/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/host/reference/reference_rowwise_quantization2d.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/host/reference/reference_rowwise_quantization2d.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_rowwise_quantization2d.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"
 #include <thread>
  
 namespace ck_tile {
 template <typename XDataType, typename ScaleDataType, typename QXDataType>
 CK_TILE_HOST void reference_rowwise_quantization2d(const HostTensor<XDataType>& x_m_n,
                                                    const HostTensor<ScaleDataType>& scale_m,
                                                    HostTensor<QXDataType>& qx_m_n)
 {
     auto f = [&](auto m) {
         const int N = x_m_n.mDesc.get_lengths()[1];
  
         for(int n = 0; n < N; ++n)
         {
             auto v_x = x_m_n(m, n);
             // scale = amax / 127 for int8
             auto v_scale = type_convert<XDataType>(scale_m(m));
             auto v_qx    = v_x / v_scale;
             qx_m_n(m, n) = type_convert<QXDataType>(saturates<QXDataType>{}(v_qx));
         }
     };
  
     make_ParallelTensorFunctor(f,
                                scale_m.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
 }
  
 } // namespace ck_tile