tilelang.tileop.gemm.gemm_wgmma
===============================

.. py:module:: tilelang.tileop.gemm.gemm_wgmma


Classes
-------

.. autoapisummary::

   tilelang.tileop.gemm.gemm_wgmma.GemmWGMMA


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

.. py:class:: GemmWGMMA

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


   .. py:method:: infer_shared_layout(continuity)

      Infer the swizzle layout for shared memory based on continuity.

      WGMMA can directly use shared memory as input, so the swizzle layout must
      match the tensor core's access pattern. The swizzle granularity is determined
      by the continuous dimension size:
        - 128B swizzle (Full):    continuity % (vectorized_size * 8) == 0
        - 64B swizzle (Half):     continuity % (vectorized_size * 4) == 0
        - 32B swizzle (Quarter):  continuity % (vectorized_size * 2) == 0
        - Linear (no swizzle):    otherwise

      See: https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__TENSOR__MEMORY.html



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


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


   .. py:method:: is_gemm_ss()


   .. py:method:: is_gemm_sr()


   .. py:method:: is_gemm_rs()


   .. py:method:: is_gemm_rr()