Merge branch 'develop' into gridwise_2d

example/44_conv2d_fwd_quant/CMakeLists.txt

+add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
+add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)

example/44_conv2d_fwd_quant/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using InDataType       = int8_t;
+using WeiDataType      = int8_t;
+using BiasDataType     = int32_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using OutDataType      = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<BiasDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        8>;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& bias_g_k_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<BiasDataType> bias(bias_g_k_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+    bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        in_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
+        out_device_buf.GetDeviceBuffer(),
+        a_g_n_c_wis_lengths,
+        a_g_n_c_wis_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d0_g_n_k_wos_lengths}},
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d0_g_n_k_wos_strides}},
+        e_g_n_k_wos_lengths,
+        e_g_n_k_wos_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     CShuffleDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     PassThrough>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach(
+            [&](auto&, auto idx) { out_element_op(out_host(idx), c_host(idx), bias(idx)); });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &= ck::utils::check_err(
+            out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int main()
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        32,           // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{0.5f, ActivationOp{}};
+
+    using InLayout   = ck::tensor_layout::convolution::GNHWC;
+    using WeiLayout  = ck::tensor_layout::convolution::GKYXC;
+    using BiasLayout = ck::tensor_layout::convolution::G_K;
+    using OutLayout  = ck::tensor_layout::convolution::GNHWK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    // TODO - make_bias_host_tensor_descriptor_g_n_k_wos_packed()
+    const auto bias_g_k_desc = HostTensorDescriptor({conv_param.G_,
+                                                     conv_param.N_,
+                                                     conv_param.K_,
+                                                     conv_param.output_spatial_lengths_[0],
+                                                     conv_param.output_spatial_lengths_[1]},
+                                                    {
+                                                        conv_param.K_, // g
+                                                        0,             // n
+                                                        1,             // k
+                                                        0,             // ho
+                                                        0              // wo
+                                                    });
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    std::cout << out_g_n_k_wos_desc << std::endl;
+
+    return run_grouped_conv_fwd<
+        ndim_spatial,
+        InDataType,
+        WeiDataType,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        DeviceGroupedConvNDFwdInstance<ndim_spatial, InLayout, WeiLayout, BiasLayout, OutLayout>>(
+        do_verification,
+        time_kernel,
+        conv_param,
+        in_g_n_c_wis_desc,
+        wei_g_k_c_xs_desc,
+        bias_g_k_desc,
+        out_g_n_k_wos_desc,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+}

example/44_conv2d_fwd_quant/conv2d_fwd_xdl_perlayer_quantization_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using InDataType       = int8_t;
+using WeiDataType      = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using OutDataType      = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        16>;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &= ck::utils::check_err(
+            out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int main()
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        32,           // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{0.5f, ActivationOp{}};
+
+    using InLayout  = ck::tensor_layout::convolution::GNHWC;
+    using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+    using OutLayout = ck::tensor_layout::convolution::GNHWK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    return run_grouped_conv_fwd<
+        ndim_spatial,
+        InDataType,
+        WeiDataType,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        DeviceGroupedConvNDFwdInstance<ndim_spatial, InLayout, WeiLayout, OutLayout>>(
+        do_verification,
+        time_kernel,
+        conv_param,
+        in_g_n_c_wis_desc,
+        wei_g_k_c_xs_desc,
+        out_g_n_k_wos_desc,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+}

example/45_elementwise_normalization/CMakeLists.txt

+add_example_executable(example_elementwise_layernorm_blockwise elementwise_layernorm_blockwise.cpp)

example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+using ADataType             = ck::half_t; // Input 1
+using BDataType             = ck::half_t; // Input 2
+using XDataType             = ck::half_t;
+using GammaDataType         = ck::half_t;
+using BetaDataType          = ck::half_t;
+using YDataType             = ck::half_t;
+using AccDataType           = float;
+using XElementwiseOperation = ck::tensor_operation::element_wise::Add;
+using YElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+// X = Elementwise(input1, input2, input3, ...)
+// Y = Layernorm(X, beta, gamma)
+using DeviceInstance = ck::tensor_operation::device::DeviceElementwiseNormalizationImpl<
+    ck::Tuple<ADataType, BDataType>,
+    GammaDataType,
+    BetaDataType,
+    AccDataType,
+    YDataType,
+    XElementwiseOperation,
+    YElementwiseOperation,
+    Rank,
+    NumReduceDim,
+    256, // BlockSize
+    8,   // ClusterM
+    32,  // ClusterK
+    1,   // SliceM
+    32,  // SliceK
+    1,   // SrcVecDim (0=M, 1=K)
+    8,   // SrcScalarPerVector
+    1,   // GammaVecDim (0=M, 1=K)
+    8,   // GammaScalarPerVector
+    1,   // BetaVecDim (0=M, 1=K)
+    8,   // BetaScalarPerVector
+    8>;  // OutScalarPerVector
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise2D(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))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+        {
+            auto a_val  = A(m, n);
+            auto b_val  = B(m, n);
+            ctype c_val = 0;
+            functor(c_val, a_val, b_val);
+            C(m, n) = c_val;
+        }
+}
+
+int main()
+{
+    bool time_kernel = true;
+
+    ck::index_t M      = 48 * 256;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    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<ADataType> a(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<BDataType> b(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<GammaDataType> gamma(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta(f_host_tensor_descriptor1d(N, 1));
+    Tensor<YDataType> y(f_host_tensor_descriptor2d(M, N, Stride));
+
+    a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+    b.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+    gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+    beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+
+    DeviceMem a_dev(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_dev(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    a_dev.ToDevice(a.mData.data());
+    b_dev.ToDevice(b.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    std::array<const void*, 2> input = {a_dev.GetDeviceBuffer(), b_dev.GetDeviceBuffer()};
+
+    auto device_instance = DeviceInstance{};
+    auto argument_ptr    = device_instance.MakeArgumentPointer(
+        {M, N},
+        {
+            std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
+            std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
+        },
+        {0, 1},
+        {0, 1},
+        std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+        {1},
+        1e-4,
+        input,
+        gamma_dev.GetDeviceBuffer(),
+        beta_dev.GetDeviceBuffer(),
+        y_dev.GetDeviceBuffer(),
+        XElementwiseOperation{},
+        YElementwiseOperation{});
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+    float ela_time   = 0;
+    ela_time         = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    float data_mem_size = M * N * sizeof(ADataType) + M * N * sizeof(BDataType) +
+                          M * N * sizeof(YDataType) + N * sizeof(GammaDataType) +
+                          N * sizeof(BetaDataType);
+    float bandwidth = data_mem_size * 1000 / ela_time / 1024 / 1024 / 1024;
+
+    std::cout << "Bandwidth is : " << bandwidth << "GB/s . " << std::endl;
+    std::cout << "Time elapase is : " << ela_time << " ms . " << std::endl;
+
+    bool pass = true;
+    {
+        std::vector<std::size_t> mn = {static_cast<unsigned long>(M),
+                                       static_cast<unsigned long>(N)};
+        Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
+        host_elementwise2D<Tensor<ADataType>,
+                           Tensor<BDataType>,
+                           Tensor<XDataType>,
+                           XElementwiseOperation>(x, a, b, mn, XElementwiseOperation{});
+
+        Tensor<YDataType> host_y(f_host_tensor_descriptor2d(M, N, Stride));
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           YDataType,
+                                                           AccDataType,
+                                                           YElementwiseOperation,
+                                                           Rank,
+                                                           NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, YElementwiseOperation{}, {M, N}, {1}, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        y_dev.FromDevice(y.mData.data());
+        pass &=
+            ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+        if(!(pass))
+        {
+            std::cout << "layernorm wrong" << std::endl;
+        }
+    }
+    return (pass ? 0 : 1);
+}

include/ck/ck.hpp

@@ -126,8 +126,14 @@
 #define CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST 1
 
 // experimental feature: optimize for inter-wave scheduling policy
-#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 0
+#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 1
 #define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS 1
+// this will let make_default_loop_scheduler() return interwave scheduling flag by default
+#define CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING 0
+// experimental feature: add instances using interwave scheduling
+#define CK_EXPERIMENTAL_INTER_WAVE_INSTANCES 1
+// experimental feature: add instances using pipeline v2
+#define CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES 1
 
 // hack: have underlying assumption that need to be satsified, otherwise it's a bug
 // hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
@@ -144,26 +150,10 @@
 // workaround: compiler gnerating inefficient ds_write instructions
 #define CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE 1
 
-// (gfx908 only) workaround: compiler crash in fused kernels on mainline #9110; #10738 seems ok
-// error message was "fatal error: error in backend: Error while trying to spill VGPR0 from class
-// VGPR_32: Cannot scavenge register without an emergency spill slot!"
-// this fall back to less ideal way of handle NPadding in fused attention kernel
-#ifdef __gfx908__
-#define CK_WORKAROUND_SWDEV_XXXXXX_ATTN_KERNEL_CLANG_CANNOT_SCAVENGE_REGISTER 1
-#else
-// for __gfx90a__, ...
-#define CK_WORKAROUND_SWDEV_XXXXXX_ATTN_KERNEL_CLANG_CANNOT_SCAVENGE_REGISTER 0
-#endif // __gfx908__
-
 // workaround: verifaction failure, due to compiler regression, for conv bwd-data fp16 using some
 // tuning parameter
 #define CK_WORKAROUND_SWDEV_325164 0
 
-// workaround: disable broken fused attention kernel instance that does not pass validation
-// issue found on mi100/#10738 combo when irregular KPerBlock attention kernel has acc0 scaling
-// enabled
-#define CK_WORKAROUND_DISABLE_BROKEN_ATTN_KERNEL_INSTANCE 1
-
 namespace ck {
 
 enum struct InMemoryDataOperationEnum

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

@@ -18,11 +18,11 @@ enum struct LoopScheduler
 
 constexpr LoopScheduler make_default_loop_scheduler()
 {
-#if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
+#if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
     return LoopScheduler::Interwave;
 #else
     return LoopScheduler::Default;
-#endif // if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
+#endif // if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
 }
 
 template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename XElementwiseOperation,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceElementwiseNormalization : public BaseOperator
+{
+    static constexpr int NumInput = InDataTypeTuple::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        AccDataType epsilon,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        XElementwiseOperation x_elementwise_op,
+                        YElementwiseOperation y_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename XElementwiseOperation,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+using DeviceElementwiseNormalizationPtr =
+    std::unique_ptr<DeviceElementwiseNormalization<InDataTypeTuple,
+                                                   GammaDataType,
+                                                   BetaDataType,
+                                                   AccDataType,
+                                                   YDataType,
+                                                   XElementwiseOperation,
+                                                   YElementwiseOperation,
+                                                   Rank,
+                                                   NumReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::index_t NDimSpatial,
+          typename InputLayout,
+          typename WeightLayout,
+          typename OutputLayout,
+          typename InputDataType,
+          typename WeightDataType,
+          typename OutputDataType,
+          typename InputElementwiseOperation,
+          typename WeightElementwiseOperation,
+          typename OutputElementwiseOperation>
+struct DeviceGroupedConvBwdData : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_input,
+                        const void* p_weight,
+                        const void* p_output,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& weight_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& weight_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& output_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& output_g_n_k_wos_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const InputElementwiseOperation& input_element_op,
+                        const WeightElementwiseOperation& weight_element_op,
+                        const OutputElementwiseOperation& output_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -6,6 +6,7 @@
 #include <vector>
 
 #include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -62,6 +63,100 @@ struct DeviceGroupedConvBwdDataMultipleD : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
+                                         ALayout,
+                                         BLayout,
+                                         Tuple<>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         Tuple<>,
+                                         EDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CDEElementwiseOperation>
+    : public DeviceGroupedConvBwdData<NDimSpatial,
+                                      ELayout,
+                                      BLayout,
+                                      ALayout,
+                                      EDataType,
+                                      BDataType,
+                                      ADataType,
+                                      CDEElementwiseOperation,
+                                      BElementwiseOperation,
+                                      AElementwiseOperation>
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,                                                 // output image
+        const void* p_b,                                                 // weight
+        const std::array<const void*, 0>&,                               // bias
+        void* p_e,                                                       // input image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides, // output image
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,  // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,  // weight
+        const std::array<std::array<index_t, NDimSpatial + 3>, 0>&,      // bias
+        const std::array<std::array<index_t, NDimSpatial + 3>, 0>&,      // bias
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, // input image
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides, // input image
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) = 0;
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_input,
+                        const void* p_weight,
+                        const void* p_output,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& weight_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& weight_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& output_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& output_g_n_k_wos_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const CDEElementwiseOperation& input_element_op,
+                        const BElementwiseOperation& weight_element_op,
+                        const AElementwiseOperation& output_element_op) override final
+    {
+        return MakeArgumentPointer(p_output,
+                                   p_weight,
+                                   std::array<const void*, 0>{},
+                                   p_input,
+                                   output_g_n_k_wos_lengths,
+                                   output_g_n_k_wos_strides,
+                                   weight_g_k_c_xs_lengths,
+                                   weight_g_k_c_xs_strides,
+                                   std::array<std::array<index_t, NDimSpatial + 3>, 0>{},
+                                   std::array<std::array<index_t, NDimSpatial + 3>, 0>{},
+                                   input_g_n_c_wis_lengths,
+                                   input_g_n_c_wis_strides,
+                                   conv_filter_strides,
+                                   conv_filter_dilations,
+                                   input_left_pads,
+                                   input_right_pads,
+                                   output_element_op,
+                                   weight_element_op,
+                                   input_element_op);
+    }
+};
+
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

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

@@ -33,6 +33,8 @@ struct DeviceNormalization : public BaseOperator
                         const void* p_gamma,
                         const void* p_beta,
                         void* p_y,
+                        void* p_savedMean,
+                        void* p_savedInvVar,
                         AccElementwiseOperation acc_elementwise_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;

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

@@ -6,6 +6,7 @@
 #include <memory>
 #include <vector>
 
+#include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/device_base.hpp"
 
 namespace ck {

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp

@@ -150,7 +150,10 @@ template <typename ADataType,
           ck::index_t BBlockTransferDstScalarPerVector_K1,
           bool BBlockLdsAddExtraN,
           ck::index_t CThreadTransferSrcDstVectorDim,
-          ck::index_t CThreadTransferDstScalarPerVector>
+          ck::index_t CThreadTransferDstScalarPerVector,
+          ck::index_t NumGemmKPrefetchStage = 1,
+          ck::LoopScheduler LoopSched       = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer   = ck::PipelineVersion::v1>
 struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
                                                        BLayout,
                                                        CLayout,
@@ -323,7 +326,10 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
                                                 BBlockLdsAddExtraN,
                                                 Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
                                                 CThreadTransferSrcDstVectorDim,
-                                                CThreadTransferDstScalarPerVector>;
+                                                CThreadTransferDstScalarPerVector,
+                                                NumGemmKPrefetchStage,
+                                                LoopSched,
+                                                PipelineVer>;
 
     using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
         decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
@@ -622,6 +628,12 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
     {
         auto str = std::stringstream();
 
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
         // clang-format off
         str << "DeviceBatchedGemmXdl"
             << "<"
@@ -629,7 +641,13 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
             << MPerBlock << ", "
             << NPerBlock << ", "
             << K0PerBlock
-            << ">";
+            << ">"
+            << " NumGemmKPrefetchStage: "
+            << NumGemmKPrefetchStage << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+// X = Elementwise(input1, input2, input3, ...)
+// Y = Normalization(X, beta, gamma)
+namespace ck {
+template <typename GridwiseElementwiseReduction,
+          typename InDataTypePointerTuple, // Datatype tuple of inputs
+          typename XDataType,              // Datatype of X
+          typename GammaDataType,          // Datatype of Gamma
+          typename BetaDataType,           // Datatype of Beta
+          typename YDataType,              // Datatype of Y
+          typename AccDataType,            // AccDatatype
+          typename XElementwiseOperation,  // Operation of input
+          typename YElementwiseOperation,  // Operation of output of normalization
+          typename InGrid2dDescTuple,      // Descriptor tuple of inputs
+          typename GridDesc_M_K>           // Descriptor of inputs, Gamma, Beta
+__global__ void kernel_elementwise_layernorm(
+    const InGrid2dDescTuple in_grid_2d_desc_tuple,          // Descriptor tuple of inputs
+    const GridDesc_M_K x_grid_desc_m_k,                     // Descriptor of X
+    const GridDesc_M_K gamma_grid_desc_m_k,                 // Descriptor of gamma
+    const GridDesc_M_K beta_grid_desc_m_k,                  // Descriptor of beta
+    const GridDesc_M_K y_grid_desc_m_k,                     // Descriptor of Y
+    index_t num_k_block_tile_iteration,                     //
+    AccDataType epsilon,                                    // Datatype of epsilon
+    const InDataTypePointerTuple p_in_global_tuple,         // Ptr tuple of input matrixs
+    const GammaDataType* const __restrict__ p_gamma_global, // Ptr of gamma
+    const BetaDataType* const __restrict__ p_beta_global,   // Ptr of beta
+    YDataType* const __restrict__ p_y_global,               // Ptr of y
+    const XElementwiseOperation x_elementwise_op,           // Operation of input
+    const YElementwiseOperation y_elementwise_op)           // Operation of output of normalization
+{
+    extern __shared__ XDataType p_x_lds[];
+    GridwiseElementwiseReduction::Run(in_grid_2d_desc_tuple,      // Descriptor tuple of inputs
+                                      x_grid_desc_m_k,            // Descriptor of X
+                                      gamma_grid_desc_m_k,        // Descriptor of Gamma
+                                      beta_grid_desc_m_k,         // Descriptor of Beta
+                                      y_grid_desc_m_k,            // Descriptor of Y
+                                      num_k_block_tile_iteration, //
+                                      epsilon,                    // epsilon
+                                      p_in_global_tuple,          // Ptr tuple of inputs
+                                      p_x_lds,                    // Ptr of X
+                                      p_gamma_global,             // Ptr of gamma
+                                      p_beta_global,              // Ptr of beta
+                                      p_y_global,                 // Ptr of Y
+                                      x_elementwise_op,           // Operation of input
+                                      y_elementwise_op); // Operation of output of normalization
+};
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Y = LayerNorm(A + B, Beta, Gamma)
+template <typename InDataTypeTuple,       // Datatype of inputs
+          typename GammaDataType,         // Datatype of gamma
+          typename BetaDataType,          // Datatype of beta
+          typename AccDataType,           //
+          typename YDataType,             //
+          typename XElementwiseOperation, //
+          typename YElementwiseOperation, //
+          index_t Rank,                   //
+          index_t NumReduceDim,           //
+          index_t BlockSize,              //
+          index_t MThreadClusterSize,     // Num of threads in a block on M direction
+          index_t KThreadClusterSize,     // Num of threads in a block on N direction
+          index_t MThreadSliceSize,       // Each thread calculate rows
+          index_t KThreadSliceSize,       // Each thread calculate columns
+          index_t XYSrcVectorDim,         // Dimension to do reduce
+          index_t XSrcVectorSize,         // Size to fetch source x
+          index_t GammaSrcVectorDim,      // Dimension for gamma to do reduce
+          index_t GammaSrcVectorSize,     // Size to fetch source gamma
+          index_t BetaSrcVectorDim,       // Dimension for beta to do reduce
+          index_t BetaSrcVectorSize,      // Size to fetch source beta
+          index_t YDstVectorSize>         // Size to write destination Y
+struct DeviceElementwiseNormalizationImpl
+    : public DeviceElementwiseNormalization<InDataTypeTuple,
+                                            GammaDataType,
+                                            BetaDataType,
+                                            AccDataType,
+                                            YDataType,
+                                            XElementwiseOperation,
+                                            YElementwiseOperation,
+                                            Rank,
+                                            NumReduceDim>
+{
+    static constexpr int NumInput = InDataTypeTuple::Size();
+
+    using XDataType = YDataType;
+
+    static_assert(
+        (KThreadSliceSize % GammaSrcVectorSize == 0),
+        "Invalid thread slice sizes and/or gamma vector sizes configuration, please check!");
+
+    static_assert(
+        (KThreadSliceSize % BetaSrcVectorSize == 0),
+        "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
+
+    static constexpr index_t M_BlockTileSize =
+        MThreadClusterSize * MThreadSliceSize; // num of rows calculated in a block
+    static constexpr index_t K_BlockTileSize =
+        KThreadClusterSize * KThreadSliceSize; // num of columns calculated in a block
+
+    static auto GenerateInDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                return static_cast<const DataType*>(nullptr);
+            },
+            Number<NumInput>{});
+    };
+
+    using InDataTypePointerTuple = decltype(GenerateInDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
+        static constexpr index_t numSrcDim = Rank;
+        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+        const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
+        const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
+                const auto invariantDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    template <index_t TupleSize>
+    static auto GenerateSrcGrid2dDescTuple(Number<TupleSize>)
+    {
+        return generate_tuple([&](auto) { return MakeSrc2dDescriptor({1}, {1}, 1, 1); },
+                              Number<TupleSize>{});
+    };
+
+    using InGrid2dDescTuple = decltype(GenerateSrcGrid2dDescTuple(Number<NumInput>{}));
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+
+    using GridwiseReduceLayernormGeneric =
+        GridwiseElementwiseLayernormWelfordVariance_mk_to_mk<InDataTypePointerTuple,
+                                                             XDataType,
+                                                             GammaDataType,
+                                                             BetaDataType,
+                                                             YDataType,
+                                                             AccDataType,
+                                                             XElementwiseOperation,
+                                                             YElementwiseOperation,
+                                                             InGrid2dDescTuple,
+                                                             GridDesc_M_K,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             XYSrcVectorDim,
+                                                             XSrcVectorSize,
+                                                             GammaSrcVectorDim,
+                                                             GammaSrcVectorSize,
+                                                             BetaSrcVectorDim,
+                                                             BetaSrcVectorSize,
+                                                             XYSrcVectorDim,
+                                                             YDstVectorSize,
+                                                             false>;
+
+    using GridwiseReduceLayernormSweepOnce =
+        GridwiseElementwiseLayernormWelfordVariance_mk_to_mk<InDataTypePointerTuple,
+                                                             XDataType,
+                                                             GammaDataType,
+                                                             BetaDataType,
+                                                             YDataType,
+                                                             AccDataType,
+                                                             XElementwiseOperation,
+                                                             YElementwiseOperation,
+                                                             InGrid2dDescTuple,
+                                                             GridDesc_M_K,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             XYSrcVectorDim,
+                                                             XSrcVectorSize,
+                                                             GammaSrcVectorDim,
+                                                             GammaSrcVectorSize,
+                                                             BetaSrcVectorDim,
+                                                             BetaSrcVectorSize,
+                                                             XYSrcVectorDim,
+                                                             YDstVectorSize,
+                                                             true>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> lengths,
+                 const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                 const std::vector<index_t> gammaStrides,
+                 const std::vector<index_t> betaStrides,
+                 const std::vector<index_t> yStrides,
+                 const std::vector<index_t> reduceDims,
+                 XElementwiseOperation x_elementwise_op,
+                 YElementwiseOperation y_elementwise_op,
+                 AccDataType epsilon,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 YDataType* p_y)
+            : epsilon_(epsilon),
+              p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              p_y_(p_y),
+              x_elementwise_op_(x_elementwise_op),
+              y_elementwise_op_(y_elementwise_op)
+        {
+
+            Lengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
+            for(int i = 0; i < NumInput; i++)
+            {
+                inStridesArray_[i] =
+                    shuffle_tensor_dimensions<Rank, NumReduceDim>(inStridesArray[i], reduceDims);
+            }
+
+            yStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+            xStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+
+            gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
+            betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+
+            in_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                    return static_cast<const DataType*>(in_dev_buffers[I.value]);
+                },
+                Number<NumInput>{});
+
+            long_index_t invariant_total_length;
+            long_index_t reduce_total_length;
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(Lengths_);
+
+            blkGroupSize_          = 1;
+            numBlockTileIteration_ = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize_ = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                        M_BlockTileSize * blkGroupSize_;
+
+            in_grid_2d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeSrc2dDescriptor(
+                        Lengths_, inStridesArray_[I.value], blkGroupSize_, numBlockTileIteration_);
+                },
+                Number<NumInput>{});
+
+            x_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            gamma_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, gammaStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            beta_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, betaStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            y_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            sweep_once_ =
+                x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+
+            if(!sweep_once_) // if not sweep once, compute memory size for matrix X in lds for
+                             // store Intermediate results
+            {
+                int block_TileSize = M_BlockTileSize * reduce_total_length;
+                x_lds_size_        = block_TileSize * sizeof(XDataType);
+            }
+            else
+                x_lds_size_ = 0;
+        }
+
+        AccDataType epsilon_;
+
+        InDataTypePointerTuple in_dev_buffers_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        YDataType* p_y_;
+
+        std::vector<index_t> Lengths_;
+        std::array<std::vector<index_t>, NumInput> inStridesArray_;
+        std::vector<index_t> xStrides_;
+        std::vector<index_t> gammaStrides_;
+        std::vector<index_t> betaStrides_;
+        std::vector<index_t> yStrides_;
+
+        XElementwiseOperation x_elementwise_op_;
+        YElementwiseOperation y_elementwise_op_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        InGrid2dDescTuple in_grid_2d_desc_tuple_;
+        GridDesc_M_K x_grid_desc_m_k_;
+        GridDesc_M_K gamma_grid_desc_m_k_;
+        GridDesc_M_K beta_grid_desc_m_k_;
+        GridDesc_M_K y_grid_desc_m_k_;
+        bool sweep_once_;
+        int x_lds_size_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel_main =
+                arg.sweep_once_ ? kernel_elementwise_layernorm<GridwiseReduceLayernormSweepOnce,
+                                                               InDataTypePointerTuple,
+                                                               XDataType,
+                                                               GammaDataType,
+                                                               BetaDataType,
+                                                               YDataType,
+                                                               AccDataType,
+                                                               XElementwiseOperation,
+                                                               YElementwiseOperation,
+                                                               InGrid2dDescTuple,
+                                                               GridDesc_M_K>
+                                : kernel_elementwise_layernorm<GridwiseReduceLayernormGeneric,
+                                                               InDataTypePointerTuple,
+                                                               XDataType,
+                                                               GammaDataType,
+                                                               BetaDataType,
+                                                               YDataType,
+                                                               AccDataType,
+                                                               XElementwiseOperation,
+                                                               YElementwiseOperation,
+                                                               InGrid2dDescTuple,
+                                                               GridDesc_M_K>;
+
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize_),
+                                               dim3(BlockSize),
+                                               arg.x_lds_size_,
+                                               arg.in_grid_2d_desc_tuple_,
+                                               arg.x_grid_desc_m_k_,
+                                               arg.gamma_grid_desc_m_k_,
+                                               arg.beta_grid_desc_m_k_,
+                                               arg.y_grid_desc_m_k_,
+                                               arg.numBlockTileIteration_,
+                                               arg.epsilon_,
+                                               arg.in_dev_buffers_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               arg.p_y_,
+                                               arg.x_elementwise_op_,
+                                               arg.y_elementwise_op_);
+
+            return (avg_time);
+        };
+
+        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* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                for(int i = 0; i < NumInput; i++)
+                {
+                    if(p_arg_->inStridesArray_[i][NumInvariantDim - 1] != 1)
+                        return false;
+                }
+
+                if(p_arg_->inStridesArray_[0][NumInvariantDim - 1] != 1 &&
+                   p_arg_->inStridesArray_[1][NumInvariantDim - 1] != 1)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                    return false;
+            };
+        }
+        else
+        {
+            for(int i = 0; i < NumInput; i++)
+            {
+                if(p_arg_->inStridesArray_[i][Rank - 1] != 1)
+                    return false;
+            }
+
+            if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
+                return false;
+        };
+
+        if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        auto IsScalarPerVectorValid = [](bool isLastDimensionCoalesced, int scalarPerVector) {
+            bool ret = true;
+
+            if(!isLastDimensionCoalesced)
+                ret = scalarPerVector == 1;
+            else
+                ret = KThreadSliceSize % scalarPerVector == 0;
+
+            return ret;
+        };
+
+        if(!IsScalarPerVectorValid(p_arg_->gammaStrides_.back() == 1, GammaSrcVectorSize))
+            return false;
+
+        if(!IsScalarPerVectorValid(p_arg_->betaStrides_.back() == 1, BetaSrcVectorSize))
+            return false;
+
+        // if fastest dim is not reduced
+        if constexpr(XYSrcVectorDim == 0) //
+        {
+            if(p_arg_->gammaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->gammaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->betaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % BetaSrcVectorSize != 0)
+                return (false);
+        }
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        AccDataType epsilon,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        XElementwiseOperation x_elementwise_op,
+                        YElementwiseOperation y_elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          inStridesArray,
+                                          gammaStrides,
+                                          betaStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          x_elementwise_op,
+                                          y_elementwise_op,
+                                          epsilon,
+                                          in_dev_buffers,
+                                          static_cast<const GammaDataType*>(p_gamma),
+                                          static_cast<const BetaDataType*>(p_beta),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceElementwiseNormalizationImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XYSrcVectorDim_" << XYSrcVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_Gamma" << GammaSrcVectorSize << "_Beta" << BetaSrcVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp

@@ -141,7 +141,8 @@ template <typename ALayout,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEBlockTransferScalarPerVector_NPerBlock,
-          LoopScheduler LoopSched = make_default_loop_scheduler()>
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1>
 struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
                                                                      BLayout,
                                                                      DsLayout,
@@ -282,7 +283,8 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
         CShuffleNXdlPerWavePerShuffle,
         CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         CDEBlockTransferScalarPerVector_NPerBlock,
-        LoopSched>;
+        LoopSched,
+        PipelineVer>;
 
     // desc for blockwise copy
     using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
@@ -664,6 +666,12 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
     {
         auto str = std::stringstream();
 
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
         // clang-format off
         str << "DeviceGemmMultipleD_Xdl_CShuffle"
             << "<"
@@ -674,7 +682,11 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
             << AK1 << ", "
             << BK1 << ", "
             << getGemmSpecializationString(GemmSpec)
-            << ">";
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp

@@ -56,7 +56,9 @@ template <typename ADataType,
           bool BBlockLdsAddExtraN,
           ck::index_t CThreadTransferSrcDstVectorDim,
           ck::index_t CThreadTransferDstScalarPerVector,
-          ck::index_t NumPrefetch = 1>
+          ck::index_t NumPrefetch         = 1,
+          ck::LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
 struct DeviceGemmXdl : public DeviceGemm<ALayout,
                                          BLayout,
                                          CLayout,
@@ -230,7 +232,9 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
         Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
         CThreadTransferSrcDstVectorDim,
         CThreadTransferDstScalarPerVector,
-        NumPrefetch>;
+        NumPrefetch,
+        LoopSched,
+        PipelineVer>;
 
     // Argument
     struct Argument : public BaseArgument
@@ -523,6 +527,12 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
     {
         auto str = std::stringstream();
 
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
         // clang-format off
         str << "DeviceGemmXdl"
             << "<"
@@ -535,7 +545,13 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
             << NPerXDL << ", "
             << MXdlPerWave << ", "
             << NXdlPerWave
-            << ">";
+            << ">"
+            << " NumPrefetch: "
+            << NumPrefetch << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp

@@ -64,7 +64,8 @@ template <typename ALayout,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
-          LoopScheduler LoopSched = make_default_loop_scheduler()>
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1>
 struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
                                                    BLayout,
                                                    CLayout,
@@ -393,7 +394,8 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
         CShuffleNXdlPerWavePerShuffle,
         CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         CShuffleBlockTransferScalarPerVector_NPerBlock,
-        LoopSched>;
+        LoopSched,
+        PipelineVer>;
 
     // Argument
     struct Argument : public BaseArgument
@@ -656,6 +658,12 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
     {
         auto str = std::stringstream();
 
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
         // clang-format off
         str << "DeviceGemm_Xdl_CShuffle"
             << "<"
@@ -665,7 +673,11 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
             << KPerBlock << ", "
             << AK1 << ", "
             << BK1
-            << ">";
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];;
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp

@@ -10,7 +10,7 @@
 #include "ck/tensor_operation/gpu/device/device_normalization.hpp"
 #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_layernorm_welford_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
@@ -24,17 +24,17 @@ template <typename GridwiseReduction,
           typename AccDataType,
           typename AccElementwiseOperation,
           typename GridDesc_M_K>
-__global__ void kernel_layernorm(const GridDesc_M_K x_grid_desc_m_k,
-                                 const GridDesc_M_K gamma_grid_desc_m_k,
-                                 const GridDesc_M_K beta_grid_desc_m_k,
-                                 const GridDesc_M_K y_grid_desc_m_k,
-                                 index_t num_k_block_tile_iteration,
-                                 AccDataType epsilon,
-                                 const XDataType* const __restrict__ p_x_global,
-                                 const GammaDataType* const __restrict__ p_gamma_global,
-                                 const BetaDataType* const __restrict__ p_beta_global,
-                                 YDataType* const __restrict__ p_y_global,
-                                 const AccElementwiseOperation acc_elementwise_op)
+__global__ void kernel_normalization(const GridDesc_M_K x_grid_desc_m_k,
+                                     const GridDesc_M_K gamma_grid_desc_m_k,
+                                     const GridDesc_M_K beta_grid_desc_m_k,
+                                     const GridDesc_M_K y_grid_desc_m_k,
+                                     index_t num_k_block_tile_iteration,
+                                     AccDataType epsilon,
+                                     const XDataType* const __restrict__ p_x_global,
+                                     const GammaDataType* const __restrict__ p_gamma_global,
+                                     const BetaDataType* const __restrict__ p_beta_global,
+                                     YDataType* const __restrict__ p_y_global,
+                                     const AccElementwiseOperation acc_elementwise_op)
 {
     GridwiseReduction::Run(x_grid_desc_m_k,
                            gamma_grid_desc_m_k,
@@ -54,7 +54,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-// Y = LayerNorm(X, Beta, Gamma)
+// Y = Normalization(X, Beta, Gamma)
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
@@ -168,49 +168,49 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
     using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
 
     using GridwiseReduceLayernormGeneric =
-        GridwiseLayernormWelfordVariance_mk_to_mk<XDataType,
-                                                  GammaDataType,
-                                                  BetaDataType,
-                                                  YDataType,
-                                                  AccDataType,
-                                                  AccElementwiseOperation,
-                                                  GridDesc_M_K,
-                                                  BlockSize,
-                                                  MThreadClusterSize,
-                                                  KThreadClusterSize,
-                                                  MThreadSliceSize,
-                                                  KThreadSliceSize,
-                                                  XYSrcVectorDim,
-                                                  XSrcVectorSize,
-                                                  GammaSrcVectorDim,
-                                                  GammaSrcVectorSize,
-                                                  BetaSrcVectorDim,
-                                                  BetaSrcVectorSize,
-                                                  XYSrcVectorDim,
-                                                  YDstVectorSize,
-                                                  false>;
-    using GridwiseReduceLayernormSweepOnce =
-        GridwiseLayernormWelfordVariance_mk_to_mk<XDataType,
-                                                  GammaDataType,
-                                                  BetaDataType,
-                                                  YDataType,
-                                                  AccDataType,
-                                                  AccElementwiseOperation,
-                                                  GridDesc_M_K,
-                                                  BlockSize,
-                                                  MThreadClusterSize,
-                                                  KThreadClusterSize,
-                                                  MThreadSliceSize,
-                                                  KThreadSliceSize,
-                                                  XYSrcVectorDim,
-                                                  XSrcVectorSize,
-                                                  GammaSrcVectorDim,
-                                                  GammaSrcVectorSize,
-                                                  BetaSrcVectorDim,
-                                                  BetaSrcVectorSize,
-                                                  XYSrcVectorDim,
-                                                  YDstVectorSize,
-                                                  true>;
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      AccDataType,
+                                                      AccElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XYSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      XYSrcVectorDim,
+                                                      YDstVectorSize,
+                                                      false>;
+    using GridwiseNormalizationSweepOnce =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      AccDataType,
+                                                      AccElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XYSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      XYSrcVectorDim,
+                                                      YDstVectorSize,
+                                                      true>;
 
     struct Argument : public BaseArgument
     {
@@ -295,22 +295,22 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
             const auto kernel_main = arg.isSweeponce_
-                                         ? kernel_layernorm<GridwiseReduceLayernormSweepOnce,
-                                                            XDataType,
-                                                            GammaDataType,
-                                                            BetaDataType,
-                                                            YDataType,
-                                                            AccDataType,
-                                                            AccElementwiseOperation,
-                                                            GridDesc_M_K>
-                                         : kernel_layernorm<GridwiseReduceLayernormGeneric,
-                                                            XDataType,
-                                                            GammaDataType,
-                                                            BetaDataType,
-                                                            YDataType,
-                                                            AccDataType,
-                                                            AccElementwiseOperation,
-                                                            GridDesc_M_K>;
+                                         ? kernel_normalization<GridwiseNormalizationSweepOnce,
+                                                                XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                YDataType,
+                                                                AccDataType,
+                                                                AccElementwiseOperation,
+                                                                GridDesc_M_K>
+                                         : kernel_normalization<GridwiseReduceLayernormGeneric,
+                                                                XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                YDataType,
+                                                                AccDataType,
+                                                                AccElementwiseOperation,
+                                                                GridDesc_M_K>;
 
             float avg_time = 0;
             avg_time += launch_and_time_kernel(stream_config,
@@ -426,8 +426,16 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
                         const void* p_gamma,
                         const void* p_beta,
                         void* p_y,
+                        void* p_saveMean,
+                        void* p_saveInvVar,
                         AccElementwiseOperation acc_elementwise_op) override
     {
+        // TODO
+        // Optional cache of the intermediate results (mean and InvVariance) during the
+        // forward pass could speedup in the backward
+        ignore = p_saveMean;
+        ignore = p_saveInvVar;
+
         return std::make_unique<Argument>(lengths,
                                           xStrides,
                                           gammaStrides,

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

@@ -226,6 +226,30 @@ struct DeviceReduceMultiBlock
               in_elementwise_op_{in_elementwise_op},
               acc_elementwise_op_{acc_elementwise_op}
         {
+            if(Rank != inLengths.size() || Rank != inStrides.size() ||
+               NumReduceDim != reduceDims.size())
+            {
+                throw std::runtime_error(
+                    "One of inLengths/inStrides/reduceDims has invalid size!"
+                    "\nExpected size inLengths: " +
+                    std::to_string(Rank) + ", inStrides: " + std::to_string(Rank) +
+                    ", reduceDims: " + std::to_string(NumReduceDim) +
+                    "\nBut have inLengths: " + std::to_string(inLengths.size()) +
+                    ", inStrides: " + std::to_string(inStrides.size()) +
+                    ", reduceDims: " + std::to_string(reduceDims.size()));
+            }
+
+            for(std::size_t i = 0; i < reduceDims.size(); ++i)
+            {
+                if(reduceDims[i] < 0 || reduceDims[i] >= Rank)
+                {
+                    throw std::runtime_error("Provided reduce dimension exceed input tensor Rank!"
+                                             "\nHave reduceDims[" +
+                                             std::to_string(i) +
+                                             "]: " + std::to_string(reduceDims[i]));
+                }
+            }
+
             inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
             inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
 

include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp

@@ -40,8 +40,9 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
                                                 AccElementwiseOp,
                                                 Rank>
 {
-    static constexpr index_t kRank         = Rank;
-    static constexpr index_t kNumReduceDim = NumReduceDim;
+    static constexpr index_t kRank            = Rank;
+    static constexpr index_t kNumReduceDim    = NumReduceDim;
+    static constexpr index_t kNumInvariantDim = Rank - NumReduceDim;
 
     virtual index_t GetRank() const override { return kRank; }
 
@@ -168,6 +169,30 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
               in_elementwise_op_{in_elementwise_op},
               acc_elementwise_op_{acc_elementwise_op}
         {
+            if(Rank != inLengths.size() || Rank != inStrides.size() ||
+               NumReduceDim != reduceDims.size())
+            {
+                throw std::runtime_error(
+                    "One of inLengths/inStrides/reduceDims has invalid size!"
+                    "\nExpected size inLengths: " +
+                    std::to_string(Rank) + ", inStrides: " + std::to_string(Rank) +
+                    ", reduceDims: " + std::to_string(NumReduceDim) +
+                    "\nBut have inLengths: " + std::to_string(inLengths.size()) +
+                    ", inStrides: " + std::to_string(inStrides.size()) +
+                    ", reduceDims: " + std::to_string(reduceDims.size()));
+            }
+
+            for(std::size_t i = 0; i < reduceDims.size(); ++i)
+            {
+                if(reduceDims[i] < 0 || reduceDims[i] >= Rank)
+                {
+                    throw std::runtime_error("Provided reduce dimension exceed input tensor Rank!"
+                                             "\nHave reduceDims[" +
+                                             std::to_string(i) +
+                                             "]: " + std::to_string(reduceDims[i]));
+                }
+            }
+
             inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
             inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
 
@@ -257,40 +282,78 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
         };
     };
 
-    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    static bool IsSupportedArgument(const Argument& arg)
     {
-        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
-
         if constexpr(InSrcVectorDim == 0)
         {
-            if constexpr(NumInvariantDim == 0)
+            if constexpr(kNumInvariantDim == 0)
             {
                 return false;
             }
             else
             {
-                if(p_arg_->inStrides_[NumInvariantDim - 1] != 1)
+                if(arg.inStrides_[kNumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                {
                     return false;
-
-                if(p_arg_->invariant_lowest_length_ % InSrcVectorSize != 0)
+                }
+                if(arg.invariant_lowest_length_ % InSrcVectorSize != 0)
+                {
                     return false;
-            };
+                }
+            }
         }
         else
         {
-            if(p_arg_->inStrides_[Rank - 1] != 1)
+            if(arg.inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+            {
                 return false;
-
-            if(p_arg_->inLengths_[Rank - 1] % InSrcVectorSize != 0)
+            }
+            if(arg.inLengths_[Rank - 1] % InSrcVectorSize != 0)
+            {
                 return false;
-        };
+            }
+        }
+
+        // To improve
+        if(kNumInvariantDim > 0 && arg.invariant_lowest_length_ % OutDstVectorSize != 0)
+        {
+            return false;
+        }
 
-        if(p_arg_->invariant_lowest_length_ % OutDstVectorSize != 0)
+        if(arg.inLengths_[Rank - 1] % OutDstVectorSize != 0)
+        {
             return false;
+        }
 
         return true;
     };
 
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const std::vector<index_t> inLengths,
+                             const std::vector<index_t> inStrides,
+                             const std::vector<int> reduceDims,
+                             const AccDataType alpha,
+                             const AccDataType beta,
+                             const InDataType* in_dev,
+                             OutDataType* out_dev,
+                             InElementwiseOp in_elementwise_op,
+                             AccElementwiseOp acc_elementwise_op)
+    {
+        return Argument{inLengths,
+                        inStrides,
+                        reduceDims,
+                        alpha,
+                        beta,
+                        in_dev,
+                        out_dev,
+                        in_elementwise_op,
+                        acc_elementwise_op};
+    };
+
     //
     // @brief      Makes a pointer to Argument class.
     //
@@ -330,6 +393,8 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
                                           acc_elementwise_op);
     };
 
+    static auto MakeInvoker() { return Invoker{}; }
+
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
         return std::make_unique<Invoker>();
@@ -340,10 +405,13 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
         auto str = std::stringstream();
 
         // clang-format off
-        str << "DeviceReduceSoftmax<" << BlockSize << ",";
-        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
-        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
-        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << "_OutDstVectorSize_" << OutDstVectorSize << ">";
+        str << "DeviceReduceSoftmax<" 
+            << Rank << "," << NumReduceDim << "," << BlockSize << ","
+            << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ","
+            << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ","
+            << "InSrcVectorDim_" << InSrcVectorDim 
+            << "_InSrcVectorSize_" << InSrcVectorSize 
+            << "_OutDstVectorSize_" << OutDstVectorSize << ">";
         // clang-format on
 
         return str.str();