[Build] Is it possible to build onnxruntime with DirectML EP statically?

[xeeetu]

Describe the issue

Urgency

As soon as possible, the project is under development.

Target platform

Windows

Build script

./build.bat --config Release --parallel --enable_msvc_static_runtime --use_dml (I used this command to build, I attach a file with test results. There were no errors when building a single onnxruntime.)

Error / output

tests.txt

Visual Studio Version

2022

GCC / Compiler Version

No response

[wschin]

Your build log looks ok -- it finished the build but failed one post-build test.

1: D:\onnxruntime\onnxruntime\test\providers\checkers.cc(265): error: The difference between cur_expected[i] and cur_actual[i] is inf, where
1: cur_expected[i] evaluates to 3.4028234663852886e+38,
1: cur_actual[i] evaluates to inf.
1: The abs_error parameter threshold evaluates to 0.004999999888241291 which is smaller than the minimum distance between doubles for numbers of this magnitude which is 3.7778931862957162e+22, thus making this EXPECT_NEAR check equivalent to EXPECT_EQUAL. Consider using EXPECT_DOUBLE_EQ instead.
1: i:10
1: Google Test trace:
1: D:\onnxruntime\onnxruntime\test\providers\checkers.cc(484): provider type: DmlExecutionProvider
1: D:\onnxruntime\onnxruntime\test\providers\base_tester.cc(774): registered execution providers: DmlExecutionProvider
1: [  FAILED  ] ActivationOpTest.Gelu (656 ms)

Since this kind of error is usually reasonable, I'd ignore it for now. Do you see other problems during the build?

[wschin]

I am not sure. Maybe you can find answer in the section of building DirectML execution provider at

onnxruntime/cmake/onnxruntime_providers.cmake

Line 1318 in 992f3e4

target_link_libraries(${target} PRIVATE DirectML)

. It allows custom build of DirectML, so it's possible you can statically link to DirectML if you build DirectML from source or the installed DirectML has a static library.

onnxruntime/build\Windows\packages\Microsoft.AI.DirectML.1.12.1 looks like a third-party library --- If you want to statically link against DirectML, you need to make sure it includes a DirectML.lib for static linking. I am not sure if DirectML has that DirectML.lib.

[fdwr]

xeetu: So there's the:

• "DML EP" (DirectML execution provider) which is built as part of ONNX Runtime and translates ONNX operators to DirectML.dll calls.
• DirectML.dll which is a separate library holds the necessary logic to run D3D shaders on the GPU, available via Nuget in redistributable form and as an older system version in Windows/System32 (but ORT uses the newer redist versions). https://www.nuget.org/packages/Microsoft.AI.DirectML
• DirectML.lib here just is the import library just to call DMLCreateDevice and doesn't contain DML code.

Like Wei-Sheng said, I'm not seeing any actual build errors, just a test failure that's not unreasonably far away. If you run with --skip_tests, does the build complete?

[wschin]

I guess you just need to use DirectML.dll as a DLL for building onnxruntime. If you want to execute onnxruntime on another machine, just install DirectML nuget on that machine.

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 7 days if no further activity occurs. If further support is needed, please provide an update and/or more details.