Merge branch 'develop' into ck_profiler_m_instances

include/ck_tile/ops/gemm/kernel/gemm_kernel.hpp

@@ -8,6 +8,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/common.hpp"
+#include "ck_tile/host/concat.hpp"
 
 namespace ck_tile {
 
@@ -75,6 +76,13 @@ struct GemmKernel
     static constexpr auto I1 = number<1>();
     static constexpr auto I2 = number<2>();
 
+    [[nodiscard]] CK_TILE_HOST static const std::string GetName()
+    {
+        // clang-format off
+        return concat('_', "gemm", gemm_prec_str<ADataType, BDataType>, GemmPipeline::GetName());
+        // clang-format on
+    }
+
     CK_TILE_HOST static constexpr auto GridSize(index_t M, index_t N, index_t KBatch)
     {
         return dim3(TilePartitioner::GridSize(M, N), 1, KBatch);
@@ -455,7 +463,9 @@ struct GemmKernel
      * @param a_ptr input A pointer
      * @param b_ptr input B pointer
      * @param c_ptr output C pointer
+     * @param smem_ptr_0 The start memory pointer of the shared memory block.
      * @param kargs GEMM kernel arguments
+     * @param splitk_batch_offset splitk_batch_offset Utility structure used to calculate k batch.
      * @param block_idx_m The GEMM's output M dimension tile index processed by this workgroup.
      * @param block_idx_n The GEMM's output N dimension tile index processed by this workgroup.
      *
@@ -465,7 +475,7 @@ struct GemmKernel
     CK_TILE_DEVICE static void RunGemm(const ADataType* a_ptr,
                                        const BDataType* b_ptr,
                                        CDataType* c_ptr,
-                                       void* smem_ptr,
+                                       void* smem_ptr_0,
                                        const GemmKernelArgs& kargs,
                                        const SplitKBatchOffset& splitk_batch_offset,
                                        const index_t block_idx_m,
@@ -483,15 +493,67 @@ struct GemmKernel
         // Run GEMM cooperatively by whole workgroup.
         const auto& a_block_window = gemm_tile_windows.at(I0);
         const auto& b_block_window = gemm_tile_windows.at(I1);
-        const auto& c_block_tile =
-            GemmPipeline{}.template operator()(a_block_window, b_block_window, num_loop, smem_ptr);
+
+        const auto& c_block_tile = GemmPipeline{}.template operator()(
+            a_block_window, b_block_window, num_loop, smem_ptr_0);
 
         // Run Epilogue Pipeline
         auto& c_block_window = gemm_tile_windows.at(I2);
 
         EpiloguePipeline{}
             .template operator()<decltype(c_block_window), decltype(c_block_tile), DstInMemOp>(
-                c_block_window, c_block_tile, smem_ptr);
+                c_block_window, c_block_tile, smem_ptr_0);
+    }
+
+    /**
+     * @brief Runs single GEMM problem cooperatively by whole workgroup.
+     *
+     * @note RunGEMM2LDS in with two shared memory buffers using the ping pong buffer mechanism.
+     *
+     * @param a_ptr input A pointer
+     * @param b_ptr input B pointer
+     * @param c_ptr output C pointer
+     * @param smem_ptr_0 The starting pointer of 1st shared memory block.
+     * @param smem_ptr_1 The starting pointer of 2nd shared memory block.
+     * @param kargs GEMM kernel arguments
+     * @param splitk_batch_offset Utility structure used to calculate k batch.
+     * @param block_idx_m The GEMM's output M dimension tile index processed by this workgroup.
+     * @param block_idx_n The GEMM's output N dimension tile index processed by this workgroup.
+     *
+     * @tparam DstInMemOp Destination memory operation (default: set).
+     */
+    template <memory_operation_enum DstInMemOp = memory_operation_enum::set>
+    CK_TILE_DEVICE static void RunGemm2LDS(const ADataType* a_ptr,
+                                           const BDataType* b_ptr,
+                                           CDataType* c_ptr,
+                                           void* __restrict__ smem_ptr_0,
+                                           void* __restrict__ smem_ptr_1,
+                                           const GemmKernelArgs& kargs,
+                                           const SplitKBatchOffset& splitk_batch_offset,
+                                           const index_t block_idx_m,
+                                           const index_t block_idx_n)
+    {
+        // Create Gemm tensor views, pad views and tile windows
+        const auto& gemm_tensor_views_tuple =
+            MakeGemmTensorViews<DstInMemOp>(a_ptr, b_ptr, c_ptr, kargs, splitk_batch_offset);
+        const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
+        auto gemm_tile_windows     = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
+
+        const index_t num_loop = TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k);
+
+        // Run GEMM cooperatively by whole workgroup.
+        const auto& a_block_window = gemm_tile_windows.at(I0);
+        const auto& b_block_window = gemm_tile_windows.at(I1);
+
+        const auto& c_block_tile = GemmPipeline{}.template operator()(
+            a_block_window, b_block_window, num_loop, smem_ptr_0, smem_ptr_1);
+
+        // Run Epilogue Pipeline
+        auto& c_block_window = gemm_tile_windows.at(I2);
+
+        EpiloguePipeline{}
+            .template operator()<decltype(c_block_window), decltype(c_block_tile), DstInMemOp>(
+                c_block_window, c_block_tile, smem_ptr_0);
     }
 
     CK_TILE_DEVICE void operator()(GemmKernelArgs kargs) const
@@ -509,11 +571,27 @@ struct GemmKernel
         CDataType* c_ptr = static_cast<CDataType*>(kargs.c_ptr);
 
         // allocate LDS
-        __shared__ char smem_ptr[GetSmemSize()];
+        __shared__ char smem_ptr_0[GetSmemSize()];
+        __shared__ char smem_ptr_1[GetSmemSize()];
 
         if(kargs.k_batch == 1)
         {
-            RunGemm(a_ptr, b_ptr, c_ptr, smem_ptr, kargs, splitk_batch_offset, i_m, i_n);
+            if constexpr(GemmPipeline::DoubleSmemBuffer == true)
+            {
+                RunGemm2LDS(a_ptr,
+                            b_ptr,
+                            c_ptr,
+                            smem_ptr_0,
+                            smem_ptr_1,
+                            kargs,
+                            splitk_batch_offset,
+                            i_m,
+                            i_n);
+            }
+            else
+            {
+                RunGemm(a_ptr, b_ptr, c_ptr, smem_ptr_0, kargs, splitk_batch_offset, i_m, i_n);
+            }
         }
         else
         {
@@ -522,8 +600,23 @@ struct GemmKernel
             if constexpr(!(EpiloguePipeline::template GetVectorSizeC<CDataType>() % 2 != 0 &&
                            is_any_of<CDataType, fp16_t, bf16_t>::value))
             {
-                RunGemm<memory_operation_enum::atomic_add>(
-                    a_ptr, b_ptr, c_ptr, smem_ptr, kargs, splitk_batch_offset, i_m, i_n);
+                if constexpr(GemmPipeline::DoubleSmemBuffer == true)
+                {
+                    RunGemm2LDS<memory_operation_enum::atomic_add>(a_ptr,
+                                                                   b_ptr,
+                                                                   c_ptr,
+                                                                   smem_ptr_0,
+                                                                   smem_ptr_1,
+                                                                   kargs,
+                                                                   splitk_batch_offset,
+                                                                   i_m,
+                                                                   i_n);
+                }
+                else
+                {
+                    RunGemm<memory_operation_enum::atomic_add>(
+                        a_ptr, b_ptr, c_ptr, smem_ptr_0, kargs, splitk_batch_offset, i_m, i_n);
+                }
             }
         }
     }

include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp

@@ -64,6 +64,18 @@ struct GroupedGemmKernel : public GemmKernel<TilePartitioner_, GemmPipeline_, Ep
         }
     };
 
+    [[nodiscard]] CK_TILE_HOST static const std::string GetName()
+    {
+        // clang-format off
+        using P_ = GemmPipeline;
+
+        return concat('_', "gemm_grouped", gemm_prec_str<ADataType, BDataType>,
+                      concat('x', P_::kMPerBlock, P_::kNPerBlock, P_::kKPerBlock),
+                      concat('x', P_::GetVectorSizeA(), P_::GetVectorSizeB(), P_::GetVectorSizeC()),
+                      concat('x', P_::kPadM, P_::kPadN, P_::kPadK));
+        // clang-format on
+    }
+
     __host__ static auto GetWorkSpaceSize(const std::vector<GroupedGemmHostArgs>& gemm_descs)
         -> std::size_t
     {

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp

@@ -41,20 +41,26 @@ struct GemmPipelineAgBgCrImplBase
         store_tile(lds_tile_window, block_tile_tmp);
     }
 
+    template <typename DstBlockTile, typename SrcTileWindow>
+    CK_TILE_DEVICE void LocalPrefetch(DstBlockTile& dst_block_tile,
+                                      const SrcTileWindow& lds_tile_window) const
+    {
+        load_tile(dst_block_tile, lds_tile_window);
+    }
+
     CK_TILE_DEVICE auto GetABLdsTensorViews(void* p_smem) const
     {
         // A tile in LDS
-        ADataType* p_a_lds              = static_cast<ADataType*>(p_smem);
+        ADataType* __restrict__ p_a_lds = static_cast<ADataType*>(p_smem);
         constexpr auto a_lds_block_desc = Policy::template MakeALdsBlockDescriptor<Problem>();
         auto a_lds_block = make_tensor_view<address_space_enum::lds>(p_a_lds, a_lds_block_desc);
 
         // TODO: LDS alignment should come from Policy!
-        constexpr index_t a_lds_block_space_size_aligned =
-            integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16) *
-            16;
+        constexpr index_t a_lds_block_space_size_aligned = integer_least_multiple(
+            sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16);
 
         // B tile in LDS
-        BDataType* p_b_lds = static_cast<BDataType*>(
+        BDataType* __restrict__ p_b_lds = static_cast<BDataType*>(
             static_cast<void*>(static_cast<char*>(p_smem) + a_lds_block_space_size_aligned));
         constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
         auto b_lds_block = make_tensor_view<address_space_enum::lds>(p_b_lds, b_lds_block_desc);

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp

@@ -10,6 +10,7 @@
 #include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp"
+#include "ck_tile/host/concat.hpp"
 
 namespace ck_tile {
 
@@ -75,12 +76,23 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
     static constexpr bool kPadN = Problem::kPadN;
     static constexpr bool kPadK = Problem::kPadK;
 
+    static constexpr bool DoubleSmemBuffer = Problem::DoubleSmemBuffer;
+
     static constexpr bool HasHotLoop = Problem::HasHotLoop;
     static constexpr auto TailNum    = Problem::TailNum;
     static constexpr auto Scheduler  = Problem::Scheduler;
 
     using Base::PrefetchStages;
 
+    [[nodiscard]] CK_TILE_HOST static const std::string GetName()
+    {
+        // clang-format off
+        return concat('_', "pipeline_AgBgCrCompV3", BlockSize,
+                      concat('x', GetVectorSizeA(), GetVectorSizeB(),  GetVectorSizeC()),
+                      concat('x', kPadM, kPadN, kPadK));
+        // clang-format on
+    }
+
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
     {
         return Policy::template GetSmemSize<Problem>();

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4_default_policy.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
+#include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
+
+namespace ck_tile {
+// Default policy for GemmPipelineAGmemBGmemCregComputeV4, except the block gemm method, it shares
+// the same vector size implementation, SmemSize, Global memory tile distiribution as the
+// UniversalGemm Pipeline Policy.
+// Default policy class should not be templated, put template on
+// member functions instead.
+struct GemmPipelineAgBgCrCompV4DefaultPolicy
+    : public UniversalGemmBasePolicy<GemmPipelineAgBgCrCompV4DefaultPolicy>
+{
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
+    {
+        using namespace ck_tile;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+        constexpr index_t KPack      = GetSmemPackA<Problem>();
+
+        constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(number<kKPerBlock / KPack>{}, number<kMPerBlock>{}, number<KPack>{}),
+            make_tuple(number<kMPerBlock * KPack>{}, number<KPack>{}, number<1>{}),
+            number<KPack>{},
+            number<1>{});
+
+        constexpr auto a_lds_block_desc = transform_tensor_descriptor(
+            a_lds_block_desc_0,
+            make_tuple(
+                make_pass_through_transform(number<kMPerBlock>{}),
+                make_merge_transform(make_tuple(number<kKPerBlock>{} / KPack, number<KPack>{}))),
+            make_tuple(sequence<1>{}, sequence<0, 2>{}),
+            make_tuple(sequence<0>{}, sequence<1>{}));
+
+        return a_lds_block_desc;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+        constexpr index_t KPack      = GetSmemPackB<Problem>();
+
+        constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(number<kKPerBlock / KPack>{}, number<kNPerBlock>{}, number<KPack>{}),
+            make_tuple(number<(kNPerBlock)*KPack>{}, number<KPack>{}, number<1>{}),
+            number<KPack>{},
+            number<1>{});
+
+        constexpr auto b_lds_block_desc = transform_tensor_descriptor(
+            b_lds_block_desc_0,
+            make_tuple(
+                make_pass_through_transform(number<kNPerBlock>{}),
+                make_merge_transform(make_tuple(number<kKPerBlock / KPack>{}, number<KPack>{}))),
+            make_tuple(sequence<1>{}, sequence<0, 2>{}),
+            make_tuple(sequence<0>{}, sequence<1>{}));
+
+        return b_lds_block_desc;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
+    {
+        using AccDataType     = float;
+        using BlockWarps      = typename Problem::BlockGemmShape::BlockWarps;
+        using WarpTile        = typename Problem::BlockGemmShape::WarpTile;
+        using WarpGemm        = WarpGemmMfmaDispatcher<typename Problem::ADataType,
+                                                typename Problem::BDataType,
+                                                AccDataType,
+                                                WarpTile::at(I0),
+                                                WarpTile::at(I1),
+                                                WarpTile::at(I2),
+                                                Problem::TransposeC>;
+        using BlockGemmPolicy = BlockGemmARegBRegCRegV1CustomPolicy<typename Problem::ADataType,
+                                                                    typename Problem::BDataType,
+                                                                    typename Problem::CDataType,
+                                                                    BlockWarps,
+                                                                    WarpGemm>;
+
+        return BlockGemmARegBRegCRegV1<Problem, BlockGemmPolicy>{};
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp

@@ -7,6 +7,7 @@
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp"
+#include "ck_tile/host/concat.hpp"
 
 namespace ck_tile {
 
@@ -123,11 +124,23 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
     static constexpr bool kPadN = Problem::kPadN;
     static constexpr bool kPadK = Problem::kPadK;
 
+    static constexpr bool DoubleSmemBuffer = Problem::DoubleSmemBuffer;
+
     // Where is the right place for HasHotLoop and TailNum ???
     static constexpr bool HasHotLoop = Problem::HasHotLoop;
     static constexpr auto TailNum    = Problem::TailNum;
     static constexpr auto Scheduler  = Problem::Scheduler;
 
+    [[nodiscard]] CK_TILE_HOST static const std::string GetName()
+    {
+        // clang-format off
+        return concat('_', "pipeline_AgBgCrMe",
+                      concat('x', MPerBlock, NPerBlock, KPerBlock),
+                      concat('x', GetVectorSizeA(), GetVectorSizeB(), GetVectorSizeC()),
+                      concat('x', kPadM, kPadN, kPadK));
+        // clang-format on
+    }
+
     using Base::PrefetchStages;
 
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
 #include <ostream>
+#include <sstream>
 
 #include "ck_tile/core.hpp"
 

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1.hpp

@@ -5,6 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
+#include "ck_tile/host/concat.hpp"
 
 namespace ck_tile {
 
@@ -39,6 +40,21 @@ struct GemmPipelineAGmemBGmemCRegV1
     static constexpr bool kPadN = Problem::kPadN;
     static constexpr bool kPadK = Problem::kPadK;
 
+    static constexpr index_t kLdsAlignmentInBytes = 16;
+
+    [[nodiscard]] CK_TILE_HOST static const std::string GetName()
+    {
+        // clang-format off
+        return concat('_', "pipeline_AGmemBGmemCRegV1", 
+                      concat('x', kMPerBlock, kNPerBlock, kKPerBlock,  BlockSize),
+                      concat('x', GetVectorSizeA(), GetVectorSizeB(), GetVectorSizeC()),
+                      concat('x', kPadM, kPadN, kPadK));
+        // clang-format on
+    }
+
+    // For the basic gemm pipelien DoubleSmemBuffer set to be false naturally.
+    static constexpr bool DoubleSmemBuffer = false;
+
     CK_TILE_HOST_DEVICE static constexpr auto TransposeC() { return Problem::TransposeC; }
 
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
@@ -75,8 +91,9 @@ struct GemmPipelineAGmemBGmemCRegV1
         auto a_lds_block = make_tensor_view<address_space_enum::lds>(p_a_lds, a_lds_block_desc);
 
         constexpr index_t a_lds_block_space_size_aligned =
-            integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16) *
-            16;
+            integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(),
+                                kLdsAlignmentInBytes) *
+            kLdsAlignmentInBytes;
 
         // B tile in LDS
         BDataType* p_b_lds = static_cast<BDataType*>(

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp

@@ -338,7 +338,7 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
     {
         using ALayout   = remove_cvref_t<typename Problem::ALayout>;
         using ADataType = remove_cvref_t<typename Problem::ADataType>;
-        static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>);
+        static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
         constexpr index_t kBlockSize = Problem::kBlockSize;
         constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
         constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2.hpp

@@ -5,6 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp"
+#include "ck_tile/host/concat.hpp"
 
 namespace ck_tile {
 
@@ -25,6 +26,13 @@ struct GemmPipelineAGmemBGmemCRegV2
     static constexpr index_t kNPerBlock = BlockGemmShape::kN;
     static constexpr index_t kKPerBlock = BlockGemmShape::kK;
 
+    [[nodiscard]] CK_TILE_HOST static const std::string GetName()
+    {
+        // clang-format off
+        return concat('_', "pipeline_AGmemBGmemCRegV2",
+                      concat('x', kMPerBlock, kNPerBlock, kKPerBlock, kBlockSize));
+        // clang-format on
+    }
     CK_TILE_HOST_DEVICE static constexpr auto TransposeC() { return Problem::TransposeC; }
 
     CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_problem.hpp

@@ -5,6 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
+#include "ck_tile/host/concat.hpp"
 
 namespace ck_tile {
 
@@ -35,9 +36,21 @@ struct GemmPipelineProblemBase
     static constexpr bool kPadN = Traits::kPadN;
     static constexpr bool kPadK = Traits::kPadK;
 
-    static constexpr auto Scheduler = GemmPipelineScheduler::Default;
+    static constexpr bool DoubleSmemBuffer = Traits::DoubleSmemBuffer;
 
+    static constexpr auto Scheduler         = GemmPipelineScheduler::Default;
     static constexpr index_t VectorLoadSize = Traits::_VectorSize;
+
+    [[nodiscard]] CK_TILE_HOST static const std::string GetName()
+    {
+        // clang-format off
+        return concat('_', "gemm_problem", 
+                      concat('x', VectorLoadSize, kBlockSize),
+                      concat('x', kPadM, kPadN, kPadK),
+                      Scheduler);
+        // clang-format on
+    }
+
     CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentA()
     {
         if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
@@ -162,6 +175,8 @@ struct UniversalGemmPipelineProblem
     static constexpr bool kPadN = Traits::kPadN;
     static constexpr bool kPadK = Traits::kPadK;
 
+    static constexpr bool DoubleSmemBuffer = Traits::DoubleSmemBuffer;
+
     static constexpr auto Scheduler  = Scheduler_;
     static constexpr auto HasHotLoop = HasHotLoop_;
     static constexpr auto TailNum    = TailNum_;

include/ck_tile/ops/gemm/pipeline/tile_gemm_shape.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
 #include "ck_tile/core.hpp"
+#include "ck_tile/host/concat.hpp"
 
 namespace ck_tile {
 
@@ -19,6 +20,16 @@ struct TileGemmShape
     static constexpr index_t kM = BlockTile::at(number<0>{});
     static constexpr index_t kN = BlockTile::at(number<1>{});
     static constexpr index_t kK = BlockTile::at(number<2>{});
+
+    CK_TILE_HOST static std::string GetName()
+    {
+        // clang-format off
+        return concat('_', "tile_gemm_shape",
+                      concat('x', kM, kN, kK, NumWarps),
+                      concat('x', BlockWarps::at(number<0>{}), BlockWarps::at(number<1>{}), BlockWarps::at(number<2>{})),
+                      concat('x', (WarpTile::at(number<0>{})), WarpTile::at(number<1>{}), WarpTile::at(number<2>{})));
+        // clang-format on
+    }
 };
 
 } // namespace ck_tile

include/ck_tile/ops/gemm/pipeline/tile_gemm_traits.hpp

@@ -32,6 +32,7 @@ struct TileGemmTraits
 template <bool kPadM_,
           bool kPadN_,
           bool kPadK_,
+          bool DoubleSmemBuffer_,
           typename ALayout_,
           typename BLayout_,
           typename CLayout_,
@@ -42,6 +43,8 @@ struct TileGemmUniversalTraits
     static constexpr bool kPadN = kPadN_;
     static constexpr bool kPadK = kPadK_;
 
+    static constexpr bool DoubleSmemBuffer = DoubleSmemBuffer_;
+
     using ALayout = ALayout_;
     using BLayout = BLayout_;
     using CLayout = CLayout_;

include/ck_tile/ops/image_to_column.hpp

@@ -8,3 +8,4 @@
 #include "ck_tile/ops/image_to_column/pipeline/tile_image_to_column_shape.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/common/utils.hpp"

include/ck_tile/ops/layernorm2d.hpp

@@ -11,3 +11,4 @@
 #include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_traits.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/common/utils.hpp"

include/ck_tile/ops/norm_reduce.hpp

@@ -8,3 +8,4 @@
 #include "ck_tile/ops/norm_reduce/thread/thread_welford.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/common/utils.hpp"

include/ck_tile/ops/permute.hpp

@@ -7,3 +7,4 @@
 #include "ck_tile/ops/permute/pipeline/generic_petmute_problem.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/common/utils.hpp"

include/ck_tile/ops/reduce.hpp

@@ -9,3 +9,4 @@
 #include "ck_tile/ops/reduce/block/block_reduce2d_problem.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/common/utils.hpp"