Merge branch 'gfx950' into lwpck-2390

test/data_type/CMakeLists.txt

@@ -22,4 +22,9 @@ if(result EQUAL 0)
   target_link_libraries(test_fp4 PRIVATE utility)
 endif()
 
+add_gtest_executable(test_custom_type test_custom_type.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_custom_type PRIVATE utility)
+endif()
+
 add_gtest_executable(test_type_convert_const type_convert_const.cpp)

test/data_type/test_custom_type.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type_convert.hpp"
+
+using ck::bf8_t;
+using ck::bhalf_t;
+using ck::f8_t;
+using ck::half_t;
+using ck::Number;
+using ck::type_convert;
+using ck::vector_type;
+
+TEST(Custom_bool, TestSize)
+{
+    struct custom_bool_t
+    {
+        bool data;
+    };
+    ASSERT_EQ(sizeof(custom_bool_t), sizeof(bool));
+    ASSERT_EQ(sizeof(vector_type<custom_bool_t, 2>), sizeof(vector_type<bool, 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_bool_t, 4>), sizeof(vector_type<bool, 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_bool_t, 8>), sizeof(vector_type<bool, 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_bool_t, 16>), sizeof(vector_type<bool, 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_bool_t, 32>), sizeof(vector_type<bool, 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_bool_t, 64>), sizeof(vector_type<bool, 64>));
+}
+
+TEST(Custom_bool, TestAsType)
+{
+    struct custom_bool_t
+    {
+        using type = bool;
+        type data;
+        custom_bool_t() : data{type{}} {}
+        custom_bool_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size             = 4;
+    std::vector<bool> test_vec = {false, true, false, true};
+    // reference vector
+    vector_type<custom_bool_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_bool_t>()(Number<i>{}).data, false);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_bool_t>()(Number<i>{}) = custom_bool_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_bool_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_bool_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_bool, TestAsTypeReshape)
+{
+    struct custom_bool_t
+    {
+        using type = bool;
+        type data;
+        custom_bool_t() : data{type{}} {}
+        custom_bool_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size             = 4;
+    std::vector<bool> test_vec = {false, true, false, true};
+    // reference vector
+    vector_type<custom_bool_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_bool_t>()(Number<i>{}).data, false);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_bool_t>()(Number<i>{}) = custom_bool_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_bool_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_bool_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_bool_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_int8, TestSize)
+{
+    struct custom_int8_t
+    {
+        int8_t data;
+    };
+    ASSERT_EQ(sizeof(custom_int8_t), sizeof(int8_t));
+    ASSERT_EQ(sizeof(vector_type<custom_int8_t, 2>), sizeof(vector_type<int8_t, 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_int8_t, 4>), sizeof(vector_type<int8_t, 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_int8_t, 8>), sizeof(vector_type<int8_t, 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_int8_t, 16>), sizeof(vector_type<int8_t, 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_int8_t, 32>), sizeof(vector_type<int8_t, 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_int8_t, 64>), sizeof(vector_type<int8_t, 64>));
+}
+
+TEST(Custom_int8, TestAsType)
+{
+    struct custom_int8_t
+    {
+        using type = int8_t;
+        type data;
+        custom_int8_t() : data{type{}} {}
+        custom_int8_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size               = 4;
+    std::vector<int8_t> test_vec = {3, -6, 8, -2};
+    // reference vector
+    vector_type<custom_int8_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_int8_t>()(Number<i>{}).data, 0);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_int8_t>()(Number<i>{}) = custom_int8_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_int8_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_int8_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_int8, TestAsTypeReshape)
+{
+    struct custom_int8_t
+    {
+        using type = int8_t;
+        type data;
+        custom_int8_t() : data{type{}} {}
+        custom_int8_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size               = 4;
+    std::vector<int8_t> test_vec = {3, -6, 8, -2};
+    // reference vector
+    vector_type<custom_int8_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_int8_t>()(Number<i>{}).data, 0);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_int8_t>()(Number<i>{}) = custom_int8_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_int8_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_int8_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_int8_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_uint8, TestSize)
+{
+    struct custom_uint8_t
+    {
+        uint8_t data;
+    };
+    ASSERT_EQ(sizeof(custom_uint8_t), sizeof(uint8_t));
+    ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 2>), sizeof(vector_type<uint8_t, 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 4>), sizeof(vector_type<uint8_t, 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 8>), sizeof(vector_type<uint8_t, 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 16>), sizeof(vector_type<uint8_t, 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 32>), sizeof(vector_type<uint8_t, 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_uint8_t, 64>), sizeof(vector_type<uint8_t, 64>));
+}
+
+TEST(Custom_uint8, TestAsType)
+{
+    struct custom_uint8_t
+    {
+        using type = uint8_t;
+        type data;
+        custom_uint8_t() : data{type{}} {}
+        custom_uint8_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size                = 4;
+    std::vector<uint8_t> test_vec = {3, 6, 8, 2};
+    // reference vector
+    vector_type<custom_uint8_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, 0);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_uint8_t>()(Number<i>{}) = custom_uint8_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_uint8_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_uint8, TestAsTypeReshape)
+{
+    struct custom_uint8_t
+    {
+        using type = uint8_t;
+        type data;
+        custom_uint8_t() : data{type{}} {}
+        custom_uint8_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size                = 4;
+    std::vector<uint8_t> test_vec = {3, 6, 8, 2};
+    // reference vector
+    vector_type<custom_uint8_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, 0);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_uint8_t>()(Number<i>{}) = custom_uint8_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_uint8_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_uint8_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_uint8_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_f8, TestSize)
+{
+    struct custom_f8_t
+    {
+        _BitInt(8) data;
+    };
+    ASSERT_EQ(sizeof(custom_f8_t), sizeof(_BitInt(8)));
+    ASSERT_EQ(sizeof(vector_type<custom_f8_t, 2>), sizeof(vector_type<_BitInt(8), 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_f8_t, 4>), sizeof(vector_type<_BitInt(8), 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_f8_t, 8>), sizeof(vector_type<_BitInt(8), 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_f8_t, 16>), sizeof(vector_type<_BitInt(8), 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_f8_t, 32>), sizeof(vector_type<_BitInt(8), 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_f8_t, 64>), sizeof(vector_type<_BitInt(8), 64>));
+}
+
+TEST(Custom_f8, TestAsType)
+{
+    struct custom_f8_t
+    {
+        using type = _BitInt(8);
+        type data;
+        custom_f8_t() : data{type{}} {}
+        custom_f8_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size                   = 4;
+    std::vector<_BitInt(8)> test_vec = {type_convert<_BitInt(8)>(0.3f),
+                                        type_convert<_BitInt(8)>(-0.6f),
+                                        type_convert<_BitInt(8)>(0.8f),
+                                        type_convert<_BitInt(8)>(-0.2f)};
+    // reference vector
+    vector_type<custom_f8_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}(
+        [&](auto i) { ASSERT_EQ(right_vec.template AsType<custom_f8_t>()(Number<i>{}).data, 0); });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_f8_t>()(Number<i>{}) = custom_f8_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_f8_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_f8_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_f8, TestAsTypeReshape)
+{
+    struct custom_f8_t
+    {
+        using type = _BitInt(8);
+        type data;
+        custom_f8_t() : data{type{}} {}
+        custom_f8_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size                   = 4;
+    std::vector<_BitInt(8)> test_vec = {type_convert<_BitInt(8)>(0.3f),
+                                        type_convert<_BitInt(8)>(-0.6f),
+                                        type_convert<_BitInt(8)>(0.8f),
+                                        type_convert<_BitInt(8)>(-0.2f)};
+    // reference vector
+    vector_type<custom_f8_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}(
+        [&](auto i) { ASSERT_EQ(right_vec.template AsType<custom_f8_t>()(Number<i>{}).data, 0); });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_f8_t>()(Number<i>{}) = custom_f8_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_f8_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_f8_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_f8_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_bf8, TestSize)
+{
+    struct custom_bf8_t
+    {
+        unsigned _BitInt(8) data;
+    };
+    ASSERT_EQ(sizeof(custom_bf8_t), sizeof(unsigned _BitInt(8)));
+    ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 2>), sizeof(vector_type<unsigned _BitInt(8), 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 4>), sizeof(vector_type<unsigned _BitInt(8), 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 8>), sizeof(vector_type<unsigned _BitInt(8), 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 16>), sizeof(vector_type<unsigned _BitInt(8), 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 32>), sizeof(vector_type<unsigned _BitInt(8), 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_bf8_t, 64>), sizeof(vector_type<unsigned _BitInt(8), 64>));
+}
+
+TEST(Custom_bf8, TestAsType)
+{
+    struct custom_bf8_t
+    {
+        using type = unsigned _BitInt(8);
+        type data;
+        custom_bf8_t() : data{type{}} {}
+        custom_bf8_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size                            = 4;
+    std::vector<unsigned _BitInt(8)> test_vec = {type_convert<unsigned _BitInt(8)>(0.3f),
+                                                 type_convert<unsigned _BitInt(8)>(-0.6f),
+                                                 type_convert<unsigned _BitInt(8)>(0.8f),
+                                                 type_convert<unsigned _BitInt(8)>(-0.2f)};
+    // reference vector
+    vector_type<custom_bf8_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}(
+        [&](auto i) { ASSERT_EQ(right_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, 0); });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_bf8_t>()(Number<i>{}) = custom_bf8_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_bf8_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_bf8, TestAsTypeReshape)
+{
+    struct custom_bf8_t
+    {
+        using type = unsigned _BitInt(8);
+        type data;
+        custom_bf8_t() : data{type{}} {}
+        custom_bf8_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size                            = 4;
+    std::vector<unsigned _BitInt(8)> test_vec = {type_convert<unsigned _BitInt(8)>(0.3f),
+                                                 type_convert<unsigned _BitInt(8)>(-0.6f),
+                                                 type_convert<unsigned _BitInt(8)>(0.8f),
+                                                 type_convert<unsigned _BitInt(8)>(-0.2f)};
+    // reference vector
+    vector_type<custom_bf8_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}(
+        [&](auto i) { ASSERT_EQ(right_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, 0); });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_bf8_t>()(Number<i>{}) = custom_bf8_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_bf8_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_bf8_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_bf8_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_half, TestSize)
+{
+    struct custom_half_t
+    {
+        half_t data;
+    };
+    ASSERT_EQ(sizeof(custom_half_t), sizeof(half_t));
+    ASSERT_EQ(sizeof(vector_type<custom_half_t, 2>), sizeof(vector_type<half_t, 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_half_t, 4>), sizeof(vector_type<half_t, 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_half_t, 8>), sizeof(vector_type<half_t, 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_half_t, 16>), sizeof(vector_type<half_t, 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_half_t, 32>), sizeof(vector_type<half_t, 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_half_t, 64>), sizeof(vector_type<half_t, 64>));
+}
+
+TEST(Custom_half, TestAsType)
+{
+    struct custom_half_t
+    {
+        using type = half_t;
+        type data;
+        custom_half_t() : data{type{}} {}
+        custom_half_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size               = 4;
+    std::vector<half_t> test_vec = {half_t{0.3f}, half_t{-0.6f}, half_t{0.8f}, half_t{-0.2f}};
+    // reference vector
+    vector_type<custom_half_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_half_t>()(Number<i>{}).data,
+                  type_convert<half_t>(0.0f));
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_half_t>()(Number<i>{}) = custom_half_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_half_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_half_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_half, TestAsTypeReshape)
+{
+    struct custom_half_t
+    {
+        using type = half_t;
+        type data;
+        custom_half_t() : data{type{}} {}
+        custom_half_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size               = 4;
+    std::vector<half_t> test_vec = {half_t{0.3f}, half_t{-0.6f}, half_t{0.8f}, half_t{-0.2f}};
+    // reference vector
+    vector_type<custom_half_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_half_t>()(Number<i>{}).data,
+                  type_convert<half_t>(0.0f));
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_half_t>()(Number<i>{}) = custom_half_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_half_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_half_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_half_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_bhalf, TestSize)
+{
+    struct custom_bhalf_t
+    {
+        bhalf_t data;
+    };
+    ASSERT_EQ(sizeof(custom_bhalf_t), sizeof(bhalf_t));
+    ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 2>), sizeof(vector_type<bhalf_t, 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 4>), sizeof(vector_type<bhalf_t, 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 8>), sizeof(vector_type<bhalf_t, 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 16>), sizeof(vector_type<bhalf_t, 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 32>), sizeof(vector_type<bhalf_t, 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_bhalf_t, 64>), sizeof(vector_type<bhalf_t, 64>));
+}
+
+TEST(Custom_bhalf, TestAsType)
+{
+    struct custom_bhalf_t
+    {
+        using type = bhalf_t;
+        type data;
+        custom_bhalf_t() : data{type{}} {}
+        custom_bhalf_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size                = 4;
+    std::vector<bhalf_t> test_vec = {type_convert<bhalf_t>(0.3f),
+                                     type_convert<bhalf_t>(-0.6f),
+                                     type_convert<bhalf_t>(0.8f),
+                                     type_convert<bhalf_t>(-0.2f)};
+    // reference vector
+    vector_type<custom_bhalf_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data,
+                  type_convert<bhalf_t>(0.0f));
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_bhalf_t>()(Number<i>{}) = custom_bhalf_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_bhalf_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_bhalf, TestAsTypeReshape)
+{
+    struct custom_bhalf_t
+    {
+        using type = bhalf_t;
+        type data;
+        custom_bhalf_t() : data{type{}} {}
+        custom_bhalf_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size                = 4;
+    std::vector<bhalf_t> test_vec = {type_convert<bhalf_t>(0.3f),
+                                     type_convert<bhalf_t>(-0.6f),
+                                     type_convert<bhalf_t>(0.8f),
+                                     type_convert<bhalf_t>(-0.2f)};
+    // reference vector
+    vector_type<custom_bhalf_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data,
+                  type_convert<bhalf_t>(0.0f));
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_bhalf_t>()(Number<i>{}) = custom_bhalf_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_bhalf_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_bhalf_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_bhalf_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_float, TestSize)
+{
+    struct custom_float_t
+    {
+        float data;
+    };
+    ASSERT_EQ(sizeof(custom_float_t), sizeof(float));
+    ASSERT_EQ(sizeof(vector_type<custom_float_t, 2>), sizeof(vector_type<float, 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_float_t, 4>), sizeof(vector_type<float, 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_float_t, 8>), sizeof(vector_type<float, 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_float_t, 16>), sizeof(vector_type<float, 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_float_t, 32>), sizeof(vector_type<float, 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_float_t, 64>), sizeof(vector_type<float, 64>));
+}
+
+TEST(Custom_float, TestAsType)
+{
+    struct custom_float_t
+    {
+        using type = float;
+        type data;
+        custom_float_t() : data{type{}} {}
+        custom_float_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size              = 4;
+    std::vector<float> test_vec = {0.3f, -0.6f, 0.8f, -0.2f};
+    // reference vector
+    vector_type<custom_float_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_float_t>()(Number<i>{}).data, 0.0f);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_float_t>()(Number<i>{}) = custom_float_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_float_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_float_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_float, TestAsTypeReshape)
+{
+    struct custom_float_t
+    {
+        using type = float;
+        type data;
+        custom_float_t() : data{type{}} {}
+        custom_float_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size              = 4;
+    std::vector<float> test_vec = {0.3f, -0.6f, 0.8f, -0.2f};
+    // reference vector
+    vector_type<custom_float_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_float_t>()(Number<i>{}).data, 0.0f);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_float_t>()(Number<i>{}) = custom_float_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_float_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_float_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_float_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_double, TestSize)
+{
+    struct custom_double_t
+    {
+        double data;
+    };
+    ASSERT_EQ(sizeof(custom_double_t), sizeof(double));
+    ASSERT_EQ(sizeof(vector_type<custom_double_t, 2>), sizeof(vector_type<double, 2>));
+    ASSERT_EQ(sizeof(vector_type<custom_double_t, 4>), sizeof(vector_type<double, 4>));
+    ASSERT_EQ(sizeof(vector_type<custom_double_t, 8>), sizeof(vector_type<double, 8>));
+    ASSERT_EQ(sizeof(vector_type<custom_double_t, 16>), sizeof(vector_type<double, 16>));
+    ASSERT_EQ(sizeof(vector_type<custom_double_t, 32>), sizeof(vector_type<double, 32>));
+    ASSERT_EQ(sizeof(vector_type<custom_double_t, 64>), sizeof(vector_type<double, 64>));
+}
+
+TEST(Custom_double, TestAsType)
+{
+    struct custom_double_t
+    {
+        using type = double;
+        type data;
+        custom_double_t() : data{type{}} {}
+        custom_double_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size               = 4;
+    std::vector<double> test_vec = {0.3, 0.6, 0.8, 0.2};
+    // reference vector
+    vector_type<custom_double_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_double_t>()(Number<i>{}).data, 0.0);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_double_t>()(Number<i>{}) = custom_double_t{test_vec.at(i)};
+    });
+    // copy the vector
+    vector_type<custom_double_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_double_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Custom_double, TestAsTypeReshape)
+{
+    struct custom_double_t
+    {
+        using type = double;
+        type data;
+        custom_double_t() : data{type{}} {}
+        custom_double_t(type init) : data{init} {}
+    };
+
+    // test size
+    const int size               = 4;
+    std::vector<double> test_vec = {0.3, 0.6, 0.8, 0.2};
+    // reference vector
+    vector_type<custom_double_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<custom_double_t>()(Number<i>{}).data, 0.0);
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<custom_double_t>()(Number<i>{}) = custom_double_t{test_vec.at(i)};
+    });
+    // copy the first half of a vector
+    vector_type<custom_double_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<custom_double_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<custom_double_t>()(Number<i>{}).data, test_vec.at(i));
+    });
+}
+
+TEST(Complex_half, TestSize)
+{
+    struct complex_half_t
+    {
+        half_t real;
+        half_t img;
+    };
+    ASSERT_EQ(sizeof(complex_half_t), sizeof(half_t) + sizeof(half_t));
+    ASSERT_EQ(sizeof(vector_type<complex_half_t, 2>),
+              sizeof(vector_type<half_t, 2>) + sizeof(vector_type<half_t, 2>));
+    ASSERT_EQ(sizeof(vector_type<complex_half_t, 4>),
+              sizeof(vector_type<half_t, 4>) + sizeof(vector_type<half_t, 4>));
+    ASSERT_EQ(sizeof(vector_type<complex_half_t, 8>),
+              sizeof(vector_type<half_t, 8>) + sizeof(vector_type<half_t, 8>));
+    ASSERT_EQ(sizeof(vector_type<complex_half_t, 16>),
+              sizeof(vector_type<half_t, 16>) + sizeof(vector_type<half_t, 16>));
+    ASSERT_EQ(sizeof(vector_type<complex_half_t, 32>),
+              sizeof(vector_type<half_t, 32>) + sizeof(vector_type<half_t, 32>));
+    ASSERT_EQ(sizeof(vector_type<complex_half_t, 64>),
+              sizeof(vector_type<half_t, 64>) + sizeof(vector_type<half_t, 64>));
+}
+
+TEST(Complex_half, TestAlignment)
+{
+    struct complex_half_t
+    {
+        half_t real;
+        half_t img;
+    };
+    ASSERT_EQ(alignof(vector_type<complex_half_t, 2>),
+              alignof(vector_type<half_t, 2>) + alignof(vector_type<half_t, 2>));
+    ASSERT_EQ(alignof(vector_type<complex_half_t, 4>),
+              alignof(vector_type<half_t, 4>) + alignof(vector_type<half_t, 4>));
+    ASSERT_EQ(alignof(vector_type<complex_half_t, 8>),
+              alignof(vector_type<half_t, 8>) + alignof(vector_type<half_t, 8>));
+    ASSERT_EQ(alignof(vector_type<complex_half_t, 16>),
+              alignof(vector_type<half_t, 16>) + alignof(vector_type<half_t, 16>));
+    ASSERT_EQ(alignof(vector_type<complex_half_t, 32>),
+              alignof(vector_type<half_t, 32>) + alignof(vector_type<half_t, 32>));
+    ASSERT_EQ(alignof(vector_type<complex_half_t, 64>),
+              alignof(vector_type<half_t, 64>) + alignof(vector_type<half_t, 64>));
+}
+
+TEST(Complex_half, TestAsType)
+{
+    struct complex_half_t
+    {
+        using type = half_t;
+        type real;
+        type img;
+        complex_half_t() : real{type{}}, img{type{}} {}
+        complex_half_t(type real_init, type img_init) : real{real_init}, img{img_init} {}
+    };
+
+    // test size
+    const int size = 4;
+    // custom type number of elements
+    const int num_elem           = sizeof(complex_half_t) / sizeof(complex_half_t::type);
+    std::vector<half_t> test_vec = {half_t{0.3f},
+                                    half_t{-0.6f},
+                                    half_t{0.8f},
+                                    half_t{-0.2f},
+                                    half_t{0.5f},
+                                    half_t{-0.7f},
+                                    half_t{0.9f},
+                                    half_t{-0.3f}};
+    // reference vector
+    vector_type<complex_half_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<complex_half_t>()(Number<i>{}).real,
+                  type_convert<half_t>(0.0f));
+        ASSERT_EQ(right_vec.template AsType<complex_half_t>()(Number<i>{}).img,
+                  type_convert<half_t>(0.0f));
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<complex_half_t>()(Number<i>{}) =
+            complex_half_t{test_vec.at(num_elem * i), test_vec.at(num_elem * i + 1)};
+    });
+    // copy the vector
+    vector_type<complex_half_t, size> left_vec{right_vec};
+    // check if values were copied correctly
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).real,
+                  test_vec.at(num_elem * i));
+        ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).img,
+                  test_vec.at(num_elem * i + 1));
+    });
+}
+
+TEST(Complex_half, TestAsTypeReshape)
+{
+    struct complex_half_t
+    {
+        using type = half_t;
+        type real;
+        type img;
+        complex_half_t() : real{type{}}, img{type{}} {}
+        complex_half_t(type real_init, type img_init) : real{real_init}, img{img_init} {}
+    };
+
+    // test size
+    const int size = 4;
+    // custom type number of elements
+    const int num_elem           = sizeof(complex_half_t) / sizeof(complex_half_t::type);
+    std::vector<half_t> test_vec = {half_t{0.3f},
+                                    half_t{-0.6f},
+                                    half_t{0.8f},
+                                    half_t{-0.2f},
+                                    half_t{0.5f},
+                                    half_t{-0.7f},
+                                    half_t{0.9f},
+                                    half_t{-0.3f}};
+    // reference vector
+    vector_type<complex_half_t, size> right_vec;
+    // check default CTOR
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        ASSERT_EQ(right_vec.template AsType<complex_half_t>()(Number<i>{}).real,
+                  type_convert<half_t>(0.0f));
+        ASSERT_EQ(right_vec.template AsType<complex_half_t>()(Number<i>{}).img,
+                  type_convert<half_t>(0.0f));
+    });
+    // assign test values to the vector
+    ck::static_for<0, size, 1>{}([&](auto i) {
+        right_vec.template AsType<complex_half_t>()(Number<i>{}) =
+            complex_half_t{test_vec.at(num_elem * i), test_vec.at(num_elem * i + 1)};
+    });
+    // copy the first half of a vector
+    vector_type<complex_half_t, size / 2> left_vec{
+        right_vec.template AsType<vector_type<complex_half_t, size / 2>::type>()(Number<0>{})};
+    // check if values were copied correctly
+    ck::static_for<0, size / 2, 1>{}([&](auto i) {
+        ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).real,
+                  test_vec.at(num_elem * i));
+        ASSERT_EQ(left_vec.template AsType<complex_half_t>()(Number<i>{}).img,
+                  test_vec.at(num_elem * i + 1));
+    });
+}

test/gemm_universal/test_gemm_universal_xdl.cpp

@@ -56,7 +56,7 @@ class TestGemmUniversal_KM_NK
 using KernelTypes_MK_KN = ::testing::Types<
     //         ADataType, BDataType, ComputeDataType, CDataType
     std::tuple<      F16,       F16,             F16,     F16>,
-#if (defined CK_ENABLE_FP8)
+#if defined(CK_ENABLE_FP8) && defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
     std::tuple<      F16,        F8,             F16,     F16>,
     std::tuple<       F8,       F16,             F16,     F16>,
     std::tuple<       F8,        F8,              F8,    BF16>,
@@ -66,7 +66,7 @@ using KernelTypes_MK_KN = ::testing::Types<
 using KernelTypes_MK_NK = ::testing::Types<
     //         ADataType, BDataType, ComputeDataType, CDataType
     std::tuple<      F16,       F16,             F16,     F16>,
-#if (defined CK_ENABLE_FP8)
+#if defined(CK_ENABLE_FP8) && defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
     std::tuple<      F16,        F8,             F16,     F16>,
     std::tuple<       F8,       F16,             F16,     F16>,
     std::tuple<       F8,        F8,              F8,    BF16>,

test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp

@@ -58,13 +58,13 @@ using KernelTypes1d = ::testing::Types<std::tuple<float, GNWC, GKXC, GNWK>,
 using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWC, GKYXC, GNHWK>,
                                        std::tuple<ck::half_t, GNHWC, GKYXC, GNHWK>,
                                        std::tuple<ck::bhalf_t, GNHWC, GKYXC, GNHWK>,
-                                       std::tuple<int8_t, GNHWC, GKYXC, GNHWK>,
                                        std::tuple<float, NHWGC, GKYXC, NHWGK>,
                                        std::tuple<ck::half_t, NHWGC, GKYXC, NHWGK>,
                                        std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK>,
                                        std::tuple<int8_t, NHWGC, GKYXC, NHWGK>,
                                        std::tuple<float, NGCHW, GKYXC, NGKHW>,
-                                       std::tuple<ck::half_t, NGCHW, GKYXC, NGKHW>>;
+                                       std::tuple<ck::half_t, NGCHW, GKYXC, NGKHW>,
+                                       std::tuple<int8_t, NGCHW, GKYXC, NGKHW>>;
 
 using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWC, GKZYXC, GNDHWK>,
                                        std::tuple<ck::half_t, GNDHWC, GKZYXC, GNDHWK>,

test/grouped_convnd_fwd/test_grouped_convnd_fwd_large_cases_xdl.cpp

@@ -52,7 +52,8 @@ using namespace ck::tensor_layout::convolution;
 
 using KernelTypes2d = ::testing::Types<std::tuple<float, NHWGC, GKYXC, NHWGK>,
                                        std::tuple<ck::half_t, NHWGC, GKYXC, NHWGK>,
-                                       std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK>>;
+                                       std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK>,
+                                       std::tuple<int8_t, NHWGC, GKYXC, NHWGK>>;
 
 using KernelTypes3d = ::testing::Types<std::tuple<float, NDHWGC, GKZYXC, NDHWGK>,
                                        std::tuple<ck::half_t, NDHWGC, GKZYXC, NDHWGK>,

test/scatter_gather/CMakeLists.txt

+add_test_executable(test_scatter_gather scatter_gather.cpp)
+# target_compile_options(test_scatter_gather PRIVATE -v --save-temps -Wno-gnu-line-marker)

test/scatter_gather/scatter_gather.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+#include <iostream>
+#include <numeric>
+#include <cassert>
+#include <cstdlib>
+#include <iostream>
+#include <time.h>
+#include <unordered_set>
+
+#include "ck_tile/core.hpp"
+
+#ifndef TEST_SCATTER_GATHER_VERBOSE
+#define TEST_SCATTER_GATHER_VERBOSE 1
+#endif
+
+#define HIP_CALL(call)                                                              \
+    do                                                                              \
+    {                                                                               \
+        hipError_t err = call;                                                      \
+        if(err != hipSuccess)                                                       \
+        {                                                                           \
+            printf("[hiperror](%d) fail to call %s", static_cast<int>(err), #call); \
+            exit(0);                                                                \
+        }                                                                           \
+    } while(0)
+
+/*
+TODO:
+This is a simple design of scatter/gather through indexing transform, with limitations
+We may design a scatter/gather adaptor layer directly inside tile window
+*/
+template <ck_tile::index_t ROW_TILE_SIZE = 8,
+          ck_tile::index_t COL_TILE_SIZE = 32 * 8,
+          ck_tile::index_t BLOCK_SIZE    = 256,
+          ck_tile::index_t ALIGNMENT     = 8,
+          typename INDEX_BUF_TYPE        = ck_tile::index_t,
+          typename DATA_TYPE             = ck_tile::fp16_t>
+__global__ void row_scatter_gather(const INDEX_BUF_TYPE* src_row_idx_ptr,
+                                   const INDEX_BUF_TYPE* dst_row_idx_ptr,
+                                   const DATA_TYPE* src_ptr,
+                                   DATA_TYPE* dst_ptr,
+                                   ck_tile::index_t n_row_total,
+                                   ck_tile::index_t /*n_row_select*/,
+                                   ck_tile::index_t n_cols)
+{
+    using namespace ck_tile;
+
+    // some constexpr vars
+    constexpr index_t vec = ALIGNMENT;
+    static_assert(COL_TILE_SIZE % vec == 0);
+    constexpr index_t col_lanes = COL_TILE_SIZE / vec;
+    constexpr index_t warp_size = ck_tile::get_warp_size();
+    static_assert(warp_size % col_lanes == 0);
+    constexpr index_t row_lanes = warp_size / col_lanes;
+    constexpr index_t num_warps = BLOCK_SIZE / warp_size;
+    static_assert(ROW_TILE_SIZE % (num_warps * row_lanes) == 0);
+    constexpr index_t row_repeat = ROW_TILE_SIZE / (num_warps * row_lanes);
+    static_assert(
+        row_repeat == 1,
+        "currently indexing not support(and would be not performant) if row_repeat has more");
+
+    // tile partitioner
+    index_t tile_col_idx = 0;
+    index_t tile_row_idx = blockIdx.x * ROW_TILE_SIZE;
+
+    // create our tild distribution, which tell us the location of different threads
+    constexpr auto src_dist = make_static_tile_distribution(
+        tile_distribution_encoding<
+            sequence<1>,
+            tuple<sequence<row_repeat, num_warps, row_lanes>, sequence<col_lanes, vec>>,
+            tuple<sequence<1>, sequence<1, 2>>,
+            tuple<sequence<1>, sequence<2, 0>>,
+            sequence<1, 2>,
+            sequence<0, 1>>{});
+    const auto coord     = src_dist.calculate_index();
+    const auto row_coord = coord[number<0>{}] + tile_row_idx;
+
+    // load the current row index from the indexing buffer. we do not use ck_tile utility here
+    INDEX_BUF_TYPE src_row_id = src_row_idx_ptr[row_coord];
+    INDEX_BUF_TYPE dst_row_id = dst_row_idx_ptr[row_coord];
+
+    // printf("-- tid:%d, src_row_id:%d, dst_row_id:%d\n", static_cast<int>(threadIdx.x),
+    // static_cast<int>(src_row_id), static_cast<int>(dst_row_id));
+
+    const auto src_view =
+        make_naive_tensor_view<address_space_enum::global>(src_ptr,
+                                                           make_tuple(n_row_total, n_cols),
+                                                           make_tuple(n_cols, 1),
+                                                           number<vec>{}, // alignement
+                                                           number<1>{});
+
+    const auto src_gather_view = transform_tensor_view(
+        src_view,
+        make_tuple(make_indexing_transform(
+                       n_row_total,
+                       src_row_id), // here we replace row_idx  which is loaded from another buffer
+                   make_pass_through_transform(n_cols)),
+        make_tuple(sequence<0>{}, sequence<1>{}),
+        make_tuple(sequence<0>{}, sequence<1>{}));
+
+    auto src_tile = make_tile_window(src_gather_view,
+                                     make_tuple(number<ROW_TILE_SIZE>{}, number<COL_TILE_SIZE>{}),
+                                     {tile_row_idx, tile_col_idx},
+                                     src_dist);
+
+    const auto dst_view =
+        make_naive_tensor_view<address_space_enum::global>(dst_ptr,
+                                                           make_tuple(n_row_total, n_cols),
+                                                           make_tuple(n_cols, 1),
+                                                           number<vec>{},
+                                                           number<1>{});
+
+    const auto dst_scatter_view = transform_tensor_view(
+        dst_view,
+        make_tuple(make_indexing_transform(
+                       n_row_total,
+                       dst_row_id), // here we replace row_idx  which is loaded from another buffer
+                   make_pass_through_transform(n_cols)),
+        make_tuple(sequence<0>{}, sequence<1>{}),
+        make_tuple(sequence<0>{}, sequence<1>{}));
+
+    auto dst_tile = make_tile_window(dst_scatter_view,
+                                     make_tuple(number<ROW_TILE_SIZE>{}, number<COL_TILE_SIZE>{}),
+                                     {tile_row_idx, tile_col_idx},
+                                     src_dist /*reuse distribution*/);
+
+    // we finished descriptor construction and index calculation, now start load/store
+    for(auto i = 0; i < n_cols; i += COL_TILE_SIZE)
+    {
+        // note that scatter/gather are just the same API when doing load store as normal memory
+        // operation
+        auto data = load_tile(src_tile);
+        store_tile(dst_tile, data);
+
+        move_tile_window(src_tile, {number<0>{}, number<COL_TILE_SIZE>{}});
+        move_tile_window(dst_tile, {number<0>{}, number<COL_TILE_SIZE>{}});
+    }
+}
+
+union pixel
+{
+    struct __attribute__((packed))
+    {
+        unsigned int r : 6;
+        unsigned int c : 10;
+    };
+    ushort data;
+};
+
+struct unique_linear_rand
+{
+    unique_linear_rand(int capacity_) : capacity(capacity_) {}
+    std::unordered_set<int> set;
+    int gen()
+    {
+        if(static_cast<int>(set.size()) >= capacity)
+        {
+            printf("overflow, but will give you an number as well\n");
+            return std::rand() % capacity;
+        }
+        while(1)
+        {
+            int r = std::rand() % capacity;
+            if(set.count(r) == 1)
+            {
+                continue;
+            }
+            set.insert(r);
+            return r;
+        }
+    }
+
+    int capacity;
+};
+
+int main()
+{
+    int row_total  = 64;
+    int row_select = 8 * 2;
+    int col        = 256 * 2;
+    using fp16_t   = ck_tile::fp16_t;
+
+    constexpr int row_tile = 8;
+    constexpr int col_tile = 256;
+
+    fp16_t* src = reinterpret_cast<fp16_t*>(malloc(row_total * col * sizeof(fp16_t)));
+    for(int i_r = 0; i_r < row_total; i_r++)
+    {
+        for(int i_c = 0; i_c < col; i_c++)
+        {
+            int i = i_r * col + i_c;
+            pixel p;
+            p.r      = i_r;
+            p.c      = i_c;
+            ushort d = p.data;
+            src[i]   = ck_tile::bit_cast<fp16_t>(d); // for simplicity, just cast
+        }
+    }
+
+    fp16_t* dst  = reinterpret_cast<fp16_t*>(malloc(row_total * col * sizeof(fp16_t)));
+    int* src_idx = reinterpret_cast<int*>(malloc(row_select * sizeof(int)));
+    int* dst_idx = reinterpret_cast<int*>(malloc(row_select * sizeof(int)));
+    // std::srand(std::time(std::nullptr));
+    // std::srand(11935);
+    std::srand(std::time(nullptr));
+    auto src_gen = unique_linear_rand(row_total);
+    auto dst_gen = unique_linear_rand(row_total); // dst index must be unique. src is fine
+    for(int i_r = 0; i_r < row_select; i_r++)
+    {
+        src_idx[i_r] = src_gen.gen();
+        dst_idx[i_r] = dst_gen.gen();
+    }
+
+    void* dev_src;
+    void* dev_dst;
+    void* dev_src_idx;
+    void* dev_dst_idx;
+    HIP_CALL(hipMalloc(&dev_src, row_total * col * sizeof(fp16_t)));
+    HIP_CALL(hipMalloc(&dev_dst, row_total * col * sizeof(fp16_t)));
+    HIP_CALL(hipMalloc(&dev_src_idx, row_select * sizeof(int)));
+    HIP_CALL(hipMalloc(&dev_dst_idx, row_select * sizeof(int)));
+
+    HIP_CALL(hipMemcpy(dev_src, src, row_total * col * sizeof(fp16_t), hipMemcpyHostToDevice));
+    HIP_CALL(hipMemcpy(dev_src_idx, src_idx, row_select * sizeof(int), hipMemcpyHostToDevice));
+    HIP_CALL(hipMemcpy(dev_dst_idx, dst_idx, row_select * sizeof(int), hipMemcpyHostToDevice));
+
+    constexpr int bdim = 256;
+    int gdim           = (row_select + row_tile - 1) / row_tile;
+    row_scatter_gather<row_tile, col_tile><<<gdim, bdim>>>(reinterpret_cast<int*>(dev_src_idx),
+                                                           reinterpret_cast<int*>(dev_dst_idx),
+                                                           reinterpret_cast<fp16_t*>(dev_src),
+                                                           reinterpret_cast<fp16_t*>(dev_dst),
+                                                           row_total,
+                                                           row_select,
+                                                           col);
+
+    HIP_CALL(hipMemcpy(dst, dev_dst, row_total * col * sizeof(fp16_t), hipMemcpyDeviceToHost));
+
+#if TEST_SCATTER_GATHER_VERBOSE
+    printf("select row:");
+    for(int i_r = 0; i_r < row_select; i_r++)
+    {
+        printf("%d->%d->%d ", i_r, src_idx[i_r], dst_idx[i_r]);
+    }
+    printf("\n");
+#endif
+
+    int err_cnt = 0;
+    for(int i_r = 0; i_r < row_select; i_r++)
+    {
+        for(int i_c = 0; i_c < col; i_c++)
+        {
+            int i      = dst_idx[i_r] * col + i_c;
+            pixel p    = ck_tile::bit_cast<pixel>(dst[i]);
+            bool is_ok = p.r == src_idx[i_r] && p.c == i_c;
+            if(!is_ok)
+            {
+                if(i_c == 0)
+                    printf("(%d)pixel: %dx%d -> %d\n", i_r, p.r, p.c, dst_idx[i_r]);
+                err_cnt++;
+            }
+        }
+    }
+#if TEST_SCATTER_GATHER_VERBOSE
+    printf("err:%d\n", err_cnt);
+#endif
+
+    free(src);
+    free(dst);
+    free(src_idx);
+    free(dst_idx);
+    return err_cnt == 0 ? 0 : -1;
+}