* add some docs about speed [pr] * better torch gemm * enable locals on llvm/clang * disable locals for beam speed on LLVM/CLANG * 0x20 alignment in llvm allows ymm use
