[CK_TILE] fused-moe first version (#1634)

* moe pipeline

* update code

* compile OK

* update

* update cpu reference

* update pipeline_gemm0

* compiler ok

* update pipeline

* rename to ex pipeline

* block-asm

* update

* update

* update first gemm ok

* compute correct

* update file structure

* update README

* update

* update

* update code

* update API

* return unsupport case

* add comment

* update readme

* update

* uncomment

* update

* fix build err

---------
Co-authored-by: valarLip <340077269@qq.com>

include/ck_tile/host/reference/reference_fused_moe.hpp

+// 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
+    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;
+    ck_tile::index_t intermediate_size_1 = intermediate_size / (gate_only ? 1 : 2);
+
+    // 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::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];
+        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
+        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_permute.hpp

@@ -16,7 +16,7 @@ namespace ck_tile {
 */
 template <typename DataType>
 CK_TILE_HOST void
-reference_permute(const HostTensor<DataType>& x, HostTensor<DataType>& y, std::vector<index_t> dims)
+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();
@@ -43,7 +43,7 @@ reference_permute(const HostTensor<DataType>& x, HostTensor<DataType>& y, std::v
             std::vector<size_t> tmp(rank, 0);
             for(index_t i = 0; i < rank; i++)
             {
-                tmp[dims[i]] = y_coord[i];
+                tmp[perm[i]] = y_coord[i];
             }
             return tmp;
         }();
@@ -54,4 +54,23 @@ reference_permute(const HostTensor<DataType>& x, HostTensor<DataType>& y, std::v
 
     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/ops/elementwise/unary_element_wise_operation.hpp

@@ -572,6 +572,105 @@ struct FastGelu
     }
 };
 
+struct FastGeluAsm
+{
+    template <typename Y, typename X>
+    CK_TILE_HOST void operator()(Y& y, const X& x) const;
+
+    template <typename Y, typename X>
+    CK_TILE_DEVICE void operator()(Y& y, const X& x) const;
+
+    template <>
+    CK_TILE_HOST void operator()<float, float>(float& y, const float& x) const
+    {
+        // const float u   = -2.f * x * (0.035677f * x * x + 0.797885f);
+        const float c1  = -2.0 * 0.035677f;
+        const float c2  = -2.0 * 0.797885f;
+        const float u   = x * (c1 * x * x + c2);
+        const float emu = exp(u);
+        y               = x / (1.f + emu);
+    }
+
+    // device code, use lower precision "__ocml_exp_f32" and "rcp"
+    template <>
+    CK_TILE_DEVICE void operator()<float, float>(float& y, const float& x) const
+    {
+        const uint32_t c1     = 0xbd92220c; // -2.0 * 0.035677f;
+        const float c2        = -2.0 * 0.797885f;
+        const uint32_t log2e_ = 0x3fb8aa3b; // log2e_v<float>;
+        float tmp;
+
+        asm volatile("v_mul_f32 %[v_tmp], %[v_x], %[v_x]        ; x*x\n"
+                     "v_fma_f32 %[v_tmp], %[v_tmp], %[s_c1], %[v_c2]  ; c1*x*x+c2\n"
+                     "v_mul_f32 %[v_tmp], %[v_tmp], %[v_x]      ; x*(c1*x*x+c2)\n"
+                     "v_mul_f32 %[v_tmp], %[v_tmp], %[s_log2e]  ; log2e*x*(c1*x*x+c2)\n"
+                     "v_exp_f32 %[v_tmp], %[v_tmp]              ; emu = exp2(log2e*x*(c1*x*x+c2))\n"
+                     "s_nop 0                                   ; hazard for exp\n"
+                     "v_add_f32 %[v_tmp], %[v_tmp], 1.0         ; emu+1.0f\n"
+                     "v_rcp_f32 %[v_tmp], %[v_tmp]              ; 1/(emu+1.0f)\n"
+                     "s_nop 0                                   ; hazard for rcp \n"
+                     "v_mul_f32 %[v_y], %[v_tmp], %[v_x]        ; x * 1/(emu+1f)\n"
+                     : [v_y] "=v"(y), [v_tmp] "+v"(tmp)
+                     : [v_x] "v"(x), [s_c1] "s"(c1), [v_c2] "v"(c2), [s_log2e] "s"(log2e_)
+                     :);
+    }
+
+    template <>
+    CK_TILE_HOST void operator()<fp32x2_t, fp32x2_t>(fp32x2_t& y, const fp32x2_t& x) const
+    {
+        const float c1   = -2.0 * 0.035677f;
+        const float c2   = -2.0 * 0.797885f;
+        const float u0   = x.x * (c1 * x.x * x.x + c2);
+        const float emu0 = exp(u0);
+        y.x              = x.x / (1.f + emu0);
+        const float u1   = x.y * (c1 * x.y * x.y + c2);
+        const float emu1 = exp(u1);
+        y.y              = x.y / (1.f + emu1);
+    }
+
+    // this is packed verion to remove data hazard for trans
+    template <>
+    CK_TILE_DEVICE void operator()<fp32x2_t, fp32x2_t>(fp32x2_t& y, const fp32x2_t& x) const
+    {
+        const uint32_t c1     = 0xbd92220c; // -2.0 * 0.035677f;
+        float c2              = -2.0 * 0.797885f;
+        const uint32_t log2e_ = 0x3fb8aa3b; // log2e_v<float>;
+        float tmp0, tmp1;
+        float y0 = x.x, y1 = x.y;
+
+        asm volatile(
+            "v_mul_f32 %[v_tmp0], %[v_y0], %[v_y0]        ; x*x\n"
+            "v_mul_f32 %[v_tmp1], %[v_y1], %[v_y1]        ; x*x\n"
+            "v_fma_f32 %[v_tmp0], %[v_tmp0], %[s_c1], %[v_c2]  ; c1*x*x+c2\n"
+            "v_fma_f32 %[v_tmp1], %[v_tmp1], %[s_c1], %[v_c2]  ; c1*x*x+c2\n"
+            "v_mul_f32 %[v_tmp0], %[v_tmp0], %[v_y0]      ; x*(c1*x*x+c2)\n"
+            "v_mul_f32 %[v_tmp1], %[v_tmp1], %[v_y1]      ; x*(c1*x*x+c2)\n"
+            "v_mul_f32 %[v_tmp0], %[v_tmp0], %[s_log2e]  ; log2e*x*(c1*x*x+c2)\n"
+            "v_mul_f32 %[v_tmp1], %[v_tmp1], %[s_log2e]  ; log2e*x*(c1*x*x+c2)\n"
+            "v_exp_f32 %[v_tmp0], %[v_tmp0]              ; emu = exp2(log2e*x*(c1*x*x+c2))\n"
+            "v_exp_f32 %[v_tmp1], %[v_tmp1]              ; emu = exp2(log2e*x*(c1*x*x+c2))\n"
+            "v_add_f32 %[v_tmp0], %[v_tmp0], 1.0         ; emu+1.0f\n"
+            "v_add_f32 %[v_tmp1], %[v_tmp1], 1.0         ; emu+1.0f\n"
+            "v_rcp_f32 %[v_tmp0], %[v_tmp0]              ; 1/(emu+1.0f)\n"
+            "v_rcp_f32 %[v_tmp1], %[v_tmp1]              ; 1/(emu+1.0f)\n"
+            "v_mul_f32 %[v_y0], %[v_tmp0], %[v_y0]        ; x * 1/(emu+1f)\n"
+            "v_mul_f32 %[v_y1], %[v_tmp1], %[v_y1]        ; x * 1/(emu+1f)\n"
+            : [v_y0] "+v"(y0),
+              [v_y1] "+v"(y1),
+              [v_c2] "+v"(c2),
+              // NOTE! it is totally possible that c2/y0/y1 share same register, they are all local
+              // tmp variables we need to expicitly hint compiler they may read+write, to allow
+              // allocate different register , the side effect is c2=** may issue for every such
+              // inline asm block
+              [v_tmp0] "+v"(tmp0),
+              [v_tmp1] "+v"(tmp1)
+            : [s_c1] "s"(c1), [s_log2e] "s"(log2e_)
+            :);
+        y.x = y0;
+        y.y = y1;
+    }
+};
+
 // https://paperswithcode.com/method/gelu
 // y = 0.5*x*(1+erf(x/sqrt(2)))
 struct Gelu

include/ck_tile/ops/moe_sorting.hpp → include/ck_tile/ops/flatmm.hpp

@@ -3,9 +3,8 @@
 
 #pragma once
 
-#include "ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp"
-#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp"
-#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
-#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp"
+#include "ck_tile/ops/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp"
+#include "ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp"
+#include "ck_tile/ops/flatmm/block/flatmm_uk_config.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/flatmm/block/flatmm_uk_config.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#define CK_TILE_FLATMM_UK_MFMA_FP16 0
+#define CK_TILE_FLATMM_UK_MFMA_BF16 1
+#define CK_TILE_FLATMM_UK_MFMA_INT8 2
+#define CK_TILE_FLATMM_UK_MFMA_FP8 3
+#define CK_TILE_FLATMM_UK_MFMA_BF8 4

include/ck_tile/ops/flatmm/block/uk/README.md

+the files under this folder should not be included directly!
\ No newline at end of file

include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp

@@ -331,7 +331,8 @@ struct BlockFmhaPipelineQRKSVSAsync
                              Policy::template MakeVDramTileDistribution<Problem>());
 
         // prefetch K tile
-        async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})), k_dram_window, k_oob_ck, k_pre_np);
+        async_load_tile_raw(
+            k_lds_store(LdsSeq.at(number<0>{})), k_dram_window, number<-1>{}, k_oob_ck, k_pre_np);
         move_tile_window(k_dram_window, {0, kK0});
         __builtin_amdgcn_sched_barrier(0);
 
@@ -355,6 +356,7 @@ struct BlockFmhaPipelineQRKSVSAsync
                 static_for<0, k0_loops - 1, 1>{}([&](auto i_k0) {
                     async_load_tile_raw(k_lds_store(number<LdsSeq.at(number<i_k0 + 1>{})>{}),
                                         k_dram_window,
+                                        number<-1>{},
                                         k_oob_ck,
                                         k_pre_np);
                     if constexpr(i_k0 < k0_loops - 1)
@@ -386,7 +388,7 @@ struct BlockFmhaPipelineQRKSVSAsync
             __builtin_amdgcn_s_barrier();
 
             const auto bias_tile = load_tile(bias_dram_window); // load bias tile
-            auto v_buf           = load_tile(v_dram_window, bool_constant<false>{});
+            auto v_buf           = load_tile(v_dram_window, number<-1>{}, bool_constant<false>{});
             __builtin_amdgcn_sched_barrier(0);
             { // tail
                 gemm_0(s_acc,
@@ -514,7 +516,8 @@ struct BlockFmhaPipelineQRKSVSAsync
                 move_tile_window(
                     v_dram_window,
                     {0, kK1}); // will have scratch if move this right after load_tile(v_dram)...
-                v_buf = load_tile(v_dram_window, bool_constant<false>{}); // load next v_buf
+                v_buf = load_tile(
+                    v_dram_window, number<-1>{}, bool_constant<false>{}); // load next v_buf
             }
             __builtin_amdgcn_sched_barrier(0);
 
@@ -618,7 +621,8 @@ struct BlockFmhaPipelineQRKSVSAsync
                 static_for<0, k1_loops - 1, 1>{}([&](auto i_k1) {
                     if constexpr(i_k1 != 0 && i_k1 < k1_loops - 1)
                     {
-                        v_buf = load_tile(v_dram_window, bool_constant<false>{}); // load next v_buf
+                        v_buf = load_tile(
+                            v_dram_window, number<-1>{}, bool_constant<false>{}); // load next v_buf
                     }
                     block_sync_lds();
                     gemm_1(o_acc,
@@ -665,8 +669,11 @@ struct BlockFmhaPipelineQRKSVSAsync
                 if constexpr(k1_loops >= 2 &&
                              LdsSeq.at(number<0>{}) == LdsSeq.at(number<k0_loops + k1_loops - 2>{}))
                     __builtin_amdgcn_s_barrier();
-                async_load_tile_raw(
-                    k_lds_store(LdsSeq.at(number<0>{})), k_dram_window, k_oob_ck, k_pre_np);
+                async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})),
+                                    k_dram_window,
+                                    number<-1>{},
+                                    k_oob_ck,
+                                    k_pre_np);
                 move_tile_window(k_dram_window, {0, kK0});
             }
             // tail

include/ck_tile/ops/fused_moe.hpp

@@ -3,7 +3,15 @@
 
 #pragma once
 
+#include "ck_tile/ops/fused_moe/kernel/fused_moegemm_kernel.hpp"
+#include "ck_tile/ops/fused_moe/kernel/fused_moegemm_shape.hpp"
+#include "ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp"
 #include "ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp"
+#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp"
+#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp"
+#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp"
+#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_problem.hpp"
+#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp"
 #include "ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp"
 #include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
 #include "ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp"

include/ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck_tile {
+
+template <typename BlockShape_>
+struct FusedMoeGemmTilePartitioner_Linear
+{
+    //  FusedMoeGemmShape
+    using BlockShape = ck_tile::remove_cvref_t<BlockShape_>;
+
+    static constexpr const char* name = "lin";
+
+    CK_TILE_DEVICE auto operator()(ck_tile::index_t /*num_sorted_tiles*/,
+                                   ck_tile::index_t /*intermediate_size*/)
+    {
+        index_t i_n = blockIdx.x;
+        index_t i_m = blockIdx.y;
+
+        return ck_tile::make_tuple(i_m, i_n);
+    }
+
+    CK_TILE_HOST static constexpr auto GridSize(index_t max_tokens, index_t intermediate_size)
+    {
+        // TODO: this may need tuning
+        index_t ms = ck_tile::integer_divide_ceil(max_tokens, BlockShape::Block_M0);
+        index_t ns = ck_tile::integer_divide_ceil(intermediate_size, BlockShape::Block_N0);
+        return dim3(ns, ms, 1);
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_problem.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+// TODO: alow 2 gemm have different type
+template <typename ADataType_,
+          typename GDataType_,
+          typename DDataType_,
+          typename AccDataType_,
+          typename ODataType_,
+          typename AScaleDataType_,
+          typename GScaleDataType_,
+          typename DScaleDataType_,
+          typename YSmoothScaleDataType_,
+          typename TopkWeightDataType_,
+          typename IndexDataType_,  // data type for all indexing
+          typename GateActivation_, // = ck_tile::element_wise::Silu,
+          typename BlockShape_,     // shoule be FusedMoeGemmShape
+          typename Traits_>
+struct FusedMoeGemmPipelineProblem
+{
+    using ADataType            = remove_cvref_t<ADataType_>;
+    using GDataType            = remove_cvref_t<GDataType_>;
+    using DDataType            = remove_cvref_t<DDataType_>;
+    using AccDataType          = remove_cvref_t<AccDataType_>;
+    using ODataType            = remove_cvref_t<ODataType_>;
+    using AScaleDataType       = remove_cvref_t<AScaleDataType_>;
+    using GScaleDataType       = remove_cvref_t<GScaleDataType_>;
+    using DScaleDataType       = remove_cvref_t<DScaleDataType_>;
+    using YSmoothScaleDataType = remove_cvref_t<YSmoothScaleDataType_>;
+    using TopkWeightDataType   = remove_cvref_t<TopkWeightDataType_>;
+    using IndexDataType        = remove_cvref_t<IndexDataType_>;
+
+    // the input for next gemm should have same time as
+    using YDataType = ADataType;
+
+    using GateActivation = remove_cvref_t<GateActivation_>;
+    using BlockShape     = remove_cvref_t<BlockShape_>;
+    using Traits         = remove_cvref_t<Traits_>;
+};
+} // namespace ck_tile