/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.0/include/ck_tile/core/tensor/buffer_view.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.0/include/ck_tile/core/tensor/buffer_view.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.0/include/ck_tile/core/tensor/buffer_view.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck_tile/core/config.hpp"
 #include "ck_tile/core/arch/arch.hpp"
 #if __clang_major__ == 20
 #include "ck_tile/core/arch/amd_buffer_addressing_builtins.hpp"
 #else
 #include "ck_tile/core/arch/amd_buffer_addressing.hpp"
 #endif
 #include "ck_tile/core/arch/generic_memory_space_atomic.hpp"
 #include "ck_tile/core/container/array.hpp"
 #include "ck_tile/core/numeric/integer.hpp"
 #include "ck_tile/core/numeric/integral_constant.hpp"
 #include "ck_tile/core/numeric/float8.hpp"
 #include "ck_tile/core/numeric/half.hpp"
 #include "ck_tile/core/numeric/bfloat16.hpp"
 #include "ck_tile/core/utility/type_traits.hpp"
 #include "ck_tile/core/utility/ignore.hpp"
  
 namespace ck_tile {
  
 // T may be scalar or vector
 // X may be scalar or vector
 // T and X have same scalar type
 // X contains multiple T
 // FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
 //        transforms of tensor_view/Tensor
 // FIXME: amd_buffer_coherence_enum is only meaningful for buffer addressing. Need to split
 // buffer_view definition for different memory address space (Global/GenericLds/Vgpr)
 template <address_space_enum BufferAddressSpace,
           typename T,
           typename BufferSizeType,
           bool InvalidElementUseNumericalZeroValue,
           amd_buffer_coherence_enum Coherence = amd_buffer_coherence_enum::coherence_default>
 struct buffer_view;
  
 // Address Space: generic
 // T may be scalar or vector
 // X may be scalar or vector
 // T and X have same scalar type
 // X contains multiple T
 // FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
 //        transforms of tensor_view/Tensor
 template <typename T, typename BufferSizeType, bool InvalidElementUseNumericalZeroValue>
 struct buffer_view<address_space_enum::generic,
                    T,
                    BufferSizeType,
                    InvalidElementUseNumericalZeroValue,
                    amd_buffer_coherence_enum::coherence_default>
 {
     using type = T;
  
     T* p_data_ = nullptr;
     BufferSizeType buffer_size_;
     remove_cvref_t<T> invalid_element_value_ = T{0};
  
     CK_TILE_HOST_DEVICE constexpr buffer_view()
         : p_data_{}, buffer_size_{}, invalid_element_value_{}
     {
     }
  
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data, BufferSizeType buffer_size)
         : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{0}
     {
     }
  
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data,
                                               BufferSizeType buffer_size,
                                               T invalid_element_value)
         : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{invalid_element_value}
     {
     }
  
     CK_TILE_HOST_DEVICE void init_raw() {}
  
     CK_TILE_DEVICE static constexpr address_space_enum get_address_space()
     {
         return address_space_enum::generic;
     }
  
     // i is offset of T
     // FIXME: doesn't do is_valid check
     CK_TILE_DEVICE constexpr const T& operator[](index_t i) const { return p_data_[i]; }
  
     // i is offset of T
     // FIXME: doesn't do is_valid check
     CK_TILE_DEVICE constexpr T& operator()(index_t i) { return p_data_[i]; }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto get(index_t i,
                                       index_t linear_offset,
                                       bool is_valid_element,
                                       bool_constant<oob_conditional_check> = {}) const
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         if(is_valid_element)
         {
 #if CK_TILE_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
             X tmp;
  
             __builtin_memcpy(&tmp, &(p_data_[i + linear_offset]), sizeof(X));
  
             return tmp;
 #else
             return *c_style_pointer_cast<const X*>(&p_data_[i + linear_offset]);
 #endif
         }
         else
         {
             if constexpr(InvalidElementUseNumericalZeroValue)
             {
                 return X{numeric<remove_cvref_t<T>>::zero()};
             }
             else
             {
                 return X{invalid_element_value_};
             }
         }
     }
  
     /*
     In the generic address space, we do not support the transpose instruction in the buffer view.
     Will report compilation error when developer wants to use it.
     */
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto transpose_get(index_t i,
                                                 index_t linear_offset,
                                                 bool is_valid_element,
                                                 bool_constant<oob_conditional_check> = {}) const
     {
         static_assert(false, "Error: transpose load not supported in global memory space.");
         ignore = i;
         ignore = linear_offset;
         ignore = is_valid_element;
         return;
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <memory_operation_enum Op,
               typename X,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void update(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         if constexpr(Op == memory_operation_enum::set)
         {
             this->template set<X>(i, linear_offset, is_valid_element, x);
         }
         // FIXME: remove memory_operation_enum::add
         else if constexpr(Op == memory_operation_enum::add)
         {
             auto tmp = this->template get<X>(i, linear_offset, is_valid_element);
             this->template set<X>(i, linear_offset, is_valid_element, x + tmp);
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void set(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         if(is_valid_element)
         {
 #if CK_TILE_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
             X tmp = x;
  
             __builtin_memcpy(&(p_data_[i + linear_offset]), &tmp, sizeof(X));
 #else
             *c_style_pointer_cast<X*>(&p_data_[i + linear_offset]) = x;
 #endif
         }
     }
  
     // FIXME: remove
     CK_TILE_DEVICE static constexpr bool is_static_buffer() { return false; }
  
     // FIXME: remove
     CK_TILE_DEVICE static constexpr bool is_dynamic_buffer() { return true; }
  
     CK_TILE_HOST_DEVICE void print() const
     {
         printf("buffer_view{");
  
         // AddressSpace
         printf("AddressSpace: generic, ");
  
         // p_data_
         printf("p_data_: %p, ", static_cast<void*>(const_cast<remove_cvref_t<T>*>(p_data_)));
  
         // buffer_size_
         printf("buffer_size_: ");
         print(buffer_size_);
         printf(", ");
  
         // invalid_element_value_
         printf("invalid_element_value_: ");
         print(invalid_element_value_);
  
         printf("}");
     }
 };
  
 // Address Space: Global
 // T may be scalar or vector
 // X may be scalar or vector
 // T and X have same scalar type
 // X contains multiple T
 // FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
 //        transforms of tensor_view/Tensor
 template <typename T,
           typename BufferSizeType,
           bool InvalidElementUseNumericalZeroValue,
           amd_buffer_coherence_enum Coherence>
 struct buffer_view<address_space_enum::global,
                    T,
                    BufferSizeType,
                    InvalidElementUseNumericalZeroValue,
                    Coherence>
 {
     using type = T;
  
     T* p_data_ = nullptr;
     BufferSizeType buffer_size_;
     int32x4_t cached_buf_res_;
     remove_cvref_t<T> invalid_element_value_ = T{0};
  
     static constexpr index_t PackedSize = ck_tile::numeric_traits<remove_cvref_t<T>>::PackedSize;
  
     CK_TILE_HOST_DEVICE constexpr buffer_view()
         : p_data_{}, buffer_size_{}, cached_buf_res_{0}, invalid_element_value_{}
     {
     }
  
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data, BufferSizeType buffer_size)
         : p_data_{p_data},
           buffer_size_{buffer_size / PackedSize},
           cached_buf_res_{0},
           invalid_element_value_{0}
     {
     }
  
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data,
                                               BufferSizeType buffer_size,
                                               T invalid_element_value)
         : p_data_{p_data},
           buffer_size_{buffer_size / PackedSize},
           cached_buf_res_{0},
           invalid_element_value_{invalid_element_value}
     {
     }
  
     // this is non constexpr intentially (will call some intrinsic internally)
     // Must call for buffers that need *_raw load/store
     CK_TILE_HOST_DEVICE void init_raw()
     {
         cached_buf_res_ = make_wave_buffer_resource(p_data_, (buffer_size_) * sizeof(type));
     }
  
     CK_TILE_DEVICE static constexpr address_space_enum get_address_space()
     {
         return address_space_enum::global;
     }
  
     // i is offset of T
     // FIXME: doesn't do is_valid check
     CK_TILE_DEVICE constexpr const T& operator[](index_t i) const { return p_data_[i]; }
  
     // i is offset of T
     // FIXME: doesn't do is_valid check
     CK_TILE_DEVICE constexpr T& operator()(index_t i) { return p_data_[i]; }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto get(index_t i,
                                       index_t linear_offset,
                                       bool is_valid_element,
                                       bool_constant<oob_conditional_check> = {}) const
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
 #if CK_TILE_USE_AMD_BUFFER_LOAD
         bool constexpr use_amd_buffer_addressing = true;
 #else
         bool constexpr use_amd_buffer_addressing = false;
 #endif
  
         if constexpr(use_amd_buffer_addressing)
         {
             constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
  
             if constexpr(InvalidElementUseNumericalZeroValue)
             {
                 return amd_buffer_load_invalid_element_return_zero<remove_cvref_t<T>,
                                                                    t_per_x,
                                                                    Coherence,
                                                                    oob_conditional_check>(
                     p_data_, i + linear_offset, is_valid_element, buffer_size_);
             }
             else
             {
                 return amd_buffer_load_invalid_element_return_customized_value<
                     remove_cvref_t<T>,
                     t_per_x,
                     Coherence,
                     oob_conditional_check>(p_data_,
                                            i + linear_offset,
                                            is_valid_element,
                                            buffer_size_,
                                            invalid_element_value_);
             }
         }
         else
         {
             if(is_valid_element)
             {
 #if CK_TILE_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
                 X tmp;
  
                 __builtin_memcpy(&tmp, &(p_data_[i + linear_offset]), sizeof(X));
  
                 return tmp;
 #else
                 return *c_style_pointer_cast<const X*>(&p_data_[i + linear_offset]);
 #endif
             }
             else
             {
                 if constexpr(InvalidElementUseNumericalZeroValue)
                 {
                     return X{numeric<remove_cvref_t<T>>::zero()};
                 }
                 else
                 {
                     return X{invalid_element_value_};
                 }
             }
         }
     }
  
     /*
     In the global memory address space, we do not support the transpose instruction in the buffer
     view. Will report compilation error when developer wants to use it.
     */
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto transpose_get(index_t i,
                                                 index_t linear_offset,
                                                 bool is_valid_element,
                                                 bool_constant<oob_conditional_check> = {}) const
     {
         static_assert(false, "Error: transpose load not supported in global memory space.");
         ignore = i;
         ignore = linear_offset;
         ignore = is_valid_element;
         return;
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               bool pre_nop               = false,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto get_raw(remove_cvref_t<X>& dst,
                                           index_t v_offset,
                                           index_t i_offset,
                                           bool is_valid_element,
                                           bool_constant<pre_nop> = {}) const
     {
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
  
         amd_buffer_load_raw<remove_cvref_t<T>, t_per_x, Coherence, oob_conditional_check, pre_nop>(
             dst, cached_buf_res_, v_offset, i_offset, is_valid_element, bool_constant<pre_nop>{});
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto async_get(CK_TILE_LDS_ADDR remove_cvref_t<T>* smem,
                                             index_t i,
                                             index_t linear_offset,
                                             bool is_valid_element,
                                             bool_constant<oob_conditional_check> = {}) const
     {
         // X is vector of T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
         const int32x4_t src_wave_buffer_resource =
             make_wave_buffer_resource(p_data_, (buffer_size_) * sizeof(type));
  
         amd_async_buffer_load_with_oob<remove_cvref_t<T>, t_per_x, Coherence>(
             smem,
             src_wave_buffer_resource,
             i,
             linear_offset,
             is_valid_element,
             bool_constant<oob_conditional_check>{});
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool pre_nop = false,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto async_get_raw(remove_cvref_t<T>* smem,
                                                 index_t i,
                                                 index_t linear_offset,
                                                 bool /*is_valid_element*/,
                                                 bool_constant<pre_nop> = {}) const
     {
         // X is vector of T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
  
         amd_async_buffer_load_with_oob_raw<remove_cvref_t<T>, t_per_x, Coherence>(
             smem, cached_buf_res_, i, linear_offset, bool_constant<pre_nop>{});
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <memory_operation_enum Op,
               typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void update(index_t i,
                                index_t linear_offset,
                                bool is_valid_element,
                                const X& x,
                                bool_constant<oob_conditional_check> = {})
     {
         if constexpr(Op == memory_operation_enum::set)
         {
             this->template set<X, oob_conditional_check>(i, linear_offset, is_valid_element, x);
         }
         else if constexpr(Op == memory_operation_enum::atomic_add)
         {
             this->template atomic_add<X, oob_conditional_check>(
                 i, linear_offset, is_valid_element, x);
         }
         else if constexpr(Op == memory_operation_enum::atomic_max)
         {
             this->template atomic_max<X, oob_conditional_check>(
                 i, linear_offset, is_valid_element, x);
         }
         // FIXME: remove memory_operation_enum::add
         else if constexpr(Op == memory_operation_enum::add)
         {
             auto tmp =
                 this->template get<X, oob_conditional_check>(i, linear_offset, is_valid_element);
             this->template set<X, oob_conditional_check>(
                 i, linear_offset, is_valid_element, x + tmp);
             // tmp += x;
             // this->template set<X>(i, is_valid_element, tmp);
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <memory_operation_enum Op,
               typename X,
               bool oob_conditional_check = true,
               bool pre_nop               = false,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void update_raw(index_t i,
                                    index_t linear_offset,
                                    bool is_valid_element,
                                    const X& x,
                                    bool_constant<oob_conditional_check> = {},
                                    bool_constant<pre_nop>               = {})
     {
         if constexpr(Op == memory_operation_enum::set)
         {
             this->template set_raw<X, oob_conditional_check>(i, linear_offset, is_valid_element, x);
         }
         else if constexpr(Op == memory_operation_enum::atomic_add)
         {
             this->template atomic_add_raw<X, oob_conditional_check, pre_nop>(
                 i, linear_offset, is_valid_element, x);
         }
         else if constexpr(Op == memory_operation_enum::atomic_max)
         {
             // this->template atomic_max_raw<X>(i, linear_offset, is_valid_element, x);
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void set(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
 #if CK_TILE_USE_AMD_BUFFER_STORE
         bool constexpr use_amd_buffer_addressing = true;
 #else
         bool constexpr use_amd_buffer_addressing = false;
 #endif
  
         if constexpr(use_amd_buffer_addressing)
         {
             constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
  
             amd_buffer_store<remove_cvref_t<T>, t_per_x, Coherence>(
                 x, p_data_, i + linear_offset, is_valid_element, buffer_size_);
         }
         else
         {
             if(is_valid_element)
             {
 #if CK_TILE_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
                 X tmp = x;
  
                 __builtin_memcpy(&(p_data_[i + linear_offset]), &tmp, sizeof(X));
 #else
                 *c_style_pointer_cast<X*>(&p_data_[i + linear_offset]) = x;
 #endif
             }
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void set_raw(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
         amd_buffer_store_raw<remove_cvref_t<T>, t_per_x, Coherence, oob_conditional_check>(
             x, p_data_, i, linear_offset, is_valid_element, buffer_size_);
     }
  
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void
     atomic_add(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         using scalar_t = typename vector_traits<remove_cvref_t<T>>::scalar_type;
  
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         static_assert(get_address_space() == address_space_enum::global, "only support global mem");
  
 #if CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER && CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT
         bool constexpr use_amd_buffer_addressing =
             std::is_same_v<remove_cvref_t<scalar_t>, int32_t> ||
             std::is_same_v<remove_cvref_t<scalar_t>, float> ||
             (std::is_same_v<remove_cvref_t<scalar_t>, half_t> && scalar_per_x_vector % 2 == 0);
 #elif CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER && (!CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT)
         bool constexpr use_amd_buffer_addressing =
             std::is_same_v<remove_cvref_t<scalar_t>, int32_t>;
 #elif(!CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER) && CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT
         bool constexpr use_amd_buffer_addressing =
             std::is_same_v<remove_cvref_t<scalar_t>, float> ||
             (std::is_same_v<remove_cvref_t<scalar_t>, half_t> && scalar_per_x_vector % 2 == 0);
 #else
         bool constexpr use_amd_buffer_addressing = false;
 #endif
  
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
  
         if constexpr(use_amd_buffer_addressing)
         {
             amd_buffer_atomic_add<remove_cvref_t<T>, t_per_x>(
                 x, p_data_, i + linear_offset, is_valid_element, buffer_size_);
         }
         else
         {
             if(is_valid_element)
             {
                 atomic_add_g<remove_cvref_t<T>, t_per_x>(&p_data_[i + linear_offset], x);
             }
         }
     }
  
     template <typename X,
               bool oob_conditional_check = true,
               bool pre_nop               = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void
     atomic_add_raw(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         // using scalar_t = typename vector_traits<remove_cvref_t<T>>::scalar_type;
  
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         static_assert(get_address_space() == address_space_enum::global, "only support global mem");
  
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
  
         amd_buffer_atomic_add_raw<remove_cvref_t<T>,
                                   t_per_x,
                                   Coherence,
                                   oob_conditional_check,
                                   pre_nop>(
             x, p_data_, i, linear_offset, is_valid_element, buffer_size_);
     }
  
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void
     atomic_max(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         static_assert(get_address_space() == address_space_enum::global, "only support global mem");
  
 #if CK_TILE_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64
         using scalar_t = typename vector_traits<remove_cvref_t<T>>::scalar_type;
         bool constexpr use_amd_buffer_addressing = std::is_same_v<remove_cvref_t<scalar_t>, double>;
 #else
         bool constexpr use_amd_buffer_addressing = false;
 #endif
  
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
  
         if constexpr(use_amd_buffer_addressing)
         {
             amd_buffer_atomic_max<remove_cvref_t<T>, t_per_x>(
                 x, p_data_, i + linear_offset, is_valid_element, buffer_size_);
         }
         else if(is_valid_element)
         {
             atomic_max_g<remove_cvref_t<T>, t_per_x>(&p_data_[i + linear_offset], x);
         }
     }
  
     // FIXME: remove
     CK_TILE_DEVICE static constexpr bool is_static_buffer() { return false; }
  
     // FIXME: remove
     CK_TILE_DEVICE static constexpr bool is_dynamic_buffer() { return true; }
  
     CK_TILE_HOST_DEVICE void print() const
     {
         printf("buffer_view{");
  
         // AddressSpace
         printf("AddressSpace: Global, ");
  
         // p_data_
         printf("p_data_: %p, ", static_cast<void*>(const_cast<remove_cvref_t<T>*>(p_data_)));
  
         // buffer_size_
         printf("buffer_size_: ");
         print(buffer_size_);
         printf(", ");
  
         // invalid_element_value_
         printf("invalid_element_value_: ");
         print(invalid_element_value_);
  
         printf("}");
     }
 };
  
 // Address Space: LDS
 // T may be scalar or vector
 // X may be scalar or vector
 // T and X have same scalar type
 // X contains multiple T
 // FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
 //        transforms of tensor_view/Tensor
 template <typename T, typename BufferSizeType, bool InvalidElementUseNumericalZeroValue>
 struct buffer_view<address_space_enum::lds,
                    T,
                    BufferSizeType,
                    InvalidElementUseNumericalZeroValue,
                    amd_buffer_coherence_enum::coherence_default>
 {
     using type = T;
  
     T* p_data_ = nullptr;
     BufferSizeType buffer_size_;
     remove_cvref_t<T> invalid_element_value_ = T{0};
  
     CK_TILE_HOST_DEVICE constexpr buffer_view()
         : p_data_{}, buffer_size_{}, invalid_element_value_{}
     {
     }
  
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data, BufferSizeType buffer_size)
         : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{0}
     {
     }
  
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data,
                                               BufferSizeType buffer_size,
                                               T invalid_element_value)
         : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{invalid_element_value}
     {
     }
  
     CK_TILE_HOST_DEVICE void init_raw() {}
  
     CK_TILE_DEVICE static constexpr address_space_enum get_address_space()
     {
         return address_space_enum::lds;
     }
  
     // i is offset of T
     // FIXME: doesn't do is_valid check
     CK_TILE_DEVICE constexpr const T& operator[](index_t i) const { return p_data_[i]; }
  
     // i is offset of T
     // FIXME: doesn't do is_valid check
     CK_TILE_DEVICE constexpr T& operator()(index_t i) { return p_data_[i]; }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto get(index_t i,
                                       index_t linear_offset,
                                       bool is_valid_element,
                                       bool_constant<oob_conditional_check> = {}) const
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         if(is_valid_element)
         {
 #if CK_TILE_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
             X tmp;
  
             __builtin_memcpy(&tmp, &(p_data_[i + linear_offset]), sizeof(X));
  
             return tmp;
 #else
             using buf_t = ext_vector_t<typename vector_traits<remove_cvref_t<T>>::scalar_type,
                                        scalar_per_t_vector * scalar_per_x_vector>;
             // using buf_t = ushort __attribute__((ext_vector_type(8)));
             auto rtn = *c_style_pointer_cast<const buf_t*>(&p_data_[i + linear_offset]);
             return bit_cast<X>(rtn);
 #endif
         }
         else
         {
             if constexpr(InvalidElementUseNumericalZeroValue)
             {
                 return X{numeric<remove_cvref_t<T>>::zero()};
             }
             else
             {
                 return X{invalid_element_value_};
             }
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               bool pre_nop               = false,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto get_raw(remove_cvref_t<X>& dst,
                                           index_t v_offset,
                                           index_t i_offset,
                                           bool /*is_valid_element*/,
                                           bool_constant<pre_nop> = {}) const
     {
         smem_load<sizeof(X)>{}(dst, v_offset * sizeof(T), i_offset * sizeof(T));
     }
  
     template <typename X,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto transpose_get([[maybe_unused]] index_t i,
                                                 [[maybe_unused]] index_t linear_offset,
                                                 bool is_valid_element) const
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         if(is_valid_element)
         {
 #if defined(__gfx950__)
             constexpr index_t t_per_x               = scalar_per_x_vector / scalar_per_t_vector;
             constexpr address_space_enum addr_space = get_address_space();
             return amd_transpose_load_to_vgpr<remove_cvref_t<T>, t_per_x, addr_space>(
                 p_data_ + i + linear_offset);
 #else
             return X{numeric<remove_cvref_t<T>>::zero()};
 #endif
         }
         else
         {
             if constexpr(InvalidElementUseNumericalZeroValue)
             {
                 return X{numeric<remove_cvref_t<T>>::zero()};
             }
             else
             {
                 return X{invalid_element_value_};
             }
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <memory_operation_enum Op,
               typename X,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void update(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         if constexpr(Op == memory_operation_enum::set)
         {
             this->template set<X>(i, linear_offset, is_valid_element, x);
         }
         // FIXME: remove memory_operation_enum::add
         else if constexpr(Op == memory_operation_enum::add)
         {
             auto tmp = this->template get<X>(i, linear_offset, is_valid_element);
             this->template set<X>(i, linear_offset, is_valid_element, x + tmp);
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void set(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
 #if CK_TILE_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE
         bool constexpr workaround_int8_ds_write_issue = true;
 #else
         bool constexpr workaround_int8_ds_write_issue = false;
 #endif
  
         i += linear_offset; // simplicity
         if constexpr(std::is_same_v<typename vector_traits<remove_cvref_t<T>>::scalar_type,
                                     int8_t> &&
                      workaround_int8_ds_write_issue)
         {
             if(is_valid_element)
             {
                 // HACK: compiler would lower IR "store<i8, 16> address_space(3)" into inefficient
                 // ISA, so I try to let compiler emit IR "store<i32, 4>" which would be lower to
                 // ds_write_b128
                 // TODO: remove this after compiler fix
                 static_assert(
                     (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                      std::is_same_v<remove_cvref_t<X>, int8_t>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                          std::is_same_v<remove_cvref_t<X>, int8x2_t>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                          std::is_same_v<remove_cvref_t<X>, int8x4_t>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                          std::is_same_v<remove_cvref_t<X>, int8x8_t>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                          std::is_same_v<remove_cvref_t<X>, int8x16_t>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8x4_t> &&
                          std::is_same_v<remove_cvref_t<X>, int8x4_t>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8x8_t> &&
                          std::is_same_v<remove_cvref_t<X>, int8x8_t>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8x16_t> &&
                          std::is_same_v<remove_cvref_t<X>, int8x16_t>) ||
                         // int8 on thread buffer
                         (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<int8_t, 8>>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<int8_t, 4>>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<int8_t, 2>>) ||
                         (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<int8_t, 1>>) ||
                         // ext_vector_type for pk_int4 must use int8_t as type
                         (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 1>>) ||
                         (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 2>>) ||
                         (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 4>>) ||
                         (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 8>>) ||
                         (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 16>>) ||
                         (std::is_same_v<remove_cvref_t<T>, pk_int4x4_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 4>>) ||
                         (std::is_same_v<remove_cvref_t<T>, pk_int4x8_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 8>>) ||
                         (std::is_same_v<remove_cvref_t<T>, pk_int4x16_t> &&
                          std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 16>>),
                     "wrong! not implemented for this combination, please add "
                     "implementation");
  
                 if constexpr((std::is_same_v<remove_cvref_t<T>, int8_t> &&
                               std::is_same_v<remove_cvref_t<X>, int8_t>) ||
                              (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                               std::is_same_v<remove_cvref_t<X>, thread_buffer<int8_t, 1>>) ||
                              (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                               std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 1>>))
                 {
                     // HACK: cast pointer of x is bad
                     // TODO: remove this after compiler fix
                     *c_style_pointer_cast<int8_t*>(&p_data_[i]) =
                         *c_style_pointer_cast<const int8_t*>(&x);
                 }
                 else if constexpr((std::is_same_v<remove_cvref_t<T>, int8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, int8x2_t>) ||
                                   (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<int8_t, 2>>) ||
                                   (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 2>>))
                 {
                     // HACK: cast pointer of x is bad
                     // TODO: remove this after compiler fix
                     *c_style_pointer_cast<int16_t*>(&p_data_[i]) =
                         *c_style_pointer_cast<const int16_t*>(&x);
                 }
                 else if constexpr((std::is_same_v<remove_cvref_t<T>, int8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, int8x4_t>) ||
                                   (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<int8_t, 4>>) ||
                                   (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 4>>))
                 {
                     // HACK: cast pointer of x is bad
                     // TODO: remove this after compiler fix
                     *c_style_pointer_cast<int32_t*>(&p_data_[i]) =
                         *c_style_pointer_cast<const int32_t*>(&x);
                 }
                 else if constexpr((std::is_same_v<remove_cvref_t<T>, int8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, int8x8_t>) ||
                                   (std::is_same_v<remove_cvref_t<T>, int8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<int8_t, 8>>) ||
                                   (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 8>>))
                 {
                     // HACK: cast pointer of x is bad
                     // TODO: remove this after compiler fix
                     *c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
                         *c_style_pointer_cast<const int32x2_t*>(&x);
                 }
                 else if constexpr((std::is_same_v<remove_cvref_t<T>, int8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, int8x16_t>) ||
                                   (std::is_same_v<remove_cvref_t<T>, pk_int4_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 16>>))
                 {
                     // HACK: cast pointer of x is bad
                     // TODO: remove this after compiler fix
                     *c_style_pointer_cast<int32x4_t*>(&p_data_[i]) =
                         *c_style_pointer_cast<const int32x4_t*>(&x);
                 }
                 else if constexpr((std::is_same_v<remove_cvref_t<T>, int8x4_t> &&
                                    std::is_same_v<remove_cvref_t<X>, int8x4_t>) ||
                                   (std::is_same_v<remove_cvref_t<T>, pk_int4x4_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 4>>))
                 {
                     // HACK: cast pointer of x is bad
                     // TODO: remove this after compiler fix
                     *c_style_pointer_cast<int32_t*>(&p_data_[i]) =
                         *c_style_pointer_cast<const int32_t*>(&x);
                 }
                 else if constexpr((std::is_same_v<remove_cvref_t<T>, int8x8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, int8x8_t>) ||
                                   (std::is_same_v<remove_cvref_t<T>, pk_int4x8_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 8>>))
                 {
                     // HACK: cast pointer of x is bad
                     // TODO: remove this after compiler fix
                     *c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
                         *c_style_pointer_cast<const int32x2_t*>(&x);
                 }
                 else if constexpr((std::is_same_v<remove_cvref_t<T>, int8x16_t> &&
                                    std::is_same_v<remove_cvref_t<X>, int8x16_t>) ||
                                   (std::is_same_v<remove_cvref_t<T>, pk_int4x16_t> &&
                                    std::is_same_v<remove_cvref_t<X>, thread_buffer<pk_int4_t, 16>>))
                 {
                     // HACK: cast pointer of x is bad
                     // TODO: remove this after compiler fix
                     *c_style_pointer_cast<int32x4_t*>(&p_data_[i]) =
                         *c_style_pointer_cast<const int32x4_t*>(&x);
                 }
             }
         }
         else
         {
             if(is_valid_element)
             {
 #if CK_TILE_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
                 X tmp = x;
  
                 __builtin_memcpy(&(p_data_[i]), &tmp, sizeof(X));
 #else
                 using buf_t = ext_vector_t<typename vector_traits<remove_cvref_t<T>>::scalar_type,
                                            scalar_per_t_vector * scalar_per_x_vector>;
  
                 *c_style_pointer_cast<buf_t*>(&p_data_[i]) = reinterpret_cast<const buf_t&>(x);
 #endif
             }
         }
     }
  
     // FIXME: remove
     CK_TILE_DEVICE static constexpr bool is_static_buffer() { return false; }
  
     // FIXME: remove
     CK_TILE_DEVICE static constexpr bool is_dynamic_buffer() { return true; }
  
     CK_TILE_HOST_DEVICE void print() const
     {
         printf("buffer_view{");
  
         // AddressSpace
         printf("AddressSpace: Lds, ");
  
         // p_data_
         printf("p_data_: %p, ", static_cast<void*>(const_cast<remove_cvref_t<T>*>(p_data_)));
  
         // buffer_size_
         printf("buffer_size_: ");
         print(buffer_size_);
         printf(", ");
  
         // invalid_element_value_
         printf("invalid_element_value_: ");
         print(invalid_element_value_);
  
         printf("}");
     }
 };
  
 // Address Space: Vgpr
 // T may be scalar or vector
 // X may be scalar or vector
 // T and X have same scalar type
 // X contains multiple T
 // FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
 //        transforms of tensor_view/Tensor
 template <typename T, typename BufferSizeType, bool InvalidElementUseNumericalZeroValue>
 struct buffer_view<address_space_enum::vgpr,
                    T,
                    BufferSizeType,
                    InvalidElementUseNumericalZeroValue,
                    amd_buffer_coherence_enum::coherence_default>
 {
     using type = T;
  
     T* p_data_ = nullptr;
     BufferSizeType buffer_size_;
     remove_cvref_t<T> invalid_element_value_ = T{0};
  
     CK_TILE_HOST_DEVICE constexpr buffer_view()
         : p_data_{}, buffer_size_{}, invalid_element_value_{}
     {
     }
  
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data, BufferSizeType buffer_size)
         : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{0}
     {
     }
  
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data,
                                               BufferSizeType buffer_size,
                                               T invalid_element_value)
         : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{invalid_element_value}
     {
     }
  
     CK_TILE_HOST_DEVICE void init_raw() {}
  
     CK_TILE_DEVICE static constexpr address_space_enum get_address_space()
     {
         return address_space_enum::vgpr;
     }
  
     // i is offset of T
     // FIXME: doesn't do is_valid check
     CK_TILE_DEVICE constexpr const T& operator[](index_t i) const { return p_data_[i]; }
  
     // i is offset of T
     // FIXME: doesn't do is_valid check
     CK_TILE_DEVICE constexpr T& operator()(index_t i) { return p_data_[i]; }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE constexpr auto get(index_t i,
                                       index_t /*linear_offset*/,
                                       bool is_valid_element,
                                       bool_constant<oob_conditional_check> = {}) const
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         if(is_valid_element)
         {
 #if CK_TILE_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
             X tmp;
  
             __builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
  
             return tmp;
 #else
             return *c_style_pointer_cast<const X*>(&p_data_[i]);
 #endif
         }
         else
         {
             if constexpr(InvalidElementUseNumericalZeroValue)
             {
                 return X{numeric<remove_cvref_t<T>>::zero()};
             }
             else
             {
                 return X{invalid_element_value_};
             }
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <memory_operation_enum Op,
               typename X,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void update(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         if constexpr(Op == memory_operation_enum::set)
         {
             this->template set<X>(i, linear_offset, is_valid_element, x);
         }
         // FIXME: remove memory_operation_enum::add
         else if constexpr(Op == memory_operation_enum::add)
         {
             auto tmp = this->template get<X>(i, linear_offset, is_valid_element);
             this->template set<X>(i, linear_offset, is_valid_element, x + tmp);
         }
     }
  
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
     CK_TILE_DEVICE void set(index_t i, index_t linear_offset, bool is_valid_element, const X& x)
     {
         // X contains multiple T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
  
         constexpr index_t scalar_per_x_vector = vector_traits<remove_cvref_t<X>>::vector_size;
  
         static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
                       "wrong! X should contain multiple T");
  
         if(is_valid_element)
         {
 #if CK_TILE_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
             X tmp = x;
  
             __builtin_memcpy(&(p_data_[i + linear_offset]), &tmp, sizeof(X));
 #else
             *c_style_pointer_cast<X*>(&p_data_[i + linear_offset]) = x;
 #endif
         }
     }
  
     // FIXME: remove
     CK_TILE_DEVICE static constexpr bool is_static_buffer() { return false; }
  
     // FIXME: remove
     CK_TILE_DEVICE static constexpr bool is_dynamic_buffer() { return true; }
  
     CK_TILE_HOST_DEVICE void print() const
     {
         printf("buffer_view{");
  
         // AddressSpace
         printf("AddressSpace: Vgpr, ");
  
         // p_data_
         printf("p_data_: %p, ", static_cast<void*>(const_cast<remove_cvref_t<T>*>(p_data_)));
  
         // buffer_size_
         printf("buffer_size_: ");
         print(buffer_size_);
         printf(", ");
  
         // invalid_element_value_
         printf("invalid_element_value_: ");
         print(invalid_element_value_);
  
         printf("}");
     }
 };
  
 template <address_space_enum BufferAddressSpace,
           amd_buffer_coherence_enum Coherence = amd_buffer_coherence_enum::coherence_default,
           typename T,
           typename BufferSizeType>
 CK_TILE_HOST_DEVICE constexpr auto make_buffer_view(T* p, BufferSizeType buffer_size)
 {
     return buffer_view<BufferAddressSpace, T, BufferSizeType, true, Coherence>{p, buffer_size};
 }
  
 template <address_space_enum BufferAddressSpace,
           amd_buffer_coherence_enum Coherence = amd_buffer_coherence_enum::coherence_default,
           typename T,
           typename BufferSizeType,
           typename X,
           typename std::enable_if<std::is_same<remove_cvref_t<T>, remove_cvref_t<X>>::value,
                                   bool>::type = false>
 CK_TILE_HOST_DEVICE constexpr auto
 make_buffer_view(T* p, BufferSizeType buffer_size, X invalid_element_value)
 {
     return buffer_view<BufferAddressSpace, T, BufferSizeType, false, Coherence>{
         p, buffer_size, invalid_element_value};
 }
  
 } // namespace ck_tile