Merge branch 'amd-develop' into amd-master

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

@@ -32,6 +32,50 @@ struct TileFmhaTraits
     static constexpr index_t kBlockPerCu    = kBlockPerCu_;
 };
 
+template <bool kPadSeqLenQ /* padding for seqlen_q */,
+          bool kPadSeqLenK /* padding for seqlen_k */,
+          bool kPadHeadDimQ /* paddding for hdim_q */,
+          bool kPadHeadDimV /* paddding for hdim_v */,
+          BlockAttentionBiasEnum BiasEnum,
+          bool kHasBiasGrad,
+          bool kStoreLSE,
+          bool kHasDropout,
+          bool kDoFp8StaticQuant,
+          bool kHasUnevenSplits_ = true,
+          index_t kBlockPerCu    = -1 /* overwrite occupancy if not -1 */>
+struct TileFmhaFwdSplitKVTraits : TileFmhaTraits<kPadSeqLenQ,
+                                                 kPadSeqLenK,
+                                                 kPadHeadDimQ,
+                                                 kPadHeadDimV,
+                                                 BiasEnum,
+                                                 kHasBiasGrad,
+                                                 kStoreLSE,
+                                                 kHasDropout,
+                                                 kDoFp8StaticQuant,
+                                                 kBlockPerCu>
+{
+    // determine if some split (length) is not divisible by tile size
+    static constexpr bool kHasUnevenSplits = kHasUnevenSplits_;
+};
+
+template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
+          bool kPadHeadDimV_ /* paddding for hdim_v */,
+          bool kStoreLSE_,
+          bool kDoFp8StaticQuant_,
+          index_t kLogMaxSplits_,
+          index_t kBlockPerCu_ = -1 /* overwrite occupancy if not -1 */>
+struct TileFmhaFwdSplitKVCombineTraits
+{
+    static constexpr bool kPadSeqLenQ       = kPadSeqLenQ_;
+    static constexpr bool kPadHeadDimV      = kPadHeadDimV_;
+    static constexpr bool kStoreLSE         = kStoreLSE_;
+    static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_;
+
+    static constexpr index_t kMaxSplits = (1 << kLogMaxSplits_);
+    static_assert(kMaxSplits <= get_warp_size() || kMaxSplits % get_warp_size() == 0);
+    static constexpr index_t kBlockPerCu = kBlockPerCu_;
+};
+
 template <bool kPadSeqLenQ_ /* padding for seqlen_q */,
           bool kPadHeadDimV_ /* paddding for hdim_v */,
           index_t kBlockPerCu_ = 2 /* hint to occupancy */>

include/ck_tile/ops/layernorm2d.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/ops/layernorm2d/kernel/layernorm2d_fwd_kernel.hpp"
+#include "ck_tile/ops/layernorm2d/pipeline/block_layernorm2d_fwd_problem.hpp"
+#include "ck_tile/ops/layernorm2d/pipeline/tile_layernorm2d_fwd_shape.hpp"
+#include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/layernorm2d/kernel/layernorm2d_fwd_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/welford/thread/thread_welford.hpp"
+#include "ck_tile/ops/welford/warp/warp_welford.hpp"
+
+namespace ck_tile {
+
+// TODO: Extract some type to wrapper class
+template <typename Problem_>
+struct Layernorm2dFwd
+{
+    using Problem = ck_tile::remove_cvref_t<Problem_>;
+
+    using XDataType       = ck_tile::remove_cvref_t<typename Problem::XDataType>;
+    using GammaDataType   = ck_tile::remove_cvref_t<typename Problem::GammaDataType>;
+    using BetaDataType    = ck_tile::remove_cvref_t<typename Problem::BetaDataType>;
+    using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
+    using YDataType       = ck_tile::remove_cvref_t<typename Problem::YDataType>;
+    using MeanDataType    = ck_tile::remove_cvref_t<typename Problem::MeanDataType>;
+    using InvStdDataType  = ck_tile::remove_cvref_t<typename Problem::InvStdDataType>;
+
+    static constexpr bool kHasGamma   = !std::is_same_v<GammaDataType, ck_tile::null_type>;
+    static constexpr bool kHasBeta    = !std::is_same_v<BetaDataType, ck_tile::null_type>;
+    static constexpr bool kSaveMean   = !std::is_same_v<MeanDataType, ck_tile::null_type>;
+    static constexpr bool kSaveInvStd = !std::is_same_v<InvStdDataType, ck_tile::null_type>;
+
+    static constexpr ck_tile::index_t kMPerBlock = Problem::BlockShape::kMPerBlock;
+    static constexpr ck_tile::index_t kNPerBlock = Problem::BlockShape::kNPerBlock;
+
+    static constexpr ck_tile::index_t kNThreadPerWarp = Problem::BlockShape::kNThreadPerWarp;
+
+    struct Kargs
+    {
+        const void* p_x;
+        const void* p_gamma;
+        const void* p_beta;
+
+        void* p_y;
+        void* p_mean;
+        void* p_invStd;
+
+        float epsilon;
+
+        ck_tile::index_t M;
+        ck_tile::index_t N;
+    };
+
+    CK_TILE_HOST static constexpr Kargs MakeKargs(const void* p_x,
+                                                  const void* p_gamma,
+                                                  const void* p_beta,
+                                                  void* p_y,
+                                                  void* p_mean,
+                                                  void* p_invStd,
+                                                  float epsilon,
+                                                  ck_tile::index_t M,
+                                                  ck_tile::index_t N)
+    {
+        return Kargs{p_x, p_gamma, p_beta, p_y, p_mean, p_invStd, epsilon, M, N};
+    }
+
+    CK_TILE_HOST static constexpr auto GridSize(ck_tile::index_t M) { return M / kMPerBlock; }
+
+    CK_TILE_HOST static constexpr auto BlockSize() { return Problem::BlockShape::kBlockSize; }
+
+    CK_TILE_DEVICE static constexpr auto MakeXBlockTileDistribution()
+    {
+        using S = typename Problem::BlockShape;
+
+        return make_static_tile_distribution(
+            tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<S::kMWarpPerBlock, S::kMThreadPerWarp, S::kMPerThread>,
+                      sequence<S::kNWarpPerBlock, S::kNThreadPerWarp, S::kNPerThread>>,
+                tuple<sequence<1, 2>, sequence<1, 2>>,
+                tuple<sequence<0, 0>, sequence<1, 1>>,
+                sequence<1, 2>,
+                sequence<2, 2>>{});
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeGammaBetaBlockTileDistribution()
+    {
+        using S = typename Problem::BlockShape;
+
+        return make_static_tile_distribution(
+            tile_distribution_encoding<
+                sequence<S::kMWarpPerBlock, S::kMThreadPerWarp>,
+                tuple<sequence<S::kNWarpPerBlock, S::kNThreadPerWarp, S::kNPerThread>>,
+                tuple<sequence<0, 1>, sequence<0, 1>>,
+                tuple<sequence<0, 0>, sequence<1, 1>>,
+                sequence<1>,
+                sequence<2>>{});
+    }
+
+    template <typename Dstr>
+    CK_TILE_DEVICE static constexpr auto GetNPerThread(Dstr)
+    {
+        constexpr auto nDstrSpan = Dstr::get_distributed_spans().template at<1>();
+
+        using Lengths = decltype(nDstrSpan.impl_);
+
+        ck_tile::index_t ret = 1;
+
+        ck_tile::static_for<0, Lengths::size(), 1>{}(
+            [&](auto idx) { ret *= Lengths::template at(idx); });
+
+        return ret;
+    }
+
+    template <typename DistributedTensor>
+    CK_TILE_DEVICE static auto InvSqrt(const DistributedTensor& in_dstr_tensor,
+                                       const ComputeDataType epsilon)
+    {
+        // TODO: Investigate fast inverse square root algorithm with epsilon
+        constexpr auto spans = DistributedTensor::get_distributed_spans();
+
+        DistributedTensor out_dstr_tensor;
+
+        sweep_tile_span(spans[number<0>{}], [&](auto idx0) {
+            constexpr auto i_idx   = make_tuple(idx0);
+            out_dstr_tensor(i_idx) = type_convert<ComputeDataType>(1.0f) /
+                                     ck_tile::sqrt(in_dstr_tensor[i_idx] + epsilon);
+        });
+
+        return out_dstr_tensor;
+    }
+
+    template <bool Cond = (kHasGamma && kHasBeta)>
+    CK_TILE_DEVICE std::enable_if_t<Cond> TwoPassLayernorm2dFwd(const XDataType* p_x,
+                                                                const GammaDataType* p_gamma,
+                                                                const BetaDataType* p_beta,
+                                                                YDataType* p_y,
+                                                                MeanDataType* p_mean,
+                                                                InvStdDataType* p_invStd,
+                                                                const ComputeDataType epsilon,
+                                                                ck_tile::index_t M,
+                                                                ck_tile::index_t N) const
+    {
+        constexpr auto I0 = number<0>{};
+        constexpr auto I1 = number<1>{};
+
+        const auto x_m_n = make_naive_tensor_view<address_space_enum::global>(
+            p_x, make_tuple(M, N), make_tuple(N, 1), number<32>{}, number<1>{});
+
+        const auto gamma_n = make_naive_tensor_view<address_space_enum::global>(
+            p_gamma, make_tuple(N), make_tuple(1), number<32>{}, number<1>{});
+
+        const auto beta_n = make_naive_tensor_view<address_space_enum::global>(
+            p_beta, make_tuple(N), make_tuple(1), number<32>{}, number<1>{});
+
+        const auto iM = get_block_id() * kMPerBlock;
+
+        constexpr auto xDstr = MakeXBlockTileDistribution();
+
+        auto x_block_window = make_tile_window(
+            x_m_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {iM, 0}, xDstr);
+
+        index_t num_n_tile_iteration = __builtin_amdgcn_readfirstlane(N / kNPerBlock);
+
+        // TODO: padding - handle max_count if N % kNPerBlock != 0
+        constexpr auto NPerThread = GetNPerThread(xDstr);
+        ThreadWelford<ComputeDataType, XDataType> thread_welford{
+            type_convert<int>(NPerThread * N / kNPerBlock)};
+
+        using XTensorType = decltype(load_tile(x_block_window));
+        auto mean_compute_block_tensor =
+            thread_welford.template MakeInitialMeanVarDistributedTensor<XTensorType>();
+        auto var_compute_block_tensor =
+            thread_welford.template MakeInitialMeanVarDistributedTensor<XTensorType>();
+
+        clear_tile(mean_compute_block_tensor);
+        clear_tile(var_compute_block_tensor);
+
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
+        {
+            const auto x_block_tensor = load_tile(x_block_window);
+
+            thread_welford(x_block_tensor, mean_compute_block_tensor, var_compute_block_tensor);
+            move_tile_window(x_block_window, {0, kNPerBlock});
+        }
+
+        // TODO: support cross warp Welford
+        WarpMergeWelford<ComputeDataType, true>{}(
+            mean_compute_block_tensor, var_compute_block_tensor, thread_welford.cur_count_);
+
+        auto inv_std_compute_block_tensor = InvSqrt(var_compute_block_tensor, epsilon);
+
+        if constexpr(kSaveMean)
+        {
+            const auto mean_m = make_naive_tensor_view_packed<address_space_enum::global>(
+                p_mean, make_tuple(M), number<32>{});
+
+            auto mean_block_window =
+                make_tile_window(mean_m, make_tuple(number<kMPerBlock>{}), {iM});
+
+            store_tile(mean_block_window, cast_tile<MeanDataType>(mean_compute_block_tensor));
+        }
+        if constexpr(kSaveInvStd)
+        {
+            const auto inv_std_m = make_naive_tensor_view_packed<address_space_enum::global>(
+                p_invStd, make_tuple(M), number<32>{});
+
+            auto inv_std_block_window =
+                make_tile_window(inv_std_m, make_tuple(number<kMPerBlock>{}), {iM});
+
+            store_tile(inv_std_block_window, cast_tile<MeanDataType>(inv_std_compute_block_tensor));
+        }
+
+        // TODO: Extract normalize pipeline
+        const auto y_m_n = make_naive_tensor_view<address_space_enum::global>(
+            p_y, make_tuple(M, N), make_tuple(N, 1), number<32>{}, number<1>{});
+
+        auto y_block_window = make_tile_window(
+            y_m_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {iM, 0});
+
+        constexpr auto gammaDstr = MakeGammaBetaBlockTileDistribution();
+        constexpr auto betaDstr  = gammaDstr;
+
+        auto gamma_block_window =
+            make_tile_window(gamma_n, make_tuple(number<kNPerBlock>{}), {0}, gammaDstr);
+
+        auto beta_block_window = make_tile_window(
+            beta_n, make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}), {0}, betaDstr);
+
+        // reverse read x to reuse cache
+        ck_tile::index_t stride_to_right_most_window = N - kNPerBlock;
+
+        move_tile_window(x_block_window, {0, -kNPerBlock});
+        move_tile_window(gamma_block_window, {stride_to_right_most_window});
+        move_tile_window(beta_block_window, {stride_to_right_most_window});
+        move_tile_window(y_block_window, {0, stride_to_right_most_window});
+
+        // Normalization
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
+        {
+            const auto x_block_tensor     = load_tile(x_block_window);
+            const auto gamma_block_tensor = load_tile(gamma_block_window);
+            const auto beta_block_tensor  = load_tile(beta_block_window);
+
+            constexpr auto x_spans = decltype(x_block_tensor)::get_distributed_spans();
+
+            auto y_block_tensor =
+                make_static_distributed_tensor<YDataType>(x_block_tensor.get_tile_distribution());
+
+            sweep_tile_span(x_spans[I1], [&](auto idx1) {
+                constexpr auto j_idx = make_tuple(idx1);
+                const auto gamma     = type_convert<ComputeDataType>(gamma_block_tensor[j_idx]);
+                const auto beta      = type_convert<ComputeDataType>(beta_block_tensor[j_idx]);
+
+                sweep_tile_span(x_spans[I0], [&](auto idx0) {
+                    constexpr auto i_idx   = make_tuple(idx0);
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                    const auto mean    = mean_compute_block_tensor[i_idx];
+                    const auto inv_std = inv_std_compute_block_tensor[i_idx];
+
+                    const auto x = type_convert<ComputeDataType>(x_block_tensor[i_j_idx]);
+                    auto y       = (x - mean) * inv_std * gamma + beta;
+
+                    y_block_tensor(i_j_idx) = type_convert<YDataType>(y);
+                });
+            });
+
+            store_tile(y_block_window, y_block_tensor);
+
+            move_tile_window(x_block_window, {0, -kNPerBlock});
+            move_tile_window(gamma_block_window, {-kNPerBlock});
+            move_tile_window(beta_block_window, {-kNPerBlock});
+            move_tile_window(y_block_window, {0, -kNPerBlock});
+        }
+    }
+
+    CK_TILE_DEVICE void operator()(Kargs kargs) const
+    {
+        TwoPassLayernorm2dFwd(static_cast<const XDataType*>(kargs.p_x),
+                              static_cast<const GammaDataType*>(kargs.p_gamma),
+                              static_cast<const BetaDataType*>(kargs.p_beta),
+                              static_cast<YDataType*>(kargs.p_y),
+                              static_cast<MeanDataType*>(kargs.p_mean),
+                              static_cast<InvStdDataType*>(kargs.p_invStd),
+                              static_cast<const ComputeDataType>(kargs.epsilon),
+                              kargs.M,
+                              kargs.N);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/layernorm2d/pipeline/block_layernorm2d_fwd_problem.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core/utility/type_traits.hpp"
+
+namespace ck_tile {
+
+template <typename XDataType_,
+          typename GammaDataType_,
+          typename BetaDataType_,
+          typename ComputeDataType_,
+          typename YDataType_,
+          typename MeanDataType_,
+          typename InvStdDataType_,
+          typename BlockShape_>
+struct BlockLayernorm2dFwdProblem
+{
+    using XDataType       = remove_cvref_t<XDataType_>;
+    using GammaDataType   = remove_cvref_t<GammaDataType_>;
+    using BetaDataType    = remove_cvref_t<BetaDataType_>;
+    using ComputeDataType = remove_cvref_t<ComputeDataType_>;
+    using YDataType       = remove_cvref_t<YDataType_>;
+    using MeanDataType    = remove_cvref_t<MeanDataType_>;
+    using InvStdDataType  = remove_cvref_t<InvStdDataType_>;
+    using BlockShape      = remove_cvref_t<BlockShape_>;
+};
+
+} // namespace ck_tile

include/ck_tile/ops/layernorm2d/pipeline/tile_layernorm2d_fwd_shape.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+template <typename ThreadTile, // Sequence<...
+          typename WarpTile,   // Sequence<...
+          typename BlockTile>  // Sequence<...
+struct TileLayernorm2dShape
+{
+    static constexpr index_t kMPerThread = ThreadTile::at(number<0>{});
+    static constexpr index_t kNPerThread = ThreadTile::at(number<1>{});
+
+    static constexpr index_t kMPerWarp = WarpTile::at(number<0>{});
+    static constexpr index_t kNPerWarp = WarpTile::at(number<1>{});
+
+    static constexpr index_t kMThreadPerWarp = kMPerWarp / kMPerThread;
+    static constexpr index_t kNThreadPerWarp = kNPerWarp / kNPerThread;
+
+    static constexpr index_t kMPerBlock = BlockTile::at(number<0>{});
+    static constexpr index_t kNPerBlock = BlockTile::at(number<1>{});
+
+    static constexpr index_t kMWarpPerBlock = kMPerBlock / kMPerWarp;
+    static constexpr index_t kNWarpPerBlock = kNPerBlock / kNPerWarp;
+
+    // TODO - kNNumWarps can only be 1 if we don't support cross warp welford
+    static_assert(kNWarpPerBlock == 1);
+
+    static constexpr index_t kBlockSize = warpSize * kMWarpPerBlock * kNWarpPerBlock;
+};
+
+} // namespace ck_tile

include/ck_tile/ops/welford.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/ops/welford/thread/thread_welford.hpp"
+#include "ck_tile/ops/welford/warp/warp_welford.hpp"
+#include "ck_tile/ops/common/tensor_layout.hpp"

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+template <typename ComputeDataType_, typename XDataType_>
+struct ThreadWelford
+{
+    using XDataType       = remove_cvref_t<XDataType_>;
+    using ComputeDataType = remove_cvref_t<ComputeDataType_>;
+
+    template <typename T>
+    CK_TILE_DEVICE void Update(T& mean, T& var, T x)
+    {
+        if(ck_tile::isnan(x))
+        {
+            mean = x;
+            var  = x;
+        }
+        else
+        {
+            T delta = x - mean;
+            mean += delta / cur_count_;
+            T delta2 = x - mean;
+            var += delta * delta2;
+        }
+    }
+
+    // [CAUSION] - max_count_ is to deal with the padding problem
+    // max_count_ is depend on caller, eg: naive and splitN welford will have different
+    // calculation of max_count_
+    CK_TILE_DEVICE constexpr ThreadWelford(int max_count) : cur_count_(0), max_count_(max_count) {}
+
+    template <typename XDistributedTensor_,
+              typename MeanDistributedTensor_,
+              typename VarDistributedTensor_>
+    CK_TILE_DEVICE void operator()(const XDistributedTensor_& x_tensor,
+                                   MeanDistributedTensor_& mean_tensor,
+                                   VarDistributedTensor_& var_tensor)
+    {
+        constexpr auto I0 = number<0>{};
+        constexpr auto I1 = number<1>{};
+
+        constexpr auto spans = XDistributedTensor_::get_distributed_spans();
+
+        sweep_tile_span(spans[I1], [&](auto dstr_idx_i1) {
+            if(cur_count_ < max_count_)
+            {
+                ++cur_count_;
+
+                sweep_tile_span(spans[I0], [&](auto dstr_idx_i0) {
+                    constexpr auto in_dstr_idx  = make_tuple(dstr_idx_i0, dstr_idx_i1);
+                    constexpr auto out_dstr_idx = make_tuple(dstr_idx_i0);
+
+                    auto x = ck_tile::type_convert<ComputeDataType>(x_tensor[in_dstr_idx]);
+
+                    Update(mean_tensor(out_dstr_idx), var_tensor(out_dstr_idx), x);
+                });
+            }
+        });
+    }
+
+    template <typename XDistributedTensor_>
+    CK_TILE_DEVICE static auto MakeInitialMeanVarDistributedTensor()
+    {
+        static_assert(std::is_same_v<XDataType, typename XDistributedTensor_::DataType>, "wrong!");
+
+        constexpr auto reduce_dims = sequence<1>{};
+
+        constexpr auto dstr =
+            make_static_tile_distribution(detail::make_reduce_tile_distribution_encoding(
+                XDistributedTensor_::get_tile_distribution()
+                    .get_static_tile_distribution_encoding(),
+                reduce_dims));
+
+        auto tensor = make_static_distributed_tensor<ComputeDataType>(dstr);
+        clear_tile(tensor);
+
+        return tensor;
+    }
+
+    template <typename XDistributedTensor_>
+    CK_TILE_DEVICE auto operator()(const XDistributedTensor_& x_tensor)
+    {
+        auto mean_tensor = MakeInitialMeanVarDistributedTensor<XDistributedTensor_>();
+        auto var_tensor  = MakeInitialMeanVarDistributedTensor<XDistributedTensor_>();
+
+        (*this)(x_tensor, mean_tensor, var_tensor);
+
+        return ck_tile::make_tuple(mean_tensor, var_tensor);
+    }
+
+    int cur_count_;
+    int max_count_;
+};
+
+} // namespace ck_tile

include/ck_tile/ops/welford/warp/warp_welford.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+template <typename ComputeDataType_, bool BroadcastLane = true, bool GetActualVariance = true>
+struct WarpMergeWelford
+{
+    using ComputeDataType = remove_cvref_t<ComputeDataType_>;
+
+    template <typename T>
+    CK_TILE_DEVICE static void
+    Merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
+    {
+        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_;
+
+        T delta = mean_b - mean_a;
+        mean_a += delta * count_b_over_count;
+        var_a += var_b + delta * delta * count_a_ * count_b_over_count;
+        count_a = count;
+    }
+
+    template <typename MeanDistributedTensor_, typename VarDistributedTensor_>
+    CK_TILE_DEVICE void
+    operator()(MeanDistributedTensor_& mean_tensor, VarDistributedTensor_& var_tensor, int& count)
+    {
+        using Dstr             = typename MeanDistributedTensor_::StaticTileDistribution;
+        using DstrEncode       = typename Dstr::DstrEncode;
+        using DstrEncodeDetail = typename DstrEncode::detail;
+
+        static_assert(std::is_same_v<Dstr, typename VarDistributedTensor_::StaticTileDistribution>,
+                      "wrong!");
+
+        constexpr index_t NDimP = Dstr::get_num_of_dimension_p();
+        constexpr index_t NDimR = Dstr::get_num_of_dimension_r();
+
+        constexpr index_t idim_p_lane = NDimP - 1;
+
+        const auto ps_idx = make_array<index_t>(get_warp_id(), get_lane_id());
+        const auto rs_idx =
+            mean_tensor.get_tile_distribution().calculate_rs_index_from_ps_index(ps_idx);
+
+        constexpr index_t thread_buf_size = MeanDistributedTensor_::get_thread_buffer_size();
+        static_assert(thread_buf_size == VarDistributedTensor_::get_thread_buffer_size());
+
+        const int original_count = count;
+
+        // loop over thread data
+        static_for<0, thread_buf_size, 1>{}([&](auto i) {
+            auto v_local_mean  = mean_tensor.get_thread_buffer()[i];
+            auto v_local_var   = var_tensor.get_thread_buffer()[i];
+            auto v_local_count = original_count;
+
+            // cross-lane reduce for replication
+            // only reduce on R dimension correspond to lane
+            // (lane id maps to this R dimension)
+            static_for<0, NDimR, 1>{}([&](auto idim_r) {
+                // FIXME: nasty to use does_p_own_r_
+                if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_lane][idim_r])
+                {
+                    constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
+
+                    constexpr index_t lid_over_rid_derivative =
+                        DstrEncodeDetail::ps_over_rs_derivative_[idim_p_lane][idim_r];
+
+                    static_assert(is_power_of_two_integer(r_length),
+                                  "wrong! only support power of 2 reduction");
+
+                    constexpr index_t nstage = integer_log2_floor(r_length);
+
+                    // reduction sweep forward
+                    static_for<0, nstage, 1>{}([&](auto istage) {
+                        constexpr index_t lid_delta =
+                            lid_over_rid_derivative * (1 << (nstage - istage - 1));
+
+                        // pull data from remote lane
+                        const auto v_remote_mean  = warp_shuffle_down(v_local_mean, lid_delta);
+                        const auto v_remote_var   = warp_shuffle_down(v_local_var, lid_delta);
+                        const auto v_remote_count = warp_shuffle_down(v_local_count, lid_delta);
+
+                        // welford merge
+                        Merge(v_local_mean,
+                              v_local_var,
+                              v_local_count,
+                              v_remote_mean,
+                              v_remote_var,
+                              v_remote_count);
+                    });
+                }
+            });
+
+            // cross-lane broadcast for replication
+            // only broadcast on R dimension correspond to lane
+            // (lane id maps to this R dimension)
+            if constexpr(BroadcastLane)
+            {
+                static_for<0, NDimR, 1>{}([&](auto idim_r) {
+                    // FIXME: nasty to use does_p_own_r_
+                    if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_lane][idim_r])
+                    {
+                        const index_t r_id = rs_idx[idim_r];
+
+                        constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
+
+                        constexpr index_t lid_over_rid_derivative =
+                            DstrEncodeDetail::ps_over_rs_derivative_[NDimP - 1][idim_r];
+
+                        static_assert(is_power_of_two_integer(r_length),
+                                      "wrong! only support power of 2 reduction");
+
+                        constexpr index_t nstage = integer_log2_floor(r_length);
+
+                        // broadcast sweep backward
+                        static_for<0, nstage, 1>{}([&](auto istage) {
+                            // do I hold reduced data?
+                            const bool do_i_hold_reduced_data = r_id < (1 << istage);
+
+                            constexpr index_t lid_delta = lid_over_rid_derivative * (1 << istage);
+
+                            // pull data from remote lane
+                            const auto v_remote_mean  = warp_shuffle_up(v_local_mean, lid_delta);
+                            const auto v_remote_var   = warp_shuffle_up(v_local_var, lid_delta);
+                            const auto v_remote_count = warp_shuffle_up(v_local_count, lid_delta);
+
+                            // decide whether to update local data with remote data
+                            v_local_mean  = do_i_hold_reduced_data ? v_local_mean : v_remote_mean;
+                            v_local_var   = do_i_hold_reduced_data ? v_local_var : v_remote_var;
+                            v_local_count = do_i_hold_reduced_data ? v_local_count : v_remote_count;
+                        });
+                    }
+                });
+            }
+
+            mean_tensor.get_thread_buffer()(i) = v_local_mean;
+
+            if constexpr(GetActualVariance)
+                var_tensor.get_thread_buffer()(i) = v_local_var / v_local_count;
+            else
+                var_tensor.get_thread_buffer()(i) = v_local_var;
+
+            count = v_local_count;
+        });
+    }
+};
+
+} // namespace ck_tile

library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_streamk.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+#endif
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemm_Streamk_V2<
+    ALayout,
+    BLayout,
+    CLayout,
+    ADataType,
+    BDataType,
+    CDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemm_Streamk_V2<ALayout,
+                                           BLayout,
+                                           CLayout,
+                                           ADataType,
+                                           BDataType,
+                                           CDataType,
+                                           ck::tensor_operation::element_wise::PassThrough,
+                                           ck::tensor_operation::element_wise::PassThrough,
+                                           ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_outelementop_instance.hpp

@@ -15,6 +15,7 @@ namespace instance {
 
 using F32 = float;
 using F8  = ck::f8_t;
+using BF8 = ck::bf8_t;
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -72,6 +73,80 @@ using device_grouped_conv_fwd_xdl_outelementop_f8_instances = std::tuple<
     // clang-format on
     >;
 
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionForwardSpecialization ConvSpec,
+          typename OutElementOp>
+using device_grouped_conv_fwd_xdl_outelementop_bf8_instances = std::tuple<
+// clang-format off
+        //########################################|     NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|       Ds| EData|           A|           B|           CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer| Compute Type|
+        //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType| DataType|  Type| Elementwise| Elementwise|   Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|             |
+        //########################################|           |       |       |            |       |      |      |        |         |         |      |   Operation|   Operation|     Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|             |
+        //########################################|           |       |       |            |       |      |      |        |         |         |      |            |            |              |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |             |
+#if(defined(CK_ENABLE_FP8) && defined(CK_ENABLE_BF8))
+        // generic instance
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1,         BF8>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,         BF8>,
+
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8,         BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,   BF8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8,         BF8>
+#endif
+    // clang-format on
+    >;
+
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionForwardSpecialization ConvSpec,
+          typename OutElementOp>
+using device_grouped_conv_fwd_xdl_outelementop_f8_bf8_instances = std::tuple<
+// clang-format off
+        //########################################|     NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|       Ds| EData|           A|           B|           CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer| Compute| Compute|
+        //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType| DataType|  Type| Elementwise| Elementwise|   Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|   TypeA|   TypeB|
+        //########################################|           |       |       |            |       |      |      |        |         |         |      |   Operation|   Operation|     Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|        |        |
+        //########################################|           |       |       |            |       |      |      |        |         |         |      |            |            |              |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |        |        |
+#if defined(CK_ENABLE_FP8) && defined(CK_ENABLE_BF8)
+        // generic instance
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1,     F8,     BF8>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,     F8,     BF8>,
+
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8,     F8,     BF8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,  DsLayout,ELayout,    F8,   BF8,     F32,      F32,  Tuple<>,    F8, PassThrough, PassThrough,  OutElementOp,       ConvSpec, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8,     F8,     BF8>
+#endif
+    // clang-format on
+    >;
+
 } // namespace instance
 } // namespace device
 } // namespace tensor_operation

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp

@@ -402,6 +402,17 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
         if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
                      is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
         {
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                         is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t> &&
+                         is_same_v<BComputeType, half_t>)
+            {
+                add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instances(op_ptrs);
+            }
+#endif
 #ifdef CK_ENABLE_INT8
             if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
                          is_same_v<OutDataType, int8_t> && is_same_v<AComputeType, int8_t> &&

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convinvscale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ConvInvscale = ck::tensor_operation::element_wise::ConvInvscale;
+
+#ifdef CK_ENABLE_FP8
+void add_device_grouped_conv3d_fwd_xdl_convinvscale_ndhwgc_gkzyxc_ndhwgk_f8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
+                                                                NDHWGC,
+                                                                GKZYXC,
+                                                                ck::Tuple<>,
+                                                                NDHWGK,
+                                                                F8,
+                                                                F8,
+                                                                ck::Tuple<>,
+                                                                F8,
+                                                                PassThrough,
+                                                                PassThrough,
+                                                                ConvInvscale,
+                                                                F8,
+                                                                F8>>>& instances);
+#endif
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DLayouts,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DDataTypes,
+          typename OutDataType,
+          typename AComputeType,
+          typename BComputeType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DLayouts,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DDataTypes,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::ConvInvscale,
+    AComputeType,
+    BComputeType>>
+{
+    using DeviceOp =
+        DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                        InLayout,
+                                        WeiLayout,
+                                        DLayouts,
+                                        OutLayout,
+                                        InDataType,
+                                        WeiDataType,
+                                        DDataTypes,
+                                        OutDataType,
+                                        ck::tensor_operation::element_wise::PassThrough,
+                                        ck::tensor_operation::element_wise::PassThrough,
+                                        ck::tensor_operation::element_wise::ConvInvscale,
+                                        AComputeType,
+                                        BComputeType>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
+                     is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
+        {
+#ifdef CK_ENABLE_FP8
+            if constexpr(is_same_v<InDataType, f8_t> && is_same_v<WeiDataType, f8_t> &&
+                         is_same_v<OutDataType, f8_t> && is_same_v<AComputeType, f8_t> &&
+                         is_same_v<BComputeType, f8_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_convinvscale_ndhwgc_gkzyxc_ndhwgk_f8_instances(
+                    op_ptrs);
+            }
+#endif
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale.hpp

@@ -39,6 +39,39 @@ void add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_f8_instanc
                                                                 F8>>>& instances);
 #endif
 
+#if(defined(CK_ENABLE_FP8) && defined(CK_ENABLE_BF8))
+void add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_bf8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
+                                                                NDHWGC,
+                                                                GKZYXC,
+                                                                ck::Tuple<>,
+                                                                NDHWGK,
+                                                                BF8,
+                                                                BF8,
+                                                                ck::Tuple<>,
+                                                                F8,
+                                                                PassThrough,
+                                                                PassThrough,
+                                                                ConvScale,
+                                                                BF8>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_f8_bf8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
+                                                                NDHWGC,
+                                                                GKZYXC,
+                                                                ck::Tuple<>,
+                                                                NDHWGK,
+                                                                F8,
+                                                                BF8,
+                                                                ck::Tuple<>,
+                                                                F8,
+                                                                PassThrough,
+                                                                PassThrough,
+                                                                ConvScale,
+                                                                F8,
+                                                                BF8>>>& instances);
+#endif
+
 template <ck::index_t NumDimSpatial,
           typename InLayout,
           typename WeiLayout,
@@ -97,6 +130,24 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                     op_ptrs);
             }
 #endif
+
+#if(defined(CK_ENABLE_FP8) && defined(CK_ENABLE_BF8))
+            if constexpr(is_same_v<InDataType, BF8> && is_same_v<WeiDataType, BF8> &&
+                         is_same_v<OutDataType, F8> && is_same_v<AComputeType, BF8> &&
+                         is_same_v<BComputeType, BF8>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_bf8_instances(
+                    op_ptrs);
+            }
+
+            if constexpr(is_same_v<InDataType, f8_t> && is_same_v<WeiDataType, bf8_t> &&
+                         is_same_v<OutDataType, f8_t> && is_same_v<AComputeType, f8_t> &&
+                         is_same_v<BComputeType, bf8_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_f8_bf8_instances(
+                    op_ptrs);
+            }
+#endif
         }
         return op_ptrs;
     }

library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multply.hpp → library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multiply.hpp

@@ -17,7 +17,150 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instances(
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_comp_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_comp_mnpadding_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v1_mnpadding_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v2_mnpadding_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
     std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
                                                           Row,
                                                           Row_Tuple,
@@ -67,14 +210,35 @@ struct DeviceOperationInstanceFactory<
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
 
-        // fp16_output
         if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
                      is_same_v<EDataType, bhalf_t>)
         {
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<ELayout, Row>)
             {
-                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instances(
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_comp_default_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_comp_mnkpadding_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_comp_mnpadding_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_comp_kpadding_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v1_default_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v1_mnkpadding_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v1_mnpadding_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v1_kpadding_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v2_default_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v2_mnpadding_instances(
+                    op_ptrs);
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
                     op_ptrs);
             }
         }
@@ -132,7 +296,6 @@ struct DeviceOperationInstanceFactory<
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
 
-        // fp16_output
         if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
                      is_same_v<EDataType, bhalf_t>)
         {
@@ -199,7 +362,6 @@ struct DeviceOperationInstanceFactory<
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
 
-        // fp16_output
         if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
                      is_same_v<EDataType, bhalf_t>)
         {
@@ -266,7 +428,6 @@ struct DeviceOperationInstanceFactory<
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
 
-        // fp16_output
         if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
                      is_same_v<EDataType, bhalf_t>)
         {

library/include/ck/library/utility/fill.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -133,5 +133,40 @@ struct FillConstant
     }
 };
 
+template <typename T>
+struct TransformIntoStructuralSparsity
+{
+    // clang-format off
+    static constexpr T valid_sequences[] = {
+        0, 0, 1, 1,
+        0, 1, 0, 1,
+        0, 1, 1, 0,
+        1, 0, 0, 1,
+        1, 0, 1, 0,
+        1, 1, 0, 0,
+    };
+    // clang-format on
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::for_each(first, last, [=, idx = 0](T& elem) mutable {
+            auto tmp_idx = idx;
+            idx += 1;
+            return elem *= valid_sequences[tmp_idx % (sizeof(valid_sequences) / sizeof(T))];
+        });
+    }
+
+    template <typename ForwardRange>
+    auto operator()(ForwardRange&& range) const
+        -> std::void_t<decltype(std::declval<const TransformIntoStructuralSparsity&>()(
+            std::begin(std::forward<ForwardRange>(range)),
+            std::end(std::forward<ForwardRange>(range))))>
+    {
+        (*this)(std::begin(std::forward<ForwardRange>(range)),
+                std::end(std::forward<ForwardRange>(range)));
+    }
+};
+
 } // namespace utils
 } // namespace ck

library/include/ck/library/utility/host_tensor.hpp

@@ -43,7 +43,15 @@ std::ostream& LogRangeAsType(std::ostream& os, Range&& range, std::string delim)
             first = false;
         else
             os << delim;
-        os << static_cast<T>(v);
+
+        if constexpr(std::is_same_v<T, ck::f8_t> || std::is_same_v<T, ck::bf8_t>)
+        {
+            os << ck::type_convert<float>(v);
+        }
+        else
+        {
+            os << static_cast<T>(v);
+        }
     }
     return os;
 }

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -59,7 +59,7 @@ function(add_instance_library INSTANCE_NAME)
     endforeach()
     # Do not build WMMA instances if gfx11 targets are not on the target list
     foreach(source IN LISTS ARGN)
-        if(NOT INST_TARGETS MATCHES "gfx11" AND source MATCHES "_wmma")
+	if(NOT INST_TARGETS MATCHES "gfx11" AND NOT INST_TARGETS MATCHES "gfx12" AND source MATCHES "_wmma")
             message("removing wmma instance ${source} ")
             list(REMOVE_ITEM ARGN "${source}")
         endif()
@@ -177,7 +177,7 @@ FOREACH(subdir_path ${dir_list})
             message("Found only xdl instances, but gfx9 is not on the targets list. Skipping.")
             set(add_inst 0)
         endif()
-        if(("${cmake_instance}" MATCHES "ONLY WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11"))
+	if(("${cmake_instance}" MATCHES "ONLY WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12"))
             message("Found only wmma instances, but gfx11 is not on the targets list. Skipping.")
             set(add_inst 0)
         endif()
@@ -185,11 +185,11 @@ FOREACH(subdir_path ${dir_list})
             message("Found only xdl and dl instances, but gfx9 is not on the targets listand DL_KERNELS is not set. Skipping.")
             set(add_inst 0)
         endif()
-        if(("${cmake_instance}" MATCHES "ONLY XDL_AND_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx9"))
+	if(("${cmake_instance}" MATCHES "ONLY XDL_AND_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12") AND (NOT INST_TARGETS MATCHES "gfx9"))
             message("Found only xdl and wmma instances, but gfx11 and gfx9 are not on the targets list. Skipping.")
             set(add_inst 0)
         endif()
-        if(("${cmake_instance}" MATCHES "XDL_DL_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx9") AND (NOT DEFINED DL_KERNELS))
+	if(("${cmake_instance}" MATCHES "XDL_DL_WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12") AND (NOT INST_TARGETS MATCHES "gfx9") AND (NOT DEFINED DL_KERNELS))
             message("Found xdl, dl, and wmma instances, but none of those meet the target list. Skipping.")
             set(add_inst 0)
         endif()

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/CMakeLists.txt

+# ONLY XDL_KERNELS
+set(GEMM_UNIVERSAL_STREAMK_INSTANCES)
+
+list(APPEND GEMM_UNIVERSAL_STREAMK_INSTANCES 
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instance.cpp)
+
+add_instance_library(device_gemm_universal_streamk_instance ${GEMM_UNIVERSAL_STREAMK_INSTANCES})

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = half_t;
+using F32 = float;
+
+using Row = tensor_layout::gemm::RowMajor;
+using Col = tensor_layout::gemm::ColumnMajor;
+
+template <index_t... Is>
+using S = Sequence<Is...>;
+
+using PassThrough = element_wise::PassThrough;
+
+static constexpr auto GemmDefault    = GemmSpecialization::Default;
+static constexpr auto GemmKPadding   = GemmSpecialization::KPadding;
+static constexpr auto GemmMNPadding  = GemmSpecialization::MNPadding;
+static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
+
+static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
+static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
+
+template <GemmSpecialization GemmSpec>
+using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances = std::tuple<
+    // clang-format off
+        //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
+        //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
+        //#########################|        |        |        |     |      |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
+        //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+        
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   224,   256,    64,   8,   8,  16,   16,    7,    8,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,          0,          1,           2,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   256,    32,   8,   4,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   128,    32,   8,   4,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+    // clang-format on
+    >;
+
+template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
+using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_instances = std::tuple<
+    // clang-format off
+        //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|    Block-wiseGemm|               Block-wiseGemm|
+        //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|          Pipeline|                     Pipeline|
+        //#########################|        |        |        |     |      |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|         Scheduler|                     Verision|
+        //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
+
+        // Latency friendly
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   8,   4,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,   128,   8,   4,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   8,   4,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   8,   4,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        // Memory friendly
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   32,    64,   8,   2,  32,   32,    2,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,          0,          1,           1,                   S<1, 32, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   16,    64,   8,   2,  16,   16,    4,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,          0,          1,           1,                   S<1, 32, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,   128,   32,    64,   8,   4,  32,   32,    2,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,   128,   16,    64,   8,   4,  16,   16,    4,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    64,   32,    64,   8,   4,  32,   32,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    64,   16,    64,   8,   4,  16,   16,    2,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   8,   4,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,   128,   8,   4,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   8,   4,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   8,   4,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    64,   8,   4,  16,   16,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   64,    64,   8,   4,  32,   32,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,  128,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,  128,    64,   8,   4,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    32,  256,    64,   8,   4,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
+    // clang-format on
+    >;
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances<GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck