for curiosity, because I am a grad student working on AI accelerators, I just want to know why these AI manufacturers are going with XOC which is new to me instead of wide used toolchain like GCC or LLVM especially with the last one using MLIR .
Good question! In a simply word, GCC&LLVM compilers are good at traditional RISC-like hardware, namely, IR expected to be load/store/compute/branch, etc. More important is, these compilers do not allow user to add custom IR at will. AI chip operator evoloing rapidly, adding IR means the compiler needs to be overhauled. XOC is multi-level IR compiler, MLIR is not. MLIR open source later than XOC. Moreover, MLIR lack s of powerful optimizations.