/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/library/utility/host_gemm.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/library/utility/host_gemm.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/library/utility/host_gemm.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 "host_tensor.hpp"
  
 template <typename AType,
           typename BType,
           typename CType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>
 void host_gemm_mk_kn_mn(const Tensor<AType>& a_m_k,
                         const Tensor<BType>& b_k_n,
                         Tensor<CType>& c_m_n,
                         const AElementwiseOperation& a_element_op,
                         const BElementwiseOperation& b_element_op,
                         const CElementwiseOperation& c_element_op)
 {
     auto f_mk_kn_mn = [&](auto m, auto n) {
         const int K = a_m_k.mDesc.GetLengths()[1];
  
         float v_acc = 0;
  
         for(int k = 0; k < K; ++k)
         {
             float v_a;
             float v_b;
  
             a_element_op(v_a, static_cast<const float>(a_m_k(m, k)));
             b_element_op(v_b, static_cast<const float>(b_k_n(k, n)));
  
             v_acc += v_a * v_b;
         }
  
         float v_c;
  
         c_element_op(v_c, v_acc);
  
         c_m_n(m, n) = v_c;
     };
  
     make_ParallelTensorFunctor(f_mk_kn_mn,
                                c_m_n.mDesc.GetLengths()[0],
                                c_m_n.mDesc.GetLengths()[1])(std::thread::hardware_concurrency());
 }