Merge remote-tracking branch 'origin/develop' into mozga-amd/universal_gemm_weight

example/ck_tile/01_fmha/codegen/ops/fmha_fwd_appendkv.py

@@ -46,9 +46,7 @@ using fmha_pipeline_problem_{F_idx} = ck_tile::BlockFmhaFwdAppendKVPipelineProbl
 using fmha_pipeline_{F_idx} = ck_tile::BlockFmhaFwdAppendKVPipeline<
     fmha_pipeline_problem_{F_idx}>;
 
-using fmha_kernel_{F_idx} =
-    ck_tile::FmhaFwdAppendKVKernel<ck_tile::FmhaFwdAppendKVTilePartitioner<{F_bs}, {F_bsk}, {F_bd}, {F_bdv}>,
-                  fmha_pipeline_{F_idx}>;
+using fmha_kernel_{F_idx} = ck_tile::FmhaFwdAppendKVKernel<fmha_pipeline_{F_idx}>;
 
 using trait_{F_idx} = fmha_fwd_appendkv_traits_<{F_hdim}, {F_dtype}, {F_bs}, {F_bsk}, {F_bd}, {F_bdv}, {F_vlayout},
                         {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}, {F_rope}, {F_pagedkv}>;

example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py

@@ -96,9 +96,7 @@ using fmha_epilogue =
                                            {F_spad}, {F_dvpad}>>;
 
 using fmha_kernel =
-    ck_tile::FmhaFwdSplitKVKernel<ck_tile::FmhaFwdSplitKVTilePartitioner<fmha_shape>,
-                  fmha_pipeline,
-                  fmha_epilogue>;
+    ck_tile::FmhaFwdSplitKVKernel<fmha_pipeline, fmha_epilogue>;
 
 static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
 {{
@@ -176,11 +174,7 @@ using fmha_epilogue =
                                            false, false>>;
 
 using fmha_kernel =
-    ck_tile::FmhaFwdSplitKVCombineKernel<
-        ck_tile::FmhaFwdSplitKVCombineTilePartitioner<
-            fmha_pipeline_problem::kM0, fmha_pipeline_problem::kN1>,
-        fmha_pipeline,
-        fmha_epilogue>;
+    ck_tile::FmhaFwdSplitKVCombineKernel<fmha_pipeline, fmha_epilogue>;
 
 static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
 {{
@@ -261,7 +255,7 @@ FMHA_FWD_SPLITKV_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F
                 static_assert({F_bn1} % 32 == 0);
 
                 if (t.has_lse) {{
-                    if constexpr (std::is_same_v<{F_dtype}, ck_tile::fp8_t>) {{
+                    if constexpr (std::is_same_v<{F_dtype}, FmhaFwdFp8>) {{
                         return -1;
                     }} else {{
                         using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, /*F_bn1=*/32, true, {F_squant}, {F_spad}, {F_dvpad}>;

example/ck_tile/01_fmha/fmha_fwd.cpp

@@ -1140,6 +1140,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
         naive_t.v_layout   = i_perm == 1 ? "bhsd" : "bshd";
         naive_t.o_layout   = o_perm == 1 ? "bhsd" : "bshd";
         naive_t.variation  = 0; // TODO?
+        naive_t.quant_algo = 0;
 
         ck_tile::DeviceMem o_naive_buf(o_host.get_element_space_size_in_bytes());
 

example/ck_tile/01_fmha/fmha_fwd.hpp

@@ -400,8 +400,18 @@ auto fmha_fwd_create_kargs_and_grids(fmha_fwd_args args)
         }
     }();
 
-    dim3 grids = FmhaKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v);
+    if constexpr(FmhaKernel::kIsGroupMode)
+    {
+        dim3 grids = FmhaKernel::GridSize(
+            args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v, args.seqlen_k_ptr != nullptr);
+        return ck_tile::make_tuple(kargs, grids);
+    }
+    else
+    {
+        dim3 grids =
+            FmhaKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v, false);
         return ck_tile::make_tuple(kargs, grids);
+    }
 }
 
 template <typename Kernel>

example/ck_tile/02_layernorm2d/script/smoke_test.sh

@@ -27,7 +27,8 @@ $EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7   -n=2734
 $EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=3182
 $EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=9   -n=4096
 $EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=8192
-#$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=10547
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=9120
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=10547
 #$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=17134
 done
 done

example/ck_tile/03_gemm/gemm_basic.hpp

@@ -54,8 +54,7 @@ using CDataType   = Types::CDataType;
 auto create_args(int argc, char* argv[])
 {
     ck_tile::ArgParser arg_parser;
-    arg_parser.insert("b", "1", "batch size")
-        .insert("m", "3840", "m dimension")
+    arg_parser.insert("m", "3840", "m dimension")
         .insert("n", "4096", "n dimension")
         .insert("k", "2048", "k dimension")
         .insert("a_layout", "R", "A tensor data layout - Row by default")
@@ -68,7 +67,8 @@ auto create_args(int argc, char* argv[])
         .insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
         .insert("warmup", "50", "number of iterations before benchmark the kernel")
         .insert("repeat", "100", "number of iterations to benchmark the kernel")
-        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer");
+        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
+        .insert("split_k", "1", "splitK value");
 
     bool result = arg_parser.parse(argc, argv);
     return std::make_tuple(result, arg_parser);

example/ck_tile/03_gemm/run_gemm_example.inc

@@ -64,7 +64,7 @@ int run_gemm_example_with_layouts(int argc,
     ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
     ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
 
-    ck_tile::index_t batch_size = arg_parser.get_int("b");
+    ck_tile::index_t kbatch = arg_parser.get_int("split_k");
     int n_warmup            = arg_parser.get_int("warmup");
     int n_repeat            = arg_parser.get_int("repeat");
 
@@ -133,7 +133,7 @@ int run_gemm_example_with_layouts(int argc,
                                            stride_A,
                                            stride_B,
                                            stride_C,
-                                           batch_size,
+                                           kbatch,
                                            n_warmup,
                                            n_repeat);
 

example/ck_tile/03_gemm/universal_gemm.cpp

@@ -22,7 +22,7 @@
 #endif
 
 template <typename ALayout, typename BLayout, typename CLayout>
-float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
+float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
 {
 #if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
     // Memory friendly for Interwave scheduler
@@ -78,7 +78,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
 #endif
         ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
 
-    const ck_tile::index_t num_loop    = TilePartitioner::GetLoopNum(args.K);
+    const ck_tile::index_t k_grain     = args.k_batch * K_Tile;
+    const ck_tile::index_t K_split     = (args.K + k_grain - 1) / k_grain * K_Tile;
+    const ck_tile::index_t num_loop    = TilePartitioner::GetLoopNum(K_split);
     const bool has_hot_loop            = BaseGemmPipeline::BlockHasHotloop(num_loop);
     const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
 
@@ -106,17 +108,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
                                                   has_hot_loop_v,
                                                   tail_number_v>>;
         using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
-        auto kargs   = Kernel::MakeKargs(args.p_a,
-                                       args.p_b,
-                                       args.p_c,
-                                       args.M,
-                                       args.N,
-                                       args.K,
-                                       args.stride_A,
-                                       args.stride_B,
-                                       args.stride_C);
-
-        const dim3 grids      = Kernel::GridSize(args.M, args.N, args.kbatch);
+        auto kargs   = Kernel::MakeKernelArgs(args);
+
+        const dim3 grids      = Kernel::GridSize(args.M, args.N, args.k_batch);
         constexpr dim3 blocks = Kernel::BlockSize();
 
         if(!Kernel::IsSupportedArgument(kargs))

example/ck_tile/16_batched_gemm/batched_gemm.cpp

@@ -70,20 +70,25 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
 
     using CodegenGemmTraits =
         ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
-
     using CodegenPipelineProblem = ck_tile::
         GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
-
-    using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
+    using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
+    using CodegenGemmPipeline =
+        ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
     // ToDo: Will add the codegen part to test different pipeline policies in GEMM.
     // Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
     using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
 
     auto kargs = Kernel::MakeKernelArgs(args);
 
-    const dim3 grids      = Kernel::GridSize(args.M, args.N, args.batch_count);
+    const dim3 grids      = Kernel::GridSize(args.M, args.N, args.k_batch, args.batch_count);
     constexpr dim3 blocks = Kernel::BlockSize();
 
+    if(!Kernel::IsSupportedArgument(kargs))
+    {
+        throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
+    }
+
     if(s.log_level_ > 0)
     {
         std::cout << "Launching kernel with args:"

example/ck_tile/16_batched_gemm/batched_gemm.hpp

@@ -49,7 +49,8 @@ auto create_args(int argc, char* argv[])
         .insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
         .insert("warmup", "50", "number of iterations before benchmark the kernel")
         .insert("repeat", "100", "number of iterations to benchmark the kernel")
-        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer");
+        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
+        .insert("split_k", "1", "splitK value");
 
     bool result = arg_parser.parse(argc, argv);
     return std::make_tuple(result, arg_parser);

example/ck_tile/16_batched_gemm/run_batched_gemm_example.inc

@@ -17,6 +17,7 @@ float invoke_batched_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
                           ck_tile::index_t batch_stride_B,
                           ck_tile::index_t batch_stride_C,
                           ck_tile::index_t batch_count,
+                          ck_tile::index_t kbatch,
                           int n_warmup,
                           int n_repeat)
 {
@@ -24,6 +25,7 @@ float invoke_batched_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
     args.a_ptr          = a_m_k_dev_buf.GetDeviceBuffer();
     args.b_ptr          = b_k_n_dev_buf.GetDeviceBuffer();
     args.c_ptr          = c_m_n_dev_buf.GetDeviceBuffer();
+    args.k_batch        = kbatch;
     args.M              = M;
     args.N              = N;
     args.K              = K;
@@ -79,6 +81,7 @@ int run_batched_gemm_example_with_layouts(int argc,
     ck_tile::index_t batch_stride_B = arg_parser.get_int("batch_stride_b");
     ck_tile::index_t batch_stride_C = arg_parser.get_int("batch_stride_c");
     ck_tile::index_t batch_count    = arg_parser.get_int("batch_count");
+    ck_tile::index_t kbatch         = arg_parser.get_int("split_k");
 
     int n_warmup = arg_parser.get_int("warmup");
     int n_repeat = arg_parser.get_int("repeat");
@@ -159,6 +162,7 @@ int run_batched_gemm_example_with_layouts(int argc,
                                                    batch_stride_B,
                                                    batch_stride_C,
                                                    batch_count,
+                                                   kbatch,
                                                    n_warmup,
                                                    n_repeat);
 

include/ck/library/utility/host_tensor.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -44,10 +44,19 @@ std::ostream& LogRangeAsType(std::ostream& os, Range&& range, std::string delim)
         else
             os << delim;
 
-        if constexpr(std::is_same_v<T, ck::f8_t> || std::is_same_v<T, ck::bf8_t>)
+        using RangeType = ck::remove_cvref_t<decltype(v)>;
+        if constexpr(std::is_same_v<RangeType, ck::f8_t> || std::is_same_v<RangeType, ck::bf8_t> ||
+                     std::is_same_v<RangeType, ck::bhalf_t>)
         {
             os << ck::type_convert<float>(v);
         }
+        else if constexpr(std::is_same_v<RangeType, ck::pk_i4_t>)
+        {
+            const auto packed_floats = ck::type_convert<ck::float2_t>(v);
+            const ck::vector_type<float, 2> vector_of_floats{packed_floats};
+            os << vector_of_floats.template AsType<float>()[ck::Number<0>{}] << delim
+               << vector_of_floats.template AsType<float>()[ck::Number<1>{}];
+        }
         else
         {
             os << static_cast<T>(v);
@@ -266,18 +275,18 @@ struct Tensor
     using Data       = std::vector<T>;
 
     template <typename X>
-    Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(mDesc.GetElementSpaceSize())
+    Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(GetElementSpaceSize())
     {
     }
 
     template <typename X, typename Y>
     Tensor(std::initializer_list<X> lens, std::initializer_list<Y> strides)
-        : mDesc(lens, strides), mData(mDesc.GetElementSpaceSize())
+        : mDesc(lens, strides), mData(GetElementSpaceSize())
     {
     }
 
     template <typename Lengths>
-    Tensor(const Lengths& lens) : mDesc(lens), mData(mDesc.GetElementSpaceSize())
+    Tensor(const Lengths& lens) : mDesc(lens), mData(GetElementSpaceSize())
     {
     }
 
@@ -287,7 +296,7 @@ struct Tensor
     {
     }
 
-    Tensor(const Descriptor& desc) : mDesc(desc), mData(mDesc.GetElementSpaceSize()) {}
+    Tensor(const Descriptor& desc) : mDesc(desc), mData(GetElementSpaceSize()) {}
 
     template <typename OutT>
     Tensor<OutT> CopyAsType() const
@@ -322,7 +331,17 @@ struct Tensor
 
     std::size_t GetElementSize() const { return mDesc.GetElementSize(); }
 
-    std::size_t GetElementSpaceSize() const { return mDesc.GetElementSpaceSize(); }
+    std::size_t GetElementSpaceSize() const
+    {
+        if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
+        {
+            return (mDesc.GetElementSpaceSize() + 1) / 2;
+        }
+        else
+        {
+            return mDesc.GetElementSpaceSize();
+        }
+    }
 
     std::size_t GetElementSpaceSizeInBytes() const { return sizeof(T) * GetElementSpaceSize(); }
 
@@ -468,31 +487,66 @@ struct Tensor
 
     template <typename... Is>
     std::size_t GetOffsetFromMultiIndex(Is... is) const
+    {
+        if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
+        {
+            return mDesc.GetOffsetFromMultiIndex(is...) / 2;
+        }
+        else
         {
             return mDesc.GetOffsetFromMultiIndex(is...);
         }
+    }
 
     template <typename... Is>
     T& operator()(Is... is)
+    {
+        if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
+        {
+            return mData[mDesc.GetOffsetFromMultiIndex(is...) / 2];
+        }
+        else
         {
             return mData[mDesc.GetOffsetFromMultiIndex(is...)];
         }
+    }
 
     template <typename... Is>
     const T& operator()(Is... is) const
+    {
+        if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
+        {
+            return mData[mDesc.GetOffsetFromMultiIndex(is...) / 2];
+        }
+        else
         {
             return mData[mDesc.GetOffsetFromMultiIndex(is...)];
         }
+    }
 
     T& operator()(std::vector<std::size_t> idx)
+    {
+        if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
+        {
+            return mData[mDesc.GetOffsetFromMultiIndex(idx) / 2];
+        }
+        else
         {
             return mData[mDesc.GetOffsetFromMultiIndex(idx)];
         }
+    }
 
     const T& operator()(std::vector<std::size_t> idx) const
+    {
+        if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
+        {
+            return mData[mDesc.GetOffsetFromMultiIndex(idx) / 2];
+        }
+        else
         {
             return mData[mDesc.GetOffsetFromMultiIndex(idx)];
         }
+    }
 
     typename Data::iterator begin() { return mData.begin(); }
 

include/ck/library/utility/host_tensor_generator.hpp

@@ -81,6 +81,20 @@ struct GeneratorTensor_1<int8_t>
     }
 };
 
+template <>
+struct GeneratorTensor_1<ck::pk_i4_t>
+{
+    int8_t value = 1;
+
+    template <typename... Is>
+    ck::pk_i4_t operator()(Is...)
+    {
+        int t         = value + 8;
+        ck::pk_i4_t r = ((t << 4) + t) & 0xff;
+        return r;
+    }
+};
+
 template <typename T>
 struct GeneratorTensor_2
 {
@@ -121,6 +135,22 @@ struct GeneratorTensor_2<int8_t>
     }
 };
 
+template <>
+struct GeneratorTensor_2<ck::pk_i4_t>
+{
+    int min_value = 0;
+    int max_value = 1;
+
+    template <typename... Is>
+    ck::pk_i4_t operator()(Is...)
+    {
+        int hi        = std::rand() % (max_value - min_value) + min_value + 8;
+        int lo        = std::rand() % (max_value - min_value) + min_value + 8;
+        ck::pk_i4_t r = ((hi << 4) + lo) & 0xff;
+        return r;
+    }
+};
+
 #if defined CK_ENABLE_FP8
 template <>
 struct GeneratorTensor_2<ck::f8_t>

include/ck/tensor/static_tensor.hpp

@@ -167,7 +167,7 @@ struct StaticTensorTupleOfVectorBuffer
     // Idx is for S, not X. Idx should be aligned with X
     template <typename X,
               typename Idx,
-              typename enable_if<has_same_scalar_type<S, X>::value &&
+              typename enable_if<(has_same_scalar_type<S, X>::value || !is_native_type<S>()) &&
                                      is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
                                  bool>::type = false>
     __host__ __device__ constexpr X GetAsType(Idx) const
@@ -201,7 +201,7 @@ struct StaticTensorTupleOfVectorBuffer
     // Idx is for S, not X. Idx should be aligned with X
     template <typename X,
               typename Idx,
-              typename enable_if<has_same_scalar_type<S, X>::value &&
+              typename enable_if<(has_same_scalar_type<S, X>::value || !is_native_type<S>()) &&
                                      is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
                                  bool>::type = false>
     __host__ __device__ constexpr void SetAsType(Idx, X x)

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp"
+
+namespace ck {
+
+enum struct BlockGemmPipelineVersion
+{
+    v1, // Naive
+    v2, // Mem
+    v3, // Comp
+    v4, // Comp, double lds buffer
+    v5, // Comp, double global prefetch register buffer
+};
+
+template <BlockGemmPipelineVersion BlkGemmPipelineVer,
+          BlockGemmPipelineScheduler BlkGemmPipeSche,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+constexpr auto BlockGemmPipeline_Selector()
+{
+    if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+    {
+        return BlockwiseGemmXdlops_pipeline_v1_b_scale<BlkGemmPipeSche,
+                                                       BlockSize,
+                                                       ADataType,
+                                                       BDataType,
+                                                       ComputeDataType,
+                                                       AccDataType,
+                                                       ATileDesc,
+                                                       BTileDesc,
+                                                       AMmaTileDesc,
+                                                       BMmaTileDesc,
+                                                       ABlockTransferSrcScalarPerVector,
+                                                       BBlockTransferSrcScalarPerVector,
+                                                       MPerBlock,
+                                                       NPerBlock,
+                                                       KPerBlock,
+                                                       MPerXDL,
+                                                       NPerXDL,
+                                                       MRepeat,
+                                                       NRepeat,
+                                                       KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
+    {
+        return BlockwiseGemmXdlops_pipeline_v2_b_scale<BlkGemmPipeSche,
+                                                       BlockSize,
+                                                       ADataType,
+                                                       BDataType,
+                                                       ComputeDataType,
+                                                       AccDataType,
+                                                       ATileDesc,
+                                                       BTileDesc,
+                                                       AMmaTileDesc,
+                                                       BMmaTileDesc,
+                                                       ABlockTransferSrcScalarPerVector,
+                                                       BBlockTransferSrcScalarPerVector,
+                                                       MPerBlock,
+                                                       NPerBlock,
+                                                       KPerBlock,
+                                                       MPerXDL,
+                                                       NPerXDL,
+                                                       MRepeat,
+                                                       NRepeat,
+                                                       KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+    {
+        return BlockwiseGemmXdlops_pipeline_v3_b_scale<BlkGemmPipeSche,
+                                                       BlockSize,
+                                                       ADataType,
+                                                       BDataType,
+                                                       ComputeDataType,
+                                                       AccDataType,
+                                                       ATileDesc,
+                                                       BTileDesc,
+                                                       AMmaTileDesc,
+                                                       BMmaTileDesc,
+                                                       ABlockTransferSrcScalarPerVector,
+                                                       BBlockTransferSrcScalarPerVector,
+                                                       MPerBlock,
+                                                       NPerBlock,
+                                                       KPerBlock,
+                                                       MPerXDL,
+                                                       NPerXDL,
+                                                       MRepeat,
+                                                       NRepeat,
+                                                       KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
+    {
+        return BlockwiseGemmXdlops_pipeline_v4_b_scale<BlkGemmPipeSche,
+                                                       BlockSize,
+                                                       ADataType,
+                                                       BDataType,
+                                                       ComputeDataType,
+                                                       AccDataType,
+                                                       ATileDesc,
+                                                       BTileDesc,
+                                                       AMmaTileDesc,
+                                                       BMmaTileDesc,
+                                                       ABlockTransferSrcScalarPerVector,
+                                                       BBlockTransferSrcScalarPerVector,
+                                                       MPerBlock,
+                                                       NPerBlock,
+                                                       KPerBlock,
+                                                       MPerXDL,
+                                                       NPerXDL,
+                                                       MRepeat,
+                                                       NRepeat,
+                                                       KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v5)
+    {
+        return BlockwiseGemmXdlops_pipeline_v5<BlkGemmPipeSche,
+                                               BlockSize,
+                                               ADataType,
+                                               BDataType,
+                                               ComputeDataType,
+                                               AccDataType,
+                                               ATileDesc,
+                                               BTileDesc,
+                                               AMmaTileDesc,
+                                               BMmaTileDesc,
+                                               ABlockTransferSrcScalarPerVector,
+                                               BBlockTransferSrcScalarPerVector,
+                                               MPerBlock,
+                                               NPerBlock,
+                                               KPerBlock,
+                                               MPerXDL,
+                                               NPerXDL,
+                                               MRepeat,
+                                               NRepeat,
+                                               KPack>{};
+    }
+    else
+    {
+        std::cerr << "BlockGemmPipeline configuration is not available" << std::endl;
+    }
+}
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_b_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Naive pipeline with lowest resource request per WGP
+// GlobalPrefetchStages: 1
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 0
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPacks>
+struct BlockwiseGemmXdlops_pipeline_v1_b_scale
+{
+};
+
+template <index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack
+          // ,bool TransposeC //disable transposec right now...
+          >
+struct BlockwiseGemmXdlops_pipeline_v1_b_scale<BlockGemmPipelineScheduler::Intrawave,
+                                               BlockSize,
+                                               ADataType,
+                                               BDataType,
+                                               ComputeDataType,
+                                               AccDataType,
+                                               ATileDesc,
+                                               BTileDesc,
+                                               AMmaTileDesc,
+                                               BMmaTileDesc,
+                                               ABlockTransferSrcScalarPerVector,
+                                               BBlockTransferSrcScalarPerVector,
+                                               MPerBlock,
+                                               NPerBlock,
+                                               KPerBlock,
+                                               MPerXDL,
+                                               NPerXDL,
+                                               MRepeat,
+                                               NRepeat,
+                                               KPack>
+    : BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                        ADataType,
+                                        BDataType,
+                                        ComputeDataType,
+                                        AccDataType,
+                                        ATileDesc,
+                                        BTileDesc,
+                                        AMmaTileDesc,
+                                        BMmaTileDesc,
+                                        ABlockTransferSrcScalarPerVector,
+                                        BBlockTransferSrcScalarPerVector,
+                                        MPerBlock,
+                                        NPerBlock,
+                                        KPerBlock,
+                                        MPerXDL,
+                                        NPerXDL,
+                                        MRepeat,
+                                        NRepeat,
+                                        KPack>
+
+{
+    using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>;
+    using Base::I0;
+    using Base::KRepeat;
+    using Base::xdlops_gemm;
+
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::CalculateCThreadOriginDataIndex8D;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::b_block_desc_n0_n1_n2_k;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+
+    static constexpr index_t PrefetchStages  = 1;
+    static constexpr index_t PrefillStages   = 1;
+    static constexpr index_t GlobalBufferNum = 1;
+
+    __host__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        ignore = num_loop;
+        return TailNumber::Full;
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer,
+              // BScale Thread Copy
+              typename BScaleGridBuffer,
+              typename BScaleGridDesc,
+              typename BScaleThreadDesc,
+              typename BScaleThreadTransfer,
+              typename BScaleThreadTransferStep>
+    __device__ void Run(
+        // ABlockCopy
+        const AGridDesc& a_grid_desc,
+        const ABlockDesc& a_block_desc,
+        ABlockTransfer& a_blockwise_copy,
+        const AGridBuffer& a_grid_buf,
+        ABlockBuffer& a_block_buf,
+        const ABlockTransferStep& a_block_copy_step,
+        // BBlockCopy
+        const BGridDesc& b_grid_desc,
+        const BBlockDesc& b_block_desc,
+        BBlockTransfer& b_blockwise_copy,
+        const BGridBuffer& b_grid_buf,
+        BBlockBuffer& b_block_buf,
+        const BBlockTransferStep& b_block_copy_step,
+        // CThread
+        CThreadBuffer& c_thread_buf,
+        // BScaleThreadCopy
+        const BScaleGridDesc& b_scale_grid_desc,
+        const BScaleThreadDesc& b_scale_thread_desc,
+        BScaleThreadTransfer& b_scale_thread_copy,
+        const BScaleGridBuffer& b_scale_grid_buf,
+        const BScaleThreadTransferStep& b_scale_thread_copy_step,
+        // num_loop
+        index_t num_loop,
+        index_t num_loop_per_scale) const
+    {
+        // assume kperblock = scaleblockk
+        ignore            = num_loop_per_scale;
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_scale_thread_desc.GetElementSpaceSize());
+
+        // Global prefetch 1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            b_scale_thread_copy.Run(b_scale_grid_desc,
+                                    b_scale_grid_buf,
+                                    b_scale_thread_desc,
+                                    make_tuple(n0, I0),
+                                    b_scale_thread_buf);
+
+            b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                                   b_scale_thread_copy_step.At(Number<0>{}));
+        });
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               b_scale_thread_copy_step.At(Number<1>{}));
+
+        // Local prefill 1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                // -------------------------------------------------------------------------------------------
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                block_sync_lds();
+                static_for<0, KRepeat, 1>{}([&](auto k) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, I0),
+                                           a_thread_buf);
+                    });
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf,
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k, I0),
+                                           b_thread_buf);
+                    });
+                });
+
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        c_thread_buf_per_scale.Clear();
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            vector_type<ComputeDataType, KPack> a_thread_vec;
+                            vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, I0, k0, ik))>{}];
+                                b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, I0, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<ComputeDataType,
+                                                     xdlops_gemm.K1PerXdlops>::type;
+
+                            xdlops_gemm.template Run<>(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                        });
+                        static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
+                            c_thread_buf(Number<c_offset>{}) +=
+                                c_thread_buf_per_scale[Number<t>{}] *
+                                type_convert<AccDataType>(b_scale_thread_buf[n0]);
+                        });
+                    });
+                });
+
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc,
+                                            make_tuple(n0, I0),
+                                            b_scale_thread_buf);
+
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
+                });
+
+                b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                                       b_scale_thread_copy_step.At(Number<1>{}));
+
+                block_sync_lds();
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                i += 1;
+
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        if constexpr(TailNum == TailNumber::Full)
+        {
+            block_sync_lds();
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, I0),
+                                       a_thread_buf);
+                });
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(n0, I0, k, I0),
+                                       b_thread_buf);
+                });
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    c_thread_buf_per_scale.Clear();
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                    });
+                    static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
+                        c_thread_buf(Number<c_offset>{}) +=
+                            c_thread_buf_per_scale[Number<t>{}] *
+                            type_convert<AccDataType>(b_scale_thread_buf[n0]);
+                    });
+                });
+            });
+        }
+    }
+
+    protected:
+    using Base::a_thread_copy_;
+    using Base::a_thread_desc_;
+    using Base::b_thread_copy_;
+    using Base::b_thread_desc_;
+    using Base::c_thread_desc_;
+};
+
+} // namespace ck