Reorganize project folders (#6)

include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp

@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+// Assume
+//  1) SrcDesc is known at compile-time
+//  2) DstDesc is known at compile-time
+//  3) SrcBuffer is static buffer
+//  4) DstBuffer is static buffer
+template <typename AccDataType,
+          typename SrcThreadDesc_M_K,
+          typename DstThreadDesc_M,
+          typename OpReduce,
+          bool PropagateNan,
+          typename Accumulation =
+              detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
+struct ThreadwiseReduction
+{
+    static constexpr auto src_thread_desc_m_k = SrcThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    using Op = OpReduce;
+
+    template <typename SrcBufferType, typename DstBufferType>
+    __device__ static void Reduce(const SrcBufferType& src_buf, DstBufferType& dst_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            constexpr index_t out_offset = dst_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Accumulation::Calculate(dst_buf(Number<out_offset>{}), src_buf[Number<offset>{}]);
+            });
+        });
+    };
+};
+
+// Assume
+//  1) SrcDesc is known at compile-time
+//  2) DstDesc is known at compile-time
+//  3) SrcBuffer is static buffer
+//  4) DstBuffer is static buffer
+template <
+    typename AccDataType,
+    typename IndexDataType,
+    typename SrcThreadDesc_M_K,
+    typename DstThreadDesc_M,
+    typename OpReduce,
+    bool PropagateNan,
+    typename Accumulation =
+        detail::AccumulateWithIndexAndNanCheck<PropagateNan, OpReduce, AccDataType, IndexDataType>>
+struct ThreadwiseReductionWithIndex
+{
+    static constexpr auto src_thread_desc_m_k = SrcThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    template <typename SrcValueBufferType,
+              typename SrcIndexBufferType,
+              typename DstValueBufferType,
+              typename DstIndexBufferType>
+    __device__ static void Reduce(const SrcValueBufferType& src_val_buf,
+                                  const SrcIndexBufferType& src_idx_buf,
+                                  DstValueBufferType& dst_val_buf,
+                                  DstIndexBufferType& dst_idx_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            constexpr index_t out_offset = dst_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Accumulation::Calculate(dst_val_buf(Number<out_offset>{}),
+                                        src_val_buf[Number<offset>{}],
+                                        dst_idx_buf(Number<out_offset>{}),
+                                        src_idx_buf[Number<offset>{}]);
+            });
+        });
+    };
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp

@@ -0,0 +1,224 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+
+namespace ck {
+
+// C[TM0, TM1, TN0, TN1] += A[TK, TM0, TM1] * B[TK, TN0, TN1]
+//   Tensor element can be vectorized data
+// Assume:
+//   1. AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1, CThreadDesc_TM0_TM1_TN0_TN1 are
+//   known at compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_TK0_TM0_TM1_TK1,
+          typename BThreadDesc_TK0_TN0_TN1_TK1,
+          typename CThreadDesc_TM0_TM1_TN0_TN1,
+          typename TKLengths,
+          typename TMLengths,
+          typename TNLengths,
+          typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                                 BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                                 CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseGemmDl_km0m1_kn0n1_m0m1n0n1
+{
+    __device__ constexpr ThreadwiseGemmDl_km0m1_kn0n1_m0m1n0n1()
+    {
+        static_assert(AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                          BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                          CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: sanity-check: compare AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1,
+        // CThreadDesc_TM0_TM1_TN0_TN1 Size with KLenghts, TMLengths and TNLengths
+
+        // TODO remove this restriction
+        static_assert(TKLengths::Size() == 1 && TMLengths::Size() == 2 && TNLengths::Size() == 2,
+                      "wrong!");
+    }
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        constexpr auto TK  = TKLengths{}[I0];
+        constexpr auto TM0 = TMLengths{}[I0];
+        constexpr auto TM1 = TMLengths{}[I1];
+        constexpr auto TN0 = TNLengths{}[I0];
+        constexpr auto TN1 = TNLengths{}[I1];
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        static_for<0, TK, 1>{}([&](auto tk) {
+            static_for<0, TM0, 1>{}([&](auto tm0) {
+                static_for<0, TM1, 1>{}([&](auto tm1) {
+                    static_for<0, TN0, 1>{}([&](auto tn0) {
+                        static_for<0, TN1, 1>{}([&](auto tn1) {
+                            constexpr index_t a_offset =
+                                AThreadDesc_TK0_TM0_TM1_TK1{}.CalculateOffset(
+                                    a_origin_idx + make_multi_index(tk, tm0, tm1));
+                            constexpr index_t b_offset =
+                                BThreadDesc_TK0_TN0_TN1_TK1{}.CalculateOffset(
+                                    b_origin_idx + make_multi_index(tk, tn0, tn1));
+                            constexpr index_t c_offset =
+                                CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
+                                    c_origin_idx + make_multi_index(tm0, tm1, tn0, tn1));
+
+                            inner_product<FloatA, FloatB, FloatC>(a_buf[Number<a_offset>{}],
+                                                                  b_buf[Number<b_offset>{}],
+                                                                  c_buf(Number<c_offset>{}));
+                        });
+                    });
+                });
+            });
+        });
+    }
+};
+
+// C[TM0, TM1, TN0, TN1] += A[TK0, TM0, TM1, TK1] * B[TK0, TN0, TN1, TK1]
+//   Tensor element can be vectorized data
+// Assume:
+//   1. AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1, CThreadDesc_TM0_TM1_TN0_TN1 are
+//      known at compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_TK0_TM0_TM1_TK1,
+          typename BThreadDesc_TK0_TN0_TN1_TK1,
+          typename CThreadDesc_TM0_TM1_TN0_TN1,
+          typename TKLengths,
+          typename TMLengths,
+          typename TNLengths,
+          typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                                 BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                                 CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1
+{
+    __device__ constexpr ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1()
+    {
+        static_assert(AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                          BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                          CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: sanity-check: compare AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1,
+        // CThreadDesc_TM0_TM1_TN0_TN1 Size with KLenghts, TMLengths and TNLengths
+
+        // TODO remove this restriction
+        static_assert(TKLengths::Size() == 2 && TMLengths::Size() == 2 && TNLengths::Size() == 2,
+                      "wrong!");
+    }
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        constexpr index_t TK0 = TKLengths{}[I0];
+        constexpr index_t TK1 = TKLengths{}[I1];
+        constexpr index_t TM0 = TMLengths{}[I0];
+        constexpr index_t TM1 = TMLengths{}[I1];
+        constexpr index_t TN0 = TNLengths{}[I0];
+        constexpr index_t TN1 = TNLengths{}[I1];
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        static_for<0, TK0, 1>{}([&](auto tk0) {
+            static_for<0, TM0, 1>{}([&](auto tm0) {
+                static_for<0, TM1, 1>{}([&](auto tm1) {
+                    static_for<0, TN0, 1>{}([&](auto tn0) {
+                        static_for<0, TN1, 1>{}([&](auto tn1) {
+                            vector_type<FloatA, TK1> a_vec;
+                            vector_type<FloatB, TK1> b_vec;
+
+                            static_for<0, TK1, 1>{}([&](auto tk1) {
+                                constexpr index_t a_offset =
+                                    AThreadDesc_TK0_TM0_TM1_TK1{}.CalculateOffset(
+                                        a_origin_idx + make_multi_index(tk0, tm0, tm1, tk1));
+
+                                constexpr index_t b_offset =
+                                    BThreadDesc_TK0_TN0_TN1_TK1{}.CalculateOffset(
+                                        b_origin_idx + make_multi_index(tk0, tn0, tn1, tk1));
+
+                                a_vec.template AsType<FloatA>()(tk1) = a_buf[Number<a_offset>{}];
+                                b_vec.template AsType<FloatB>()(tk1) = b_buf[Number<b_offset>{}];
+                            });
+
+                            using a_vector_t = typename vector_type<FloatA, TK1>::type;
+                            using b_vector_t = typename vector_type<FloatB, TK1>::type;
+
+                            constexpr index_t c_offset =
+                                CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
+                                    c_origin_idx + make_multi_index(tm0, tm1, tn0, tn1));
+
+                            inner_product<a_vector_t, b_vector_t, FloatC>(
+                                a_vec.template AsType<a_vector_t>()[I0],
+                                b_vec.template AsType<b_vector_t>()[I0],
+                                c_buf(Number<c_offset>{}));
+                        });
+                    });
+                });
+            });
+        });
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp

@@ -0,0 +1,168 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_THREADWISE_GEMM_DLOPS_V3_HPP
+#define CK_THREADWISE_GEMM_DLOPS_V3_HPP
+
+#include "common_header.hpp"
+#include "math.hpp"
+
+namespace ck {
+
+// C[M, N] += transpose(A[K, M]) * B[K, N]
+//   Element of matrix can be vectorized data
+// Assume:
+//   1. AThreadDesc_E1_K_E2, BThreadDesc_E1_N_Ho_Wo_E2, CThreadDesc_K_N_Ho_Wo are known at
+//   compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_E1_K_E2,
+          typename BThreadDesc_E1_N_Ho_Wo_E2,
+          typename CThreadDesc_K_N_Ho_Wo,
+          typename enable_if<AThreadDesc_E1_K_E2::IsKnownAtCompileTime() &&
+                                 BThreadDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
+                                 CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseGemmDlops_km_kn_mn_v3
+{
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+
+        static_assert(AThreadDesc_E1_K_E2::IsKnownAtCompileTime() &&
+                          BThreadDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
+                          CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+        constexpr auto I2 = Number<2>{};
+        constexpr auto I3 = Number<3>{};
+
+        constexpr auto E1 = AThreadDesc_E1_K_E2{}.GetLength(I0);
+        constexpr auto K  = AThreadDesc_E1_K_E2{}.GetLength(I1);
+        constexpr auto E2 = AThreadDesc_E1_K_E2{}.GetLength(I2);
+
+        constexpr auto Ho = BThreadDesc_E1_N_Ho_Wo_E2{}.GetLength(I2);
+        constexpr auto Wo = BThreadDesc_E1_N_Ho_Wo_E2{}.GetLength(I3);
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        if constexpr((Ho % 2 == 0) && (Wo % 2 == 0))
+        {
+            constexpr auto SubHW = 2;
+
+            static_for<0, K, 1>{}([&](auto k) {
+                static_for<0, Ho, SubHW>{}([&](auto h) {
+                    static_for<0, Wo, SubHW>{}([&](auto w) {
+                        static_for<0, E1, 1>{}([&](auto e1) {
+                            static_for<0, E2, 1>{}([&](auto e2) {
+                                constexpr index_t a_offset = AThreadDesc_E1_K_E2{}.CalculateOffset(
+                                    a_origin_idx + make_tuple(e1, k, e2));
+
+                                constexpr index_t b0_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w, e2));
+
+                                constexpr index_t b1_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w + 1, e2));
+
+                                constexpr index_t b2_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h + 1, w, e2));
+
+                                constexpr index_t b3_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h + 1, w + 1, e2));
+
+                                constexpr index_t c0_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(c_origin_idx +
+                                                                            make_tuple(k, 0, h, w));
+
+                                constexpr index_t c1_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h, w + 1));
+
+                                constexpr index_t c2_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h + 1, w));
+
+                                constexpr index_t c3_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h + 1, w + 1));
+
+                                amd_assembly_outer_product_1x4(a_buf[Number<a_offset>{}],
+                                                               b_buf[Number<b0_offset>{}],
+                                                               b_buf[Number<b1_offset>{}],
+                                                               b_buf[Number<b2_offset>{}],
+                                                               b_buf[Number<b3_offset>{}],
+                                                               c_buf(Number<c0_offset>{}),
+                                                               c_buf(Number<c1_offset>{}),
+                                                               c_buf(Number<c2_offset>{}),
+                                                               c_buf(Number<c3_offset>{}));
+                            });
+                        });
+                    });
+                });
+            });
+        }
+        else
+        {
+
+            static_for<0, K, 1>{}([&](auto k) {
+                static_for<0, Ho, 1>{}([&](auto h) {
+                    static_for<0, Wo, 1>{}([&](auto w) {
+                        static_for<0, E1, 1>{}([&](auto e1) {
+                            static_for<0, E2, 1>{}([&](auto e2) {
+                                constexpr index_t a_offset = AThreadDesc_E1_K_E2{}.CalculateOffset(
+                                    a_origin_idx + make_tuple(e1, k, e2));
+
+                                constexpr index_t b_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w, e2));
+
+                                constexpr index_t c_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(c_origin_idx +
+                                                                            make_tuple(k, 0, h, w));
+
+                                inner_product<FloatA, FloatB, FloatC>(a_buf[Number<a_offset>{}],
+                                                                      b_buf[Number<b_offset>{}],
+                                                                      c_buf(Number<c_offset>{}));
+                            });
+                        });
+                    });
+                });
+            });
+        }
+    }
+};
+
+} // namespace ck
+#endif

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp

@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. Desc is known at compile-time
+//   2. Buffer is StaticBuffer
+//   3. OriginIdx is known at compile-time
+//   4. use #-step
+template <typename Data,
+          typename Desc,
+          typename SliceLengths,
+          typename enable_if<Desc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceSet_v1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    template <typename OriginIdx, typename Buffer>
+    __device__ void Run(const Desc&, const OriginIdx&, Buffer& buf, const Data& initial_value) const
+    {
+        static_assert(Desc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(Buffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<OriginIdx>>::value,
+                      "wrong! OriginIdx need to be known at compile-time");
+
+        // Desc is known at compile-time
+        constexpr auto desc = remove_cvref_t<Desc>{};
+
+        // OriginIdx is known at compile-time
+        constexpr auto origin_idx = to_multi_index(OriginIdx{});
+
+        static_ford<SliceLengths>{}([&](auto access_idx) {
+            constexpr auto coord = make_tensor_coordinate(desc, origin_idx + access_idx);
+
+            constexpr bool is_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(desc, coord);
+
+            constexpr index_t offset = coord.GetOffset();
+
+            if constexpr(is_valid)
+            {
+                buf(Number<offset>{}) = initial_value;
+            }
+        });
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp

@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+#pragma once
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+namespace detail {
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor
+template <index_t VectorDim, index_t ScalarPerVector>
+struct lambda_scalar_per_access
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        return (i == VectorDim) ? ScalarPerVector : 1;
+    }
+};
+
+template <index_t VectorDim>
+struct lambda_scalar_step_in_vector
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        return (i == VectorDim) ? 1 : 0;
+    }
+};
+
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor
+template <index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector>
+struct lambda_scalar_per_access_for_src_and_dst
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        if(i == SrcVectorDim && i == DstVectorDim)
+        {
+            return math::lcm(SrcScalarPerVector, DstScalarPerVector);
+        }
+        else if(i == SrcVectorDim)
+        {
+            return SrcScalarPerVector;
+        }
+        else if(i == DstVectorDim)
+        {
+            return DstScalarPerVector;
+        }
+        else
+        {
+            return 1;
+        }
+    }
+};
+
+} // namespace detail
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp

@@ -0,0 +1,972 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor/static_tensor.hpp"
+#include "ck/utility/is_detected.hpp"
+
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector_,
+          index_t DstScalarPerVector_,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun, // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v3r1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    static constexpr auto I0  = Number<0>{};
+    static constexpr auto I1  = Number<1>{};
+    static constexpr auto I2  = Number<2>{};
+    static constexpr auto I3  = Number<3>{};
+    static constexpr auto I4  = Number<4>{};
+    static constexpr auto I5  = Number<5>{};
+    static constexpr auto I6  = Number<6>{};
+    static constexpr auto I7  = Number<7>{};
+    static constexpr auto I8  = Number<8>{};
+    static constexpr auto I10 = Number<10>{};
+    static constexpr auto I12 = Number<12>{};
+    static constexpr auto I13 = Number<13>{};
+    static constexpr auto I14 = Number<14>{};
+    static constexpr auto I16 = Number<16>{};
+
+    static constexpr index_t PackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    static constexpr auto SrcScalarPerVector = Number<SrcScalarPerVector_ / PackedSize>{};
+    static constexpr auto DstScalarPerVector = Number<DstScalarPerVector_ / PackedSize>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3r1(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const DstElementwiseOperation& dst_element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          src_element_op_(src_element_op),
+          dst_element_op_(dst_element_op)
+    {
+        if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t>)
+        {
+            static_assert(is_same_v<remove_cvref_t<SrcData>, remove_cvref_t<DstData>>,
+                          "SrcData != DstData");
+
+            static_assert(
+                SrcScalarPerVector_ % PackedSize == 0 && DstScalarPerVector_ % PackedSize == 0,
+                "SrcScalarPerVector_ and DstScalarPerVector_ cannot be 1 for packed data type");
+
+            static_assert(SrcVectorDim == DstVectorDim, "pk_i4_t does not support transpose");
+        }
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDesc& src_desc,
+                            const SrcBuffer& src_buf,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        static_assert(SliceLengths::At(SrcVectorDim) % (SrcScalarPerVector_) == 0,
+                      "SliceLengths[SrcVectorDim] must be divisible by SrcScalarPerVector");
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            constexpr auto src_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
+
+            // maintain a container record is_src_valid, waiting for RunWrite use.
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+            src_oob_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<bool>(src_data_idx_seq, is_src_valid);
+
+            using dst_vector_type = vector_type_maker_t<DstData, SrcScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+            dst_vector_type op_r_v;
+
+            constexpr auto get_elem_op_vec_len = []() {
+                if constexpr(is_detected<is_pack8_invocable_t, decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack8_invocable)
+                        return math::min(8, SrcScalarPerVector);
+                }
+                else if constexpr(is_detected<is_pack4_invocable_t,
+                                              decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack4_invocable)
+                        return math::min(4, SrcScalarPerVector);
+                }
+                else if constexpr(is_detected<is_pack2_invocable_t,
+                                              decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack2_invocable)
+                        return math::min(2, SrcScalarPerVector);
+                }
+                else
+                {
+                    return 1;
+                }
+            };
+
+            constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
+
+            using src_elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
+            using dst_elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
+
+            using VectorSizeLookupTable    = Tuple<Sequence<>,
+                                                Sequence<I1>,
+                                                Sequence<I2>,
+                                                Sequence<I2, I1>,
+                                                Sequence<I4>,
+                                                Sequence<I4, I1>,
+                                                Sequence<I4, I2>,
+                                                Sequence<I4, I2, I1>,
+                                                Sequence<I8>,
+                                                Sequence<I8, I1>,
+                                                Sequence<I8, I2>,
+                                                Sequence<I8, I2, I1>,
+                                                Sequence<I8, I4>,
+                                                Sequence<I8, I4, I1>,
+                                                Sequence<I8, I4, I2>,
+                                                Sequence<I8, I4, I2, I1>,
+                                                Sequence<I16>>;
+            using VectorOffsetsLookupTable = Tuple<Sequence<>,
+                                                   Sequence<I0>,
+                                                   Sequence<I0>,
+                                                   Sequence<I0, I2>,
+                                                   Sequence<I0>,
+                                                   Sequence<I0, I4>,
+                                                   Sequence<I0, I4>,
+                                                   Sequence<I0, I4, I6>,
+                                                   Sequence<I0>,
+                                                   Sequence<I0, I8>,
+                                                   Sequence<I0, I8>,
+                                                   Sequence<I0, I8, I10>,
+                                                   Sequence<I0, I8>,
+                                                   Sequence<I0, I8, I12>,
+                                                   Sequence<I0, I8, I12>,
+                                                   Sequence<I0, I8, I12, I14>,
+                                                   Sequence<I0>>;
+
+            static_for<0, tuple_element_t<SrcScalarPerVector, VectorSizeLookupTable>::Size(), 1>{}(
+                [&](auto v_idx) {
+                    constexpr auto VectorLoadSize =
+                        tuple_element_t<SrcScalarPerVector, VectorSizeLookupTable>::At(v_idx);
+                    constexpr auto LoadOffset =
+                        tuple_element_t<SrcScalarPerVector, VectorOffsetsLookupTable>::At(v_idx);
+
+                    using src_vector_container   = vector_type_maker_t<SrcData, VectorLoadSize>;
+                    using src_vector_container_t = typename src_vector_container::type;
+
+                    src_vector_container src_vector =
+                        src_vector_container{src_buf.template Get<src_vector_container_t>(
+                            src_coord_.GetOffset() / PackedSize + LoadOffset, true)};
+
+                    static_for<0, VectorLoadSize / elem_op_vec_len, 1>{}([&](auto idx) {
+                        // apply the src elementwise op and convert to DstData under the hood if
+                        // needed
+                        src_element_op_(
+                            op_r_v.template AsType<dst_elem_op_vec_t>()(idx + LoadOffset),
+                            src_vector.template AsType<src_elem_op_vec_t>()[idx]);
+                    });
+                });
+
+            // copy data from src_vector_container into src_thread_scratch_
+            src_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<dst_vector_t>(src_data_idx_seq,
+                                                  op_r_v.template AsType<dst_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move src coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <typename SeqIdx, index_t ThreadScratchId = 0>
+    __device__ constexpr auto
+    GetSrcThreadScratchIdx(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        using vector_t = typename vector_type_maker<SrcData, SrcScalarPerVector>::type::type;
+        return src_thread_scratch_tuple_(thread_scratch_id).template GetAsType<vector_t>(SeqIdx{});
+    }
+
+    template <index_t ThreadScratchId>
+    __device__ void
+    TransferDataFromSrcThreadScratchToDstThreadScratch(Number<ThreadScratchId> thread_scratch_id)
+    {
+#if !CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
+        static_ford<SliceLengths>{}([&](auto idx) {
+            dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
+        });
+#else
+        // OOB Check
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            constexpr auto src_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
+
+            using vector_t = typename vector_type_maker<DstData, SrcScalarPerVector>::type::type;
+
+            auto op_r = src_thread_scratch_tuple_(thread_scratch_id)
+                            .template GetAsType<vector_t>(src_data_idx_seq);
+
+            const bool is_src_valid = src_oob_thread_scratch_tuple_(thread_scratch_id)
+                                          .template GetAsType<bool>(src_data_idx_seq);
+
+            auto op_r_v = is_src_valid ? op_r : vector_t(0);
+
+            src_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<vector_t>(src_data_idx_seq, op_r_v);
+        });
+
+        // sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
+        // TODO make this logic more generic for more sub-dword datatype
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) ||
+                      (is_same<f8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
+        {
+            static_assert(!is_same_v<remove_cvref_t<SrcData>, pk_i4_t>,
+                          "in-register transpose is not supported for pk_i4_t");
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+            static_assert(SrcVectorDim != DstVectorDim, "wrong");
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst<SrcVectorDim,
+                                                                 SrcScalarPerVector,
+                                                                 DstVectorDim,
+                                                                 DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                using src_vector_t = vector_type_maker_t<DstData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return src_thread_scratch_tuple_[thread_scratch_id].GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                transpose_vectors<DstData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+        else
+        {
+            constexpr auto packed_per_access = generate_sequence(
+                detail::lambda_scalar_per_access<SrcVectorDim, PackedSize>{}, Number<nDim>{});
+
+            constexpr auto packed_access_lengths = SliceLengths{} / packed_per_access;
+
+            static_ford<decltype(packed_access_lengths)>{}([&](auto idx) {
+                dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
+            });
+        }
+#endif
+    }
+
+    template <typename DstBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // if there is transpose, it's done here
+        // if there is oob check, it's done here
+        // TODO move this elsewhere
+        TransferDataFromSrcThreadScratchToDstThreadScratch(thread_scratch_id);
+
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        // src scalar per access on each dim
+        // TODO: don't use this
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_scalar_per_access;
+            }();
+
+            constexpr auto dst_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<dst_data_idx[i]>{}; }, Number<dst_data_idx.Size()>{});
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            using dst_vector_type = vector_type_maker_t<DstData, DstScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            // copy data from dst_thread_scratch_ into dst_vector_container
+            auto dst_vector_container = dst_vector_type{
+                dst_thread_scratch_.template GetAsType<dst_vector_t>(dst_data_idx_seq)};
+
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                DstData dst_v;
+
+                // apply DstElementwiseOperation
+                dst_element_op_(dst_v, dst_vector_container.template AsType<DstData>()[i]);
+            });
+
+            // copy data from dst_vector_container to dst_buf
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset() / PackedSize,
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move dst coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetSrcOOBThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        return make_naive_tensor_descriptor_packed(src_access_lengths);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    private:
+    static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto src_oob_thread_scratch_desc_ =
+        decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
+
+    using SrcThreadScratch =
+        StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                        DstData, // apply data_convert with SrcThreadScratch
+                                        SrcScalarPerVector,
+                                        decltype(src_thread_scratch_desc_),
+                                        true>;
+
+    using SrcOOBThreadScratch =
+        StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                        bool, // apply data_convert with SrcThreadScratch
+                                        1,
+                                        decltype(src_oob_thread_scratch_desc_),
+                                        true>;
+
+    using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                             DstData,
+                                                             DstScalarPerVector,
+                                                             decltype(dst_thread_scratch_desc_),
+                                                             true>;
+
+    StaticallyIndexedArray<SrcThreadScratch, NumThreadScratch> src_thread_scratch_tuple_;
+    StaticallyIndexedArray<SrcOOBThreadScratch, NumThreadScratch> src_oob_thread_scratch_tuple_;
+
+    DstThreadScratch dst_thread_scratch_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const SrcElementwiseOperation src_element_op_;
+    const DstElementwiseOperation dst_element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1_gather.hpp

@@ -0,0 +1,942 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor/static_tensor.hpp"
+#include "ck/utility/is_detected.hpp"
+
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector_,
+          index_t DstScalarPerVector_,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun, // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+          typename IndexType,
+          index_t GatherDim        = 1,
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v3r1_gather
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    static constexpr auto I0  = Number<0>{};
+    static constexpr auto I1  = Number<1>{};
+    static constexpr auto I2  = Number<2>{};
+    static constexpr auto I3  = Number<3>{};
+    static constexpr auto I4  = Number<4>{};
+    static constexpr auto I5  = Number<5>{};
+    static constexpr auto I6  = Number<6>{};
+    static constexpr auto I7  = Number<7>{};
+    static constexpr auto I8  = Number<8>{};
+    static constexpr auto I10 = Number<10>{};
+    static constexpr auto I12 = Number<12>{};
+    static constexpr auto I13 = Number<13>{};
+    static constexpr auto I14 = Number<14>{};
+    static constexpr auto I16 = Number<16>{};
+
+    static constexpr index_t PackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    static constexpr auto SrcScalarPerVector = Number<SrcScalarPerVector_ / PackedSize>{};
+    static constexpr auto DstScalarPerVector = Number<DstScalarPerVector_ / PackedSize>{};
+
+    static constexpr index_t gather_num = SliceLengths{}.At(Number<GatherDim>{});
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3r1_gather(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const DstElementwiseOperation& dst_element_op,
+        const StaticallyIndexedArray<IndexType, gather_num>& gather_offsets)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          src_element_op_(src_element_op),
+          dst_element_op_(dst_element_op),
+          gather_offsets_(gather_offsets)
+    {
+        if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t>)
+        {
+            static_assert(is_same_v<remove_cvref_t<SrcData>, remove_cvref_t<DstData>>,
+                          "SrcData != DstData");
+
+            static_assert(
+                SrcScalarPerVector_ % PackedSize == 0 && DstScalarPerVector_ % PackedSize == 0,
+                "SrcScalarPerVector_ and DstScalarPerVector_ cannot be 1 for packed data type");
+
+            static_assert(SrcVectorDim == DstVectorDim, "pk_i4_t does not support transpose");
+        }
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+
+        auto adjusted_origin_idx = [&]() {
+            Index idx;
+            static_for<0, nDim, 1>{}([&](auto i) {
+                idx(i) = i.value == GatherDim ? 0 : src_slice_origin_idx[Number<i>{}];
+            });
+            return idx;
+        }();
+        src_coord_ = make_tensor_coordinate(src_desc, adjusted_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDesc& src_desc,
+                            const SrcBuffer& src_buf,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        static_assert(SliceLengths::At(SrcVectorDim) % (SrcScalarPerVector_) == 0,
+                      "SliceLengths[SrcVectorDim] must be divisible by SrcScalarPerVector");
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+        constexpr auto ordered_gather_dim   = src_dim_access_order[GatherDim];
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            constexpr auto src_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
+
+            auto gather_offset =
+                gather_offsets_(ordered_src_access_idx[Number<ordered_gather_dim>{}]);
+
+            const IndexType ld_offset = src_coord_.GetOffset() + gather_offset;
+            src_oob_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<bool>(src_data_idx_seq, true);
+
+            using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            auto src_vector_container =
+                src_vector_type{src_buf.template Get<src_vector_t>(ld_offset, true)};
+
+            using dst_vector_type = vector_type_maker_t<DstData, SrcScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+            dst_vector_type op_r_v;
+
+            constexpr auto get_elem_op_vec_len = []() {
+                if constexpr(is_detected<is_pack8_invocable_t, decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack8_invocable)
+                        return math::min(8, SrcScalarPerVector);
+                }
+                else if constexpr(is_detected<is_pack4_invocable_t,
+                                              decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack4_invocable)
+                        return math::min(4, SrcScalarPerVector);
+                }
+                else if constexpr(is_detected<is_pack2_invocable_t,
+                                              decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack2_invocable)
+                        return math::min(2, SrcScalarPerVector);
+                }
+                else
+                {
+                    return 1;
+                }
+            };
+
+            constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
+
+            using src_elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
+            using dst_elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
+
+            static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto idx) {
+                // apply the src elementwise op and convert to DstData under the hood if needed
+                src_element_op_(op_r_v.template AsType<dst_elem_op_vec_t>()(idx),
+                                src_vector_container.template AsType<src_elem_op_vec_t>()[idx]);
+            });
+
+            // copy data from src_vector_container into src_thread_scratch_
+            src_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<dst_vector_t>(src_data_idx_seq,
+                                                  op_r_v.template AsType<dst_vector_t>()[I0]);
+
+            auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                    move_on_dim_(i) &= i.value != ordered_gather_dim;
+                });
+
+                return move_on_dim_;
+            }
+            ();
+            // move src coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <typename SeqIdx, index_t ThreadScratchId = 0>
+    __device__ constexpr auto
+    GetSrcThreadScratchIdx(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        using vector_t = typename vector_type_maker<SrcData, SrcScalarPerVector>::type::type;
+        return src_thread_scratch_tuple_(thread_scratch_id).template GetAsType<vector_t>(SeqIdx{});
+    }
+
+    template <index_t ThreadScratchId>
+    __device__ void
+    TransferDataFromSrcThreadScratchToDstThreadScratch(Number<ThreadScratchId> thread_scratch_id)
+    {
+#if !CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
+        static_ford<SliceLengths>{}([&](auto idx) {
+            dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
+        });
+#else
+
+        // OOB Check
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            constexpr auto src_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
+
+            using vector_t = typename vector_type_maker<DstData, SrcScalarPerVector>::type::type;
+
+            auto op_r = src_thread_scratch_tuple_(thread_scratch_id)
+                            .template GetAsType<vector_t>(src_data_idx_seq);
+
+            auto op_r_v = op_r;
+
+            src_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<vector_t>(src_data_idx_seq, op_r_v);
+        });
+
+        // sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
+        // TODO make this logic more generic for more sub-dword datatype
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) ||
+                      (is_same<f8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
+        {
+            static_assert(!is_same_v<remove_cvref_t<SrcData>, pk_i4_t>,
+                          "in-register transpose is not supported for pk_i4_t");
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+            static_assert(SrcVectorDim != DstVectorDim, "wrong");
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst<SrcVectorDim,
+                                                                 SrcScalarPerVector,
+                                                                 DstVectorDim,
+                                                                 DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                using src_vector_t = vector_type_maker_t<DstData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return src_thread_scratch_tuple_[thread_scratch_id].GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                transpose_vectors<DstData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+        else
+        {
+            constexpr auto packed_per_access = generate_sequence(
+                detail::lambda_scalar_per_access<SrcVectorDim, PackedSize>{}, Number<nDim>{});
+
+            constexpr auto packed_access_lengths = SliceLengths{} / packed_per_access;
+
+            static_ford<decltype(packed_access_lengths)>{}([&](auto idx) {
+                dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
+            });
+        }
+#endif
+    }
+
+    template <typename DstBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // if there is transpose, it's done here
+        // if there is oob check, it's done here
+        // TODO move this elsewhere
+        TransferDataFromSrcThreadScratchToDstThreadScratch(thread_scratch_id);
+
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        // src scalar per access on each dim
+        // TODO: don't use this
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_scalar_per_access;
+            }();
+
+            constexpr auto dst_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<dst_data_idx[i]>{}; }, Number<dst_data_idx.Size()>{});
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            using dst_vector_type = vector_type_maker_t<DstData, DstScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            // copy data from dst_thread_scratch_ into dst_vector_container
+            auto dst_vector_container = dst_vector_type{
+                dst_thread_scratch_.template GetAsType<dst_vector_t>(dst_data_idx_seq)};
+
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                DstData dst_v;
+
+                // apply DstElementwiseOperation
+                dst_element_op_(dst_v, dst_vector_container.template AsType<DstData>()[i]);
+
+                dst_vector_container.template AsType<DstData>()(i) = dst_v;
+            });
+
+            // copy data from dst_vector_container to dst_buf
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset() / PackedSize,
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move dst coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                reset_src_data_step_(i) = i.value == GatherDim ? 0 : -src_data_idx[i];
+            });
+
+            return reset_src_data_step_;
+        }();
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetSrcOOBThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        return make_naive_tensor_descriptor_packed(src_access_lengths);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector_>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    private:
+    static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto src_oob_thread_scratch_desc_ =
+        decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
+
+    using SrcThreadScratch =
+        StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                        DstData, // apply data_convert with SrcThreadScratch
+                                        SrcScalarPerVector,
+                                        decltype(src_thread_scratch_desc_),
+                                        true>;
+
+    using SrcOOBThreadScratch =
+        StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                        bool, // apply data_convert with SrcThreadScratch
+                                        1,
+                                        decltype(src_oob_thread_scratch_desc_),
+                                        true>;
+
+    using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                             DstData,
+                                                             DstScalarPerVector,
+                                                             decltype(dst_thread_scratch_desc_),
+                                                             true>;
+
+    StaticallyIndexedArray<SrcThreadScratch, NumThreadScratch> src_thread_scratch_tuple_;
+    StaticallyIndexedArray<SrcOOBThreadScratch, NumThreadScratch> src_oob_thread_scratch_tuple_;
+
+    DstThreadScratch dst_thread_scratch_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const SrcElementwiseOperation src_element_op_;
+    const DstElementwiseOperation dst_element_op_;
+    StaticallyIndexedArray<IndexType, gather_num> gather_offsets_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r2.hpp

@@ -0,0 +1,804 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor/static_tensor.hpp"
+#include "ck/utility/is_detected.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence
+          typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          typename SrcsScalarPerVector,         // Sequence
+          typename DstsScalarPerVector,         // Sequence
+          typename SrcsScalarStrideInVector,    // Sequence
+          typename DstsScalarStrideInVector,    // Sequence
+          typename SrcsResetCoordinateAfterRun, // control whether to move back src coordinate after
+                                                // each RunRead(),  will be fused with
+                                                // MoveSrcSliceWindow to save addr computation
+          typename DstsResetCoordinateAfterRun, // control whether to move back dst coordinate after
+                                                // each RunWrite(),  will be fused with
+                                                // MoveDstSliceWindow to save addr computation
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v3r2
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3r2(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto src_i) {
+            src_coords_(src_i) =
+                make_tensor_coordinate(src_descs.At(src_i), src_slice_origin_idxs[src_i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto dst_i) {
+            dst_coords_(dst_i) =
+                make_tensor_coordinate(dst_descs.At(dst_i), dst_slice_origin_idxs[dst_i]);
+        });
+    }
+
+    template <typename SrcBuffers, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDescs& src_descs,
+                            const SrcBuffers& src_bufs,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access_tuple = generate_tuple(
+            [&](auto src_i) {
+                return generate_sequence(
+                    detail::lambda_scalar_per_access<SrcVectorDim,
+                                                     SrcsScalarPerVector::At(src_i)>{},
+                    Number<nDim>{});
+            },
+            Number<nSrc>{});
+
+        constexpr auto src_access_lengths_tuple = generate_tuple(
+            [&](auto src_i) {
+                return SliceLengths{} / src_scalar_per_access_tuple.At(src_i);
+                static_assert(
+                    SliceLengths::At(SrcVectorDim) % SrcsScalarPerVector::At(src_i) == 0,
+                    "SliceLengths[SrcVectorDim] must be divisible by SrcsScalarPerVector");
+            },
+            Number<nSrc>{});
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths_tuple = generate_tuple(
+            [&](auto src_i) {
+                return container_reorder_given_new2old(src_access_lengths_tuple.At(src_i),
+                                                       src_dim_access_order);
+            },
+            Number<nSrc>{});
+
+        // make forward steps
+        const auto src_forward_steps_tuple = generate_tuple(
+            [&](auto src_i) {
+                return generate_tuple(
+                    [&](auto i) {
+                        Index forward_step_idx;
+
+                        static_for<0, nDim, 1>{}([&](auto j) {
+                            forward_step_idx(j) =
+                                (i.value == j.value) ? src_scalar_per_access_tuple.At(src_i)[i] : 0;
+                        });
+
+                        return make_tensor_coordinate_step(src_descs.At(src_i), forward_step_idx);
+                    },
+                    Number<nDim>{});
+            },
+            Number<nSrc>{});
+
+        // make backward steps
+        const auto src_backward_steps_tuple = generate_tuple(
+            [&](auto src_i) {
+                return generate_tuple(
+                    [&](auto i) {
+                        Index backward_step_idx;
+
+                        static_for<0, nDim, 1>{}([&](auto j) {
+                            backward_step_idx(j) = (i.value == j.value)
+                                                       ? -src_scalar_per_access_tuple.At(src_i)[i]
+                                                       : 0;
+                        });
+
+                        return make_tensor_coordinate_step(src_descs.At(src_i), backward_step_idx);
+                    },
+                    Number<nDim>{});
+            },
+            Number<nSrc>{});
+
+        // loop over tensor and copy
+        static_for<0, nSrc, 1>{}([&](auto src_i) {
+            static_ford<remove_cvref_t<decltype(ordered_src_access_lengths_tuple.At(src_i))>>{}(
+                [&](auto ordered_src_access_idx) {
+                    // judge move forward or move backward
+                    constexpr auto forward_sweep = [&]() {
+                        StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                        forward_sweep_(I0) = true;
+
+                        static_for<1, nDim, 1>{}([&](auto i) {
+                            index_t tmp = ordered_src_access_idx[I0];
+
+                            static_for<1, i, 1>{}([&](auto j) {
+                                tmp = tmp * ordered_src_access_lengths_tuple[j] +
+                                      ordered_src_access_idx[j];
+                            });
+
+                            forward_sweep_(i) = tmp % 2 == 0;
+                        });
+
+                        return forward_sweep_;
+                    }();
+
+                    // calculate src data index
+                    constexpr auto src_data_idx = [&]() {
+                        Index ordered_idx;
+
+                        static_for<0, nDim, 1>{}([&](auto i) {
+                            ordered_idx(i) = forward_sweep[i]
+                                                 ? ordered_src_access_idx[i]
+                                                 : ordered_src_access_lengths_tuple.At(src_i)[i] -
+                                                       1 - ordered_src_access_idx[i];
+                        });
+
+                        return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                               src_scalar_per_access_tuple.At(src_i);
+                    }();
+
+                    constexpr auto src_data_idx_seq =
+                        generate_sequence_v2([&](auto i) { return Number<src_data_idx[i]>{}; },
+                                             Number<src_data_idx.Size()>{});
+
+                    const bool is_src_valid =
+                        coordinate_has_valid_offset_assuming_visible_index_is_valid(
+                            src_descs.At(src_i), src_coords_.At(src_i));
+
+                    using src_vector_type = vector_type_maker_t<tuple_element_t<src_i, SrcDatas>,
+                                                                SrcsScalarPerVector::At(src_i)>;
+                    using src_vector_t    = typename src_vector_type::type;
+
+                    // copy data from src_buf into src_vector_container
+                    auto src_vector_container =
+                        src_vector_type{src_bufs.At(src_i).template Get<src_vector_t>(
+                            src_coords_.At(src_i).GetOffset(), is_src_valid)};
+
+                    // copy data from src_vector_container into src_thread_scratch_
+                    src_thread_scratch_tuple_(thread_scratch_id)
+                        .At(src_i)
+                        .template SetAsType<src_vector_t>(
+                            src_data_idx_seq,
+                            src_vector_container.template AsType<src_vector_t>()[I0]);
+
+                    constexpr auto move_on_dim = [&]() constexpr
+                    {
+                        StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                        static_for<0, nDim, 1>{}([&](auto i) {
+                            move_on_dim_(i) = ordered_src_access_idx[i] <
+                                              ordered_src_access_lengths_tuple.At(src_i)[i] - 1;
+
+                            static_for<i + 1, nDim, 1>{}([&](auto j) {
+                                move_on_dim_(i) &=
+                                    ordered_src_access_idx[j] ==
+                                    ordered_src_access_lengths_tuple.At(src_i)[j] - 1;
+                            });
+                        });
+
+                        return move_on_dim_;
+                    }
+                    ();
+
+                    // move src coord
+                    static_for<0, nDim, 1>{}([&](auto i) {
+                        if constexpr(move_on_dim[i])
+                        {
+                            if constexpr(forward_sweep[i])
+                            {
+                                move_tensor_coordinate(
+                                    src_descs.At(src_i),
+                                    src_coords_.At(src_i),
+                                    src_forward_steps_tuple.At(src_i)[src_dim_access_order[i]]);
+                            }
+                            else
+                            {
+                                move_tensor_coordinate(
+                                    src_descs.At(src_i),
+                                    src_coords_.At(src_i),
+                                    src_backward_steps_tuple.At(src_i)[src_dim_access_order[i]]);
+                            }
+                        }
+                    });
+                });
+        });
+
+        static_for<0, nSrc, 1>{}([&](auto src_i) {
+            // move src coordinate back to slice origin (or not)
+            if constexpr(SrcsResetCoordinateAfterRun::At(src_i))
+            {
+                const auto src_reset_step = make_tensor_coordinate_step(
+                    src_descs.At(src_i), GetSrcCoordinateResetStep<src_i>());
+
+                move_tensor_coordinate(src_descs.At(src_i), src_coords_.At(src_i), src_reset_step);
+            }
+        });
+    }
+
+    template <index_t ThreadScratchId>
+    __device__ void
+    TransferDataFromSrcThreadScratchToDstThreadScratch(Number<ThreadScratchId> thread_scratch_id)
+    {
+        // TODO: Add support for CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
+        // (it requires to add Elementwise support in transpose_vectors)
+        static_ford<SliceLengths>{}([&](auto idx) {
+            const auto src_data_refs = generate_tie(
+                [&](auto src_i) -> const auto& {
+                    return src_thread_scratch_tuple_[thread_scratch_id].At(src_i)[idx];
+                },
+                Number<nSrc>{});
+
+            auto dst_data_refs = generate_tie(
+                [&](auto dst_i) -> auto& { return dst_thread_scratch_tuple_.At(dst_i)(idx); },
+                Number<nDst>{});
+            unpack2(element_op_, dst_data_refs, src_data_refs);
+        });
+    }
+
+    template <typename DstBuffers, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDescs& dst_descs,
+                             DstBuffers& dst_bufs,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // if there is transpose, it's done here
+        // TODO move this elsewhere
+        TransferDataFromSrcThreadScratchToDstThreadScratch(thread_scratch_id);
+
+        // src scalar per access on each dim
+        // TODO: don't use this
+        constexpr auto dst_scalar_per_access_tuple = generate_tuple(
+            [&](auto dst_i) {
+                return generate_sequence(
+                    detail::lambda_scalar_per_access<DstVectorDim,
+                                                     DstsScalarPerVector::At(dst_i)>{},
+                    Number<nDim>{});
+            },
+            Number<nDst>{});
+
+        constexpr auto dst_access_lengths_tuple = generate_tuple(
+            [&](auto dst_i) { return SliceLengths{} / dst_scalar_per_access_tuple.At(dst_i); },
+            Number<nDst>{});
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths_tuple = generate_tuple(
+            [&](auto dst_i) {
+                return container_reorder_given_new2old(dst_access_lengths_tuple.At(dst_i),
+                                                       dst_dim_access_order);
+            },
+            Number<nDst>{});
+
+        // make forward steps
+        const auto dst_forward_steps_tuple = generate_tuple(
+            [&](auto dst_i) {
+                return generate_tuple(
+                    [&](auto i) {
+                        Index forward_step_idx;
+
+                        static_for<0, nDim, 1>{}([&](auto j) {
+                            forward_step_idx(j) =
+                                (i.value == j.value) ? dst_scalar_per_access_tuple.At(dst_i)[i] : 0;
+                        });
+
+                        return make_tensor_coordinate_step(dst_descs.At(dst_i), forward_step_idx);
+                    },
+                    Number<nDim>{});
+            },
+            Number<nDst>{});
+
+        // make backward steps
+        const auto dst_backward_steps_tuple = generate_tuple(
+            [&](auto dst_i) {
+                return generate_tuple(
+                    [&](auto i) {
+                        Index backward_step_idx;
+
+                        static_for<0, nDim, 1>{}([&](auto j) {
+                            backward_step_idx(j) = (i.value == j.value)
+                                                       ? -dst_scalar_per_access_tuple.At(dst_i)[i]
+                                                       : 0;
+                        });
+
+                        return make_tensor_coordinate_step(dst_descs.At(dst_i), backward_step_idx);
+                    },
+                    Number<nDim>{});
+            },
+            Number<nDst>{});
+
+        // loop over tensor and copy
+        static_for<0, nDst, 1>{}([&](auto dst_i) {
+            static_ford<remove_cvref_t<decltype(ordered_dst_access_lengths_tuple.At(dst_i))>>{}(
+                [&](auto ordered_dst_access_idx) {
+                    // judge move forward or move backward
+                    constexpr auto forward_sweep = [&]() {
+                        StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                        forward_sweep_(I0) = true;
+
+                        static_for<1, nDim, 1>{}([&](auto i) {
+                            index_t tmp = ordered_dst_access_idx[I0];
+
+                            static_for<1, i, 1>{}([&](auto j) {
+                                tmp = tmp * ordered_dst_access_lengths_tuple.At(dst_i)[j] +
+                                      ordered_dst_access_idx[j];
+                            });
+
+                            forward_sweep_(i) = tmp % 2 == 0;
+                        });
+
+                        return forward_sweep_;
+                    }();
+
+                    // calculate dst data index
+                    constexpr auto dst_data_idx = [&]() {
+                        Index ordered_idx;
+
+                        static_for<0, nDim, 1>{}([&](auto i) {
+                            ordered_idx(i) = forward_sweep[i]
+                                                 ? ordered_dst_access_idx[i]
+                                                 : ordered_dst_access_lengths_tuple.At(dst_i)[i] -
+                                                       1 - ordered_dst_access_idx[i];
+                        });
+
+                        return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                               dst_scalar_per_access_tuple.At(dst_i);
+                    }();
+
+                    constexpr auto dst_data_idx_seq =
+                        generate_sequence_v2([&](auto i) { return Number<dst_data_idx[i]>{}; },
+                                             Number<dst_data_idx.Size()>{});
+
+                    const bool is_dst_valid =
+                        coordinate_has_valid_offset_assuming_visible_index_is_valid(
+                            dst_descs.At(dst_i), dst_coords_.At(dst_i));
+
+                    using dst_vector_type = vector_type_maker_t<tuple_element_t<dst_i, DstDatas>,
+                                                                DstsScalarPerVector::At(dst_i)>;
+                    using dst_vector_t    = typename dst_vector_type::type;
+
+                    // copy data from dst_thread_scratch_ into dst_vector_container
+                    auto dst_vector_container = dst_vector_type{
+                        dst_thread_scratch_tuple_.At(dst_i).template GetAsType<dst_vector_t>(
+                            dst_data_idx_seq)};
+
+                    constexpr InMemoryDataOperationEnum DstInMemOp =
+                        static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(dst_i.value));
+
+                    // copy data from dst_vector_container to dst_buf
+                    dst_bufs.At(dst_i).template Update<DstInMemOp, dst_vector_t>(
+                        dst_coords_.At(dst_i).GetOffset(),
+                        is_dst_valid,
+                        dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+                    constexpr auto move_on_dim = [&]() constexpr
+                    {
+                        StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                        static_for<0, nDim, 1>{}([&](auto i) {
+                            move_on_dim_(i) = ordered_dst_access_idx[i] <
+                                              ordered_dst_access_lengths_tuple.At(dst_i)[i] - 1;
+
+                            static_for<i + 1, nDim, 1>{}([&](auto j) {
+                                move_on_dim_(i) &=
+                                    ordered_dst_access_idx[j] ==
+                                    ordered_dst_access_lengths_tuple.At(dst_i)[j] - 1;
+                            });
+                        });
+
+                        return move_on_dim_;
+                    }
+                    ();
+
+                    // move dst coord
+                    static_for<0, nDim, 1>{}([&](auto i) {
+                        if constexpr(move_on_dim[i])
+                        {
+                            if constexpr(forward_sweep[i])
+                            {
+                                move_tensor_coordinate(
+                                    dst_descs.At(dst_i),
+                                    dst_coords_.At(dst_i),
+                                    dst_forward_steps_tuple.At(dst_i)[dst_dim_access_order[i]]);
+                            }
+                            else
+                            {
+                                move_tensor_coordinate(
+                                    dst_descs.At(dst_i),
+                                    dst_coords_.At(dst_i),
+                                    dst_backward_steps_tuple.At(dst_i)[dst_dim_access_order[i]]);
+                            }
+                        }
+                    });
+                });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        static_for<0, nDst, 1>{}([&](auto dst_i) {
+            if constexpr(DstsResetCoordinateAfterRun::At(dst_i))
+            {
+                const auto dst_reset_step = make_tensor_coordinate_step(
+                    dst_descs.At(dst_i), GetDstCoordinateResetStep<dst_i>());
+
+                move_tensor_coordinate(dst_descs.At(dst_i), dst_coords_.At(dst_i), dst_reset_step);
+            }
+        });
+    }
+
+    template <index_t src_i>
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcsScalarPerVector::At(src_i)>{},
+            Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    template <index_t dst_i>
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstsScalarPerVector::At(dst_i)>{},
+            Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_scalar_per_access.At(dst_i);
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        static_for<0, nSrc, 1>{}([&](auto src_i) {
+            // if src coord was not reset by RunRead(), then need to adjust the step here
+            const auto adjusted_step_idx =
+                SrcsResetCoordinateAfterRun::At(src_i)
+                    ? src_slice_origin_step_idx
+                    : src_slice_origin_step_idx + GetSrcCoordinateResetStep<src_i>();
+
+            // is it OK to construct a new step every time?
+            const auto adjusted_step =
+                make_tensor_coordinate_step(src_descs.At(src_i), adjusted_step_idx);
+
+            move_tensor_coordinate(src_descs.At(src_i), src_coords_.At(src_i), adjusted_step);
+        });
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        static_for<0, nDst, 1>{}([&](auto dst_i) {
+            // if dst coord was not reset by RunWrite(), then need to adjust the step here
+            const auto adjusted_step_idx =
+                DstsResetCoordinateAfterRun::At(dst_i)
+                    ? dst_slice_origin_step_idx
+                    : dst_slice_origin_step_idx + GetDstCoordinateResetStep<dst_i>();
+
+            // is it OK to construct a new step every time?
+            const auto adjusted_step =
+                make_tensor_coordinate_step(dst_descs.At(dst_i), adjusted_step_idx);
+
+            move_tensor_coordinate(dst_descs.At(dst_i), dst_coords_.At(dst_i), adjusted_step);
+        });
+    }
+
+    template <index_t src_i>
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcsScalarPerVector::At(src_i)>{},
+            Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length =
+            container_push_back(sequence_to_tuple_of_number(src_access_lengths),
+                                Number<SrcsScalarPerVector::At(src_i)>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    template <index_t dst_i>
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstsScalarPerVector::At(dst_i)>{},
+            Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length =
+            container_push_back(sequence_to_tuple_of_number(dst_access_lengths),
+                                Number<DstsScalarPerVector::At(dst_i)>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto MakeSrcThreadScratchTuple()
+    {
+        return generate_tuple(
+            [&](auto src_i) {
+                constexpr auto src_thread_scratch_desc =
+                    decltype(GetSrcThreadScratchDescriptor<src_i>()){};
+                using SrcThreadScratch =
+                    StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                    tuple_element_t<src_i, SrcDatas>,
+                                                    SrcsScalarPerVector::At(src_i),
+                                                    decltype(src_thread_scratch_desc),
+                                                    true>;
+                return SrcThreadScratch{};
+            },
+            Number<nSrc>{});
+    }
+
+    __device__ static constexpr auto MakeDstThreadScratchTuple()
+    {
+        return generate_tuple(
+            [&](auto dst_i) {
+                constexpr auto dst_thread_scratch_desc =
+                    decltype(GetDstThreadScratchDescriptor<dst_i>()){};
+                using DstThreadScratch =
+                    StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                    tuple_element_t<dst_i, DstDatas>,
+                                                    DstsScalarPerVector::At(dst_i),
+                                                    decltype(dst_thread_scratch_desc),
+                                                    true>;
+                return DstThreadScratch{};
+            },
+            Number<nDst>{});
+    }
+
+    private:
+    using SrcThreadScratchTuple = decltype(MakeSrcThreadScratchTuple());
+    using DstThreadScratchTuple = decltype(MakeDstThreadScratchTuple());
+
+    StaticallyIndexedArray<SrcThreadScratchTuple, NumThreadScratch> src_thread_scratch_tuple_;
+
+    DstThreadScratchTuple dst_thread_scratch_tuple_;
+
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp

@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+// Assume:
+//   1. src:
+//     1. SrcDesc is known at compile-time
+//     2. SrcBuffer is DynamicBuffer
+//     3. src_ref_idx is known at run-time
+//     4. SrcRefToOriginDisplacement is known at compile-time
+//     5. use #-step
+//   2. dst:
+//     1. DstDesc is known at compile-time
+//     2. DstBuffer is StaticBuffer
+//     3. DstOriginIdx is known at compile-time
+//     4. use direct address calculation
+//   3. vector access on src
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v4r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v4r1(const Index& src_ref_idx)
+        : src_ref_coord_(make_tensor_coordinate(SrcDesc{}, src_ref_idx))
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_for<0, nDim, 1>{}([](auto i) {
+            static_assert(SliceLengths::At(i) % SrcVectorTensorLengths::At(i) == 0, "wrong!");
+        });
+    }
+
+    template <typename SrcRefToOriginDisplacement,
+              typename DstOriginIdx,
+              typename SrcBuffer,
+              typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcRefToOriginDisplacement&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstOriginIdx&,
+                        DstBuffer& dst_buf) const
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
+                is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
+                      "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
+                      "at compile-time");
+
+        // SrcDesc and DstDesc are known at compile-time
+        constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
+        constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
+
+        // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
+        constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
+        constexpr auto dst_origin_idx             = to_multi_index(DstOriginIdx{});
+
+        // tensor descriptor for src_vector
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(src_vector_tensor_lengths,
+                                                SrcVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            SrcVectorTensorContiguousDimOrder{});
+
+        constexpr auto src_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(src_vector_tensor_strides));
+
+        // access order and lengths
+        constexpr auto access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto dim_access_order = DimAccessOrder{};
+
+        constexpr auto ordered_access_lengths =
+            container_reorder_given_new2old(access_lengths, dim_access_order);
+
+        static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
+            // position in slice window
+            constexpr auto data_to_origin_disp_idx =
+                ordered_access_idx.ReorderGivenOld2New(dim_access_order) *
+                src_vector_tensor_lengths;
+
+            // src coordinate at starting point of src_vector
+            constexpr auto src_ref_to_data_disp_idx =
+                src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
+
+            constexpr auto src_ref_to_data_disp_coord_step =
+                make_tensor_coordinate_step(src_desc, src_ref_to_data_disp_idx);
+
+            auto src_data_coord = src_ref_coord_;
+
+            move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_coord_step);
+
+            vector_type_maker_t<SrcData, src_vector_desc.GetElementSpaceSize()> src_vector;
+
+            using src_vector_t = typename decltype(src_vector)::type;
+
+            const bool is_src_valid = coordinate_has_valid_offset_assuming_visible_index_is_valid(
+                src_desc, src_data_coord);
+
+            // copy data from src_buf into src_vector
+            src_vector.template AsType<src_vector_t>()(I0) =
+                src_buf.template Get<src_vector_t>(src_data_coord.GetOffset(), is_src_valid);
+
+            // copy data from src_vector into dst_buf (also cast from SrcData to DstData)
+            static_ford<SrcVectorTensorLengths>{}([&](auto src_vector_idx_) {
+                constexpr auto src_vector_idx = to_multi_index(src_vector_idx_);
+
+                constexpr index_t src_vector_offset =
+                    src_vector_desc.CalculateOffset(src_vector_idx);
+
+                constexpr index_t dst_offset = dst_desc.CalculateOffset(
+                    dst_origin_idx + data_to_origin_disp_idx + src_vector_idx);
+
+                dst_buf(Number<dst_offset>{}) = type_convert<DstData>(
+                    src_vector.template AsType<DstData>()[Number<src_vector_offset>{}]);
+            });
+        });
+    }
+
+    template <typename SrcSliceMoveStepIdx>
+    __device__ void MoveSrcSliceWindow(const SrcDesc&,
+                                       const SrcSliceMoveStepIdx& src_slice_move_step_idx)
+    {
+        constexpr auto src_desc = SrcDesc{};
+
+        const auto src_slice_move_step_iter =
+            make_tensor_coordinate_step(src_desc, to_multi_index(src_slice_move_step_idx));
+
+        move_tensor_coordinate(SrcDesc{}, src_ref_coord_, src_slice_move_step_iter);
+    }
+
+    private:
+    SrcCoord src_ref_coord_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp

@@ -0,0 +1,614 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename DstVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename DstVectorTensorContiguousDimOrder,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun> // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+struct ThreadwiseTensorSliceTransfer_v5r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v5r1(const SrcDesc& src_desc,
+                                                            const Index& src_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin))
+    {
+        // TODO: fix this
+        static_assert(is_same<SrcData, DstData>::value,
+                      "wrong! current implementation assume SrcData and DstData are same type");
+
+        static_for<0, nDim, 1>{}([](auto i) {
+            static_assert(SliceLengths::At(i) % SrcVectorTensorLengths::At(i) == 0 &&
+                              SliceLengths::At(i) % DstVectorTensorLengths::At(i) == 0,
+                          "wrong!");
+        });
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename SrcStepHacks>
+    __device__ void
+    RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        // tensor descriptor for src_vector
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(src_vector_tensor_lengths,
+                                                SrcVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            SrcVectorTensorContiguousDimOrder{});
+
+        constexpr auto src_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(src_vector_tensor_strides));
+
+        // access order and lengths
+        constexpr auto src_access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, forward_step_idx, src_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, backward_step_idx, src_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<0, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_vector_tensor_lengths;
+            }();
+
+            vector_type_maker_t<SrcData, src_vector_desc.GetElementSpaceSize()> src_vector;
+
+            using src_vector_t = typename decltype(src_vector)::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf to src_vector
+            src_vector.template AsType<src_vector_t>()(I0) =
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid);
+
+            // copy data from src_vector to buffer_
+            static_ford<SrcVectorTensorLengths>{}([&](auto src_vector_idx_) {
+                constexpr auto src_vector_idx = to_multi_index(src_vector_idx_);
+
+                constexpr index_t src_vector_offset =
+                    src_vector_desc.CalculateOffset(src_vector_idx);
+
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(src_data_idx + src_vector_idx);
+
+                buffer_(Number<buffer_offset>{}) =
+                    src_vector.template AsType<SrcData>()[Number<src_vector_offset>{}];
+            });
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <typename DstBuffer, typename DstStepHacks>
+    __device__ void
+    RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf, const DstStepHacks& dst_step_hacks)
+    {
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        // tensor descriptor for dst_vector
+        constexpr auto dst_vector_tensor_lengths = DstVectorTensorLengths{};
+
+        constexpr auto dst_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(dst_vector_tensor_lengths,
+                                                DstVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            DstVectorTensorContiguousDimOrder{});
+
+        constexpr auto dst_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(dst_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(dst_vector_tensor_strides));
+
+        // dst access order and lengths
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_vector_tensor_lengths;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = 0;
+
+                    static_for<0, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_vector_tensor_lengths;
+            }();
+
+            vector_type_maker_t<DstData, dst_vector_desc.GetElementSpaceSize()> dst_vector;
+
+            // copy data from buffer_ to dst_vector (also cast from SrcData to DstData)
+            static_ford<DstVectorTensorLengths>{}([&](auto dst_vector_idx_) {
+                constexpr auto dst_vector_idx = to_multi_index(dst_vector_idx_);
+
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(dst_data_idx + dst_vector_idx);
+
+                constexpr index_t dst_vector_offset =
+                    dst_vector_desc.CalculateOffset(dst_vector_idx);
+
+                dst_vector.template AsType<DstData>()(Number<dst_vector_offset>{}) =
+                    type_convert<DstData>(buffer_[Number<buffer_offset>{}]);
+            });
+
+            using dst_vector_t = typename decltype(dst_vector)::type;
+
+            // copy data from dst_vector to dst_buf
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    template <typename SrcBuffer>
+    __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
+    {
+        constexpr index_t ntransform_src = SrcDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_src, 0>::type{};
+
+        constexpr auto src_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunRead(src_desc, src_buf, src_step_hacks);
+    }
+
+    template <typename DstBuffer>
+    __device__ void RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        constexpr index_t ntransform_dst = DstDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_dst, 0>::type{};
+
+        constexpr auto dst_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunWrite(dst_desc, dst_buf, dst_step_hacks);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<0, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_vector_tensor_lengths;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto dst_vector_tensor_lengths = DstVectorTensorLengths{};
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_vector_tensor_lengths;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<0, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_vector_tensor_lengths;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& src_slice_origin_step_idx,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(
+            src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    static constexpr auto buffer_desc_ =
+        make_naive_tensor_descriptor_packed(sequence_to_tuple_of_number(SliceLengths{}));
+
+    static constexpr auto buffer_size_ = buffer_desc_.GetElementSpaceSize();
+
+    StaticBuffer<AddressSpaceEnum::Vgpr, SrcData, buffer_size_, true> buffer_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp

@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool SrcResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r1(const SrcDesc& src_desc,
+                                                            const Index& src_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        // loop over space-filling curve
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                DstData v;
+
+                // apply element-wise operation
+                element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
+
+                // apply type convert
+                dst_vector_container.template AsType<DstData>()(i) = v;
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRun
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp

@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool SrcResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r1r2
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r1r2(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const ElementwiseOperation& element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, InMemoryDataOperationEnum DstInMemOp>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        // loop over space-filling curve
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
+
+                // apply type convert
+                dst_vector_container.template AsType<DstData>()(i) = type_convert<DstData>(v);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRun
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp

@@ -0,0 +1,260 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src0_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename Src0Data,
+          typename Src1Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool Src0ResetCoordinateAfterRun,
+          bool Src1ResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r2
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using Src0Coord = decltype(make_tensor_coordinate(Src0Desc{}, Index{}));
+    using Src1Coord = decltype(make_tensor_coordinate(Src1Desc{}, Index{}));
+    using DstCoord  = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r2(const Src0Desc& src0_desc,
+                                                            const Index& src0_slice_origin,
+                                                            const Src1Desc& src1_desc,
+                                                            const Index& src1_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src0_coord_(make_tensor_coordinate(src0_desc, src0_slice_origin)),
+          src1_coord_(make_tensor_coordinate(src1_desc, src1_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrc0SliceOrigin(const Src0Desc& src0_desc,
+                                       const Index& src0_slice_origin_idx)
+    {
+        src0_coord_ = make_tensor_coordinate(src0_desc, src0_slice_origin_idx);
+    }
+
+    __device__ void SetSrc1SliceOrigin(const Src1Desc& src1_desc,
+                                       const Index& src1_slice_origin_idx)
+    {
+        src1_coord_ = make_tensor_coordinate(src1_desc, src1_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>;
+            using src0_vector_t    = typename src0_vector_type::type;
+
+            using src1_vector_type = vector_type_maker_t<Src1Data, ScalarPerVector>;
+            using src1_vector_t    = typename src1_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src0_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src0_desc, src0_coord_);
+
+            const bool is_src1_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src1_desc, src1_coord_);
+
+            // copy data from src0_buf into src0_vector_container
+            auto src0_vector_container = src0_vector_type{
+                src0_buf.template Get<src0_vector_t>(src0_coord_.GetOffset(), is_src0_valid)};
+
+            auto src1_vector_container = src1_vector_type{
+                src1_buf.template Get<src1_vector_t>(src1_coord_.GetOffset(), is_src1_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                element_op_(dst_vector_container.template AsType<DstData>()(i),
+                            src0_vector_container.template AsType<Src0Data>()[i],
+                            src1_vector_container.template AsType<Src1Data>()[i]);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src0_desc, src0_coord_, make_tensor_coordinate_step(src0_desc, forward_step));
+                move_tensor_coordinate(
+                    src1_desc, src1_coord_, make_tensor_coordinate_step(src1_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(Src0ResetCoordinateAfterRun)
+        {
+            const auto src0_reset_step =
+                make_tensor_coordinate_step(src0_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src0_desc, src0_coord_, src0_reset_step);
+        }
+
+        if constexpr(Src1ResetCoordinateAfterRun)
+        {
+            const auto src1_reset_step =
+                make_tensor_coordinate_step(src1_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src1_desc, src1_coord_, src1_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc,
+                                        const Index& src0_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src0ResetCoordinateAfterRun
+                                           ? src0_slice_origin_step_idx
+                                           : src0_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src0_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src0_desc, src0_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc,
+                                        const Index& src1_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src1ResetCoordinateAfterRun
+                                           ? src1_slice_origin_step_idx
+                                           : src1_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src1_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src1_desc, src1_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    Src0Coord src0_coord_;
+    Src1Coord src1_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp

@@ -0,0 +1,310 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src0_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename Src0Data,
+          typename Src1Data,
+          typename Src2Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename Src2Desc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool Src0ResetCoordinateAfterRun,
+          bool Src1ResetCoordinateAfterRun,
+          bool Src2ResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r3
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using Src0Coord = decltype(make_tensor_coordinate(Src0Desc{}, Index{}));
+    using Src1Coord = decltype(make_tensor_coordinate(Src1Desc{}, Index{}));
+    using Src2Coord = decltype(make_tensor_coordinate(Src2Desc{}, Index{}));
+    using DstCoord  = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r3(const Src0Desc& src0_desc,
+                                                            const Index& src0_slice_origin,
+                                                            const Src1Desc& src1_desc,
+                                                            const Index& src1_slice_origin,
+                                                            const Src2Desc& src2_desc,
+                                                            const Index& src2_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src0_coord_(make_tensor_coordinate(src0_desc, src0_slice_origin)),
+          src1_coord_(make_tensor_coordinate(src1_desc, src1_slice_origin)),
+          src2_coord_(make_tensor_coordinate(src2_desc, src2_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrc0SliceOrigin(const Src0Desc& src0_desc,
+                                       const Index& src0_slice_origin_idx)
+    {
+        src0_coord_ = make_tensor_coordinate(src0_desc, src0_slice_origin_idx);
+    }
+
+    __device__ void SetSrc1SliceOrigin(const Src1Desc& src1_desc,
+                                       const Index& src1_slice_origin_idx)
+    {
+        src1_coord_ = make_tensor_coordinate(src1_desc, src1_slice_origin_idx);
+    }
+
+    __device__ void SetSrc2SliceOrigin(const Src2Desc& src2_desc,
+                                       const Index& src2_slice_origin_idx)
+    {
+        src2_coord_ = make_tensor_coordinate(src2_desc, src2_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename Src2Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const Src2Desc& src2_desc,
+                        const Src2Buffer& src2_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>;
+            using src0_vector_t    = typename src0_vector_type::type;
+
+            using src1_vector_type = vector_type_maker_t<Src1Data, ScalarPerVector>;
+            using src1_vector_t    = typename src1_vector_type::type;
+
+            using src2_vector_type = vector_type_maker_t<Src2Data, ScalarPerVector>;
+            using src2_vector_t    = typename src2_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src0_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src0_desc, src0_coord_);
+
+            const bool is_src1_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src1_desc, src1_coord_);
+
+            const bool is_src2_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src2_desc, src2_coord_);
+
+            // copy data from src0_buf into src0_vector_container
+            auto src0_vector_container = src0_vector_type{
+                src0_buf.template Get<src0_vector_t>(src0_coord_.GetOffset(), is_src0_valid)};
+
+            auto src1_vector_container = src1_vector_type{
+                src1_buf.template Get<src1_vector_t>(src1_coord_.GetOffset(), is_src1_valid)};
+
+            auto src2_vector_container = src2_vector_type{
+                src2_buf.template Get<src2_vector_t>(src2_coord_.GetOffset(), is_src2_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                element_op_(dst_vector_container.template AsType<DstData>()(i),
+                            src0_vector_container.template AsType<Src0Data>()[i],
+                            src1_vector_container.template AsType<Src1Data>()[i],
+                            src2_vector_container.template AsType<Src2Data>()[i]);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src0_desc, src0_coord_, make_tensor_coordinate_step(src0_desc, forward_step));
+                move_tensor_coordinate(
+                    src1_desc, src1_coord_, make_tensor_coordinate_step(src1_desc, forward_step));
+                move_tensor_coordinate(
+                    src2_desc, src2_coord_, make_tensor_coordinate_step(src2_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(Src0ResetCoordinateAfterRun)
+        {
+            const auto src0_reset_step =
+                make_tensor_coordinate_step(src0_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src0_desc, src0_coord_, src0_reset_step);
+        }
+
+        if constexpr(Src1ResetCoordinateAfterRun)
+        {
+            const auto src1_reset_step =
+                make_tensor_coordinate_step(src1_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src1_desc, src1_coord_, src1_reset_step);
+        }
+
+        if constexpr(Src2ResetCoordinateAfterRun)
+        {
+            const auto src2_reset_step =
+                make_tensor_coordinate_step(src2_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src2_desc, src2_coord_, src2_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc,
+                                        const Index& src0_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src0ResetCoordinateAfterRun
+                                           ? src0_slice_origin_step_idx
+                                           : src0_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src0_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src0_desc, src0_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc,
+                                        const Index& src1_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src1ResetCoordinateAfterRun
+                                           ? src1_slice_origin_step_idx
+                                           : src1_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src1_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src1_desc, src1_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc2SliceWindow(const Src2Desc& src2_desc,
+                                        const Index& src2_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src2ResetCoordinateAfterRun
+                                           ? src2_slice_origin_step_idx
+                                           : src2_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src2_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src2_desc, src2_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    Src0Coord src0_coord_;
+    Src1Coord src1_coord_;
+    Src2Coord src2_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp

@@ -0,0 +1,298 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags> // Sequence<bool ...>
+struct ThreadwiseTensorSliceTransfer_v7
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+    using SpaceFillingCurve =
+        SpaceFillingCurve<SliceLengths, DimAccessOrder, remove_cv_t<decltype(scalar_per_access)>>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        auto generate_vectors = [&](auto data_types) {
+            constexpr index_t num = data_types.Size();
+
+            return generate_tuple(
+                [&](auto i) {
+                    using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                    return vector_type_maker_t<DataType, ScalarPerVector>{};
+                },
+                Number<num>{});
+        };
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors(SrcDatas{});
+            auto dst_vectors = generate_vectors(DstDatas{});
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
+                                                                                src_coords_[i]);
+
+                src_vectors(i).template AsType<src_vector_t>()(I0) =
+                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(),
+                                                           is_src_valid);
+            });
+
+            // apply pointwise function
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        return src_vectors[iSrc].template AsType<SrcData>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        return dst_vectors(iDst).template AsType<DstData>()(i);
+                    },
+                    Number<nDst>{});
+
+                // apply pointwise function
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
+
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+
+                const bool is_dst_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
+                                                                                dst_coords_[i]);
+
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_coords_[i].GetOffset(),
+                    is_dst_valid,
+                    dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i],
+                                           src_coords_(i),
+                                           make_tensor_coordinate_step(src_descs[i], forward_step));
+                });
+
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(dst_descs[i],
+                                           dst_coords_(i),
+                                           make_tensor_coordinate_step(dst_descs[i], forward_step));
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto src_reset_step =
+                    make_tensor_coordinate_step(src_descs[i], GetCoordinateResetStep());
+
+                move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto dst_reset_step =
+                    make_tensor_coordinate_step(dst_descs[i], GetCoordinateResetStep());
+
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
+            }
+        });
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRunFlags::At(iSrc)
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRunFlags::At(iDst)
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
+    }
+
+    private:
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp

@@ -0,0 +1,630 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/utility/is_detected.hpp"
+#include "ck/tensor/static_tensor.hpp"
+
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
+
+namespace ck {
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v7r2
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto src_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+    static constexpr auto dst_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+    using SrcSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   SrcDimAccessOrder,
+                                                   remove_cv_t<decltype(src_scalar_per_access)>,
+                                                   false>;
+
+    using DstSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   DstDimAccessOrder,
+                                                   remove_cv_t<decltype(dst_scalar_per_access)>,
+                                                   false>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7r2(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename DataTypes, index_t ScalarPerVector>
+    __device__ static auto generate_vectors()
+    {
+        auto data_types = DataTypes{};
+
+        constexpr index_t num = data_types.Size();
+
+        return generate_tuple(
+            [&](auto i) {
+                using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                return vector_type_maker_t<DataType, ScalarPerVector>{};
+            },
+            Number<num>{});
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    template <typename SrcBuffers,
+              index_t ThreadScratchId                                   = 0,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size(), bool> = false>
+    __device__ void RunRead(const SrcDescs& src_descs,
+                            const SrcBuffers& src_bufs,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors<SrcDatas, SrcScalarPerVector>();
+            auto elm_vectors = generate_vectors<DstDatas, SrcScalarPerVector>();
+
+            bool oob_val = true;
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
+                                                                                src_coords_[i]);
+
+                oob_val = oob_val & is_src_valid;
+
+                src_vectors(i).template AsType<src_vector_t>()(I0) =
+                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(), true);
+            });
+
+            constexpr auto get_elem_op_vec_len = []() {
+                if constexpr(is_detected<is_pack8_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack8_invocable)
+                        return math::min(8, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack4_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack4_invocable)
+                        return math::min(4, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack2_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack2_invocable)
+                        return math::min(2, SrcScalarPerVector);
+                }
+                return 1;
+            };
+
+            constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
+
+            // apply pointwise function
+            static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
+
+                        return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
+
+                        return elm_vectors(iDst).template AsType<elem_op_vec_t>()(i);
+                    },
+                    Number<nDst>{});
+
+                // apply pointwise function
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+            oob_vectors_tuple_(thread_scratch_id)(iAccess) = oob_val;
+
+            // move coordinate
+            if constexpr(iAccess.value != src_num_access - 1)
+            {
+                constexpr auto forward_step = SrcSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i],
+                                           src_coords_(i),
+                                           make_tensor_coordinate_step(src_descs[i], forward_step));
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto src_reset_step =
+                    make_tensor_coordinate_step(src_descs[i], GetSrcCoordinateResetStep());
+
+                move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
+            }
+        });
+    }
+
+#if 1
+    template <index_t ThreadScratchId = 0>
+    __device__ void OOBCheck(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto elm_vectors = elm_vectors_tuple_[thread_scratch_id][iAccess];
+            auto oob_val     = oob_vectors_tuple_[thread_scratch_id][iAccess];
+
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using elm_vector_t = typename remove_cvref_t<decltype(elm_vectors[i])>::type;
+                elm_vectors(i).template AsType<elm_vector_t>()(I0) =
+                    oob_val ? elm_vectors(i).template AsType<elm_vector_t>()[I0] : elm_vector_t{0};
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+        });
+    }
+#endif
+
+    template <index_t ThreadScratchId = 0>
+    __device__ void
+    TransposeFromElmToDst(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        using DstData = remove_cvref_t<decltype(DstDatas{}[I0])>;
+
+        using ElmThreadScratch =
+            StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                            DstData,
+                                            SrcScalarPerVector,
+                                            decltype(GetSrcThreadScratchDescriptor()),
+                                            true>;
+        using DstThreadScratch =
+            StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                            DstData,
+                                            DstScalarPerVector,
+                                            decltype(GetDstThreadScratchDescriptor()),
+                                            true>;
+
+        ElmThreadScratch elm_thread_scratch_;
+        DstThreadScratch dst_thread_scratch_;
+
+        elm_thread_scratch_.data_ =
+            bit_cast<decltype(elm_thread_scratch_.data_)>(elm_vectors_tuple_[thread_scratch_id]);
+
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<f8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) ||
+                      (is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
+        {
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst<SrcVectorDim,
+                                                                 SrcScalarPerVector,
+                                                                 DstVectorDim,
+                                                                 DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                using src_vector_t = vector_type_maker_t<DstData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return elm_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                transpose_vectors<DstData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+        else
+        {
+            static_ford<SliceLengths>{}(
+                [&](auto idx) { dst_thread_scratch_(idx) = elm_thread_scratch_[idx]; });
+        }
+
+        dst_vectors_tuple_(thread_scratch_id) = bit_cast<DstVectorTuple>(dst_thread_scratch_.data_);
+    }
+
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename DstBuffers,
+              index_t ThreadScratchId                                             = 0,
+              enable_if_t<DstDescs::Size() == 1 && DstBuffers::Size() == 1, bool> = false>
+    __device__ void RunWrite(const DstDescs& dst_descs,
+                             DstBuffers dst_bufs,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        OOBCheck(thread_scratch_id);
+        TransposeFromElmToDst(thread_scratch_id);
+
+        // loop over space-filling curve
+        static_for<0, dst_num_access, 1>{}([&](auto iAccess) {
+            auto dst_vectors = dst_vectors_tuple_[thread_scratch_id][iAccess];
+
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+
+                const bool is_dst_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
+                                                                                dst_coords_[i]);
+
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_coords_[i].GetOffset(),
+                    is_dst_valid,
+                    dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != dst_num_access - 1)
+            {
+                constexpr auto forward_step = DstSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(dst_descs[i],
+                                           dst_coords_(i),
+                                           make_tensor_coordinate_step(dst_descs[i], forward_step));
+                });
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto dst_reset_step =
+                    make_tensor_coordinate_step(dst_descs[i], GetDstCoordinateResetStep());
+
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
+            }
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        RunRead(src_descs, src_bufs);
+        RunWrite(dst_descs, dst_bufs);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        if constexpr(src_num_access == 0)
+        {
+            return typename SrcSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return SrcSpaceFillingCurve::GetStepBetween(Number<src_num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        if constexpr(dst_num_access == 0)
+        {
+            return typename DstSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return DstSpaceFillingCurve::GetStepBetween(Number<dst_num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        // constexpr auto src_scalar_per_access = generate_sequence(
+        // detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        // constexpr auto dst_scalar_per_access = generate_sequence(
+        // detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRunFlags::At(iSrc)
+                ? src_slice_origin_step_idx
+                : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRunFlags::At(iDst)
+                ? dst_slice_origin_step_idx
+                : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
+    }
+
+    private:
+    using SrcVectorsType = decltype(generate_vectors<SrcDatas, SrcScalarPerVector>());
+    using ElmVectorsType = decltype(generate_vectors<DstDatas, SrcScalarPerVector>());
+    using DstVectorsType = decltype(generate_vectors<DstDatas, DstScalarPerVector>());
+
+    static constexpr auto src_num_access = SrcSpaceFillingCurve::GetNumOfAccess();
+    static constexpr auto dst_num_access = DstSpaceFillingCurve::GetNumOfAccess();
+
+    using ElmVectorTuple = StaticallyIndexedArray<ElmVectorsType, src_num_access>;
+    using DstVectorTuple = StaticallyIndexedArray<DstVectorsType, dst_num_access>;
+
+    StaticallyIndexedArray<ElmVectorTuple, NumThreadScratch> elm_vectors_tuple_;
+    StaticallyIndexedArray<DstVectorTuple, NumThreadScratch> dst_vectors_tuple_;
+
+    using OOBVectorTuple = StaticallyIndexedArray<bool, src_num_access>;
+    StaticallyIndexedArray<OOBVectorTuple, NumThreadScratch> oob_vectors_tuple_;
+
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r3.hpp

@@ -0,0 +1,648 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/utility/is_detected.hpp"
+#include "ck/tensor/static_tensor.hpp"
+
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
+
+namespace ck {
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          typename SrcScalarPerVectors,
+          index_t DstScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v7r3
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr auto SrcScalarPerVector = SrcScalarPerVectors{}[I0];
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto src_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+    static constexpr auto dst_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+    using SrcSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   SrcDimAccessOrder,
+                                                   remove_cv_t<decltype(src_scalar_per_access)>,
+                                                   false>;
+
+    using DstSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   DstDimAccessOrder,
+                                                   remove_cv_t<decltype(dst_scalar_per_access)>,
+                                                   false>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7r3(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename DataTypes, index_t ScalarPerVector>
+    __device__ static auto generate_vectors()
+    {
+        auto data_types = DataTypes{};
+
+        constexpr index_t num = data_types.Size();
+
+        return generate_tuple(
+            [&](auto i) {
+                using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                return vector_type_maker_t<DataType, ScalarPerVector>{};
+            },
+            Number<num>{});
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    template <typename SrcBuffers,
+              index_t ThreadScratchId                                   = 0,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size(), bool> = false>
+    __device__ void RunRead(const SrcDescs& src_descs,
+                            const SrcBuffers& src_bufs,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors<SrcDatas, SrcScalarPerVector>();
+            auto elm_vectors = generate_vectors<DstDatas, SrcScalarPerVector>();
+
+            bool oob_val = true;
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
+                                                                                src_coords_[i]);
+
+                oob_val = oob_val & is_src_valid;
+
+                if constexpr(SrcScalarPerVectors{}[i] == 1)
+                {
+                    auto data_types = SrcDatas{};
+                    using DataType  = remove_cvref_t<decltype(data_types[i])>;
+                    const auto tmp =
+                        src_bufs[i].template Get<DataType>(src_coords_[i].GetOffset(), true);
+
+                    static_for<0, SrcScalarPerVector, 1>{}(
+                        [&](auto j) { src_vectors(i).template AsType<DataType>()(j) = tmp; });
+                }
+                else
+                {
+                    src_vectors(i).template AsType<src_vector_t>()(I0) =
+                        src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(), true);
+                }
+            });
+
+            constexpr auto get_elem_op_vec_len = []() {
+                if constexpr(is_detected<is_pack8_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack8_invocable)
+                        return math::min(8, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack4_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack4_invocable)
+                        return math::min(4, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack2_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack2_invocable)
+                        return math::min(2, SrcScalarPerVector);
+                }
+                return 1;
+            };
+
+            constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
+
+            // apply pointwise function
+            static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
+
+                        return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
+
+                        return elm_vectors(iDst).template AsType<elem_op_vec_t>()(i);
+                    },
+                    Number<nDst>{});
+
+                // apply pointwise function
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+            oob_vectors_tuple_(thread_scratch_id)(iAccess) = oob_val;
+
+            // move coordinate
+            if constexpr(iAccess.value != src_num_access - 1)
+            {
+                constexpr auto forward_step = SrcSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i],
+                                           src_coords_(i),
+                                           make_tensor_coordinate_step(src_descs[i], forward_step));
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto src_reset_step =
+                    make_tensor_coordinate_step(src_descs[i], GetSrcCoordinateResetStep());
+
+                move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
+            }
+        });
+    }
+
+#if 1
+    template <index_t ThreadScratchId = 0>
+    __device__ void OOBCheck(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto elm_vectors = elm_vectors_tuple_[thread_scratch_id][iAccess];
+            auto oob_val     = oob_vectors_tuple_[thread_scratch_id][iAccess];
+
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using elm_vector_t = typename remove_cvref_t<decltype(elm_vectors[i])>::type;
+                elm_vectors(i).template AsType<elm_vector_t>()(I0) =
+                    oob_val ? elm_vectors(i).template AsType<elm_vector_t>()[I0] : elm_vector_t{0};
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+        });
+    }
+#endif
+
+    template <index_t ThreadScratchId = 0>
+    __device__ void
+    TransposeFromElmToDst(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        using DstData = remove_cvref_t<decltype(DstDatas{}[I0])>;
+
+        using ElmThreadScratch =
+            StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                            DstData,
+                                            SrcScalarPerVector,
+                                            decltype(GetSrcThreadScratchDescriptor()),
+                                            true>;
+        using DstThreadScratch =
+            StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                            DstData,
+                                            DstScalarPerVector,
+                                            decltype(GetDstThreadScratchDescriptor()),
+                                            true>;
+
+        ElmThreadScratch elm_thread_scratch_;
+        DstThreadScratch dst_thread_scratch_;
+
+        elm_thread_scratch_.data_ =
+            bit_cast<decltype(elm_thread_scratch_.data_)>(elm_vectors_tuple_[thread_scratch_id]);
+
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<f8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) ||
+                      (is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
+        {
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst<SrcVectorDim,
+                                                                 SrcScalarPerVector,
+                                                                 DstVectorDim,
+                                                                 DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                using src_vector_t = vector_type_maker_t<DstData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return elm_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from
+                // dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                transpose_vectors<DstData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+        else
+        {
+            static_ford<SliceLengths>{}(
+                [&](auto idx) { dst_thread_scratch_(idx) = elm_thread_scratch_[idx]; });
+        }
+
+        dst_vectors_tuple_(thread_scratch_id) = bit_cast<DstVectorTuple>(dst_thread_scratch_.data_);
+    }
+
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename DstBuffers,
+              index_t ThreadScratchId                                             = 0,
+              enable_if_t<DstDescs::Size() == 1 && DstBuffers::Size() == 1, bool> = false>
+    __device__ void RunWrite(const DstDescs& dst_descs,
+                             DstBuffers dst_bufs,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        OOBCheck(thread_scratch_id);
+        TransposeFromElmToDst(thread_scratch_id);
+
+        // loop over space-filling curve
+        static_for<0, dst_num_access, 1>{}([&](auto iAccess) {
+            auto dst_vectors = dst_vectors_tuple_[thread_scratch_id][iAccess];
+
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+
+                const bool is_dst_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
+                                                                                dst_coords_[i]);
+
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_coords_[i].GetOffset(),
+                    is_dst_valid,
+                    dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != dst_num_access - 1)
+            {
+                constexpr auto forward_step = DstSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(dst_descs[i],
+                                           dst_coords_(i),
+                                           make_tensor_coordinate_step(dst_descs[i], forward_step));
+                });
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto dst_reset_step =
+                    make_tensor_coordinate_step(dst_descs[i], GetDstCoordinateResetStep());
+
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
+            }
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        RunRead(src_descs, src_bufs);
+        RunWrite(dst_descs, dst_bufs);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        if constexpr(src_num_access == 0)
+        {
+            return typename SrcSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return SrcSpaceFillingCurve::GetStepBetween(Number<src_num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        if constexpr(dst_num_access == 0)
+        {
+            return typename DstSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return DstSpaceFillingCurve::GetStepBetween(Number<dst_num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        // constexpr auto src_scalar_per_access = generate_sequence(
+        // detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{},
+        // Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        // constexpr auto dst_scalar_per_access = generate_sequence(
+        // detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{},
+        // Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRunFlags::At(iSrc)
+                ? src_slice_origin_step_idx
+                : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRunFlags::At(iDst)
+                ? dst_slice_origin_step_idx
+                : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
+    }
+
+    private:
+    using SrcVectorsType = decltype(generate_vectors<SrcDatas, SrcScalarPerVector>());
+    using ElmVectorsType = decltype(generate_vectors<DstDatas, SrcScalarPerVector>());
+    using DstVectorsType = decltype(generate_vectors<DstDatas, DstScalarPerVector>());
+
+    static constexpr auto src_num_access = SrcSpaceFillingCurve::GetNumOfAccess();
+    static constexpr auto dst_num_access = DstSpaceFillingCurve::GetNumOfAccess();
+
+    using ElmVectorTuple = StaticallyIndexedArray<ElmVectorsType, src_num_access>;
+    using DstVectorTuple = StaticallyIndexedArray<DstVectorsType, dst_num_access>;
+
+    StaticallyIndexedArray<ElmVectorTuple, NumThreadScratch> elm_vectors_tuple_;
+    StaticallyIndexedArray<DstVectorTuple, NumThreadScratch> dst_vectors_tuple_;
+
+    using OOBVectorTuple = StaticallyIndexedArray<bool, src_num_access>;
+    StaticallyIndexedArray<OOBVectorTuple, NumThreadScratch> oob_vectors_tuple_;
+
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r3_scatter.hpp

@@ -0,0 +1,681 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/utility/is_detected.hpp"
+#include "ck/tensor/static_tensor.hpp"
+
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
+
+namespace ck {
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          typename SrcScalarPerVectors,
+          index_t DstScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename IndexType,
+          index_t ScatterDim       = 1,
+          bool OutputScatter       = true,
+          index_t ScatterWeightIdx = 3,
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v7r3_scatter
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto SrcScalarPerVector = SrcScalarPerVectors{}[I0];
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index                          = MultiIndex<nDim>;
+    static constexpr index_t scatter_num = SliceLengths{}.At(Number<ScatterDim>{});
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto src_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+    static constexpr auto dst_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+    using SrcSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   SrcDimAccessOrder,
+                                                   remove_cv_t<decltype(src_scalar_per_access)>,
+                                                   false>;
+
+    using DstSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   DstDimAccessOrder,
+                                                   remove_cv_t<decltype(dst_scalar_per_access)>,
+                                                   false>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7r3_scatter(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename DataTypes, index_t ScalarPerVector>
+    __device__ static auto generate_vectors()
+    {
+        auto data_types = DataTypes{};
+
+        constexpr index_t num = data_types.Size();
+
+        return generate_tuple(
+            [&](auto i) {
+                using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                return vector_type_maker_t<DataType, ScalarPerVector>{};
+            },
+            Number<num>{});
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    template <typename SrcBuffers,
+              index_t ThreadScratchId                                   = 0,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size(), bool> = false>
+    __device__ void RunRead(const SrcDescs& src_descs,
+                            const SrcBuffers& src_bufs,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors<SrcDatas, SrcScalarPerVector>();
+            auto elm_vectors = generate_vectors<DstDatas, SrcScalarPerVector>();
+
+            bool oob_val = true;
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
+                                                                                src_coords_[i]);
+
+                oob_val = oob_val & is_src_valid;
+                src_vectors(i).template AsType<src_vector_t>()(I0) =
+                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(), true);
+            });
+
+            constexpr auto get_elem_op_vec_len = []() {
+                if constexpr(is_detected<is_pack8_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack8_invocable)
+                        return math::min(8, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack4_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack4_invocable)
+                        return math::min(4, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack2_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack2_invocable)
+                        return math::min(2, SrcScalarPerVector);
+                }
+                return 1;
+            };
+
+            constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
+
+            // apply pointwise function
+            static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
+
+                        return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
+
+                        return elm_vectors(iDst).template AsType<elem_op_vec_t>()(i);
+                    },
+                    Number<nDst>{});
+
+                // apply pointwise function
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+            oob_vectors_tuple_(thread_scratch_id)(iAccess) = oob_val;
+
+            // move coordinate
+            if constexpr(iAccess.value != src_num_access - 1)
+            {
+                constexpr auto forward_step = SrcSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i],
+                                           src_coords_(i),
+                                           make_tensor_coordinate_step(src_descs[i], forward_step));
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto src_reset_step =
+                    make_tensor_coordinate_step(src_descs[i], GetSrcCoordinateResetStep());
+
+                move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
+            }
+        });
+    }
+
+#if 1
+    template <index_t ThreadScratchId = 0>
+    __device__ void OOBCheck(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto elm_vectors = elm_vectors_tuple_[thread_scratch_id][iAccess];
+            auto oob_val     = oob_vectors_tuple_[thread_scratch_id][iAccess];
+
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using elm_vector_t = typename remove_cvref_t<decltype(elm_vectors[i])>::type;
+                elm_vectors(i).template AsType<elm_vector_t>()(I0) =
+                    oob_val ? elm_vectors(i).template AsType<elm_vector_t>()[I0] : elm_vector_t{0};
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+        });
+    }
+#endif
+
+    template <index_t ThreadScratchId = 0>
+    __device__ void
+    TransposeFromElmToDst(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        using DstData = remove_cvref_t<decltype(DstDatas{}[I0])>;
+
+        using ElmThreadScratch =
+            StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                            DstData,
+                                            SrcScalarPerVector,
+                                            decltype(GetSrcThreadScratchDescriptor()),
+                                            true>;
+        using DstThreadScratch =
+            StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                            DstData,
+                                            DstScalarPerVector,
+                                            decltype(GetDstThreadScratchDescriptor()),
+                                            true>;
+
+        ElmThreadScratch elm_thread_scratch_;
+        DstThreadScratch dst_thread_scratch_;
+
+        elm_thread_scratch_.data_ =
+            bit_cast<decltype(elm_thread_scratch_.data_)>(elm_vectors_tuple_[thread_scratch_id]);
+
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<f8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) ||
+                      (is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
+        {
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst<SrcVectorDim,
+                                                                 SrcScalarPerVector,
+                                                                 DstVectorDim,
+                                                                 DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                using src_vector_t = vector_type_maker_t<DstData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return elm_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from
+                // dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                transpose_vectors<DstData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+        else
+        {
+            static_ford<SliceLengths>{}(
+                [&](auto idx) { dst_thread_scratch_(idx) = elm_thread_scratch_[idx]; });
+        }
+
+        dst_vectors_tuple_(thread_scratch_id) = bit_cast<DstVectorTuple>(dst_thread_scratch_.data_);
+    }
+
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename DstBuffers,
+              index_t ThreadScratchId                                             = 0,
+              enable_if_t<DstDescs::Size() == 1 && DstBuffers::Size() == 1, bool> = false>
+    __device__ void RunWrite(const DstDescs& dst_descs,
+                             DstBuffers dst_bufs,
+                             StaticallyIndexedArray<IndexType, scatter_num>& scatter_offsets,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        OOBCheck(thread_scratch_id);
+        TransposeFromElmToDst(thread_scratch_id);
+
+        // loop over space-filling curve
+        static_for<0, dst_num_access, 1>{}([&](auto iAccess) {
+            auto dst_vectors         = dst_vectors_tuple_[thread_scratch_id][iAccess];
+            IndexType scatter_offset = 0;
+            if constexpr(OutputScatter)
+            {
+                constexpr auto iScatter =
+                    DstSpaceFillingCurve::GetIndex(iAccess)(Number<ScatterDim>{});
+                scatter_offset = scatter_offsets(Number<iScatter>{});
+            }
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t      = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+                IndexType dst_offset    = scatter_offset + (dst_coords_[i].GetOffset());
+                const bool is_dst_valid = dst_offset < dst_descs[i].GetElementSpaceSize();
+                // coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
+                //                                                             dst_coords_[i]);
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_offset, is_dst_valid, dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != dst_num_access - 1)
+            {
+                constexpr auto forward_step = DstSpaceFillingCurve::GetForwardStep(iAccess);
+
+                auto forward_step_scatter = [&]() constexpr
+                {
+                    Index step_;
+
+                    static_for<0, nDim, 1>{}([&](auto i) {
+                        step_(i) = (i.value == ScatterDim && OutputScatter) ? 0 : forward_step[i];
+                    });
+
+                    return step_;
+                }
+                ();
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(
+                        dst_descs[i],
+                        dst_coords_(i),
+                        make_tensor_coordinate_step(dst_descs[i], forward_step_scatter));
+                });
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto dst_reset_step =
+                    make_tensor_coordinate_step(dst_descs[i], GetDstCoordinateResetStep());
+
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
+            }
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs,
+                        StaticallyIndexedArray<IndexType, scatter_num>& scatter_offsets)
+    {
+        RunRead(src_descs, src_bufs);
+        RunWrite(dst_descs, dst_bufs, scatter_offsets);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        if constexpr(src_num_access == 0)
+        {
+            return typename SrcSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return SrcSpaceFillingCurve::GetStepBetween(Number<src_num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        if constexpr(dst_num_access == 0)
+        {
+            return typename DstSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                DstSpaceFillingCurve::GetStepBetween(Number<dst_num_access - 1>{}, Number<0>{});
+            auto reset_step_scatter = [&]() constexpr
+            {
+                Index step_;
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    step_(i) =
+                        (i.value == ScatterDim && OutputScatter) ? 0 : reset_step[Number<i>{}];
+                });
+
+                return step_;
+            }
+            ();
+            return reset_step_scatter;
+        }
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        // constexpr auto src_scalar_per_access = generate_sequence(
+        // detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{},
+        // Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        // constexpr auto dst_scalar_per_access = generate_sequence(
+        // detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{},
+        // Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRunFlags::At(iSrc)
+                ? src_slice_origin_step_idx
+                : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRunFlags::At(iDst)
+                ? dst_slice_origin_step_idx
+                : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        auto adjusted_step_idx_scatter = [&]() {
+            Index step_;
+            static_for<0, nDim, 1>{}([&](auto i) {
+                step_(i) =
+                    (i.value == ScatterDim && OutputScatter) ? 0 : adjusted_step_idx[Number<i>{}];
+            });
+
+            return step_;
+        }();
+        // is it OK to construct a new step every time?
+        const auto adjusted_step =
+            make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx_scatter);
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
+    }
+
+    private:
+    using SrcVectorsType = decltype(generate_vectors<SrcDatas, SrcScalarPerVector>());
+    using ElmVectorsType = decltype(generate_vectors<DstDatas, SrcScalarPerVector>());
+    using DstVectorsType = decltype(generate_vectors<DstDatas, DstScalarPerVector>());
+
+    static constexpr auto src_num_access = SrcSpaceFillingCurve::GetNumOfAccess();
+    static constexpr auto dst_num_access = DstSpaceFillingCurve::GetNumOfAccess();
+
+    using ElmVectorTuple = StaticallyIndexedArray<ElmVectorsType, src_num_access>;
+    using DstVectorTuple = StaticallyIndexedArray<DstVectorsType, dst_num_access>;
+
+    StaticallyIndexedArray<ElmVectorTuple, NumThreadScratch> elm_vectors_tuple_;
+    StaticallyIndexedArray<DstVectorTuple, NumThreadScratch> dst_vectors_tuple_;
+
+    using OOBVectorTuple = StaticallyIndexedArray<bool, src_num_access>;
+    StaticallyIndexedArray<OOBVectorTuple, NumThreadScratch> oob_vectors_tuple_;
+
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp

@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/math_v2.hpp"
+
+namespace ck {
+
+// Assume
+//  1) XDesc is known at compile-time
+//  2) MeanVarDesc is known at compile-time
+//  3) XBuffer is static buffer
+//  4) MeanBuffer is static buffer
+//  5) VarBuffer is static buffer
+template <typename T, typename XThreadDesc_M_K, typename MeanVarThreadDesc_M>
+struct ThreadwiseWelford
+{
+    static constexpr auto x_thread_desc_m_k      = XThreadDesc_M_K{};
+    static constexpr auto mean_var_thread_desc_m = MeanVarThreadDesc_M{};
+
+    static constexpr auto thread_x_length_m        = x_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto thread_x_length_k        = x_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto thread_mean_var_length_m = mean_var_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(thread_x_length_m == thread_mean_var_length_m,
+                  "lengths of source and mean/var buffer must match!");
+
+    __device__ constexpr ThreadwiseWelford() : cur_count_(0), max_count_(0) {}
+
+    __device__ inline void Update(T& mean, T& var, T x)
+    {
+        using ck::math::isnan;
+
+        if(isnan(x))
+        {
+            mean = x;
+            var  = x;
+        }
+        else
+        {
+            T delta = x - mean;
+            mean += delta / cur_count_;
+            T delta2 = x - mean;
+            var += delta * delta2;
+        }
+    }
+
+    template <typename XBufferType, typename MeanBufferType, typename VarBufferType>
+    __device__ void
+    Run(const XBufferType& x_buf_m_k, MeanBufferType& mean_buf_m, VarBufferType& var_buf_m)
+    {
+        // FIXME - Better naming for var_buf_m
+
+        static_for<0, thread_x_length_k, 1>{}([&](auto iK) {
+            if(cur_count_ < max_count_)
+            {
+                ++cur_count_;
+
+                static_for<0, thread_x_length_m, 1>{}([&](auto iM) {
+                    constexpr index_t out_offset =
+                        mean_var_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+                    constexpr auto in_offset =
+                        x_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                    Update(mean_buf_m(Number<out_offset>{}),
+                           var_buf_m(Number<out_offset>{}),
+                           x_buf_m_k[Number<in_offset>{}]);
+                });
+            }
+        });
+    };
+
+    int cur_count_;
+    int max_count_;
+};
+
+template <typename T,
+          typename SrcMeanVarCountThreadDesc_M_K,
+          typename DstMeanVarThreadDesc_M,
+          bool GetActualVariance = false>
+struct ThreadwiseWelfordMerge
+{
+    static constexpr auto src_thread_desc_m_k = SrcMeanVarCountThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstMeanVarThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    __device__ static void
+    Merge(T& mean_a, T& var_a, int32_t& count_a, T mean_b, T var_b, int32_t count_b)
+    {
+        int count            = count_a + count_b;
+        T count_b_over_count = count == 0 ? type_convert<T>(0) : type_convert<T>(count_b) / count;
+        T delta              = mean_b - mean_a;
+        mean_a += delta * count_b_over_count;
+        var_a += var_b + delta * delta * count_a * count_b_over_count;
+        count_a = count;
+    }
+
+    template <typename SrcMeanBufferType,
+              typename SrcVarBufferType,
+              typename SrcCountBufferType,
+              typename DstMeanBufferType,
+              typename DstVarBufferType,
+              typename DstCountBufferType>
+    __device__ static void Run(const SrcMeanBufferType& src_mean_buf,
+                               const SrcVarBufferType& src_var_buf,
+                               const SrcCountBufferType& src_count_buf,
+                               DstMeanBufferType& dst_mean_buf,
+                               DstVarBufferType& dst_var_buf,
+                               DstCountBufferType& dst_count_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto src_offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Merge(dst_mean_buf(iM),
+                      dst_var_buf(iM),
+                      dst_count_buf(iM),
+                      src_mean_buf[Number<src_offset>{}],
+                      src_var_buf[Number<src_offset>{}],
+                      src_count_buf[Number<src_offset>{}]);
+            });
+
+            if constexpr(GetActualVariance)
+            {
+                dst_var_buf(iM) = dst_var_buf[iM] / dst_count_buf[iM];
+            };
+        });
+    };
+};
+
+} // namespace ck