Reorganize project folders (#6)

include/ck_tile/host/reference/reference_batched_dropout.hpp

@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename DataType, typename RandValOutputDataType>
+CK_TILE_HOST void reference_batched_dropout(HostTensor<DataType>& in_out_b_m_n,
+                                            const HostTensor<RandValOutputDataType>& randval_b_m_n,
+                                            const uint8_t& p_undrop_in_uint8_t,
+                                            const float scale)
+{
+    const int N = in_out_b_m_n.mDesc.get_lengths()[2];
+    auto f      = [&](auto batch, auto m) {
+        for(int n = 0; n < N; ++n)
+        {
+            float tmp = ck_tile::type_convert<float>(in_out_b_m_n(batch, m, n)) * scale;
+            in_out_b_m_n(batch, m, n) = randval_b_m_n(batch, m, n) <= p_undrop_in_uint8_t
+                                                 ? ck_tile::type_convert<DataType>(tmp)
+                                                 : DataType(0);
+        }
+    };
+
+    make_ParallelTensorFunctor(
+        f, randval_b_m_n.mDesc.get_lengths()[0], randval_b_m_n.mDesc.get_lengths()[1])(
+        std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_batched_elementwise.hpp

@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename AElementOp      = ck_tile::identity,
+          typename BElementOp      = ck_tile::identity,
+          typename BinaryElementOp = ck_tile::plus<AccDataType>>
+CK_TILE_HOST void reference_batched_elementwise(const HostTensor<ADataType>& a_b_m_n,
+                                                const HostTensor<BDataType>& b_b_m_n,
+                                                HostTensor<CDataType>& c_b_m_n,
+                                                const AElementOp& a_element_op           = {},
+                                                const BElementOp& b_element_op           = {},
+                                                const BinaryElementOp& binary_element_op = {})
+{
+    const ck_tile::index_t N = c_b_m_n.mDesc.get_lengths()[2];
+
+    const bool broadcast_a_dim_b = (a_b_m_n.get_lengths()[0] == 1);
+    const bool broadcast_a_dim_m = (a_b_m_n.get_lengths()[1] == 1);
+    const bool broadcast_a_dim_n = (a_b_m_n.get_lengths()[2] == 1);
+
+    const bool broadcast_b_dim_b = (b_b_m_n.get_lengths()[0] == 1);
+    const bool broadcast_b_dim_m = (b_b_m_n.get_lengths()[1] == 1);
+    const bool broadcast_b_dim_n = (b_b_m_n.get_lengths()[2] == 1);
+
+    auto f = [&](auto batch, auto m) {
+        for(ck_tile::index_t n = 0; n < N; ++n)
+        {
+            AccDataType v_a{};
+            {
+                ck_tile::index_t i_b = (broadcast_a_dim_b ? 0 : batch);
+                ck_tile::index_t i_m = (broadcast_a_dim_m ? 0 : m);
+                ck_tile::index_t i_n = (broadcast_a_dim_n ? 0 : n);
+
+                v_a = ck_tile::type_convert<AccDataType>(a_element_op(a_b_m_n(i_b, i_m, i_n)));
+            }
+
+            AccDataType v_b{};
+            {
+                ck_tile::index_t i_b = (broadcast_b_dim_b ? 0 : batch);
+                ck_tile::index_t i_m = (broadcast_b_dim_m ? 0 : m);
+                ck_tile::index_t i_n = (broadcast_b_dim_n ? 0 : n);
+
+                v_b = ck_tile::type_convert<AccDataType>(b_element_op(b_b_m_n(i_b, i_m, i_n)));
+            }
+
+            c_b_m_n(batch, m, n) = ck_tile::type_convert<CDataType>(binary_element_op(v_a, v_b));
+        }
+    };
+
+    make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(
+        std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_batched_gemm.hpp

@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename AElementOp   = ck_tile::identity,
+          typename BElementOp   = ck_tile::identity,
+          typename ACCElementOp = ck_tile::identity>
+CK_TILE_HOST void reference_batched_gemm(const HostTensor<ADataType>& a_b_m_k,
+                                         const HostTensor<BDataType>& b_b_n_k,
+                                         HostTensor<CDataType>& c_b_m_n,
+                                         const AElementOp& a_element_op     = {},
+                                         const BElementOp& b_element_op     = {},
+                                         const ACCElementOp& acc_element_op = {})
+{
+    const int N = b_b_n_k.mDesc.get_lengths()[1];
+    const int K = b_b_n_k.mDesc.get_lengths()[2];
+
+    auto f = [&](auto batch, auto m) {
+        for(int n = 0; n < N; ++n)
+        {
+            AccDataType v_acc = 0;
+
+            for(int k = 0; k < K; ++k)
+            {
+                ADataType v_a = a_element_op(a_b_m_k(batch, m, k));
+                BDataType v_b = b_element_op(b_b_n_k(batch, n, k));
+
+                v_acc += ck_tile::type_convert<AccDataType>(v_a) *
+                         ck_tile::type_convert<AccDataType>(v_b);
+            }
+
+            c_b_m_n(batch, m, n) = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
+        }
+    };
+
+    make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(
+        std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_batched_masking.hpp

@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename CDataType, typename MaskingType>
+CK_TILE_HOST void reference_batched_masking(HostTensor<CDataType>& c_b_m_n, const MaskingType& mask)
+{
+    const int M = c_b_m_n.mDesc.get_lengths()[1];
+    const int N = c_b_m_n.mDesc.get_lengths()[2];
+
+    auto f = [&](auto batch) {
+        for(int n = 0; n < N; ++n)
+        {
+            for(int m = 0; m < M; ++m)
+            {
+                if(mask.IsOutOfBound(m, n))
+                    c_b_m_n(batch, m, n) = -ck_tile::numeric<CDataType>::infinity();
+            }
+        }
+    };
+
+    make_ParallelTensorFunctor(f,
+                               c_b_m_n.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_batched_rotary_position_embedding.hpp

@@ -0,0 +1,73 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+
+#include <cassert>
+#include <thread>
+
+namespace ck_tile {
+
+template <typename DataType, typename ComputeDataType = float>
+CK_TILE_HOST void reference_batched_rotary_position_embedding(const HostTensor<DataType>& input_bsd,
+                                                              const HostTensor<DataType>& cos_sd,
+                                                              const HostTensor<DataType>& sin_sd,
+                                                              bool interleaved,
+                                                              HostTensor<DataType>& output_bsd,
+                                                              bool use_1_row_sin_cos = false)
+{
+    assert(cos_sd.get_num_of_dimension() == 2 && sin_sd.get_num_of_dimension() == 2);
+    assert(cos_sd.get_length(0) == sin_sd.get_length(0) &&
+           cos_sd.get_length(1) == sin_sd.get_length(1));
+
+    const index_t rotary_dim = cos_sd.get_length(1) * 2;
+    assert(static_cast<std::size_t>(rotary_dim) <= input_bsd.get_length(2));
+
+    output_bsd.ForEach([&](auto& self, auto i) {
+        const index_t i_d = i[2];
+        if(rotary_dim <= i_d)
+        {
+            self(i) = input_bsd(i);
+            return;
+        }
+        assert(i_d < rotary_dim);
+
+        const index_t i_s         = i[1];
+        const index_t i_s_cos_sin = (use_1_row_sin_cos ? 0 : i_s);
+
+        const ComputeDataType cos = type_convert<ComputeDataType>(
+            interleaved ? cos_sd(i_s_cos_sin, i_d / 2)
+                        : cos_sd(i_s_cos_sin, i_d % cos_sd.get_length(1)));
+        const ComputeDataType sin = type_convert<ComputeDataType>(
+            interleaved ? sin_sd(i_s_cos_sin, i_d / 2)
+                        : sin_sd(i_s_cos_sin, i_d % sin_sd.get_length(1)));
+
+        const ComputeDataType half_rotated_input = [&] {
+            const index_t i_b = i[0];
+
+            if(interleaved)
+            {
+                const bool is_even         = (i_d % 2 == 0);
+                const index_t pos          = i_d + (is_even ? 1 : -1);
+                const ComputeDataType sign = (is_even ? -1 : 1);
+                return sign * type_convert<ComputeDataType>(input_bsd(i_b, i_s, pos));
+            }
+            else
+            {
+                const index_t half_rdim    = (rotary_dim / 2);
+                const index_t pos          = (i_d + half_rdim) % rotary_dim;
+                const ComputeDataType sign = (pos < half_rdim ? 1 : -1);
+                return sign * type_convert<ComputeDataType>(input_bsd(i_b, i_s, pos));
+            }
+        }();
+        ComputeDataType result =
+            type_convert<ComputeDataType>(input_bsd(i)) * cos + half_rotated_input * sin;
+
+        self(i) = type_convert<DataType>(result);
+    });
+}
+
+} // namespace ck_tile

include/ck_tile/host/reference/reference_batched_softmax.hpp

@@ -0,0 +1,71 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename ADataType,
+          typename CompDataType,
+          typename BDataType,
+          typename CompElementOp = ck_tile::identity>
+CK_TILE_HOST void reference_batched_softmax(
+    const HostTensor<ADataType>& a_b_m_n,
+    HostTensor<BDataType>& b_b_m_n,
+    const CompElementOp& comp_element_op                                    = {},
+    std::optional<std::reference_wrapper<HostTensor<CompDataType>>> lse_b_m = std::nullopt)
+{
+    const int N = a_b_m_n.mDesc.get_lengths()[2];
+
+    auto f = [&](auto batch, auto m) {
+        CompDataType v_max = -ck_tile::numeric<CompDataType>::infinity();
+
+        // max
+        for(int n = 0; n < N; ++n)
+        {
+            const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
+
+            v_max = v_max < v_a ? v_a : v_max;
+        }
+
+        CompDataType v_exp_sum = 0;
+        // validate v_max if all the elements within a row are -INF
+        if(std::isinf(v_max) && v_max < 0)
+        {
+            v_max = ck_tile::type_convert<CompDataType>(0.f);
+        }
+
+        // sum
+        for(int n = 0; n < N; ++n)
+        {
+            const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
+
+            v_exp_sum += ck_tile::exp(v_a - v_max);
+        }
+
+        // if sum is zero(masked), or nan/inf(other computation error), don't do divide
+        CompDataType inv_sum = (v_exp_sum == 0.f ? 1.f : 1.f / v_exp_sum);
+
+        // elementwise
+        for(int n = 0; n < N; ++n)
+        {
+            const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
+            const CompDataType v_b = ck_tile::exp(v_a - v_max) * inv_sum;
+
+            b_b_m_n(batch, m, n) = ck_tile::type_convert<BDataType>(comp_element_op(v_b));
+        }
+        // lse
+        if(lse_b_m)
+        {
+            lse_b_m->get()(batch, m) = v_max + ck_tile::log(v_exp_sum);
+        }
+    };
+
+    make_ParallelTensorFunctor(f, b_b_m_n.mDesc.get_lengths()[0], b_b_m_n.mDesc.get_lengths()[1])(
+        std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_batched_transpose.hpp

@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename Type>
+CK_TILE_HOST void reference_batched_transpose(const HostTensor<Type>& x,
+                                              HostTensor<Type>& y,
+                                              std::string layout_in  = "NCHW",
+                                              std::string layout_out = "NHWC")
+{
+    const int N = x.mDesc.get_lengths()[0];
+
+    auto f = [&](auto batch) {
+        if(layout_in == "NCHW" && layout_out == "NHWC")
+        {
+            const int C = x.mDesc.get_lengths()[1];
+            const int H = x.mDesc.get_lengths()[2];
+            const int W = x.mDesc.get_lengths()[3];
+            for(int c = 0; c < C; ++c)
+            {
+                for(int h = 0; h < H; ++h)
+                {
+                    for(int w = 0; w < W; ++w)
+                    {
+                        Type v_x          = x(batch, c, h, w);
+                        y(batch, h, w, c) = v_x;
+                    }
+                }
+            }
+        }
+        else if(layout_in == "NHWC" && layout_out == "NCHW")
+        {
+            const int H = x.mDesc.get_lengths()[1];
+            const int W = x.mDesc.get_lengths()[2];
+            const int C = x.mDesc.get_lengths()[3];
+            for(int h = 0; h < H; ++h)
+            {
+                for(int w = 0; w < W; ++w)
+                {
+                    for(int c = 0; c < C; ++c)
+                    {
+                        Type v_x          = x(batch, h, w, c);
+                        y(batch, c, h, w) = v_x;
+                    }
+                }
+            }
+        }
+    };
+
+    make_ParallelTensorFunctor(f, N)(std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_elementwise.hpp

@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+template <typename ADataType, typename BDataType, typename ComputeDataType, typename ElementOp>
+CK_TILE_HOST void reference_unary_elementwise(const HostTensor<ADataType>& a,
+                                              HostTensor<BDataType>& b,
+                                              ElementOp element_op)
+{
+    // TODO: imeplement gpu version reference function
+    auto f = [&](auto i) {
+        auto v_a   = type_convert<ComputeDataType>(a.mData[i]);
+        auto v_b   = element_op(v_a);
+        b.mData[i] = ck_tile::type_convert<BDataType>(v_b);
+    };
+
+    make_ParallelTensorFunctor(f, b.get_element_space_size())(std::thread::hardware_concurrency());
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ComputeDataType,
+          typename ElementOp>
+CK_TILE_HOST void reference_binary_elementwise(const HostTensor<ADataType>& a,
+                                               const HostTensor<BDataType>& b,
+                                               HostTensor<CDataType>& c,
+                                               ElementOp element_op)
+{
+    // TODO: imeplement gpu version reference function
+    auto f = [&](auto i) {
+        auto v_a   = type_convert<ComputeDataType>(a.mData[i]);
+        auto v_b   = type_convert<ComputeDataType>(b.mData[i]);
+        auto v_c   = element_op(v_a, v_b);
+        c.mData[i] = ck_tile::type_convert<CDataType>(v_c);
+    };
+
+    make_ParallelTensorFunctor(f, c.get_element_space_size())(std::thread::hardware_concurrency());
+}
+
+} // namespace ck_tile

include/ck_tile/host/reference/reference_fused_moe.hpp

@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+
+namespace ck_tile {
+// [indexing implementation-1]
+// using M_a as constexpr block_size to partition all tokens into different slices
+// each slice map to one expert, and one expert can have multiple slices
+// e.g. num_experts = 6, topk=3, M_a = 4, input_tokens = 5
+// before sort, topk_ids is : [[0, 3, 5], [2, 3, 5], [1, 3, 5], [1, 2, 3], [1, 3, 5]]
+//                            tok-0      tok-1      tok-2      tok-3      tok-4
+//           topk_weight is : [[a, b, c], [d, e, f], [g, h, i], [j, k, l], [m, n, o]] (some float
+//           number)
+//
+// token_id_per_expert is : [[0], [2, 3, 4], [1, 3], [0, 1, 2, 3, 4], [], [0, 1, 2, 5]]
+//  (only for reference)    exp-0  exp-1     exp-2   exp-3          exp-4  exp-5
+// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]]
+//
+// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1)
+// max_num_tokens_padded : topk * input_tokens + num_experts * M_a - topk (updated)
+// * this could be larger than actual, since actual tokens are on GPU
+//
+// sorted_token_ids_ptr   : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6,
+// 0, 1, 2, 5]
+//                          |-  exp-0  -|-  exp-1  -|-  exp-2  -|-      exp-3          -|-  exp-4
+//                          -|-  exp-5  -|
+// sorted_weight_ptr      : [a, *, *, *, g, j, m, *, d, k, *, *, b, e, h, l, n, *, *, *, *, *, *, *,
+// c, f, i, o]
+//
+// * length is max_num_tokens_padded, actual size is num_tokens_post_padded_ptr
+//
+// sorted_expert_ids_ptr  : [0, 1, 2, 3, 3, 4, 5]
+// * length is (max_num_tokens_padded + block_size - 1) / block_size
+///
+// num_tokens_post_padded_ptr : [28]
+// num_sorted_tiles_ptr : [7]
+
+template <typename AccDataType, // you only need to explcitly set this one
+          typename Activation,  // ck_tile::element_wise::Gelu
+          typename ADataType,
+          typename GDataType,
+          typename DDataType,
+          typename ODataType,
+          typename AScaleDataType,
+          typename GScaleDataType,
+          typename DScaleDataType,
+          typename YSmoothScaleDataType,
+          typename TopkWeightDataType,
+          typename IndexDataType>
+void reference_fused_moe(
+    const ck_tile::HostTensor<ADataType>& a_host,       // [tokens, hidden_size]
+    const ck_tile::HostTensor<GDataType>& g_host,       // [experts, interme_size_0, hidden_size]
+    const ck_tile::HostTensor<DDataType>& d_host,       // [experts, hidden_size, interme_size_1]
+    const ck_tile::HostTensor<AScaleDataType>& sa_host, // [tokens, 1],
+    const ck_tile::HostTensor<GScaleDataType>& sg_host, // [experts, 1, interme_size_0]
+    const ck_tile::HostTensor<DScaleDataType>& sd_host, // [experts, 1, hidden_size],
+    const ck_tile::HostTensor<YSmoothScaleDataType>& sy_host,        // [experts, 1, interme_size_0]
+    ck_tile::HostTensor<ODataType>& o_host,                          // [tokens, hidden_size]
+    const ck_tile::HostTensor<IndexDataType>& sorted_token_ids_host, // [max_num_tokens_padded]
+    const ck_tile::HostTensor<TopkWeightDataType>& sorted_weight_host, // [max_num_tokens_padded]
+    const ck_tile::HostTensor<IndexDataType>&
+        sorted_expert_ids_host, // [(max_num_tokens_padded + block_size - 1) / block_size]
+    const ck_tile::HostTensor<IndexDataType>& num_sorted_tiles_host, // [1]
+
+    const ck_tile::HostTensor<IndexDataType>&
+        token_ids_host, // [tokens, topk] --> ugly!!! remove in the future
+
+    ck_tile::index_t block_m,
+    ck_tile::index_t tokens,
+    ck_tile::index_t experts,
+    ck_tile::index_t hidden_size,
+    ck_tile::index_t intermediate_size, // this size is for gate/up/down
+    ck_tile::index_t topk,
+    ck_tile::index_t gate_only)
+{
+    assert(sorted_token_ids_host.get_num_of_dimension() == 1);
+    assert(sorted_weight_host.get_num_of_dimension() == 1);
+    assert(sorted_expert_ids_host.get_num_of_dimension() == 1);
+    assert(num_sorted_tiles_host.get_element_size() == 1);
+    ck_tile::index_t num_sorted_tiles    = num_sorted_tiles_host.mData[0] / block_m;
+    ck_tile::index_t intermediate_size_0 = intermediate_size * (gate_only ? 1 : 2);
+    ck_tile::index_t intermediate_size_1 = intermediate_size;
+
+    ck_tile::HostTensor<AccDataType> out_topk_tokens({tokens, topk, hidden_size});
+
+    int max_num_tokens_padded = topk * tokens + experts * block_m - topk;
+    // assert();
+    auto f = [&](auto i_flatten) {
+        ck_tile::index_t i_tile = i_flatten / block_m;
+        if(i_tile >= num_sorted_tiles)
+            return;
+        ck_tile::index_t i_expert = sorted_expert_ids_host.mData[i_tile];
+
+#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
+        ck_tile::index_t i_token = sorted_token_ids_host.mData[i_flatten];
+        ck_tile::index_t i_topk  = i_token >> 24;
+        i_token &= 0xffffff;
+        if(i_token >= tokens)
+            return;
+        (void)token_ids_host;
+#else
+        // TODO: better remove this in the future, or modify the token_id value
+        auto get_topk_id = [&](ck_tile::index_t token_id_, ck_tile::index_t expert_id_) {
+            for(ck_tile::index_t i_ = 0; i_ < topk; i_++)
+            {
+                if(token_ids_host(token_id_, i_) == expert_id_)
+                    return i_;
+            }
+            throw std::runtime_error("not correct token/expert pair\n");
+            return -1; // TODO: not correct!!
+        };
+        ck_tile::index_t i_token = sorted_token_ids_host.mData[i_flatten];
+        if(i_token >= tokens)
+            return;
+        ck_tile::index_t i_topk = get_topk_id(i_token, i_expert); // TODO: ugly
+#endif
+        auto weight = sorted_weight_host.mData[i_flatten];
+
+        ck_tile::HostTensor<AccDataType> acc_0({1, intermediate_size_0});
+        // first gemm
+        for(ck_tile::index_t i_n = 0; i_n < intermediate_size_0; i_n++)
+        {
+            AccDataType acc = static_cast<AccDataType>(0);
+            for(ck_tile::index_t i_k = 0; i_k < hidden_size; i_k++)
+            {
+                acc += type_convert<AccDataType>(a_host(i_token, i_k)) *
+                       type_convert<AccDataType>(g_host(i_expert, i_n, i_k));
+            }
+            acc_0(0, i_n) = acc;
+            // printf("ie:%2d, it:%3d, in:%d, %f\n", i_expert, i_token, i_n, acc);
+        }
+
+        ck_tile::HostTensor<AccDataType> y({1, intermediate_size_1});
+        if(gate_only)
+        {
+            if(intermediate_size_1 != intermediate_size_0)
+                throw std::runtime_error(
+                    "intermediate_size not correct, 0:" + std::to_string(intermediate_size_0) +
+                    ", 1:" + std::to_string(intermediate_size_1));
+            for(ck_tile::index_t i_n = 0; i_n < intermediate_size_1; i_n++)
+            {
+                Activation{}(y(0, i_n), acc_0(0, i_n));
+                // printf("ie:%2d, it:%3d, in:%d, %f\n", i_expert, i_token, i_n, y(0, i_n));
+            }
+        }
+        else
+        {
+            if(intermediate_size_1 * 2 != intermediate_size_0)
+                throw std::runtime_error(
+                    "intermediate_size not correct, 0:" + std::to_string(intermediate_size_0) +
+                    ", 1:" + std::to_string(intermediate_size_1));
+            for(ck_tile::index_t i_n = 0; i_n < intermediate_size_1; i_n++)
+            {
+                AccDataType tmp;
+                Activation{}(tmp, acc_0(0, i_n));
+                y(0, i_n) = tmp * acc_0(0, i_n + intermediate_size_1); // TODO: elementwise mul
+            }
+        }
+
+        // second gemm, loop along gemm-n
+        ck_tile::HostTensor<AccDataType> acc_1({1, hidden_size});
+        for(ck_tile::index_t i_n = 0; i_n < hidden_size; i_n++)
+        {
+            AccDataType acc = static_cast<AccDataType>(0);
+            for(ck_tile::index_t i_k = 0; i_k < intermediate_size_1; i_k++)
+            {
+                acc += y(0, i_k) * type_convert<AccDataType>(d_host(i_expert, i_n, i_k));
+            }
+            acc_1(0, i_n) = acc * weight; // multiple weight here
+        }
+
+        for(ck_tile::index_t i_n = 0; i_n < hidden_size; i_n++)
+        {
+            out_topk_tokens(i_token, i_topk, i_n) = acc_1(0, i_n);
+        }
+    };
+
+    // make_ParallelTensorFunctor(f, max_num_tokens_padded)(std::thread::hardware_concurrency());
+    make_ParallelTensorFunctor(f, max_num_tokens_padded)(1);
+
+    // reduce
+    auto r = [&](auto i_token) {
+        for(ck_tile::index_t i_n = 0; i_n < hidden_size; i_n++)
+        {
+            AccDataType acc = type_convert<AccDataType>(0);
+            for(ck_tile::index_t i_topk = 0; i_topk < topk; i_topk++)
+            {
+                acc += out_topk_tokens(i_token, i_topk, i_n);
+            }
+            o_host(i_token, i_n) = type_convert<ODataType>(acc);
+        }
+    };
+    make_ParallelTensorFunctor(r, tokens)(std::thread::hardware_concurrency());
+
+    (void)num_sorted_tiles_host;
+    (void)sa_host;
+    (void)sg_host;
+    (void)sd_host;
+    (void)sy_host;
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_gemm.hpp

@@ -0,0 +1,211 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <thread>
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+
+namespace ck_tile {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename AElementOp   = ck_tile::identity,
+          typename BElementOp   = ck_tile::identity,
+          typename ACCElementOp = ck_tile::identity>
+CK_TILE_HOST void reference_gemm(const HostTensor<ADataType>& a_m_k,
+                                 const HostTensor<BDataType>& b_k_n,
+                                 HostTensor<CDataType>& c_m_n,
+                                 const AElementOp& a_element_op     = {},
+                                 const BElementOp& b_element_op     = {},
+                                 const ACCElementOp& acc_element_op = {})
+{
+    const std::size_t M = a_m_k.get_length(0);
+    const std::size_t N = b_k_n.get_length(1);
+    const std::size_t K = a_m_k.get_length(1);
+
+    auto f_mn = [&](auto m, auto n) {
+        AccDataType v_acc = 0;
+
+        for(std::size_t k = 0; k < K; ++k)
+        {
+            AccDataType v_a;
+            AccDataType v_b;
+            if constexpr(std::is_same_v<ADataType, pk_int4_t>)
+            {
+                const pk_int4_t pk_val  = a_element_op(a_m_k(m, k));
+                const fp32x2_t fp32_val = pk_int4_t_to_fp32x2_t(pk_val);
+                if(k % 2 == 1)
+                    v_a = fp32_val.hi;
+                else
+                    v_a = fp32_val.lo;
+            }
+            else
+            {
+                v_a = ck_tile::type_convert<AccDataType>(a_element_op(a_m_k(m, k)));
+            }
+            if constexpr(std::is_same_v<BDataType, pk_int4_t>)
+            {
+                const pk_int4_t pk_val  = b_element_op(b_k_n(k, n));
+                const fp32x2_t fp32_val = pk_int4_t_to_fp32x2_t(pk_val);
+                if(k % 2 == 1)
+                    v_b = fp32_val.hi;
+                else
+                    v_b = fp32_val.lo;
+            }
+            else
+            {
+                v_b = ck_tile::type_convert<AccDataType>(b_element_op(b_k_n(k, n)));
+            }
+            v_acc += v_a * v_b;
+        }
+
+        c_m_n(m, n) = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
+    };
+
+    make_ParallelTensorFunctor(f_mn, M, N)(std::thread::hardware_concurrency());
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename LayoutA,
+          typename LayoutB,
+          typename LayoutC>
+__global__ void naive_gemm_kernel(ADataType* A,
+                                  BDataType* B,
+                                  CDataType* C,
+                                  ck_tile::index_t M,
+                                  ck_tile::index_t N,
+                                  ck_tile::index_t K,
+                                  ck_tile::index_t strideA,
+                                  ck_tile::index_t strideB,
+                                  ck_tile::index_t strideC)
+{
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int row = idx / N; // Compute row index
+    int col = idx % N; // Compute column index
+
+    if(row < M && col < N)
+    {
+        AccDataType acc = 0.0;
+        for(int k = 0; k < K; ++k)
+        {
+            constexpr index_t packed_size_a = ck_tile::numeric_traits<ADataType>::PackedSize;
+            constexpr index_t packed_size_b = ck_tile::numeric_traits<BDataType>::PackedSize;
+            // Adjust indexing based on matrix layout
+            int a_index = (std::is_same_v<LayoutA, tensor_layout::gemm::RowMajor>)
+                              ? row * strideA + k
+                              : k * strideA + row;
+            int b_index = (std::is_same_v<LayoutB, tensor_layout::gemm::ColumnMajor>)
+                              ? col * strideB + k
+                              : k * strideB + col;
+
+            AccDataType v_a;
+            AccDataType v_b;
+            if constexpr(std::is_same_v<ADataType, pk_int4_t>)
+            {
+                const fp32x2_t fp32_val = pk_int4_t_to_fp32x2_t(A[a_index / packed_size_a]);
+                if(k % 2 == 1)
+                    v_a = fp32_val.hi;
+                else
+                    v_a = fp32_val.lo;
+            }
+            else
+            {
+                v_a = ck_tile::type_convert<AccDataType>(A[a_index]);
+            }
+            if constexpr(std::is_same_v<BDataType, pk_int4_t>)
+            {
+                const fp32x2_t fp32_val = pk_int4_t_to_fp32x2_t(B[b_index / packed_size_b]);
+                if(k % 2 == 1)
+                    v_b = fp32_val.hi;
+                else
+                    v_b = fp32_val.lo;
+            }
+            else
+            {
+                v_b = ck_tile::type_convert<AccDataType>(B[b_index]);
+            }
+            acc += v_a * v_b;
+        }
+
+        int c_index = (std::is_same_v<LayoutC, tensor_layout::gemm::RowMajor>)
+                          ? row * strideC + col
+                          : col * strideC + row;
+        C[c_index]  = ck_tile::type_convert<CDataType>(acc);
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename LayoutA,
+          typename LayoutB,
+          typename LayoutC>
+void reference_gemm_gpu(ADataType* a_ptr,
+                        BDataType* b_ptr,
+                        CDataType* c_ptr,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t stride_a,
+                        index_t stride_b,
+                        index_t stride_c)
+{
+    int totalElements      = M * N;
+    int numThreadsPerBlock = 256; // Common choice for threads per block
+    int numBlocks          = (totalElements + numThreadsPerBlock - 1) / numThreadsPerBlock;
+
+    naive_gemm_kernel<ADataType, BDataType, AccDataType, CDataType, LayoutA, LayoutB, LayoutC>
+        <<<numBlocks, numThreadsPerBlock>>>(
+            a_ptr, b_ptr, c_ptr, M, N, K, stride_a, stride_b, stride_c);
+
+    return;
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename LayoutA,
+          typename LayoutB,
+          typename LayoutC>
+void reference_batched_gemm_gpu(ADataType* a_ptr,
+                                BDataType* b_ptr,
+                                CDataType* c_ptr,
+                                index_t M,
+                                index_t N,
+                                index_t K,
+                                index_t stride_a,
+                                index_t stride_b,
+                                index_t stride_c,
+                                index_t batch_stride_A,
+                                index_t batch_stride_B,
+                                index_t batch_stride_C,
+                                index_t batch_count)
+{
+    int totalElements      = M * N;
+    int numThreadsPerBlock = 256; // Common choice for threads per block
+    int numBlocks          = (totalElements + numThreadsPerBlock - 1) / numThreadsPerBlock;
+
+    for(index_t batch_id = 0; batch_id < batch_count; ++batch_id)
+    {
+        ADataType* d_ATemp = a_ptr + batch_id * batch_stride_A;
+        BDataType* d_BTemp = b_ptr + batch_id * batch_stride_B;
+        CDataType* d_CTemp = c_ptr + batch_id * batch_stride_C;
+        naive_gemm_kernel<ADataType, BDataType, AccDataType, CDataType, LayoutA, LayoutB, LayoutC>
+            <<<numBlocks, numThreadsPerBlock>>>(
+                d_ATemp, d_BTemp, d_CTemp, M, N, K, stride_a, stride_b, stride_c);
+    }
+
+    return;
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_im2col.hpp

@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename InDataType, typename OutDataType, index_t NDimSpatial>
+CK_TILE_HOST void reference_im2col(const HostTensor<InDataType>& in_host,
+                                   HostTensor<OutDataType>& out_host,
+                                   const ck_tile::conv::ConvParam& conv_params)
+{
+    const long_index_t G = in_host.get_lengths()[0];
+    const long_index_t N = in_host.get_lengths()[1];
+    const long_index_t C = in_host.get_lengths()[2];
+
+    if constexpr(NDimSpatial == 1)
+    {
+        const long_index_t Wo = conv_params.output_spatial_lengths_[0];
+        auto func             = [&](auto g, auto n, auto wo) {
+            long_index_t row    = n * Wo + wo;
+            long_index_t column = 0;
+
+            for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[0]; ++x)
+            {
+                auto wi = static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[0]) +
+                          static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[0]) -
+                          static_cast<long_index_t>(conv_params.input_left_pads_[0]);
+
+                for(long_index_t c = 0; c < C; ++c)
+                {
+                    if(wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[3])
+                    {
+                        InDataType v_in          = in_host(g, n, c, wi);
+                        out_host(g, row, column) = type_convert<OutDataType>(v_in);
+                    }
+                    column++;
+                }
+            }
+        };
+
+        make_ParallelTensorFunctor(func, G, N, Wo)(std::thread::hardware_concurrency());
+    }
+    else if constexpr(NDimSpatial == 2)
+    {
+        const long_index_t Ho = conv_params.output_spatial_lengths_[0];
+        const long_index_t Wo = conv_params.output_spatial_lengths_[1];
+
+        auto func = [&](auto g, auto n, auto ho, auto wo) {
+            long_index_t row    = n * Ho * Wo + ho * Wo + wo;
+            long_index_t column = 0;
+
+            for(long_index_t y = 0; y < conv_params.filter_spatial_lengths_[0]; ++y)
+            {
+                auto hi = static_cast<long_index_t>(ho * conv_params.conv_filter_strides_[0]) +
+                          static_cast<long_index_t>(y * conv_params.conv_filter_dilations_[0]) -
+                          static_cast<long_index_t>(conv_params.input_left_pads_[0]);
+
+                for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[1]; ++x)
+                {
+                    auto wi = static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[1]) +
+                              static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[1]) -
+                              static_cast<long_index_t>(conv_params.input_left_pads_[1]);
+
+                    for(long_index_t c = 0; c < C; ++c)
+                    {
+
+                        if(hi >= 0 && type_convert<std::size_t>(hi) < in_host.get_lengths()[3] &&
+                           wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[4])
+                        {
+                            InDataType v_in          = in_host(g, n, c, hi, wi);
+                            out_host(g, row, column) = type_convert<OutDataType>(v_in);
+                        }
+                        column++;
+                    }
+                }
+            }
+        };
+
+        make_ParallelTensorFunctor(func, G, N, Ho, Wo)(std::thread::hardware_concurrency());
+    }
+    else if constexpr(NDimSpatial == 3)
+    {
+        const long_index_t Do = conv_params.output_spatial_lengths_[0];
+        const long_index_t Ho = conv_params.output_spatial_lengths_[1];
+        const long_index_t Wo = conv_params.output_spatial_lengths_[2];
+
+        auto func = [&](auto g, auto n, auto d_o, auto ho, auto wo) {
+            long_index_t row    = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo;
+            long_index_t column = 0;
+
+            for(long_index_t z = 0; z < conv_params.filter_spatial_lengths_[0]; ++z)
+            {
+                auto di = static_cast<long_index_t>(d_o * conv_params.conv_filter_strides_[0]) +
+                          static_cast<long_index_t>(z * conv_params.conv_filter_dilations_[0]) -
+                          static_cast<long_index_t>(conv_params.input_left_pads_[0]);
+                for(long_index_t y = 0; y < conv_params.filter_spatial_lengths_[1]; ++y)
+                {
+                    auto hi = static_cast<long_index_t>(ho * conv_params.conv_filter_strides_[1]) +
+                              static_cast<long_index_t>(y * conv_params.conv_filter_dilations_[1]) -
+                              static_cast<long_index_t>(conv_params.input_left_pads_[1]);
+                    for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[2]; ++x)
+                    {
+                        auto wi =
+                            static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[2]) +
+                            static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[2]) -
+                            static_cast<long_index_t>(conv_params.input_left_pads_[2]);
+                        for(long_index_t c = 0; c < C; ++c)
+                        {
+                            if(di >= 0 &&
+                               type_convert<std::size_t>(di) < in_host.get_lengths()[3] &&
+                               hi >= 0 &&
+                               type_convert<std::size_t>(hi) < in_host.get_lengths()[4] &&
+                               wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[5])
+                            {
+                                InDataType v_in          = in_host(g, n, c, di, hi, wi);
+                                out_host(g, row, column) = type_convert<OutDataType>(v_in);
+                            }
+                            column++;
+                        }
+                    }
+                }
+            }
+        };
+
+        make_ParallelTensorFunctor(func, G, N, Do, Ho, Wo)(std::thread::hardware_concurrency());
+    }
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_layernorm2d_fwd.hpp

@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+
+namespace ck_tile {
+
+// Note: for simplicity, each functor only care about single M
+struct reference_layernorm2d_default_epilogue
+{
+    template <typename OutDataType, typename AccDataType>
+    void operator()(int m, HostTensor<OutDataType>& o, const HostTensor<AccDataType>& acc)
+    {
+        const int N = acc.mDesc.get_lengths()[1];
+        for(int n = 0; n < N; ++n)
+        {
+            o(m, n) = ck_tile::type_convert<OutDataType>(acc(m, n));
+        }
+    }
+
+    template <typename OutDataType, typename AccDataType>
+    auto operator()(int m, const HostTensor<AccDataType>& acc)
+    {
+        HostTensor<OutDataType> o(acc.get_lengths(), acc.get_strides());
+        operator()(m, o, acc);
+        return o;
+    }
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename YDataType,
+          typename MeanDataType,
+          typename InvStdDataType,
+          typename Epilogue = reference_layernorm2d_default_epilogue>
+void reference_layernorm2d_fwd(const HostTensor<XDataType>& x_m_n,
+                               const HostTensor<GammaDataType>& gamma_n,
+                               const HostTensor<BetaDataType>& beta_n,
+                               HostTensor<YDataType>& y_m_n,
+                               HostTensor<MeanDataType>& mean_m,
+                               HostTensor<InvStdDataType>& invStd_m,
+                               ComputeDataType epsilon,
+                               Epilogue epilogue_functor = {})
+{
+    auto layernorm2d_fwd_func = [&](auto m) {
+        const int N = x_m_n.mDesc.get_lengths()[1];
+
+        int count                = 0;
+        ComputeDataType mean     = 0;
+        ComputeDataType variance = 0;
+        ComputeDataType divisor  = 0;
+
+        for(int n = 0; n < N; ++n)
+        {
+            ++count;
+            ComputeDataType x     = ck_tile::type_convert<ComputeDataType>(x_m_n(m, n));
+            ComputeDataType delta = x - mean;
+            mean += delta / count;
+            ComputeDataType delta2 = x - mean;
+            variance += delta * delta2;
+        }
+
+        // actual variance
+        variance = variance / count;
+        divisor  = ck_tile::type_convert<ComputeDataType>(1) / ck_tile::sqrt(variance + epsilon);
+
+        if constexpr(!std::is_same_v<MeanDataType, ck_tile::null_type>)
+            mean_m(m) = ck_tile::type_convert<MeanDataType>(mean);
+
+        if constexpr(!std::is_same_v<InvStdDataType, ck_tile::null_type>)
+            invStd_m(m) = ck_tile::type_convert<InvStdDataType>(divisor);
+
+        HostTensor<ComputeDataType> acc(x_m_n.get_lengths(), x_m_n.get_strides());
+        for(int n = 0; n < N; ++n)
+        {
+            ComputeDataType x     = ck_tile::type_convert<ComputeDataType>(x_m_n(m, n));
+            ComputeDataType gamma = ck_tile::type_convert<ComputeDataType>(gamma_n(n));
+            ComputeDataType beta  = ck_tile::type_convert<ComputeDataType>(beta_n(n));
+            auto a_               = (x - mean) * divisor;
+            a_                    = a_ * gamma + beta;
+
+            acc(m, n) = a_;
+        }
+
+        epilogue_functor(m, y_m_n, acc);
+    };
+
+    make_ParallelTensorFunctor(layernorm2d_fwd_func,
+                               mean_m.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_moe_sorting.hpp

@@ -0,0 +1,119 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+
+namespace ck_tile {
+
+#define MOE_SORTING_MOCK_ID(token_id_, topk_id_) \
+    static_cast<uint32_t>(((token_id_)&0x00ffffff) | (((topk_id_)&0xff) << 24))
+
+template <typename WeightType, typename IndexType = index_t>
+CK_TILE_HOST void reference_moe_sorting(const HostTensor<IndexType>& topk_ids,
+                                        const HostTensor<WeightType>& weights,
+                                        const HostTensor<IndexType>& local_expert_mask,
+                                        HostTensor<IndexType>& p_sorted_token_ids,
+                                        HostTensor<WeightType>& sorted_weight,
+                                        HostTensor<IndexType>& sorted_expert_ids,
+                                        index_t& unit_cnt,
+                                        const index_t experts,
+                                        const index_t unit_size,
+                                        bool local_expert_masking,
+                                        bool skip_experts_with_zero_token = true)
+{
+    const index_t num_token = topk_ids.mDesc.get_lengths()[0];
+    const index_t topk      = topk_ids.mDesc.get_lengths()[1];
+    // allocate a temp buffer, and fill the value with [number_token|topk]
+    std::vector<std::vector<IndexType>> expert_tokens(
+        experts,
+#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
+        std::vector<IndexType>(unit_size, MOE_SORTING_MOCK_ID(num_token, topk)));
+#else
+        std::vector<IndexType>(unit_size, num_token));
+#endif
+    std::vector<std::vector<WeightType>> expert_token_weights(
+        experts, std::vector<WeightType>(unit_size, 0));
+    // count number of unit-size slices in this expert
+    std::vector<IndexType> expert_slices(experts, 1);
+    // count the tokens used in this expert
+    std::vector<IndexType> expert_slice_idxs(experts, 0);
+    // TODO: above 2 buffer seems duplicated
+
+    for(index_t t = 0; t < num_token; t++)
+    {
+        for(index_t k = 0; k < topk; k++)
+        {
+            IndexType e  = topk_ids(t, k);
+            WeightType w = weights(t, k);
+            index_t idx  = expert_slice_idxs[e];
+            if(idx > expert_slices[e] * unit_size - 1)
+            {
+                expert_slices[e]++;
+                index_t new_size = expert_slices[e] * unit_size;
+                expert_tokens[e].resize(new_size);
+                expert_token_weights[e].resize(new_size);
+                for(index_t i = (expert_slices[e] - 1) * unit_size; i < new_size; i++)
+                {
+#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
+                    expert_tokens[e][i] = MOE_SORTING_MOCK_ID(num_token, topk);
+#else
+                    expert_tokens[e][i] = num_token;
+#endif
+                    expert_token_weights[e][i] = 0;
+                }
+            }
+#if CK_TILE_REFERENCE_MOE_SORTING_MOCK_ID
+            expert_tokens[e][idx] = MOE_SORTING_MOCK_ID(t, k);
+#else
+            expert_tokens[e][idx] = t;
+#endif
+            expert_token_weights[e][idx] = w;
+            expert_slice_idxs[e]++;
+        }
+    }
+
+    IndexType* out_tokens    = p_sorted_token_ids.data();
+    WeightType* out_weights  = sorted_weight.data();
+    IndexType* out_expert_id = sorted_expert_ids.data();
+    int curr_expert_id       = 0;
+    for(index_t e = 0; e < experts; e++)
+    {
+        if(local_expert_masking)
+        {
+            if(local_expert_mask(e) == 0)
+                continue;
+        }
+        if(skip_experts_with_zero_token)
+        {
+            if(expert_slice_idxs[e] == 0)
+            {
+                curr_expert_id++;
+                continue;
+            }
+        }
+
+        memcpy(out_tokens, expert_tokens[e].data(), sizeof(index_t) * expert_slices[e] * unit_size);
+        out_tokens += expert_slices[e] * unit_size;
+        memcpy(out_weights,
+               expert_token_weights[e].data(),
+               sizeof(WeightType) * expert_slices[e] * unit_size);
+        out_weights += expert_slices[e] * unit_size;
+
+        for(index_t s = 0; s < expert_slices[e]; s++)
+        {
+            out_expert_id[s] = curr_expert_id;
+            unit_cnt++;
+        }
+        out_expert_id += expert_slices[e];
+        curr_expert_id++;
+    }
+    unit_cnt *= unit_size;
+    return;
+}
+
+#undef MOE_SORTING_MOCK_ID
+
+} // namespace ck_tile

include/ck_tile/host/reference/reference_permute.hpp

@@ -0,0 +1,76 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+#include <numeric>
+#include <functional>
+
+namespace ck_tile {
+
+/*
+    this will do permute + contiguous like functionality in pytorch
+*/
+template <typename DataType>
+CK_TILE_HOST void
+reference_permute(const HostTensor<DataType>& x, HostTensor<DataType>& y, std::vector<index_t> perm)
+{
+    const auto x_len = x.mDesc.get_lengths();
+    const auto y_len = y.mDesc.get_lengths();
+    assert(x_len.size() == y_len.size());
+    index_t rank     = x_len.size();
+    const auto x_elm = std::accumulate(x_len.begin(), x_len.end(), 1, std::multiplies<index_t>());
+    const auto y_elm = std::accumulate(y_len.begin(), y_len.end(), 1, std::multiplies<index_t>());
+    assert(x_elm == y_elm);
+    (void)y_elm;
+
+    auto f = [&](auto i_element) {
+        std::vector<size_t> y_coord = [&]() {
+            std::vector<size_t> tmp(rank, 0);
+            size_t r = i_element;
+            for(index_t i = rank - 1; i >= 0; i--)
+            {
+                tmp[i] = r % y_len[i];
+                r      = r / y_len[i];
+            }
+            return tmp;
+        }();
+
+        std::vector<size_t> x_coord = [&]() {
+            std::vector<size_t> tmp(rank, 0);
+            for(index_t i = 0; i < rank; i++)
+            {
+                tmp[perm[i]] = y_coord[i];
+            }
+            return tmp;
+        }();
+
+        // do permute
+        y(y_coord) = x(x_coord);
+    };
+
+    make_ParallelTensorFunctor(f, x_elm)(std::thread::hardware_concurrency());
+}
+
+template <typename DataType>
+CK_TILE_HOST auto reference_permute(const HostTensor<DataType>& x, std::vector<index_t> perm)
+{
+    auto x_shape                          = x.get_lengths();
+    ck_tile::index_t rank                 = perm.size();
+    std::vector<ck_tile::index_t> y_shape = [&]() {
+        std::vector<ck_tile::index_t> tmp(rank, 0);
+        for(int i = 0; i < static_cast<int>(rank); i++)
+        {
+            tmp[i] = x_shape[perm[i]];
+        }
+        return tmp;
+    }();
+
+    HostTensor<DataType> y(y_shape);
+    reference_permute(x, y, perm);
+    return y;
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_reduce.hpp

@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename XDataType, typename ComputeDataType, typename YDataType, typename ReduceOp>
+CK_TILE_HOST void
+reference_reduce(const HostTensor<XDataType>& x_m_n, HostTensor<YDataType>& y_m, ReduceOp reduce_op)
+{
+    auto f = [&](auto m) {
+        const int N = x_m_n.mDesc.get_lengths()[1];
+
+        ComputeDataType v_acc = reduce_op.template GetIdentityValue<ComputeDataType>();
+
+        for(int n = 0; n < N; ++n)
+        {
+            const ComputeDataType v_a = type_convert<ComputeDataType>(x_m_n(m, n));
+
+            v_acc = reduce_op(v_acc, v_a);
+        }
+
+        y_m(m) = ck_tile::type_convert<YDataType>(v_acc);
+    };
+
+    make_ParallelTensorFunctor(f, y_m.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_rmsnorm2d_fwd.hpp

@@ -0,0 +1,87 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+
+namespace ck_tile {
+
+// Note: for simplicity, each functor only care about single M
+struct reference_rmsnorm2d_default_epilogue
+{
+    template <typename OutDataType, typename AccDataType>
+    void operator()(int m, HostTensor<OutDataType>& o, const HostTensor<AccDataType>& acc)
+    {
+        const int N = acc.mDesc.get_lengths()[1];
+        for(int n = 0; n < N; ++n)
+        {
+            o(m, n) = ck_tile::type_convert<OutDataType>(acc(m, n));
+        }
+    }
+
+    template <typename OutDataType, typename AccDataType>
+    auto operator()(int m, const HostTensor<AccDataType>& acc)
+    {
+        HostTensor<OutDataType> o(acc.get_lengths(), acc.get_strides());
+        operator()(m, o, acc);
+        return o;
+    }
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename ComputeDataType,
+          typename YDataType,
+          typename InvRmsDataType,
+          typename UnquantYDataType,
+          typename Epilogue = reference_rmsnorm2d_default_epilogue>
+void reference_rmsnorm2d_fwd(const HostTensor<XDataType>& x_m_n,
+                             const HostTensor<GammaDataType>& gamma_n,
+                             HostTensor<YDataType>& y_m_n,
+                             HostTensor<InvRmsDataType>& invRms_m,
+                             HostTensor<UnquantYDataType>& unquant_y_m_n,
+                             ComputeDataType epsilon,
+                             Epilogue epilogue_functor = {})
+{
+    auto rmsnorm2d_fwd_func = [&](auto m) {
+        const int N = x_m_n.mDesc.get_lengths()[1];
+
+        ComputeDataType mean_square = 0;
+        ComputeDataType divisor     = 0;
+
+        for(int n = 0; n < N; ++n)
+        {
+            ComputeDataType x = ck_tile::type_convert<ComputeDataType>(x_m_n(m, n));
+            mean_square += x * x;
+        }
+
+        mean_square = mean_square / N;
+        divisor = ck_tile::type_convert<ComputeDataType>(1) / ck_tile::sqrt(mean_square + epsilon);
+
+        if constexpr(!std::is_same_v<InvRmsDataType, ck_tile::null_type>)
+            invRms_m(m) = ck_tile::type_convert<InvRmsDataType>(divisor);
+
+        HostTensor<ComputeDataType> acc(x_m_n.get_lengths(), x_m_n.get_strides());
+        for(int n = 0; n < N; ++n)
+        {
+            ComputeDataType x     = ck_tile::type_convert<ComputeDataType>(x_m_n(m, n));
+            ComputeDataType gamma = ck_tile::type_convert<ComputeDataType>(gamma_n(n));
+            acc(m, n)             = x * divisor * gamma;
+        }
+
+        if constexpr(!std::is_same_v<UnquantYDataType, ck_tile::null_type>)
+        {
+            epilogue_functor(m, unquant_y_m_n, y_m_n, acc);
+        }
+        else
+        {
+            epilogue_functor(m, y_m_n, acc);
+        }
+    };
+
+    make_ParallelTensorFunctor(rmsnorm2d_fwd_func, invRms_m.mDesc.get_lengths()[0])(
+        std::thread::hardware_concurrency());
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_rowwise_quantization2d.hpp

@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+template <typename XDataType, typename ScaleDataType, typename QXDataType>
+CK_TILE_HOST void reference_rowwise_quantization2d(const HostTensor<XDataType>& x_m_n,
+                                                   const HostTensor<ScaleDataType>& scale_m,
+                                                   HostTensor<QXDataType>& qx_m_n)
+{
+    auto f = [&](auto m) {
+        const int N = x_m_n.mDesc.get_lengths()[1];
+
+        for(int n = 0; n < N; ++n)
+        {
+            auto v_x = x_m_n(m, n);
+            // scale = amax / 127 for int8
+            auto v_scale = type_convert<XDataType>(scale_m(m));
+            auto v_qx    = v_x / v_scale;
+            qx_m_n(m, n) = type_convert<QXDataType>(saturates<QXDataType>{}(v_qx));
+        }
+    };
+
+    make_ParallelTensorFunctor(f,
+                               scale_m.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
+}
+
+} // namespace ck_tile

include/ck_tile/host/reference/reference_softmax.hpp

@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+
+namespace ck_tile {
+
+template <typename InputType, typename ComputeType, typename OutputType = ComputeType>
+CK_TILE_HOST void
+reference_softmax(const HostTensor<InputType>& x, HostTensor<OutputType>& y, index_t dim = -1)
+{
+    index_t rank = x.get_num_of_dimension();
+    assert(rank == y.get_num_of_dimension());
+    assert(dim == -1 || dim < rank);
+
+    index_t target_dim  = dim == -1 ? (rank - 1) : dim;
+    index_t softmax_len = x.get_length(target_dim);
+    index_t n_parallel  = x.get_element_size() / softmax_len;
+    auto x_len          = x.get_lengths();
+
+    auto f = [&](auto i_element) {
+        std::vector<size_t> coord = [&]() {
+            std::vector<size_t> t_(rank, 0);
+            size_t r = i_element;
+            for(index_t i = rank - 1; i >= 0; i--)
+            {
+                if(i == target_dim)
+                    continue;
+                t_[i] = r % x_len[i];
+                r     = r / x_len[i];
+            }
+            return t_;
+        }();
+
+        ComputeType v_max = -ck_tile::numeric<ComputeType>::infinity();
+
+        // compute max
+        for(auto idx = 0; idx < softmax_len; idx++)
+        {
+            auto c_               = coord;
+            c_[target_dim]        = idx;
+            const ComputeType v_x = ck_tile::type_convert<ComputeType>(x(c_));
+            v_max                 = v_max < v_x ? v_x : v_max;
+        }
+
+        ComputeType v_exp_sum = static_cast<ComputeType>(0);
+
+        // sum
+        for(auto idx = 0; idx < softmax_len; idx++)
+        {
+            auto c_        = coord;
+            c_[target_dim] = idx;
+
+            const ComputeType v_x = ck_tile::type_convert<ComputeType>(x(c_));
+
+            v_exp_sum += ck_tile::exp(v_x - v_max);
+        }
+
+        // elementwise
+        for(auto idx = 0; idx < softmax_len; idx++)
+        {
+            auto c_        = coord;
+            c_[target_dim] = idx;
+
+            const ComputeType v_x = ck_tile::type_convert<ComputeType>(x(c_));
+
+            auto out = ck_tile::exp(v_x - v_max) / v_exp_sum;
+
+            y(c_) = ck_tile::type_convert<OutputType>(out);
+        }
+    };
+
+    make_ParallelTensorFunctor(f, n_parallel)(std::thread::hardware_concurrency());
+}
+
+template <typename InputType, typename ComputeType, typename OutputType = ComputeType>
+CK_TILE_HOST auto reference_softmax(const HostTensor<InputType>& x, index_t dim = -1)
+{
+    HostTensor<OutputType> y(x.get_lengths(), x.get_strides());
+
+    reference_softmax<InputType, ComputeType, OutputType>(x, y, dim);
+
+    return y;
+}
+} // namespace ck_tile

include/ck_tile/host/reference/reference_topk.hpp

@@ -0,0 +1,124 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+#include <thread>
+#include <numeric>
+#include <functional>
+#include <utility>
+#include <algorithm>
+
+namespace ck_tile {
+
+/*
+    similiar to torch.topk()
+    x (Tensor) – the input tensor.
+    k (int) – the k in “top-k”
+    dim (int, optional) – the dimension to sort along
+    largest (bool, optional) – largest or smallest elements
+    sorted (bool, optional) – elements in sorted order or not
+
+    output:
+    y_values
+    y_indices
+
+    https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/TopKImpl.h
+*/
+template <typename DataType, typename IndexType = index_t>
+CK_TILE_HOST void reference_topk(const HostTensor<DataType>& x,
+                                 HostTensor<DataType>& y_values,
+                                 HostTensor<IndexType>& y_indices,
+                                 index_t k,
+                                 index_t dim  = -1,
+                                 bool largest = true,
+                                 bool sorted  = true)
+{
+    // rank must be the same
+    index_t rank = x.get_num_of_dimension();
+    assert(rank == y_values.get_num_of_dimension());
+    assert(rank == y_indices.get_num_of_dimension());
+    assert(dim == -1 || dim < rank);
+
+    index_t topk_dim     = dim == -1 ? (rank - 1) : dim;
+    index_t topk_src_len = x.get_length(topk_dim);
+    auto x_len           = x.get_lengths();
+
+    assert(k <= topk_src_len);
+    assert(k == y_values.get_length(topk_dim) && k == y_indices.get_length(topk_dim));
+
+    index_t n_parallel = x.get_element_size() / topk_src_len;
+
+    // clang-format off
+    auto f = [&](auto i_element) {
+        std::vector<size_t> topk_coord = [&](){
+            std::vector<size_t> t_(rank, 0);
+            size_t r = i_element;
+            for(index_t i = rank - 1; i >= 0; i--) {
+                if(i == topk_dim)          continue; // topk dim should be zero
+                t_[i] = r % x_len[i];      r = r / x_len[i];
+            }
+            return t_;
+        }();
+
+        using elem_t = std::pair<DataType, IndexType>;
+        std::vector<elem_t> q = [&](){
+            std::vector<elem_t> t_(topk_src_len);
+            for(index_t i = 0; i < topk_src_len; i++) {
+                auto c_ = topk_coord;  c_[topk_dim] = i;
+                t_[i].first = x(c_);   t_[i].second = i;
+            }
+            return t_;
+        }();
+
+        // run topk
+        if(largest) {
+            std::nth_element(q.begin(), q.begin() + k - 1, q.end(),
+            [](const elem_t& lhs, const elem_t& rhs) -> bool { return lhs.first > rhs.first; });
+            if(sorted) {
+                std::sort(q.begin(), q.begin() + k - 1,
+                [](const elem_t& lhs, const elem_t& rhs) -> bool { return lhs.first > rhs.first; });
+            }
+        } else {
+            std::nth_element(q.begin(), q.begin() + k - 1, q.end(),
+            [](const elem_t& lhs, const elem_t& rhs) -> bool { return lhs.first < rhs.first; });
+            if(sorted) {
+                std::sort(q.begin(), q.begin() + k - 1,
+                [](const elem_t& lhs, const elem_t& rhs) -> bool { return lhs.first < rhs.first; });
+            }
+        }
+
+        // write out
+        for(index_t i = 0; i < k; i++) {
+            auto c_ = topk_coord;  c_[topk_dim] = i;
+            y_values(c_) = q[i].first;  y_indices(c_) = q[i].second;
+        }
+    };
+    // clang-format on
+
+    make_ParallelTensorFunctor(f, n_parallel)(std::thread::hardware_concurrency());
+}
+
+// TODO: if using this method, the return tensor would be dense(no stride)
+template <typename DataType, typename IndexType = index_t>
+CK_TILE_HOST auto reference_topk(const HostTensor<DataType>& x,
+                                 index_t k,
+                                 index_t dim  = -1,
+                                 bool largest = true,
+                                 bool sorted  = true)
+{
+    auto lens          = x.get_lengths();
+    index_t target_dim = (dim == -1) ? (lens.size() - 1) : dim;
+    assert(target_dim < lens.size());
+    assert(k <= lens[target_dim]);
+    lens[target_dim] = k;
+    HostTensor<DataType> y_values(lens);
+    HostTensor<IndexType> y_indices(lens);
+
+    reference_topk<DataType, IndexType>(x, y_values, y_indices, k, dim, largest, sorted);
+
+    return ck_tile::make_tuple(y_values, y_indices);
+}
+} // namespace ck_tile