Grouped Gemm with looping over the tiles. (#788)

* Introduce LocalBlockToCTileMap.

* Change the signature of CalculateBottomIndex() function which now does
not accept any argument. The B2C map which is already passed as an
argument to the kernel Run function is calculating block's local id
already outside at kernel entry point __global__ function.
The LocalB2C map stores as members local block ID.

* Use LocalBlockToCTile map in device ops.

* First draft of tile loop work distribution.

* Fix typo.

* Simplify kernel arguments.

Calculate descriptors & B2C maps on the device.

* Use looping kernel.

* Fix B2C constructor.

* Fix Navi21 errors.

* Calculate tile start/end in device kernel.

* Change Run API to accept user provided workspace buffer.

* Add new line at EOF.

* Move Gemm KernelArguments to device op interface.

* Remove unused code.

* Update API.

* Launch grid size which is min of occupancy vs tile count

* Get back to use constant memory for gemm descriptors.

* Remove unused code.

* Add default virtual method implementation.

* Update comments to conform with doxygen style.

* Fix doc style and unused parameters.

* Add thread cluster lengths to kernel name.

* Remove old splitk impl and replace it with tile looping one.

* Modify instances.

* set KPerBlock to 64
* maximize wherever possible vector load size.

* Fix instances cluster lengths.

* Change comment style.

* Use 128b store where possible in instances.

* Update test cases, since KPerBlock has doubled.

* Update output stream operator for Sequence.

* Add pipeline version to GroupedGEMM device op type string.

* Fix pipeline version type logging.

* Fix input tensors type after merge.

* Fix compiler error.

* Fix output stream operator for Pipeline version.

* Store using 128b.

* Set of instances with kpb 32/64

* Limit number of instances

* Remove commented out instances.

* Fix function name.

* Limit the number of instances.

Add pipline version to the regular instances

* Change thr cluster layout for reading B tensor.

* disabled failed instances

---------
Co-authored-by: Adam Osewski <aosewski@amd.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>
Co-authored-by: Jing Zhang <jizha@amd.com>

include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp

@@ -8,6 +8,57 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
+///
+/// @brief      Structure representing single GEMM problem arguments.
+///
+///             The pointer to the vector of those structures is passed
+///             to the GroupedGEMM entry point kernel.
+///
+struct GroupedGemmKernelArguments
+{
+    __host__ __device__ GroupedGemmKernelArguments(const void* p_a_grid_,
+                                                   const void* p_b_grid_,
+                                                   void* p_c_grid_,
+                                                   index_t M_,
+                                                   index_t N_,
+                                                   index_t K_,
+                                                   index_t StrideA_,
+                                                   index_t StrideB_,
+                                                   index_t StrideC_)
+        : p_a_grid{p_a_grid_},
+          p_b_grid{p_b_grid_},
+          p_c_grid{p_c_grid_},
+          M{M_},
+          N{N_},
+          K{K_},
+          StrideA{StrideA_},
+          StrideB{StrideB_},
+          StrideC{StrideC_}
+    {
+    }
+
+    const void* p_a_grid;
+    const void* p_b_grid;
+    void* p_c_grid;
+    index_t M;
+    index_t N;
+    index_t K;
+    index_t StrideA;
+    index_t StrideB;
+    index_t StrideC;
+
+    void Print() const
+    {
+        std::cout << "arg {"
+                  << "M:" << M << ", "
+                  << "N:" << N << ", "
+                  << "K:" << K << ", "
+                  << "SA:" << StrideA << ", "
+                  << "SB:" << StrideB << ", "
+                  << "SC:" << StrideC << "}" << std::endl;
+    }
+};
+
 template <typename ALayout,
           typename BLayout,
           typename DsLayout,
@@ -31,7 +82,28 @@ struct DeviceGroupedGemmSplitK : public DeviceGroupedGemm<ALayout,
                                                           BElementwiseOperation,
                                                           CElementwiseOperation>
 {
-    virtual void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const = 0;
+    //----------------------------------------------------------------------------------------------
+    /// @brief      Sets the k batch size.
+    ///
+    /// @param      p_arg   Pointer to the Argument we're going to change.
+    /// @param[in]  kbatch  The kbatch value.
+    ///
+    virtual void SetKBatchSize([[maybe_unused]] BaseArgument* p_arg,
+                               [[maybe_unused]] index_t kbatch) const
+    {
+    }
+
+    //----------------------------------------------------------------------------------------------
+    /// @brief      Sets the device kernel arguments pointer.
+    ///
+    /// @param      p_arg              The pointer to the Argument we're going to update.
+    /// @param[in]  p_dev_kernel_args  The pointer to the device memory which contains kernel
+    ///                                arguments.
+    ///
+    virtual void SetDeviceKernelArgs([[maybe_unused]] BaseArgument* p_arg,
+                                     [[maybe_unused]] const void* p_dev_kernel_args) const
+    {
+    }
 };
 
 } // namespace device

include/ck/tensor_operation/gpu/device/device_softmax.hpp

@@ -22,22 +22,22 @@ template <typename InDataType,
           index_t NumReduceDim>
 struct DeviceSoftmax : public BaseOperator
 {
-    //
-    // @brief      Makes a pointer to Argument class.
-    //
-    // @param[in]  inLengths           Input tensor extent(s) from high to low dimension
-    // @param[in]  inStrides           Input tensor stride(s) from high to low dimension
-    // @param[in]  reduceDims          The dimension(s) the normalization operation is applied
-    // @param[in]  alpha               double type value
-    // @param[in]  beta                double type value
-    // @param[in]  in_dev              Typeless const pointer in device memory storing the input
-    //                                 tensor
-    // @param      out_dev             Typeless pointer in device memory storing the output tensor
-    // @param[in]  in_elementwise_op   The input elementwise operation.
-    // @param[in]  acc_elementwise_op  The accumulation elementwise operation.
-    //
-    // @return     Unique pointer to the Argument class.
-    //
+    ///
+    /// @brief      Makes a pointer to Argument class.
+    ///
+    /// @param[in]  inLengths           Input tensor extent(s) from high to low dimension
+    /// @param[in]  inStrides           Input tensor stride(s) from high to low dimension
+    /// @param[in]  reduceDims          The dimension(s) the normalization operation is applied
+    /// @param[in]  alpha               double type value
+    /// @param[in]  beta                double type value
+    /// @param[in]  in_dev              Typeless const pointer in device memory storing the input
+    ///                                 tensor
+    /// @param      out_dev             Typeless pointer in device memory storing the output tensor
+    /// @param[in]  in_elementwise_op   The input elementwise operation.
+    /// @param[in]  acc_elementwise_op  The accumulation elementwise operation.
+    ///
+    /// @return     Unique pointer to the Argument class.
+    ///
     virtual std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const std::vector<index_t> inLengths,
                         const std::vector<index_t> inStrides,

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp

@@ -168,7 +168,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                                                      stream_config.stream_id_));
 
                 ave_time = launch_and_time_kernel(
-                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg, b2c_map);
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
             };
 
             if(has_main_k0_block_loop)

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp

@@ -157,22 +157,22 @@ __global__ void
 }
 } // namespace
 
-//
-// @brief      Device Convolution operation.
-//
-// Supports:
-//  @li         Forward convolution with up to 3 spatial dimentions
-//  @li         Input tensor in GNWC data format
-//  @li         Weight tensor in GKXC data format
-//  @li         Output tensor in GNWK data format
-//
-// 1D:
-// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
-// 2D:
-// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
-// 3D:
-// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
-//
+///
+/// @brief      Device Convolution operation.
+///
+/// Supports:
+///  @li         Forward convolution with up to 3 spatial dimentions
+///  @li         Input tensor in GNWC data format
+///  @li         Weight tensor in GKXC data format
+///  @li         Output tensor in GNWK data format
+///
+/// 1D:
+/// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+/// 2D:
+/// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+/// 3D:
+/// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+///
 template <index_t NDimSpatial,
           typename ADataType,
           typename BDataType,

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp

@@ -154,22 +154,22 @@ __global__ void
 
 } // namespace
 
-//
-// @brief      Device Convolution operation.
-//
-// Supports:
-//  @li         Forward convolution with up to 3 spatial dimentions
-//  @li         Input tensor in GNWC data format
-//  @li         Weight tensor in GKXC data format
-//  @li         Output tensor in GNWK data format
-//
-// 1D:
-// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
-// 2D:
-// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
-// 3D:
-// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
-//
+///
+/// @brief      Device Convolution operation.
+///
+/// Supports:
+///  @li         Forward convolution with up to 3 spatial dimentions
+///  @li         Input tensor in GNWC data format
+///  @li         Weight tensor in GKXC data format
+///  @li         Output tensor in GNWK data format
+///
+/// 1D:
+/// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+/// 2D:
+/// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+/// 3D:
+/// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+///
 template <
     index_t NDimSpatial,
     typename ADataType,

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp

@@ -150,22 +150,22 @@ __global__ void
 
 } // namespace
 
-//
-// @brief      Device Convolution operation.
-//
-// Supports:
-//  @li         Forward convolution with up to 3 spatial dimentions
-//  @li         Input tensor in GNWC data format
-//  @li         Weight tensor in GKXC data format
-//  @li         Output tensor in GNWK data format
-//
-// 1D:
-// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
-// 2D:
-// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
-// 3D:
-// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
-//
+///
+/// @brief      Device Convolution operation.
+///
+/// Supports:
+///  @li         Forward convolution with up to 3 spatial dimentions
+///  @li         Input tensor in GNWC data format
+///  @li         Weight tensor in GKXC data format
+///  @li         Output tensor in GNWK data format
+///
+/// 1D:
+/// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+/// 2D:
+/// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+/// 3D:
+/// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+///
 template <index_t NDimSpatial,
           typename ALayout,
           typename BLayout,

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp

@@ -5,11 +5,13 @@
 
 #include <iostream>
 #include <sstream>
+#include <tuple>
 
 #include "ck/ck.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 #include "ck/host_utility/hip_check_error.hpp"
+#include "ck/host_utility/stream_utility.hpp"
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/tuple.hpp"
 #include "ck/tensor_description/tensor_descriptor.hpp"
@@ -23,8 +25,28 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
+///
+/// @brief      Entry point kernel for device-wide Grouped GEMM operation.
+///
+/// @param[in]  gemm_descs_const            The pointer to the array of GEMM descriptor structures.
+/// @param[in]  tile_count                  The overall number of output tiles we divided all groups
+///                                         into.
+/// @param[in]  k_batch                     The number of batches we split the K dimension into.
+///
+/// @tparam     GridwiseGemm                The specific GridwiseGEMM algorithm implementation.
+/// @tparam     GemmDesc                    The structure holding all necessary descriptors and
+///                                         other data needed for groupd gemm calculation and work
+///                                         distribution.
+/// @tparam     HasMainKBlockLoop           Flag indicating whether all GEMM problem configurations
+///                                         need to loop over tiles in K dimension.
+/// @tparam     CGlobalMemoryDataOperation  The functor used to store data in output C matrix.
+///                                         In example could be: AtomicAdd or Store.
+///
 template <typename GridwiseGemm,
           typename GemmDesc,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
           bool HasMainKBlockLoop,
           InMemoryDataOperationEnum CGlobalMemoryDataOperation>
 __global__ void
@@ -32,42 +54,99 @@ __global__ void
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
         kernel_grouped_gemm_xdl_splitk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
-                                       const index_t group_count)
+                                       const index_t tile_count,
+                                       const index_t k_batch)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
     defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+
     constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
     __shared__ uint8_t p_shared[shared_size];
 
-    const index_t block_id = get_block_1d_id();
+    index_t tile_id         = get_block_1d_id();
+    const index_t grid_size = get_grid_size();
     const auto gemm_desc_ptr =
         reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
 
-    index_t left     = 0;
-    index_t right    = group_count;
-    index_t group_id = index_t((left + right) / 2);
-    while((!(block_id >= gemm_desc_ptr[group_id].block_start_ &&
-             block_id < gemm_desc_ptr[group_id].block_end_)) &&
-          left <= right)
+    static constexpr index_t MPerBlock = GridwiseGemm::GetMPerBlock();
+    static constexpr index_t NPerBlock = GridwiseGemm::GetNPerBlock();
+    static constexpr index_t B2E_M01   = 8;
+
+    using CGridDesc_M_N = typename GridwiseGemm::CGridDesc_M_N;
+    using Block2ETileMapKSplit =
+        BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>;
+
+    index_t group_id = 0;
+    index_t offset   = 0;
+
+    auto M                = gemm_desc_ptr[group_id].M;
+    auto N                = gemm_desc_ptr[group_id].N;
+    auto StrideC          = gemm_desc_ptr[group_id].StrideC;
+    auto c_grid_desc_m_n  = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, StrideC);
+    auto b2c_tile_map     = Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, k_batch};
+    index_t grid_size_grp = b2c_tile_map.CalculateGridSize(c_grid_desc_m_n);
+
+    index_t gemm_tile_id_start = 0;
+    index_t gemm_tile_id_end   = grid_size_grp;
+
+    while(tile_id < tile_count)
     {
-        if(block_id < gemm_desc_ptr[group_id].block_start_)
+        // Find corresponding GEMM group for out tile
+        while(!(tile_id >= gemm_tile_id_start && tile_id < gemm_tile_id_end))
         {
-            right = group_id;
-        }
-        else
-        {
-            left = group_id;
-        }
-        group_id = index_t((left + right) / 2);
+            offset += grid_size_grp;
+            group_id++;
+
+            M               = gemm_desc_ptr[group_id].M;
+            N               = gemm_desc_ptr[group_id].N;
+            StrideC         = gemm_desc_ptr[group_id].StrideC;
+            c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, StrideC);
+            b2c_tile_map    = Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, k_batch};
+            grid_size_grp   = b2c_tile_map.CalculateGridSize(c_grid_desc_m_n);
+
+            gemm_tile_id_start = offset;
+            gemm_tile_id_end   = offset + grid_size_grp;
         }
 
+        const auto p_a_grid = reinterpret_cast<const FloatA*>(gemm_desc_ptr[group_id].p_a_grid);
+        const auto p_b_grid = reinterpret_cast<const FloatB*>(gemm_desc_ptr[group_id].p_b_grid);
+        const auto p_c_grid = reinterpret_cast<FloatC*>(gemm_desc_ptr[group_id].p_c_grid);
+
+        const auto K       = gemm_desc_ptr[group_id].K;
+        const auto StrideA = gemm_desc_ptr[group_id].StrideA;
+        const auto StrideB = gemm_desc_ptr[group_id].StrideB;
+
+        const auto MPadded = GridwiseGemm::CalculateMPadded(M);
+        const auto NPadded = GridwiseGemm::CalculateNPadded(N);
+        const auto KPadded = GridwiseGemm::CalculateKPadded(K, k_batch);
+        const auto K0      = GridwiseGemm::CalculateK0(K, k_batch);
+
+        LocalBlockToCTileMap<Block2ETileMapKSplit> local_b2c{b2c_tile_map, tile_id - offset};
         GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
-        gemm_desc_ptr[group_id].karg_,
+            p_a_grid,
+            p_b_grid,
+            p_c_grid,
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            StrideC,
+            MPadded,
+            NPadded,
+            KPadded,
+            K0,
+            k_batch,
             static_cast<void*>(p_shared),
-        gemm_desc_ptr[group_id].block_2_ctile_map_);
+            local_b2c);
+
+        tile_id += grid_size;
+    }
+
 #else
     ignore = gemm_descs_const;
-    ignore = group_count;
+    ignore = tile_count;
+    ignore = k_batch;
 #endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
@@ -187,32 +266,12 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
         PipelineVer>;
 
     using CGridDesc_M_N   = typename GridwiseGemm::CGridDesc_M_N;
+    using GridwiseGemmArg = typename GridwiseGemm::Argument;
+    using KernelArguments = GroupedGemmKernelArguments;
     using Block2ETileMapKSplit =
         BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>;
     // Block2CTileMap configuration parameter.
     static constexpr index_t B2E_M01       = 8;
-    using GroupedGemmBlock2ETileMap  = OffsettedBlockToCTileMap<Block2ETileMapKSplit>;
-    using KernelArgument             = typename GridwiseGemm::Argument;
-
-    struct GemmTransKernelArg
-    {
-        KernelArgument karg_;
-        GroupedGemmBlock2ETileMap block_2_ctile_map_;
-        index_t block_start_, block_end_;
-
-        GemmTransKernelArg() = default;
-        GemmTransKernelArg(KernelArgument&& karg,
-                           GroupedGemmBlock2ETileMap&& b2c_map,
-                           index_t block_start,
-                           index_t block_end)
-            : karg_{karg},
-              block_2_ctile_map_{b2c_map},
-              block_start_{block_start},
-              block_end_{block_end}
-        {
-        }
-    };
-
     static constexpr index_t DefaultKBatch = 1;
 
     // Argument
@@ -225,7 +284,6 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                  std::vector<GemmDesc>& gemm_descs)
             : Argument(p_As, p_Bs, p_Es, gemm_descs, DefaultKBatch)
         {
-            // TODO: use occupancy api to calculate appropriate batch size.
         }
 
         Argument(std::vector<const void*>& p_As,
@@ -233,9 +291,8 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                  std::vector<void*>& p_Es,
                  std::vector<GemmDesc>& gemm_descs,
                  index_t kbatch)
-            : K_BATCH{kbatch}
+            : K_BATCH{kbatch}, group_count_{0}, skipped_group_count_{0}, grid_size_{0}
         {
-            grid_size_   = 0;
             group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
 
             if(!(group_count_ == ck::type_convert<ck::index_t>(p_As.size()) &&
@@ -247,8 +304,6 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
 
             gemm_kernel_args_.reserve(group_count_);
 
-            skipped_group_count_ = 0;
-
             for(std::size_t i = 0; i < gemm_descs.size(); ++i)
             {
                 const index_t M = gemm_descs[i].M_;
@@ -265,27 +320,13 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                 const index_t stride_b = gemm_descs[i].stride_B_;
                 const index_t stride_c = gemm_descs[i].stride_C_;
 
-                const index_t m_padded = GridwiseGemm::CalculateMPadded(M);
-                const index_t n_padded = GridwiseGemm::CalculateNPadded(N);
-                const index_t k_padded = GridwiseGemm::CalculateKPadded(K, K_BATCH);
-                const index_t k0       = GridwiseGemm::CalculateK0(K, K_BATCH);
-
                 const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, stride_c);
 
-                const auto local_b2c_tile_map =
-                    Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
+                auto local_b2c_tile_map = Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
                 const index_t grid_size_grp = local_b2c_tile_map.CalculateGridSize(c_grid_desc_m_n);
-
-                const index_t block_start = grid_size_;
-                const index_t block_end   = grid_size_ + grid_size_grp;
-
                 grid_size_ += grid_size_grp;
 
-                // block-to-e-tile map
-                auto grouped_block_2_ctile_map =
-                    GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
-
-                auto karg = KernelArgument{type_convert<const ADataType*>(p_As[i]),
+                gemm_kernel_args_.emplace_back(type_convert<const ADataType*>(p_As[i]),
                                                type_convert<const BDataType*>(p_Bs[i]),
                                                type_convert<EDataType*>(p_Es[i]),
                                                M,
@@ -293,23 +334,15 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                                                K,
                                                stride_a,
                                                stride_b,
-                                           stride_c,
-                                           m_padded,
-                                           n_padded,
-                                           k_padded,
-                                           k0,
-                                           K_BATCH};
-
-                gemm_kernel_args_.emplace_back(
-                    std::move(karg), std::move(grouped_block_2_ctile_map), block_start, block_end);
+                                               stride_c);
             }
         }
 
-        /**
-         * @brief      Recalculate group grid size for all gemms and update B2C maps.
-         *
-         * @param[in]  kbatch  The new splitK parameter value.
-         */
+        ///
+        /// @brief      Set new kbatch value.
+        ///
+        /// @param[in]  kbatch  The new splitK parameter value.
+        ///
         void UpdateKBatch(index_t kbatch)
         {
             K_BATCH    = kbatch;
@@ -318,33 +351,14 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
             for(std::size_t i = 0; i < gemm_kernel_args_.size(); ++i)
             {
 
-                auto& karg = gemm_kernel_args_[i].karg_;
-
-                const index_t k_padded = GridwiseGemm::CalculateKPadded(karg.K, K_BATCH);
-                const index_t k0       = GridwiseGemm::CalculateK0(karg.K, K_BATCH);
+                auto& gemm_arg = gemm_kernel_args_[i];
 
                 const auto c_grid_desc_m_n =
-                    GridwiseGemm::MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
+                    GridwiseGemm::MakeCGridDescriptor_M_N(gemm_arg.M, gemm_arg.N, gemm_arg.StrideC);
 
-                const auto local_b2c_tile_map =
-                    Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
+                auto local_b2c_tile_map = Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
                 const index_t grid_size_grp = local_b2c_tile_map.CalculateGridSize(c_grid_desc_m_n);
-
-                const index_t block_start = grid_size_;
-                const index_t block_end   = grid_size_ + grid_size_grp;
-
                 grid_size_ += grid_size_grp;
-
-                // block-to-e-tile map
-                auto grouped_block_2_ctile_map =
-                    GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
-
-                karg.KPadded                            = k_padded;
-                karg.K0                                 = k0;
-                karg.k_batch                            = K_BATCH;
-                gemm_kernel_args_[i].block_2_ctile_map_ = grouped_block_2_ctile_map;
-                gemm_kernel_args_[i].block_start_       = block_start;
-                gemm_kernel_args_[i].block_end_         = block_end;
             }
         }
 
@@ -352,31 +366,167 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
         index_t K_BATCH;
         index_t group_count_;
         index_t skipped_group_count_;
-
-        std::vector<GemmTransKernelArg> gemm_kernel_args_;
+        // The overall number of output tiles to be processed.
         index_t grid_size_;
+        const void* p_dev_gemm_args_;
+
+        std::vector<KernelArguments> gemm_kernel_args_;
     };
 
     // Invoker
     struct Invoker : public BaseInvoker
     {
+        // The oversubscription factor for the number of blocks that can simultaneously reside on
+        // GPU.
+        static constexpr int BLOCK_SUBSCRIPTION_FACTOR = 1;
+        static constexpr int BLOCK_WAVES               = BlockSize / get_warp_size();
+        static constexpr int CU_SIMDS                  = 4;
+        // Assume we want to have at most 2 waves per SIMD
+        static constexpr int CU_BLOCKS = math::integer_divide_floor(2 * CU_SIMDS, BLOCK_WAVES);
+
+        ///
+        /// @brief      Launch Grouped Gemm kernel.
+        ///
+        /// @note       This function overload is using user provided device buffer for kernel
+        ///             arguments.
+        ///
+        /// @param[in]  arg            The structure containing kernel arguments (in host memory).
+        /// @param[in]  dev_gemm_args  The point to device memory with kernel arguments.
+        /// @param[in]  stream_config  The device stream configuration.
+        ///
+        /// @return     The average kernel execution time (if time measurement is enabled.)
+        ///
+        float Run(const Argument& arg,
+                  const void* dev_gemm_args,
+                  const StreamConfig& stream_config = StreamConfig{})
+        {
+            auto [all_have_kbatch_gt_one, all_have_main_k0_block_loop] =
+                CheckArgument(arg, stream_config);
+
+            if(dev_gemm_args == nullptr)
+            {
+                std::ostringstream err;
+                err << "The gemm arguments workspace buffer is not allocated!"
+                    << " In " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+                throw std::runtime_error(err.str());
+            }
+
+            if(all_have_kbatch_gt_one)
+            {
+                for(const auto& gemm_arg : arg.gemm_kernel_args_)
+                {
+                    hip_check_error(hipMemsetAsync(gemm_arg.p_c_grid,
+                                                   0,
+                                                   gemm_arg.M * gemm_arg.N * sizeof(EDataType),
+                                                   stream_config.stream_id_));
+                }
+            }
+
+            float ave_time = 0;
+
+            if(all_have_main_k0_block_loop)
+            {
+                if(all_have_kbatch_gt_one)
+                {
+                    ave_time = DispatchKernel<InMemoryDataOperationEnum::AtomicAdd, true>(
+                        arg, dev_gemm_args, stream_config);
+                }
+                else
+                {
+                    ave_time = DispatchKernel<InMemoryDataOperationEnum::Set, true>(
+                        arg, dev_gemm_args, stream_config);
+                }
+            }
+            else
+            {
+                if(all_have_kbatch_gt_one)
+                {
+                    ave_time = DispatchKernel<InMemoryDataOperationEnum::AtomicAdd, false>(
+                        arg, dev_gemm_args, stream_config);
+                }
+                else
+                {
+                    ave_time = DispatchKernel<InMemoryDataOperationEnum::Set, false>(
+                        arg, dev_gemm_args, stream_config);
+                }
+            }
+
+            return ave_time;
+        }
+
+        ///
+        /// @brief      Launch Grouped Gemm kernel.
+        ///
+        /// @note       This function overload is using device workspace buffer for kernel
+        ///             arguments. The user should call @see GetWorkSpaceSize and @see
+        ///             SetWorkSpacePointer on arg parameter to properly allocate this buffer.
+        ///
+        /// @param[in]  arg            The structure containing kernel arguments (in host memory).
+        /// @param[in]  stream_config  The device stream configuration.
+        ///
+        /// @return     The average kernel execution time (if time measurement is enabled.)
+        ///
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-            index_t K0                       = arg.gemm_kernel_args_[0].karg_.K0;
-            bool all_have_kbatch_gt_one      = arg.gemm_kernel_args_[0].karg_.k_batch > 1;
+            if(arg.p_workspace_ != nullptr)
+            {
+                hip_check_error(
+                    hipMemcpyWithStream(arg.p_workspace_,
+                                        arg.gemm_kernel_args_.data(),
+                                        arg.gemm_kernel_args_.size() * sizeof(KernelArguments),
+                                        hipMemcpyHostToDevice,
+                                        stream_config.stream_id_));
+            }
+            else
+            {
+                std::ostringstream err;
+                err << "The gemm arguments workspace buffer is not allocated!"
+                    << " In " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+                throw std::runtime_error(err.str());
+            }
+
+            return Run(arg, arg.p_workspace_, stream_config);
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+
+        private:
+        auto CheckArgument(const Argument& arg, const StreamConfig& stream_config) const
+        {
+            index_t K0 = GridwiseGemm::CalculateK0(arg.gemm_kernel_args_[0].K, arg.K_BATCH);
+            bool all_have_kbatch_gt_one      = arg.K_BATCH > 1;
             bool all_have_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
 
             for(std::size_t i = 0; i < arg.gemm_kernel_args_.size(); ++i)
             {
-                const auto& karg = arg.gemm_kernel_args_[i].karg_;
+                const auto& gemm_arg = arg.gemm_kernel_args_[i];
                 if(stream_config.log_level_ > 0)
                 {
-                    karg.Print();
+                    gemm_arg.Print();
                 }
 
-                auto kbatch = karg.k_batch;
+                // Currently all groups use same kbatch value.
+                auto kbatch = arg.K_BATCH;
+                K0          = GridwiseGemm::CalculateK0(arg.gemm_kernel_args_[i].K, arg.K_BATCH);
 
-                if(!GridwiseGemm::CheckValidity(karg))
+                if(!GridwiseGemm::CheckValidity(GridwiseGemmArg{nullptr,
+                                                                nullptr,
+                                                                nullptr,
+                                                                gemm_arg.M,
+                                                                gemm_arg.N,
+                                                                gemm_arg.K,
+                                                                gemm_arg.StrideA,
+                                                                gemm_arg.StrideB,
+                                                                gemm_arg.StrideC,
+                                                                0, // MPadded
+                                                                0, // NPadded
+                                                                0, // KPadded
+                                                                K0,
+                                                                kbatch}))
                 {
                     std::ostringstream err;
                     err << "Group id: " << i << " has invalid GridwiseGemm settings!" << __FILE__
@@ -384,7 +534,6 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                     throw std::runtime_error(err.str());
                 }
 
-                K0 = karg.K0;
                 bool not_all_have_main_k0_block_loop_same =
                     all_have_main_k0_block_loop xor GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
                 bool not_all_have_kbatch_value_same = all_have_kbatch_gt_one xor (kbatch > 1);
@@ -402,99 +551,75 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                     std::ostringstream err;
                     err << "Not all gemms have same kbatch value (=1 or >1)! "
                         << "group [" << i << "], kbatch: " << kbatch
-                        << ", group [0], kbatch: " << arg.gemm_kernel_args_[0].karg_.k_batch
-                        << " in " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+                        << ", group [0], kbatch: " << arg.K_BATCH << " in " << __FILE__ << ":"
+                        << __LINE__ << ", in function: " << __func__;
                     throw std::runtime_error(err.str());
                 }
             }
-
-            hip_check_error(
-                hipMemcpyWithStream(arg.p_workspace_,
-                                    arg.gemm_kernel_args_.data(),
-                                    arg.gemm_kernel_args_.size() * sizeof(GemmTransKernelArg),
-                                    hipMemcpyHostToDevice,
-                                    stream_config.stream_id_));
-
-            float ave_time = 0;
-
-            const auto Run = [&](const auto& kernel) {
-                if(all_have_kbatch_gt_one)
-                {
-                    for(const auto& trans_arg : arg.gemm_kernel_args_)
-                    {
-                        const auto& karg = trans_arg.karg_;
-                        hip_check_error(hipMemsetAsync(karg.p_c_grid,
-                                                       0,
-                                                       karg.M * karg.N * sizeof(EDataType),
-                                                       stream_config.stream_id_));
-                    }
+            return std::make_tuple(all_have_kbatch_gt_one, all_have_main_k0_block_loop);
         }
 
-                ave_time =
-                    launch_and_time_kernel(stream_config,
-                                           kernel,
-                                           dim3(arg.grid_size_),
-                                           dim3(BlockSize),
-                                           0,
-                                           cast_pointer_to_constant_address_space(arg.p_workspace_),
-                                           arg.gemm_kernel_args_.size());
-            };
-
-            if(all_have_main_k0_block_loop)
-            {
-                if(all_have_kbatch_gt_one)
+        template <InMemoryDataOperationEnum CGlobalMemoryDataOperation, bool HasMainKBlockLoop>
+        float DispatchKernel(const Argument& arg,
+                             const void* dev_gemm_args,
+                             const StreamConfig& stream_config) const
         {
-                    const auto kernel =
-                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
-                                                       GemmTransKernelArg,
-                                                       true,
-                                                       InMemoryDataOperationEnum::AtomicAdd>;
-
-                    Run(kernel);
+            const auto kernel = kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                               KernelArguments,
+                                                               ADataType,
+                                                               BDataType,
+                                                               EDataType,
+                                                               HasMainKBlockLoop,
+                                                               CGlobalMemoryDataOperation>;
+            return LaunchKernel(kernel, arg, dev_gemm_args, stream_config);
         }
-                else
-                {
-                    const auto kernel =
-                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
-                                                       GemmTransKernelArg,
-                                                       true,
-                                                       InMemoryDataOperationEnum::Set>;
 
-                    Run(kernel);
-                }
-            }
-            else
-            {
-                if(all_have_kbatch_gt_one)
+        template <typename KernelFunction>
+        int CalculateMaxOccupancyGridSize(const KernelFunction& kernel,
+                                          const StreamConfig& stream_config) const
         {
-                    const auto kernel =
-                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
-                                                       GemmTransKernelArg,
-                                                       false,
-                                                       InMemoryDataOperationEnum::AtomicAdd>;
+            // Calculate max number of workgroups that can simultaneously reside on the CU.
+            int num_blocks                = 0;
+            size_t dyn_shared_mem_per_blk = 0;
+            hip_check_error(hipOccupancyMaxActiveBlocksPerMultiprocessor(
+                &num_blocks, kernel, BlockSize, dyn_shared_mem_per_blk));
 
-                    Run(kernel);
-                }
-                else
-                {
-                    const auto kernel =
-                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
-                                                       GemmTransKernelArg,
-                                                       false,
-                                                       InMemoryDataOperationEnum::Set>;
+            int cu_count = getAvailableComputeUnitCount(stream_config);
 
-                    Run(kernel);
-                }
+            if(stream_config.log_level_ > 0)
+            {
+                std::cout << "MaxActiveBlocksPerCU: " << num_blocks
+                          << ", available CUs count: " << cu_count << ", occup. grid size: "
+                          << ck::math::min(num_blocks, CU_BLOCKS) * cu_count *
+                                 BLOCK_SUBSCRIPTION_FACTOR
+                          << std::endl;
             }
 
-            return ave_time;
+            return cu_count * ck::math::min(num_blocks, CU_BLOCKS) * BLOCK_SUBSCRIPTION_FACTOR;
         }
 
-        // polymorphic
-        float Run(const BaseArgument* p_arg,
-                  const StreamConfig& stream_config = StreamConfig{}) override
+        template <typename KernelFunction>
+        float LaunchKernel(const KernelFunction& kernel,
+                           const Argument& arg,
+                           const void* dev_gemm_args,
+                           const StreamConfig& stream_config) const
         {
-            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+            int max_occupancy_grid_size = CalculateMaxOccupancyGridSize(kernel, stream_config);
+
+            // We launch the smaller number of workgroups from acutally needed tiles and the
+            // number of workgroups that maximize the GPU occupancy. That is because for some tile
+            // configuration the first is smaller than the latter. Launching too many workgroups
+            // mean some of them will have to iterate through all gemm problem descriptors just to
+            // find out they have nothing to do which is of course waste of GPU cycles.
+            return launch_and_time_kernel(
+                stream_config,
+                kernel,
+                dim3(ck::math::min(arg.grid_size_, max_occupancy_grid_size)),
+                dim3(BlockSize),
+                0,
+                cast_pointer_to_constant_address_space(dev_gemm_args),
+                arg.grid_size_,
+                arg.K_BATCH);
         }
     };
 
@@ -506,11 +631,6 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!ck::is_xdl_supported())
-        {
-            return false;
-        }
-
         if((ck::type_convert<ck::index_t>(arg.gemm_kernel_args_.size()) +
             arg.skipped_group_count_) != arg.group_count_)
         {
@@ -525,14 +645,28 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
         bool supported = true;
         for(std::size_t i = 0; i < arg.gemm_kernel_args_.size(); ++i)
         {
-            const auto& a        = arg.gemm_kernel_args_[i].karg_;
-            bool group_arg_valid = GridwiseGemm::CheckValidity(a);
+            const auto& gemm_arg = arg.gemm_kernel_args_[i];
+            const auto K0        = GridwiseGemm::CalculateK0(gemm_arg.K, arg.K_BATCH);
+            bool group_arg_valid = GridwiseGemm::CheckValidity(GridwiseGemmArg{nullptr,
+                                                                               nullptr,
+                                                                               nullptr,
+                                                                               gemm_arg.M,
+                                                                               gemm_arg.N,
+                                                                               gemm_arg.K,
+                                                                               gemm_arg.StrideA,
+                                                                               gemm_arg.StrideB,
+                                                                               gemm_arg.StrideC,
+                                                                               0, // MPadded
+                                                                               0, // NPadded
+                                                                               0, // KPadded
+                                                                               K0,
+                                                                               arg.K_BATCH});
             if(not group_arg_valid)
             {
 #if DEBUG_LOG
                 std::cout << "[" << __func__ << "] group id: " << i
                           << " has invalid GridwiseGemm settings!" << std::endl;
-                a.Print();
+                gemm_arg.Print();
 #endif // DEBUG_LOG
             }
             supported = supported && group_arg_valid;
@@ -540,7 +674,6 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
         return supported;
     }
 
-    // polymorphic
     bool IsSupportedArgument(const BaseArgument* p_arg) override
     {
         return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
@@ -560,7 +693,6 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
 
     static auto MakeInvoker() { return Invoker{}; }
 
-    // polymorphic
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(std::vector<const void*>& p_As,
                         std::vector<const void*>& p_Bs,
@@ -574,19 +706,17 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
         return std::make_unique<Argument>(p_As, p_Bs, p_Es, gemm_descs);
     }
 
-    // polymorphic
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
         return std::make_unique<Invoker>(Invoker{});
     }
 
-    // polymorphic
     std::string GetTypeString() const override
     {
         auto str = std::stringstream();
 
         // clang-format off
-        str << "DeviceGroupedGemm_XdlSplitK"
+        str << "DeviceGroupedGemm_XdlSplitKTileLoop"
             << "<"
             << std::string(ALayout::name)[0] << ","
             << std::string(BLayout::name)[0] << ","
@@ -605,7 +735,9 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
             << BBlockTransferSrcScalarPerVector << ", "
             << CShuffleMXdlPerWavePerShuffle << ", "
             << CShuffleNXdlPerWavePerShuffle << ", "
-            << getGemmSpecializationString(GemmSpec)
+            << ABlockTransferThreadClusterLengths_K0_M_K1{} << ", "
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << PipelineVer
             << ">";
         // clang-format on
 
@@ -615,16 +747,24 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
     size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
     {
         return dynamic_cast<const Argument*>(p_arg)->gemm_kernel_args_.size() *
-               sizeof(GemmTransKernelArg);
+               sizeof(KernelArguments);
     }
 
     static void SetKBatchSize(Argument& arg, index_t kbatch) { arg.UpdateKBatch(kbatch); }
+    static void SetDeviceKernelArgs(Argument& arg, const void* p_dev_kernel_args)
+    {
+        arg.p_dev_gemm_args_ = p_dev_kernel_args;
+    }
 
-    // polymorphic
     void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const override
     {
         return SetKBatchSize(*dynamic_cast<Argument*>(p_arg), kbatch);
     }
+
+    void SetDeviceKernelArgs(BaseArgument* p_arg, const void* p_dev_kernel_args) const override
+    {
+        return SetDeviceKernelArgs(*dynamic_cast<Argument*>(p_arg), p_dev_kernel_args);
+    }
 };
 
 } // namespace device

include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp

@@ -348,24 +348,24 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
                         acc_elementwise_op};
     };
 
-    //
-    // @brief      Makes a pointer to Argument class.
-    //
-    // @param[in]  inLengths           Input tensor extent(s) from high to low dimension
-    // @param[in]  inStrides           Input tensor stride(s) from high to low dimension
-    // @param[in]  reduceDims          The dimension(s) the normalization operation is applied
-    // @param[in]  alpha               Typeless pointer in host memory storing the alpha scaling
-    //                                 value as type AccDataType
-    // @param[in]  beta                Typeless pointer in host memory storing the beta scaling
-    //                                 value as type AccDataType
-    // @param[in]  in_dev              Typeless const pointer in device memory storing the input
-    //                                 tensor
-    // @param      out_dev             Typeless pointer in device memory storing the output tensor
-    // @param[in]  in_elementwise_op   The input elementwise operation.
-    // @param[in]  acc_elementwise_op  The accumulation elementwise operation.
-    //
-    // @return     Unique pointer to the Argument class.
-    //
+    ///
+    /// @brief      Makes a pointer to Argument class.
+    ///
+    /// @param[in]  inLengths           Input tensor extent(s) from high to low dimension
+    /// @param[in]  inStrides           Input tensor stride(s) from high to low dimension
+    /// @param[in]  reduceDims          The dimension(s) the normalization operation is applied
+    /// @param[in]  alpha               Typeless pointer in host memory storing the alpha scaling
+    ///                                 value as type AccDataType
+    /// @param[in]  beta                Typeless pointer in host memory storing the beta scaling
+    ///                                 value as type AccDataType
+    /// @param[in]  in_dev              Typeless const pointer in device memory storing the input
+    ///                                 tensor
+    /// @param      out_dev             Typeless pointer in device memory storing the output tensor
+    /// @param[in]  in_elementwise_op   The input elementwise operation.
+    /// @param[in]  acc_elementwise_op  The accumulation elementwise operation.
+    ///
+    /// @return     Unique pointer to the Argument class.
+    ///
     std::unique_ptr<BaseArgument> MakeArgumentPointer(const std::vector<index_t> inLengths,
                                                       const std::vector<index_t> inStrides,
                                                       const std::vector<int> reduceDims,

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

@@ -271,7 +271,8 @@ struct BlockToCTileMap_KSplit_M00_N0_M01Adapt
     {
     }
 
-    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    __host__ __device__ constexpr index_t
+    CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
     {
         const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
         const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
@@ -624,24 +625,36 @@ struct OffsettedBlockToCTileMap
     index_t block_start_;
 };
 
-/**
- * @brief      Simple tile mapping which creates 3D grid of block of threads.
- *
- * @paragraph  Description
- *             This Block-to-C-tile-map creates a 3D grid (n_blocks, m_blocks, z_blocks) of thread
- *             blocks. The first 2D are regular 2D tiles created by division of output GEMM
- *             dimenions by corresponding tile size. The third dimension (Z) is a k-split dimension,
- *             which denotes the number of blocks we use to divide work on GEMM K dimension onto.
- *
- * @tparam     MPerBlock  Output block tile size in M dimension.
- * @tparam     NPerBlock  Output block tile size in N dimension.
- */
+///
+/// @brief      Simple tile mapping which creates 3D grid of block of threads.
+///
+/// @paragraph  Description
+///             This Block-to-C-tile-map creates a 3D grid (n_blocks, m_blocks, z_blocks) of thread
+///             blocks. The first 2D are regular 2D tiles created by division of output GEMM
+///             dimenions by corresponding tile size. The third dimension (Z) is a k-split
+///             dimension, which denotes the number of blocks we use to divide work on GEMM K
+///             dimension onto.
+///
+/// @tparam     MPerBlock  Output block tile size in M dimension.
+/// @tparam     NPerBlock  Output block tile size in N dimension.
+///
 template <index_t MPerBlock, index_t NPerBlock>
 struct BlockToCTileMap_3DGrid_KSplit
 {
-
     __host__ __device__ BlockToCTileMap_3DGrid_KSplit() = default;
 
+    ///
+    /// @brief      Constructs a new instance.
+    ///
+    /// @param[in]  top_idx  Swallow blockIdx.
+    ///
+    /// @tparam     TopIdx   The type of block index.
+    ///
+    template <typename TopIdx>
+    __host__ __device__ BlockToCTileMap_3DGrid_KSplit([[maybe_unused]] TopIdx top_idx)
+    {
+    }
+
     __host__ __device__ constexpr auto
     CalculateGridSize(index_t M, index_t N, index_t k_split) const
     {
@@ -652,8 +665,7 @@ struct BlockToCTileMap_3DGrid_KSplit
         return std::make_tuple(N0, M0, k_split);
     }
 
-    template <typename TopIdx>
-    __device__ constexpr auto CalculateBottomIndex(const TopIdx&) const
+    __device__ constexpr auto CalculateBottomIndex() const
     {
         return make_tuple(blockIdx.z, blockIdx.y, blockIdx.x);
     }
@@ -672,6 +684,53 @@ struct BlockToCTileMap_3DGrid_KSplit
     }
 };
 
+///
+/// @brief      Block to CTile Map which foster external mechanism for setting up local block id.
+///
+///             In example this type can be easily used to implement tile looping work distribution
+///             scheme.
+///
+/// @tparam     UnderlyingBlockToCTileMap  The type of the local tile mapp.
+///
+template <typename UnderlyingBlockToCTileMap>
+struct LocalBlockToCTileMap
+{
+    using underlying_type = UnderlyingBlockToCTileMap;
+
+    __host__ __device__ LocalBlockToCTileMap(UnderlyingBlockToCTileMap block_to_ctile_map,
+                                             index_t local_id)
+        : block_to_ctile_map_{block_to_ctile_map}, local_block_id_{local_id}
+    {
+    }
+
+    __host__ __device__ constexpr auto CalculateBottomIndex() const
+    {
+        return block_to_ctile_map_.CalculateBottomIndex(make_multi_index(local_block_id_));
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        return block_to_ctile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        return block_to_ctile_map_.CheckValidity(c_grid_desc_m_n);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        return block_to_ctile_map_.CalculateGridSize(c_grid_desc_m_n);
+    }
+
+    UnderlyingBlockToCTileMap block_to_ctile_map_;
+    index_t local_block_id_;
+};
+
 enum StreamKReductionStrategy
 {
     Atomic = 0, // sk block use atomic to do reduction

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

@@ -4,6 +4,8 @@
 #pragma once
 
 #include <iostream>
+#include <ostream>
+#include <string>
 
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp"
@@ -42,4 +44,20 @@ constexpr auto GridwiseGemmPipeline_Selector()
     }
 }
 
+inline std::string getPipelineVersionString(const PipelineVersion& pv)
+{
+    switch(pv)
+    {
+    case PipelineVersion::v1: return "PipelineVersion::v1";
+    case PipelineVersion::v2: return "PipelineVersion::v2";
+    default: return "Unrecognized pipeline version!";
+    }
+}
+
 } // namespace ck
+
+inline std::ostream& operator<<(std::ostream& os, const ck::PipelineVersion pv)
+{
+    os << ck::getPipelineVersionString(pv);
+    return os;
+}

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp

@@ -27,8 +27,7 @@ __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
-                                             const Block2CTileMap& b2c_map)
+        kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
     defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
@@ -36,11 +35,12 @@ __global__ void
 
     __shared__ uint8_t p_shared[shared_size];
 
+    Block2CTileMap b2c_map{get_block_1d_id()};
+
     GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
         karg, static_cast<void*>(p_shared), b2c_map);
 #else
     ignore = karg;
-    ignore = b2c_map;
 #endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
@@ -541,15 +541,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
             make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
     }
 
-    // return block_id to C matrix tile idx (m0, n0) mapping
-    template <typename CGridDesc>
-    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
-        const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
-    {
-        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>(
-            c_m_n_grid_desc, 8, KBatch);
-    }
-
     __host__ __device__ static constexpr auto
     GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
     {
@@ -575,18 +566,28 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
     template <bool HasMainKBlockLoop,
               InMemoryDataOperationEnum CGlobalMemoryDataOperation,
               typename Block2CTileMap>
-    __device__ static void Run(const Argument& karg,
+    __device__ static void Run(const FloatA* p_a_grid,
+                               const FloatB* p_b_grid,
+                               FloatC* p_c_grid,
+                               index_t M,
+                               index_t N,
+                               index_t K,
+                               index_t StrideA,
+                               index_t StrideB,
+                               index_t StrideC,
+                               index_t MPadded,
+                               index_t NPadded,
+                               index_t KPadded,
+                               index_t K0,
+                               index_t k_batch,
                                void* __restrict__ p_shared_block,
                                const Block2CTileMap& block_2_ctile_map)
     {
-        const FloatA* p_a_grid           = karg.p_a_grid;
-        const FloatB* p_b_grid           = karg.p_b_grid;
-        FloatC* p_c_grid                 = karg.p_c_grid;
-        const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(
-            karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0, karg.KPadded);
-        const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(
-            karg.K, karg.NPadded, karg.N, karg.StrideB, karg.k_batch, karg.K0, karg.KPadded);
-        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
+        const auto a_b_k0_m_k1_grid_desc =
+            MakeAGridDescriptor_KBatch_K0_M_K1(M, MPadded, K, StrideA, k_batch, K0, KPadded);
+        const auto b_b_k0_n_k1_grid_desc =
+            MakeBGridDescriptor_KBatch_K0_N_K1(K, NPadded, N, StrideB, k_batch, K0, KPadded);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(M, N, StrideC);
 
         const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
             MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
@@ -602,8 +603,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
             p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
 
         // divide block work by [KBatch, M, N]
-        const auto block_work_idx =
-            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+        const auto block_work_idx = block_2_ctile_map.CalculateBottomIndex();
 
         if(!block_2_ctile_map.ValidCTileIndex(
                block_work_idx,
@@ -1010,6 +1010,34 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         }
     }
 
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              typename Block2CTileMap>
+    __device__ static void Run(const Argument& karg,
+                               void* __restrict__ p_shared_block,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, Block2CTileMap>(karg.p_a_grid,
+                                                                           karg.p_b_grid,
+                                                                           karg.p_c_grid,
+                                                                           karg.M,
+                                                                           karg.N,
+                                                                           karg.K,
+                                                                           karg.StrideA,
+                                                                           karg.StrideB,
+                                                                           karg.StrideC,
+                                                                           karg.MPadded,
+                                                                           karg.NPadded,
+                                                                           karg.KPadded,
+                                                                           karg.K0,
+                                                                           karg.k_batch,
+                                                                           p_shared_block,
+                                                                           block_2_ctile_map);
+    }
+
+    static constexpr auto GetMPerBlock() { return MPerBlock; }
+    static constexpr auto GetNPerBlock() { return NPerBlock; }
+
     static std::string GetTypeString()
     {
         auto str = std::stringstream();

include/ck/utility/sequence.hpp

@@ -897,3 +897,14 @@ template <index_t NSize, index_t I>
 using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type;
 
 } // namespace ck
+
+template <ck::index_t... Is>
+std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>)
+{
+    using S = ck::Sequence<Is...>;
+    os << "{";
+    ck::static_for<0, S::Size() - ck::Number<1>{}, 1>{}(
+        [&](auto i) { os << S::At(i).value << ", "; });
+    os << S::At(S::Size() - ck::Number<1>{}).value << "}";
+    return os;
+}

library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp

@@ -14,27 +14,27 @@ namespace ck {
 namespace tensor_operation {
 namespace host {
 
-//
-// @brief      Reference implementation for forward convolution.
-//
-// @paragraph
-//             Tensor descriptor in GNCHW/GKCXY/GNKHW dimensional order
-//             Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout
-//             as long as dimensions in tensor descriptor is in GNCHW order
-//
-// @tparam     InDataType               Input tensor data type.
-// @tparam     WeiDataType              Weights tensor data type.
-// @tparam     OutDataType              Output tensor data type.
-// @tparam     InElementwiseOperation   Functor for input tensor elementwise
-//                                      operation.
-// @tparam     WeiElementwiseOperation  Functor for weights tensor elementwise
-//                                      operation.
-// @tparam     NDimSpatial  Number of spatial dimensions.
-//
-// input descriptor in [G, N, C, Do, Ho, Wo] order
-// weight descriptor in [G, K, C, Z, Y, X] order
-// output descriptor in [G, N, K, Di, Hi, Wi] order
-// phyiscal layout is irrelavent
+///
+/// @brief      Reference implementation for forward convolution.
+///
+/// @paragraph
+///             Tensor descriptor in GNCHW/GKCXY/GNKHW dimensional order
+///             Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout
+///             as long as dimensions in tensor descriptor is in GNCHW order
+///
+/// @tparam     InDataType               Input tensor data type.
+/// @tparam     WeiDataType              Weights tensor data type.
+/// @tparam     OutDataType              Output tensor data type.
+/// @tparam     InElementwiseOperation   Functor for input tensor elementwise
+///                                      operation.
+/// @tparam     WeiElementwiseOperation  Functor for weights tensor elementwise
+///                                      operation.
+/// @tparam     NDimSpatial  Number of spatial dimensions.
+///
+/// input descriptor in [G, N, C, Do, Ho, Wo] order
+/// weight descriptor in [G, K, C, Z, Y, X] order
+/// output descriptor in [G, N, K, Di, Hi, Wi] order
+/// phyiscal layout is irrelavent
 template <ck::index_t NDimSpatial,
           typename InDataType,
           typename WeiDataType,

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp

@@ -33,27 +33,41 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
 // a[m, k] * b[k, n] = e[m, n]
 using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
     // clang-format off
-        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
-        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
-        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
-        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        // Currently AK1 must equal BK1 !
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|     PipelineVersion|
+    //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                    |
+    //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                    |
+    //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                    |
+    // Pipline v1
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
+
+    // Pipline v2
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp

@@ -31,57 +31,41 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
 
 using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
     // clang-format off
-        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
-        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
-        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
-        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        // Currently AK1 must equal BK1 !
-     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
-     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
-        
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+    //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|     PipelineVersion|
+    //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                    |
+    //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                    |
+    //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                    |
+    // Pipline v1
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+  //DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+  //DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+  //DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
     
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+    // Pipline v2
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+  //DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+  //DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+  //DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
     DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp

@@ -32,23 +32,31 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
 // a[m, k] * b[n, k] = e[m, n]
 using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances = std::tuple<
     // clang-format off
-        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
-        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
-        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
-        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         PipelineVersion|
+    //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                        |
+    //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                        |
+    //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                        |
+    // Pipeline v1
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,   8,   8,   32,   32,    2,    4,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,    64,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,    32,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    64,   8,   8,   32,   32,    1,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, ck::PipelineVersion::v1>,
+    // Pipeline v2
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,   8,   8,   32,   32,    2,    4,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,    64,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,    32,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    64,   8,   8,   32,   32,    1,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, ck::PipelineVersion::v2>
+
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp

@@ -31,30 +31,31 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
 
 using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
     // clang-format off
-        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
-        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
-        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
-        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 48, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,    
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   256,    32,   8,   8,   32,   32,    1,    4,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
-        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         PipelineVersion|
+    //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                        |
+    //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                        |
+    //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                        |
+    // Pipeline v1
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    64,   8,   8,   32,   32,    2,    4,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, ck::PipelineVersion::v1>,
+  //DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v1>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    64,   8,   8,   32,   32,    1,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, ck::PipelineVersion::v1>,
+    
+    // Pipeline v2
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    64,   8,   8,   32,   32,    2,    4,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, ck::PipelineVersion::v2>,
+  //DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, ck::PipelineVersion::v2>,
+    DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    64,   8,   8,   32,   32,    1,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, ck::PipelineVersion::v2>
     // clang-format on
     >;
 

script/cmake-ck-dev.sh

@@ -12,7 +12,9 @@ cmake
 -save-temps=$PWD"                                                                                 \
 -D CMAKE_BUILD_TYPE=Release                                                                       \
 -D BUILD_DEV=ON                                                                                   \
--D GPU_TARGETS="gfx908;gfx90a;gfx940"                                                             \
+-D GPU_TARGETS="gfx90a"                                                             \
 -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                 \
 -D USE_BITINT_EXTENSION_INT4=OFF                                                                  \
 ${MY_PROJECT_SOURCE}
+
+#-D GPU_TARGETS="gfx908;gfx90a;gfx940"                                                             \