[CoreML] more performace flag (#22975)

### Description
refactor unsquzee's implementation
add more flags to boost peformance.
add profile flag


### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->

---------

Co-authored-by: jicwen <[email protected]>
Co-authored-by: wejoncy <[email protected]>
Co-authored-by: Scott McKay <[email protected]>

include/onnxruntime/core/providers/coreml/coreml_provider_factory.h

@@ -47,8 +47,20 @@ enum COREMLFlags {
 // and SessionOptionsAppendExecutionProvider (C API). For the old API, use COREMLFlags instead.
 static const char* const kCoremlProviderOption_MLComputeUnits = "MLComputeUnits";
 static const char* const kCoremlProviderOption_ModelFormat = "ModelFormat";
+// same as COREML_FLAG_ONLY_ALLOW_STATIC_INPUT_SHAPES
 static const char* const kCoremlProviderOption_RequireStaticInputShapes = "RequireStaticInputShapes";
 static const char* const kCoremlProviderOption_EnableOnSubgraphs = "EnableOnSubgraphs";
+// provided by https://developer.apple.com/documentation/coreml/mloptimizationhints-swift.struct/specializationstrategy-swift.property
+// Core ML segments the model’s compute graph and specializes each segment for the target compute device.
+// This process can affect the model loading time and the prediction latency.
+// Use this option to tailor the specialization strategy for your model.
+static const char* const kCoremlProviderOption_SpecializationStrategy = "SpecializationStrategy";
+// Profile the Core ML MLComputePlan.
+// This logs the hardware each operator is dispatched to and the estimated execution time.
+// Intended for developer usage but provide useful diagnostic information if performance is not as expected.
+static const char* const kCoremlProviderOption_ProfileComputePlan = "ProfileComputePlan";
+// please refer to https://developer.apple.com/documentation/coreml/mlmodelconfiguration/allowlowprecisionaccumulationongpu
+static const char* const kCoremlProviderOption_AllowLowPrecisionAccumulationOnGPU = "AllowLowPrecisionAccumulationOnGPU";
 
 #ifdef __cplusplus
 extern "C" {

onnxruntime/core/providers/coreml/builders/impl/batch_norm_op_builder.cc

@@ -151,7 +151,7 @@ bool BatchNormalizationOpBuilder::IsOpSupportedImpl(const Node& node, const OpBu
     return false;
   }
 
-#if defined(TARGET_OS_IOS) && defined(TARGET_CPU_X86_64)
+#if defined(TARGET_OS_IOS) && defined(TARGET_CPU_X86_64) && TARGET_OS_IOS && TARGET_CPU_X86_64
   // To Pass IOS pipeline https://dev.azure.com/onnxruntime/onnxruntime/_build?definitionId=134&_a=summary
   auto input_dtype = input_defs[0]->TypeAsProto()->tensor_type().elem_type();
   if (input_dtype == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16 && input_params.coreml_version < 7) {

onnxruntime/core/providers/coreml/builders/impl/reduction_op_builder.cc

@@ -133,9 +133,8 @@ bool ReductionOpBuilder::IsOpSupportedImpl(const Node& node, const OpBuilderInpu
     return false;
   }
 
-#if defined(TARGET_OS_IOS) && defined(TARGET_CPU_X86_64)
-  // to pass https://dev.azure.com/onnxruntime/onnxruntime/_build/results?buildId=1563483&view=logs&j=f7cc61a9-cc70-56e7-b06c-4668ca17e426
-  // ReductionOpTest.ReduceSum_half_bert
+#if defined(TARGET_OS_IOS) && defined(TARGET_CPU_X86_64) && TARGET_OS_IOS && TARGET_CPU_X86_64
+  // skip ReductionOpTest.ReduceSum_half_bert because reduce_sum will output all zeros
   int32_t input_type;
   GetType(*input_defs[0], input_type, logger);
   if (node.OpType() == "ReduceSum" && input_type == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16) {

onnxruntime/core/providers/coreml/builders/impl/squeeze_op_builder.cc

@@ -13,6 +13,10 @@
 #include "core/optimizer/initializer.h"
 #include "core/providers/cpu/tensor/unsqueeze.h"
 
+#ifdef __APPLE__
+#include <TargetConditionals.h>
+#endif
+
 namespace onnxruntime {
 namespace coreml {
 
@@ -54,32 +58,50 @@ void SqueezeOpBuilder::AddInitializersToSkip(ModelBuilder& model_builder, const
   }
 }
 
+#if defined(COREML_ENABLE_MLPROGRAM)
+void HandleX86ArchUnsqueezeScalarInput(ModelBuilder& model_builder,
+                                       const Node& node, const logging::Logger& logger) {
+  const auto& input_defs(node.InputDefs());
+  TensorShapeVector axes;
+  GetAxes(model_builder, node, axes);
+
+  std::vector<int64_t> input_shape;
+  GetShape(*input_defs[0], input_shape, logger);
+  auto op = model_builder.CreateOperation(node, "reshape");
+  AddOperationInput(*op, "x", input_defs[0]->Name());
+  TensorShapeVector output_shape = UnsqueezeBase::ComputeOutputShape(TensorShape(input_shape), axes);
+  AddOperationInput(*op, "shape", model_builder.AddConstant(op->type(), "shape", AsSpan(output_shape)));
+  AddOperationOutput(*op, *node.OutputDefs()[0]);
+  model_builder.AddOperation(std::move(op));
+}
+#endif
+
 Status SqueezeOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder,
                                                const Node& node,
                                                [[maybe_unused]] const logging::Logger& logger) const {
   std::unique_ptr<COREML_SPEC::NeuralNetworkLayer> layer = model_builder.CreateNNLayer(node);
-  const auto& input_defs(node.InputDefs());
   auto* coreml_squeeze = layer->mutable_squeeze();
   TensorShapeVector axes;
   GetAxes(model_builder, node, axes);
-  std::vector<int64_t> input_shape;
-  GetShape(*input_defs[0], input_shape, logger);
 #if defined(COREML_ENABLE_MLPROGRAM)
+  const auto& input_defs(node.InputDefs());
   if (model_builder.CreateMLProgram()) {
     using namespace CoreML::Specification::MILSpec;
 
-    std::string_view coreml_op_type = node.OpType() == "Squeeze" ? "squeeze" : "reshape";
+#if defined(TARGET_CPU_X86_64) && TARGET_CPU_X86_64
+    // expand_dims has limited requirements for static shape, however, X86_64 has a bug that it can't handle scalar input
+    if (node.OpType() == "Unsqueeze" && input_defs[0]->Shape()->dim_size() < 2) {
+      HandleX86ArchUnsqueezeScalarInput(model_builder, node, logger);
+      return Status::OK();
+    }
+#endif
+    std::string_view coreml_op_type = node.OpType() == "Squeeze" ? "squeeze" : "expand_dims";
     std::unique_ptr<Operation> op = model_builder.CreateOperation(node, coreml_op_type);
     AddOperationInput(*op, "x", input_defs[0]->Name());
 
-    if (coreml_op_type == "squeeze") {
-      if (!axes.empty()) {
-        // coreml squeeze op does support negative axes
-        AddOperationInput(*op, "axes", model_builder.AddConstant(op->type(), "axes", AsSpan(axes)));
-      }
-    } else {
-      TensorShapeVector output_shape = UnsqueezeBase::ComputeOutputShape(TensorShape(input_shape), axes);
-      AddOperationInput(*op, "shape", model_builder.AddConstant(op->type(), "shape", AsSpan(output_shape)));
+    if (!axes.empty()) {
+      // coreml supports negative axes
+      AddOperationInput(*op, "axes", model_builder.AddConstant(op->type(), "axes", AsSpan(axes)));
     }
     AddOperationOutput(*op, *node.OutputDefs()[0]);
     model_builder.AddOperation(std::move(op));

onnxruntime/core/providers/coreml/builders/model_builder.cc

@@ -408,7 +408,7 @@ ModelBuilder::ModelBuilder(const GraphViewer& graph_viewer, const logging::Logge
     : graph_viewer_(graph_viewer),
       logger_(logger),
       coreml_version_(coreml_version),
-      coreml_compute_unit_(coreml_options.ComputeUnits()),
+      coreml_options_(coreml_options),
       create_ml_program_(coreml_options.CreateMLProgram()),
       model_output_path_(GetModelOutputPath(create_ml_program_)),
       onnx_input_names_(std::move(onnx_input_names)),
@@ -989,7 +989,7 @@ Status ModelBuilder::LoadModel(std::unique_ptr<Model>& model) {
                                     get_sanitized_io_info(std::move(input_output_info_)),
                                     std::move(scalar_outputs_),
                                     std::move(int64_outputs_),
-                                    logger_, coreml_compute_unit_);
+                                    logger_, coreml_options_);
   } else
 #endif
   {
@@ -999,7 +999,7 @@ Status ModelBuilder::LoadModel(std::unique_ptr<Model>& model) {
                                     std::move(input_output_info_),
                                     std::move(scalar_outputs_),
                                     std::move(int64_outputs_),
-                                    logger_, coreml_compute_unit_);
+                                    logger_, coreml_options_);
   }
 
   return model->LoadModel();  // load using CoreML API, including compilation

onnxruntime/core/providers/coreml/builders/model_builder.h

@@ -7,6 +7,7 @@
 #include "core/graph/graph_viewer.h"
 #include "core/providers/coreml/builders/coreml_spec.h"
 #include "core/providers/coreml/model/model.h"
+#include "core/providers/coreml/coreml_options.h"
 
 #if defined(COREML_ENABLE_MLPROGRAM)
 // coremltools classes
@@ -22,8 +23,6 @@ class StorageWriter;
 #endif
 
 namespace onnxruntime {
-class CoreMLOptions;
-
 namespace coreml {
 
 class IOpBuilder;
@@ -218,7 +217,7 @@ class ModelBuilder {
   const GraphViewer& graph_viewer_;
   const logging::Logger& logger_;
   const int32_t coreml_version_;
-  const uint32_t coreml_compute_unit_;
+  CoreMLOptions coreml_options_;
   const bool create_ml_program_;         // ML Program (CoreML5, iOS 15+, macOS 12+) or NeuralNetwork (old)
   const std::string model_output_path_;  // create_ml_program_ ? dir for mlpackage : filename for mlmodel
 

onnxruntime/core/providers/coreml/coreml_options.cc

@@ -63,11 +63,14 @@ void CoreMLOptions::ValidateAndParseProviderOption(const ProviderOptions& option
       {"MLProgram", COREML_FLAG_CREATE_MLPROGRAM},
       {"NeuralNetwork", COREML_FLAG_USE_NONE},
   };
-  std::unordered_set<std::string> valid_options = {
+  const std::unordered_set<std::string_view> valid_options = {
       kCoremlProviderOption_MLComputeUnits,
       kCoremlProviderOption_ModelFormat,
       kCoremlProviderOption_RequireStaticInputShapes,
       kCoremlProviderOption_EnableOnSubgraphs,
+      kCoremlProviderOption_SpecializationStrategy,
+      kCoremlProviderOption_ProfileComputePlan,
+      kCoremlProviderOption_AllowLowPrecisionAccumulationOnGPU,
   };
   // Validate the options
   for (const auto& option : options) {
@@ -90,6 +93,16 @@ void CoreMLOptions::ValidateAndParseProviderOption(const ProviderOptions& option
       require_static_shape_ = option.second == "1";
     } else if (kCoremlProviderOption_EnableOnSubgraphs == option.first) {
       enable_on_subgraph_ = option.second == "1";
+    } else if (kCoremlProviderOption_SpecializationStrategy == option.first) {
+      if (option.second != "Default" && option.second != "FastPrediction") {
+        ORT_THROW("Invalid value for option ", option.first, ": ", option.second,
+                  ". Valid values are Default and FastPrediction.");
+      }
+      strategy_ = option.second;
+    } else if (kCoremlProviderOption_ProfileComputePlan == option.first) {
+      profile_compute_plan_ = option.second == "1";
+    } else if (kCoremlProviderOption_AllowLowPrecisionAccumulationOnGPU == option.first) {
+      allow_low_precision_accumulation_on_gpu_ = option.second == "1";
     }
   }
 }

onnxruntime/core/providers/coreml/coreml_options.h

@@ -14,6 +14,9 @@ class CoreMLOptions {
   bool create_mlprogram_{false};
   bool enable_on_subgraph_{false};
   uint32_t compute_units_{0};
+  std::string strategy_;
+  bool profile_compute_plan_{false};
+  bool allow_low_precision_accumulation_on_gpu_{false};
 
  public:
   explicit CoreMLOptions(uint32_t coreml_flags);
@@ -25,6 +28,9 @@ class CoreMLOptions {
   bool CreateMLProgram() const { return create_mlprogram_; }
   bool EnableOnSubgraph() const { return enable_on_subgraph_; }
   uint32_t ComputeUnits(uint32_t specific_flag = 0xffffffff) const { return compute_units_ & specific_flag; }
+  bool AllowLowPrecisionAccumulationOnGPU() const { return allow_low_precision_accumulation_on_gpu_; }
+  bool UseStrategy(std::string_view strategy) const { return strategy_ == strategy; }
+  bool ProfileComputePlan() const { return profile_compute_plan_ && create_mlprogram_; }
 
  private:
   void ValidateAndParseProviderOption(const ProviderOptions& options);

onnxruntime/core/providers/coreml/model/model.h

@@ -18,6 +18,7 @@
 #endif
 
 namespace onnxruntime {
+class CoreMLOptions;
 namespace coreml {
 
 class Execution;
@@ -53,7 +54,7 @@ class Model {
         std::unordered_map<std::string, OnnxTensorInfo>&& input_output_info,
         std::unordered_set<std::string>&& scalar_outputs,
         std::unordered_set<std::string>&& int64_outputs,
-        const logging::Logger& logger, uint32_t coreml_compute_unit);
+        const logging::Logger& logger, const CoreMLOptions& coreml_options);
 
   ~Model();
   ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Model);

onnxruntime/core/providers/coreml/model/model.mm

@@ -25,6 +25,7 @@
 #include "core/providers/coreml/model/host_utils.h"
 #include "core/providers/coreml/model/objc_str_utils.h"
 #include "core/providers/coreml/shape_utils.h"
+#include "core/providers/coreml/coreml_options.h"
 
 // force the linker to create a dependency on the CoreML framework so that in MAUI usage we don't need
 // to manually do this
@@ -300,14 +301,61 @@ Status GetMLMultiArrayCopyInfo(const MLMultiArray* _Nonnull array,
   return Status::OK();
 }
 
+// since __clang_major__ >= 15, MLComputePlan is introduced in <CoreML/CoreML.h>
+// We are actually ensure the MacOS/IOS version and Xcode version is greater than `macOS 14.4, iOS 17.4`.
+// The macro API_AVAILABLE should also be fine.
+// Otherwise, the compiler will complain `MLComputePlan` is not defined.
+// we define __clang_analyzer__ here is for bypass static analysis
+void ProfileComputePlan(NSURL* compileUrl, MLModelConfiguration* config) {
+#if defined(__APPLE__) && defined(__clang__) && __clang_major__ >= 15 && !defined(__clang_analyzer__)
+  if (@available(macOS 14.4, iOS 17.4, *)) {
+    [MLComputePlan loadContentsOfURL:compileUrl
+                       configuration:config
+                   completionHandler:^(MLComputePlan* _Nullable computePlan, NSError* _Nullable error) {
+                     if (!computePlan) {
+                       NSLog(@"Error loading compute plan: %@", error);
+                       // Handle error.
+                       return;
+                     }
+                     MLModelStructureProgram* program = computePlan.modelStructure.program;
+                     if (!program) {
+                       NSLog(@"Error loading program from compute plan., this is not a mlprogram model");
+                       return;
+                     }
+
+                     MLModelStructureProgramFunction* mainFunction = program.functions[@"main"];
+                     if (!mainFunction) {
+                       NSLog(@"Error loading main function from program");
+                       return;
+                     }
+
+                     NSArray<MLModelStructureProgramOperation*>* operations = mainFunction.block.operations;
+                     NSLog(@"Number of operations, 'const' node is included. : %lu", operations.count);
+                     for (MLModelStructureProgramOperation* operation in operations) {
+                       // Get the compute device usage for the operation.
+                       MLComputePlanDeviceUsage* computeDeviceUsage = [computePlan computeDeviceUsageForMLProgramOperation:operation];
+                       id<MLComputeDeviceProtocol> preferredDevice = computeDeviceUsage.preferredComputeDevice;
+                       // Get the estimated cost of executing the operation.
+                       MLComputePlanCost* estimatedCost = [computePlan estimatedCostOfMLProgramOperation:operation];
+                       if (![operation.operatorName isEqualToString:@"const"]) {
+                         NSLog(@"Operation: %@, Device Usage: %@, Estimated Cost: %f", operation.operatorName, preferredDevice, estimatedCost.weight);
+                       }
+                     }
+                   }];
+  } else {
+    NSLog(@"iOS 17.4+/macOS 14.4+ or later is required to use the compute plan API");
+  }
+#endif
+}
+
 // Internal Execution class
 // This class is part of the model class and handles the calls into CoreML. Specifically, it performs
 // 1. Compile the model by given path for execution
 // 2. Predict using given OnnxTensorFeatureProvider input and copy the output data back ORT
 // 3. The compiled model will be removed in dealloc or removed using cleanup function
 class Execution {
  public:
-  Execution(const std::string& path, const logging::Logger& logger, uint32_t coreml_flags);
+  Execution(const std::string& path, const logging::Logger& logger, const CoreMLOptions& coreml_options);
   ~Execution();
 
   Status LoadModel();
@@ -320,13 +368,13 @@ Status Predict(const std::unordered_map<std::string, OnnxTensorData>& inputs,
   NSString* coreml_model_path_{nil};
   NSString* compiled_model_path_{nil};
   const logging::Logger& logger_;
-  uint32_t coreml_compute_unit_{0};
+  CoreMLOptions coreml_options_;
   MLModel* model_{nil};
 };
 
-Execution::Execution(const std::string& path, const logging::Logger& logger, uint32_t coreml_compute_unit)
+Execution::Execution(const std::string& path, const logging::Logger& logger, const CoreMLOptions& coreml_options)
     : logger_(logger),
-      coreml_compute_unit_(coreml_compute_unit) {
+      coreml_options_(coreml_options) {
   @autoreleasepool {
     coreml_model_path_ = util::Utf8StringToNSString(path.c_str());
   }
@@ -395,17 +443,41 @@ Status Predict(const std::unordered_map<std::string, OnnxTensorData>& inputs,
       compiled_model_path_ = [compileUrl path];
 
       MLModelConfiguration* config = [[MLModelConfiguration alloc] init];
-
-      if (coreml_compute_unit_ & COREML_FLAG_USE_CPU_ONLY) {
+      uint32_t coreml_compute_unit = coreml_options_.ComputeUnits();
+      if (coreml_compute_unit & COREML_FLAG_USE_CPU_ONLY) {
         config.computeUnits = MLComputeUnitsCPUOnly;
-      } else if (coreml_compute_unit_ & COREML_FLAG_USE_CPU_AND_GPU) {
+      } else if (coreml_compute_unit & COREML_FLAG_USE_CPU_AND_GPU) {
         config.computeUnits = MLComputeUnitsCPUAndGPU;
-      } else if (coreml_compute_unit_ & COREML_FLAG_ONLY_ENABLE_DEVICE_WITH_ANE) {
+      } else if (coreml_compute_unit & COREML_FLAG_ONLY_ENABLE_DEVICE_WITH_ANE) {
         config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;  // Apple Neural Engine
       } else {
         config.computeUnits = MLComputeUnitsAll;
       }
 
+      if (coreml_options_.AllowLowPrecisionAccumulationOnGPU()) {
+        config.allowLowPrecisionAccumulationOnGPU = YES;
+      }
+
+// Set the specialization strategy to FastPrediction  for macOS 10.15+
+// since __clang_major__ >= 15, optimizationHints is introduced in <CoreML/CoreML.h>
+// Same as above comments for why we are checking __clang_major__.
+// we define __clang_analyzer__ here is for bypass static analysis
+#if defined(__APPLE__) && defined(__clang__) && __clang_major__ >= 15 && !defined(__clang_analyzer__)
+      if (HAS_COREML8_OR_LATER) {
+        MLOptimizationHints* optimizationHints = [[MLOptimizationHints alloc] init];
+        if (coreml_options_.UseStrategy("FastPrediction")) {
+          optimizationHints.specializationStrategy = MLSpecializationStrategyFastPrediction;
+          config.optimizationHints = optimizationHints;
+        } else if (coreml_options_.UseStrategy("Default")) {
+          optimizationHints.specializationStrategy = MLSpecializationStrategyDefault;
+          config.optimizationHints = optimizationHints;
+        }
+      }
+#endif
+      if (coreml_options_.ProfileComputePlan()) {
+        ProfileComputePlan(compileUrl, config);
+      }
+
       model_ = [MLModel modelWithContentsOfURL:compileUrl configuration:config error:&error];
 
       if (error != nil || model_ == nil) {
@@ -524,8 +596,8 @@ Status Predict(const std::unordered_map<std::string, OnnxTensorData>& inputs,
              std::unordered_set<std::string>&& scalar_outputs,
              std::unordered_set<std::string>&& int64_outputs,
              const logging::Logger& logger,
-             uint32_t coreml_flags)
-    : execution_(std::make_unique<Execution>(path, logger, coreml_flags)),
+             const CoreMLOptions& coreml_options)
+    : execution_(std::make_unique<Execution>(path, logger, coreml_options)),
       model_input_names_(std::move(model_input_names)),
       model_output_names_(std::move(model_output_names)),
       input_output_info_(std::move(input_output_info)),