Compiler version

[jeyblu]

Which version of compilers are used to build the ONNX Runtime binaries released on the repository's release page?

• Which MSVC version is used to build the Windows binaries?
• Which gcc version is used to build the Linux binaries?

What about the python pip package for ONNX Runtime?

• Which MSVC version is used to build the Windows pip package?
• Which gcc version is used to build the Linux pip package?

Thanks

[snnn]

The latest release used Visual Studio 2019. The next release will use Visual Studio 2022 instead.
GCC version is 11.

[jeyblu]

Which gcc 11 version? Is it gcc 11.4.0? Thanks

[snnn]

I think the minor version doesn't matter. It doesn't change the ABI. I didn't keep it in record.

[jeyblu]

Do you know which assembler version was used? Command to check as --version Thanks

[snnn]

I didn't look deep into this. We just use what CentOS provides. Our Linux builds run in docker. Here is the docker file: https://github.com/microsoft/onnxruntime/blob/main/tools/ci_build/github/linux/docker/inference/x64/default/cpu/Dockerfile You can build the docker image and run the command by yourself. But, the output doesn't guarantee to be always the same.

[jeyblu]

The container built with the current Dockerfile has

gcc (GCC) 12.2.1 20221121 (Red Hat 12.2.1-7)

In the ONNX Runtime release announcements, in v1.15.0, it says

At compiler level we will drop the support for
GCC version <= 9
Visual Studio 2019

The documentation on "Build ONNX Runtime for inferencing" says

GCC 8.x and below are not supported.

Is it possible to change the gcc requirement to at least gcc 12, which is what your current build container has? Newer versions of compiler are required to build better and more optimal code, including compiling new instructions for deep learning. For example, compiler support for AMX instructions which ONNX Runtime uses was added in gcc 11.

[jeyblu]

Got it. As you suggested, if we scope the AMX code within #ifdef, then we can use the gcc version required for AMX for the CPU execution provider, and the gcc version required for the CUDA execution provider. We can try this. Thanks

[jeyblu]

CUDA documentation states that gcc 6.x - 13.2 is supported. Does this address the reason for not using latest gcc?

[snnn]

You were referring CUDA 12.4's document. Not 11.x.

[jeyblu]

Do you have plans to update to CUDA 12.4?

[snnn]

We need to support both CUDA 11.x and CUDA 12.x , until Nvidia stop supporting CUDA 11.x.