block_norm_reduce.hpp Source File

block_norm_reduce.hpp Source File#

Composable Kernel: block_norm_reduce.hpp Source File
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
#pragma once
 
#include "ck_tile/core.hpp"
#include "ck_tile/ops/norm_reduce/thread/thread_welford.hpp"
 
namespace ck_tile {
 
template <typename Problem_, typename Policy_ = void>
struct BlockNormReduce
{
    using Problem                   = remove_cvref_t<Problem_>;
    using XDataType                 = typename Problem::XDataType;
    using ComputeDataType           = typename Problem::ComputeDataType;
    static constexpr bool kFastFDiv = Problem::kFastFDiv;
    static constexpr bool kWelford  = Problem::kWelford;
 
    CK_TILE_DEVICE constexpr BlockNormReduce() {}
 
    // [CAUSION] - max_count_ is to deal with the padding problem
    // max_count_ is depend on caller, eg: naive and splitN norm_reduce will have different
    // calculation of max_count_
    // -> use block_welford_calculate_max_count to compute
    template <typename XDistributedTensor_,
              typename MeanDistributedTensor_,
              typename VarDistributedTensor_>
    CK_TILE_DEVICE void operator()(const XDistributedTensor_& x_tensor,
                                   MeanDistributedTensor_& mean_tensor,
                                   VarDistributedTensor_& var_tensor,
                                   int& cur_count_, // -> prefer init as zero
                                   const int& max_count_)
    {
        constexpr auto I0 = number<0>{};
        constexpr auto I1 = number<1>{};
 
        constexpr auto spans = XDistributedTensor_::get_distributed_spans();
 
        sweep_tile_span(spans[I1], [&](auto dstr_idx_i1) {
            if(cur_count_ < max_count_)
            {
                ++cur_count_;
                sweep_tile_span(spans[I0], [&](auto dstr_idx_i0) {
                    constexpr auto in_dstr_idx  = make_tuple(dstr_idx_i0, dstr_idx_i1);
                    constexpr auto out_dstr_idx = make_tuple(dstr_idx_i0);
 
                    auto x = ck_tile::type_convert<ComputeDataType>(x_tensor[in_dstr_idx]);
                    if(kWelford)
                    {
                        welford_update(mean_tensor(out_dstr_idx),
                                       var_tensor(out_dstr_idx),
                                       x,
                                       cur_count_,
                                       constant<kFastFDiv>{});
                    }
                    else
                    {
                        mean_tensor(out_dstr_idx) += x;
                        var_tensor(out_dstr_idx) += x * x;
                    }
                });
            }
        });
    }
 
    template <typename XDistributedTensor_>
    CK_TILE_DEVICE static auto MakeMeanVarBlockTile()
    {
        static_assert(std::is_same_v<XDataType, typename XDistributedTensor_::DataType>, "wrong!");
 
        constexpr auto reduce_dims = sequence<1>{};
 
        constexpr auto dstr =
            make_static_tile_distribution(detail::make_reduce_tile_distribution_encoding(
                XDistributedTensor_::get_tile_distribution()
                    .get_static_tile_distribution_encoding(),
                reduce_dims));
 
        auto tensor = make_static_distributed_tensor<ComputeDataType>(dstr);
 
        return tensor;
    }
 
    template <typename XDistributedTensor_>
    CK_TILE_DEVICE auto
    operator()(const XDistributedTensor_& x_tensor, int& cur_count_, const int& max_count_)
    {
        auto mean_tensor = MakeMeanVarBlockTile<XDistributedTensor_>();
        auto var_tensor  = MakeMeanVarBlockTile<XDistributedTensor_>();
        clear_tile(mean_tensor);
        clear_tile(var_tensor);
 
        (*this)(x_tensor, mean_tensor, var_tensor, cur_count_, max_count_);
 
        return ck_tile::make_tuple(mean_tensor, var_tensor);
    }
};
 
template <typename Problem_, typename Policy_ = void>
struct BlockNormReduceSync
{
    using Problem                   = remove_cvref_t<Problem_>;
    static constexpr bool kFastFDiv = Problem::kFastFDiv;
    static constexpr bool kWelford  = Problem::kWelford;
 
    template <typename MeanDistributedTensor_, typename VarDistributedTensor_>
    CK_TILE_DEVICE void
    operator()(MeanDistributedTensor_& mean_tensor, VarDistributedTensor_& var_tensor, int& count)
    {
        using Dstr             = typename MeanDistributedTensor_::StaticTileDistribution;
        using DstrEncode       = typename Dstr::DstrEncode;
        using DstrEncodeDetail = typename DstrEncode::detail;
 
        static_assert(std::is_same_v<Dstr, typename VarDistributedTensor_::StaticTileDistribution>,
                      "wrong!");
 
        constexpr index_t NDimP = Dstr::get_num_of_dimension_p();
        constexpr index_t NDimR = Dstr::get_num_of_dimension_r();
 
        constexpr index_t idim_p_lane = NDimP - 1;
 
        // const auto ps_idx = make_array<index_t>(get_warp_id(), get_lane_id());
        // const auto rs_idx =
        //     mean_tensor.get_tile_distribution().calculate_rs_index_from_ps_index(ps_idx);
 
        constexpr index_t thread_buf_size = MeanDistributedTensor_::get_thread_buffer_size();
        static_assert(thread_buf_size == VarDistributedTensor_::get_thread_buffer_size());
 
        const int original_count = count;
 
        // loop over thread data
        static_for<0, thread_buf_size, 1>{}([&](auto i) {
            auto v_local_mean  = mean_tensor.get_thread_buffer()[i];
            auto v_local_var   = var_tensor.get_thread_buffer()[i];
            auto v_local_count = original_count;
 
            // cross-lane reduce for replication
            // only reduce on R dimension correspond to lane
            // (lane id maps to this R dimension)
            static_for<0, NDimR, 1>{}([&](auto idim_r) {
                // FIXME: nasty to use does_p_own_r_
                if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_lane][idim_r])
                {
                    constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
 
                    constexpr index_t lid_over_rid_derivative =
                        DstrEncodeDetail::ps_over_rs_derivative_[idim_p_lane][idim_r];
 
                    static_assert(is_power_of_two_integer(r_length),
                                  "wrong! only support power of 2 reduction");
 
                    constexpr index_t nstage = integer_log2_floor(r_length);
 
                    // reduction sweep forward
                    static_for<0, nstage, 1>{}([&](auto istage) {
                        // xor
                        index_t src_lane =
                            (__lane_id()) ^
                            (number<lid_over_rid_derivative << istage.value>{}.value);
 
                        // pull data from remote lane
                        const auto v_remote_mean = warp_shuffle(v_local_mean, src_lane);
                        const auto v_remote_var  = warp_shuffle(v_local_var, src_lane);
                        if(kWelford)
                        {
                            const auto v_remote_count = warp_shuffle(v_local_count, src_lane);
 
                            // norm_reduce merge
                            welford_merge(v_local_mean,
                                          v_local_var,
                                          v_local_count,
                                          v_remote_mean,
                                          v_remote_var,
                                          v_remote_count,
                                          constant<kFastFDiv>{});
                        }
                        else
                        {
                            v_local_mean += v_remote_mean;
                            v_local_var += v_remote_var;
                        }
                    });
                }
            });
 
            mean_tensor.get_thread_buffer()(i) = v_local_mean;
            var_tensor.get_thread_buffer()(i)  = v_local_var;
            if(kWelford)
            {
                count = v_local_count;
            }
        });
    }
};
 
template <typename Problem_, typename Policy_ = void>
struct BlockNormReduceCrossWarpSync
{
    using Problem                   = remove_cvref_t<Problem_>;
    using BlockShape                = typename Problem::BlockShape;
    static constexpr bool kFastFDiv = Problem::kFastFDiv;
    static constexpr bool kWelford  = Problem::kWelford;
    using smem_dtype                = std::conditional_t<kWelford, fp32x4_t, fp32x2_t>;
 
    template <typename MeanDistributedTensor_>
    CK_TILE_DEVICE static constexpr index_t GetReduceWarps()
    {
        constexpr index_t num_reduce_warps = [&]() {
            using Dstr             = typename MeanDistributedTensor_::StaticTileDistribution;
            using DstrEncode       = typename Dstr::DstrEncode;
            using DstrEncodeDetail = typename DstrEncode::detail;
 
            constexpr index_t NDimR = Dstr::get_num_of_dimension_r();
 
            constexpr index_t idim_p_warp = 0;
 
            index_t len_ = 1;
            static_for<0, NDimR, 1>{}([&](auto idim_r) {
                if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_warp][idim_r])
                {
                    constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
                    len_ *= r_length;
                }
            });
            return len_;
        }();
        return num_reduce_warps;
    }
 
    // return in byte
    template <typename MeanDistributedTensor_>
    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
    {
        // constexpr auto num_reduce_warps = GetReduceWarps<MeanDistributedTensor_>();
 
        // data need to exchange is very small, we just pack mean+var+count -> 4dword
        constexpr index_t thread_buf_size = MeanDistributedTensor_::get_thread_buffer_size();
 
        // we need to store all data from every wave into smem
        // e.g. 2x2 reduce along N
        //     -------------> reduce N
        //    | w0 | w1 |   ___>      | w01 |
        //    | w2 | w3 |             | w23 |
        //
        //   -> store data from every wave into LDS
        //
        //
        //     -------------> reduce N
        //    | w0 | w1 | w2 | w3 |   ----->  | w0123 |
        //
        //   -> also store data from every wave into LDS
        constexpr index_t num_warps = BlockShape::BlockSize / get_warp_size();
        return num_warps * 4 * thread_buf_size * sizeof(float);
    }
 
    template <typename MeanDistributedTensor_, typename VarDistributedTensor_>
    CK_TILE_DEVICE void operator()(MeanDistributedTensor_& mean_tensor,
                                   VarDistributedTensor_& var_tensor,
                                   int& count,
                                   void* smem)
    {
        using DataType = typename MeanDistributedTensor_::DataType;
        using Dstr     = typename MeanDistributedTensor_::StaticTileDistribution;
        // using DstrEncode       = typename Dstr::DstrEncode;
        // using DstrEncodeDetail = typename DstrEncode::detail;
 
        static_assert(std::is_same_v<Dstr, typename VarDistributedTensor_::StaticTileDistribution>,
                      "wrong!");
 
        constexpr index_t thread_buf_size = MeanDistributedTensor_::get_thread_buffer_size();
        static_assert(thread_buf_size == VarDistributedTensor_::get_thread_buffer_size());
 
        // Note: we always pack everything into fp32x4
        smem_dtype* smem_ptr            = reinterpret_cast<smem_dtype*>(smem);
        const index_t lane_id           = get_lane_id();
        const index_t warp_id           = get_warp_id();
        constexpr auto num_reduce_warps = GetReduceWarps<MeanDistributedTensor_>();
        constexpr index_t num_warps     = BlockShape::BlockSize / get_warp_size();
        const index_t smem_offset       = warp_id;
 
        // skip if nonthing to do
        if constexpr(num_reduce_warps == 1)
            return;
 
        // store into smem only for lane-0 within one warp
        if(lane_id == 0)
        {
            static_for<0, thread_buf_size, 1>{}([&](auto i) {
                smem_dtype local_scratch_;
                local_scratch_[0] = bit_cast<float>(mean_tensor.get_thread_buffer()[i]);
                local_scratch_[1] = bit_cast<float>(var_tensor.get_thread_buffer()[i]);
                if(kWelford)
                {
                    local_scratch_[2] = bit_cast<float>(count);
                }
                smem_ptr[smem_offset + i * num_warps] = local_scratch_;
            });
        }
        block_sync_lds();
 
        // load from smem. here we let everythread to do compute :)
        index_t local_warp_id = warp_id / num_reduce_warps;
        index_t local_smem_os = local_warp_id * num_reduce_warps;
        smem_dtype all_scratch[thread_buf_size * num_reduce_warps];
        static_for<0, thread_buf_size, 1>{}([&](auto i_0) {
            static_for<0, num_reduce_warps, 1>{}([&](auto i_1) {
                all_scratch[i_0 * num_reduce_warps + i_1] =
                    smem_ptr[i_0 * num_warps + local_smem_os + i_1];
            });
        });
        block_sync_lds(); // TODO: we don't need sync here
 
        // const int original_count = count;
 
        static_for<0, thread_buf_size, 1>{}([&](auto i_0) {
            // TODO: use descriptor for this
            auto v_local      = all_scratch[i_0 * num_reduce_warps];
            auto v_local_mean = bit_cast<DataType>(v_local[0]);
            auto v_local_var  = bit_cast<DataType>(v_local[1]);
            int v_local_count = kWelford ? bit_cast<int>(v_local[2]) : 0;
 
            // further reduce mean/var
            static_for<0, num_reduce_warps - 1, 1>{}([&](auto i_1_n1) {
                constexpr auto i_1        = number<i_1_n1 + 1>{};
                const smem_dtype v_remote = all_scratch[i_0 * num_reduce_warps + i_1];
                const auto v_remote_mean  = bit_cast<DataType>(v_remote[0]);
                const auto v_remote_var   = bit_cast<DataType>(v_remote[1]);
                if(kWelford)
                {
                    const auto v_remote_count = bit_cast<int>(v_remote[2]);
 
                    welford_merge(v_local_mean,
                                  v_local_var,
                                  v_local_count,
                                  v_remote_mean,
                                  v_remote_var,
                                  v_remote_count,
                                  constant<kFastFDiv>{});
                }
                else
                {
                    v_local_mean += v_remote_mean;
                    v_local_var += v_remote_var;
                }
            });
 
            mean_tensor.get_thread_buffer()(i_0) = v_local_mean;
            var_tensor.get_thread_buffer()(i_0)  = v_local_var;
            if(kWelford)
                count = v_local_count;
        });
    }
};
 
// compute the max count for a last dim reduce
// everything may have vector/repeat, so the max count could be uneven
// TODO: specify which dim to compute and proper set the problem
// TODO: BlockShape we reuse layernorm_fwd_shape :)
template <typename BlockShape>
CK_TILE_DEVICE constexpr index_t block_tile_welford_calculate_max_count(int row_size)
{
#if 0
    using S                   = BlockShape;
    index_t LastloopN         = row_size % S::Block_N == 0 ? S::Block_N : row_size % S::Block_N;
    constexpr index_t NThread = S::WarpPerBlock_N * S::ThreadPerWarp_N;
    index_t iNLane            = get_thread_id() % NThread;
    index_t iN0               = LastloopN / (S::Vector_N * S::ThreadPerWarp_N);
    index_t iN1               = (LastloopN % (S::Vector_N * S::ThreadPerWarp_N)) / S::Vector_N;
    index_t N2                = (LastloopN % (S::Vector_N * S::ThreadPerWarp_N)) % S::Vector_N;
    index_t iN3               = iNLane < iN1 ? S::Vector_N : iNLane == iN1 ? N2 : 0;
    return iN0 * S::Vector_N + iN3;
#endif
    using S_                            = BlockShape;
    constexpr index_t ThreadsPerBlock_N = S_::WarpPerBlock_N * S_::ThreadPerWarp_N;
 
    // TODO: we always check vector size, need be evenly devidable by vector-n
    const index_t element_per_row = row_size / S_::Vector_N;
    index_t lane_id_n             = get_thread_id() % ThreadsPerBlock_N;
 
    index_t cnt = 0;
    // TODO: Repeat_N can not be too long, otherwise this is not good
    static_for<0, S_::Repeat_N, 1>{}([&](auto) {
        index_t _a = lane_id_n < element_per_row ? 1 : 0;
        cnt += _a;
        lane_id_n += ThreadsPerBlock_N;
    });
    return cnt * S_::Vector_N;
}
 
// Note: this function must be called after all the computation
template <typename VarDistributedTensor_, bool FastFdiv_ = false>
CK_TILE_DEVICE constexpr void block_tile_welford_post_scale_var(VarDistributedTensor_& var_tensor,
                                                                int count,
                                                                bool_constant<FastFdiv_> = {})
{
    using DataType = typename VarDistributedTensor_::DataType;
    tile_elementwise_inout(
        [&count](auto& x) {
            if(FastFdiv_ && std::is_same_v<DataType, float>)
            {
                x = x * __builtin_amdgcn_rcpf(type_convert<DataType>(count));
            }
            else
            {
                x = x / type_convert<DataType>(count);
            }
        },
        var_tensor);
}
} // namespace ck_tile