tilelang.tileop.gemm.gemm_cutedsl
=================================

.. py:module:: tilelang.tileop.gemm.gemm_cutedsl

.. autoapi-nested-parse::

   GEMM implementation for CuTeDSL backend - directly calls tl::gemm intrinsic.



Classes
-------

.. autoapisummary::

   tilelang.tileop.gemm.gemm_cutedsl.GemmCuTeDSL


Module Contents
---------------

.. py:class:: GemmCuTeDSL

   Bases: :py:obj:`tilelang.tileop.gemm.gemm_base.GemmBase`


   GEMM implementation for CuTeDSL that directly calls tl::gemm intrinsic.

   This implementation bypasses the complex lowering logic of MMA/WGMMA
   and directly emits a call to tl::gemm, similar to gemm_v1 behavior.
   This is necessary for CuTeDSL backend which requires simpler IR.


   .. py:method:: infer_layout(target, thread_nums)

      For CuTeDSL, we still need proper layout inference for A, B, C buffers.

      CuTeDSL uses the same underlying hardware instructions (WGMMA/MMA),
      so it needs the same layout information. We delegate to the appropriate
      implementation based on the instruction type.



   .. py:method:: lower(layout_map, target, thread_nums, thread_var)

      Lower to a direct gemm_v1 call without complex MMA/WGMMA lowering.