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

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

Go to the documentation of this file.

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/host_tensor.hpp"
 #include <thread>
 #include <numeric>
 #include <functional>
 #include <utility>
 #include <algorithm>
  
 namespace ck_tile {
  
 /*
     similiar to torch.topk()
     x (Tensor) – the input tensor.
     k (int) – the k in “top-k”
     dim (int, optional) – the dimension to sort along
     largest (bool, optional) – largest or smallest elements
     sorted (bool, optional) – elements in sorted order or not
  
     output:
     y_values
     y_indices
  
     https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/TopKImpl.h
 */
 template <typename DataType, typename IndexType = index_t>
 CK_TILE_HOST void reference_topk(const HostTensor<DataType>& x,
                                  HostTensor<DataType>& y_values,
                                  HostTensor<IndexType>& y_indices,
                                  index_t k,
                                  index_t dim  = -1,
                                  bool largest = true,
                                  bool sorted  = true)
 {
     // rank must be the same
     index_t rank = x.get_num_of_dimension();
     assert(static_cast<std::size_t>(rank) == y_values.get_num_of_dimension());
     assert(static_cast<size_t>(rank) == y_indices.get_num_of_dimension());
     assert(dim == -1 || dim < rank);
  
     index_t topk_dim     = dim == -1 ? (rank - 1) : dim;
     index_t topk_src_len = x.get_length(topk_dim);
     auto x_len           = x.get_lengths();
  
     assert(k <= topk_src_len);
     assert(static_cast<size_t>(k) == y_values.get_length(topk_dim) &&
            static_cast<size_t>(k) == y_indices.get_length(topk_dim));
  
     index_t n_parallel = x.get_element_size() / topk_src_len;
  
     // clang-format off
     auto f = [&](auto i_element) {
         std::vector<size_t> topk_coord = [&](){
             std::vector<size_t> t_(rank, 0);
             size_t r = i_element;
             for(index_t i = rank - 1; i >= 0; i--) {
                 if(i == topk_dim)          continue; // topk dim should be zero
                 t_[i] = r % x_len[i];      r = r / x_len[i];
             }
             return t_;
         }();
  
         using elem_t = std::pair<DataType, IndexType>;
         std::vector<elem_t> q = [&](){
             std::vector<elem_t> t_(topk_src_len);
             for(index_t i = 0; i < topk_src_len; i++) {
                 auto c_ = topk_coord;  c_[topk_dim] = i;
                 t_[i].first = x(c_);   t_[i].second = i;
             }
             return t_;
         }();
  
         // run topk
         if(largest) {
             std::nth_element(q.begin(), q.begin() + k - 1, q.end(),
             [](const elem_t& lhs, const elem_t& rhs) -> bool { return lhs.first > rhs.first; });
             if(sorted) {
                 std::sort(q.begin(), q.begin() + k - 1,
                 [](const elem_t& lhs, const elem_t& rhs) -> bool { return lhs.first > rhs.first; });
             }
         } else {
             std::nth_element(q.begin(), q.begin() + k - 1, q.end(),
             [](const elem_t& lhs, const elem_t& rhs) -> bool { return lhs.first < rhs.first; });
             if(sorted) {
                 std::sort(q.begin(), q.begin() + k - 1,
                 [](const elem_t& lhs, const elem_t& rhs) -> bool { return lhs.first < rhs.first; });
             }
         }
  
         // write out
         for(index_t i = 0; i < k; i++) {
             auto c_ = topk_coord;  c_[topk_dim] = i;
             y_values(c_) = q[i].first;  y_indices(c_) = q[i].second;
         }
     };
     // clang-format on
  
     make_ParallelTensorFunctor(f, n_parallel)(std::thread::hardware_concurrency());
 }
  
 // TODO: if using this method, the return tensor would be dense(no stride)
 template <typename DataType, typename IndexType = index_t>
 CK_TILE_HOST auto reference_topk(const HostTensor<DataType>& x,
                                  index_t k,
                                  index_t dim  = -1,
                                  bool largest = true,
                                  bool sorted  = true)
 {
     auto lens          = x.get_lengths();
     index_t target_dim = (dim == -1) ? (lens.size() - 1) : dim;
     assert(target_dim < lens.size());
     assert(k <= lens[target_dim]);
     lens[target_dim] = k;
     HostTensor<DataType> y_values(lens);
     HostTensor<IndexType> y_indices(lens);
  
     reference_topk<DataType, IndexType>(x, y_values, y_indices, k, dim, largest, sorted);
  
     return ck_tile::make_tuple(y_values, y_indices);
 }
 } // namespace ck_tile