LLVM's OpenMP Offloads Liboffload Into Oblivion
The liboffload library has been dropped from LLVM's OpenMP repository.
Liboffload is/was the Intel runtime library for offloading and geared for supporting the Xeon Phi co-processors. But liboffload within LLVM hasn't been receiving updates, it wasn't properly integrated within the LLVM build system, and unfortunately Xeon Phi co-processors appear to be discontinued. The liboffload library has also confused some with LLVM's libomptarget library for OpenMP support that is in much better shape.
While Xeon Phi was doing well for a while and even helped accelerate some of the world's fastest super-computers, the PCI Express add-on cards and standalone CPUs have waned in popularity in recent years. Intel cancelled the planned 10nm "Knights Hill" Xeon Phi and last year released the "Knights Mill" 14nm product that was essentially the same as the older "Knights Landing" architecture but with some deep learning optimizations. Since then it's been quiet.
Intel developers did not object to the call to drop liboffload from the LLVM OpenMP tree and as of last week it's now been removed. Of course, it could still be revived via Git should Xeon Phi restart again in the future and Intel's own offloading library outside of LLVM will presumably still be maintained.
Liboffload is/was the Intel runtime library for offloading and geared for supporting the Xeon Phi co-processors. But liboffload within LLVM hasn't been receiving updates, it wasn't properly integrated within the LLVM build system, and unfortunately Xeon Phi co-processors appear to be discontinued. The liboffload library has also confused some with LLVM's libomptarget library for OpenMP support that is in much better shape.
While Xeon Phi was doing well for a while and even helped accelerate some of the world's fastest super-computers, the PCI Express add-on cards and standalone CPUs have waned in popularity in recent years. Intel cancelled the planned 10nm "Knights Hill" Xeon Phi and last year released the "Knights Mill" 14nm product that was essentially the same as the older "Knights Landing" architecture but with some deep learning optimizations. Since then it's been quiet.
Intel developers did not object to the call to drop liboffload from the LLVM OpenMP tree and as of last week it's now been removed. Of course, it could still be revived via Git should Xeon Phi restart again in the future and Intel's own offloading library outside of LLVM will presumably still be maintained.
3 Comments