Test Bundled Binaries Generated: 2025-12-29 22:59:19.373659
ROCm Path: /home/stella/claude-rocm-workspace/rocm/gfx1100
Code Object Version: 5
Platform: linux
Compiler: HIP version: 7.1.25432-ba73a9bf4c

Bundled Executables (with GPU device code):
- test_kernel_single.exe: Single architecture (gfx1100)
- test_kernel_multi.exe: Multiple architectures (gfx1100, gfx1101)
- test_kernel_compressed.exe: Multiple architectures with compression (if supported)
- test_kernel_wide.exe: Wide architecture coverage (gfx900,906,908,90a,1100)

Bundled Shared Libraries (with GPU device code):
- libtest_kernel_single.so: Single architecture (gfx1100)
- libtest_kernel_multi.so: Multiple architectures (gfx1100, gfx1101)
- libtest_multi_wrapper.so: RDC build with 2 __CudaFatBinaryWrapper structs (gfx1100)

Host-Only Binaries (NO GPU device code, for negative testing):
- host_only.exe: Host-only executable
- libhost_only.so: Host-only shared library

Bundled binaries contain HIP kernels (vectorAdd, scalarMultiply) with .hip_fatbin sections.
Host-only binaries contain only CPU code with no .hip_fatbin sections.

Sources:
- test_generation/simple_kernel.hip (GPU kernels)
- test_generation/multi_wrapper_kernel1.hip (RDC kernel 1)
- test_generation/multi_wrapper_kernel2.hip (RDC kernel 2)
- test_generation/host_only.cpp (host-only code)
Build Script: test_generation/build_test_bundles.py

These assets are used to test unbundling functionality across:
- Executables and shared libraries
- Uncompressed vs compressed code objects
- Single vs multiple architectures
- Different code object versions
- ELF (Linux) vs PE/COFF (Windows) formats
- Positive tests (bundled binaries) vs negative tests (host-only binaries)
