/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.3/include/ck/tensor_operation/gpu/device/impl/device_avgpool2d_bwd_nhwc_nhwc.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.3/include/ck/tensor_operation/gpu/device/impl/device_avgpool2d_bwd_nhwc_nhwc.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.3/include/ck/tensor_operation/gpu/device/impl/device_avgpool2d_bwd_nhwc_nhwc.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include <iostream>
 #include <sstream>
  
 #include "ck/tensor_description/tensor_descriptor.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
 #include "ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
  
 namespace ck {
 namespace tensor_operation {
 namespace device {
  
 // In and Din = [N, C, Hi, Wi]
 // Out and Dout = [N, C, Ho, Wo]
 // Out = AvgPool2dFwd(In)
 // Din = AvgPool2dBwd(Dout)
 // Pooling dimension = H, W
 template <typename DOutDataType,
           typename DInDataType,
           typename ComputeDataType,
           ck::index_t BlockSize,
           ck::index_t MThreadClusterSize,
           ck::index_t KThreadClusterSize,
           ck::index_t MThreadSliceSize,
           ck::index_t KThreadSliceSize,
           ck::index_t InSrcOutDstVectorSize>
 struct DeviceAvgPool2dBwd_NHWC_NHWC : public DeviceAvgPoolBwd<2,
                                                               DOutDataType,
                                                               DInDataType,
                                                               tensor_layout::convolution::NHWC,
                                                               tensor_layout::convolution::NHWC>
 {
  
     static constexpr ck::index_t NDimSpatial = 2;
  
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
  
     static constexpr ck::index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
     static constexpr ck::index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
  
     static auto
     Make2DGridDescriptor_Out_M_K_In_M(const std::vector<ck::index_t>& dout_n_c_wos_lengths,
                                       const std::vector<ck::index_t>& din_n_c_wos_length,
                                       const std::vector<ck::index_t>& dout_n_c_wos_strides,
                                       const std::vector<ck::index_t>& din_n_c_wos_strides,
                                       const std::vector<ck::index_t>& window_lengths,
                                       const std::vector<ck::index_t>& window_strides,
                                       const std::vector<ck::index_t>& window_dilations,
                                       const std::vector<ck::index_t>& input_left_pads,
                                       const std::vector<ck::index_t>& input_right_pads,
                                       const std::vector<ck::index_t>& tildes)
     {
         index_t i_ytilde = tildes[0];
         index_t i_xtilde = tildes[1];
  
         const index_t N  = dout_n_c_wos_lengths[0];
         const index_t C  = dout_n_c_wos_lengths[1];
         const index_t Ho = dout_n_c_wos_lengths[2];
         const index_t Wo = dout_n_c_wos_lengths[3];
  
         const index_t Hi = din_n_c_wos_length[2];
         const index_t Wi = din_n_c_wos_length[3];
  
         const index_t Y = window_lengths[0];
         const index_t X = window_lengths[1];
  
         const index_t InLeftPadH = input_left_pads[0];
         const index_t InLeftPadW = input_left_pads[1];
  
         const index_t InRightPadH = input_right_pads[0];
         const index_t InRightPadW = input_right_pads[1];
  
         const index_t ConvStrideH = window_strides[0];
         const index_t ConvStrideW = window_strides[1];
  
         const index_t ConvDilationH = window_dilations[0];
         const index_t ConvDilationW = window_dilations[1];
  
         const index_t Ni_stride = dout_n_c_wos_strides[0];
         const index_t Ci_stride = dout_n_c_wos_strides[1];
         const index_t Ho_stride = dout_n_c_wos_strides[2];
         const index_t Wo_stride = dout_n_c_wos_strides[3];
  
         const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
         const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
  
         const auto YTilde = ConvStrideH / GcdStrideDilationH;
         const auto XTilde = ConvStrideW / GcdStrideDilationW;
  
         const auto YDot = math::integer_divide_ceil(Y, YTilde);
         const auto XDot = math::integer_divide_ceil(X, XTilde);
  
         const auto HTilde = Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
         const auto WTilde = Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
  
         // only work on Tildes that contribute to non-padding area of input tensor
         const auto IHTildeSliceBegin = math::integer_divide_floor(
             math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
         const auto IWTildeSliceBegin = math::integer_divide_floor(
             math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
  
         const auto IHTildeSliceEnd =
             math::min(HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
         const auto IWTildeSliceEnd =
             math::min(WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
  
         const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
         const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
  
         // ReduceK is different for each Reduce
         const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
         const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
  
         // Problem size of reduction kernel
         const index_t MRaw = N * HTildeSlice * WTildeSlice * C;
         const index_t MPad = math::integer_least_multiple(MRaw, M_BlockTileSize) - MRaw;
  
         const index_t KRaw = YDotSlice * XDotSlice;
         const index_t KPad = math::integer_least_multiple(KRaw, K_BlockTileSize) - KRaw;
  
         const auto out_n_ho_wo_c_grid_desc = make_naive_tensor_descriptor(
             make_tuple(N, Ho, Wo, C), make_tuple(Ni_stride, Ho_stride, Wo_stride, Ci_stride));
  
         // Out[ReduceM, ReduceK]
         const auto out_n_hop_wop_c_grid_desc = transform_tensor_descriptor(
             out_n_ho_wo_c_grid_desc,
             make_tuple(make_pass_through_transform(N),
                        make_pad_transform(Ho, I0, I0),
                        make_pad_transform(Wo, I0, I0),
                        make_pass_through_transform(C)),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
  
         const auto out_n_ydot_htilde_xdot_wtilde_c_grid_desc = transform_tensor_descriptor(
             out_n_hop_wop_c_grid_desc,
             make_tuple(make_pass_through_transform(N),
                        make_embed_transform(make_tuple(YDot, HTilde),
                                             make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
                        make_embed_transform(make_tuple(XDot, WTilde),
                                             make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
                        make_pass_through_transform(C)),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
             make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
  
         const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_c_grid_desc =
             transform_tensor_descriptor(
                 out_n_ydot_htilde_xdot_wtilde_c_grid_desc,
                 make_tuple(make_pass_through_transform(N),
                            make_slice_transform(YDot, I0, YDotSlice),
                            make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
                            make_slice_transform(XDot, I0, XDotSlice),
                            make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
                            make_pass_through_transform(C)),
                 make_tuple(Sequence<0>{},
                            Sequence<1>{},
                            Sequence<2>{},
                            Sequence<3>{},
                            Sequence<4>{},
                            Sequence<5>{}),
                 make_tuple(Sequence<0>{},
                            Sequence<1>{},
                            Sequence<2>{},
                            Sequence<3>{},
                            Sequence<4>{},
                            Sequence<5>{}));
  
         const auto out_grid_desc_reducemraw_reducekraw = transform_tensor_descriptor(
             out_n_ydotslice_htildeslice_xdotslice_wtildeslice_c_grid_desc,
             make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice, C)),
                        make_merge_transform(make_tuple(YDotSlice, XDotSlice))),
             make_tuple(Sequence<0, 2, 4, 5>{}, Sequence<1, 3>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}));
  
         const auto out_grid_desc_reducem_reducek = transform_tensor_descriptor(
             out_grid_desc_reducemraw_reducekraw,
             make_tuple(make_right_pad_transform(MRaw, MPad), make_right_pad_transform(KRaw, KPad)),
             make_tuple(Sequence<0>{}, Sequence<1>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}));
  
         // In[ReduceM]
         const auto in_n_hi_wi_c_grid_desc =
             make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
                                          make_tuple(din_n_c_wos_strides[0],
                                                     din_n_c_wos_strides[2],
                                                     din_n_c_wos_strides[3],
                                                     din_n_c_wos_strides[1]));
  
         const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
             in_n_hi_wi_c_grid_desc,
             make_tuple(make_pass_through_transform(N),
                        make_pad_transform(Hi, InLeftPadH, InRightPadH),
                        make_pad_transform(Wi, InLeftPadW, InRightPadW),
                        make_pass_through_transform(C)),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
  
         const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
             in_n_hip_wip_c_grid_desc,
             make_tuple(make_pass_through_transform(N),
                        make_embed_transform(make_tuple(YTilde, HTilde),
                                             make_tuple(ConvDilationH, ConvStrideH)),
                        make_embed_transform(make_tuple(XTilde, WTilde),
                                             make_tuple(ConvDilationW, ConvStrideW)),
                        make_pass_through_transform(C)),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
             make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
  
         const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
             in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
             make_tuple(make_pass_through_transform(N),
                        make_freeze_transform(i_ytilde),
                        make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
                        make_freeze_transform(i_xtilde),
                        make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
                        make_pass_through_transform(C)),
             make_tuple(Sequence<0>{},
                        Sequence<1>{},
                        Sequence<2>{},
                        Sequence<3>{},
                        Sequence<4>{},
                        Sequence<5>{}),
             make_tuple(Sequence<0>{},
                        Sequence<>{},
                        Sequence<1>{},
                        Sequence<>{},
                        Sequence<2>{},
                        Sequence<3>{}));
  
         const auto in_grid_desc_reducemraw = transform_tensor_descriptor(
             in_n_htildeslice_wtildeslice_c_grid_desc,
             make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice, C))),
             make_tuple(Sequence<0, 1, 2, 3>{}),
             make_tuple(Sequence<0>{}));
  
         const auto in_grid_desc_reducem =
             transform_tensor_descriptor(in_grid_desc_reducemraw,
                                         make_tuple(make_right_pad_transform(MRaw, MPad)),
                                         make_tuple(Sequence<0>{}),
                                         make_tuple(Sequence<0>{}));
  
         return make_tuple(out_grid_desc_reducem_reducek, in_grid_desc_reducem);
     }
  
     using DoutDinGridDesc = decltype(Make2DGridDescriptor_Out_M_K_In_M({0, 0, 0, 0},
                                                                        {0, 0, 0, 0},
                                                                        {0, 0, 0, 0},
                                                                        {0, 0, 0, 0},
                                                                        {0, 0},
                                                                        {0, 0},
                                                                        {0, 0},
                                                                        {0, 0},
                                                                        {0, 0},
                                                                        {0, 0}));
  
     using DoutGridDesc_M_K = remove_cvref_t<tuple_element_t<0, DoutDinGridDesc>>;
     using DinGridDesc_M    = remove_cvref_t<tuple_element_t<1, DoutDinGridDesc>>;
  
     // FIXME
     // for NHWC, the dim C is the fastest dimension, and is not reduced.
     // Hence, it is in M dimension for reduction kernel.
     static constexpr index_t OutSrcInDstVectorDim = 0; // 0: M, 1: K
  
     using PassThrough = tensor_operation::element_wise::PassThrough;
     using Div         = tensor_operation::element_wise::UnaryDivide;
  
     using gridwise_reduce = GridwiseReduction_mk_to_m_threadwise<DOutDataType,
                                                                  DInDataType,
                                                                  ComputeDataType,
                                                                  int,
                                                                  DoutGridDesc_M_K,
                                                                  DinGridDesc_M,
                                                                  reduce::Add,
                                                                  PassThrough,
                                                                  Div,
                                                                  InMemoryDataOperationEnum::Set,
                                                                  false, // propagate_nan
                                                                  BlockSize,
                                                                  MThreadSliceSize,
                                                                  KThreadSliceSize,
                                                                  OutSrcInDstVectorDim,
                                                                  InSrcOutDstVectorSize,
                                                                  InSrcOutDstVectorSize>;
  
     struct Argument : public BaseArgument
     {
         Argument(const DOutDataType* p_dout,
                  DInDataType* p_din,
                  std::vector<ck::index_t> dout_n_c_wos_lengths,
                  std::vector<ck::index_t> din_n_c_wos_length,
                  std::vector<ck::index_t> dout_n_c_wos_strides,
                  std::vector<ck::index_t> din_n_c_wos_strides,
                  std::vector<ck::index_t> window_lengths,
                  std::vector<ck::index_t> window_strides,
                  std::vector<ck::index_t> window_dilations,
                  std::vector<ck::index_t> input_left_pads,
                  std::vector<ck::index_t> input_right_pads)
             : p_dout_grid_{p_dout},
               p_din_grid_{p_din},
               dout_n_c_wos_lengths_{dout_n_c_wos_lengths},
               din_n_c_wos_length_{din_n_c_wos_length},
               dout_n_c_wos_strides_{dout_n_c_wos_strides},
               din_n_c_wos_strides_{din_n_c_wos_strides},
               num_reduce_{1},
               div_element_op_{window_lengths[0] * window_lengths[1]}
         {
             std::vector<ck::index_t> Tildes(NDimSpatial);
             for(int i = 0; i < NDimSpatial; ++i)
             {
                 int GcdStrideDilation = math::gcd(window_strides[i], window_dilations[i]);
                 Tildes[i]             = window_strides[i] / GcdStrideDilation;
                 num_reduce_ *= Tildes[i];
             }
  
             for(index_t i_ytilde = 0; i_ytilde < Tildes[0]; ++i_ytilde)
             {
                 for(index_t i_xtilde = 0; i_xtilde < Tildes[1]; ++i_xtilde)
                 {
                     const auto YDotSlice =
                         math::integer_divide_ceil(window_lengths[0] - i_ytilde, Tildes[0]);
                     const auto XDotSlice =
                         math::integer_divide_ceil(window_lengths[1] - i_xtilde, Tildes[1]);
  
                     if(YDotSlice * XDotSlice <= 0)
                     {
                         continue;
                     }
  
                     const auto dout_din_grid_desc =
                         Make2DGridDescriptor_Out_M_K_In_M(dout_n_c_wos_lengths,
                                                           din_n_c_wos_length,
                                                           dout_n_c_wos_strides,
                                                           din_n_c_wos_strides,
                                                           window_lengths,
                                                           window_strides,
                                                           window_dilations,
                                                           input_left_pads,
                                                           input_right_pads,
                                                           {i_ytilde, i_xtilde});
  
                     dout_grid_desc_m_k_container_.push_back(dout_din_grid_desc[I0]);
                     din_grid_desc_m_container_.push_back(dout_din_grid_desc[I1]);
                 }
             }
         }
  
         const DOutDataType* p_dout_grid_;
         DInDataType* p_din_grid_;
         std::vector<ck::index_t> dout_n_c_wos_lengths_;
         std::vector<ck::index_t> din_n_c_wos_length_;
         std::vector<ck::index_t> dout_n_c_wos_strides_;
         std::vector<ck::index_t> din_n_c_wos_strides_;
  
         int num_reduce_;
         std::vector<DoutGridDesc_M_K> dout_grid_desc_m_k_container_;
         std::vector<DinGridDesc_M> din_grid_desc_m_container_;
  
         Div div_element_op_;
     };
  
     struct Invoker : public BaseInvoker
     {
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
             float ave_time = 0;
  
             for(index_t i = 0; i < arg.num_reduce_; i++)
             {
                 const auto kernel = kernel_reduce_threadwise<gridwise_reduce,
                                                              false,
                                                              false,
                                                              false, // don't have index input
                                                              DOutDataType,
                                                              DInDataType,
                                                              ComputeDataType,
                                                              int,
                                                              DoutGridDesc_M_K,
                                                              DinGridDesc_M,
                                                              PassThrough,
                                                              Div>;
  
                 ck::index_t M           = arg.dout_grid_desc_m_k_container_[i].GetLength(I0);
                 const index_t grid_size = (M / M_BlockTileSize);
  
                 ave_time += launch_and_time_kernel(stream_config,
                                                    kernel,
                                                    dim3(grid_size),
                                                    dim3(BlockSize),
                                                    0,
                                                    arg.dout_grid_desc_m_k_container_[i],
                                                    arg.din_grid_desc_m_container_[i],
                                                    PassThrough{},
                                                    arg.div_element_op_,
                                                    float(1),
                                                    arg.p_dout_grid_,
                                                    nullptr,
                                                    float(0),
                                                    arg.p_din_grid_,
                                                    nullptr);
             }
  
             return ave_time;
         }
  
         float Run(const BaseArgument* p_arg,
                   const StreamConfig& stream_config = StreamConfig{}) override
         {
             return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
         }
     };
  
     static bool IsSupportedArgument(const Argument& arg)
     {
         constexpr index_t Rank = NDimSpatial + 2;
         int doutFastestDim     = -1;
         int dinFastestDim      = -1;
  
         for(int i = 0; i < Rank; ++i)
         {
             if(arg.dout_n_c_wos_strides_[i] == 1)
                 doutFastestDim = i;
             if(arg.din_n_c_wos_strides_[i] == 1)
                 dinFastestDim = i;
         }
         if(InSrcOutDstVectorSize != 1 && (dinFastestDim != 1 || doutFastestDim != 1))
         {
             return false;
         }
         if(doutFastestDim == -1 || dinFastestDim == -1)
         {
             if constexpr(InSrcOutDstVectorSize != 1)
                 return false;
         }
         else
         {
             if(arg.dout_n_c_wos_lengths_[doutFastestDim] % InSrcOutDstVectorSize != 0)
                 return false;
             if(arg.din_n_c_wos_length_[dinFastestDim] % InSrcOutDstVectorSize != 0)
                 return false;
         }
         return true;
     }
  
     bool IsSupportedArgument(const BaseArgument* p_arg) override
     {
         return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
     }
  
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_dout,
                         void* p_din,
                         std::vector<ck::index_t> dout_n_c_wos_lengths,
                         std::vector<ck::index_t> din_n_c_wos_length,
                         std::vector<ck::index_t> dout_n_c_wos_strides,
                         std::vector<ck::index_t> din_n_c_wos_strides,
                         std::vector<ck::index_t> window_lengths,
                         std::vector<ck::index_t> window_strides,
                         std::vector<ck::index_t> window_dilations,
                         std::vector<ck::index_t> input_left_pads,
                         std::vector<ck::index_t> input_right_pads) override
     {
         constexpr index_t Rank = NDimSpatial + 2;
  
         if(dout_n_c_wos_strides.size() != Rank || din_n_c_wos_strides.size() != Rank ||
            dout_n_c_wos_lengths.size() != Rank || din_n_c_wos_length.size() != Rank)
         {
             throw std::runtime_error("dimension of [dout|din]_n_c_wos_strides or "
                                      "[dout|din]_n_c_wos_lengths is not equal to Rank");
         }
  
         if(window_lengths.size() != NDimSpatial || window_strides.size() != NDimSpatial ||
            window_dilations.size() != NDimSpatial || input_left_pads.size() != NDimSpatial ||
            input_right_pads.size() != NDimSpatial)
         {
             throw std::runtime_error(
                 "dimension of [window_lengths, window_strides, window_dilations, input_left_pads, "
                 "input_right_pads] is not equal to Rank");
         }
         return std::make_unique<Argument>(static_cast<const DOutDataType*>(p_dout),
                                           static_cast<DInDataType*>(p_din),
                                           dout_n_c_wos_lengths,
                                           din_n_c_wos_length,
                                           dout_n_c_wos_strides,
                                           din_n_c_wos_strides,
                                           window_lengths,
                                           window_strides,
                                           window_dilations,
                                           input_left_pads,
                                           input_right_pads);
     }
  
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
         return std::make_unique<Invoker>(Invoker{});
     }
  
     std::string GetTypeString() const override
     {
         auto str = std::stringstream();
  
         // clang-format off
         str << "DeviceAvgPool2dBwd<" << BlockSize << ",";
         str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
         str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
         str <<"InSrcOutDstVectorSize_" << InSrcOutDstVectorSize << ">";
         // clang-format on
  
         return str.str();
     }
 };
  
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck