Trying to use Cpu provider on Windows ARM64 : execution engine exception

[oising]

I tried to run the ClassificationDemo as it is using the Cpu provider. I'm on a snapdragon surface laptop 7 -- and while there is a QNN provider with Microsoft.ML.OnnxRuntime, this project doesn't support it (yet?).

I've validated that there are runtimes/win-arm64/native/ binaries for onxxruntime and skiasharp native in the output folder and yet, this crashes the CLR oddly:

Fatal error.
0xC0000005
   at Microsoft.ML.OnnxRuntime.CompileApi.NativeMethods..ctor(DOrtGetCompileApi)
   at Microsoft.ML.OnnxRuntime.NativeMethods..cctor()
   at System.Runtime.CompilerServices.InitHelpers.InitClassSlow(System.Runtime.CompilerServices.MethodTable*)
   at System.Runtime.CompilerServices.StaticsHelpers.GetGCStaticBaseSlow(System.Runtime.CompilerServices.MethodTable*)
   at Microsoft.ML.OnnxRuntime.OrtEnv.CreateDefaultEnv(Microsoft.ML.OnnxRuntime.OrtLoggingLevel, Byte[])
   at Microsoft.ML.OnnxRuntime.OrtEnv.CreateInstance()
   at System.Lazy`1[[System.__Canon, System.Private.CoreLib, Version=10.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].ViaFactory(System.Threading.LazyThreadSafetyMode)
   at System.Lazy`1[[System.__Canon, System.Private.CoreLib, Version=10.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].ExecutionAndPublication(System.LazyHelper, Boolean)
   at System.Lazy`1[[System.__Canon, System.Private.CoreLib, Version=10.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].CreateValue()
   at Microsoft.ML.OnnxRuntime.OrtEnv.CreateInstanceWithOptions(Microsoft.ML.OnnxRuntime.EnvironmentCreationOptions ByRef)
   at YoloDotNet.ExecutionProvider.Cpu.CpuExecutionProvider.ConfigureOrtEnv()
   at YoloDotNet.ExecutionProvider.Cpu.CpuExecutionProvider.InitializeYolo(System.Object)
   at YoloDotNet.ExecutionProvider.Cpu.CpuExecutionProvider..ctor(System.String)
   at ClassificationDemo.Program.Main(System.String[])

[oising]

I found something possibly related on the ONNX runtime repo and added a comment:

microsoft/onnxruntime#27304

As an aside -- any plans to port/wrap the QNN provider? I would take on the job myself but I'm so new to this stuff. If I was to try implementing it, which of your existing providers would be the best baseline to work from?