Merge remote-tracking branch 'upstream/develop' into gemm-layernorm-4

include/ck/tensor_operation/gpu/device/device_pool2d_fwd_nhwc_nhwc.hpp

@@ -35,14 +35,13 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
 
     using IndexDataType = int32_t;
 
-    using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
 
     using InElementwiseOperation =
-        typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
 
     using AccElementwiseOperation =
-        typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
-            AccElementwiseOperation;
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
 
     static constexpr index_t InSrcOutDstVectorDim =
         0; // for NHWC, the dim C is the vector Dim for both input and output in memory, which is
@@ -178,13 +177,10 @@ struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePool2dFwd
             invariant_lowest_length_ = C;
             reduce_lowest_length_    = window_spatial_lengths[1];
 
-            // TODO: is this correct?
-            if constexpr(ReduceOpId == ck::ReduceTensorOp::AVG)
-            {
-                ck::index_t divider = window_spatial_lengths[0] * window_spatial_lengths[1];
-                in_element_op_      = InElementwiseOperation{divider};
-                acc_element_op_     = AccElementwiseOperation{divider};
-            }
+            int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
+
+            std::tie(in_element_op_, acc_element_op_) =
+                reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
         }
 
         const InDataType* p_in_dev_;

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

@@ -61,12 +61,9 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
     static constexpr bool use_multiblock =
         (OutMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd);
 
-    static constexpr bool out_type_compatible_with_atomic_op =
-        std::is_same<OutDataType, float>::value || std::is_same<OutDataType, double>::value;
-
-    static_assert(
-        !use_multiblock || (use_multiblock && out_type_compatible_with_atomic_op),
-        "The OutDataType must support the atomic operation for using MultiBlock reduction");
+    static_assert(ck::reduce::InMemoryDataOperatonSupportedOnDataType<OutMemoryDataOperation,
+                                                                      OutDataType>::value,
+                  "The OutDataType must support the specified OutMemoryDataOperation!");
 
     static_assert(!use_multiblock || (use_multiblock && !OutputIndex),
                   "MultiBlock reduction can only be used when outputing index is not required");
@@ -349,7 +346,7 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
             if constexpr(use_multiblock)
             {
                 const auto identityVal =
-                    ck::reduce::GetIdentityValueueForInMemoryDataOperation<OutDataType>(
+                    ck::reduce::GetIdentityValueForInMemoryDataOperation<OutDataType>(
                         OutMemoryDataOperation);
 
                 const auto kernel_pre =
@@ -492,7 +489,7 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
         auto str = std::stringstream();
 
         // clang-format off
-        str << "DeviceReduceMultiBlockAtomicAdd<" << BlockSize << ",";
+        str << (OutMemoryDataOperation == InMemoryDataOperationEnum::Set? "DeviceReduceBlockWise<" : "DeviceReduceMultiBlock<") << BlockSize << ",";
         str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
         str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
         str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << "_OutDstVectorSize_" << OutDstVectorSize << ">";

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

+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device.hpp"
+#include "device_base.hpp"
+#include "gridwise_unary_elementwise_1d.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename ElementwiseFunctor,
+          index_t Dim,
+          index_t ScalarPerVector>
+struct DeviceUnaryElementwise : public BaseOperator
+{
+    static constexpr auto I0 = Number<0>{};
+
+    template <typename Desc_M0>
+    static auto PadDescriptor_M0_1d(Desc_M0 desc_m0, index_t gridSize, index_t blockSize)
+    {
+        const auto m0           = desc_m0.GetLength(I0);
+        const index_t loop_step = gridSize * blockSize * ScalarPerVector;
+        const auto pad          = math::integer_least_multiple(m0, loop_step) - m0;
+        const auto desc_m0_pad =
+            transform_tensor_descriptor(desc_m0,
+                                        make_tuple(make_right_pad_transform(m0, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m0_pad;
+    }
+
+    static auto MakeDescriptor_M0(const std::vector<index_t>& shape,
+                                  const std::vector<index_t>& stride,
+                                  index_t gridSize,
+                                  index_t blockSize)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return shape[I]; }, Number<Dim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<Dim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        // merge nd to 1d desc - [s0 * s1 * ...]
+        if constexpr(Dim > 1)
+        {
+            const auto desc_m0 = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(tupleOfShape)),
+                make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<Dim>{})),
+                make_tuple(Sequence<0>{}));
+
+            return PadDescriptor_M0_1d(desc_m0, gridSize, blockSize);
+        }
+        else
+            return PadDescriptor_M0_1d(desc, gridSize, blockSize);
+    }
+
+    using GridDesc_M0      = decltype(MakeDescriptor_M0({1, 1}, {1, 1}, 1, 1));
+    using GridwiseUEltwise = GridwiseUnaryElementwise_1D<ADataType,
+                                                         BDataType,
+                                                         GridDesc_M0,
+                                                         ElementwiseFunctor,
+                                                         ScalarPerVector>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a,
+                 BDataType* p_b,
+                 const std::vector<index_t>& shape,
+                 const std::vector<index_t>& stride_a,
+                 const std::vector<index_t>& stride_b,
+                 ElementwiseFunctor functor)
+            : p_a_(p_a),
+              p_b_(p_b),
+              shape_(shape),
+              functor_(functor),
+              blockSize_(256) // FIXME - Calculate the grid size by number of CU in the future
+        {
+            index_t tensor_size =
+                std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>{});
+            gridSize_       = GridwiseUEltwise::CalculateGridSize(tensor_size);
+            a_grid_desc_m0_ = MakeDescriptor_M0(shape, stride_a, gridSize_, blockSize_);
+            b_grid_desc_m0_ = MakeDescriptor_M0(shape, stride_b, gridSize_, blockSize_);
+        }
+
+        const ADataType* p_a_;
+        BDataType* p_b_;
+        std::vector<int> shape_;
+        GridDesc_M0 a_grid_desc_m0_;
+        GridDesc_M0 b_grid_desc_m0_;
+        ElementwiseFunctor functor_;
+        index_t blockSize_;
+        index_t gridSize_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel = kernel_unary_elementwise_1d<GridwiseUEltwise,
+                                                            ADataType,
+                                                            BDataType,
+                                                            GridDesc_M0,
+                                                            ElementwiseFunctor>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(arg.gridSize_),
+                                                        dim3(arg.blockSize_),
+                                                        0,
+                                                        arg.p_a_,
+                                                        arg.p_b_,
+                                                        arg.a_grid_desc_m0_,
+                                                        arg.b_grid_desc_m0_,
+                                                        arg.functor_);
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if(pArg == nullptr)
+            return false;
+
+        if(pArg->shape_.back() % ScalarPerVector != 0)
+            return false;
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      void* p_b,
+                                                      std::vector<index_t> shape,
+                                                      std::vector<index_t> stride_a,
+                                                      std::vector<index_t> stride_b,
+                                                      ElementwiseFunctor functor)
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<BDataType*>(p_b),
+                                          shape,
+                                          stride_a,
+                                          stride_b,
+                                          functor);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() { return std::make_unique<Invoker>(); }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBinaryElementwise"
+            << "<"
+            << "ScalarPerVector = " << ScalarPerVector
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -29,6 +29,7 @@
 #include "reduction_operator.hpp"
 #include "reduction_enums.hpp"
 #include "element_wise_operation.hpp"
+#include <tuple>
 
 namespace ck {
 
@@ -37,77 +38,69 @@ namespace ck {
 // The boolean member "indexable" are also provided in reduce_binary_operactor for
 // easier checking by the upper-layer codes in the kernels.
 
-template <typename T, ReduceTensorOp Op>
+template <ReduceTensorOp Op>
 struct reduce_binary_operator;
 
-template <typename T>
-struct reduce_binary_operator<T, ReduceTensorOp::ADD>
+template <>
+struct reduce_binary_operator<ReduceTensorOp::ADD>
 {
-    using opType   = reduce::Add<T>;
-    using dataType = T;
+    using opType = reduce::Add;
 
     static constexpr bool indexable = false;
 };
 
-template <typename T>
-struct reduce_binary_operator<T, ReduceTensorOp::MUL>
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MUL>
 {
-    using opType   = reduce::Mul<T>;
-    using dataType = T;
+    using opType = reduce::Mul;
 
     static constexpr bool indexable = false;
 };
 
-template <typename T>
-struct reduce_binary_operator<T, ReduceTensorOp::MIN>
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MIN>
 {
-    using opType   = reduce::Min<T>;
-    using dataType = T;
+    using opType = reduce::Min;
 
     static constexpr bool indexable = true;
 };
 
-template <typename T>
-struct reduce_binary_operator<T, ReduceTensorOp::MAX>
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MAX>
 {
-    using opType   = reduce::Max<T>;
-    using dataType = T;
+    using opType = reduce::Max;
 
     static constexpr bool indexable = true;
 };
 
-template <typename T>
-struct reduce_binary_operator<T, ReduceTensorOp::AMAX>
+template <>
+struct reduce_binary_operator<ReduceTensorOp::AMAX>
 {
-    using opType   = reduce::AMax<T>;
-    using dataType = T;
+    using opType = reduce::AMax;
 
     static constexpr bool indexable = true;
 };
 
-template <typename T>
-struct reduce_binary_operator<T, ReduceTensorOp::AVG>
+template <>
+struct reduce_binary_operator<ReduceTensorOp::AVG>
 {
-    using opType   = reduce::Add<T>;
-    using dataType = T;
+    using opType = reduce::Add;
 
     static constexpr bool indexable = false;
 };
 
-template <typename T>
-struct reduce_binary_operator<T, ReduceTensorOp::NORM1>
+template <>
+struct reduce_binary_operator<ReduceTensorOp::NORM1>
 {
-    using opType   = reduce::Add<T>;
-    using dataType = T;
+    using opType = reduce::Add;
 
     static constexpr bool indexable = false;
 };
 
-template <typename T>
-struct reduce_binary_operator<T, ReduceTensorOp::NORM2>
+template <>
+struct reduce_binary_operator<ReduceTensorOp::NORM2>
 {
-    using opType   = reduce::Add<T>;
-    using dataType = T;
+    using opType = reduce::Add;
 
     static constexpr bool indexable = false;
 };
@@ -115,53 +108,101 @@ struct reduce_binary_operator<T, ReduceTensorOp::NORM2>
 // The templated struct reduce_unary_operator maps the enum Ids of Reduce operators to two unary
 // functor classes.
 // The two unary functors are called before and afer the Reduction is executed respectively
-template <typename T, ReduceTensorOp Op, bool IsFirstReduce, bool IsLastReduce>
+template <ReduceTensorOp Op, bool IsFirstReduce, bool IsLastReduce>
 struct reduce_unary_operator
 {
-    using InElementwiseOperation  = tensor_operation::element_wise::UnaryIdentic<T, T>;
-    using AccElementwiseOperation = tensor_operation::element_wise::UnaryIdentic<T, T>;
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
 };
 
-template <typename T, bool IsFirstReduce>
-struct reduce_unary_operator<T, ReduceTensorOp::AVG, IsFirstReduce, true>
+template <bool IsFirstReduce>
+struct reduce_unary_operator<ReduceTensorOp::AVG, IsFirstReduce, true>
 {
-    using InElementwiseOperation  = tensor_operation::element_wise::UnaryIdentic<T, T>;
-    using AccElementwiseOperation = tensor_operation::element_wise::UnaryIdentic<T, T, true>;
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnaryDivide;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{reduceLength});
+    };
 };
 
-template <typename T, bool IsLastReduce>
-struct reduce_unary_operator<T, ReduceTensorOp::NORM1, true, IsLastReduce>
+template <bool IsLastReduce>
+struct reduce_unary_operator<ReduceTensorOp::NORM1, true, IsLastReduce>
 {
-    using InElementwiseOperation  = tensor_operation::element_wise::UnaryAbs<T, T>;
-    using AccElementwiseOperation = tensor_operation::element_wise::UnaryIdentic<T, T>;
+    using InElementwiseOperation  = tensor_operation::element_wise::UnaryAbs;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
 };
 
-template <typename T, bool IsLastReduce>
-struct reduce_unary_operator<T, ReduceTensorOp::AMAX, true, IsLastReduce>
+template <bool IsLastReduce>
+struct reduce_unary_operator<ReduceTensorOp::AMAX, true, IsLastReduce>
 {
-    using InElementwiseOperation  = tensor_operation::element_wise::UnaryAbs<T, T>;
-    using AccElementwiseOperation = tensor_operation::element_wise::UnaryIdentic<T, T>;
+    using InElementwiseOperation  = tensor_operation::element_wise::UnaryAbs;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
 };
 
-template <typename T>
-struct reduce_unary_operator<T, ReduceTensorOp::NORM2, true, false>
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, true, false>
 {
-    using InElementwiseOperation  = tensor_operation::element_wise::UnarySquare<T, T>;
-    using AccElementwiseOperation = tensor_operation::element_wise::UnaryIdentic<T, T>;
+    using InElementwiseOperation  = tensor_operation::element_wise::UnarySquare;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
 };
 
-template <typename T>
-struct reduce_unary_operator<T, ReduceTensorOp::NORM2, true, true>
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, true, true>
 {
-    using InElementwiseOperation  = tensor_operation::element_wise::UnarySquare<T, T>;
-    using AccElementwiseOperation = tensor_operation::element_wise::UnarySqrt<T, T>;
+    using InElementwiseOperation  = tensor_operation::element_wise::UnarySquare;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnarySqrt;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
 };
 
-template <typename T>
-struct reduce_unary_operator<T, ReduceTensorOp::NORM2, false, true>
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, false, true>
 {
-    using InElementwiseOperation  = tensor_operation::element_wise::UnaryIdentic<T, T>;
-    using AccElementwiseOperation = tensor_operation::element_wise::UnarySqrt<T, T>;
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnarySqrt;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
 };
 
 } // end of namespace ck

include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp

@@ -24,104 +24,192 @@
  *
  *******************************************************************************/
 #pragma once
+
 #include "data_type.hpp"
 
 namespace ck {
 namespace tensor_operation {
-namespace binary_element_wise {
-
-template <typename Y, typename X1, typename X2>
-struct Add;
+namespace element_wise {
 
-template <>
-struct Add<double, double, double>
+struct Add
 {
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
     __host__ __device__ constexpr void
-    operator()(double& dst, const double& src1, const double& src2) const
+    operator()<float>(float& y, const float& x0, const float& x1) const
     {
-        dst = src1 + src2;
-    }
-};
+        y = x0 + x1;
+    };
 
-template <>
-struct Add<float, float, float>
-{
+    template <>
     __host__ __device__ constexpr void
-    operator()(float& dst, const float& src1, const float& src2) const
+    operator()<double>(double& y, const double& x0, const double& x1) const
     {
-        dst = src1 + src2;
-    }
-};
+        y = x0 + x1;
+    };
 
-template <>
-struct Add<half_t, half_t, half_t>
-{
+    // Question: should half_t be supported ?
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = x0 + x1;
+    };
+
+    // Question: should bhalf_t be supported ?
+    template <>
     __host__ __device__ constexpr void
-    operator()(half_t& dst, const half_t& src1, const half_t& src2) const
+    operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
     {
-        dst = src1 + src2;
+        const float x1_tmp = ck::type_convert<float>(x0);
+        const float x2_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = x1_tmp + x2_tmp;
+        y                  = ck::type_convert<bhalf_t>(y_tmp);
     }
 };
 
-template <>
-struct Add<bhalf_t, bhalf_t, bhalf_t>
+struct Subtract
 {
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
     __host__ __device__ constexpr void
-    operator()(bhalf_t& dst, const bhalf_t& src1, const bhalf_t& src2) const
+    operator()<float>(float& y, const float& x0, const float& x1) const
     {
-        const float x1 = ck::type_convert<float>(src1);
-        const float x2 = ck::type_convert<float>(src2);
-        const float y  = x1 + x2;
-        dst            = ck::type_convert<bhalf_t>(y);
-    }
-};
+        y = x0 - x1;
+    };
 
-template <typename Y, typename X1, typename X2>
-struct Substract;
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        y = x0 - x1;
+    };
 
-template <>
-struct Substract<double, double, double>
-{
+    // Question: should half_t be supported ?
+    template <>
     __host__ __device__ constexpr void
-    operator()(double& dst, const double& src1, const double& src2) const
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
     {
-        dst = src1 - src2;
+        y = x0 - x1;
+    };
+
+    // Question: should bhalf_t be supported ?
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x0);
+        const float x2_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = x1_tmp - x2_tmp;
+        y                  = ck::type_convert<bhalf_t>(y_tmp);
     }
 };
 
-template <>
-struct Substract<float, float, float>
+struct AlphaBetaAdd
 {
+    AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
     __host__ __device__ constexpr void
-    operator()(float& dst, const float& src1, const float& src2) const
+    operator()<float>(float& y, const float& x0, const float& x1) const
     {
-        dst = src1 - src2;
-    }
+        y = alpha_ * x0 + beta_ * x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        y = static_cast<double>(alpha_) * x0 + static_cast<double>(beta_) * x1;
+    };
+
+    // Question: should half_t be supported ?
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = static_cast<half_t>(alpha_ * static_cast<float>(x0) + beta_ * static_cast<float>(x1));
+    };
+
+    float alpha_;
+    float beta_;
 };
 
-template <>
-struct Substract<half_t, half_t, half_t>
+struct AddRelu
 {
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
     __host__ __device__ constexpr void
-    operator()(half_t& dst, const half_t& src1, const half_t& src2) const
+    operator()<float>(float& y, const float& x0, const float& x1) const
     {
-        dst = src1 - src2;
-    }
+        const float a = x0 + x1;
+        y             = a > 0.0f ? a : 0.0f;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        const double a = x0 + x1;
+        y              = a > 0.0 ? a : 0.0;
+    };
+
+    // Question: should half_t be supported ?
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        const half_t a = x0 + x1;
+        y              = a > static_cast<half_t>(0.0f) ? a : static_cast<half_t>(0.0f);
+    };
 };
 
-template <>
-struct Substract<bhalf_t, bhalf_t, bhalf_t>
+struct AddHardswish
 {
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
     __host__ __device__ constexpr void
-    operator()(bhalf_t& dst, const bhalf_t& src1, const bhalf_t& src2) const
+    operator()<float>(float& y, const float& x0, const float& x1) const
     {
-        const float x1 = ck::type_convert<float>(src1);
-        const float x2 = ck::type_convert<float>(src2);
-        const float y  = x1 - x2;
-        dst            = ck::type_convert<bhalf_t>(y);
-    }
+        float a = x0 + x1;
+        float b = a + float{3};
+        float c = (b > 0) * (b > 6.0f ? 6.0f : b) * a * 0.166667f;
+        y       = c;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        double a = x0 + x1;
+        double b = a + 3.0;
+        double c = (b > 0) * (b > 6.0 ? 6.0 : b) * a * 0.166667;
+        y        = c;
+    };
+
+    // Question: should half_t be supported ?
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        float a = x0 + x1;
+        float b = a + 3.0f;
+        float c = (b > 0) * (b > 6.0f ? 6.0f : b) * a * 0.166667f;
+        y       = c;
+    };
 };
 
-} // namespace binary_element_wise
+} // namespace element_wise
 } // namespace tensor_operation
 } // namespace ck

include/ck/tensor_operation/gpu/element/element_wise_operation.hpp

 #pragma once
+
 #include "data_type.hpp"
 #include "math_v2.hpp"
+#include "unary_element_wise_operation.hpp"
+#include "binary_element_wise_operation.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace element_wise {
 
-struct PassThrough
-{
-    __host__ __device__ void operator()(float& y, const float& x) const { y = x; }
-
-    __host__ __device__ void operator()(half_t& y, const half_t& x) const { y = x; }
-
-    __host__ __device__ void operator()(bhalf_t& y, const bhalf_t& x) const { y = x; }
-
-    __host__ __device__ void operator()(int32_t& y, const int32_t& x) const { y = x; }
-
-    __host__ __device__ void operator()(int8_t& y, const int8_t& x) const { y = x; }
-
-    __host__ __device__ void operator()(double& y, const double& x) const { y = x; }
-};
-
-struct Add
-{
-    __host__ __device__ constexpr void operator()(float& y, const float& x0, const float& x1) const
-    {
-        y = x0 + x1;
-    }
-
-    __host__ __device__ constexpr void
-    operator()(half_t& y, const half_t& x0, const half_t& x1) const
-    {
-        // FIXME - Use float (acc type) bias in the future.
-        y = x0 + x1;
-    }
-};
-
-struct AlphaBetaAdd
-{
-    AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta) {}
-
-    __host__ __device__ constexpr void operator()(float& y, const float& x0, const float& x1) const
-    {
-        y = alpha_ * x0 + beta_ * x1;
-    }
-
-    __host__ __device__ constexpr void
-    operator()(half_t& y, const half_t& x0, const half_t& x1) const
-    {
-        // FIXME - Let x0 be acc type
-        y = static_cast<half_t>(alpha_ * static_cast<float>(x0) + beta_ * static_cast<float>(x1));
-    }
-
-    float alpha_;
-    float beta_;
-};
-
-struct AddRelu
-{
-    __host__ __device__ constexpr void operator()(float& y, const float& x0, const float& x1) const
-    {
-        const float a = x0 + x1;
-        y             = a > 0 ? a : 0;
-    }
-
-    __host__ __device__ constexpr void
-    operator()(half_t& y, const half_t& x0, const half_t& x1) const
-    {
-        const half_t a = x0 + x1;
-        y              = a > 0 ? a : 0;
-    }
-};
-
-struct AddHardswish
-{
-    __host__ __device__ constexpr void operator()(float& y, const float& x0, const float& x1) const
-    {
-        float a = x0 + x1;
-        float b = a + float{3};
-        float c = (b > 0) * (b > float{6} ? float{6} : b) * a * float{0.166667};
-        y       = c;
-    }
-
-    __host__ __device__ constexpr void
-    operator()(half_t& y, const half_t& x0, const half_t& x1) const
-    {
-        float a = x0 + x1;
-        float b = a + float{3};
-        float c = (b > 0) * (b > float{6} ? float{6} : b) * a * float{0.166667};
-        y       = c;
-    }
-};
+// Need to ensure compiler will fail if there is no matching candidate, instead of compiler
+// siliently do implicit type conversion
+//
+// Method 1:
+//
+// struct ExampleElementwiseOp
+// {
+//     template<typename Y, typename X>
+//     __host__ __device__ constexpr void
+//     operator()(Y&, const X) const;
+//
+//     template<>
+//     __host__ __device__ constexpr void
+//     operator()<half_t, half_t>(half_t& y, const half_t& x) const
+//     {
+//     }
+// };
+//
+// Method 2:
+//
+// template <typename Y, typename X>
+// struct ExampleElementwiseOp;
+//
+// template <>
+// struct ExampleElementwiseOp<float, ck::bhalf_t>
+// {
+//     __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
+//     {
+//     }
+// };
 
 struct AddReluAdd
 {
-    __host__ __device__ constexpr void
-    operator()(half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
+    template <typename Y, typename X0, typename X1, typename X2>
+    __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&, const X2&) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
     {
         half_t a = x0 + x1;
         half_t b = a > 0 ? a : 0;
         y        = b + x2;
     }
 
-    __host__ __device__ constexpr void
-    operator()(float& y, const float& x0, const float& x1, const float& x2) const
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& y,
+                                                                              const float& x0,
+                                                                              const float& x1,
+                                                                              const float& x2) const
     {
         float a = x0 + x1;
         float b = a > 0 ? a : 0;
@@ -111,8 +66,9 @@ struct AddReluAdd
         y       = c;
     }
 
-    __host__ __device__ constexpr void
-    operator()(half_t& y, const float& x0, const half_t& x1, const half_t& x2) const
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
+        half_t& y, const float& x0, const half_t& x1, const half_t& x2) const
     {
         float a = x0 + x1;
         float b = a > 0 ? a : 0;
@@ -123,8 +79,14 @@ struct AddReluAdd
 
 struct AddHardswishAdd
 {
-    __host__ __device__ constexpr void
-    operator()(float& y, const float& x0, const float& x1, const float& x2) const
+    template <typename Y, typename X0, typename X1, typename X2>
+    __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&, const X2&) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& y,
+                                                                              const float& x0,
+                                                                              const float& x1,
+                                                                              const float& x2) const
     {
         float a = x0 + x1;
         float b = a + float{3};
@@ -133,8 +95,9 @@ struct AddHardswishAdd
         y       = d;
     }
 
-    __host__ __device__ constexpr void
-    operator()(half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
     {
         float a = x0 + x1;
         float b = a + float{3};
@@ -144,206 +107,95 @@ struct AddHardswishAdd
     }
 };
 
-struct Normalize
+// C = A * B
+// E = FastGelu(C + D0 + D1)
+struct AddAddFastGelu
 {
-    Normalize(float epsilon = 1e-4) : epsilon_(epsilon) {}
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ void operator()(E&, const C&, const D0&, const D1&) const;
 
-    __host__ __device__ constexpr void operator()(float& y,
-                                                  const float& x,
-                                                  const float& mean,
-                                                  const float& mean_square,
-                                                  const float& gamma,
-                                                  const float& beta) const
+    template <>
+    __host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
+                                                                       const float& c,
+                                                                       const half_t& d0,
+                                                                       const half_t& d1) const
     {
-        float variance = mean_square - (mean * mean);
-        y              = ((x - mean) / sqrtf(variance + epsilon_)) * gamma + beta;
-    }
-
-    float epsilon_;
-};
+        // Fast GeLU
+        // https://paperswithcode.com/method/gelu
+        // y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+        const auto fast_gelu = [&](float x) {
+            const float u   = float(2) * x * (float(0.035677) * x * x + float(0.797885));
+            const float emu = exp(-u);
+            const float cdf = float(0.5) + float(0.5) * (float(2) / (float(1) + emu) - float(1));
+            return x * cdf;
+        };
 
-// Unary operators are usually called element-wisely before/after the reduction is executed on the
-// elements. They are needed for easy implementation of reduction types of AVG, NRM1, NRM2
+        const float y = fast_gelu(c + float(d0) + float(d1));
 
-template <typename Y, typename X, bool HasDividing = false>
-struct UnaryIdentic;
-
-template <>
-struct UnaryIdentic<float, float, false>
-{
-    __host__ __device__ UnaryIdentic(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(float& y, const float& x) const { y = x; };
-};
-
-template <>
-struct UnaryIdentic<float, float, true>
-{
-    __host__ __device__ UnaryIdentic(const int32_t divider = 1) { divider_ = divider; };
-
-    __host__ __device__ void operator()(float& y, const float& x) const
-    {
-        y = x / type_convert<float>(divider_);
-    };
-
-    int32_t divider_ = 1;
-};
-
-template <>
-struct UnaryIdentic<half_t, half_t, false>
-{
-    __host__ __device__ UnaryIdentic(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(half_t& y, const half_t& x) const { y = x; };
+        e = type_convert<half_t>(y);
+    }
 };
 
-template <>
-struct UnaryIdentic<double, double, false>
+struct Normalize
 {
-    __host__ __device__ UnaryIdentic(const int32_t divider = 1) { (void)divider; };
+    // FIXME: is double absolutely necessary?
+    Normalize(double epsilon = 1e-4) : epsilon_(epsilon) {}
 
-    __host__ __device__ void operator()(double& y, const double& x) const { y = x; };
-};
-
-template <>
-struct UnaryIdentic<double, double, true>
-{
-    __host__ __device__ UnaryIdentic(const int32_t divider = 1) { divider_ = divider; };
+    template <typename T>
+    __host__ __device__ constexpr void operator()(
+        T& y, const T& x, const T& mean, const T& mean_square, const T& gamma, const T& beta) const;
 
-    __host__ __device__ void operator()(double& y, const double& x) const
+    template <>
+    __host__ __device__ constexpr void operator()<float>(float& y,
+                                                         const float& x,
+                                                         const float& mean,
+                                                         const float& mean_square,
+                                                         const float& gamma,
+                                                         const float& beta) const
     {
-        y = x / type_convert<double>(divider_);
-    };
-
-    int32_t divider_ = 1;
-};
-
-template <>
-struct UnaryIdentic<int32_t, int32_t, false>
-{
-    __host__ __device__ UnaryIdentic(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(int32_t& y, const int32_t& x) const { y = x; };
-};
-
-template <>
-struct UnaryIdentic<int32_t, int32_t, true>
-{
-    __host__ __device__ UnaryIdentic(const int32_t divider = 1) { divider_ = divider; };
-
-    __host__ __device__ void operator()(int32_t& y, const int32_t& x) const { y = x / divider_; };
-
-    int32_t divider_ = 1;
-};
-
-template <>
-struct UnaryIdentic<int8_t, int8_t, false>
-{
-    __host__ __device__ UnaryIdentic(const int8_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(int8_t& y, const int8_t& x) const { y = x; };
-};
-
-template <typename Y, typename X, bool HasDividing = false>
-struct UnarySquare;
-
-template <>
-struct UnarySquare<float, float, false>
-{
-    __host__ __device__ UnarySquare(const int32_t divider = 1) { (void)divider; };
+        using ck::math::sqrt;
 
-    __host__ __device__ void operator()(float& y, const float& x) const { y = x * x; };
-};
-
-template <>
-struct UnarySquare<float, float, true>
-{
-    __host__ __device__ UnarySquare(const int32_t divider = 1) { divider_ = divider; };
-
-    __host__ __device__ void operator()(float& y, const float& x) const
-    {
-        y = x * x / type_convert<float>(divider_);
+        float variance = mean_square - (mean * mean);
+        y = ((x - mean) / sqrt(variance + static_cast<float>(epsilon_))) * gamma + beta;
     };
 
-    int32_t divider_ = 1;
-};
-
-template <>
-struct UnarySquare<double, double, false>
-{
-    __host__ __device__ UnarySquare(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(double& y, const double& x) const { y = x * x; };
-};
-
-template <>
-struct UnarySquare<double, double, true>
-{
-    __host__ __device__ UnarySquare(const int32_t divider = 1) { divider_ = divider; };
-
-    __host__ __device__ void operator()(double& y, const double& x) const
+    template <>
+    __host__ __device__ constexpr void operator()<double>(double& y,
+                                                          const double& x,
+                                                          const double& mean,
+                                                          const double& mean_square,
+                                                          const double& gamma,
+                                                          const double& beta) const
     {
-        y = x * x / type_convert<double>(divider_);
-    };
+        using ck::math::sqrt;
 
-    int32_t divider_ = 1;
-};
-
-template <typename Y, typename X>
-struct UnaryAbs;
-
-template <>
-struct UnaryAbs<float, float>
-{
-    __host__ __device__ UnaryAbs(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(float& y, const float& x) const { y = ck::math::abs(x); };
-};
-
-template <>
-struct UnaryAbs<half_t, half_t>
-{
-    __host__ __device__ UnaryAbs(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(half_t& y, const half_t& x) const { y = ck::math::abs(x); };
-};
-
-template <>
-struct UnaryAbs<double, double>
-{
-    __host__ __device__ UnaryAbs(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(double& y, const double& x) const { y = ck::math::abs(x); };
-};
-
-template <>
-struct UnaryAbs<int8_t, int8_t>
-{
-    __host__ __device__ UnaryAbs(const int32_t divider = 1) { (void)divider; };
+        double variance = mean_square - (mean * mean);
+        y               = ((x - mean) / sqrt(variance + epsilon_)) * gamma + beta;
+    };
 
-    __host__ __device__ void operator()(int8_t& y, const int8_t& x) const { y = ck::math::abs(x); };
+    // FIXME: is double absolutely necessary?
+    double epsilon_;
 };
 
 template <typename Y, typename X>
-struct UnarySqrt;
+struct UnaryTypeConvert;
 
 template <>
-struct UnarySqrt<float, float>
+struct UnaryTypeConvert<float, ck::bhalf_t>
 {
-    __host__ __device__ UnarySqrt(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(float& y, const float& x) const { y = ck::math::sqrt(x); };
+    __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
+    {
+        y = ck::type_convert<float, ck::bhalf_t>(x);
+    }
 };
 
 template <>
-struct UnarySqrt<double, double>
+struct UnaryTypeConvert<ck::bhalf_t, float>
 {
-    __host__ __device__ UnarySqrt(const int32_t divider = 1) { (void)divider; };
-
-    __host__ __device__ void operator()(double& y, const double& x) const
+    __host__ __device__ void operator()(ck::bhalf_t& y, float& x) const
     {
-        y = ck::math::sqrt(x);
-    };
+        y = ck::type_convert<ck::bhalf_t, float>(x);
+    }
 };
 
 } // namespace element_wise

include/ck/tensor_operation/gpu/element/element_wise_reduce_operation.hpp

-#pragma once
-#include "data_type.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace element_wise {
-
-} // namespace element_wise
-} // namespace tensor_operation
-} // namespace ck

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

+#pragma once
+
+#include "data_type.hpp"
+#include "math_v2.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+struct PassThrough
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, bhalf_t>::value ||
+                          is_same<T, int32_t>::value || is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = x;
+    };
+};
+
+struct UnaryDivide
+{
+    __host__ __device__ UnaryDivide(const int32_t divider = 1) : divider_(divider){};
+
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, int32_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = x / type_convert<T>(divider_);
+    };
+
+    int32_t divider_ = 1;
+};
+
+struct UnarySquare
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value,
+                      "Data type is not supported by this operation!");
+
+        y = x * x;
+    };
+};
+
+struct UnaryAbs
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = ck::math::abs(x);
+    };
+};
+
+struct UnarySqrt
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value,
+                      "Data type is not supported by this operation!");
+
+        y = ck::math::sqrt(x);
+    };
+};
+
+struct Relu
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+        y = x > 0 ? x : 0;
+    }
+
+    template <>
+    __host__ __device__ void operator()(bhalf_t& y, const bhalf_t& x) const
+    {
+        float x_f32 = ck::type_convert<float>(x);
+        float y_f32 = x_f32 > 0 ? x_f32 : 0;
+        y           = ck::type_convert<bhalf_t>(y_f32);
+    }
+};
+
+// https://paperswithcode.com/method/gelu
+// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+struct FastGelu
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        const float u   = float(2) * x * (float(0.035677) * x * x + float(0.797885));
+        const float emu = exp(-u);
+        const float cdf = float(0.5) + float(0.5) * (float(2) / (float(1) + emu) - float(1));
+
+        y = x * cdf;
+    }
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp

@@ -171,15 +171,15 @@ struct GridwiseReduction_mk_to_m_multiblock
                                AccDataType beta,
                                OutDataType* const __restrict__ p_out_value_global)
     {
-        const auto identityVal = ReduceOperation::GetIdentityValue();
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
 
         // LDS
         __shared__ AccDataType p_reduce_work_buffer[BlockSize];
 
-        const auto in_global_val_buf =
-            make_dynamic_buffer<AddressSpaceEnum::Global>(p_in_value_global,
-                                                          in_grid_desc_m_k.GetElementSpaceSize(),
-                                                          type_convert<InDataType>(identityVal));
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
         auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
 
@@ -358,12 +358,12 @@ struct GridwiseReduction_mk_to_m_multiblock
         __shared__ AccDataType p_reduce_work_val_buffer[BlockSize];
         __shared__ IndexDataType p_reduce_work_idx_buffer[BlockSize];
 
-        const auto identityVal = ReduceOperation::GetIdentityValue();
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
 
-        const auto in_global_val_buf =
-            make_dynamic_buffer<AddressSpaceEnum::Global>(p_in_value_global,
-                                                          in_grid_desc_m_k.GetElementSpaceSize(),
-                                                          type_convert<InDataType>(identityVal));
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
         const auto in_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_in_index_global, in_grid_desc_m_k.GetElementSpaceSize());
         auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(

include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp

@@ -135,12 +135,12 @@ struct GridwiseReduction_mk_to_m_threadwise
                                                      ReduceOperation,
                                                      PropagateNan>;
 
-        const auto identityVal = ReduceOperation::GetIdentityValue();
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
 
-        const auto in_global_val_buf =
-            make_dynamic_buffer<AddressSpaceEnum::Global>(p_in_value_global,
-                                                          in_grid_desc_m_k.GetElementSpaceSize(),
-                                                          type_convert<InDataType>(identityVal));
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
         auto dst_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
 
@@ -276,12 +276,12 @@ struct GridwiseReduction_mk_to_m_threadwise
 
         (void)acc_elementwise_op;
 
-        const auto identityVal = ReduceOperation::GetIdentityValue();
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
 
-        const auto in_global_val_buf =
-            make_dynamic_buffer<AddressSpaceEnum::Global>(p_in_value_global,
-                                                          in_grid_desc_m_k.GetElementSpaceSize(),
-                                                          type_convert<InDataType>(identityVal));
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
         const auto in_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_in_index_global, in_grid_desc_m_k.GetElementSpaceSize());
 

include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp

+#pragma once
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "blockwise_gemm_xdlops.hpp"
+#include "thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "gridwise_gemm_pipeline_v1.hpp"
+#include "reduction_functions_threadwise.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename FloatC0,
+          typename FloatC1,
+          typename DPtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename C1ElementwiseOperation,
+          typename DxsInElementwiseOperation,
+          typename DxsReduceAccElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename DGridDescriptor_MBlock_MPerBlock,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_bias_add_reduce_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const FloatC0* __restrict__ p_c0_grid,
+            const FloatC1* __restrict__ p_c1_grid,
+            DPtrsGlobal p_ds_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const C1ElementwiseOperation c1_element_op,
+            const DxsInElementwiseOperation dxs_in_element_op,
+            const DxsReduceAccElementwiseOperation dxs_out_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c0_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c1_grid_desc_mblock_mperblock_nblock_nperblock,
+            const DGridDescriptor_MBlock_MPerBlock d_grid_desc_mblock_mperblock,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_c0_grid,
+                                                  p_c1_grid,
+                                                  p_ds_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  c1_element_op,
+                                                  dxs_in_element_op,
+                                                  dxs_out_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  d_grid_desc_mblock_mperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_c0_grid;
+    ignore = p_c1_grid;
+    ignore = p_ds_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = c1_element_op;
+    ignore = dxs_in_element_op;
+    ignore = dxs_out_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c0_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c1_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = d_grid_desc_mblock_mperblock;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename FloatC0,
+          typename FloatC1,
+          typename FloatReduceAcc,
+          typename DPtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename C1ElementwiseOperation,
+          typename DxsReduceOperation,
+          typename DxsInElementwiseOperation,
+          typename DxsReduceAccElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename DGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          typename C0GridDesc_M_N,
+          typename C1GridDesc_M_N,
+          typename DGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched>
+struct GridwiseGemmBiasAddReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        // static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
+        //               is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
+        //               "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N_& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeDGridDescriptor_MBlock_MPerBlock(const DGridDesc_M& d_grid_desc_m)
+    {
+        const auto M      = d_grid_desc_m.GetLength(I0);
+        const auto MBlock = M / MPerBlock;
+
+        const auto d_grid_desc_mblock_mperblock = transform_tensor_descriptor(
+            d_grid_desc_m,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return d_grid_desc_mblock_mperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(C0GridDesc_M_N{}))>;
+
+    using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(C1GridDesc_M_N{}))>;
+
+    using DGridDescriptor_MBlock_MPerBlock =
+        remove_cvref_t<decltype(MakeDGridDescriptor_MBlock_MPerBlock(DGridDesc_M{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               const FloatC0* __restrict__ p_c0_grid,
+                               const FloatC1* __restrict__ p_c1_grid,
+                               DPtrsGlobal p_ds_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const C1ElementwiseOperation& c1_element_op,
+                               const DxsInElementwiseOperation& dxs_in_element_op,
+                               const DxsReduceAccElementwiseOperation& dxs_out_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const DGridDescriptor_MBlock_MPerBlock& d_grid_desc_mblock_mperblock,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_grid, c0_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c1_grid, c1_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C + reduction + write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            // TODO: this should be implemented as a blockwise reduction
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // VGPR c_reduce_thread_desc_mperblock_nperblock
+            constexpr auto c_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            // VGPR d_reduce_thread_desc_mperblock
+            constexpr auto d_reduce_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            // VGPR d_reduce_thread_desc_mblock_mperblock
+            constexpr auto d_reduce_thread_desc_mblock_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
+
+            auto c_reduce_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            auto dxs_reduce_thread_copy_vgpr_to_global = generate_tuple(
+                [&](auto I) {
+                    auto p_d_grid         = p_ds_grid[I];
+                    auto d_out_element_op = dxs_out_element_op[I];
+
+                    return ThreadwiseTensorSliceTransfer_v1r3<
+                        FloatReduceAcc,
+                        remove_pointer_t<decltype(p_d_grid)>,
+                        decltype(d_reduce_thread_desc_mblock_mperblock),
+                        decltype(d_grid_desc_mblock_mperblock),
+                        decltype(d_out_element_op),
+                        Sequence<1, mreduce_per_thread>,
+                        Sequence<0, 1>,
+                        1,
+                        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+                        DGlobalMemoryDataOperation::At(I),
+                        1,
+                        false>{d_grid_desc_mblock_mperblock,
+                               make_multi_index(block_work_idx[I0],                  // mblock
+                                                c_reduce_thread_data_idx_begin[I0]), // mperblock
+                               d_out_element_op};
+                },
+                Number<p_ds_grid.Size()>{});
+
+            // c0 and c1
+            constexpr auto c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            constexpr auto c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock;
+
+            auto c01_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            auto c0_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC0,
+                FloatReduceAcc,
+                decltype(c0_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(
+                c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
+
+            auto c1_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC1,
+                FloatReduceAcc,
+                decltype(c1_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(
+                c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
+
+            constexpr auto c_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            auto c_reduce_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                FloatReduceAcc,
+                FloatC,
+                decltype(c_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>, // SliceLengths
+                Sequence<0, 1, 2, 3>,                                     // DimAccessOrder
+                3,                                                        // DstVectorDim
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]),
+                tensor_operation::element_wise::PassThrough{}};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+                {
+                    c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf,
+                                                         c_reduce_thread_desc_mperblock_nperblock,
+                                                         make_tuple(I0, I0),
+                                                         c_reduce_thread_buf);
+
+                    c0_thread_copy_global_to_vgpr.Run(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c0_grid_buf,
+                        c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c01_thread_buf);
+
+                    // c = activation(c + bias)
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            FloatReduceAcc out;
+                            c_element_op(out, c_reduce_thread_buf(i) + c01_thread_buf(i));
+                            c_reduce_thread_buf(i) = out;
+                        });
+
+                    c1_thread_copy_global_to_vgpr.Run(
+                        c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c1_grid_buf,
+                        c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c01_thread_buf);
+
+                    // c = c + c1_functior(c1)
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            c1_element_op(c01_thread_buf(i), c01_thread_buf(i));
+                            c_reduce_thread_buf(i) += c01_thread_buf(i);
+                        });
+
+                    c_reduce_thread_copy_vgpr_to_global.Run(
+                        c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c_reduce_thread_buf,
+                        c_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c_grid_buf);
+
+                    static_for<0, p_ds_grid.Size(), 1>{}([&](auto In) {
+                        auto& p_d_grid = p_ds_grid[In];
+
+                        auto d_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                            p_d_grid, d_grid_desc_mblock_mperblock.GetElementSpaceSize());
+
+                        auto d_thread_buf =
+                            make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                                d_reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+                        auto& d_in_element_op = dxs_in_element_op[In];
+
+                        auto& d_reduce_thread_copy_vgpr_to_global =
+                            dxs_reduce_thread_copy_vgpr_to_global(In);
+
+                        using DReduceOperation = remove_cvref_t<decltype(DxsReduceOperation{}[In])>;
+                        using ThreadwiseReduce =
+                            ThreadwiseReduction<FloatReduceAcc,
+                                                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                                decltype(d_reduce_thread_desc_mperblock),
+                                                DReduceOperation,
+                                                false>;
+
+                        // Global write Gemm shuffle + reduction
+                        const auto d_zeroVal =
+                            DReduceOperation::template GetIdentityValue<FloatReduceAcc>();
+
+                        static_for<0, mreduce_per_thread, 1>{}(
+                            [&](auto I) { d_thread_buf(I) = d_zeroVal; });
+
+                        // reduce in VGPR
+                        static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                            static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                                constexpr auto offset =
+                                    Number<c_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                        make_tuple(im, in))>{};
+
+                                d_in_element_op(c_reduce_thread_buf(offset),
+                                                c_reduce_thread_buf(offset));
+                            });
+                        });
+
+                        ThreadwiseReduce::Reduce(c_reduce_thread_buf, d_thread_buf);
+
+                        // copy from VGPR to Global
+                        d_reduce_thread_copy_vgpr_to_global.Run(
+                            d_reduce_thread_desc_mblock_mperblock,
+                            make_tuple(I0, I0),
+                            d_thread_buf,
+                            d_grid_desc_mblock_mperblock,
+                            d_grid_buf);
+
+                        if constexpr(access_id < num_access - 1)
+                        {
+                            constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+                            d_reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                                d_grid_desc_mblock_mperblock,
+                                make_tuple(c_global_step[I0], c_global_step[I1]));
+                        }
+                    });
+                }
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C0
+                    c0_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C1
+                    c1_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c1_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        } // Reduction
+    }
+};
+
+} // namespace ck

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

+#pragma once
+
+#include "common_header.hpp"
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "blockwise_gemm_xdlops.hpp"
+#include "thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "thread_group_tensor_slice_transfer_v7.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "gridwise_gemm_pipeline_v1.hpp"
+
+namespace ck {
+
+// input : A[AK0, M, AK1]
+// input : B[AK0, N, AK1]
+// input : D0[M, N], D1[M, N], ...
+// output : E[M, N]
+// C = a_op(A) * b_op(B)
+// E = cde_op(C, D0, D1, ...)
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename DsDataType,
+          typename FloatE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename EGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched>
+struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2ETileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const EGridDesc_M_N& e_grid_desc_m_n,
+                  const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+
+    using DefaultBlock2ETileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <bool HasMainKBlockLoop, typename Block2ETileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        DsGridPointer p_ds_grid,
+        FloatE* __restrict__ p_e_grid,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const StaticallyIndexedArray<EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                                     NumDTensor>&
+            ds_grid_desc_mblock_mperblock_nblock_nperblock, // FIXME: Ds desc may be of different
+                                                            // type from E
+        const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            e_grid_desc_mblock_mperblock_nblock_nperblock,
+        const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(FloatCShuffle{}), DsDataType{})),
+                Tuple<FloatE>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                            // support arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 cde_element_op};
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp

@@ -21,7 +21,7 @@ template <typename GridwiseGemm,
           typename BElementwiseOperation,
           typename CElementwiseOperation,
           typename DxsInElementwiseOperation,
-          typename DxsAccElementwiseOperation,
+          typename DxsReduceAccElementwiseOperation,
           typename AGridDesc_AK0_M_AK1,
           typename BGridDesc_BK0_N_BK1,
           typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
@@ -41,7 +41,7 @@ __global__ void
             const BElementwiseOperation b_element_op,
             const CElementwiseOperation c_element_op,
             const DxsInElementwiseOperation dxs_in_element_op,
-            const DxsAccElementwiseOperation dxs_out_element_op,
+            const DxsReduceAccElementwiseOperation dxs_out_element_op,
             const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
             const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
             const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
@@ -96,7 +96,7 @@ template <typename FloatAB,
           typename CElementwiseOperation,
           typename DxsReduceOperation,
           typename DxsInElementwiseOperation,
-          typename DxsAccElementwiseOperation,
+          typename DxsReduceAccElementwiseOperation,
           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
           typename DGlobalMemoryDataOperation,
           typename AGridDesc_AK0_M_AK1,
@@ -329,7 +329,7 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
                                const BElementwiseOperation& b_element_op,
                                const CElementwiseOperation& c_element_op,
                                const DxsInElementwiseOperation& dxs_in_element_op,
-                               const DxsAccElementwiseOperation& dxs_out_element_op,
+                               const DxsReduceAccElementwiseOperation& dxs_out_element_op,
                                const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
                                const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
                                const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
@@ -816,7 +816,8 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
                                                 false>;
 
                         // Global write Gemm shuffle + reduction
-                        const auto d_identityVal = DReduceOperation::GetIdentityValue();
+                        const auto d_identityVal =
+                            DReduceOperation::template GetIdentityValue<FloatReduceAcc>();
 
                         static_for<0, mreduce_per_thread, 1>{}(
                             [&](auto I) { d_thread_buf(I) = d_identityVal; });

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

@@ -791,8 +791,10 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
             constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
                 GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
 
+            void* p_shared = static_cast<void*>(p_shared_block);
+
             auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-                static_cast<FloatC*>(p_shared_block),
+                static_cast<FloatC*>(p_shared),
                 c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
 
             static_assert(M1 == MWave, "");

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp

@@ -340,7 +340,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
     using DefaultBlock2CTileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
 
-    template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
     __device__ static void
     Run(const FloatAB* __restrict__ p_a_grid,
         const FloatAB* __restrict__ p_b_grid,

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

@@ -37,7 +37,7 @@ __global__ void kernel_buffer_set_value(const Grid1dBufferDescType grid_1d_buffe
 
 {
 
-    using PassThroughOp = tensor_operation::element_wise::UnaryIdentic<DataType, DataType>;
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
 
     constexpr auto I0 = Number<0>{};
 

include/ck/tensor_operation/gpu/grid/gridwise_unary_elementwise_1d.hpp

+#pragma once
+
+#include "cluster_descriptor.hpp"
+#include "data_type.hpp"
+#include "element_wise_operation.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+
+namespace ck {
+
+template <typename GridwiseUEltwise,
+          typename ADataType,
+          typename BDataType,
+          typename GridDesc_M0,
+          typename ElementwiseFunctor>
+__global__ void kernel_unary_elementwise_1d(const ADataType* __restrict__ p_a_global,
+                                            BDataType* __restrict__ p_b_global,
+                                            const GridDesc_M0 a_grid_desc_m0,
+                                            const GridDesc_M0 b_grid_desc_m0,
+                                            const ElementwiseFunctor functor)
+{
+    GridwiseUEltwise::Run(p_a_global, p_b_global, a_grid_desc_m0, b_grid_desc_m0, functor);
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename GridDesc_M0,
+          typename ElementwiseFunctor,
+          index_t ScalarPerVector>
+struct GridwiseUnaryElementwise_1D
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto thread_desc_m0 =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<ScalarPerVector>{}));
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+
+    static __device__ auto CalculateElementwiseIndex()
+    {
+        const index_t global_thread_id = get_thread_global_1d_id();
+        return make_multi_index(global_thread_id * ScalarPerVector);
+    }
+
+    __host__ __device__ static constexpr bool CheckValidity(const GridDesc_M0 a_grid_desc_m0,
+                                                            const GridDesc_M0 b_grid_desc_m0)
+    {
+        return a_grid_desc_m0.GetLength(I0) == b_grid_desc_m0.GetLength(I0);
+    }
+
+    __host__ __device__ static constexpr index_t CalculateGridSize(const index_t tensor_size)
+    {
+        const index_t grid_size = math::integer_divide_ceil(tensor_size, 256 * ScalarPerVector);
+
+        return grid_size;
+    }
+
+    __device__ static void Run(const ADataType* __restrict__ p_a_global,
+                               BDataType* __restrict__ p_b_global,
+                               const GridDesc_M0 a_grid_desc_m0,
+                               const GridDesc_M0 b_grid_desc_m0,
+                               const ElementwiseFunctor functor)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_global, a_grid_desc_m0.GetElementSpaceSize());
+        auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_global, b_grid_desc_m0.GetElementSpaceSize());
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ADataType, ScalarPerVector, true> a_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, BDataType, ScalarPerVector, true> b_thread_buf;
+
+        const auto thread_store_global_offset = CalculateElementwiseIndex();
+
+        auto a_global_load =
+            ThreadwiseTensorSliceTransfer_v2<ADataType,
+                                             ADataType,
+                                             GridDesc_M0,
+                                             decltype(thread_desc_m0),
+                                             Sequence<ScalarPerVector>, // SliceLengths
+                                             Sequence<0>,               // DimAccessOrder
+                                             0,                         // SrcVectorDim
+                                             ScalarPerVector,
+                                             1, // SrcScalarStrideInVector
+                                             false>{a_grid_desc_m0, thread_store_global_offset};
+
+        auto b_global_write =
+            ThreadwiseTensorSliceTransfer_v1r3<BDataType,
+                                               BDataType,
+                                               decltype(thread_desc_m0),
+                                               GridDesc_M0,
+                                               PassThrough,
+                                               Sequence<ScalarPerVector>, // SliceLengths
+                                               Sequence<0>,               // DimAccessOrder
+                                               0,                         // DstVectorDim
+                                               ScalarPerVector,
+                                               InMemoryDataOperationEnum::Set,
+                                               1, // DstScalarStrideInVector
+                                               false>{
+                b_grid_desc_m0, thread_store_global_offset, PassThrough{}};
+
+        const index_t blockSize    = get_block_size();
+        const index_t blockPerGrid = get_grid_size();
+        const auto m0              = b_grid_desc_m0.GetLength(I0);
+        const index_t loop_step    = blockPerGrid * blockSize * ScalarPerVector;
+        const auto loop_step_index = make_multi_index(loop_step);
+
+        index_t num_iter = m0 / (loop_step);
+        do
+        {
+            // read and process ScalarPerVector elements
+            a_global_load.Run(
+                a_grid_desc_m0, a_global_buf, thread_desc_m0, make_tuple(I0), a_thread_buf);
+
+            static_for<0, ScalarPerVector, 1>{}([&](auto m) {
+                constexpr auto offset = thread_desc_m0.CalculateOffset(make_tuple(m));
+                functor(b_thread_buf(Number<offset>{}), a_thread_buf(Number<offset>{}));
+            });
+
+            b_global_write.Run(thread_desc_m0,
+                               make_tuple(I0), // SrcSliceOriginIdx
+                               b_thread_buf,
+                               b_grid_desc_m0,
+                               b_global_buf);
+
+            a_global_load.MoveSrcSliceWindow(a_grid_desc_m0, loop_step_index);
+            b_global_write.MoveDstSliceWindow(b_grid_desc_m0, loop_step_index);
+        } while(--num_iter);
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp

+#pragma once
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags> // Sequence<bool ...>
+struct ThreadwiseTensorSliceTransfer_v7
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+    using SpaceFillingCurve =
+        SpaceFillingCurve<SliceLengths, DimAccessOrder, remove_cv_t<decltype(scalar_per_access)>>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        auto generate_vectors = [&](auto data_types) {
+            constexpr index_t num = data_types.Size();
+
+            return generate_tuple(
+                [&](auto i) {
+                    using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                    return vector_type_maker_t<DataType, ScalarPerVector>{};
+                },
+                Number<num>{});
+        };
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors(SrcDatas{});
+            auto dst_vectors = generate_vectors(DstDatas{});
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
+                                                                                src_coords_[i]);
+
+                src_vectors(i).template AsType<src_vector_t>()(I0) =
+                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(),
+                                                           is_src_valid);
+            });
+
+            // apply pointwise function
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        return src_vectors[iSrc].template AsType<SrcData>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        return dst_vectors(iDst).template AsType<DstData>()(i);
+                    },
+                    Number<nDst>{});
+
+                // apply pointwise function
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
+
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+
+                const bool is_dst_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
+                                                                                dst_coords_[i]);
+
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_coords_[i].GetOffset(),
+                    is_dst_valid,
+                    dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i],
+                                           src_coords_(i),
+                                           make_tensor_coordinate_step(src_descs[i], forward_step));
+                });
+
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(dst_descs[i],
+                                           dst_coords_(i),
+                                           make_tensor_coordinate_step(dst_descs[i], forward_step));
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto src_reset_step =
+                    make_tensor_coordinate_step(src_descs[i], GetCoordinateResetStep());
+
+                move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto dst_reset_step =
+                    make_tensor_coordinate_step(dst_descs[i], GetCoordinateResetStep());
+
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
+            }
+        });
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRunFlags::At(iSrc)
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRunFlags::At(iDst)
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
+    }
+
+    private:
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/utility/amd_buffer_addressing.hpp

@@ -6,6 +6,8 @@ namespace ck {
 template <typename T>
 union BufferResource
 {
+    __device__ constexpr BufferResource() : content{} {}
+
     // 128 bit SGPRs to supply buffer resource in buffer instructions
     // https://rocm-documentation.readthedocs.io/en/latest/GCN_ISA_Manuals/testdocbook.html#vector-memory-buffer-instructions
     int32x4_t content;

include/ck/utility/data_type.hpp

 #pragma once
+
 #include "statically_indexed_array.hpp"
 
 namespace ck {