Skip to content

[UR][HIP] Select the correct binary in a multi-arch bundle #19092

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 8 commits into from
Jun 24, 2025

Conversation

frasercrmck
Copy link
Contributor

This commit changes the HIP adapter to select the correct binary for the device when a bundle contains binaries built for multiple AMDGPU architectures.

Similarly to other adapaters, the HIP adapter would previously select the first 'amdgcn' binary it came across. This works fine for the common case where the program was compiled for one architecture but may fail otherwise.

To aid in this, the SYCL runtime passes some extra information into urDeviceSelectBinary via the pre-existing 'pNext' field of ur_device_binary_t. It does this only for the HIP backend. The HIP adapater then parses this binary information as a clang offload bundle, which conveniently contains specific triple & architecture information for each binary. For this we re-use the code that the offload adapter was using, making it common and fixing a bug in the version matching logic.

This commit changes the HIP adapter to select the correct binary for the
device when a bundle contains binaries built for multiple AMDGPU
architectures.

Similarly to other adapaters, the HIP adapter would previously select
the first 'amdgcn' binary it came across. This works fine for the common
case where the program was compiled for one architecture but may fail
otherwise.

To aid in this, the SYCL runtime passes some extra information into
urDeviceSelectBinary via the pre-existing 'pNext' field of
ur_device_binary_t. It does this only for the HIP backend. The HIP
adapater then parses this binary information as a clang offload bundle,
which conveniently contains specific triple & architecture information
for each binary. For this we re-use the code that the offload adapter
was using, making it common and fixing a bug in the version matching
logic.
Copy link
Contributor

@steffenlarsen steffenlarsen left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM!

@frasercrmck
Copy link
Contributor Author

@intel/llvm-gatekeepers this is ready to merge - thank you.

@sarnex sarnex merged commit d829db7 into intel:sycl Jun 24, 2025
58 of 62 checks passed
@frasercrmck frasercrmck deleted the hip-multi-arch branch June 24, 2025 15:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

5 participants