I've attached a diff of the passing and failing traces with AMD_LOG_LEVEL=3 set in the environment. Timestamps, time durations, and pointers were replaced with 'x' before diffing.
The most noteworthy change is: -:3:hip_fatbin.cpp :386 : xxxxxxxxxxxx us: Inserting bundle entry of amdgcn-amd-amdhsa--gfx1201 : size: 18784, data: 0xXXXXXXXX +:3:hip_fatbin.cpp :386 : xxxxxxxxxxxx us: Inserting bundle entry of amdgcn-amd-amdhsa--gfx1201 : size: 68936, data: 0xXXXXXXXX To me, it looks like hip_fatbin.cpp (part of src:rocm-hipamd) is confused about the size of the code object bundle. I'm suspicious that perhaps the compressed size is being returned from libllvm21 somewhere that the uncompressed size is expected. The underlying bug is presumably in libllvm21, as both the pass and fail runs use the same version of libamd_comgr.so.3 with minor patches. The biggest difference between pass and fail is that the passing version uses the AMD fork and the failing version uses upstream LLVM 21. ** Patch added: "rocsparse-test-amdloglevel3.diff" https://bugs.launchpad.net/ubuntu/+source/rocm-llvm/+bug/2142813/+attachment/5948978/+files/rocsparse-test-amdloglevel3.diff -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2142813 Title: missing symbol error loading compressed bundles To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/rocm-llvm/+bug/2142813/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
