[QST] SmemCopyAtom and MMA_Atom for fp32?

What is your question?
hello, I am developing a full precision attention backward kernel using cutlass, and get stuck in the use of ldmatrix and mma instructions for fp32.

My Gemm calculation is based on fp32 matrix, i.e. the datatype of D/A/B/C are all fp32. But the structs providied in mma_sm80.hpp take half-precision/mixed precision inputs so I am pretty confused about how to do things right in full precision. Here is my current setting for MMA, smem and gmem. Is there a way to use SM75_U32x4_LDSM_N and one of the mma instructions in my case?

  // MMA
  using TiledMma = TiledMMA<MMA_Atom<UniversalFMA<float, float, float>>, Layout<Shape<Int<16>, Int<8>, _1>>>;

  // Smem
  using SmemLayoutAtom = decltype(
    composition(Swizzle<3,3,3>{},
                Layout<Shape < _16,_32>,
                       Stride<_32, _1>>{}));
  using SmemCopyAtom = Copy_Atom<DefaultCopy, float>;

  // Gmem
  using GmemTiledCopy = decltype(
    make_tiled_copy(Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<cute::uint128_t>, float>{},
                    Layout<Shape <_16,_8>,
                           Stride< _8,_1>>{},
                    Layout<Shape < _1,_4>>{}));

Start with the fp32-related definitions in mma_sm80.hpp and the shown TiledMma, SmemCopyAtom, and GmemTiledCopy configuration. Check whether SM75_U32x4_LDSM_N and an available MMA instruction support the requested fp32 operands, and document a confirmed configuration or the missing support.