.. meta:: :description: StinkyTofu is an LLVM-inspired pass-based IR optimizer for AMD GPU assembly kernels. :keywords: StinkyTofu, ROCm, documentation, hipBLASLt, TensileLite ************************* StinkyTofu documentation ************************* StinkyTofu is an LLVM-inspired pass-based IR optimizer for AMD GPU assembly kernels. It is used by hipBLASLt/TensileLite to schedule and optimize generated GPU code for gfx1250. StinkyTofu uses a two-level IR: a high-level, architecture-agnostic Logical IR and a low-level, architecture-specific Asm IR, connected by a pass pipeline covering DAG scheduling, wait-count insertion, dead code elimination, redundant mov elimination, and peephole optimization. The public repository for StinkyTofu is located at `https://github.com/ROCm/rocm-libraries/tree/develop/shared/stinkytofu `_. .. grid:: 2 :gutter: 3 .. grid-item-card:: User guides * :doc:`Global parameters ` * :doc:`IR converter ` * :doc:`Assembly emitter ` * :doc:`Virtual registers ` * :doc:`Error codes ` .. grid-item-card:: Developer guides * :doc:`Architecture overview ` * :doc:`Adding instructions ` * :doc:`Adding a GPU architecture ` * :doc:`Pattern grammar reference ` .. grid-item-card:: Reference * :doc:`API reference ` * :doc:`Known issues ` .. grid-item-card:: About * :doc:`License ` You can find licensing information on the `Licensing `_ page.