adding bias add

composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v1r4.hpp

+#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V1R4_HPP
+#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V1R4_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src:
+//     1. Src0Desc is known at compile-time
+//     2. Src0Buffer is StaticBuffer
+//     3. SrcSliceOrginIdx is known at compile-time
+//   2. dst:
+//     1. DstDesc is not known at compile-time
+//     2. DstBuffer is DynamicBuffer
+//     3. DstSliceOrginIdx is not known at compile time
+template <typename Src0Data,
+          typename DstData,
+          typename Src0Desc,
+          typename DstDesc,
+          typename SrcElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector,
+          InMemoryDataOperationEnum_t DstInMemOp,
+          index_t DstScalarStrideInVector,
+          bool DstResetCoordinateAfterRun,
+          typename enable_if<Src0Desc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v1r4
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v1r4(
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin_idx,
+        const SrcElementwiseOperation src_element_op)
+        : dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin_idx)),
+          src_element_op_{src_element_op}
+    {
+        static_assert(Src0Desc::IsKnownAtCompileTime(),
+                      "wrong! Src0Desc need to known at compile-time");
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcSliceOriginIdx,
+              typename Src0Buffer,
+              typename DstBuffer,
+              typename DstStepHacks>
+    __device__ void Run(const Src0Desc&,
+                        const SrcSliceOriginIdx&,
+                        const Src0Buffer& src0_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf,
+                        const DstStepHacks& dst_step_hacks)
+    {
+        static_assert(Src0Desc::IsKnownAtCompileTime(),
+                      "wrong! Src0Desc need to known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcSliceOriginIdx>>::value,
+                      "wrong! SrcSliceOrigin need to known at compile-time");
+
+        static_assert(Src0Buffer::IsStaticBuffer(), "wrong! Src0Buffer need to be StaticBuffer");
+
+        // Src0Desc and src_slice_origin_idx are known at compile-time
+        constexpr auto src_desc             = remove_cvref_t<Src0Desc>{};
+        constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{});
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+        constexpr auto access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dim_access_order = DimAccessOrder{};
+
+        constexpr auto ordered_access_lengths =
+            container_reorder_given_new2old(access_lengths, dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_access_idx[I0];
+
+                    static_for<0, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i]
+                                         ? ordered_access_idx[i]
+                                         : ordered_access_lengths[i] - 1 - ordered_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dim_access_order) *
+                       dst_scalar_per_access;
+            }();
+
+            typename vector_type_maker<DstData, DstScalarPerVector>::type dst_vector;
+
+            using dst_vector_t =
+                typename vector_type_maker<DstData, DstScalarPerVector>::type::type;
+
+            // copy data from src0_buf into dst_vector
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t src_offset = src_desc.CalculateOffset(
+                    src_slice_origin_idx + dst_data_idx + i * dst_scalar_step_in_vector);
+
+                // apply element-wise operation and type convert
+                dst_vector.template AsType<DstData>()(i) =
+                    type_convert<DstData>(src_element_op_(src0_buf[Number<src_offset>{}]));
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Set)
+            {
+                dst_buf.template Set<dst_vector_t>(
+                    dst_coord_.GetOffset(),
+                    is_dst_valid,
+                    dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+            }
+            else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::AtomicAdd)
+            {
+                dst_buf.template AtomicAdd<dst_vector_t>(
+                    dst_coord_.GetOffset(),
+                    is_dst_valid,
+                    dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+            }
+            else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Add)
+            {
+
+                typename vector_type_maker<DstData, DstScalarPerVector>::type tmp;
+                tmp.template AsType<dst_vector_t>()(Number<0>{}) =
+                    dst_buf.template Get<dst_vector_t>(dst_coord_.GetOffset(), is_dst_valid);
+
+                static_for<0, DstScalarPerVector, 1>{}([&](auto t) {
+                    dst_vector.template AsType<DstData>()(t) += tmp.template AsType<DstData>()[t];
+                });
+
+                dst_buf.template Set<dst_vector_t>(
+                    dst_coord_.GetOffset(),
+                    is_dst_valid,
+                    dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+            }
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_access_idx[i] < ordered_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &= ordered_access_idx[j] == ordered_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    template <typename SrcSliceOriginIdx, typename Src0Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc&,
+                        const SrcSliceOriginIdx&,
+                        const Src0Buffer& src0_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        constexpr index_t ntransform_dst = DstDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_dst, 0>::type{};
+
+        constexpr auto dst_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        Run(Src0Desc{}, SrcSliceOriginIdx{}, src0_buf, dst_desc, dst_buf, dst_step_hacks);
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto I0 = Number<0>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dim_access_order = DimAccessOrder{};
+
+        constexpr auto ordered_access_lengths =
+            container_reorder_given_new2old(access_lengths, dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_access_lengths[I0] - 1;
+
+                static_for<0, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_access_lengths[j] + ordered_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in Run(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dim_access_order) *
+                   dst_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       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 =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    DstCoord dst_coord_;
+    SrcElementwiseOperation src_element_op_;
+}; // namespace ck
+
+} // namespace ck
+#endif

example/2_gemm_xdl_bias_add/README.md

+# Instructions for ```gemm_xdl``` Example
+
+## Docker script
+```bash
+docker run                                                                   \
+-it                                                                          \
+--rm                                                                         \
+--privileged                                                                 \
+--group-add sudo                                                             \
+-w /root/workspace                                                           \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace                                \
+rocm/tensorflow:rocm4.3.1-tf2.6-dev                                          \
+/bin/bash
+```
+
+## Build ``gemm_xdl```
+```bash
+mkdir build && cd build
+```
+
+```bash
+# Need to specify target ID, example below is gfx908
+cmake                                                                  \
+-D BUILD_DEV=OFF                                                       \
+-D CMAKE_BUILD_TYPE=Release                                            \
+-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 "   \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                              \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                         \
+..
+```
+
+```bash
+ make -j gemm_xdl
+```
+
+## Run ```gemm_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+./example/gemm_xdl.sh 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
+arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 5 times...
+Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
+```

example/2_gemm_xdl_bias_add/gemm_xdl_bias_add.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "host_gemm.hpp"
+#include "device_tensor.hpp"
+#include "device_base.hpp"
+#include "example/2_gemm_xdl_bias_add/include/device_gemm_xdl_bias_add.hpp"
+
+// C[m, n] = alpha(A[m, k] * B[k, n]) + beta * D[m] + gamma * E[m, n]
+
+struct PassThrough
+{
+    template <typename T>
+    __host__ __device__ constexpr T operator()(T v) const
+    {
+        return v;
+    }
+};
+
+struct Relu
+{
+    float alpha = 0.1;
+
+    // ReLU
+    template <typename T>
+    __host__ __device__ constexpr T operator()(T v) const
+    {
+        T tmp = alpha * v;
+        return tmp > 0 ? tmp : 0;
+    }
+};
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmInstance;
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmInstance<ck::half_t,
+                          ck::half_t,
+                          ck::half_t,
+                          ck::tensor_layout::gemm::RowMajor,
+                          ck::tensor_layout::gemm::ColumnMajor,
+                          ck::tensor_layout::gemm::RowMajor,
+                          AElementwiseOperation,
+                          BElementwiseOperation,
+                          CElementwiseOperation>
+{
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using AOp = AElementwiseOperation;
+    using BOp = BElementwiseOperation;
+    using COp = CElementwiseOperation;
+
+    // Compilation parameters for NT problem
+    // clang-format off
+    using type =
+        //########################################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| AElementwise| BElementwise| CElementwise| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
+        //########################################|  Type|  Type|  Type|    Type|        |        |        |    Operation|    Operation|    Operation|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|     ThreadSlice|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|     ThreadSlice|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| SrcDstVectorDim|       DstScalar| AddExtraM| AddExtraN|
+        //########################################|      |      |      |        |        |        |        |             |             |             |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|                |       PerVector|          |          |
+        //########################################|      |      |      |        |        |        |        |             |             |             |      |      |      |      |   |     |     |     |     |                |                |               |               |               |               |               |                |                |               |               |              |               |               |                |                |          |          |
+        ck::tensor_operation::device::DeviceGemmXdl_two_extra_source_reduce<  F16,   F16,   F16,     F32,     Row,     Col,     Row,          AOp,          BOp,          COp,   256,   256,   128,     4,  8,   32,   32,    4,    2,      S<1, 4, 8>,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      S<1, 2, 8>,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,               7,               1,      true,      true>;
+    // clang-format on
+};
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmInstance<float,
+                          float,
+                          float,
+                          ck::tensor_layout::gemm::RowMajor,
+                          ck::tensor_layout::gemm::ColumnMajor,
+                          ck::tensor_layout::gemm::RowMajor,
+                          AElementwiseOperation,
+                          BElementwiseOperation,
+                          CElementwiseOperation>
+{
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using AOp = AElementwiseOperation;
+    using BOp = BElementwiseOperation;
+    using COp = CElementwiseOperation;
+
+    // Compilation parameters for NT problem
+    // clang-format off
+    using type =
+    //########################################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| AElementwise| BElementwise| CElementwise| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
+    //########################################|  Type|  Type|  Type|    Type|        |        |        |    Operation|    Operation|    Operation|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|     ThreadSlice|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|     ThreadSlice|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| SrcDstVectorDim|       DstScalar| AddExtraM| AddExtraN|
+    //########################################|      |      |      |        |        |        |        |             |             |             |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|                |       PerVector|          |          |
+    //########################################|      |      |      |        |        |        |        |             |             |             |      |      |      |      |   |     |     |     |     |                |                |               |               |               |               |               |                |                |               |               |              |               |               |                |                |          |          |
+    ck::tensor_operation::device::DeviceGemmXdl_two_extra_source_reduce<  F32,   F32,   F32,     F32,     Row,     Col,     Row,          AOp,          BOp,          COp,   256,   256,   128,     4,  4,   32,   32,    4,    2,      S<1, 4, 4>,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      S<1, 2, 4>,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,               7,               1,      true,      true>;
+    // clang-format on
+};
+
+int main(int argc, char* argv[])
+{
+    if(argc != 4)
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: run kernel # of times (>1)\n");
+        exit(0);
+    }
+
+    const bool do_verification = std::stoi(argv[1]);
+    const int init_method      = std::stoi(argv[2]);
+    const int nrepeat          = std::stoi(argv[3]);
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    // matrix data type
+    using ADataType = ck::half_t;
+    using BDataType = ck::half_t;
+    using CDataType = ck::half_t;
+
+    // matrix layout
+    using ALayout = ck::tensor_layout::gemm::RowMajor;
+    using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+    using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<BDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<BDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+    c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
+
+    // do GEMM
+    auto gemm = typename DeviceGemmInstance<ADataType,
+                                            BDataType,
+                                            CDataType,
+                                            ALayout,
+                                            BLayout,
+                                            CLayout,
+                                            PassThrough,
+                                            PassThrough,
+                                            Relu>::type{};
+
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      PassThrough{},
+                                      PassThrough{},
+                                      Relu{});
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, nrepeat);
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * M + sizeof(CDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        host_gemm_mk_kn_mn(a_m_k, b_k_n, c_m_n_host_result, PassThrough{}, PassThrough{}, Relu{});
+
+        check_error(c_m_n_host_result, c_m_n_device_result);
+    }
+}

example/2_gemm_xdl_bias_add/include/device_gemm_bias_add.hpp

+#ifndef DEVICE_GEMM_HPP
+#define DEVICE_GEMM_HPP
+
+#include <iostream>
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceGemmPtr = std::unique_ptr<
+    DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

example/CMakeLists.txt

 include_directories(BEFORE
-    include
+    ${PROJECT_SOURCE_DIR}
     ${PROJECT_SOURCE_DIR}/host/host_tensor/include
     ${PROJECT_SOURCE_DIR}/host/device/include
     ${PROJECT_SOURCE_DIR}/device_operation/include
@@ -12,7 +12,10 @@ include_directories(BEFORE
 )
 
 set(GEMM_XDL_SOURCE 1_gemm_xdl/gemm_xdl.cpp)
+set(GEMM_XDL_BIAS_ADD_SOURCE 2_gemm_xdl_bias_add/gemm_xdl_bias_add.cpp)
 
 add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
+add_executable(gemm_xdl_bias_add ${GEMM_XDL_BIAS_ADD_SOURCE})
 
 target_link_libraries(gemm_xdl PRIVATE host_tensor)
+target_link_libraries(gemm_xdl_bias_add PRIVATE host_tensor)