Merge branch 'develop' into muozturk_bf16fp8_streamk

include/ck_tile/ops/common/utils.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <string>
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+// clang-format off
+template <typename T> struct typeToStr;
+template <> struct typeToStr<float> { static constexpr const char * name = "fp32"; };
+template <> struct typeToStr<fp16_t> { static constexpr const char * name = "fp16"; };
+template <> struct typeToStr<bf16_t> { static constexpr const char * name = "bf16"; };
+template <> struct typeToStr<fp8_t> { static constexpr const char * name = "fp8"; };
+template <> struct typeToStr<bf8_t> { static constexpr const char * name = "bf8"; };
+template <> struct typeToStr<int8_t> { static constexpr const char * name = "int8"; };
+// clang-format on
+
+template <typename ADataType_, typename BDataType_>
+std::string gemm_prec_str()
+{
+    std::string base_str = std::string(typeToStr<ADataType_>::name);
+    if(!std::is_same_v<ADataType_, BDataType_>)
+    {
+        base_str += "_" + std::string(typeToStr<BDataType_>::name);
+    }
+    return base_str;
+}
+
+} // namespace ck_tile

include/ck_tile/ops/elementwise.hpp

@@ -6,3 +6,4 @@
 #include "ck_tile/ops/elementwise/unary_element_wise_operation.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/epilogue.hpp

@@ -8,3 +8,4 @@
 #include "ck_tile/ops/epilogue/dynamic_quant_epilogue.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/flatmm.hpp

@@ -9,3 +9,4 @@
 #include "ck_tile/ops/flatmm/block/flatmm_uk_config.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/fmha.hpp

@@ -44,3 +44,4 @@
 #include "ck_tile/ops/fmha/pipeline/tile_fmha_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/fused_moe.hpp

@@ -7,6 +7,7 @@
 #include "ck_tile/ops/fused_moe/kernel/fused_moegemm_shape.hpp"
 #include "ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp"
 #include "ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp"
+#include "ck_tile/ops/fused_moe/kernel/moe_sorting_problem.hpp"
 #include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp"
 #include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp"
 #include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp"
@@ -14,6 +15,6 @@
 #include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp"
 #include "ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp"
 #include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
-#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp"
 #include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/common/utils.hpp"

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

@@ -22,7 +22,7 @@
 //  (only for reference)    exp-0  exp-1     exp-2   exp-3          exp-4  exp-5
 // weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]]
 //
-// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1)
+// max_num_tokens_padded : topk * input_tokens + num_experts * M_a - topk (updated)
 // * this could be larger than actual, since actual tokens are on GPU
 //
 // sorted_token_ids_ptr   : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5]

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

@@ -25,4 +25,28 @@ struct MoeSortingProblem
         InternalLoadUnroll_;                           // TODO: need better design(like tile size)
     static constexpr index_t ExpertTile = ExpertTile_; // TODO: only used in store out
 };
+
+template <typename IndexType_,
+          typename WeightType_,
+          index_t SubTokenTile_,    // 1,2,4,8, or 0 in the future
+          bool SubTokenOneShot_,    // if we only loop over once or not
+          bool LocalExpertMasking_, // used in EP case
+          bool SkipExpertsWithZeroTokens_ = true,
+          index_t ExpertTile_             = 0>
+struct MoeSortingProblemEx
+{
+    // TODO: this kernel only support warp per row
+    using WeightType = remove_cvref_t<WeightType_>;
+    using IndexType  = remove_cvref_t<IndexType_>;
+
+    static constexpr index_t WarpSize               = get_warp_size();
+    static constexpr index_t WarpsPerBlock          = 1;
+    static constexpr index_t SubTokenTile           = SubTokenTile_;
+    static constexpr bool SubTokenOneShot           = SubTokenOneShot_;
+    static constexpr bool LocalExpertMasking        = LocalExpertMasking_;
+    static constexpr bool SkipExpertsWithZeroTokens = SkipExpertsWithZeroTokens_;
+    static_assert(SubTokenTile == 1 || SubTokenTile == 2 || SubTokenTile == 4 || SubTokenTile == 8);
+    static constexpr index_t ExpertTile = ExpertTile_; // TODO: only used in store out
+};
+
 } // namespace ck_tile

include/ck_tile/ops/gemm.hpp

@@ -46,3 +46,4 @@
 #include "ck_tile/ops/gemm/warp/warp_gemm_impl.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/gemm/kernel/batched_gemm_kernel.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 "ck_tile/ops/gemm/kernel/gemm_kernel.hpp"
+#include "ck_tile/ops/common.hpp"
+#include "ck_tile/host/concat.hpp"
 
 namespace ck_tile {
 
@@ -57,6 +59,18 @@ struct BatchedGemmKernel : public GemmKernel<TilePartitioner_, GemmPipeline_, Ep
     using BLayout          = typename Base::BLayout;
     using CLayout          = typename Base::CLayout;
 
+    [[nodiscard]] CK_TILE_HOST static const std::string GetName()
+    {
+        // clang-format off
+        using P_ = GemmPipeline;
+
+        return concat('_', "gemm_batched", 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
+    }
+
     struct BatchedGemmKernelArgs : GemmKernelArgs
     {
         index_t batch_stride_A;

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);

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_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 {
 
@@ -81,6 +82,15 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
 
     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_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 {
 
@@ -128,6 +129,16 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
     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 +40,18 @@ 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
+    }
+
     CK_TILE_HOST_DEVICE static constexpr auto TransposeC() { return Problem::TransposeC; }
 
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
@@ -75,8 +88,9 @@ 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_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 +36,19 @@ struct GemmPipelineProblemBase
     static constexpr bool kPadN = Traits::kPadN;
     static constexpr bool kPadK = Traits::kPadK;
 
-    static constexpr auto Scheduler = GemmPipelineScheduler::Default;
-
+    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>)

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