/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/host/reference/reference_layernorm2d_fwd.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/host/reference/reference_layernorm2d_fwd.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/host/reference/reference_layernorm2d_fwd.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 "ck_tile/core.hpp"
 #include "ck_tile/host/host_tensor.hpp"
  
 namespace ck_tile {
  
 // Note: for simplicity, each functor only care about single M
 struct reference_layernorm2d_default_epilogue
 {
     template <typename OutDataType, typename AccDataType>
     void operator()(int m, HostTensor<OutDataType>& o, const HostTensor<AccDataType>& acc)
     {
         const int N = acc.mDesc.get_lengths()[1];
         for(int n = 0; n < N; ++n)
         {
             o(m, n) = ck_tile::type_convert<OutDataType>(acc(m, n));
         }
     }
  
     template <typename OutDataType, typename AccDataType>
     auto operator()(int m, const HostTensor<AccDataType>& acc)
     {
         HostTensor<OutDataType> o(acc.get_lengths(), acc.get_strides());
         operator()(m, o, acc);
         return o;
     }
 };
  
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
           typename ComputeDataType,
           typename YDataType,
           typename MeanDataType,
           typename InvStdDataType,
           typename Epilogue = reference_layernorm2d_default_epilogue>
 void reference_layernorm2d_fwd(const HostTensor<XDataType>& x_m_n,
                                const HostTensor<GammaDataType>& gamma_n,
                                const HostTensor<BetaDataType>& beta_n,
                                HostTensor<YDataType>& y_m_n,
                                HostTensor<MeanDataType>& mean_m,
                                HostTensor<InvStdDataType>& invStd_m,
                                ComputeDataType epsilon,
                                Epilogue epilogue_functor = {})
 {
     auto layernorm2d_fwd_func = [&](auto m) {
         const int N = x_m_n.mDesc.get_lengths()[1];
  
         int count                = 0;
         ComputeDataType mean     = 0;
         ComputeDataType variance = 0;
         ComputeDataType divisor  = 0;
  
         for(int n = 0; n < N; ++n)
         {
             ++count;
             ComputeDataType x     = ck_tile::type_convert<ComputeDataType>(x_m_n(m, n));
             ComputeDataType delta = x - mean;
             mean += delta / count;
             ComputeDataType delta2 = x - mean;
             variance += delta * delta2;
         }
  
         // actual variance
         variance = variance / count;
         divisor  = ck_tile::type_convert<ComputeDataType>(1) / ck_tile::sqrt(variance + epsilon);
  
         if constexpr(!std::is_same_v<MeanDataType, ck_tile::null_type>)
             mean_m(m) = ck_tile::type_convert<MeanDataType>(mean);
  
         if constexpr(!std::is_same_v<InvStdDataType, ck_tile::null_type>)
             invStd_m(m) = ck_tile::type_convert<InvStdDataType>(divisor);
  
         HostTensor<ComputeDataType> acc(x_m_n.get_lengths(), x_m_n.get_strides());
         for(int n = 0; n < N; ++n)
         {
             ComputeDataType x     = ck_tile::type_convert<ComputeDataType>(x_m_n(m, n));
             ComputeDataType gamma = ck_tile::type_convert<ComputeDataType>(gamma_n(n));
             ComputeDataType beta  = ck_tile::type_convert<ComputeDataType>(beta_n(n));
             auto a_               = (x - mean) * divisor;
             a_                    = a_ * gamma + beta;
  
             acc(m, n) = a_;
         }
  
         epilogue_functor(m, y_m_n, acc);
     };
  
     make_ParallelTensorFunctor(layernorm2d_fwd_func,
                                mean_m.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
 }
 } // namespace ck_tile