Merge branch 'develop' into layernorm/instance_support

include/ck_tile/ops/layernorm2d/kernel/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 {
+/*
+// clang-format off
+
+4-level descriptor: BlockTile-> WarpPerBlock-> WarpTile-> Vector
+
+                         Block_N (Warp_N * WarpPerBlock_N * Repeat_N )
+        +<----------------------< Repeat_N(2)>--------------------->+
+        |                                                           |
+        +<--    <WarpPerBlock_N(2)>  -->+
+            Warp_N
+        +--------------+--------------+--------------+--------------+----+----------------+
+ Warp_M | wrap_0       | wrap_1       |                             |    ^                ^
+        +--------------+--------------+                             |   <WarpPerBlock_M(2)> |
+        | wrap_2       | wrap_3       |                             |    v
+        +--------------+--------------+--------------+--------------+----+           Block_M
+        |                             |                             |
+        +                             +                             |
+        |                             |                             |                     v
+        +--------------+--------------+--------------+--------------+                     +
+
+        each Warp-tile (e.g 16 thrd per row)
+
+         Vector_N (contiguous pixels each thrd holds along N, or vector size)
+        +-----------+-----------+-----------+-----------+-----------+
+        | thrd_0    | thrd_1    | thrd_2    | thrd_3    | ...         Vector_M
+        +-----------+-----------+-----------+-----------+-----------+
+        | thrd_16   | thrd_17   | thrd_18   | thrd_19   | ...
+        +-----------+-----------+-----------+-----------+-----------+
+// clang-format on
+*/
+template <typename BlockTile_,    // block size, seq<M, N>
+          typename WarpPerBlock_, // num warps along seq<M, N>
+          typename WarpTile_,     // warp size, seq<M, N>
+          typename Vector_,       // contiguous pixels(vector size) along seq<M, N>
+          index_t BlockSize_ =
+              warpSize* reduce_on_sequence(WarpPerBlock_{}, multiplies{}, number<1>{})>
+struct Layernorm2dShape
+{
+    // block size
+    static constexpr index_t Block_M = BlockTile_::at(number<0>{});
+    static constexpr index_t Block_N = BlockTile_::at(number<1>{});
+
+    // num warps along seq<M, N>, within each block
+    static constexpr index_t WarpPerBlock_M = WarpPerBlock_::at(number<0>{});
+    static constexpr index_t WarpPerBlock_N = WarpPerBlock_::at(number<1>{});
+
+    // warp size
+    static constexpr index_t Warp_M = WarpTile_::at(number<0>{});
+    static constexpr index_t Warp_N = WarpTile_::at(number<1>{});
+
+    static_assert(Block_M % (WarpPerBlock_M * Warp_M) == 0);
+    static_assert(Block_N % (WarpPerBlock_N * Warp_N) == 0);
+    // repeat of each thread along seq<M, N>
+    static constexpr index_t Repeat_M = Block_M / (WarpPerBlock_M * Warp_M);
+    static constexpr index_t Repeat_N = Block_N / (WarpPerBlock_N * Warp_N);
+
+    // vector size along seq<M, N>
+    static constexpr index_t Vector_M = Vector_::at(number<0>{});
+    static constexpr index_t Vector_N = Vector_::at(number<1>{});
+
+    static_assert(Warp_M % Vector_M == 0);
+    static_assert(Warp_N % Vector_N == 0);
+    // num of threads along seq<M, N>, within each warp
+    static constexpr index_t ThreadPerWarp_M = Warp_M / Vector_M;
+    static constexpr index_t ThreadPerWarp_N = Warp_N / Vector_N;
+
+    static constexpr index_t BlockSize = BlockSize_;
+};
+
+} // namespace ck_tile

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/welford/block/block_welford_problem.hpp"
+#include "ck_tile/ops/welford/block/block_welford.hpp"
+
+namespace ck_tile {
+
+struct Layernorm2dFwdPipelineDefaultPolicy
+{
+    template <typename Problem>
+    CK_TILE_DEVICE static constexpr auto MakeXBlockTileDistribution()
+    {
+        using S = typename Problem::BlockShape;
+
+        return make_static_tile_distribution(
+            tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<S::Repeat_M, S::WarpPerBlock_M, S::ThreadPerWarp_M, S::Vector_M>,
+                      sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
+                tuple<sequence<1, 2>, sequence<1, 2>>,
+                tuple<sequence<1, 1>, sequence<2, 2>>,
+                sequence<1, 1, 2, 2>,
+                sequence<0, 3, 0, 3>>{});
+    }
+    template <typename Problem>
+    CK_TILE_DEVICE static constexpr auto MakeGammaBetaBlockTileDistribution()
+    {
+        using S = typename Problem::BlockShape;
+
+        return make_static_tile_distribution(
+            tile_distribution_encoding<
+                sequence<S::WarpPerBlock_M, S::ThreadPerWarp_M>,
+                tuple<sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
+                tuple<sequence<0, 1>, sequence<0, 1>>,
+                tuple<sequence<0, 1>, sequence<1, 2>>,
+                sequence<1, 1>,
+                sequence<0, 3>>{});
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelford()
+    {
+        using P_ = BlockWelfordProblem<typename Problem::XDataType,
+                                       typename Problem::ComputeDataType,
+                                       typename Problem::BlockShape>;
+
+        return BlockWelford<P_>{};
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelfordSync()
+    {
+        using P_ = BlockWelfordProblem<typename Problem::XDataType,
+                                       typename Problem::ComputeDataType,
+                                       typename Problem::BlockShape>;
+
+        return BlockWelfordSync<P_>{};
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockWelfordCrossWarpSync()
+    {
+        using P_ = BlockWelfordProblem<typename Problem::XDataType,
+                                       typename Problem::ComputeDataType,
+                                       typename Problem::BlockShape>;
+
+        return BlockWelfordCrossWarpSync<P_>{};
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        if constexpr(Problem::kNeedCrossWarpSync)
+        {
+            using P_ = BlockWelfordProblem<typename Problem::XDataType,
+                                           typename Problem::ComputeDataType,
+                                           typename Problem::BlockShape>;
+
+            using block_welford = BlockWelford<P_>;
+            using x_block_tile =
+                decltype(make_static_distributed_tensor<typename Problem::XDataType>(
+                    MakeXBlockTileDistribution<Problem>()));
+            using mean_var_block_tile =
+                decltype(block_welford::template MakeMeanVarBlockTile<x_block_tile>());
+
+            return GetBlockWelfordCrossWarpSync<Problem>()
+                .template GetSmemSize<mean_var_block_tile>();
+        }
+        else
+        {
+            return 1; // zero size arrays are an extension
+        }
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.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/layernorm2d/pipeline/layernorm2d_fwd_pipeline_default_policy.hpp"
+#include <string>
+#include <type_traits>
+
+namespace ck_tile {
+
+template <typename Problem_, typename Policy_ = Layernorm2dFwdPipelineDefaultPolicy>
+struct Layernorm2dFwdPipelineOnePass
+{
+    using Problem = ck_tile::remove_cvref_t<Problem_>;
+    using Policy  = ck_tile::remove_cvref_t<Policy_>;
+
+    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   = Problem::kSaveMeanInvStd;
+    static constexpr bool kSaveInvStd = Problem::kSaveMeanInvStd;
+
+    static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
+    static constexpr bool kPadM              = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
+    static constexpr bool kPadN              = Problem::kPadN;
+
+    static constexpr const char* name = []() {
+        if constexpr(kNeedCrossWarpSync)
+            return "bpr"; // block per row
+        else
+            return "wpr"; // warp per row
+    }();
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return Policy::template GetSmemSize<Problem>();
+    }
+
+    template <typename XWindow,
+              typename GammaWindow,
+              typename BetaWindow,
+              typename YWindow,
+              typename MeanWindow,
+              typename InvStdWindow>
+    CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
+                                   const GammaWindow& gamma_window_,
+                                   const BetaWindow& beta_window_,
+                                   YWindow& y_window,
+                                   MeanWindow& mean_window,
+                                   InvStdWindow& inv_std_window,
+                                   ComputeDataType epsilon,
+                                   ck_tile::index_t row_size,
+                                   void* smem) const
+    {
+        const auto x_window =
+            make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
+        const auto gamma_window = make_tile_window(
+            gamma_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
+        const auto beta_window = make_tile_window(
+            beta_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
+
+        const auto x  = load_tile(x_window);
+        int cur_count = 0;
+        int max_count =
+            block_tile_welford_calculate_max_count<typename Problem::BlockShape>(row_size);
+        auto block_welford      = Policy::template GetBlockWelford<Problem>();
+        auto block_welford_sync = Policy::template GetBlockWelfordSync<Problem>();
+        auto block_welford_cross_warp_sync =
+            Policy::template GetBlockWelfordCrossWarpSync<Problem>();
+
+        // load gamma/beta (TODO: support no gamma/beta?)
+        const auto gamma = load_tile(gamma_window);
+        const auto beta  = load_tile(beta_window);
+
+        // compute welford each-thread->cross-lane->cross-warp
+        auto [mean, var] = block_welford(x, cur_count, max_count);
+        block_welford_sync(mean, var, cur_count);
+        block_welford_cross_warp_sync(mean, var, cur_count, smem);
+        block_tile_welford_post_scale_var(var, cur_count);
+
+        // compute inv-std
+        auto inv_std = tile_elementwise_in(
+            [&](const auto& v_) {
+                return type_convert<ComputeDataType>(1.0f) / (sqrt(v_) + epsilon);
+            },
+            var);
+
+        if constexpr(kSaveMean)
+            store_tile(mean_window, cast_tile<MeanDataType>(mean));
+        if constexpr(kSaveInvStd)
+            store_tile(inv_std_window, cast_tile<InvStdDataType>(inv_std));
+
+        // layernorm computation
+        auto y = make_static_distributed_tensor<YDataType>(x.get_tile_distribution());
+        sweep_tile(y, [&, mean_ = mean](auto idx) {
+            constexpr auto i_idx = make_tuple(idx[number<0>{}]);
+            constexpr auto j_idx = make_tuple(idx[number<1>{}]);
+
+            const auto gamma_ = type_convert<ComputeDataType>(gamma[j_idx]);
+            const auto beta_  = type_convert<ComputeDataType>(beta[j_idx]);
+
+            const auto x_ = type_convert<ComputeDataType>(x[idx]);
+            auto y_       = (x_ - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
+
+            y(idx) = type_convert<YDataType>(y_);
+        });
+        store_tile(y_window, y);
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/layernorm2d/pipeline/block_layernorm2d_fwd_problem.hpp → include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_problem.hpp

@@ -15,20 +15,26 @@ template <typename XDataType_,
           typename MeanDataType_,
           typename InvStdDataType_,
           typename BlockShape_,
-          bool kPadM_,
-          bool kPadN_>
-struct BlockLayernorm2dFwdProblem
+          bool kPadN_,
+          bool kSaveMeanInvStd_,
+          bool kTwoPass_>
+struct Layernorm2dFwdPipelineProblem
 {
-    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_>;
-    static constexpr bool kPadM = kPadM_;
-    static constexpr bool kPadN = kPadN_;
+    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_>;
+
+    static constexpr bool kNeedCrossLaneSync = BlockShape::ThreadPerWarp_N > 1;
+    static constexpr bool kNeedCrossWarpSync = BlockShape::WarpPerBlock_N > 1;
+
+    static constexpr bool kPadN           = kPadN_;
+    static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
+    static constexpr bool kTwoPass        = kTwoPass_;
 };
 
 } // namespace ck_tile

include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_two_pass.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/layernorm2d/pipeline/layernorm2d_fwd_pipeline_default_policy.hpp"
+#include <string>
+#include <type_traits>
+
+namespace ck_tile {
+
+template <typename Problem_, typename Policy_ = Layernorm2dFwdPipelineDefaultPolicy>
+struct Layernorm2dFwdPipelineTwoPass
+{
+    using Problem = ck_tile::remove_cvref_t<Problem_>;
+    using Policy  = ck_tile::remove_cvref_t<Policy_>;
+
+    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   = Problem::kSaveMeanInvStd;
+    static constexpr bool kSaveInvStd = Problem::kSaveMeanInvStd;
+
+    static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
+    static constexpr bool kPadM              = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
+    static constexpr bool kPadN              = Problem::kPadN;
+
+    static constexpr const char* name = []() {
+        if constexpr(kNeedCrossWarpSync)
+            return "bpr"; // block per row
+        else
+            return "wpr"; // warp per row
+    }();
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return Policy::template GetSmemSize<Problem>();
+    }
+
+    template <typename XWindow,
+              typename GammaWindow,
+              typename BetaWindow,
+              typename YWindow,
+              typename MeanWindow,
+              typename InvStdWindow>
+    CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
+                                   const GammaWindow& gamma_window_,
+                                   const BetaWindow& beta_window_,
+                                   YWindow& y_window,
+                                   MeanWindow& mean_window,
+                                   InvStdWindow& inv_std_window,
+                                   ComputeDataType epsilon,
+                                   ck_tile::index_t row_size,
+                                   void* smem) const
+    {
+        auto x_window =
+            make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
+        auto gamma_window = make_tile_window(
+            gamma_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
+        auto beta_window = make_tile_window(
+            beta_window_, Policy::template MakeGammaBetaBlockTileDistribution<Problem>());
+
+        // Problem::BlockShape
+        static constexpr index_t Block_N = Problem::BlockShape::Block_N;
+        index_t num_n_tile_iteration =
+            __builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
+
+        // total number of count assume current iter have no pad(only last iter has pad)
+        constexpr index_t count_per_iter =
+            Problem::BlockShape::Repeat_N * Problem::BlockShape::Vector_N;
+        const index_t last_iter_n = row_size - (num_n_tile_iteration - 1) * Block_N;
+
+        int cur_count = 0;
+        int max_count =
+            (num_n_tile_iteration - 1) * count_per_iter +
+            block_tile_welford_calculate_max_count<typename Problem::BlockShape>(last_iter_n);
+        auto block_welford      = Policy::template GetBlockWelford<Problem>();
+        auto block_welford_sync = Policy::template GetBlockWelfordSync<Problem>();
+        auto block_welford_cross_warp_sync =
+            Policy::template GetBlockWelfordCrossWarpSync<Problem>();
+
+        using XTensorType = decltype(load_tile(x_window));
+        auto mean         = block_welford.template MakeMeanVarBlockTile<XTensorType>();
+        auto var          = block_welford.template MakeMeanVarBlockTile<XTensorType>();
+
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
+        {
+            const auto x = load_tile(x_window);
+            block_welford(x, mean, var, cur_count, max_count);
+            move_tile_window(x_window, {0, Block_N});
+        }
+
+        block_welford_sync(mean, var, cur_count);
+        block_welford_cross_warp_sync(mean, var, cur_count, smem);
+        block_tile_welford_post_scale_var(var, cur_count);
+
+        // compute inv-std
+        auto inv_std = tile_elementwise_in(
+            [&](const auto& v_) {
+                return type_convert<ComputeDataType>(1.0f) / (sqrt(v_) + epsilon);
+            },
+            var);
+
+        if constexpr(kSaveMean)
+            store_tile(mean_window, cast_tile<MeanDataType>(mean));
+        if constexpr(kSaveInvStd)
+            store_tile(inv_std_window, cast_tile<InvStdDataType>(inv_std));
+
+        // reverse read x to reuse cache
+        ck_tile::index_t stride_to_right_most_window =
+            row_size % Block_N == 0 ? row_size - Block_N : row_size - row_size % Block_N;
+
+        // x_window.foo();
+        // gamma_window.foo();
+        move_tile_window(x_window, {0, -Block_N});
+        move_tile_window(gamma_window, {stride_to_right_most_window});
+        move_tile_window(beta_window, {stride_to_right_most_window});
+        move_tile_window(y_window, {0, stride_to_right_most_window});
+
+        // layernorm computation
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
+        {
+            const auto x = load_tile(x_window);
+            // load gamma/beta (TODO: support no gamma/beta?)
+            const auto gamma = load_tile(gamma_window);
+            const auto beta  = load_tile(beta_window);
+
+            auto y = make_static_distributed_tensor<YDataType>(x.get_tile_distribution());
+
+            sweep_tile(y, [&, mean_ = mean](auto idx) {
+                constexpr auto i_idx = make_tuple(idx[number<0>{}]);
+                constexpr auto j_idx = make_tuple(idx[number<1>{}]);
+
+                const auto gamma_ = type_convert<ComputeDataType>(gamma[j_idx]);
+                const auto beta_  = type_convert<ComputeDataType>(beta[j_idx]);
+
+                const auto x_ = type_convert<ComputeDataType>(x[idx]);
+                auto y_       = (x_ - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
+
+                y(idx) = type_convert<YDataType>(y_);
+            });
+
+            store_tile(y_window, y);
+
+            move_tile_window(x_window, {0, -Block_N});
+            move_tile_window(gamma_window, {-Block_N});
+            move_tile_window(beta_window, {-Block_N});
+            move_tile_window(y_window, {0, -Block_N});
+        }
+    }
+};
+} // 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/reduce/block/block_reduce.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
 

include/ck_tile/ops/welford.hpp

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#include "ck_tile/ops/welford/block/block_welford.hpp"
+#include "ck_tile/ops/welford/block/block_welford_problem.hpp"
 #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/block/block_welford.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/welford/thread/thread_welford.hpp"
+
+namespace ck_tile {
+
+template <typename Problem_, typename Policy_ = void>
+struct BlockWelford
+{
+    using Problem         = remove_cvref_t<Problem_>;
+    using XDataType       = typename Problem::XDataType;
+    using ComputeDataType = typename Problem::ComputeDataType;
+
+    CK_TILE_DEVICE constexpr BlockWelford() {}
+
+    // [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_
+    // -> use block_welford_calculate_max_count to compute
+    template <typename XDistributedTensor_,
+              typename MeanDistributedTensor_,
+              typename VarDistributedTensor_>
+    CK_TILE_DEVICE void operator()(const XDistributedTensor_& x_tensor,
+                                   MeanDistributedTensor_& mean_tensor,
+                                   VarDistributedTensor_& var_tensor,
+                                   int& cur_count_, // -> prefer init as zero
+                                   const int& max_count_)
+    {
+        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]);
+
+                    welford_update(
+                        mean_tensor(out_dstr_idx), var_tensor(out_dstr_idx), x, cur_count_);
+                });
+            }
+        });
+    }
+
+    template <typename XDistributedTensor_>
+    CK_TILE_DEVICE static auto MakeMeanVarBlockTile()
+    {
+        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);
+
+        return tensor;
+    }
+
+    template <typename XDistributedTensor_>
+    CK_TILE_DEVICE auto
+    operator()(const XDistributedTensor_& x_tensor, int& cur_count_, const int& max_count_)
+    {
+        auto mean_tensor = MakeMeanVarBlockTile<XDistributedTensor_>();
+        auto var_tensor  = MakeMeanVarBlockTile<XDistributedTensor_>();
+        clear_tile(mean_tensor);
+        clear_tile(var_tensor);
+
+        (*this)(x_tensor, mean_tensor, var_tensor, cur_count_, max_count_);
+
+        return ck_tile::make_tuple(mean_tensor, var_tensor);
+    }
+};
+
+template <typename Problem_, typename Policy_ = void>
+struct BlockWelfordSync
+{
+    using Problem = remove_cvref_t<Problem_>;
+
+    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) {
+                        // xor
+                        index_t src_lane =
+                            (__lane_id()) ^
+                            (number<lid_over_rid_derivative << istage.value>{}.value);
+
+                        // pull data from remote lane
+                        const auto v_remote_mean  = warp_shuffle(v_local_mean, src_lane);
+                        const auto v_remote_var   = warp_shuffle(v_local_var, src_lane);
+                        const auto v_remote_count = warp_shuffle(v_local_count, src_lane);
+
+                        // welford merge
+                        welford_merge(v_local_mean,
+                                      v_local_var,
+                                      v_local_count,
+                                      v_remote_mean,
+                                      v_remote_var,
+                                      v_remote_count);
+                    });
+                }
+            });
+
+            mean_tensor.get_thread_buffer()(i) = v_local_mean;
+            var_tensor.get_thread_buffer()(i)  = v_local_var;
+
+            count = v_local_count;
+        });
+    }
+};
+
+template <typename Problem_, typename Policy_ = void>
+struct BlockWelfordCrossWarpSync
+{
+    using Problem    = remove_cvref_t<Problem_>;
+    using BlockShape = typename Problem::BlockShape;
+
+    template <typename MeanDistributedTensor_>
+    CK_TILE_DEVICE static constexpr index_t GetReduceWarps()
+    {
+        constexpr index_t num_reduce_warps = [&]() {
+            using Dstr             = typename MeanDistributedTensor_::StaticTileDistribution;
+            using DstrEncode       = typename Dstr::DstrEncode;
+            using DstrEncodeDetail = typename DstrEncode::detail;
+
+            constexpr index_t NDimR = Dstr::get_num_of_dimension_r();
+
+            constexpr index_t idim_p_warp = 0;
+
+            index_t len_ = 1;
+            static_for<0, NDimR, 1>{}([&](auto idim_r) {
+                if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_warp][idim_r])
+                {
+                    constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
+                    len_ *= r_length;
+                }
+            });
+            return len_;
+        }();
+        return num_reduce_warps;
+    }
+
+    // return in byte
+    template <typename MeanDistributedTensor_>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        // constexpr auto num_reduce_warps = GetReduceWarps<MeanDistributedTensor_>();
+
+        // data need to exchange is very small, we just pack mean+var+count -> 4dword
+        constexpr index_t thread_buf_size = MeanDistributedTensor_::get_thread_buffer_size();
+
+        // we need to store all data from every wave into smem
+        // e.g. 2x2 reduce along N
+        //     -------------> reduce N
+        //    | w0 | w1 |   ___>      | w01 |
+        //    | w2 | w3 |             | w23 |
+        //
+        //   -> store data from every wave into LDS
+        //
+        //
+        //     -------------> reduce N
+        //    | w0 | w1 | w2 | w3 |   ----->  | w0123 |
+        //
+        //   -> also store data from every wave into LDS
+        constexpr index_t num_warps = BlockShape::BlockSize / warpSize;
+        return num_warps * 4 * thread_buf_size * sizeof(float);
+    }
+
+    template <typename MeanDistributedTensor_, typename VarDistributedTensor_>
+    CK_TILE_DEVICE void operator()(MeanDistributedTensor_& mean_tensor,
+                                   VarDistributedTensor_& var_tensor,
+                                   int& count,
+                                   void* smem)
+    {
+        using DataType = typename MeanDistributedTensor_::DataType;
+        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 thread_buf_size = MeanDistributedTensor_::get_thread_buffer_size();
+        static_assert(thread_buf_size == VarDistributedTensor_::get_thread_buffer_size());
+
+        // Note: we always pack everything into fp32x4
+        fp32x4_t* smem_ptr              = reinterpret_cast<fp32x4_t*>(smem);
+        const index_t lane_id           = get_lane_id();
+        const index_t warp_id           = get_warp_id();
+        constexpr auto num_reduce_warps = GetReduceWarps<MeanDistributedTensor_>();
+        constexpr index_t num_warps     = BlockShape::BlockSize / warpSize;
+        const index_t smem_offset       = warp_id;
+
+        // skip if nonthing to do
+        if constexpr(num_reduce_warps == 1)
+            return;
+
+        // store into smem only for lane-0 within one warp
+        if(lane_id == 0)
+        {
+            static_for<0, thread_buf_size, 1>{}([&](auto i) {
+                fp32x4_t local_scratch_;
+                local_scratch_[0] = bit_cast<float>(mean_tensor.get_thread_buffer()[i]);
+                local_scratch_[1] = bit_cast<float>(var_tensor.get_thread_buffer()[i]);
+                local_scratch_[2] = bit_cast<float>(count);
+
+                smem_ptr[smem_offset + i * num_warps] = local_scratch_;
+            });
+        }
+        block_sync_lds();
+
+        // load from smem. here we let everythread to do compute :)
+        index_t local_warp_id = warp_id / num_reduce_warps;
+        index_t local_smem_os = local_warp_id * num_reduce_warps;
+        fp32x4_t all_scratch[thread_buf_size * num_reduce_warps];
+        static_for<0, thread_buf_size, 1>{}([&](auto i_0) {
+            static_for<0, num_reduce_warps, 1>{}([&](auto i_1) {
+                all_scratch[i_0 * num_warps + i_1] =
+                    smem_ptr[i_0 * num_reduce_warps + local_smem_os + i_1];
+            });
+        });
+        block_sync_lds(); // TODO: we don't need sync here
+
+        // const int original_count = count;
+
+        static_for<0, thread_buf_size, 1>{}([&](auto i_0) {
+            // TODO: use descriptor for this
+            auto v_local       = all_scratch[i_0 * num_warps];
+            auto v_local_mean  = bit_cast<DataType>(v_local[0]);
+            auto v_local_var   = bit_cast<DataType>(v_local[1]);
+            auto v_local_count = bit_cast<int>(v_local[2]);
+
+            // further reduce mean/var
+            static_for<0, num_reduce_warps - 1, 1>{}([&](auto i_1_n1) {
+                constexpr auto i_1        = number<i_1_n1 + 1>{};
+                const fp32x4_t v_remote   = all_scratch[i_0 * num_warps + i_1];
+                const auto v_remote_mean  = bit_cast<DataType>(v_remote[0]);
+                const auto v_remote_var   = bit_cast<DataType>(v_remote[1]);
+                const auto v_remote_count = bit_cast<int>(v_remote[2]);
+
+                welford_merge(v_local_mean,
+                              v_local_var,
+                              v_local_count,
+                              v_remote_mean,
+                              v_remote_var,
+                              v_remote_count);
+            });
+
+            mean_tensor.get_thread_buffer()(i_0) = v_local_mean;
+            var_tensor.get_thread_buffer()(i_0)  = v_local_var;
+
+            count = v_local_count;
+        });
+    }
+};
+
+// compute the max count for a last dim reduce
+// everything may have vector/repeat, so the max count could be uneven
+// TODO: specify which dim to compute and proper set the problem
+// TODO: BlockShape we reuse layernorm_fwd_shape :)
+template <typename BlockShape>
+CK_TILE_DEVICE constexpr index_t block_tile_welford_calculate_max_count(int row_size)
+{
+#if 0
+    using S                   = BlockShape;
+    index_t LastloopN         = row_size % S::Block_N == 0 ? S::Block_N : row_size % S::Block_N;
+    constexpr index_t NThread = S::WarpPerBlock_N * S::ThreadPerWarp_N;
+    index_t iNLane            = get_thread_id() % NThread;
+    index_t iN0               = LastloopN / (S::Vector_N * S::ThreadPerWarp_N);
+    index_t iN1               = (LastloopN % (S::Vector_N * S::ThreadPerWarp_N)) / S::Vector_N;
+    index_t N2                = (LastloopN % (S::Vector_N * S::ThreadPerWarp_N)) % S::Vector_N;
+    index_t iN3               = iNLane < iN1 ? S::Vector_N : iNLane == iN1 ? N2 : 0;
+    return iN0 * S::Vector_N + iN3;
+#endif
+    using S_                            = BlockShape;
+    constexpr index_t ThreadsPerBlock_N = S_::WarpPerBlock_N * S_::ThreadPerWarp_N;
+
+    // TODO: we always check vector size, need be evenly devidable by vector-n
+    const index_t element_per_row = row_size / S_::Vector_N;
+    index_t lane_id_n             = get_thread_id() % ThreadsPerBlock_N;
+
+    index_t cnt = 0;
+    // TODO: Repeat_N can not be too long, otherwise this is not good
+    static_for<0, S_::Repeat_N, 1>{}([&](auto) {
+        index_t _a = lane_id_n < element_per_row ? 1 : 0;
+        cnt += _a;
+        lane_id_n += ThreadsPerBlock_N;
+    });
+    return cnt * S_::Vector_N;
+}
+
+// Note: this function must be called after all the computation
+template <typename VarDistributedTensor_>
+CK_TILE_DEVICE constexpr void block_tile_welford_post_scale_var(VarDistributedTensor_& var_tensor,
+                                                                int count)
+{
+    using DataType = typename VarDistributedTensor_::DataType;
+    tile_elementwise_inout([&count](auto& x) { x = x / type_convert<DataType>(count); },
+                           var_tensor);
+}
+} // namespace ck_tile

include/ck_tile/ops/welford/block/block_welford_problem.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 XDataType_, typename ComputeDataType_, typename BlockShape_>
+struct BlockWelfordProblem
+{
+    using XDataType       = remove_cvref_t<XDataType_>;
+    using ComputeDataType = remove_cvref_t<ComputeDataType_>;
+    using BlockShape      = remove_cvref_t<BlockShape_>;
+};
+
+} // namespace ck_tile

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

@@ -7,95 +7,30 @@
 
 namespace ck_tile {
 
-template <typename ComputeDataType_, typename XDataType_>
-struct ThreadWelford
+template <typename T>
+CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count)
 {
-    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_;
-};
+    // TODO: check nan? maybe no
+    T delta = x - mean;
+    mean += delta / count;
+    T delta2 = x - mean;
+    var += delta * delta2;
+}
+
+template <typename T>
+CK_TILE_DEVICE static void
+welford_merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
+{
+    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;
+}
 
 } // 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/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_km_kn_mn.hpp

@@ -44,8 +44,11 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_km_kn_mn_comp_instances = std::tu
         
         DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              8,              4,          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,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   4,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,               1,              8,              4,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   2,   2,  32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              4,              2,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              8,              4,          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,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              4,              2,          0,    S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              8,              4,          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,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              4,              2,          0,    S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         // Can we support this kind of odd case? 224(256) = 28*8 + (4*8)
         //DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   224,   256,    64,   8,   8,  16,   16,    7,    8,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   4,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,               1,              8,              4,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
@@ -64,10 +67,13 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_km_kn_mn_mem_instances = std::tup
 
         // Latency friendly
         DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   4,   4,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              4,              4,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   2,   2,  16,   16,    1,    1,     S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              2,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   4,   4,  16,   16,    1,    1,     S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              4,              4,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   4,   4,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              4,          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>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   2,   2,  16,   16,    1,    1,     S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              2,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         // Memory friendly
         DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   16,    64,   8,   2,  16,   16,    4,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   16,    64,   2,   2,  16,   16,    4,    1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              2,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,   128,   16,    64,   8,   4,  16,   16,    4,    1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    64,   16,    64,   4,   4,  16,   16,    2,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              8,              4,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   4,   4,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              4,              4,          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>,
@@ -75,7 +81,8 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_km_kn_mn_mem_instances = std::tup
         DeviceGemm_Xdl_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   4,   4,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              4,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    64,   4,   4,  16,   16,    1,    2,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              4,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,  128,    64,   4,   4,  16,   16,    1,    4,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              4,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   2,   4,  16,   16,    1,    4,     S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              2,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   2,   4,  16,   16,    1,    4,     S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              2,          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_CShuffleV3<  Col,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   2,   2,  16,   16,    1,    4,     S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,               1,              2,              2,          0,    S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_km_nk_mn.hpp

@@ -44,13 +44,21 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_km_nk_mn_comp_instances = std::tu
         
         // Compute friendly
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   8,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   8,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   8,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   224,    64,   8,   8,  16,   16,    8,    7,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          2,           1,                   S<1, 32, 1, 8>,                4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
-        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   4,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   2,   2,  32,   32,    2,    2,     S<32, 8, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
     // clang-format on
     >;
 
@@ -64,18 +72,23 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_km_nk_mn_mem_instances = std::tup
 
         // Latency friendly 
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   4,   8,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   4,   4,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   2,   2,  16,   16,    1,    1,     S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   4,   8,  16,   16,    1,    1,     S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   4,   8,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         // Memory friendly
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   16,    64,   8,   8,  16,   16,    4,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,   128,   16,    64,   8,   8,  16,   16,    4,    1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    64,   16,    64,   4,   8,  16,   16,    2,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    64,   16,    64,   4,   4,  16,   16,    2,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   4,   8,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   4,   8,  16,   16,    1,    1,     S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   4,   8,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    64,   4,   8,  16,   16,    1,    2,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    64,   4,   4,  16,   16,    1,    2,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,  128,    64,   4,   8,  16,   16,    1,    4,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
-        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   2,   8,  16,   16,    1,    4,     S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   2,   8,  16,   16,    1,    4,     S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   2,   2,  16,   16,    1,    4,     S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn.hpp

@@ -43,6 +43,8 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_instances = std::tu
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
         
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          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,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,    S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
@@ -63,14 +65,18 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_instances = std::tup
 
         // Latency friendly
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   4,   4,  16,   16,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   2,   2,  16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   16,    64,   2,   2,  16,   16,    4,    1,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   4,   4,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      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>,

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp

@@ -44,6 +44,8 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_instances = std::tu
         
         // Compute friendly
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+        DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,    S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         // AGPR Spill
@@ -69,8 +71,12 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_instances = std::tup
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   8,   8,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   8,   8,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   8,   8,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   4,   4,  16,   16,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   2,   2,  16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,    S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         // Memory friendly
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   16,    64,   8,   8,  16,   16,    4,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   16,    64,   4,   4,  16,   16,    4,    1,     S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   16,    64,   2,   2,  16,   16,    4,    1,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,    S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 32, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,   128,   16,    64,   8,   8,  16,   16,    4,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    64,   16,    64,   8,   8,  16,   16,    2,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   8,   8,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn.hpp

@@ -46,7 +46,7 @@ using device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_instances = std::tuple<
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,  F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   192,   256,    64,    16,   8,  32,   32,    3,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,  F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,   128,    16,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,  F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,    16,   4,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,          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_CShuffleV3<  Row,     Row,     Row,     F8,  F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,    16,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,          0,    S<8,  32, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,        
+        // DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,  F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,    16,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,          0,    S<8,  32, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,        
         // We prefer following instance, however, existing compiler bug cause it failed to generate sanity code.
         // DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,  F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,    16,   4,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,          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_CShuffleV3<  Row,     Row,     Row,     F8,  F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,    16,   4,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,          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>