == "tvm":
+        files_list.append(
+            "<file src="
+            + '"'
+            + os.path.join(args.native_build_path, nuget_dependencies["providers_shared_lib"])
+            + runtimes_target
+            + args.target_architecture
+            + '\\native" />'
+        )
+        files_list.append(
+            "<file src="
+            + '"'
+            + os.path.join(args.native_build_path, nuget_dependencies["tvm_ep_shared_lib"])
+            + runtimes_target
+            + args.target_architecture
+            + '\\native" />'
+        )
+
+        tvm_build_path = os.path.join(args.ort_build_path, args.build_config, "_deps", "tvm-build")
+        if is_windows():
+            files_list.append(
+                "<file src="
+                + '"'
+                + os.path.join(tvm_build_path, args.build_config, nuget_dependencies["tvm"])
+                + runtimes_target
+                + args.target_architecture
+                + '\\native" />'
+            )
+        else:
+            # TODO(agladyshev): Add support for Linux.
+            raise RuntimeError("Now only Windows is supported for TVM EP.")
+
     if args.execution_provider == "openvino":
         openvino_path = get_env_var("INTEL_OPENVINO_DIR")
         files_list.append(