/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/core/tensor/tile_scatter_gather.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/core/tensor/tile_scatter_gather.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/core/tensor/tile_scatter_gather.hpp Source File
Go to the documentation of this file.
 // Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 // SPDX-License-Identifier: MIT
  
 #pragma once
  
 #include "ck_tile/core/arch/utility.hpp"
 #include "ck_tile/core/algorithm/space_filling_curve.hpp"
 #include "ck_tile/core/config.hpp"
 #include "ck_tile/core/container/array.hpp"
 #include "ck_tile/core/container/sequence.hpp"
 #include "ck_tile/core/container/tuple.hpp"
 #include "ck_tile/core/container/container_helper.hpp"
 #include "ck_tile/core/tensor/static_distributed_tensor.hpp"
 #include "ck_tile/core/tensor/tensor_adaptor.hpp"
 #include "ck_tile/core/tensor/tile_distribution.hpp"
 #include "ck_tile/core/utility/functional.hpp"
 #include "ck_tile/core/utility/type_traits.hpp"
  
 namespace ck_tile {
  
 template <typename BottomTensorView_,
           typename WindowLengths_,
           typename StaticTileDistribution_,
           typename StaticPageIndexArray_,
           typename StaticValidArray_,
           index_t HsGatherDim = 0,
           index_t NumCoord    = 1,
           index_t YsGatherDim = 0>
 struct tile_scatter_gather
 {
     using BottomTensorView = remove_reference_t<BottomTensorView_>;
     using WindowLengths    = remove_cvref_t<WindowLengths_>;
     using TileDstr         = remove_cvref_t<StaticTileDistribution_>;
     using PageIdxArray     = remove_cvref_t<StaticPageIndexArray_>;
     using ValidArray       = remove_cvref_t<StaticValidArray_>;
     using WindowAdaptor    = typename TileDstr::PsYs2XsAdaptor;
     using BottomTensorDesc = typename BottomTensorView::TensorDesc;
  
     using DataType = remove_cvref_t<typename BottomTensorView::DataType>;
  
     static constexpr index_t NDimWindowAdaptorTop = WindowAdaptor::get_num_of_top_dimension();
     static constexpr index_t NDimBottomTensor     = BottomTensorDesc::get_num_of_dimension();
  
     static constexpr index_t NDimP = TileDstr::get_num_of_dimension_p();
     static constexpr index_t NDimY = TileDstr::get_num_of_dimension_y();
  
     static constexpr auto I0 = number<0>{};
     static constexpr auto I1 = number<1>{};
     static_assert(NumCoord == 1);
  
     // TODO: check WindowLengths and StaticTileDistribution are consistent
  
     static_assert(ck_tile::is_known_at_compile_time<WindowLengths>::value,
                   "wrong! lengths should be static");
     static_assert(TileDstr::is_static(), "wrong!");
  
     static_assert(NDimBottomTensor == WindowAdaptor::get_num_of_bottom_dimension(),
                   "wrong! inconsistent # of diemsnions");
  
     using AdaptorTopIndex   = array<index_t, NDimWindowAdaptorTop>;
     using BottomTensorIndex = array<index_t, NDimBottomTensor>;
  
     using WindowAdaptorCoord =
         decltype(make_tensor_adaptor_coordinate(WindowAdaptor{}, AdaptorTopIndex{}));
  
     using BottomTensorCoord =
         decltype(make_tensor_coordinate(BottomTensorDesc{}, BottomTensorIndex{}));
  
     struct load_store_traits
     {
         private:
         static constexpr auto get_vector_dim_y_scalar_per_vector()
         {
             const auto [ys_vector_lengths, ys_vector_strides] =
                 tile_scatter_gather::get_window_adaptor_ys_safe_vector_length_strides();
  
             index_t VectorDimY_      = 0;
             index_t ScalarPerVector_ = 1;
  
             for(index_t i = 0; i < NDimY; ++i)
             {
                 if(ys_vector_strides[i] == 1 && ys_vector_lengths[i] > ScalarPerVector_)
                 {
                     ScalarPerVector_ = ys_vector_lengths[i];
                     VectorDimY_      = i;
                 }
             }
  
             return make_tuple(VectorDimY_, ScalarPerVector_);
         }
  
         public:
         static constexpr index_t PackedSize =
             ck_tile::numeric_traits<remove_cvref_t<DataType>>::PackedSize;
         static constexpr index_t VectorDimY = get_vector_dim_y_scalar_per_vector().template at<0>();
         static constexpr index_t ScalarPerVector =
             get_vector_dim_y_scalar_per_vector().template at<1>();
  
         // using vector_type_t = vector_type_maker_t<DataType, ScalarPerVector>;
         // using vector_t      = typename vector_type_t::type;
         using vector_t = thread_buffer<DataType, ScalarPerVector / PackedSize>;
  
         private:
         static constexpr auto scalars_per_access_ = [] {
             constexpr auto scalars_per_access_arr = generate_array(
                 [&](auto i) { return (i == VectorDimY) ? ScalarPerVector : 1; }, number<NDimY>{});
  
             constexpr auto NDimY_ = NDimY;
  
             return TO_SEQUENCE(scalars_per_access_arr, NDimY_);
         }();
  
         static constexpr auto get_space_filling_curve()
         {
             constexpr auto tile_dstr = TileDstr{};
  
             constexpr auto thread_tensor_lengths_ys =
                 to_sequence(tile_dstr.get_ys_to_d_descriptor().get_lengths());
  
             // FIXME: need logic to judge dim access order
             using DimAccessOrder = typename arithmetic_sequence_gen<0, NDimY, 1>::type;
  
             return space_filling_curve<decltype(thread_tensor_lengths_ys),
                                        DimAccessOrder,
                                        decltype(scalars_per_access_)>{};
         }
  
         public:
         using SFC_Ys = decltype(get_space_filling_curve());
  
         static constexpr index_t NumAccess = SFC_Ys::get_num_of_access();
  
         static_assert(0 < NumAccess, "Wrong! NumAccess should be larger than 0");
         static_assert(NumAccess % NumCoord == 0, "wrong! # of access is not divisible by NumCoord");
     };
  
     static constexpr index_t NumAccessPerCoord = load_store_traits::NumAccess / NumCoord;
  
     CK_TILE_DEVICE constexpr tile_scatter_gather() = default;
  
     CK_TILE_DEVICE constexpr tile_scatter_gather(const BottomTensorView& bottom_tensor_view,
                                                  const WindowLengths& window_lengths,
                                                  const BottomTensorIndex& window_origin,
                                                  const TileDstr& tile_distribution,
                                                  const PageIdxArray& page_idx,
                                                  const ValidArray& valids)
         : bottom_tensor_view_{bottom_tensor_view},
           window_lengths_{window_lengths},
           window_origin_{window_origin},
           tile_dstr_{tile_distribution},
           page_idx_{page_idx},
           valids_{valids},
           pre_computed_coords_{}
     {
 #if 0 // debug
       // TODO: this use more register for FA, but less register for GEMM
       // need investigation
       // only support warp-tile and block-tile
         static_assert(NDimP == 1 or NDimP == 2, "wrong!");
  
         WindowAdaptorCoord window_adaptor_thread_coord_tmp;
  
         if constexpr(NDimP == 1)
         {
             window_adaptor_thread_coord_tmp = make_tensor_adaptor_coordinate(
                 tile_distribution.get_ps_ys_to_xs_adaptor(), AdaptorTopIndex{get_lane_id(), 0});
         }
         else if constexpr(NDimP == 2)
         {
             window_adaptor_thread_coord_tmp =
                 make_tensor_adaptor_coordinate(tile_distribution.get_ps_ys_to_xs_adaptor(),
                                                AdaptorTopIndex{get_warp_id(), get_lane_id(), 0});
         }
 #else
         // TODO: this use less register for FA, but more register for GEMM
         // need investigation
         const auto window_adaptor_thread_coord_tmp = make_tensor_adaptor_coordinate(
             tile_distribution.get_ps_ys_to_xs_adaptor(),
             container_concat(get_partition_index(tile_distribution), array<index_t, NDimY>{0}));
 #endif
  
         BottomTensorIndex bottom_tensor_thread_origin_idx_tmp =
             window_origin + window_adaptor_thread_coord_tmp.get_bottom_index();
         bottom_tensor_thread_origin_idx_tmp(HsGatherDim) = 0;
         const auto bottom_tensor_thread_coord_tmp        = make_tensor_coordinate(
             bottom_tensor_view_.get_tensor_descriptor(), bottom_tensor_thread_origin_idx_tmp);
  
         // pre-compute NumCoord (WindowAdaptorCoord, BottomTensorCoord) bundles to speed up
         // future load/store() calls (might allocate more registers)
         using Traits = load_store_traits;
         using SFC_Ys = typename Traits::SFC_Ys;
  
         static_for<0, NumCoord, 1>{}([&](auto iCoord) {
             auto window_adaptor_thread_coord = window_adaptor_thread_coord_tmp;
             auto bottom_tensor_thread_coord  = bottom_tensor_thread_coord_tmp;
  
             constexpr auto idx_diff_ys =
                 SFC_Ys::get_step_between(number<0>{}, number<iCoord * NumAccessPerCoord>{});
  
             constexpr auto idx_diff_ps_ys = container_concat(
                 generate_tuple([&](auto) { return number<0>{}; }, number<NDimP>{}), idx_diff_ys);
  
             move_window_adaptor_and_bottom_tensor_thread_coordinate(
                 window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
  
             pre_computed_coords_(iCoord) =
                 make_tuple(window_adaptor_thread_coord, bottom_tensor_thread_coord);
         });
     }
  
     CK_TILE_DEVICE static constexpr index_t get_num_of_dimension() { return NDimBottomTensor; }
  
     CK_TILE_DEVICE static constexpr bool has_static_tile_distribution()
     {
         return TileDstr::is_static();
     }
  
     CK_TILE_DEVICE constexpr auto get_window_lengths() const { return window_lengths_; }
  
     CK_TILE_DEVICE constexpr auto get_tile_distribution() const { return tile_dstr_; }
  
     CK_TILE_DEVICE constexpr auto get_bottom_tensor_view() const { return bottom_tensor_view_; }
  
     CK_TILE_DEVICE constexpr auto get_window_origin() const { return window_origin_; }
  
     CK_TILE_DEVICE constexpr void
     set_bottom_tensor_view_data_ptr(typename BottomTensorView::DataType* data)
     {
         bottom_tensor_view_.buf_.p_data_ = data;
     }
  
     // move thread's window adaptor coordinate and bottom tensor coordinate
     // [p0, p1, ..., y0, y1, ...] ==> [x0, x1, ...] ==> [x0', x1', ...] ==> [offset]
     template <typename ATopIndex>
     CK_TILE_DEVICE void move_window_adaptor_and_bottom_tensor_thread_coordinate(
         WindowAdaptorCoord& window_adaptor_thread_coord,
         BottomTensorCoord& bottom_tensor_thread_coord,
         const ATopIndex& idx_diff_adaptor_top) const
     {
         array<index_t, NDimBottomTensor> idx_diff_adaptor_bottom;
  
         move_tensor_adaptor_coordinate(tile_dstr_.get_ps_ys_to_xs_adaptor(),
                                        window_adaptor_thread_coord,
                                        idx_diff_adaptor_top,
                                        idx_diff_adaptor_bottom);
  
         move_tensor_coordinate(bottom_tensor_view_.get_tensor_descriptor(),
                                bottom_tensor_thread_coord,
                                idx_diff_adaptor_bottom);
     }
  
     // return vector dimension among [y0, y1, ...]
     CK_TILE_DEVICE static constexpr auto get_window_adaptor_ys_safe_vector_length_strides()
     {
         // bottom tensor top dimension vector lengths and strides
         const auto [bottom_tensor_top_dim_vector_lengths, bottom_tensor_top_dim_vector_strides] =
             BottomTensorDesc::get_top_dimension_safe_vector_length_strides();
  
         // window vector lengths/strides
         const auto window_adaptor_bottom_dim_vector_lengths = bottom_tensor_top_dim_vector_lengths;
         const auto window_adaptor_bottom_dim_vector_strides = bottom_tensor_top_dim_vector_strides;
  
         // window adaptor [p0, p1, ..., y0, y1, ...]
         array<index_t, WindowAdaptor::get_num_of_hidden_dimension()> window_adaptor_vector_lengths{
             -1};
         array<index_t, WindowAdaptor::get_num_of_hidden_dimension()> window_adaptor_vector_strides{
             -1};
  
         constexpr auto window_adaptor_bottom_dims =
             WindowAdaptor::get_bottom_dimension_hidden_ids();
  
         set_container_subset(window_adaptor_vector_lengths,
                              window_adaptor_bottom_dims,
                              window_adaptor_bottom_dim_vector_lengths);
         set_container_subset(window_adaptor_vector_strides,
                              window_adaptor_bottom_dims,
                              window_adaptor_bottom_dim_vector_strides);
  
         const auto [window_adaptor_ps_ys_vector_lengths, window_adaptor_ps_ys_vector_strides] =
             WindowAdaptor{}.get_top_dimension_safe_vector_length_strides(
                 window_adaptor_vector_lengths, window_adaptor_vector_strides);
  
         // [y0, y1, ...]
         constexpr auto y_dims = typename arithmetic_sequence_gen<TileDstr::get_num_of_dimension_p(),
                                                                  NDimWindowAdaptorTop,
                                                                  1>::type{};
  
         return make_tuple(get_container_subset(window_adaptor_ps_ys_vector_lengths, y_dims),
                           get_container_subset(window_adaptor_ps_ys_vector_strides, y_dims));
     }
  
     CK_TILE_DEVICE constexpr auto get_num_of_access() const { return load_store_traits::NumAccess; }
  
     template <index_t i_access_unsupport_ = -1, bool oob_conditional_check = true>
     CK_TILE_DEVICE auto load(number<i_access_unsupport_>          = {},
                              bool_constant<oob_conditional_check> = {}) const
     {
         constexpr auto tile_dstr = TileDstr{};
         auto dst_tensor          = make_static_distributed_tensor<DataType>(tile_dstr);
         load(dst_tensor, number<i_access_unsupport_>{}, bool_constant<oob_conditional_check>{});
         return dst_tensor;
     }
  
     template <typename DistributedTensor,
               index_t i_access_unsupport_ = -1,
               bool oob_conditional_check  = true>
     CK_TILE_DEVICE auto load(DistributedTensor& dst_tensor,
                              number<i_access_unsupport_>          = {},
                              bool_constant<oob_conditional_check> = {}) const
     {
         using Traits   = load_store_traits;
         using vector_t = typename Traits::vector_t;
         using SFC_Ys   = typename Traits::SFC_Ys;
  
         constexpr auto tile_dstr = TileDstr{};
  
         // loop over thread tensor space [y0, y1, ...]
         static_for<0, NumCoord, 1>{}([&](auto iCoord) {
             auto window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0];
             auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
  
             static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
                 constexpr auto iAccess = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
  
                 // data index [y0, y1, ...]
                 constexpr auto idx_ys_start = SFC_Ys::get_index(iAccess);
                 constexpr auto idx_gather   = idx_ys_start[number<YsGatherDim>{}];
                 const auto page_offset      = page_idx_[idx_gather];
  
                 // read from bottom tensor
                 const vector_t vec_value = [&]() {
                     if constexpr(std::is_same_v<ValidArray, std::nullptr_t>)
                     {
                         return get_bottom_tensor_view().template get_vectorized_elements<vector_t>(
                             bottom_tensor_thread_coord,
                             page_offset,
                             bool_constant<oob_conditional_check>{});
                     }
                     else
                     {
                         return get_bottom_tensor_view().template get_vectorized_elements<vector_t>(
                             bottom_tensor_thread_coord,
                             page_offset,
                             valids_[idx_gather],
                             bool_constant<oob_conditional_check>{});
                     }
                 }();
 #if 1
                 // write into distributed tensor
                 static_for<0, Traits::ScalarPerVector, Traits::PackedSize>{}([&](auto j) {
                     constexpr auto idx_ys = generate_tuple(
                         [&](auto jj) {
                             return jj == Traits::VectorDimY ? (idx_ys_start[jj] + j)
                                                             : idx_ys_start[jj];
                         },
                         number<NDimY>{});
  
                     constexpr index_t d =
                         tile_dstr.get_ys_to_d_descriptor().calculate_offset(idx_ys) /
                         Traits::PackedSize;
  
                     dst_tensor.get_thread_buffer().template at<d>() =
                         vec_value.template get_as<DataType>()[j / Traits::PackedSize];
                 });
 #else
                 constexpr index_t d =
                     tile_dstr.get_ys_to_d_descriptor().calculate_offset(idx_ys_start);
                 static_assert(d % Traits::ScalarPerVector == 0);
  
                 dst_tensor.get_thread_buffer().template get_as<vector_t>()(
                     number<d / Traits::ScalarPerVector>{}) = bit_cast<vector_t>(vec_value);
 #endif
                 // move thread coordinate
                 if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
                 {
                     constexpr auto idx_diff_ys = SFC_Ys::get_forward_step(iAccess);
  
                     constexpr auto forward_step_scatter = generate_tuple(
                         [&](auto i) { return i == YsGatherDim ? 0 : idx_diff_ys[i]; },
                         number<NDimY>{});
  
                     constexpr auto idx_diff_ps_ys = container_concat(
                         generate_tuple([&](auto) { return number<0>{}; }, number<NDimP>{}),
                         forward_step_scatter);
  
                     move_window_adaptor_and_bottom_tensor_thread_coordinate(
                         window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
                 }
             });
         });
     }
  
     template <typename LdsTileWindow_,
               index_t i_access_unsupport_ = -1,
               bool oob_conditional_check  = true>
     CK_TILE_DEVICE auto async_load(LdsTileWindow_&& lds_tile,
                                    number<i_access_unsupport_>          = {},
                                    bool_constant<oob_conditional_check> = {}) const
     {
         using LdsTileWindow = remove_cvref_t<LdsTileWindow_>;
         using LdsDataType   = typename LdsTileWindow::DataType;
         using Traits        = load_store_traits;
         using vector_t      = typename Traits::vector_t;
         using SFC_Ys        = typename Traits::SFC_Ys;
  
         constexpr auto tile_dstr = TileDstr{};
  
         // Precompute invariant values outside loops
         const auto window_origin       = lds_tile.get_window_origin();
         const auto& bottom_tensor_view = lds_tile.get_bottom_tensor_view();
         const auto& tensor_descriptor  = bottom_tensor_view.get_tensor_descriptor();
         auto smem_base_ptr             = bottom_tensor_view.get_buffer_view().p_data_;
  
         // loop over thread tensor space [y0, y1, ...]
         static_for<0, NumCoord, 1>{}([&](auto iCoord) {
             auto window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0];
             auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
  
             auto lds_window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0];
             auto lds_bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
  
             static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
                 constexpr auto iAccess = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
  
                 // Use precomputed window origin
                 auto lds_bottom_tensor_thread_idx =
                     window_origin + lds_window_adaptor_thread_coord.get_bottom_index();
                 // Use precomputed tensor descriptor
                 const auto lds_coord =
                     make_tensor_coordinate(tensor_descriptor, lds_bottom_tensor_thread_idx);
                 // Calculate SMEM address using base pointer
                 CK_TILE_LDS_ADDR LdsDataType* smem = smem_base_ptr + lds_coord.get_offset();
  
                 // data index [y0, y1, ...]
                 constexpr auto idx_ys_start = SFC_Ys::get_index(iAccess);
                 constexpr auto idx_gather   = idx_ys_start[number<YsGatherDim>{}];
                 const auto page_offset      = page_idx_[idx_gather];
  
                 // merge page_offset into bottom_coord
                 auto mixed_bottom_thread_coord = bottom_tensor_thread_coord;
                 mixed_bottom_thread_coord.get_hidden_index()[number<0>{}] += page_offset;
  
                 // read from bottom tensor
                 if constexpr(std::is_same_v<ValidArray, std::nullptr_t>)
                     this->get_bottom_tensor_view().template async_get_vectorized_elements<vector_t>(
                         smem,
                         mixed_bottom_thread_coord,
                         number<0>{},
                         bool_constant<oob_conditional_check>{});
                 else
                     this->get_bottom_tensor_view().template async_get_vectorized_elements<vector_t>(
                         smem,
                         mixed_bottom_thread_coord,
                         number<0>{},
                         valids_[idx_gather],
                         bool_constant<oob_conditional_check>{});
  
                 // move thread coordinate
                 if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
                 {
                     constexpr auto idx_diff_ys = SFC_Ys::get_forward_step(iAccess);
  
                     constexpr auto forward_step_scatter = generate_tuple(
                         [&](auto i) { return i == YsGatherDim ? 0 : idx_diff_ys[i]; },
                         number<NDimY>{});
  
                     constexpr auto idx_diff_ps_ys = container_concat(
                         generate_tuple([&](auto) { return number<0>{}; }, number<NDimP>{}),
                         forward_step_scatter);
                     // lds_diff doesn't need to mask the difference of the gather-dim.
                     constexpr auto lds_idx_diff_ps_ys = container_concat(
                         generate_tuple([&](auto) { return number<0>{}; }, number<NDimP>{}),
                         idx_diff_ys);
  
                     move_window_adaptor_and_bottom_tensor_thread_coordinate(
                         window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
                     move_window_adaptor_and_bottom_tensor_thread_coordinate(
                         lds_window_adaptor_thread_coord,
                         lds_bottom_tensor_thread_coord,
                         lds_idx_diff_ps_ys);
                 }
             });
         });
     }
  
     // TODO: currently async load only implemented in inline asm
     template <typename LdsTileWindow_,
               index_t i_access_unsupport_ = -1,
               bool oob_conditional_check  = true,
               bool pre_nop                = false>
     CK_TILE_DEVICE auto async_load_raw(LdsTileWindow_&& lds_tile,
                                        number<i_access_unsupport_>          = {},
                                        bool_constant<oob_conditional_check> = {},
                                        bool_constant<pre_nop>               = {}) const
     {
         using LdsTileWindow = remove_cvref_t<LdsTileWindow_>;
         // using LdsTensorView = typename LdsTileWindow::BottomTensorView;
         using LdsDataType = typename LdsTileWindow::DataType;
         // using LdsDescriptor = typename LdsTileWindow::BottomTensorDesc;
  
         // issues * warps * lanes
         static_assert(LdsTileWindow::get_num_of_dimension() == 3); // TODO: hard coded
  
         const index_t size_per_buf =
             lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
                 make_tuple(number<0>{}, number<0>{}, number<0>{})) *
             sizeof(LdsDataType);
  
         const index_t size_per_wave =
             lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
                 make_tuple(number<0>{}, number<1>{}, number<0>{})) *
                 sizeof(LdsDataType) -
             size_per_buf;
  
         const index_t size_per_issue =
             lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset(
                 make_tuple(number<1>{}, number<0>{}, number<0>{})) *
                 sizeof(LdsDataType) -
             size_per_buf;
  
         const index_t m0_init_value = size_per_buf + size_per_wave * get_warp_id();
         m0_set_with_memory(m0_init_value); // This should be wave independent
  
         using Traits = load_store_traits;
  
         // using vector_type_t = typename Traits::vector_type_t;
         using vector_t = typename Traits::vector_t;
         using SFC_Ys   = typename Traits::SFC_Ys;
  
         LdsDataType* smem = lds_tile.get_bottom_tensor_view().get_buffer_view().p_data_;
  
         // loop over thread tensor space [y0, y1, ...]
         static_for<0, NumCoord, 1>{}([&](auto iCoord) {
             auto window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0];
             auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
  
             static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
                 constexpr auto iAccess  = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
                 constexpr auto pre_nop_ = [&]() {
                     if constexpr(pre_nop && iCoord == 0 && iCoordAccess == 0)
                         return bool_constant<true>{};
                     else
                         return bool_constant<false>{};
                 }();
  
                 constexpr auto idx_ys_start = SFC_Ys::get_index(iAccess);
                 constexpr auto idx_gather   = idx_ys_start[number<YsGatherDim>{}];
                 const auto page_offset      = page_idx_[idx_gather];
  
                 // read from bottom tensor
                 if constexpr(std::is_same_v<ValidArray, std::nullptr_t>)
                 {
                     get_bottom_tensor_view().template async_get_vectorized_elements_raw<vector_t>(
                         smem, bottom_tensor_thread_coord, page_offset, 0, pre_nop_);
                 }
                 else
                 {
                     get_bottom_tensor_view().template async_get_vectorized_elements_raw<vector_t>(
                         smem,
                         bottom_tensor_thread_coord,
                         page_offset,
                         valids_[idx_gather],
                         0,
                         pre_nop_);
                 }
  
                 // move thread coordinate
                 if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
                 {
                     constexpr auto idx_diff_ys = SFC_Ys::get_forward_step(iAccess);
  
                     constexpr auto forward_step_scatter = generate_tuple(
                         [&](auto i) { return i == YsGatherDim ? 0 : idx_diff_ys[i]; },
                         number<NDimY>{});
  
                     constexpr auto idx_diff_ps_ys = container_concat(
                         generate_tuple([&](auto) { return number<0>{}; }, number<NDimP>{}),
                         forward_step_scatter);
  
                     move_window_adaptor_and_bottom_tensor_thread_coordinate(
                         window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
  
                     m0_inc_with_memory(size_per_issue);
                 }
             });
         });
     }
  
     template <index_t i_access_unsupport_ = -1, bool oob_conditional_check = true>
     CK_TILE_DEVICE void update(const static_distributed_tensor<DataType, TileDstr>& dstr_tensor,
                                number<i_access_unsupport_>          = {},
                                bool_constant<oob_conditional_check> = {}) const
     {
         using Traits = load_store_traits;
  
         // using vector_type_t = typename Traits::vector_type_t;
         using vector_t = typename Traits::vector_t;
         using SFC_Ys   = typename Traits::SFC_Ys;
  
         constexpr auto tile_dstr = TileDstr{};
  
         static_for<0, NumCoord, 1>{}([&](auto iCoord) {
             auto window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0];
             auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
  
             static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
                 constexpr auto iAccess = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
  
                 // data index [y0, y1, ...]
                 constexpr auto idx_ys_start = SFC_Ys::get_index(iAccess);
                 constexpr auto idx_gather   = idx_ys_start[number<0>{}];
                 const auto page_offset      = page_idx_[idx_gather];
  
                 // read from distributed tensor
                 vector_t vec_value;
  
                 static_for<0, Traits::ScalarPerVector, Traits::PackedSize>{}([&](auto j) {
                     constexpr auto idx_ys = generate_tuple(
                         [&](auto jj) {
                             return jj == Traits::VectorDimY ? (idx_ys_start[jj] + j)
                                                             : idx_ys_start[jj];
                         },
                         number<NDimY>{});
  
                     constexpr index_t d =
                         tile_dstr.get_ys_to_d_descriptor().calculate_offset(idx_ys) /
                         Traits::PackedSize;
  
                     vec_value.template get_as<DataType>()(j / Traits::PackedSize) =
                         dstr_tensor.get_thread_buffer().template at<d>();
                 });
  
                 // write into bottom tensor
                 if constexpr(std::is_same_v<ValidArray, std::nullptr_t>)
                 {
                     get_bottom_tensor_view().template update_vectorized_elements<vector_t>(
                         bottom_tensor_thread_coord,
                         page_offset,
                         vec_value,
                         bool_constant<oob_conditional_check>{});
                 }
                 else
                 {
                     get_bottom_tensor_view().template update_vectorized_elements<vector_t>(
                         bottom_tensor_thread_coord,
                         page_offset,
                         valids_[idx_gather],
                         vec_value,
                         bool_constant<oob_conditional_check>{});
                 }
  
                 if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
                 {
                     constexpr auto idx_diff_ys = SFC_Ys::get_forward_step(iAccess);
  
                     constexpr auto forward_step_scatter = generate_tuple(
                         [&](auto i) { return i == YsGatherDim ? 0 : idx_diff_ys[i]; },
                         number<NDimY>{});
  
                     constexpr auto idx_diff_ps_ys = container_concat(
                         generate_tuple([&](auto) { return number<0>{}; }, number<NDimP>{}),
                         forward_step_scatter);
  
                     move_window_adaptor_and_bottom_tensor_thread_coordinate(
                         window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
                 }
             });
         });
     }
  
     template <index_t i_access_unsupport_ = -1, bool oob_conditional_check = true>
     CK_TILE_DEVICE void store(const static_distributed_tensor<DataType, TileDstr>& dstr_tensor,
                               number<i_access_unsupport_>          = {},
                               bool_constant<oob_conditional_check> = {}) const
     {
         using Traits = load_store_traits;
  
         // using vector_type_t = typename Traits::vector_type_t;
         using vector_t = typename Traits::vector_t;
         using SFC_Ys   = typename Traits::SFC_Ys;
  
         constexpr auto tile_dstr = TileDstr{};
         // printf("off %d\n", page_idx_[I0]);
         // loop over thread tensor space [y0, y1, ...]
         static_for<0, NumCoord, 1>{}([&](auto iCoord) {
             auto window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0];
             auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
  
             static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
                 constexpr auto iAccess = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
  
                 // data index [y0, y1, ...]
                 constexpr auto idx_ys_start = SFC_Ys::get_index(iAccess);
                 constexpr auto idx_gather   = idx_ys_start[number<0>{}];
                 const auto page_offset      = page_idx_[idx_gather];
  
                 // printf("idx_ys_start[0], idx_ys_start[1](%d, %d) \n",
                 // idx_ys_start[number<0>{}]+0, idx_ys_start[number<1>{}]+0);
  
                 // read from distributed tensor
                 // vector_type_t vec;
                 vector_t vec_value;
  
                 static_for<0, Traits::ScalarPerVector, Traits::PackedSize>{}([&](auto j) {
                     constexpr auto idx_ys = generate_tuple(
                         [&](auto jj) {
                             return jj == Traits::VectorDimY ? (idx_ys_start[jj] + j)
                                                             : idx_ys_start[jj];
                         },
                         number<NDimY>{});
  
                     constexpr index_t d =
                         tile_dstr.get_ys_to_d_descriptor().calculate_offset(idx_ys) /
                         Traits::PackedSize;
                     // printf("thread_idx_m: %d j: %d\n", idx_ys[number<0>{}] + 0, 0+j);
                     vec_value.template get_as<DataType>()(j / Traits::PackedSize) =
                         dstr_tensor.get_thread_buffer().template at<d>();
                 });
  
                 // const vector_t vec_value = vec.template get_as<vector_t>().template at<0>();
  
                 // write into bottom tensor
                 if constexpr(std::is_same_v<ValidArray, std::nullptr_t>)
                 {
                     get_bottom_tensor_view().template set_vectorized_elements<vector_t>(
                         bottom_tensor_thread_coord,
                         page_offset,
                         vec_value,
                         bool_constant<oob_conditional_check>{});
                 }
                 else
                 {
                     get_bottom_tensor_view().template set_vectorized_elements<vector_t>(
                         bottom_tensor_thread_coord,
                         page_offset,
                         valids_[idx_gather],
                         vec_value,
                         bool_constant<oob_conditional_check>{});
                 }
  
                 // printf("coord_offset:%d,   scatter_offset:%d \n",
                 // bottom_tensor_thread_coord.get_offset(), offset); move thread coordinate
                 if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
                 {
                     constexpr auto idx_diff_ys = SFC_Ys::get_forward_step(iAccess);
  
                     constexpr auto forward_step_scatter = generate_tuple(
                         [&](auto i) { return i == YsGatherDim ? 0 : idx_diff_ys[i]; },
                         number<NDimY>{});
  
                     constexpr auto idx_diff_ps_ys = container_concat(
                         generate_tuple([&](auto) { return number<0>{}; }, number<NDimP>{}),
                         forward_step_scatter);
  
                     move_window_adaptor_and_bottom_tensor_thread_coordinate(
                         window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
                 }
             });
         });
     }
  
     // move thread's botom tensor coordiante
     // [x0', x1', ... ] ==> [offset]
     // also move window-origin
     CK_TILE_DEVICE void move(const BottomTensorIndex& step)
     {
         window_origin_ += step;
         BottomTensorIndex step_new = step;
         step_new(HsGatherDim)      = 0;
         static_for<0, NumCoord, 1>{}([&](auto iCoord) {
             move_tensor_coordinate(bottom_tensor_view_.get_tensor_descriptor(),
                                    pre_computed_coords_(iCoord)(I1),
                                    step_new);
         });
     }
  
     CK_TILE_DEVICE void update_page_idx(const PageIdxArray& new_idx) { page_idx_ = new_idx; }
  
     CK_TILE_DEVICE void update_valids(const ValidArray& new_valids)
     {
         if constexpr(std::is_same_v<ValidArray, std::nullptr_t> == false)
         {
             valids_ = new_valids;
         }
     }
  
     CK_TILE_DEVICE void update_page_idx_and_valids(const PageIdxArray& new_idx,
                                                    const ValidArray& new_valids)
     {
         update_page_idx(new_idx);
         update_valids(new_valids);
     }
  
     CK_TILE_DEVICE void set_window_origin(const BottomTensorIndex& new_window_origin)
     {
         window_origin_ = new_window_origin;
  
 #if 0 // debug
       // TODO: this use more register for FA, but less register for GEMM
       // need investigation
       // only support warp-tile and block-tile
         static_assert(NDimP == 1 or NDimP == 2, "wrong!");
  
         WindowAdaptorCoord window_adaptor_thread_coord_tmp;
  
         if constexpr(NDimP == 1)
         {
             window_adaptor_thread_coord_tmp = make_tensor_adaptor_coordinate(
                 tile_dstr_.get_ps_ys_to_xs_adaptor(), AdaptorTopIndex{get_lane_id(), 0});
         }
         else if constexpr(NDimP == 2)
         {
             window_adaptor_thread_coord_tmp =
                 make_tensor_adaptor_coordinate(tile_dstr_.get_ps_ys_to_xs_adaptor(),
                                                AdaptorTopIndex{get_warp_id(), get_lane_id(), 0});
         }
 #else
         // TODO: this use less register for FA, but more register for GEMM
         // need investigation
         const auto window_adaptor_thread_coord_tmp = make_tensor_adaptor_coordinate(
             tile_dstr_.get_ps_ys_to_xs_adaptor(),
             container_concat(get_partition_index(tile_dstr_), array<index_t, NDimY>{0}));
 #endif
  
         BottomTensorIndex bottom_tensor_thread_origin_idx_tmp =
             window_origin_ + window_adaptor_thread_coord_tmp.get_bottom_index();
  
         bottom_tensor_thread_origin_idx_tmp(HsGatherDim) = 0;
         const auto bottom_tensor_thread_coord_tmp        = make_tensor_coordinate(
             bottom_tensor_view_.get_tensor_descriptor(), bottom_tensor_thread_origin_idx_tmp);
  
         // pre-compute NumCoord (WindowAdaptorCoord, BottomTensorCoord) bundles to speed up
         // future load/store() calls (might allocate more registers)
         using Traits = load_store_traits;
         using SFC_Ys = typename Traits::SFC_Ys;
  
         static_for<0, NumCoord, 1>{}([&](auto iCoord) {
             auto window_adaptor_thread_coord = window_adaptor_thread_coord_tmp;
             auto bottom_tensor_thread_coord  = bottom_tensor_thread_coord_tmp;
  
             constexpr auto idx_diff_ys =
                 SFC_Ys::get_step_between(number<0>{}, number<iCoord * NumAccessPerCoord>{});
  
             constexpr auto idx_diff_ps_ys = container_concat(
                 generate_tuple([&](auto) { return number<0>{}; }, number<NDimP>{}), idx_diff_ys);
  
             move_window_adaptor_and_bottom_tensor_thread_coordinate(
                 window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
  
             pre_computed_coords_(iCoord) =
                 make_tuple(window_adaptor_thread_coord, bottom_tensor_thread_coord);
         });
     }
  
     CK_TILE_HOST_DEVICE void init_raw() { bottom_tensor_view_.init_raw(); }
  
     // this is the bottom tensor view
     // [x0', x1', ...] ==> [offset]
     BottomTensorView bottom_tensor_view_;
  
     //
     WindowLengths window_lengths_;
  
     // origin ([x0', x1', ...]) of window on bottom tensor
     BottomTensorIndex window_origin_;
  
     // Tile tensor distribution, which contains:
     //   1. adaptor for window: [p0, p1, ..., y0, y1, ...] ==> [x0, x1, ...]
     //   2. thread descriptor for thread tensor in register: [y0, y1, ...] ==> [d]
     TileDstr tile_dstr_;
  
     PageIdxArray page_idx_;
     ValidArray valids_;
  
     // this contains:
     //   per-thread coordinate for window adaptor
     //   per-thread coordinate for bottom tensor
     array<tuple<WindowAdaptorCoord, BottomTensorCoord>, NumCoord> pre_computed_coords_;
 };
  
 // TODO: use strategy
 template <typename TensorView_,
           typename WindowLengths_,
           typename StaticTileDistribution_,
           typename StaticPageIndexArray_,
           index_t HsGatherDim = 0,
           index_t NumCoord    = 1>
 CK_TILE_DEVICE constexpr auto
 make_tile_scatter_gather(const TensorView_& tensor_view,
                          const WindowLengths_& window_lengths,
                          const multi_index<TensorView_::get_num_of_dimension()>& origin,
                          const StaticTileDistribution_& tile_distribution,
                          const StaticPageIndexArray_& page_idx,
                          number<HsGatherDim> = {},
                          number<NumCoord>    = {})
 {
     return tile_scatter_gather<remove_cvref_t<TensorView_>,
                                remove_cvref_t<WindowLengths_>,
                                remove_cvref_t<StaticTileDistribution_>,
                                remove_cvref_t<StaticPageIndexArray_>,
                                std::nullptr_t,
                                HsGatherDim,
                                NumCoord>{
         tensor_view, window_lengths, origin, tile_distribution, page_idx, nullptr};
 }
  
 template <typename TensorView,
           typename WindowLengths,
           typename StaticTileDistribution,
           typename StaticPageIndexArray,
           index_t HsGatherDim>
 CK_TILE_DEVICE constexpr auto make_tile_scatter_gather(
     const tile_window_with_static_lengths<TensorView, WindowLengths>& tile_window,
     const multi_index<TensorView::get_num_of_dimension()>& origin,
     const StaticTileDistribution& tile_distribution,
     const StaticPageIndexArray& page_idx,
     number<HsGatherDim> = {})
 {
     return make_tile_scatter_gather(tile_window.get_bottom_tensor_view(),
                                     tile_window.get_window_lengths(),
                                     origin,
                                     tile_distribution,
                                     page_idx,
                                     number<HsGatherDim>{});
 }
  
 template <typename TensorView,
           typename WindowLengths,
           typename StaticTileDistribution,
           typename StaticPageIndexArray,
           index_t HsGatherDim>
 CK_TILE_DEVICE constexpr auto make_tile_scatter_gather(
     const tile_window_with_static_lengths<TensorView, WindowLengths>& tile_window,
     const StaticTileDistribution& tile_distribution,
     const StaticPageIndexArray& page_idx,
     number<HsGatherDim> = {})
 {
     return make_tile_scatter_gather(tile_window.get_bottom_tensor_view(),
                                     tile_window.get_window_lengths(),
                                     tile_window.get_window_origin(),
                                     tile_distribution,
                                     page_idx,
                                     number<HsGatherDim>{});
 }
  
 template <typename TensorView_,
           typename WindowLengths_,
           typename StaticTileDistribution_,
           typename StaticPageIndexArray_,
           typename StaticValidArray_,
           index_t HsGatherDim = 0,
           index_t NumCoord    = 1>
 CK_TILE_DEVICE constexpr auto
 make_tile_scatter_gather(const TensorView_& tensor_view,
                          const WindowLengths_& window_lengths,
                          const multi_index<TensorView_::get_num_of_dimension()>& origin,
                          const StaticTileDistribution_& tile_distribution,
                          const StaticPageIndexArray_& page_idx,
                          const StaticValidArray_& valids,
                          number<HsGatherDim> = {},
                          number<NumCoord>    = {})
 {
     return tile_scatter_gather<remove_cvref_t<TensorView_>,
                                remove_cvref_t<WindowLengths_>,
                                remove_cvref_t<StaticTileDistribution_>,
                                remove_cvref_t<StaticPageIndexArray_>,
                                remove_cvref_t<StaticValidArray_>,
                                HsGatherDim,
                                NumCoord>{
         tensor_view, window_lengths, origin, tile_distribution, page_idx, valids};
 }
  
 template <typename TensorView,
           typename WindowLengths,
           typename StaticTileDistribution,
           typename StaticPageIndexArray,
           typename StaticValidArray,
           index_t HsGatherDim>
 CK_TILE_DEVICE constexpr auto make_tile_scatter_gather(
     const tile_window_with_static_lengths<TensorView, WindowLengths>& tile_window,
     const multi_index<TensorView::get_num_of_dimension()>& origin,
     const StaticTileDistribution& tile_distribution,
     const StaticPageIndexArray& page_idx,
     const StaticValidArray& valids,
     number<HsGatherDim> = {})
 {
     return make_tile_scatter_gather(tile_window.get_bottom_tensor_view(),
                                     tile_window.get_window_lengths(),
                                     origin,
                                     tile_distribution,
                                     page_idx,
                                     valids,
                                     number<HsGatherDim>{});
 }
  
 template <typename TensorView,
           typename WindowLengths,
           typename StaticTileDistribution,
           typename StaticPageIndexArray,
           typename StaticValidArray,
           index_t HsGatherDim>
 CK_TILE_DEVICE constexpr auto make_tile_scatter_gather(
     const tile_window_with_static_lengths<TensorView, WindowLengths>& tile_window,
     const StaticTileDistribution& tile_distribution,
     const StaticPageIndexArray& page_idx,
     const StaticValidArray& valids,
     number<HsGatherDim> = {})
 {
     return make_tile_scatter_gather(tile_window.get_bottom_tensor_view(),
                                     tile_window.get_window_lengths(),
                                     tile_window.get_window_origin(),
                                     tile_distribution,
                                     page_idx,
                                     valids,
                                     number<HsGatherDim>{});
 }
  
 template <typename NewTensorView_,
           typename OldTensorView_,
           typename WindowLengths_,
           typename StaticTileDistribution_,
           typename StaticPageIndexArray_,
           typename StaticValidArray_,
           index_t HsGatherDim = 0,
           index_t NumCoord    = 1>
 CK_TILE_DEVICE auto replace_bottom_tensor_view(const NewTensorView_& new_tensor_view,
                                                const tile_scatter_gather<OldTensorView_,
                                                                          WindowLengths_,
                                                                          StaticTileDistribution_,
                                                                          StaticPageIndexArray_,
                                                                          StaticValidArray_,
                                                                          HsGatherDim,
                                                                          NumCoord>& tile_window)
 {
     return make_tile_scatter_gather(new_tensor_view,
                                     tile_window.window_lengths_,
                                     tile_window.window_origin_,
                                     tile_window.tile_dstr_,
                                     tile_window.page_idx_,
                                     tile_window.valids_);
 }
  
 } // namespace ck_tile