[CoreML ] ML Program more operators support [3/N] (microsoft#22710)

### Description
- Erf
- Round
- Max
- ReduceMax
- ReduceMean
- ReduceSum
- Unsqueeze
- Squeeze
- Softmax



### 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: Scott McKay <[email protected]>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

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

@@ -13,14 +13,6 @@ using namespace CoreML::Specification;
 namespace onnxruntime {
 namespace coreml {
 
-// Once all ops are supportted FP16, we can remove it. Before that, we keep a set of ops to
-// filter suppported ones.
-static std::set<std::string> Float16Ops = {
-    "Add", "ArgMax", "AveragePool", "BatchNormalization", "Cast", "Clip", "Concat", "Conv", "ConvTranspose",
-    "DepthToSpace", "Div", "Gelu", "Gemm", "GlobalAveragePool", "GlobalMaxPool", "GridSample", "GroupNormalization",
-    "InstanceNormalization", "LayerNormalization", "LeakyRelu", "MatMul", "MaxPool", "Mul", "PRelu", "Pow",
-    "Reciprocal", "Relu", "Reshape", "Resize", "Sigmoid", "Slice", "Split", "Sqrt", "Sub", "Tanh", "Transpose"};
-
 namespace {
 // TODO, move this to shared_library
 bool HasExternalInitializer(const InitializedTensorSet& initializers, const Node& node,

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

@@ -133,8 +133,9 @@ bool ReductionOpBuilder::IsOpSupportedImpl(const Node& node, const OpBuilderInpu
     return false;
   }
 
-#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
+#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
   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

@@ -13,10 +13,6 @@
 #include "core/optimizer/initializer.h"
 #include "core/providers/cpu/tensor/unsqueeze.h"
 
-#ifdef __APPLE__
-#include <TargetConditionals.h>
-#endif
-
 namespace onnxruntime {
 namespace coreml {
 
@@ -80,28 +76,28 @@ 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;
 
-#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::string_view coreml_op_type = node.OpType() == "Squeeze" ? "squeeze" : "reshape";
     std::unique_ptr<Operation> op = model_builder.CreateOperation(node, coreml_op_type);
     AddOperationInput(*op, "x", input_defs[0]->Name());
 
-    if (!axes.empty()) {
-      // coreml supports negative axes
-      AddOperationInput(*op, "axes", model_builder.AddConstant(op->type(), "axes", AsSpan(axes)));
+    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)));
     }
     AddOperationOutput(*op, *node.OutputDefs()[0]);
     model_builder.AddOperation(std::move(op));