• Lei Wang's avatar
    [Refactor] Introduce GemmInst for different targets handling (#688) · d2afb513
    Lei Wang authored
    * [Enhancement] Refactor GEMM operations for improved warp partitioning and target instruction handling
    
    - Introduced a new `GetGemmInst` method to determine the appropriate GEMM instruction based on block size and target architecture.
    - Updated `ComputeWarpPartition` to accept the GEMM instruction type, enhancing flexibility in warp partitioning logic.
    - Added `TargetGetWarpSize` utility to streamline warp size retrieval based on target architecture.
    - Refactored layout inference and lowering methods to utilize the new GEMM instruction handling, improving clarity and maintainability of the codebase.
    
    * bug fix
    
    * test fix
    
    * lint fix
    d2afb513
gemm.cc 17.6 KB