/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.1.0/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.1.0/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.1.0/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include <functional>
 #include <numeric>
 #include <iterator>
  
 #include "ck/tensor_description/cluster_descriptor.hpp"
 #include "ck/utility/data_type.hpp"
 #include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
  
 namespace ck {
  
 template <typename GridwisePermute,
           typename InGridDesc,
           typename OutGridDesc,
           typename InDataType,
           typename OutDataType,
           typename ElementwiseOperation,
           typename Block2TileMap>
 __global__ void kernel_nd_permute(const InGridDesc in_grid_desc,
                                   const OutGridDesc out_grid_desc,
                                   const InDataType* p_in_global,
                                   OutDataType* p_out_global,
                                   const ElementwiseOperation elementwise_op,
                                   const Block2TileMap block_2_tile_map)
 {
     __shared__ char p_shared[GridwisePermute::GetSharedMemoryNumberOfByte()];
  
     GridwisePermute::Run(in_grid_desc,
                          out_grid_desc,
                          p_in_global,
                          p_out_global,
                          p_shared,
                          elementwise_op,
                          block_2_tile_map);
 }
  
 template <typename InGridDesc,
           typename OutGridDesc,
           typename InDataType,
           typename OutDataType,
           typename ElementwiseOperation,
           index_t BlockSize,
           index_t NPerBlock,
           index_t HPerBlock,
           index_t WPerBlock,
           index_t InBlockLdsExtraW,
           typename InBlockTransferThreadClusterLengths,
           typename InBlockTransferThreadClusterArrangeOrder,
           index_t SrcVectorDim,
           index_t DstVectorDim,
           index_t SrcScalarPerVector,
           index_t DstScalarPerVector>
 struct GridwisePermute
 {
     static_assert(InGridDesc::GetNumOfDimension() == OutGridDesc::GetNumOfDimension());
     static_assert(3 <= InGridDesc::GetNumOfDimension());
     static_assert((InGridDesc::GetNumOfDimension() - 2) <= SrcVectorDim &&
                   SrcVectorDim < InGridDesc::GetNumOfDimension());
     static_assert((OutGridDesc::GetNumOfDimension() - 2) <= DstVectorDim &&
                   DstVectorDim < OutGridDesc::GetNumOfDimension());
     static_assert(SrcVectorDim != DstVectorDim);
  
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
  
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
  
     struct Block2TileMap
     {
         static constexpr index_t NumDim = InGridDesc::GetNumOfDimension();
         static_assert(3 <= NumDim);
  
         static constexpr auto I0 = Number<0>{};
  
         Block2TileMap()                     = delete;
         Block2TileMap(const Block2TileMap&) = default;
         Block2TileMap(Block2TileMap&&)      = delete;
  
         ~Block2TileMap() = default;
  
         Block2TileMap& operator=(const Block2TileMap&) = delete;
         Block2TileMap& operator=(Block2TileMap&&)      = delete;
  
         explicit Block2TileMap(const InGridDesc& desc) : desc_(desc) {}
  
         __host__ constexpr index_t CalculateGridSize(const InGridDesc& desc) const
         {
             const auto N0 =
                 math::integer_divide_ceil(desc.GetLength(Number<NumDim - 3>{}), NPerBlock);
             const auto H0 =
                 math::integer_divide_ceil(desc.GetLength(Number<NumDim - 2>{}), HPerBlock);
             const auto W0 =
                 math::integer_divide_ceil(desc.GetLength(Number<NumDim - 1>{}), WPerBlock);
  
             const index_t grid_size = N0 * H0 * W0;
  
             return grid_size;
         }
  
         template <typename TopIdx>
         __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
         {
             static_assert(TopIdx::Size() == 1);
  
             auto block_1d_id = idx_top[I0];
  
             const auto N0 =
                 math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 3>{}), NPerBlock);
             const auto H0 =
                 math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 2>{}), HPerBlock);
             const auto W0 =
                 math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 1>{}), WPerBlock);
  
             block_1d_id = block_1d_id % (N0 * H0 * W0);
  
             index_t idx_N0 = block_1d_id / (H0 * W0);
             index_t idx_H0 = (block_1d_id % (H0 * W0)) / W0;
             index_t idx_W0 = block_1d_id % W0;
  
             return make_tuple(idx_N0, idx_H0, idx_W0);
         }
  
         private:
         const InGridDesc desc_;
     };
  
     using DefaultBlock2TileMap = Block2TileMap;
  
     // use an [NPerBlock, HPerBlock, WPerBlock] tensor as element-copy relay
     __host__ __device__ static constexpr auto GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock()
     {
         return make_naive_tensor_descriptor(
             make_tuple(Number<NPerBlock>{}, Number<HPerBlock>{}, Number<WPerBlock>{}),
             make_tuple(Number<HPerBlock*(WPerBlock + InBlockLdsExtraW)>{},
                        Number<WPerBlock + InBlockLdsExtraW>{},
                        I1));
     }
  
     // for N-dimension descriptor, reserve its last 2 dimensions, then merge its leading dimensions
     // into single one. finally, form a 3D descriptor: [d(0), d(1), ..., d(N - 2), d(N - 1)] ->
     // [(d(0) x d(1) x ...), d(N - 2), d(N - 1)]
     template <typename GridDesc>
     __host__ __device__ static constexpr auto GetMergedDesc(const GridDesc& desc)
     {
         constexpr index_t NumDim = GridDesc::GetNumOfDimension();
         static_assert(3 <= NumDim);
  
         const auto merged_desc = transform_tensor_descriptor(
             desc,
             make_tuple(make_merge_transform(generate_tuple(
                            [&](auto I) { return desc.GetLength(I); }, Number<NumDim - 2>{})),
                        make_pass_through_transform(desc.GetLength(Number<NumDim - 2>{})),
                        make_pass_through_transform(desc.GetLength(Number<NumDim - 1>{}))),
             make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NumDim - 2>{}),
                        Sequence<NumDim - 2>{},
                        Sequence<NumDim - 1>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
         return merged_desc;
     }
  
     __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
     {
         constexpr auto in_block_desc_nperblock_hperblock_wperblock =
             GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock();
  
         return in_block_desc_nperblock_hperblock_wperblock.GetElementSpaceSize() *
                sizeof(InDataType);
     }
  
     __host__ __device__ static constexpr auto MakeDefaultBlock2TileMap(const InGridDesc& desc)
     {
         return DefaultBlock2TileMap{desc};
     }
  
     __host__ __device__ static constexpr bool CheckValidity(const InGridDesc& in_grid_desc,
                                                             const OutGridDesc& out_grid_desc)
     {
         constexpr index_t NumDim = InGridDesc::GetNumOfDimension();
  
         // check if we only swap last 2 dimensions
         bool valid = true;
         static_for<0, NumDim - 2, 1>{}([&](auto I) {
             if(valid && in_grid_desc.GetLength(I) != out_grid_desc.GetLength(I))
             {
                 valid = false;
             }
         });
  
         return valid &&
                (in_grid_desc.GetLength(Number<NumDim - 1>{}) ==
                 out_grid_desc.GetLength(Number<NumDim - 2>{})) &&
                (in_grid_desc.GetLength(Number<NumDim - 2>{}) ==
                 out_grid_desc.GetLength(Number<NumDim - 1>{}));
     }
  
     template <typename Block2TileMap>
     __device__ static void Run(const InGridDesc in_grid_desc,
                                const OutGridDesc out_grid_desc,
                                const InDataType* p_in_global,
                                OutDataType* p_out_global,
                                void* __restrict__ p_shared,
                                const ElementwiseOperation elementwise_op,
                                const Block2TileMap& block_2_tile_map)
     {
         auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_in_global, in_grid_desc.GetElementSpaceSize());
  
         auto out_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_out_global, out_grid_desc.GetElementSpaceSize());
  
         // each workgroup handles an [NPerBlock, HPerBlock, WPerBLock] slice-transpose problem
         const auto block_work_idx =
             block_2_tile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
  
         const index_t n_block_data_idx_on_grid =
             __builtin_amdgcn_readfirstlane(block_work_idx[I0] * NPerBlock);
  
         const index_t h_block_data_idx_on_grid =
             __builtin_amdgcn_readfirstlane(block_work_idx[I1] * HPerBlock);
  
         const index_t w_block_data_idx_on_grid =
             __builtin_amdgcn_readfirstlane(block_work_idx[I2] * WPerBlock);
  
         // create [NPerBlock, HPerBlock, WPerBLock] shaped LDS buffer
         constexpr auto in_block_desc_nperblock_hperblock_wperblock =
             GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock();
  
         auto in_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             static_cast<InDataType*>(p_shared),
             in_block_desc_nperblock_hperblock_wperblock.GetElementSpaceSize());
  
         using BlockSliceLengths          = Sequence<NPerBlock, HPerBlock, WPerBlock>;
         using InBlockTransferAccessOrder = Sequence<0, 1, 2>;
  
         constexpr index_t SrcVectorDimAfterMerge =
             SrcVectorDim - (InGridDesc::GetNumOfDimension() - 3);
         constexpr index_t DstVectorDimAfterMerge = SrcVectorDimAfterMerge;
  
         using ck::tensor_operation::element_wise::PassThrough;
  
         // merge input descriptor into [(in_grid_desc.GetLength(0) x in_grid_desc.GetLength(1) x
         // ...), in_grid_desc.GetLength(NumDim - 2), in_grid_desc.GetLength(NumDim - 1)]
         const auto in_grid_desc_n_h_w = GetMergedDesc(in_grid_desc);
  
         // a workgroup copies an [NPerBlock, HPerBlock, WPerBlock] slice from global memory to LDS
         auto in_global_load = ThreadGroupTensorSliceTransfer_v4r1<
             ThisThreadBlock,
             ElementwiseOperation,
             PassThrough,
             InMemoryDataOperationEnum::Set,
             BlockSliceLengths,
             InBlockTransferThreadClusterLengths,
             InBlockTransferThreadClusterArrangeOrder,
             InDataType,
             InDataType,
             decltype(in_grid_desc_n_h_w),
             decltype(in_block_desc_nperblock_hperblock_wperblock),
             InBlockTransferAccessOrder,
             InBlockTransferAccessOrder,
             SrcVectorDimAfterMerge,
             2,
             SrcScalarPerVector,
             1,
             1,
             1,
             true,
             true>(in_grid_desc_n_h_w,
                   make_multi_index(
                       n_block_data_idx_on_grid, h_block_data_idx_on_grid, w_block_data_idx_on_grid),
                   PassThrough{},
                   in_block_desc_nperblock_hperblock_wperblock,
                   make_multi_index(0, 0, 0),
                   PassThrough{});
  
         // merge output descriptor into [(out_grid_desc.GetLength(0) x out_grid_desc.GetLength(1) x
         // ...), out_grid_desc.GetLength(NumDim - 2), out_grid_desc.GetLength(NumDim - 1)]
         const auto out_grid_desc_n_w_h = GetMergedDesc(out_grid_desc);
  
         // create transposed view of output tensor
         const auto out_grid_desc_n_h_w = transform_tensor_descriptor(
             out_grid_desc_n_w_h,
             make_tuple(make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I0)),
                        make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I1)),
                        make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I2))),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
             make_tuple(Sequence<0>{}, Sequence<2>{}, Sequence<1>{}));
  
         // a workgroup copies an [NPerBlock, HPerBlock, WPerBlock] slice from LDS to global memory
         auto out_global_store = ThreadGroupTensorSliceTransfer_v4r1<
             ThisThreadBlock,
             ElementwiseOperation,
             PassThrough,
             InMemoryDataOperationEnum::Set,
             BlockSliceLengths,
             InBlockTransferThreadClusterLengths,
             InBlockTransferThreadClusterArrangeOrder,
             InDataType,
             OutDataType,
             decltype(in_block_desc_nperblock_hperblock_wperblock),
             decltype(out_grid_desc_n_h_w),
             InBlockTransferAccessOrder,
             InBlockTransferAccessOrder,
             2,
             DstVectorDimAfterMerge,
             1,
             DstScalarPerVector,
             1,
             1,
             true,
             true>(in_block_desc_nperblock_hperblock_wperblock,
                   make_multi_index(0, 0, 0),
                   PassThrough{},
                   out_grid_desc_n_h_w,
                   make_multi_index(
                       n_block_data_idx_on_grid, h_block_data_idx_on_grid, w_block_data_idx_on_grid),
                   elementwise_op);
  
         in_global_load.Run(in_grid_desc_n_h_w,
                            in_global_buf,
                            in_block_desc_nperblock_hperblock_wperblock,
                            in_block_buf,
                            I0);
  
         out_global_store.Run(in_block_desc_nperblock_hperblock_wperblock,
                              in_block_buf,
                              out_grid_desc_n_h_w,
                              out_global_buf,
                              I0);
     }
 };
  
 } // namespace ck