Segmentation fault when creating Session with 32-bit DirectML

[Godnoken]

Describe the issue

Hello,

I am using the Rust ort crate. Opened up an issue there about getting a segfault when creating a session, but according to the maintainer this is an issue with either Onnxruntime or DirectML directly.

Everything works just fine when I compile to 64-bit. No problems with loading models or running inference. But with 32-bit, it immediately segfaults on Session creation.

Only error message I've been able to get;

error: process didn't exit successfully: target\i686-pc-windows-msvc\debug\rapidocr_rs.exe (exit code: 0xc0000005, STATUS_ACCESS_VIOLATION)

To reproduce

Onnxruntime 1.20.0 (Microsoft.ML.OnnxRuntime.DirectML.1.20.0.nupkg)
DirectML 1.15.2 (Microsoft.AI.DirectML.1.15.2)
Model used: en_PP-OCRv3_det_infer.onnx

I have tried other Onnxruntimes & DirectML but haven't found any that works.

Urgency

No response

Platform

Windows

OS Version

10

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.20.0

ONNX Runtime API

C

Architecture

X86

Execution Provider

DirectML

Execution Provider Library Version

1.15.2

[xadupre]

• @tianleiwu

[fdwr]

@Godnoken Although I know nothing about various crates packages that may wrap ORT, if it helps to diagnose, I have small C++ ORT&DML sample from a few years back here, and it runs this model fine with DML 1.10 & ORT 1.13.1, and also DML 1.15.2 & ORT 1.20.0 🤔:

[Godnoken]

@Godnoken Although I know nothing about various crates packages that may wrap ORT, if it helps to diagnose, I have small C++ ORT&DML sample from a few years back here, and it runs this model fine with DML 1.10 & ORT 1.13.1, and also DML 1.15.2 & ORT 1.20.0 🤔:

Hi, thank you for quickly investigating. That is strange.. so it would seem the fault is indeed with the wrapper then?

I tried running your example with both sets of versions & 32-bit/64-bit, but I get this error;

$ ./FullExample en_PP-OCRv3_det_infer.onnx
Creating Direct3D device.
Adding the DirectML execution provider.
DLL path ONNX Runtime: C:\dev\programming\OnnxRuntimeDirectMLEPSample\build\FullExample_Debug_Win32\onnxruntime.dll
DLL path DirectML: C:\dev\programming\OnnxRuntimeDirectMLEPSample\build\FullExample_Debug_Win32\DirectML.dll
Loading model 'en_PP-OCRv3_det_infer.onnx'.
Binding input tensor 'x', float32[3].
Binding output tensor 'sigmoid_0.tmp_0', float32[1].
Beginning execution.
2025-03-19 17:47:14.0572858 [E:onnxruntime:, sequential_executor.cc:516 onnxruntime::ExecuteKernel] Non-zero status code r
eturned while running Sigmoid node. Name:'Sigmoid.0' Status Message: D:\a\_work\1\s\onnxruntime\core\providers\dml\DmlExec
utionProvider\src\MLOperatorAuthorImpl.cpp(2583)\onnxruntime.dll!583F03F8: (caller: 589CF6F3) Exception(2) tid(27ec) 80070
23E {Application Error}
The exception %s (0x
Error running model inference: Non-zero status code returned while running Sigmoid node. Name:'Sigmoid.0' Status Message:
D:\a\_work\1\s\onnxruntime\core\providers\dml\DmlExecutionProvider\src\MLOperatorAuthorImpl.cpp(2583)\onnxruntime.dll!583F
03F8: (caller: 589CF6F3) Exception(2) tid(27ec) 8007023E {Application Error}
The exception %s (0x

If it helps, I run on a 3570k and 980 Ti. Going to try running it on my laptop as well.

[fdwr]

Binding input tensor 'x', float32[3].

Did you try with the en_PP-OCRv3_det_infer branch I pointed to? It sets the free dimension in this commit.

[Godnoken]

Did you try with the en_PP-OCRv3_det_infer branch I pointed to? It sets the free dimension in this commit.

Yup!

[fdwr]

Then I am perplexed o_o, because I'm seeing Binding input tensor 'x', float32[3] in your output, where I expected to see float32[786432]:

Given 3*64*64*64 = 786'432

constexpr std::pair<char const*, int> NAMED_MODEL_DIMENSIONS[] =
{
    {"batch_size", 1},
    {"p2o.DynamicDimension.0", 64},
    {"p2o.DynamicDimension.1", 64},
    {"p2o.DynamicDimension.2", 64},
    // Add more here if the model has any.
};

Well at least you're getting a more standard error due to incompatible shapes (probably because the input tensor is too small for the filter sizes or some other operator along the network), rather than STATUS_ACCESS_VIOLATION.

[Godnoken]

@fdwr Well, this was slightly embarrassing. The reason I couldn't run your project was because the model I had, was in fact, not the model I linked.. 1mb difference. The one in the link works fine.

However, still getting segmentation faults using the wrapper, so I suppose there's nothing more to be done here and I need to go back there once more.

Thank you for spending time on this, my apologies.