- hipObject is now built as plain C/C++ instead of HIP. It contains no GPU
kernels, so the HIP language (and its per-architecture device compilation
pass) is no longer enabled. The HIP runtime is still used via the
hip::hostCMake target.CMAKE_HIP_COMPILERno longer needs to be set when configuring. - hipObject is now built as C++20 (without compiler extensions) instead of
C++17, so a C++20-capable compiler is required. The
HIPOBJ_CXX_STANDARDCMake cache variable now defaults to20. - hipObject is now built as a shared library (
libhipobj.so) by default. TheBUILD_SHARED_LIBSCMake option, which was previously ignored, now selects between a shared and a static library and defaults toON. Set it toOFFto build the static library (libhipobj.a) as before. The shared library exports only the public C API declared inhipobj.h. - The
hipobj::hipobjCMake target no longer passeship::hostandhsa-runtime64::hsa-runtime64on to consumers, sincehipobj.hdoesn't include HIP or HSA headers. Consumers that use HIP themselves must linkhip::hostdirectly.
- The
OFFLOAD_ARCHCMake cache variable and itsrocminfo-based GPU architecture detection. It was never used to set the target architecture and is unnecessary now that nothing is compiled for the GPU. - The
hipObjInitV2()declaration inhipobj.h. The function was never implemented, so calls to it failed at link time. They now fail to compile.