Merge branch 'develop' into codegen_hiprtc

include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp

@@ -121,10 +121,10 @@ struct GridwiseTensorRearrange
             __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
 
         // Global Memory
-        const index_t a_batch_offset =
-            __builtin_amdgcn_readfirstlane(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx));
-        const index_t c_batch_offset =
-            __builtin_amdgcn_readfirstlane(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx));
+        const index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+        const index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
 
         const auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_in_global + a_batch_offset, in_grid_desc.GetElementSpaceSize());

include/ck/utility/amd_wmma.hpp

@@ -9,7 +9,8 @@
 // TODO: Add arch limitation
 namespace ck {
 
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__)
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || \
+    defined(__gfx1103__) || defined(__gfx11_generic__)
 #define __gfx11__
 #endif
 /********************************WAVE32 MODE***********************************************/
@@ -260,7 +261,7 @@ struct intrin_wmma_i32_16x16x16_iu8_w64<16, 16, neg_a, neg_b, clamp>
 // gfx12
 /********************************WAVE32 MODE***********************************************/
 
-#if defined(__gfx1200__) || defined(__gfx1201__)
+#if defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx12_generic__)
 #define __gfx12__
 #endif
 

include/ck_tile/core/config.hpp

@@ -11,13 +11,15 @@
 #define __gfx94__
 #endif
 #if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
-    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__)
+    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || \
+    defined(__gfx10_3_generic__)
 #define __gfx103__
 #endif
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__)
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || \
+    defined(__gfx1103__) || defined(__gfx11_generic__)
 #define __gfx11__
 #endif
-#if defined(__gfx1200__) || defined(__gfx1201__)
+#if defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx12_generic__)
 #define __gfx12__
 #endif
 

include/ck_tile/host.hpp

@@ -23,6 +23,7 @@
 #include "ck_tile/host/reference/reference_gemm.hpp"
 #include "ck_tile/host/reference/reference_im2col.hpp"
 #include "ck_tile/host/reference/reference_layernorm2d_fwd.hpp"
+#include "ck_tile/host/reference/reference_moe_sorting.hpp"
 #include "ck_tile/host/reference/reference_permute.hpp"
 #include "ck_tile/host/reference/reference_reduce.hpp"
 #include "ck_tile/host/reference/reference_rmsnorm2d_fwd.hpp"

include/ck_tile/host/reference/reference_moe_sorting.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 {
+
+template <typename WeightType, typename IndexType = index_t>
+CK_TILE_HOST void reference_moe_sorting(const HostTensor<IndexType>& topk_ids,
+                                        const HostTensor<WeightType>& weights,
+                                        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)
+{
+    const index_t num_token = topk_ids.mDesc.get_lengths()[0];
+    const index_t topk      = topk_ids.mDesc.get_lengths()[1];
+    std::vector<std::vector<IndexType>> expert_tokens(experts,
+                                                      std::vector<IndexType>(unit_size, num_token));
+    std::vector<std::vector<WeightType>> expert_token_weights(
+        experts, std::vector<WeightType>(unit_size, 0));
+    std::vector<IndexType> expert_slices(experts, 1);
+    std::vector<IndexType> expert_slice_idxs(experts, 0);
+
+    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++)
+                {
+                    expert_tokens[e][i]        = num_token;
+                    expert_token_weights[e][i] = 0;
+                }
+            }
+
+            expert_tokens[e][idx]        = t;
+            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();
+    for(index_t e = 0; e < experts; e++)
+    {
+        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] = e;
+            unit_cnt++;
+        }
+        out_expert_id += expert_slices[e];
+    }
+    unit_cnt *= unit_size;
+    return;
+}
+} // namespace ck_tile

include/ck_tile/ops/common/generic_2d_block_shape.hpp

@@ -38,9 +38,7 @@ namespace ck_tile {
 template <typename BlockTile_,    // block size, seq<M, N>
           typename WarpPerBlock_, // num warps along seq<M, N>
           typename WarpTile_,     // warp size, seq<M, N>
-          typename Vector_,       // contiguous pixels(vector size) along seq<M, N>
-          index_t BlockSize_ =
-              warpSize* reduce_on_sequence(WarpPerBlock_{}, multiplies{}, number<1>{})>
+          typename Vector_>       // contiguous pixels(vector size) along seq<M, N>)>
 struct Generic2dBlockShape
 {
     // block size
@@ -68,10 +66,12 @@ struct Generic2dBlockShape
     static_assert(Warp_M % Vector_M == 0);
     static_assert(Warp_N % Vector_N == 0);
     // num of threads along seq<M, N>, within each warp
-    static constexpr index_t ThreadPerWarp_M = Warp_M / Vector_M;
-    static constexpr index_t ThreadPerWarp_N = Warp_N / Vector_N;
+    static constexpr index_t ThreadPerWarp_M  = Warp_M / Vector_M;
+    static constexpr index_t ThreadPerWarp_N  = Warp_N / Vector_N;
+    static constexpr index_t ThreadPerBlock_M = Block_M / Repeat_M / Vector_M;
+    static constexpr index_t ThreadPerBlock_N = Block_N / Repeat_N / Vector_N;
 
-    static constexpr index_t BlockSize = BlockSize_;
+    static constexpr index_t BlockSize = ThreadPerBlock_M * ThreadPerBlock_N;
 };
 
 } // namespace ck_tile

include/ck_tile/ops/fmha/block/page_block_navigator.hpp

@@ -230,7 +230,15 @@ struct PageBlockNavigator
     CK_TILE_HOST_DEVICE
     DataType* get_block_ptr(index_t block_index) const
     {
-        return physical_blocks + physical_block_indices[block_index] * block_stride + fixed_offset;
+        if(block_index < num_blocks)
+        {
+            return physical_blocks + physical_block_indices[block_index] * block_stride +
+                   fixed_offset;
+        }
+        else
+        {
+            return nullptr;
+        }
     }
 
     CK_TILE_HOST_DEVICE int32_t get_block_index(const WindowOrigin& global_window_origin) const

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

+// 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/ops/common.hpp"
+#include "ck_tile/ops/elementwise.hpp"
+#include "ck_tile/host/hip_check_error.hpp"
+#include <string>
+#include <type_traits>
+
+namespace ck_tile {
+
+struct MoeSortingHostArgs
+{
+    const void* p_topk_ids;
+    const void* p_weights;
+    void* p_sorted_token_ids;
+    void* p_sorted_weights;
+    void* p_sorted_expert_ids;
+    void* p_total_tokens_post_pad;
+    void* p_moe_buf;
+    index_t tokens;
+    index_t unit_size;
+    index_t num_experts;
+    index_t topk;
+    index_t moe_buf_bytes;
+};
+
+template <typename Problem_>
+struct MoeSortingKernel
+{
+    using Problem = remove_cvref_t<Problem_>;
+
+    using IndexType  = typename Problem::IndexType;
+    using WeightType = typename Problem::WeightType;
+
+    typedef MoeSortingHostArgs MoeSortingKargs;
+
+    using Hargs = MoeSortingHostArgs;
+
+    struct Kargs
+    {
+        const void* p_topk_ids;
+        const void* p_weights;
+        void* p_sorted_token_ids;
+        void* p_sorted_weights;
+        void* p_sorted_expert_ids;
+        void* p_total_tokens_post_pad;
+        void* p_moe_buf;
+        index_t tokens;
+        index_t num_experts;
+        index_t moe_buf_bytes;
+
+        index_t tokens_per_thread;
+        mdiv unit_size_mdiv;
+        mdiv topk_mdiv;
+    };
+
+    CK_TILE_HOST static constexpr auto GridSize(const Hargs& h)
+    {
+        // TODO: assume num-experts not too much
+        return dim3(1 + ck_tile::integer_divide_ceil(h.moe_buf_bytes, BlockSize(h).x * 16));
+    }
+
+    CK_TILE_HOST static constexpr auto BlockSize(const Hargs& h)
+    {
+        return dim3(ck_tile::integer_least_multiple(h.num_experts, ck_tile::get_warp_size()));
+    }
+
+    // in byte
+    CK_TILE_HOST static constexpr auto GetSmemSize(const Hargs& h)
+    {
+        const auto blocks = BlockSize(h);
+        return ((blocks.x + 1) * h.num_experts + (h.num_experts + 1)) * sizeof(index_t);
+    }
+
+    CK_TILE_HOST static constexpr auto MakeKargs(const Hargs& h)
+    {
+        Kargs k;
+        k.p_topk_ids              = h.p_topk_ids;
+        k.p_weights               = h.p_weights;
+        k.p_sorted_token_ids      = h.p_sorted_token_ids;
+        k.p_sorted_weights        = h.p_sorted_weights;
+        k.p_sorted_expert_ids     = h.p_sorted_expert_ids;
+        k.p_moe_buf               = h.p_moe_buf;
+        k.p_total_tokens_post_pad = h.p_total_tokens_post_pad;
+        k.tokens                  = h.tokens;
+        k.num_experts             = h.num_experts;
+        k.moe_buf_bytes           = h.moe_buf_bytes;
+
+        const auto blocks   = BlockSize(h);
+        k.tokens_per_thread = integer_divide_ceil(h.tokens * h.topk, blocks.x);
+        k.unit_size_mdiv    = mdiv{static_cast<uint32_t>(h.unit_size)};
+        k.topk_mdiv         = mdiv{static_cast<uint32_t>(h.topk)};
+        return k;
+    }
+
+    CK_TILE_DEVICE index_t calc_index(index_t total_col, index_t row, index_t col) const
+    {
+        return row * total_col + col;
+    }
+
+    CK_TILE_DEVICE void moe_buf_set_zero_kernel(uint8x16_t* buf, index_t buf_bytes) const
+    {
+        const index_t offset = (blockIdx.x - 1) * blockDim.x + threadIdx.x;
+        if(offset < buf_bytes / 16)
+        {
+            buf[offset] = uint8x16_t{0};
+        }
+    }
+
+    CK_TILE_DEVICE void moe_align_block_size_kernel(const IndexType* __restrict__ topk_id,
+                                                    const WeightType* __restrict__ weights,
+                                                    index_t* p_sorted_token_ids,
+                                                    WeightType* p_sorted_weights,
+                                                    index_t* p_sorted_expert_ids,
+                                                    index_t* p_total_tokens_post_pad,
+                                                    const index_t num_experts,
+                                                    const index_t tokens_per_thread,
+                                                    const index_t numel,
+                                                    const mdiv unit_size_mdiv,
+                                                    const mdiv topk_mdiv,
+                                                    void* smem) const
+    {
+        const index_t tid       = static_cast<index_t>(threadIdx.x);
+        const index_t start_idx = tid * tokens_per_thread;
+
+        index_t* shared_mem = reinterpret_cast<index_t*>(smem);
+
+        index_t* tokens_cnts = shared_mem; // 2d: (blockDim.x + 1, num_experts)
+        index_t* cumsum      = shared_mem + (blockDim.x + 1) * num_experts; // 1: (num_experts + 1)
+        for(int i = 0; i < num_experts; ++i)
+        {
+            tokens_cnts[calc_index(num_experts, tid + 1, i)] = 0;
+        }
+#pragma unroll Problem_::InternalLoadUnroll
+        for(int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i)
+        {
+            ++tokens_cnts[calc_index(num_experts, tid + 1, topk_id[i])];
+        }
+        __syncthreads();
+
+        if(tid < num_experts)
+        {
+            tokens_cnts[calc_index(num_experts, 0, tid)] = 0;
+            for(int i = 1; i <= static_cast<index_t>(blockDim.x); ++i)
+            {
+                tokens_cnts[calc_index(num_experts, i, tid)] +=
+                    tokens_cnts[calc_index(num_experts, i - 1, tid)];
+            }
+        }
+
+        // __syncthreads();
+        if(tid == 0)
+        {
+            cumsum[0] = 0;
+            for(int i = 1; i <= num_experts; ++i)
+            {
+                auto current_units = [&]() {
+                    index_t x_ = tokens_cnts[calc_index(num_experts, blockDim.x, i - 1)] +
+                                 unit_size_mdiv.divisor - 1;
+                    index_t y_ = unit_size_mdiv.div(x_);
+                    return max(y_, 1) * unit_size_mdiv.divisor;
+                }();
+                cumsum[i] = cumsum[i - 1] + current_units;
+            }
+            *p_total_tokens_post_pad = cumsum[num_experts];
+        }
+        __syncthreads();
+        if(tid < num_experts)
+        {
+            for(int i = cumsum[tid]; i < cumsum[tid + 1]; i += unit_size_mdiv.divisor)
+            {
+                p_sorted_expert_ids[unit_size_mdiv.div(i)] = tid;
+            }
+        }
+
+#pragma unroll Problem_::InternalLoadUnroll
+        for(int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i)
+        {
+            index_t expert_id = topk_id[i];
+            index_t rank_post_pad =
+                tokens_cnts[calc_index(num_experts, tid, expert_id)] + cumsum[expert_id];
+            p_sorted_token_ids[rank_post_pad] = topk_mdiv.div(i);
+            p_sorted_weights[rank_post_pad]   = weights[i];
+            ++tokens_cnts[calc_index(num_experts, tid, expert_id)];
+        }
+
+        const index_t prefill_token = topk_mdiv.div(numel);
+        if(tid < num_experts)
+        {
+            index_t expert_offset =
+                cumsum[tid] + tokens_cnts[calc_index(num_experts, blockDim.x, tid)];
+            while(expert_offset < cumsum[tid + 1])
+            {
+                p_sorted_token_ids[expert_offset] = prefill_token;
+                p_sorted_weights[expert_offset]   = static_cast<WeightType>(0.0);
+                expert_offset++;
+            }
+        }
+    }
+
+    CK_TILE_DEVICE void operator()(Kargs kargs) const
+    {
+        if(blockIdx.x > 0)
+        {
+            if(kargs.p_moe_buf)
+            {
+                moe_buf_set_zero_kernel(reinterpret_cast<uint8x16_t*>(kargs.p_moe_buf),
+                                        kargs.moe_buf_bytes);
+            }
+            return;
+        }
+        const size_t numel = kargs.tokens * kargs.topk_mdiv.divisor;
+        extern __shared__ char smem[];
+        return moe_align_block_size_kernel(static_cast<const IndexType*>(kargs.p_topk_ids),
+                                           static_cast<const WeightType*>(kargs.p_weights),
+                                           static_cast<IndexType*>(kargs.p_sorted_token_ids),
+                                           static_cast<WeightType*>(kargs.p_sorted_weights),
+                                           static_cast<IndexType*>(kargs.p_sorted_expert_ids),
+                                           static_cast<IndexType*>(kargs.p_total_tokens_post_pad),
+                                           kargs.num_experts,
+                                           kargs.tokens_per_thread,
+                                           numel,
+                                           kargs.unit_size_mdiv,
+                                           kargs.topk_mdiv,
+                                           smem);
+    }
+};
+} // namespace ck_tile

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

+// 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/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
+#include <string>
+#include <type_traits>
+
+#ifndef TOPK_SOFTMAX_USE_RAW_TILE_WINDOW
+#define TOPK_SOFTMAX_USE_RAW_TILE_WINDOW 0
+#endif
+
+namespace ck_tile {
+
+// template <typename Problem_, typename Policy_ = MoeSortingPolicy>
+// struct MoeSortingPipeline
+// {
+//     // TODO: this kernel only support warp per row
+//     using Problem    = remove_cvref_t<Problem_>;
+//     using Policy     = remove_cvref_t<Policy_>;
+//     using WeightType = typename Problem::WeightType;
+
+//     template <typename TopkIdWindow, typename WeightWindow>
+//     CK_TILE_DEVICE auto operator()(const TopkIdWindow& topk_id_window,
+//                                    const WeightWindow& weight_window,
+//                                     index_t* p_sorted_token_ids,
+//                                     WeightType* p_sorted_weights,
+//                                     index_t* p_sorted_expert_ids,
+//                                     index_t* p_total_tokens_post_pad,
+//                                     const index_t num_experts,
+//                                     const index_t unit_size,
+//                                     const size_t numel,
+//                                     const index_t topk)
+//     {
+//     }
+// };
+} // namespace ck_tile

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

+// 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/ops/softmax.hpp"
+#include "ck_tile/ops/topk.hpp"
+
+namespace ck_tile {
+
+struct MoeSortingPolicy
+{
+};
+} // namespace ck_tile

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include <string>
+#include <type_traits>
+
+namespace ck_tile {
+
+template <typename IndexType_, typename WeightType_, index_t InternalLoadUnroll_>
+struct MoeSortingProblem
+{
+    // TODO: this kernel only support warp per row
+    using WeightType = remove_cvref_t<WeightType_>;
+    using IndexType  = remove_cvref_t<IndexType_>;
+
+    static constexpr index_t WarpSize           = get_warp_size();
+    static constexpr index_t WarpsPerBlock      = 1;
+    static constexpr index_t InternalLoadUnroll = InternalLoadUnroll_;
+};
+} // namespace ck_tile

include/ck_tile/ops/layernorm2d/kernel/layernorm2d_fwd_kernel.hpp

@@ -28,7 +28,10 @@ struct Layernorm2dFwdHostArgs
 
     index_t m;
     index_t n;
-    index_t stride; // row_stride
+    index_t x_stride;  // x row_stride
+    index_t xr_stride; // x residule row stride
+    index_t y_stride;  // y row stride
+    index_t yr_stride; // y residule row stride
 };
 
 // TODO: Extract some type to wrapper class
@@ -93,7 +96,10 @@ struct Layernorm2dFwd
 
         index_t m;
         index_t n;
-        index_t stride; // row_stride
+        index_t x_stride;  // x row_stride
+        index_t xr_stride; // x residule row stride
+        index_t y_stride;  // y row stride
+        index_t yr_stride; // y residule row stride
     };
     using Hargs = Layernorm2dFwdHostArgs;
 
@@ -112,7 +118,10 @@ struct Layernorm2dFwd
                      hargs.epsilon,
                      hargs.m,
                      hargs.n,
-                     hargs.stride};
+                     hargs.x_stride,
+                     hargs.xr_stride,
+                     hargs.y_stride,
+                     hargs.yr_stride};
     }
 
     CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
@@ -182,7 +191,7 @@ struct Layernorm2dFwd
             const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                 static_cast<const XDataType*>(kargs.p_x),
                 make_tuple(kargs.m, kargs.n),
-                make_tuple(kargs.stride, 1),
+                make_tuple(kargs.x_stride, 1),
                 number<Vector_N>{},
                 number<1>{});
 
@@ -201,7 +210,7 @@ struct Layernorm2dFwd
                 const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                     static_cast<const XResidualDataType*>(kargs.p_x_residual),
                     make_tuple(kargs.m, kargs.n),
-                    make_tuple(kargs.stride, 1),
+                    make_tuple(kargs.xr_stride, 1),
                     number<Vector_N>{},
                     number<1>{});
 
@@ -250,7 +259,7 @@ struct Layernorm2dFwd
             auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                 static_cast<YDataType*>(kargs.p_y),
                 make_tuple(kargs.m, kargs.n),
-                make_tuple(kargs.stride, 1),
+                make_tuple(kargs.y_stride, 1),
                 number<Vector_N>{},
                 number<1>{});
 
@@ -266,7 +275,7 @@ struct Layernorm2dFwd
                 auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                     static_cast<YResidualDataType*>(kargs.p_y_residual),
                     make_tuple(kargs.m, kargs.n),
-                    make_tuple(kargs.stride, 1),
+                    make_tuple(kargs.yr_stride, 1),
                     number<Vector_N>{},
                     number<1>{});
 

include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_default_policy.hpp

@@ -47,7 +47,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
     {
         using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
                                        typename Problem::ComputeDataType,
-                                       typename Problem::BlockShape>;
+                                       typename Problem::BlockShape,
+                                       Problem::Traits::kFastFDiv>;
 
         return BlockWelford<P_>{};
     }
@@ -57,7 +58,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
     {
         using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
                                        typename Problem::ComputeDataType,
-                                       typename Problem::BlockShape>;
+                                       typename Problem::BlockShape,
+                                       Problem::Traits::kFastFDiv>;
 
         return BlockWelfordSync<P_>{};
     }
@@ -67,7 +69,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
     {
         using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
                                        typename Problem::ComputeDataType,
-                                       typename Problem::BlockShape>;
+                                       typename Problem::BlockShape,
+                                       Problem::Traits::kFastFDiv>;
 
         return BlockWelfordCrossWarpSync<P_>{};
     }
@@ -79,7 +82,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
         {
             using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
                                            typename Problem::ComputeDataType,
-                                           typename Problem::BlockShape>;
+                                           typename Problem::BlockShape,
+                                           Problem::Traits::kFastFDiv>;
 
             using block_welford = BlockWelford<P_>;
             using x_block_tile =

include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.hpp

@@ -36,6 +36,7 @@ struct Layernorm2dFwdPipelineOnePass
     static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
     static constexpr bool kPadM              = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
     static constexpr bool kPadN              = Problem::Traits::kPadN;
+    static constexpr bool kFastFDiv          = Problem::Traits::kFastFDiv;
     static constexpr auto kFusedAdd          = Problem::Traits::kFusedAdd;
     static constexpr auto kFusedQuant        = Problem::Traits::kFusedQuant;
 
@@ -125,7 +126,15 @@ struct Layernorm2dFwdPipelineOnePass
         // compute inv-std
         auto inv_std = tile_elementwise_in(
             [&](const auto& v_) {
-                return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ + epsilon));
+                if(kFastFDiv && std::is_same_v<ComputeDataType, float>)
+                {
+                    return type_convert<ComputeDataType>(1.0f) *
+                           __builtin_amdgcn_rcpf(sqrt(v_ + epsilon));
+                }
+                else
+                {
+                    return type_convert<ComputeDataType>(1.0f) / sqrt(v_ + epsilon);
+                }
             },
             var);
 

include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_traits.hpp

@@ -39,6 +39,7 @@ template<> struct Layernorm2dFusedQuantEnumName<Layernorm2dFusedQuantEnum::SMOOT
 
 template <bool kPadN_,
           bool kSaveMeanInvStd_,
+          bool kFastFDiv_,
           bool kTwoPass_,
           Layernorm2dFusedAddEnum kFusedAdd_,
           Layernorm2dFusedQuantEnum kFusedQuant_>
@@ -46,6 +47,7 @@ struct Layernorm2dFwdTraits
 {
     static constexpr bool kPadN                            = kPadN_;
     static constexpr bool kSaveMeanInvStd                  = kSaveMeanInvStd_;
+    static constexpr bool kFastFDiv                        = kFastFDiv_;
     static constexpr bool kTwoPass                         = kTwoPass_;
     static constexpr Layernorm2dFusedAddEnum kFusedAdd     = kFusedAdd_;
     static constexpr Layernorm2dFusedQuantEnum kFusedQuant = kFusedQuant_;

include/ck_tile/ops/moe_sorting.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#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/common/generic_2d_block_shape.hpp"
+#include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/welford/block/block_welford.hpp

@@ -11,9 +11,10 @@ namespace ck_tile {
 template <typename Problem_, typename Policy_ = void>
 struct BlockWelford
 {
-    using Problem         = remove_cvref_t<Problem_>;
-    using XDataType       = typename Problem::XDataType;
-    using ComputeDataType = typename Problem::ComputeDataType;
+    using Problem                   = remove_cvref_t<Problem_>;
+    using XDataType                 = typename Problem::XDataType;
+    using ComputeDataType           = typename Problem::ComputeDataType;
+    static constexpr bool kFastFDiv = Problem::kFastFDiv;
 
     CK_TILE_DEVICE constexpr BlockWelford() {}
 
@@ -89,7 +90,8 @@ struct BlockWelford
 template <typename Problem_, typename Policy_ = void>
 struct BlockWelfordSync
 {
-    using Problem = remove_cvref_t<Problem_>;
+    using Problem                   = remove_cvref_t<Problem_>;
+    static constexpr bool kFastFDiv = Problem::kFastFDiv;
 
     template <typename MeanDistributedTensor_, typename VarDistributedTensor_>
     CK_TILE_DEVICE void
@@ -173,8 +175,9 @@ struct BlockWelfordSync
 template <typename Problem_, typename Policy_ = void>
 struct BlockWelfordCrossWarpSync
 {
-    using Problem    = remove_cvref_t<Problem_>;
-    using BlockShape = typename Problem::BlockShape;
+    using Problem                   = remove_cvref_t<Problem_>;
+    using BlockShape                = typename Problem::BlockShape;
+    static constexpr bool kFastFDiv = Problem::kFastFDiv;
 
     template <typename MeanDistributedTensor_>
     CK_TILE_DEVICE static constexpr index_t GetReduceWarps()
@@ -351,12 +354,23 @@ CK_TILE_DEVICE constexpr index_t block_tile_welford_calculate_max_count(int row_
 }
 
 // Note: this function must be called after all the computation
-template <typename VarDistributedTensor_>
+template <typename VarDistributedTensor_, bool FastFdiv_ = false>
 CK_TILE_DEVICE constexpr void block_tile_welford_post_scale_var(VarDistributedTensor_& var_tensor,
-                                                                int count)
+                                                                int count,
+                                                                bool_constant<FastFdiv_> = {})
 {
     using DataType = typename VarDistributedTensor_::DataType;
-    tile_elementwise_inout([&count](auto& x) { x = x / type_convert<DataType>(count); },
-                           var_tensor);
+    tile_elementwise_inout(
+        [&count](auto& x) {
+            if(FastFdiv_ && std::is_same_v<DataType, float>)
+            {
+                x = x * __builtin_amdgcn_rcpf(type_convert<DataType>(count));
+            }
+            else
+            {
+                x = x / type_convert<DataType>(count);
+            }
+        },
+        var_tensor);
 }
 } // namespace ck_tile

include/ck_tile/ops/welford/block/block_welford_problem.hpp

@@ -7,12 +7,13 @@
 
 namespace ck_tile {
 
-template <typename XDataType_, typename ComputeDataType_, typename BlockShape_>
+template <typename XDataType_, typename ComputeDataType_, typename BlockShape_, bool kFastFDiv_>
 struct BlockWelfordProblem
 {
-    using XDataType       = remove_cvref_t<XDataType_>;
-    using ComputeDataType = remove_cvref_t<ComputeDataType_>;
-    using BlockShape      = remove_cvref_t<BlockShape_>;
+    using XDataType                 = remove_cvref_t<XDataType_>;
+    using ComputeDataType           = remove_cvref_t<ComputeDataType_>;
+    using BlockShape                = remove_cvref_t<BlockShape_>;
+    static constexpr bool kFastFDiv = kFastFDiv_;
 };
 
 } // namespace ck_tile

include/ck_tile/ops/welford/thread/thread_welford.hpp

@@ -7,25 +7,46 @@
 
 namespace ck_tile {
 
-template <typename T>
-CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count)
+template <typename T, bool kFastFDiv = false>
+CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count, bool_constant<kFastFDiv> = {})
 {
     // TODO: check nan? maybe no
     T delta = x - mean;
-    mean += delta / count;
+    if(kFastFDiv && std::is_same_v<T, float>)
+    {
+        mean += delta * __builtin_amdgcn_rcpf(count);
+    }
+    else
+    {
+        mean += delta / count;
+    }
     T delta2 = x - mean;
     var += delta * delta2;
 }
 
-template <typename T>
-CK_TILE_DEVICE static void
-welford_merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
+template <typename T, bool kFastFDiv = false>
+CK_TILE_DEVICE static void welford_merge(T& mean_a,
+                                         T& var_a,
+                                         int& count_a,
+                                         T mean_b,
+                                         T var_b,
+                                         int count_b,
+                                         bool_constant<kFastFDiv> = {})
 {
-    int count            = count_a + count_b;
-    T count_             = type_convert<T>(count);
-    T count_a_           = type_convert<T>(count_a);
-    T count_b_           = type_convert<T>(count_b);
-    T count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_;
+    int count  = count_a + count_b;
+    T count_   = type_convert<T>(count);
+    T count_a_ = type_convert<T>(count_a);
+    T count_b_ = type_convert<T>(count_b);
+    T count_b_over_count;
+    if(kFastFDiv && std::is_same_v<T, float>)
+    {
+        count_b_over_count =
+            count == 0 ? type_convert<T>(0) : count_b_ * __builtin_amdgcn_rcpf(count_);
+    }
+    else
+    {
+        count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_;
+    }
 
     T delta = mean_b - mean_a;
     mean_a += delta * count_b_over_count;

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp

@@ -39,7 +39,25 @@ template <ck::index_t NDimSpatial,
           ConvolutionBackwardWeightSpecialization ConvSpec,
           BlockGemmPipelineScheduler Scheduler,
           BlockGemmPipelineVersion PipelineVersion>
-using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances = std::tuple<
+using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
+        //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|   ToMerge|
+        //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
+        //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>
+        // clang-format on
+        >;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances = std::tuple<
     // clang-format off
         //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
         //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|   ToMerge|
@@ -64,7 +82,25 @@ template <ck::index_t NDimSpatial,
           ConvolutionBackwardWeightSpecialization ConvSpec,
           BlockGemmPipelineScheduler Scheduler,
           BlockGemmPipelineVersion PipelineVersion>
-using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances = std::tuple<
+using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
+        //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|   ToMerge|
+        //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
+        //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>
+        // clang-format on
+        >;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instances = std::tuple<
     // clang-format off
         //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
         //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|   ToMerge|
@@ -82,6 +118,24 @@ using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances = st
     // clang-format on
     >;
 
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
+        //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|   ToMerge|
+        //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
+        //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>
+        // clang-format on
+        >;
+
 // NGCHW requires transpose, we use vector loads and stores params for them
 template <ck::index_t NDimSpatial,
           typename ALayout,
@@ -122,6 +176,24 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances
     // clang-format on
     >;
 
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
+        //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|   ToMerge|
+        //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
+        //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,    F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>
+        // clang-format on
+        >;
+
 template <ck::index_t NDimSpatial,
           typename ALayout,
           typename BLayout,