/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/3643/include/ck/library/utility/host_tensor.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/3643/include/ck/library/utility/host_tensor.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/3643/include/ck/library/utility/host_tensor.hpp Source File
Go to the documentation of this file.
 // Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 // SPDX-License-Identifier: MIT
  
 #pragma once
  
 #include <algorithm>
 #include <cassert>
 #include <iostream>
 #include <fstream>
 #include <numeric>
 #include <random>
 #include <thread>
 #include <utility>
 #include <vector>
  
 #include "ck/utility/data_type.hpp"
 #include "ck/utility/span.hpp"
 #include "ck/utility/type_convert.hpp"
  
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/ranges.hpp"
 #include "ck/library/utility/thread.hpp"
  
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
  
 namespace ck {
  
 template <typename Range>
 std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
 {
     bool first = true;
     for(auto&& v : range)
     {
         if(first)
             first = false;
         else
             os << delim;
         os << v;
     }
     return os;
 }
  
 template <typename T, typename Range>
 std::ostream& LogRangeAsType(std::ostream& os, Range&& range, std::string delim)
 {
     bool first = true;
     for(auto&& v : range)
     {
         if(first)
             first = false;
         else
             os << delim;
  
         using RangeType = ck::remove_cvref_t<decltype(v)>;
         if constexpr(std::is_same_v<RangeType, ck::f8_t> || std::is_same_v<RangeType, ck::bf8_t> ||
                      std::is_same_v<RangeType, ck::bhalf_t>)
         {
             os << ck::type_convert<float>(v);
         }
         else if constexpr(std::is_same_v<RangeType, ck::pk_i4_t> ||
                           std::is_same_v<RangeType, ck::f4x2_pk_t>)
         {
             const auto packed_floats = ck::type_convert<ck::float2_t>(v);
             const ck::vector_type<float, 2> vector_of_floats{packed_floats};
             os << vector_of_floats.template AsType<float>()[ck::Number<0>{}] << delim
                << vector_of_floats.template AsType<float>()[ck::Number<1>{}];
         }
         else
         {
             os << static_cast<T>(v);
         }
     }
     return os;
 }
  
 template <typename F, typename T, std::size_t... Is>
 auto call_f_unpack_args_impl(F f, T args, std::index_sequence<Is...>)
 {
     return f(std::get<Is>(args)...);
 }
  
 template <typename F, typename T>
 auto call_f_unpack_args(F f, T args)
 {
     constexpr std::size_t N = std::tuple_size<T>{};
  
     return call_f_unpack_args_impl(f, args, std::make_index_sequence<N>{});
 }
  
 template <typename F, typename T, std::size_t... Is>
 auto construct_f_unpack_args_impl(T args, std::index_sequence<Is...>)
 {
     return F(std::get<Is>(args)...);
 }
  
 template <typename F, typename T>
 auto construct_f_unpack_args(F, T args)
 {
     constexpr std::size_t N = std::tuple_size<T>{};
  
     return construct_f_unpack_args_impl<F>(args, std::make_index_sequence<N>{});
 }
  
 struct HostTensorDescriptor
 {
     using BaseTensorLayout = ck::tensor_layout::BaseTensorLayout;
     using DefaultLayout    = BaseTensorLayout;
  
     // Runtime tag describing which layout is picked when layout is not specified explicitly at
     // construction time.
     enum class ChosenLayout
     {
         Original,
         RowMajor,
         ColumnMajor
     };
  
     // Master constructor
     template <typename Layout>
     HostTensorDescriptor(std::vector<std::size_t> lens,
                          std::vector<std::size_t> strides,
                          const Layout& layout = DefaultLayout())
         : mLens(std::move(lens)), mStrides(std::move(strides))
     {
         // To support legacy use cases, when layout is not passed in
         const auto new_layout = HandleDefaultLayout(layout);
         if(dbg)
         {
             std::cout << "Original Lens: [";
             LogRange(std::cout, mLens, ", ") << "] and Strides: [";
             LogRange(std::cout, mStrides, ", ") << "]" << std::endl;
             std::cout << "Layout: " << layout << " --> " << new_layout << std::endl;
         }
  
         // Handling the strides and validation based on the chosen layout
         DispatchChosenLayout(new_layout, layout, [&](auto selected_layout) {
             this->CalculateStrides(selected_layout);
             this->ValidateStrides(selected_layout);
         });
     }
  
     HostTensorDescriptor() : HostTensorDescriptor({}, {}, DefaultLayout()){};
  
     // Helper that invokes a callable with a concrete layout object whose type
     // matches the chosen tag (so template code depending on the layout type
     // can still leverage if constexpr branches).
     template <typename F, typename OrigLayout>
     void DispatchChosenLayout(ChosenLayout tag, const OrigLayout& orig, F&& f) const
     {
         switch(tag)
         {
         case ChosenLayout::RowMajor: f(ck::tensor_layout::gemm::RowMajor{}); break;
         case ChosenLayout::ColumnMajor: f(ck::tensor_layout::gemm::ColumnMajor{}); break;
         case ChosenLayout::Original:
         default: f(orig); break;
         }
     }
  
     template <typename Layout>
     ChosenLayout HandleDefaultLayout(const Layout&)
     {
         if constexpr(!std::is_same_v<Layout, DefaultLayout>)
         {
             return ChosenLayout::Original;
         }
         else
         {
             if(mStrides.empty())
             {
                 // No strides provided -> assume RowMajor
                 return ChosenLayout::RowMajor;
             }
  
             const auto rank = mLens.size();
  
             if(rank > 2)
             {
                 // Keep as-is - validation will warn/throw later
                 return ChosenLayout::Original;
             }
  
             if(rank == 0)
             {
                 // Keep as-is - validation will warn/throw later
                 return ChosenLayout::Original;
             }
  
             if(rank == 1)
             {
                 // Treat 1D tensor as RowMajor
                 return ChosenLayout::RowMajor;
             }
  
             // rank == 2
             if(mStrides.size() == 2)
             {
                 // RowMajor pattern (?, 1)
                 if(mStrides[1] == 1)
                 {
                     return ChosenLayout::RowMajor;
                 }
  
                 // ColumnMajor pattern (1, ?)
                 if(mStrides[0] == 1)
                 {
                     return ChosenLayout::ColumnMajor;
                 }
             }
  
             // Fallback: leave as-is
             return ChosenLayout::Original;
         }
     }
  
     template <typename Layout>
     void CalculateStrides(const Layout& layout)
     {
         if constexpr(std::is_same_v<Layout, ck::tensor_layout::BypassLayoutVerification>)
             return;
         // This is a workaround if the original stride value is -1 (which means "unknown") has been
         // passed in and casted to size_t (unsigned).
         auto strides_int = AsInt(mStrides);
  
         // case of empty strides or all-zero: auto-calculate based on layout and tensor dimensions
         if(mStrides.empty() || std::all_of(strides_int.begin(), strides_int.end(), [](int stride) {
                return stride <= 0;
            }))
         {
  
             if constexpr(!(std::is_same_v<ck::tensor_layout::gemm::RowMajor, Layout> ||
                            std::is_same_v<ck::tensor_layout::gemm::ColumnMajor, Layout>))
             {
                 if(dbg)
                 {
                     std::cerr << "Only RowMajor and ColumnMajor layouts are supported for empty "
                                  "strides, got "
                               << layout << ". Will calculate strides as RowMajor." << std::endl;
                 }
             }
  
             mStrides.clear();
             mStrides.resize(mLens.size(), 0);
             if(mStrides.empty())
                 return;
  
             mStrides.back() = 1;
             std::partial_sum(mLens.rbegin(),
                              mLens.rend() - 1,
                              mStrides.rbegin() + 1,
                              std::multiplies<std::size_t>());
  
             if constexpr(std::is_same_v<ck::tensor_layout::gemm::ColumnMajor, Layout>)
             {
                 // swap the last two strides
                 if(mStrides.size() >= 2)
                     std::swap(mStrides[mStrides.size() - 1], mStrides[mStrides.size() - 2]);
             }
         }
         // The other case is if one of the strides is unknown
         // Currently, only GEMM RowMajor and ColumnMajor layouts are supported and only in the lower
         // two dimensions, e.g. {..., 0, N} or {..., M, 0}. The higher dimensions are left
         // untouched.
         else if constexpr(std::is_same_v<ck::tensor_layout::gemm::RowMajor, Layout> ||
                           std::is_same_v<ck::tensor_layout::gemm::ColumnMajor, Layout>)
         {
             auto rank = mStrides.size();
             if(mLens.size() >= 2 && rank >= 2)
             {
                 const auto inner_idx =
                     std::is_same_v<ck::tensor_layout::gemm::RowMajor, Layout> ? rank - 1 : rank - 2;
                 const auto outer_idx = inner_idx == rank - 1 ? rank - 2 : rank - 1;
                 if(mStrides[inner_idx] <= 0)
                 {
                     mStrides[inner_idx] = 1;
                 }
                 if(mStrides[outer_idx] <= 0)
                 {
                     mStrides[outer_idx] = mLens[inner_idx] * mStrides[inner_idx];
                 }
             }
         }
     }
  
     template <typename Layout>
     void ValidateStrides(const Layout& layout) const
     {
         if constexpr(std::is_same_v<ck::tensor_layout::BypassLayoutVerification, Layout>)
         {
             return;
         }
  
         if(mLens.empty())
         {
             throw std::runtime_error(
                 "HostTensorDescriptor::ValidateStrides: empty tensor dimensions is not allowed.");
         }
  
         const int rank = mLens.size();
         if(rank == 1) // skip any 1D tensors
         {
             return;
         }
  
         if constexpr(std::is_same_v<ck::tensor_layout::BaseTensorLayout, Layout>)
         {
             // Any legacy code that doesn't pass layout to HostTensorDescriptor ctor will
             // hit this case (unless it is a special case - see `HandleDefaultLayout`).
             throw std::runtime_error("HostTensorDescriptor::ValidateStrides: Abstract tensor "
                                      "layout BaseTensorLayout can't be verified. Pls "
                                      "pass specific tensor layout to HostTensorDescriptor (or "
                                      "ck::tensor_layout::BypassLayoutVerification)");
         }
  
         // GEMM cases
         if constexpr(std::is_base_of_v<ck::tensor_layout::gemm::BaseGemmLayout, Layout>)
         {
             if(mLens.size() != mStrides.size())
             {
                 std::ostringstream oss;
                 oss << "HostTensorDescriptor::ValidateStrides: mismatch between tensor rank and "
                        "size of strides: "
                     << *this;
                 throw std::runtime_error(oss.str());
             }
  
             // in GEMM, strides must be all positive or all zeros (auto-derived from tensor
             // dimensions)
             auto strides_int = AsInt(mStrides);
             if(std::any_of(
                    strides_int.begin(), strides_int.end(), [](int stride) { return stride <= 0; }))
             {
                 std::ostringstream oss;
                 oss << "Stride values must be positive or all-zeros (auto-derived from tensor "
                        "dimensions). Instead got ";
                 std::copy(
                     strides_int.begin(), strides_int.end(), std::ostream_iterator<int>(oss, " "));
                 throw std::runtime_error(oss.str());
             }
  
             if constexpr(std::is_same_v<ck::tensor_layout::gemm::RowMajor, Layout> ||
                          std::is_same_v<ck::tensor_layout::gemm::ColumnMajor, Layout>)
             {
                 // The logic here assumes the GEMM with tensor of more than 2 dims, will always have
                 // HW dimesnsions as the inner ones e.g. batched GEMM is either BHW or BWH
                 const auto inner_idx =
                     std::is_same_v<ck::tensor_layout::gemm::RowMajor, Layout> ? rank - 1 : rank - 2;
                 const auto outer_idx = inner_idx == rank - 1 ? rank - 2 : rank - 1;
  
                 if(mStrides[outer_idx] < mLens[inner_idx] * mStrides[inner_idx])
                 {
                     std::ostringstream oss;
                     oss << "Invalid strides for " << layout << ": " << *this;
                     throw std::runtime_error(oss.str());
                 }
  
                 // For higher dimensions, validate strides assuming RowMajor
                 for(int i = 1; i < rank - 2; ++i)
                 {
                     if(mStrides[i - 1] < mStrides[i] * mLens[i])
                     {
                         std::ostringstream oss;
                         oss << "Invalid strides for higher dimensions in " << layout << ": "
                             << *this;
                         throw std::runtime_error(oss.str());
                     }
                 }
             }
             else
             {
                 std::ostringstream oss;
                 oss << "Error: Unsupported GEMM layout: " << layout;
                 throw std::runtime_error(oss.str());
             }
         }
         // Convolution cases
         else if constexpr(std::is_base_of_v<ck::tensor_layout::convolution::BaseConvolutionLayout,
                                             Layout>)
         {
             // TBD: implement verification for Conv layouts
             // For now, just print warning and return
             if(dbg)
             {
  
                 std::cerr
                     << "Warning: Tensor layout verification for ck::tensor_layout::convolution "
                        "layouts is not supported yet. Skipping..."
                     << std::endl;
             }
             return;
         }
         else
         {
             std::ostringstream oss;
             oss << "Error: Tensor layout verification for " << layout << " is not supported yet.";
             throw std::runtime_error(oss.str());
         }
     }
  
     template <typename X,
               typename Layout = DefaultLayout,
               typename        = std::enable_if_t<std::is_convertible_v<X, std::size_t> &&
                                                  std::is_convertible_v<Layout, BaseTensorLayout>>>
     HostTensorDescriptor(const std::initializer_list<X>& lens, const Layout& layout = Layout{})
         : HostTensorDescriptor(std::vector<std::size_t>(lens.begin(), lens.end()), {}, layout)
     {
         if(dbg)
             std::cout << "HostTensorDescriptor ctor (" << __LINE__ << ")" << std::endl;
     }
  
     template <typename Layout = DefaultLayout,
               typename        = std::enable_if_t<std::is_convertible_v<Layout, BaseTensorLayout>>>
     HostTensorDescriptor(const std::initializer_list<ck::long_index_t>& lens,
                          const Layout& layout = Layout{})
         : HostTensorDescriptor(std::vector<std::size_t>(lens.begin(), lens.end()), {}, layout)
     {
         if(dbg)
             std::cout << "HostTensorDescriptor ctor (" << __LINE__ << ")" << std::endl;
     }
  
     template <typename Lengths,
               typename Layout = DefaultLayout,
               typename        = std::enable_if_t<
                          (std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> ||
                    std::is_convertible_v<ck::ranges::range_value_t<Lengths>, ck::long_index_t>) &&
                          std::is_convertible_v<Layout, BaseTensorLayout>>>
     HostTensorDescriptor(const Lengths& lens, const Layout& layout = Layout{})
         : HostTensorDescriptor(std::vector<std::size_t>(lens.begin(), lens.end()), {}, layout)
     {
         if(dbg)
             std::cout << "HostTensorDescriptor ctor (" << __LINE__ << ")" << std::endl;
     }
  
     template <typename X,
               typename Y,
               typename        = std::enable_if_t<std::is_convertible_v<X, std::size_t> &&
                                                  std::is_convertible_v<Y, std::size_t>>,
               typename Layout = DefaultLayout>
     HostTensorDescriptor(const std::initializer_list<X>& lens,
                          const std::initializer_list<Y>& strides,
                          const Layout& layout = Layout{})
         : HostTensorDescriptor(std::vector<std::size_t>(lens.begin(), lens.end()),
                                std::vector<std::size_t>(strides.begin(), strides.end()),
                                layout)
     {
         if(dbg)
             std::cout << "HostTensorDescriptor ctor (" << __LINE__ << ")" << std::endl;
     }
  
     // HostTensorDescriptor({row, col}, {row_stride, col_stride})
     template <typename Layout = DefaultLayout>
     HostTensorDescriptor(const std::initializer_list<ck::long_index_t>& lens,
                          const std::initializer_list<ck::long_index_t>& strides,
                          const Layout& layout = Layout{})
         : HostTensorDescriptor(std::vector<std::size_t>(lens.begin(), lens.end()),
                                std::vector<std::size_t>(strides.begin(), strides.end()),
                                layout)
     {
         if(dbg)
             std::cout << "HostTensorDescriptor ctor (" << __LINE__ << ")" << std::endl;
     }
  
     // HostTensorDescriptor({row, col}, strides)
     template <typename Strides, typename Layout = DefaultLayout>
     HostTensorDescriptor(const std::initializer_list<std::size_t>& lens,
                          const Strides& strides,
                          const Layout& layout = Layout{})
         : HostTensorDescriptor(std::vector<std::size_t>(lens.begin(), lens.end()),
                                std::vector<std::size_t>(strides.begin(), strides.end()),
                                layout)
     {
         if(dbg)
             std::cout << "HostTensorDescriptor ctor (" << __LINE__ << ")" << std::endl;
     }
  
     template <typename Lengths,
               typename Strides,
               typename Layout = DefaultLayout,
               typename        = std::enable_if_t<
                          ((std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> &&
                     std::is_convertible_v<ck::ranges::range_value_t<Strides>, std::size_t>) ||
                    (std::is_convertible_v<ck::ranges::range_value_t<Lengths>, ck::long_index_t> &&
                     std::is_convertible_v<ck::ranges::range_value_t<Strides>, ck::long_index_t>)) &&
                          std::is_convertible_v<Layout, BaseTensorLayout>>>
     HostTensorDescriptor(const Lengths& lens,
                          const Strides& strides,
                          const Layout& layout = Layout{})
         : HostTensorDescriptor(std::vector<std::size_t>(lens.begin(), lens.end()),
                                std::vector<std::size_t>(strides.begin(), strides.end()),
                                layout)
     {
         if(dbg)
             std::cout << "HostTensorDescriptor ctor (" << __LINE__ << ")" << std::endl;
     }
  
     std::size_t GetNumOfDimension() const;
     std::size_t GetElementSize() const;
     std::size_t GetElementSpaceSize() const;
  
     const std::vector<std::size_t>& GetLengths() const;
     const std::vector<std::size_t>& GetStrides() const;
  
     template <typename... Is>
     std::size_t GetOffsetFromMultiIndex(Is... is) const
     {
         assert(sizeof...(Is) == this->GetNumOfDimension());
         std::initializer_list<std::size_t> iss{static_cast<std::size_t>(is)...};
         return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
     }
  
     std::size_t GetOffsetFromMultiIndex(const std::vector<std::size_t>& iss) const
     {
         return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
     }
  
     friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc);
     friend std::ostream& operator<<(std::ostream& os, ChosenLayout tag);
  
     private:
     std::vector<std::size_t> mLens;
     std::vector<std::size_t> mStrides;
     static constexpr bool dbg = false;
  
     std::vector<int> AsInt(const std::vector<size_t>& vec) const
     {
         std::vector<int> strides_int(vec.size());
         std::transform(vec.begin(), vec.end(), strides_int.begin(), [](std::size_t stride) {
             return static_cast<int>(stride);
         });
         return strides_int;
     }
 };
  
 template <typename New2Old, typename NewLayout = HostTensorDescriptor::BaseTensorLayout>
 HostTensorDescriptor
 transpose_host_tensor_descriptor_given_new2old(const HostTensorDescriptor& a,
                                                const New2Old& new2old,
                                                const NewLayout& new_layout = NewLayout())
 {
     std::vector<std::size_t> new_lengths(a.GetNumOfDimension());
     std::vector<std::size_t> new_strides(a.GetNumOfDimension());
  
     for(std::size_t i = 0; i < a.GetNumOfDimension(); i++)
     {
         new_lengths[i] = a.GetLengths()[new2old[i]];
         new_strides[i] = a.GetStrides()[new2old[i]];
     }
  
     return HostTensorDescriptor(new_lengths, new_strides, new_layout);
 }
  
 struct joinable_thread : std::thread
 {
     template <typename... Xs>
     joinable_thread(Xs&&... xs) : std::thread(std::forward<Xs>(xs)...)
     {
     }
  
     joinable_thread(joinable_thread&&)            = default;
     joinable_thread& operator=(joinable_thread&&) = default;
  
     ~joinable_thread()
     {
         if(this->joinable())
             this->join();
     }
 };
  
 template <typename F, typename... Xs>
 struct ParallelTensorFunctor
 {
     F mF;
     static constexpr std::size_t NDIM = sizeof...(Xs);
     std::array<std::size_t, NDIM> mLens;
     std::array<std::size_t, NDIM> mStrides;
     std::size_t mN1d;
  
     ParallelTensorFunctor(F f, Xs... xs) : mF(f), mLens({static_cast<std::size_t>(xs)...})
     {
         mStrides.back() = 1;
         std::partial_sum(mLens.rbegin(),
                          mLens.rend() - 1,
                          mStrides.rbegin() + 1,
                          std::multiplies<std::size_t>());
         mN1d = mStrides[0] * mLens[0];
     }
  
     std::array<std::size_t, NDIM> GetNdIndices(std::size_t i) const
     {
         std::array<std::size_t, NDIM> indices;
  
         for(std::size_t idim = 0; idim < NDIM; ++idim)
         {
             indices[idim] = i / mStrides[idim];
             i -= indices[idim] * mStrides[idim];
         }
  
         return indices;
     }
  
     void operator()(std::size_t num_thread = 1) const
     {
         std::size_t work_per_thread = (mN1d + num_thread - 1) / num_thread;
  
         std::vector<joinable_thread> threads(num_thread);
  
         for(std::size_t it = 0; it < num_thread; ++it)
         {
             std::size_t iw_begin = it * work_per_thread;
             std::size_t iw_end   = std::min((it + 1) * work_per_thread, mN1d);
  
             auto f = [=, *this] {
                 for(std::size_t iw = iw_begin; iw < iw_end; ++iw)
                 {
                     call_f_unpack_args(mF, GetNdIndices(iw));
                 }
             };
             threads[it] = joinable_thread(f);
         }
     }
 };
  
 template <typename F, typename... Xs>
 auto make_ParallelTensorFunctor(F f, Xs... xs)
 {
     return ParallelTensorFunctor<F, Xs...>(f, xs...);
 }
  
 template <typename T>
 struct Tensor
 {
     using Descriptor = HostTensorDescriptor;
     using Data       = std::vector<T>;
  
     template <typename X>
     Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(GetElementSpaceSize())
     {
     }
  
     template <typename X, typename Y>
     Tensor(std::initializer_list<X> lens, std::initializer_list<Y> strides)
         : mDesc(lens, strides), mData(GetElementSpaceSize())
     {
     }
  
     template <typename Lengths>
     Tensor(const Lengths& lens) : mDesc(lens), mData(GetElementSpaceSize())
     {
     }
  
     template <typename Lengths, typename Strides>
     Tensor(const Lengths& lens, const Strides& strides)
         : mDesc(lens, strides), mData(GetElementSpaceSize())
     {
     }
  
     template <typename X, typename... Rest, std::enable_if_t<(sizeof...(Rest) > 0), int> = 0>
     Tensor(std::initializer_list<X> lens, Rest&&... rest)
         : mDesc(lens, std::forward<Rest>(rest)...), mData(GetElementSpaceSize())
     {
     }
  
     template <typename X,
               typename Y,
               typename... Rest,
               std::enable_if_t<(sizeof...(Rest) > 0), int> = 0>
     Tensor(std::initializer_list<X> lens, std::initializer_list<Y> strides, Rest&&... rest)
         : mDesc(lens, strides, std::forward<Rest>(rest)...), mData(GetElementSpaceSize())
     {
     }
  
     template <typename Lengths, typename... Rest, std::enable_if_t<(sizeof...(Rest) > 0), int> = 0>
     Tensor(const Lengths& lens, Rest&&... rest)
         : mDesc(lens, std::forward<Rest>(rest)...), mData(GetElementSpaceSize())
     {
     }
  
     template <typename Lengths,
               typename Strides,
               typename... Rest,
               std::enable_if_t<(sizeof...(Rest) > 0), int> = 0>
     Tensor(const Lengths& lens, const Strides& strides, Rest&&... rest)
         : mDesc(lens, strides, std::forward<Rest>(rest)...), mData(GetElementSpaceSize())
     {
     }
  
     Tensor(const Descriptor& desc) : mDesc(desc), mData(GetElementSpaceSize()) {}
  
     template <typename OutT>
     Tensor<OutT> CopyAsType() const
     {
         Tensor<OutT> ret(mDesc);
  
         ck::ranges::transform(
             mData, ret.mData.begin(), [](auto value) { return ck::type_convert<OutT>(value); });
  
         return ret;
     }
  
     Tensor()              = delete;
     Tensor(const Tensor&) = default;
     Tensor(Tensor&&)      = default;
  
     ~Tensor() = default;
  
     Tensor& operator=(const Tensor&) = default;
     Tensor& operator=(Tensor&&)      = default;
  
     template <typename FromT>
     explicit Tensor(const Tensor<FromT>& other) : Tensor(other.template CopyAsType<T>())
     {
     }
     void savetxt(std::string file_name, std::string dtype = "float")
     {
         std::ofstream file(file_name);
  
         if(file.is_open())
         {
             for(auto& itm : mData)
             {
                 if(dtype == "float")
                     file << ck::type_convert<float>(itm) << std::endl;
                 else if(dtype == "int")
                     file << ck::type_convert<int>(itm) << std::endl;
                 else
                     // TODO: we didn't implement operator<< for all custom
                     // data types, here fall back to float in case compile error
                     file << ck::type_convert<float>(itm) << std::endl;
             }
             file.close();
         }
         else
         {
             // Print an error message to the standard error
             // stream if the file cannot be opened.
             throw std::runtime_error(std::string("unable to open file:") + file_name);
         }
     }
     decltype(auto) GetLengths() const { return mDesc.GetLengths(); }
  
     decltype(auto) GetStrides() const { return mDesc.GetStrides(); }
  
     std::size_t GetNumOfDimension() const { return mDesc.GetNumOfDimension(); }
  
     std::size_t GetElementSize() const { return mDesc.GetElementSize(); }
  
     std::size_t GetElementSpaceSize() const
     {
         if constexpr(ck::is_packed_type_v<ck::remove_cvref_t<T>>)
         {
             return (mDesc.GetElementSpaceSize() + 1) / ck::packed_size_v<ck::remove_cvref_t<T>>;
         }
         else
         {
             return mDesc.GetElementSpaceSize();
         }
     }
  
     std::size_t GetElementSpaceSizeInBytes() const { return sizeof(T) * GetElementSpaceSize(); }
  
     void SetZero() { ck::ranges::fill<T>(mData, T{0}); }
  
     template <typename F>
     void ForEach_impl(F&& f, std::vector<size_t>& idx, size_t rank)
     {
         if(rank == mDesc.GetNumOfDimension())
         {
             f(*this, idx);
             return;
         }
         // else
         for(size_t i = 0; i < mDesc.GetLengths()[rank]; i++)
         {
             idx[rank] = i;
             ForEach_impl(std::forward<F>(f), idx, rank + 1);
         }
     }
  
     template <typename F>
     void ForEach(F&& f)
     {
         std::vector<size_t> idx(mDesc.GetNumOfDimension(), 0);
         ForEach_impl(std::forward<F>(f), idx, size_t(0));
     }
  
     template <typename F>
     void ForEach_impl(const F&& f, std::vector<size_t>& idx, size_t rank) const
     {
         if(rank == mDesc.GetNumOfDimension())
         {
             f(*this, idx);
             return;
         }
         // else
         for(size_t i = 0; i < mDesc.GetLengths()[rank]; i++)
         {
             idx[rank] = i;
             ForEach_impl(std::forward<const F>(f), idx, rank + 1);
         }
     }
  
     template <typename F>
     void ForEach(const F&& f) const
     {
         std::vector<size_t> idx(mDesc.GetNumOfDimension(), 0);
         ForEach_impl(std::forward<const F>(f), idx, size_t(0));
     }
  
     template <typename G>
     void GenerateTensorValue(G g, std::size_t num_thread = 1)
     {
         switch(mDesc.GetNumOfDimension())
         {
         case 1: {
             auto f = [&](auto i) { (*this)(i) = g(i); };
             make_ParallelTensorFunctor(f, mDesc.GetLengths()[0])(num_thread);
             break;
         }
         case 2: {
             auto f = [&](auto i0, auto i1) { (*this)(i0, i1) = g(i0, i1); };
             make_ParallelTensorFunctor(f, mDesc.GetLengths()[0], mDesc.GetLengths()[1])(num_thread);
             break;
         }
         case 3: {
             auto f = [&](auto i0, auto i1, auto i2) { (*this)(i0, i1, i2) = g(i0, i1, i2); };
             make_ParallelTensorFunctor(
                 f, mDesc.GetLengths()[0], mDesc.GetLengths()[1], mDesc.GetLengths()[2])(num_thread);
             break;
         }
         case 4: {
             auto f = [&](auto i0, auto i1, auto i2, auto i3) {
                 (*this)(i0, i1, i2, i3) = g(i0, i1, i2, i3);
             };
             make_ParallelTensorFunctor(f,
                                        mDesc.GetLengths()[0],
                                        mDesc.GetLengths()[1],
                                        mDesc.GetLengths()[2],
                                        mDesc.GetLengths()[3])(num_thread);
             break;
         }
         case 5: {
             auto f = [&](auto i0, auto i1, auto i2, auto i3, auto i4) {
                 (*this)(i0, i1, i2, i3, i4) = g(i0, i1, i2, i3, i4);
             };
             make_ParallelTensorFunctor(f,
                                        mDesc.GetLengths()[0],
                                        mDesc.GetLengths()[1],
                                        mDesc.GetLengths()[2],
                                        mDesc.GetLengths()[3],
                                        mDesc.GetLengths()[4])(num_thread);
             break;
         }
         case 6: {
             auto f = [&](auto i0, auto i1, auto i2, auto i3, auto i4, auto i5) {
                 (*this)(i0, i1, i2, i3, i4, i5) = g(i0, i1, i2, i3, i4, i5);
             };
             make_ParallelTensorFunctor(f,
                                        mDesc.GetLengths()[0],
                                        mDesc.GetLengths()[1],
                                        mDesc.GetLengths()[2],
                                        mDesc.GetLengths()[3],
                                        mDesc.GetLengths()[4],
                                        mDesc.GetLengths()[5])(num_thread);
             break;
         }
         case 12: {
             auto f = [&](auto i0,
                          auto i1,
                          auto i2,
                          auto i3,
                          auto i4,
                          auto i5,
                          auto i6,
                          auto i7,
                          auto i8,
                          auto i9,
                          auto i10,
                          auto i11) {
                 (*this)(i0, i1, i2, i3, i4, i5, i6, i7, i8, i9, i10, i11) =
                     g(i0, i1, i2, i3, i4, i5, i6, i7, i8, i9, i10, i11);
             };
             make_ParallelTensorFunctor(f,
                                        mDesc.GetLengths()[0],
                                        mDesc.GetLengths()[1],
                                        mDesc.GetLengths()[2],
                                        mDesc.GetLengths()[3],
                                        mDesc.GetLengths()[4],
                                        mDesc.GetLengths()[5],
                                        mDesc.GetLengths()[6],
                                        mDesc.GetLengths()[7],
                                        mDesc.GetLengths()[8],
                                        mDesc.GetLengths()[9],
                                        mDesc.GetLengths()[10],
                                        mDesc.GetLengths()[11])(num_thread);
             break;
         }
         default: throw std::runtime_error("unspported dimension");
         }
     }
  
     // Generate random values with multiple threads. Guaranteed to give the same sequence with any
     // number of threads provided.
     template <typename Distribution = std::uniform_real_distribution<float>,
               typename Mapping      = ck::identity,
               typename Generator    = std::minstd_rand>
     void GenerateTensorDistr(Distribution dis       = {0.f, 1.f},
                              Mapping fn             = {},
                              const Generator g      = Generator(0), // default seed 0
                              std::size_t num_thread = -1)
     {
         using ck::math::integer_divide_ceil;
         using ck::math::min;
         if(num_thread == -1ULL)
             num_thread = min(ck::get_available_cpu_cores(), 80U); // max 80 threads
         // At least 2MB per thread
         num_thread = min(num_thread, integer_divide_ceil(this->GetElementSpaceSize(), 0x200000));
         constexpr std::size_t BLOCK_BYTES = 64;
         constexpr std::size_t BLOCK_SIZE  = BLOCK_BYTES / sizeof(T);
  
         const std::size_t num_blocks = integer_divide_ceil(this->GetElementSpaceSize(), BLOCK_SIZE);
         const std::size_t blocks_per_thread = integer_divide_ceil(num_blocks, num_thread);
  
         std::vector<std::thread> threads;
         threads.reserve(num_thread - 1);
         const auto dst                = const_cast<T*>(this->mData.data());
         const auto element_space_size = this->GetElementSpaceSize();
         for(int it = num_thread - 1; it >= 0; --it)
         {
             std::size_t ib_begin = it * blocks_per_thread;
             std::size_t ib_end   = min(ib_begin + blocks_per_thread, num_blocks);
  
             auto job = [=]() {
                 auto g_   = g;   // copy
                 auto dis_ = dis; // copy
                 g_.discard(ib_begin * BLOCK_SIZE * ck::packed_size_v<T>);
                 auto t_fn = [&]() {
                     // As user can pass integer distribution in dis, we must ensure that the correct
                     // constructor/converter is called at all times. For f4/f6/f8 types, to ensure
                     // correct results, we convert from float to the target type. In these cases
                     // integer constructors are interpreted as direct initialization of the internal
                     // storage with binary values instead of treating integers as subset of floats.
                     if constexpr(ck::is_same_v<T, ck::f8_t> || ck::is_same_v<T, ck::bf8_t>)
                         return ck::type_convert<T>(static_cast<float>(fn(dis_(g_))));
                     else if constexpr(ck::packed_size_v<T> == 1)
                         return ck::type_convert<T>(fn(dis_(g_)));
                     else if constexpr(ck::is_same_v<T, ck::f4x2_pk_t>)
                         return ck::f4x2_pk_t{ck::type_convert<ck::f4x2_t>(
                             ck::float2_t{ck::type_convert<float>(fn(dis_(g_))),
                                          ck::type_convert<float>(fn(dis_(g_)))})};
                     else if constexpr(ck::is_same_v<T, ck::f6x32_pk_t> ||
                                       ck::is_same_v<T, ck::bf6x32_pk_t>)
                     {
                         return ck::type_convert<T>(
                             ck::float32_t{ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_)))});
                     }
                     else if constexpr(ck::is_same_v<T, ck::f6x16_pk_t> ||
                                       ck::is_same_v<T, ck::bf6x16_pk_t>)
                     {
                         return ck::type_convert<T>(
                             ck::float16_t{ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_))),
                                           ck::type_convert<float>(fn(dis_(g_)))});
                     }
                     else
                         static_assert(false, "Unsupported packed size for T");
                 };
  
                 std::size_t ib = ib_begin;
                 for(; ib < ib_end - 1; ++ib)
                     ck::static_for<0, BLOCK_SIZE, 1>{}([&](auto iw_) {
                         constexpr size_t iw       = iw_.value;
                         dst[ib * BLOCK_SIZE + iw] = t_fn();
                     });
                 for(std::size_t iw = 0; iw < BLOCK_SIZE; ++iw)
                     if(ib * BLOCK_SIZE + iw < element_space_size)
                         dst[ib * BLOCK_SIZE + iw] = t_fn();
             };
  
             if(it > 0)
                 threads.emplace_back(std::move(job));
             else
                 job(); // last job run in the main thread
         }
         for(auto& t : threads)
             t.join();
     }
  
     template <typename... Is>
     std::size_t GetOffsetFromMultiIndex(Is... is) const
     {
         return mDesc.GetOffsetFromMultiIndex(is...) / ck::packed_size_v<ck::remove_cvref_t<T>>;
     }
  
     template <typename... Is>
     T& operator()(Is... is)
     {
         return mData[mDesc.GetOffsetFromMultiIndex(is...) /
                      ck::packed_size_v<ck::remove_cvref_t<T>>];
     }
  
     template <typename... Is>
     const T& operator()(Is... is) const
     {
         return mData[mDesc.GetOffsetFromMultiIndex(is...) /
                      ck::packed_size_v<ck::remove_cvref_t<T>>];
     }
  
     T& operator()(const std::vector<std::size_t>& idx)
     {
         return mData[mDesc.GetOffsetFromMultiIndex(idx) / ck::packed_size_v<ck::remove_cvref_t<T>>];
     }
  
     const T& operator()(const std::vector<std::size_t>& idx) const
     {
         return mData[mDesc.GetOffsetFromMultiIndex(idx) / ck::packed_size_v<ck::remove_cvref_t<T>>];
     }
  
     typename Data::iterator begin() { return mData.begin(); }
  
     typename Data::iterator end() { return mData.end(); }
  
     typename Data::pointer data() { return mData.data(); }
  
     typename Data::const_iterator begin() const { return mData.begin(); }
  
     typename Data::const_iterator end() const { return mData.end(); }
  
     typename Data::const_pointer data() const { return mData.data(); }
  
     typename Data::size_type size() const { return mData.size(); }
  
     template <typename U = T>
     auto AsSpan() const
     {
         constexpr std::size_t FromSize = sizeof(T);
         constexpr std::size_t ToSize   = sizeof(U);
  
         using Element = std::add_const_t<std::remove_reference_t<U>>;
         return ck::span<Element>{reinterpret_cast<Element*>(data()), size() * FromSize / ToSize};
     }
  
     template <typename U = T>
     auto AsSpan()
     {
         constexpr std::size_t FromSize = sizeof(T);
         constexpr std::size_t ToSize   = sizeof(U);
  
         using Element = std::remove_reference_t<U>;
         return ck::span<Element>{reinterpret_cast<Element*>(data()), size() * FromSize / ToSize};
     }
  
     Descriptor mDesc;
     Data mData;
 };
  
 } // namespace ck