tilelang.layout.gemm_sp
=======================

.. py:module:: tilelang.layout.gemm_sp

.. autoapi-nested-parse::

   Wrapping Layouts.



Functions
---------

.. autoapisummary::

   tilelang.layout.gemm_sp.decompose_col_major
   tilelang.layout.gemm_sp.make_cutlass_metadata_layout_sm90
   tilelang.layout.gemm_sp.make_cutlass_metadata_layout_sm8x
   tilelang.layout.gemm_sp.make_cutlass_metadata_layout


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

.. py:function:: decompose_col_major(index_1d, basis)

.. py:function:: make_cutlass_metadata_layout_sm90(buffer, mma_dtype, block_k)

   Make a layout of metadata that is compatible with cutlass sm90 compression kernel. Note that layout atom is the same for smem and gmem.

   :param buffer: metadata buffer shape, for sm90 it should be a 8-bit type
   :param mma_dtype: dtype of mma operand A, different dtypes result in different layout atom
   :param block_k: tiling size along K dim, different block_ks results in different layout atom.


.. py:function:: make_cutlass_metadata_layout_sm8x(buffer, mma_dtype)

   Make a layout of metadata that is compatible with cutlass sm8x compression kernel. Note that layout atom is the same for smem and gmem.
       ref: https://github.com/pytorch/pytorch/blob/d0c24b392cbb7b213d22e42c52c6c2d1ac2da1bd/torch/sparse/_semi_structured_conversions.py#L5
   :param buffer: metadata buffer shape, for sm80 it should be a 16bit type


.. py:function:: make_cutlass_metadata_layout(buffer, mma_dtype = 'float16', arch = None, **extra_args)