Layer dev paulm (#2567)

* commetns for dml graph transformer
fixed ort value passing using the allocatir info

* fixed and coded maps and sequences across the abi

* cleaned up w4's
cleaned up the model info ABI
delayload directml.dll from winml

* cleaned up namepsace aliases.
renamed _winmla to winmla
this was good PR feedback from tiago a while back.

* moved files from inc to lib\api.core
cleaned up some of the cmake

* staged changes

* making windowsAI azure dev ops work.

* code review comments.

* revert changes

include/onnxruntime/core/session/onnxruntime_cxx_api.h

@@ -75,7 +75,9 @@ ORT_DEFINE_RELEASE(Value);
 // This is used internally by the C++ API. This is the common base class used by the wrapper objects.
 template <typename T>
 struct Base {
-  Base() = default;
+  Base() {
+    p_ = nullptr;
+  }
   Base(T* p) : p_{p} {
     if (!p) throw Ort::Exception("Allocation failure", ORT_FAIL);
   }
@@ -95,7 +97,7 @@ struct Base {
   }
 
   T** put() noexcept {
-    //ASSERT(p_ == nullptr);
+    assert(p_ == nullptr);
     return &p_;
   }
 

onnxruntime/core/providers/dml/DmlExecutionProvider/inc/IWinmlExecutionProvider.h

@@ -10,8 +10,6 @@
 
 #include "core/framework/op_kernel.h"
 
-#include <DirectML.h>
-
 struct AbstractOperatorDesc;
 interface IMLOperatorTensor;
 

onnxruntime/core/session/IOBinding.cc

@@ -8,11 +8,6 @@
 #include "core/framework/utils.h"
 
 namespace onnxruntime {
-IOBinding::~IOBinding() {
-  feeds_.clear();
-  outputs_.clear();
-}
-
 IOBinding::IOBinding(const SessionState& session_state) : session_state_(session_state) {
 }
 

onnxruntime/core/session/IOBinding.h

@@ -38,7 +38,6 @@ class SessionState;
  */
 class IOBinding {
  public:
-  ~IOBinding();
   /**
    * Call repeatedly to bind as many inputs as required.
    * If called again for the same name will replace an existing value.

winml/adapter/MLValueHelpers.h

@@ -1,7 +0,0 @@
-// Copyright (c) Microsoft Corporation.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace Windows::AI::MachineLearning {
-}  // namespace Windows::AI::MachineLearning

winml/adapter/TensorBaseHelpers.h

@@ -1,8 +0,0 @@
-// Copyright (c) Microsoft Corporation.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace Windows::AI::MachineLearning {
-
-}  // namespace Windows::AI::MachineLearning

winml/lib/Api/LearningModelBinding.cpp

@@ -15,7 +15,7 @@ using namespace WinML;
 
 namespace winrt::Windows::AI::MachineLearning::implementation {
 LearningModelBinding::LearningModelBinding(
-  Windows::AI::MachineLearning::LearningModelSession const& session) try : m_session(session) {
+    Windows::AI::MachineLearning::LearningModelSession const& session) try : m_session(session) {
   session.as<winmlp::LearningModelSession>()->CheckClosed();
   WINML_THROW_IF_FAILED(OrtGetWinMLAdapter(adapter_.put()));
 }
@@ -219,12 +219,12 @@ bool LearningModelBinding::HasKey(hstring const& key) {
 void LearningModelBinding::Split(
     Windows::Foundation::Collections::IMapView<hstring, Windows::Foundation::IInspectable>& first,
     Windows::Foundation::Collections::IMapView<hstring, Windows::Foundation::IInspectable>& second) {
-  ORT_UNUSED_PARAMETER(first);
-  ORT_UNUSED_PARAMETER(second);
-  throw hresult_not_implemented();
+  // the winrt api guide states: 
+  // If the IMapView instance cannot be split, then both the first and second parameters are null when the method returns.
+  first = nullptr;
+  second = nullptr;
 }
 
-
 ONNXTensorElementDataType STDMETHODCALLTYPE GetONNXTensorElementDataType(winml::TensorKind kind) {
   if (kind == TensorKind::Float) {
     return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT;
@@ -275,7 +275,6 @@ bool LearningModelBinding::IsOfMapType(const Ort::Value& ort_value, TensorKind k
 };
 
 bool LearningModelBinding::IsOfVectorMapType(const Ort::Value& ort_value, TensorKind key_kind, TensorKind value_kind) {
-
   if (ort_value.GetTypeInfo().GetONNXType() != ONNX_TYPE_SEQUENCE)
     return false;
 
@@ -305,7 +304,7 @@ ILearningModelFeatureValue LearningModelBinding::CreateUnboundOuputFeatureValue(
       if (descriptor.Kind() == LearningModelFeatureKind::Image) {
         using namespace Windows::Graphics::Imaging;
         // TODO: this format for unbound output needs more discussion
-        BitmapPixelFormat format = descriptor.as<ImageFeatureDescriptor>()->BitmapPixelFormat();        
+        BitmapPixelFormat format = descriptor.as<ImageFeatureDescriptor>()->BitmapPixelFormat();
         uint32_t width = static_cast<uint32_t>(ort_value.GetTensorTypeAndShapeInfo().GetShape()[3]);
         uint32_t height = static_cast<uint32_t>(ort_value.GetTensorTypeAndShapeInfo().GetShape()[2]);
         uint32_t batchSize = static_cast<uint32_t>(ort_value.GetTensorTypeAndShapeInfo().GetShape()[0]);
@@ -388,7 +387,7 @@ ILearningModelFeatureValue LearningModelBinding::CreateUnboundOuputFeatureValue(
 
 Windows::Foundation::IInspectable LearningModelBinding::CreateUnboundOutput(
     const std::string& name,
-    Ort::Value& ort_value) {  
+    Ort::Value& ort_value) {
   // Find valid binding port
   auto bindingPort = FindValidBinding(
       m_session.Model(),
@@ -539,8 +538,8 @@ HRESULT LearningModelBinding::BindInput(const std::string& name, Ort::Value& ml_
   if (ml_value.IsTensor()) {
     Ort::Value new_mlvalue = Ort::Value(nullptr);
     WINML_THROW_IF_FAILED(m_session.as<LearningModelSession>()
-      ->GetIInferenceSession()
-      ->CopyOneInputAcrossDevices(name.c_str(), ml_value, new_mlvalue.put()));
+                              ->GetIInferenceSession()
+                              ->CopyOneInputAcrossDevices(name.c_str(), ml_value, new_mlvalue.put()));
     add_or_replace(rc.first, rc.second, new_mlvalue);
   } else {
     add_or_replace(rc.first, rc.second, ml_value);
@@ -574,8 +573,7 @@ const std::vector<std::string>& LearningModelBinding::GetInputNames() const {
 
 const std::vector<Ort::Value>& LearningModelBinding::GetInputs() const { return inputs_; }
 
-void LearningModelBinding::BindUnboundOutputs()
-{
+void LearningModelBinding::BindUnboundOutputs() {
   auto& bound_output_names = GetOutputNames();
   std::unordered_set<std::string> bound_output_names_set(
       bound_output_names.begin(),

winml/lib/Api/LearningModelSession.cpp

@@ -109,7 +109,7 @@ void LearningModelSession::Initialize() {
     device_impl->GetDeviceQueue(),
     session_builder.put()));
 
-  Ort::SessionOptions options;
+  Ort::SessionOptions options(nullptr);
   WINML_THROW_IF_FAILED(session_builder->CreateSessionOptions(options.put()));
 
   // Make onnxruntime apply the batch size override, if any

winml/lib/Api/impl/MapBase.h

@@ -47,7 +47,6 @@ struct MapBase : winrt::implements<
 
   using LotusKey = typename ValidLotusType<TKey>::Type;
   using LotusValue = typename ValidLotusType<TValue>::Type;
-  //using LotusMap = std::map<LotusKey, LotusValue>;
   using LotusMap = std::pair<std::vector<LotusKey>, std::vector<LotusValue>>;
   using ABIMap = ::winrt::Windows::Foundation::Collections::IMap<TKey, TValue>;
   using ABIMapView = ::winrt::Windows::Foundation::Collections::IMapView<TKey, TValue>;

winml/lib/Api/impl/Tensor.h

@@ -4,7 +4,6 @@
 #pragma once
 
 #include "TensorBuffer.h"
-#include "MLValueHelpers.h"
 
 // we further specialize these base types for a couple of extra tensor element types
 namespace Ort {