elementwise op (#238)

* Add elementwise operation kernel and example

* Add comment

* Add template argument of dim . Prepare to support multiple dimension

* Rename example

* Support 1 dimension

* Add static assert

* Add comment

* Extract pad

* Remove redundant argument

* Support any dimension for elementwise operation

* Remove line

* Let it be the multiple number of CU

* Move thread per block to the parameter of constructor

* rename threadPerBlock with blockSize

* Support double

* rename kernel function name

* remove redundant include header

* Refine type

* Need to the final dimension

* Refine variable name

* Refine type

* Use index_t instead of int in API

Co-authored-by: rocking <chunylai@amd.com>

[ROCm/composable_kernel commit: aafc3ac27a]

example/19_binary_elementwise/CMakeLists.txt

@@ -0,0 +1,3 @@
+add_example_executable(example_broadcast_add_2d broadcast_add_2d.cpp)
+add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)
+add_example_executable(example_elementwise_add_4d elementwise_add_4d.cpp)

example/19_binary_elementwise/broadcast_add_2d.cpp

@@ -0,0 +1,132 @@
+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::binary_element_wise::Add;
+
+using DeviceElementwiseAddInstance = ck::tensor_operation::device::
+    DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 2, 8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor,
+          int broadcastDim>
+void host_broadcast2D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, int N, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            ComputeDataType Amn = static_cast<ComputeDataType>(A(m, n));
+            ComputeDataType Cmn = 0;
+            if constexpr(broadcastDim == 0)
+            {
+                ComputeDataType Bn = static_cast<ComputeDataType>(B(n));
+                functor(Cmn, Amn, Bn);
+            }
+            else
+            {
+                ComputeDataType Bm = static_cast<ComputeDataType>(B(m));
+                functor(Cmn, Amn, Bm);
+            }
+            C(m, n) = static_cast<ctype>(Cmn);
+        }
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                    std::vector<std::size_t>({stride, 1}));
+    };
+
+    Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+    Tensor<ABDataType> b_n(f_host_tensor_descriptor1d(N, 1));
+
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+    a_m_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_n_device_buf(sizeof(ABDataType) * a_m_n.mDesc.GetElementSpace());
+    DeviceMem b_n_device_buf(sizeof(ABDataType) * b_n.mDesc.GetElementSpace());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpace());
+
+    a_m_n_device_buf.ToDevice(a_m_n.mData.data());
+    b_n_device_buf.ToDevice(b_n.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(a_m_n_device_buf.GetDeviceBuffer(),
+                                                     b_n_device_buf.GetDeviceBuffer(),
+                                                     c_m_n_device_buf.GetDeviceBuffer(),
+                                                     {M, N},
+                                                     {Stride, 1},
+                                                     {0, 1}, // broadcast in first dimension
+                                                     {Stride, 1},
+                                                     Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise_2D instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_n_device_buf.FromDevice(c_m_n.mData.data());
+        Tensor<CDataType> host_c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+        host_broadcast2D<Tensor<ABDataType>,
+                         Tensor<ABDataType>,
+                         Tensor<CDataType>,
+                         EltwiseComputeDataType,
+                         Add,
+                         0>(host_c_m_n, a_m_n, b_n, M, N, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m_n.mData, host_c_m_n.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/19_binary_elementwise/elementwise_add_1d.cpp

@@ -0,0 +1,110 @@
+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::binary_element_wise::Add;
+
+using DeviceElementwiseAddInstance = ck::tensor_operation::device::
+    DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 1, 8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor>
+void host_elementwise1D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        ComputeDataType Am = static_cast<ComputeDataType>(A(m));
+        ComputeDataType Bm = static_cast<ComputeDataType>(B(m));
+        ComputeDataType Cm = 0;
+        functor(Cm, Am, Bm);
+        C(m) = static_cast<ctype>(Cm);
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<ABDataType> b_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<ABDataType> c_m(f_host_tensor_descriptor1d(M, 1));
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpace());
+    DeviceMem b_m_device_buf(sizeof(ABDataType) * b_m.mDesc.GetElementSpace());
+    DeviceMem c_m_device_buf(sizeof(CDataType) * c_m.mDesc.GetElementSpace());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_device_buf.ToDevice(b_m.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(a_m_device_buf.GetDeviceBuffer(),
+                                                     b_m_device_buf.GetDeviceBuffer(),
+                                                     c_m_device_buf.GetDeviceBuffer(),
+                                                     {M},
+                                                     {1},
+                                                     {1},
+                                                     {1},
+                                                     Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise_2D instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_device_buf.FromDevice(c_m.mData.data());
+        Tensor<CDataType> host_c_m(f_host_tensor_descriptor1d(M, 1));
+
+        host_elementwise1D<Tensor<ABDataType>,
+                           Tensor<ABDataType>,
+                           Tensor<CDataType>,
+                           EltwiseComputeDataType,
+                           Add>(host_c_m, a_m, b_m, M, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m.mData, host_c_m.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/19_binary_elementwise/elementwise_add_4d.cpp

@@ -0,0 +1,113 @@
+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "host_utility.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::binary_element_wise::Add;
+
+using DeviceElementwiseAddInstance = ck::tensor_operation::device::
+    DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 4, 8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor>
+void host_elementwise4D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0, 0, 0))>;
+
+    for(std::size_t n = 0; n < shape[0]; ++n)
+        for(std::size_t c = 0; c < shape[1]; ++c)
+            for(std::size_t h = 0; h < shape[2]; ++h)
+                for(std::size_t w = 0; w < shape[3]; ++w)
+                {
+                    ComputeDataType a_val = static_cast<ComputeDataType>(A(n, c, h, w));
+                    ComputeDataType b_val = static_cast<ComputeDataType>(B(n, c, h, w));
+                    ComputeDataType c_val = 0;
+                    functor(c_val, a_val, b_val);
+                    C(n, c, h, w) = static_cast<ctype>(c_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> nchw = {4, 16, 32, 32};
+
+    Tensor<ABDataType> a_m(nchw);
+    Tensor<ABDataType> b_m(nchw);
+    Tensor<ABDataType> c_m(nchw);
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpace());
+    DeviceMem b_m_device_buf(sizeof(ABDataType) * b_m.mDesc.GetElementSpace());
+    DeviceMem c_m_device_buf(sizeof(CDataType) * c_m.mDesc.GetElementSpace());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_device_buf.ToDevice(b_m.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        a_m_device_buf.GetDeviceBuffer(),
+        b_m_device_buf.GetDeviceBuffer(),
+        c_m_device_buf.GetDeviceBuffer(),
+        ck::convert_vector_element_type<std::size_t, ck::index_t>(nchw),
+        ck::convert_vector_element_type<std::size_t, ck::index_t>(a_m.mDesc.GetStrides()),
+        ck::convert_vector_element_type<std::size_t, ck::index_t>(b_m.mDesc.GetStrides()),
+        ck::convert_vector_element_type<std::size_t, ck::index_t>(c_m.mDesc.GetStrides()),
+        Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise_2D instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_device_buf.FromDevice(c_m.mData.data());
+        Tensor<CDataType> host_c_m(nchw);
+
+        host_elementwise4D<Tensor<ABDataType>,
+                           Tensor<ABDataType>,
+                           Tensor<CDataType>,
+                           EltwiseComputeDataType,
+                           Add>(host_c_m, a_m, b_m, nchw, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m.mData, host_c_m.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/CMakeLists.txt

@@ -51,3 +51,4 @@ add_subdirectory(17_convnd_bwd_data_xdl)
 add_subdirectory(15_grouped_gemm)
 add_subdirectory(16_gemm_reduce)
 add_subdirectory(18_batched_gemm_reduce)
+add_subdirectory(19_binary_elementwise)

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

@@ -0,0 +1,204 @@
+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device.hpp"
+#include "device_base.hpp"
+#include "gridwise_binary_elementwise_1d.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ComputeDataType,
+          typename ElementwiseFunctor,
+          index_t Dim,
+          index_t ScalarPerVector>
+struct DeviceBinaryElementwise : public BaseOperator
+{
+    DeviceBinaryElementwise(index_t blockSize = 256) : BaseOperator(), blockSize_(blockSize) {}
+
+    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 GridwiseBinEltwise = GridwiseBinaryElementwise_1D<ADataType,
+                                                            BDataType,
+                                                            CDataType,
+                                                            ComputeDataType,
+                                                            GridDesc_M0,
+                                                            ElementwiseFunctor,
+                                                            ScalarPerVector>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a,
+                 const BDataType* p_b,
+                 CDataType* p_c,
+                 const std::vector<index_t>& shape,
+                 const std::vector<index_t>& stride_a,
+                 const std::vector<index_t>& stride_b,
+                 const std::vector<index_t>& stride_c,
+                 ElementwiseFunctor functor,
+                 index_t blockSize)
+            : p_a_(p_a),
+              p_b_(p_b),
+              p_c_(p_c),
+              shape_(shape),
+              functor_(functor),
+              gridSize_(120) // FIXME - Calculate the grid size by number of CU in the future
+        {
+            a_grid_desc_m0_ = MakeDescriptor_M0(shape, stride_a, gridSize_, blockSize);
+            b_grid_desc_m0_ = MakeDescriptor_M0(shape, stride_b, gridSize_, blockSize);
+            c_grid_desc_m0_ = MakeDescriptor_M0(shape, stride_c, gridSize_, blockSize);
+        }
+
+        const ADataType* p_a_;
+        const BDataType* p_b_;
+        CDataType* p_c_;
+        std::vector<int> shape_;
+        GridDesc_M0 a_grid_desc_m0_;
+        GridDesc_M0 b_grid_desc_m0_;
+        GridDesc_M0 c_grid_desc_m0_;
+        ElementwiseFunctor functor_;
+        index_t gridSize_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        Invoker(index_t blockSize) : BaseInvoker(), blockSize_(blockSize) {}
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel = kernel_binary_elementwise_1d<GridwiseBinEltwise,
+                                                             ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             GridDesc_M0,
+                                                             ElementwiseFunctor>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(arg.gridSize_),
+                                                        dim3(blockSize_),
+                                                        0,
+                                                        arg.p_a_,
+                                                        arg.p_b_,
+                                                        arg.p_c_,
+                                                        arg.a_grid_desc_m0_,
+                                                        arg.b_grid_desc_m0_,
+                                                        arg.c_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);
+        }
+
+        index_t blockSize_;
+    };
+
+    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,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      std::vector<index_t> shape,
+                                                      std::vector<index_t> stride_a,
+                                                      std::vector<index_t> stride_b,
+                                                      std::vector<index_t> stride_c,
+                                                      ElementwiseFunctor functor)
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          shape,
+                                          stride_a,
+                                          stride_b,
+                                          stride_c,
+                                          functor,
+                                          blockSize_);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{blockSize_});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBinaryElementwise"
+            << "<"
+            << "ScalarPerVector = " << ScalarPerVector
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    index_t blockSize_;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -0,0 +1,25 @@
+#pragma once
+#include "data_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace binary_element_wise {
+
+struct Add
+{
+    __host__ __device__ constexpr void
+    operator()(double& dst, const double& src1, const double& src2) const
+    {
+        dst = src1 + src2;
+    }
+
+    __host__ __device__ constexpr void
+    operator()(float& dst, const float& src1, const float& src2) const
+    {
+        dst = src1 + src2;
+    }
+};
+
+} // namespace binary_element_wise
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_binary_elementwise_1d.hpp

@@ -0,0 +1,150 @@
+#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 GridwiseBinEltwise,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GridDesc_M0,
+          typename ElementwiseFunctor>
+__global__ void kernel_binary_elementwise_1d(const ADataType* __restrict__ p_a_global,
+                                             const BDataType* __restrict__ p_b_global,
+                                             CDataType* __restrict__ p_c_global,
+                                             const GridDesc_M0 a_grid_desc_m0,
+                                             const GridDesc_M0 b_grid_desc_m0,
+                                             const GridDesc_M0 c_grid_desc_m0,
+                                             const ElementwiseFunctor functor)
+{
+    GridwiseBinEltwise::Run(p_a_global,
+                            p_b_global,
+                            p_c_global,
+                            a_grid_desc_m0,
+                            b_grid_desc_m0,
+                            c_grid_desc_m0,
+                            functor);
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ComputeDataType,
+          typename GridDesc_M0,
+          typename ElementwiseFunctor,
+          index_t ScalarPerVector>
+struct GridwiseBinaryElementwise_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);
+    }
+
+    __device__ static void Run(const ADataType* __restrict__ p_a_global,
+                               const BDataType* __restrict__ p_b_global,
+                               CDataType* __restrict__ p_c_global,
+                               const GridDesc_M0 a_grid_desc_m0,
+                               const GridDesc_M0 b_grid_desc_m0,
+                               const GridDesc_M0 c_grid_desc_m0,
+                               const ElementwiseFunctor functor)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_global, a_grid_desc_m0.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_global, b_grid_desc_m0.GetElementSpaceSize());
+        auto c_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_global, c_grid_desc_m0.GetElementSpaceSize());
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, ScalarPerVector, true> a_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, ScalarPerVector, true> b_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, ScalarPerVector, true> c_thread_buf;
+
+        const auto thread_store_global_offset = CalculateElementwiseIndex();
+
+        auto a_global_load =
+            ThreadwiseTensorSliceTransfer_v2<ADataType,
+                                             ComputeDataType,
+                                             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_load =
+            ThreadwiseTensorSliceTransfer_v2<BDataType,
+                                             ComputeDataType,
+                                             GridDesc_M0,
+                                             decltype(thread_desc_m0),
+                                             Sequence<ScalarPerVector>, // SliceLengths
+                                             Sequence<0>,               // DimAccessOrder
+                                             0,                         // SrcVectorDim
+                                             ScalarPerVector,
+                                             1, // SrcScalarStrideInVector
+                                             false>{b_grid_desc_m0, thread_store_global_offset};
+
+        auto c_global_write =
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
+                                               CDataType,
+                                               decltype(thread_desc_m0),
+                                               GridDesc_M0,
+                                               PassThrough,
+                                               Sequence<ScalarPerVector>, // SliceLengths
+                                               Sequence<0>,               // DimAccessOrder
+                                               0,                         // DstVectorDim
+                                               ScalarPerVector,
+                                               InMemoryDataOperationEnum::Set,
+                                               1, // DstScalarStrideInVector
+                                               false>{
+                c_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              = c_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);
+
+            b_global_load.Run(
+                b_grid_desc_m0, b_global_buf, thread_desc_m0, make_tuple(I0), b_thread_buf);
+
+            static_for<0, ScalarPerVector, 1>{}([&](auto m) {
+                constexpr auto offset = thread_desc_m0.CalculateOffset(make_tuple(m));
+                functor(c_thread_buf(Number<offset>{}),
+                        a_thread_buf(Number<offset>{}),
+                        b_thread_buf(Number<offset>{}));
+            });
+
+            c_global_write.Run(thread_desc_m0,
+                               make_tuple(I0), // SrcSliceOriginIdx
+                               c_thread_buf,
+                               c_grid_desc_m0,
+                               c_global_buf);
+
+            a_global_load.MoveSrcSliceWindow(a_grid_desc_m0, loop_step_index);
+            b_global_load.MoveSrcSliceWindow(b_grid_desc_m0, loop_step_index);
+            c_global_write.MoveDstSliceWindow(c_grid_desc_m0, loop_step_index);
+        } while(--num_iter);
+    }
+};
+
+} // namespace ck

include/ck/utility/get_id.hpp

@@ -11,10 +11,14 @@ __host__ __device__ constexpr index_t get_warp_size()
 
 __device__ index_t get_thread_local_1d_id() { return threadIdx.x; }
 
+__device__ index_t get_thread_global_1d_id() { return blockIdx.x * blockDim.x + threadIdx.x; }
+
 __device__ index_t get_warp_local_1d_id() { return threadIdx.x / get_warp_size(); }
 
 __device__ index_t get_block_1d_id() { return blockIdx.x; }
 
 __device__ index_t get_grid_size() { return gridDim.x; }
 
+__device__ index_t get_block_size() { return blockDim.x; }
+
 } // namespace ck

library/include/ck/library/host_tensor/host_utility.hpp

@@ -0,0 +1,17 @@
+#pragma once
+#include <vector>
+
+namespace ck {
+
+template <typename Src, typename Dst>
+inline std::vector<Dst> convert_vector_element_type(const std::vector<Src>& inData)
+{
+    std::vector<Dst> outData;
+
+    for(auto elem : inData)
+        outData.push_back(static_cast<Dst>(elem));
+
+    return (outData);
+};
+
+}; // namespace ck