Merge pull request #134 from ROCm/merge_from_public

Merge from public

example/12_reduce/reduce_threadwise_multi_d_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_threadwise_multi_d.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
+
+#include "ck/library/utility/algorithm.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/host_common_util.hpp"
+
+#include "reduce_example_common.hpp"
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t Rank,
+          ck::index_t NumReduceDim,
+          bool PropagateNan,
+          bool OutputIndex>
+int reduce_threadwise_multi_d_impl(bool do_verification,
+                                   int init_method,
+                                   bool time_kernel,
+                                   const std::vector<size_t>& inLengths,
+                                   const std::array<int, NumReduceDim>& reduceDims,
+                                   float alpha,
+                                   float beta)
+
+{
+    using namespace ck;
+    using namespace ck::tensor_operation::device;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_indices =
+        (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
+         ReduceOpId == ReduceTensorOp::AMAX);
+
+    constexpr bool invalid_reduce_1 = OutputIndex && !op_support_indices;
+
+    // 1) If InOutDataType is half_t, must use half_t as AccDataType for indexable reduction
+    // operations 2) If InOutDataType is half_t, must use float as AccDataType for non-indexable
+    // reduction operations
+    constexpr bool invalid_reduce_2 =
+        std::is_same<InOutDataType, half_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, float>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, half_t>::value));
+
+    // 1) If InOutDataType is float, must use float as AccDataType for indexable reduction
+    // operations
+    constexpr bool invalid_reduce_3 =
+        std::is_same<InOutDataType, float>::value &&
+        (op_support_indices && !std::is_same<AccDataType, float>::value);
+
+    // 1) If InOutDataType is int8_t or int4_t, must use int8_t as AccDataType for indexable
+    // reduction operations 2) If InOutDataType is int8_t or int4_t, must use int32_t as AccDataType
+    // for non-indexable reduction operations
+    constexpr bool invalid_reduce_4 =
+        std::is_same<InOutDataType, int8_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+
+    // 1) If InOutDataType is int8_t or int4_t, the supported operation must be either indexable
+    // operations or ADD/AVG
+    constexpr bool invalid_reduce_5 = std::is_same<InOutDataType, int8_t>::value &&
+                                      (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                       ReduceOpId != ReduceTensorOp::AVG);
+
+    // 1) If InOutDataType is bhalf_t, must use float as AccDataType for all reduction operations
+    constexpr bool invalid_reduce_6 =
+        std::is_same<InOutDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value;
+
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
+                                     invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
+
+    if constexpr(invalid_reduce)
+    {
+        std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
+        return (-1);
+    };
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+    using Add         = tensor_operation::element_wise::Add;
+
+    using ReduceOperation         = typename reduce_binary_operator<ReduceOpId>::opType;
+    using InElementwiseOperation  = PassThrough;
+    using OutElementwiseOperation = Add;
+
+    using InOutDataTypeInDevice = InOutDataType;
+
+    using DeviceReduceInstance =
+        ck::tensor_operation::device::DeviceReduceThreadWiseMultiD<InOutDataTypeInDevice,
+                                                                   ck::Tuple<InOutDataTypeInDevice>,
+                                                                   AccDataType,
+                                                                   InOutDataTypeInDevice,
+                                                                   Rank,
+                                                                   NumReduceDim,
+                                                                   ReduceOperation,
+                                                                   InElementwiseOperation,
+                                                                   OutElementwiseOperation,
+                                                                   256, // BlockSize
+                                                                   4,   // MThreadSliceSize
+                                                                   1,   // KThreadSliceSize
+                                                                   0,   // InSrcVectorDim
+                                                                   1,   // InSrceVectorSize
+                                                                   1,
+                                                                   Sequence<1>>; // OutDstVectorSize
+
+    Tensor<InOutDataType> in(inLengths);
+
+    std::vector<size_t> outLengths;
+
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+    if(invariantDims.empty())
+        outLengths.push_back(1);
+    else
+        for(auto dim : invariantDims)
+            outLengths.push_back(inLengths[dim]);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> out(outLengths);
+
+    Tensor<InOutDataType> d0(outLengths);
+
+    Tensor<int> out_indices_ref(outLengths);
+    Tensor<int> out_indices(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            d0.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            d0.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            d0.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InOutDataTypeInDevice) * in.mDesc.GetElementSpaceSize());
+    DeviceMem d0_dev(sizeof(InOutDataTypeInDevice) * d0.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataTypeInDevice) * out.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+    d0_dev.ToDevice(d0.mData.data());
+
+    if(beta != 0.0f)
+    {
+        out_dev.ToDevice(out.mData.data());
+    };
+
+    size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
+
+    DeviceMem out_index_dev(indicesSizeInBytes);
+
+    InElementwiseOperation in_elementwise_op;
+    OutElementwiseOperation out_elementwise_op;
+
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
+
+    ck::ranges::copy(inLengths, arrInLengths.begin());
+    ck::ranges::copy(inStrides, arrInStrides.begin());
+
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    if(do_verification)
+    {
+        using ReferenceReduceInstance =
+            ck::tensor_operation::host::ReferenceReduce<InOutDataType,
+                                                        AccDataType,
+                                                        InOutDataType,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperation,
+                                                        PassThrough,
+                                                        PropagateNan,
+                                                        OutputIndex>;
+
+        auto reduce_ref = ReferenceReduceInstance{};
+
+        auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths,
+                                                               arrInStrides,
+                                                               arrOutLengths,
+                                                               arrOutStrides,
+                                                               reduceDims,
+                                                               static_cast<double>(alpha),
+                                                               static_cast<double>(beta),
+                                                               in.mData.data(),
+                                                               nullptr,
+                                                               out_ref.mData.data(),
+                                                               out_indices_ref.mData.data(),
+                                                               in_elementwise_op,
+                                                               PassThrough{});
+
+        if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reduce reference, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
+
+        invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        for(std::size_t i = 0; i < out_ref.GetElementSize(); i++)
+            out_elementwise_op(out_ref.mData[i], out_ref.mData[i], d0.mData[i]);
+    };
+
+    auto reduce = DeviceReduceInstance{};
+
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   {arrOutLengths},
+                                                   {arrOutStrides},
+                                                   arrOutLengths,
+                                                   arrOutStrides,
+                                                   reduceDims,
+                                                   in_dev.GetDeviceBuffer(),
+                                                   {d0_dev.GetDeviceBuffer()},
+                                                   out_dev.GetDeviceBuffer(),
+                                                   in_elementwise_op,
+                                                   out_elementwise_op);
+
+    if(!reduce.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cerr << "The runtime parameters not supported by the DeviceReduce instance, exiting!"
+                  << std::endl;
+
+        return (-2);
+    };
+
+    std::string reduce_name = reduce.GetTypeString();
+
+    auto invoker_ptr = reduce.MakeInvokerPointer();
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+
+        out_dev.FromDevice(out.mData.data());
+
+        pass = pass && ck::utils::check_err(out, out_ref);
+
+        if(OutputIndex)
+        {
+            out_index_dev.FromDevice(out_indices.mData.data());
+            pass = pass && ck::utils::check_err(out_indices, out_indices_ref);
+        };
+    };
+
+    return (pass ? 0 : 1);
+}

example/17_convnd_bwd_data/convnd_bwd_data_common.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 #include <numeric>
@@ -80,6 +80,29 @@ int run_conv_bwd_data(bool do_verification,
     // reset input to zero
     in_device_buf.SetZero();
 
+    std::vector<ck::index_t> input_spatial_lengths_i32(NDimSpatial);
+    std::vector<ck::index_t> filter_spatial_lengths_i32(NDimSpatial);
+    std::vector<ck::index_t> output_spatial_lengths_i32(NDimSpatial);
+    std::vector<ck::index_t> conv_filter_strides_i32(NDimSpatial);
+    std::vector<ck::index_t> conv_filter_dilations_i32(NDimSpatial);
+    std::vector<ck::index_t> input_left_pads_i32(NDimSpatial);
+    std::vector<ck::index_t> input_right_pads_i32(NDimSpatial);
+
+    for(ck::index_t d = 0; d < NDimSpatial; d++)
+    {
+        input_spatial_lengths_i32[d] =
+            static_cast<ck::index_t>(conv_param.input_spatial_lengths_[d]);
+        filter_spatial_lengths_i32[d] =
+            static_cast<ck::index_t>(conv_param.filter_spatial_lengths_[d]);
+        output_spatial_lengths_i32[d] =
+            static_cast<ck::index_t>(conv_param.GetOutputSpatialLengths()[d]);
+        conv_filter_strides_i32[d] = static_cast<ck::index_t>(conv_param.conv_filter_strides_[d]);
+        conv_filter_dilations_i32[d] =
+            static_cast<ck::index_t>(conv_param.conv_filter_dilations_[d]);
+        input_left_pads_i32[d]  = static_cast<ck::index_t>(conv_param.input_left_pads_[d]);
+        input_right_pads_i32[d] = static_cast<ck::index_t>(conv_param.input_right_pads_[d]);
+    }
+
     // do GEMM
     auto conv    = DeviceConvNdBwdDataInstance{};
     auto invoker = conv.MakeInvoker();
@@ -87,16 +110,16 @@ int run_conv_bwd_data(bool do_verification,
         conv.MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                  static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                  static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                                 conv_param.N_,
-                                 conv_param.K_,
-                                 conv_param.C_,
-                                 conv_param.input_spatial_lengths_,
-                                 conv_param.filter_spatial_lengths_,
-                                 conv_param.GetOutputSpatialLengths(),
-                                 conv_param.conv_filter_strides_,
-                                 conv_param.conv_filter_dilations_,
-                                 conv_param.input_left_pads_,
-                                 conv_param.input_right_pads_,
+                                 static_cast<ck::index_t>(conv_param.N_),
+                                 static_cast<ck::index_t>(conv_param.K_),
+                                 static_cast<ck::index_t>(conv_param.C_),
+                                 input_spatial_lengths_i32,
+                                 filter_spatial_lengths_i32,
+                                 output_spatial_lengths_i32,
+                                 conv_filter_strides_i32,
+                                 conv_filter_dilations_i32,
+                                 input_left_pads_i32,
+                                 input_right_pads_i32,
                                  in_element_op,
                                  wei_element_op,
                                  out_element_op);

example/20_grouped_conv_bwd_weight/common.hpp

@@ -23,12 +23,8 @@
 using BF16 = ck::bhalf_t;
 using F16  = ck::half_t;
 using F32  = float;
-#ifdef CK_ENABLE_FP8
-using F8 = ck::f8_t;
-#endif
-#ifdef CK_ENABLE_BF8
-using BF8 = ck::bf8_t;
-#endif
+using F8   = ck::f8_t;
+using BF8  = ck::bf8_t;
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;

example/35_splitK_gemm/CMakeLists.txt

@@ -21,3 +21,9 @@ if(USE_BITINT_EXTENSION_INT4)
     add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
     add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
 endif()
+
+add_example_executable(example_gemm_xdl_splitk_reduce_multi_d_fp16 gemm_xdl_splitk_reduce_multi_d_fp16.cpp)
+add_example_executable(example_gemm_xdl_splitk_reduce_multi_d_bf16 gemm_xdl_splitk_reduce_multi_d_bf16.cpp)
+add_example_executable(example_gemm_xdl_splitk_reduce_bf16A_i8B gemm_xdl_splitk_reduce_bf16A_i8B.cpp)
+
+add_example_executable(example_gemm_xdl_splitk_reduce_bfp16 gemm_xdl_splitk_reduce_bf16.cpp)

example/35_splitK_gemm/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm_multiple_d.hpp"
+
+struct ProblemSizeSplitK final
+{
+    ck::index_t M = 256;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideC = N;
+
+    ck::index_t KBatch = 2;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = true;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+bool parse_cmd_args(int argc,
+                    char* argv[],
+                    ProblemSizeSplitK& problem_size,
+                    ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc >= 10)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA = std::stoi(argv[7]);
+        problem_size.StrideB = std::stoi(argv[8]);
+        problem_size.StrideC = std::stoi(argv[9]);
+
+        if(argc >= 11)
+        {
+            problem_size.KBatch = std::stoi(argv[10]);
+        }
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
+                  << "arg10: KBatch" << std::endl;
+        return false;
+    }
+
+    return true;
+}

example/35_splitK_gemm/gemm_xdl_splitk_reduce_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using ReduceDataType   = ck::bhalf_t;
+using D0DataType       = ck::bhalf_t;
+using DsDataType       = ck::Tuple<>;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using CLayout  = Row;
+using D0Layout = CLayout;
+using DsLayout = ck::Tuple<>;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3R1<
+        ALayout,   BLayout,   DsLayout,  CLayout,
+        ADataType,   BDataType, DsDataType,  CDataType, AccDataType,  CShuffleDataType,
+        AElementOp, BElementOp, CDEElementOp, GemmDefault, 
+        256,
+        128,  128,  64,
+        8,    4,
+        32,   32,
+        2,    2,
+        S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,
+        2,    8,    8,   0,
+        S<16, 16, 1>,    S<0, 2, 1>,    S<0, 2, 1>,
+        1,    8,    4,   0,
+        1,    1,    S<1, 32, 1, 8>,  8,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_splitk_reduce_multi_d_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

example/35_splitK_gemm/gemm_xdl_splitk_reduce_bf16A_i8B.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = int8_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using ReduceDataType   = float;
+using D0DataType       = ck::bhalf_t;
+using DsDataType       = ck::Tuple<>;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using CLayout  = Row;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3R1<
+        ALayout,   BLayout,   DsLayout,  CLayout,
+        ADataType,   BDataType, DsDataType,  CDataType, AccDataType,  CShuffleDataType,
+        AElementOp, BElementOp, CDEElementOp, GemmDefault, 
+        256,   
+        128,  128,  64,
+        8,    4,
+        32,   32,
+        2,    2,
+        S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,
+        2,    8,    8,   0,
+        S<16, 16, 1>,    S<0, 2, 1>,    S<0, 2, 1>,
+        1,    8,    4,   0,
+        1,    1,    S<1, 32, 1, 8>,  8,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3, ReduceDataType>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_splitk_reduce_multi_d_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

example/35_splitK_gemm/gemm_xdl_splitk_reduce_multi_d_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using ReduceDataType   = float;
+using D0DataType       = ck::bhalf_t;
+using DsDataType       = ck::Tuple<D0DataType>;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using CLayout  = Row;
+using D0Layout = CLayout;
+using DsLayout = ck::Tuple<D0Layout>;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3R1<
+        ALayout,   BLayout,   DsLayout,  CLayout,
+        ADataType,   BDataType, DsDataType,  CDataType, AccDataType,  CShuffleDataType,
+        AElementOp, BElementOp, CDEElementOp, GemmDefault, 
+        256,
+        128,  128,  64,
+        8,    4,
+        32,   32,
+        2,    2,
+        S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,
+        2,    8,    8,   0,
+        S<16, 16, 1>,    S<0, 2, 1>,    S<0, 2, 1>,
+        1,    8,    4,   0,
+        1,    1,    S<1, 32, 1, 8>,  8,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3, ReduceDataType>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_splitk_reduce_multi_d_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

example/35_splitK_gemm/gemm_xdl_splitk_reduce_multi_d_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using CDataType        = ck::half_t;
+using ReduceDataType   = float;
+using D0DataType       = ck::half_t;
+using DsDataType       = ck::Tuple<D0DataType>;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using CLayout  = Row;
+using D0Layout = CLayout;
+using DsLayout = ck::Tuple<D0Layout>;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3R1<
+        ALayout,   BLayout,   DsLayout,  CLayout,
+        ADataType,   BDataType, DsDataType,  CDataType, AccDataType,  CShuffleDataType,
+	AElementOp, BElementOp, CDEElementOp, GemmDefault, 
+        256,
+        128,  128,  64,
+        8,    4,
+        32,   32,
+        2,    2,
+        S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,
+        2,    8,    8,   0,
+        S<16, 16, 1>,    S<0, 2, 1>,    S<0, 2, 1>,
+        1,    8,    4,   0,
+        1,    1,    S<1, 32, 1, 8>,  8,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v2, ReduceDataType>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_splitk_reduce_multi_d_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

example/35_splitK_gemm/run_gemm_splitk_reduce_multi_d_example.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename ProblemType>
+bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+    auto M        = problem_size.M;
+    auto N        = problem_size.N;
+    auto K        = problem_size.K;
+    auto StrideA  = problem_size.StrideA;
+    auto StrideB  = problem_size.StrideB;
+    auto StrideC  = problem_size.StrideC;
+    auto StrideD0 = problem_size.StrideC;
+    auto KBatch   = problem_size.KBatch;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    auto f_get_default_stride =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(stride == 0)
+            {
+                // give a chance if stride is zero, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return col;
+                }
+                else
+                {
+                    return row;
+                }
+            }
+            else
+                return stride;
+        };
+
+    StrideA  = f_get_default_stride(M, K, StrideA, ALayout{});
+    StrideB  = f_get_default_stride(K, N, StrideB, BLayout{});
+    StrideC  = f_get_default_stride(M, N, StrideC, CLayout{});
+    StrideD0 = f_get_default_stride(M, N, StrideD0, D0Layout{});
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+#if 0
+    printf("B matrix:\n");
+    for (int in = 0; in < N; in++)
+    {
+        for (int ik = 0; ik < K; ik++)
+        {
+            printf("%02x ", *(reinterpret_cast<uint8_t*>(&b_k_n(ik,in))));
+            if(ik%8==7) printf("|");
+        }
+        printf("\n");
+    }
+#endif
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "init method: " << config.init_method << std::endl;
+    std::cout << "KBatch: " << KBatch << std::endl;
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto gemm        = DeviceGemmV2Instance{};
+    auto invoker     = gemm.MakeInvoker();
+    float ave_time   = 0;
+    auto get_argment = [&]() {
+        if constexpr(DsDataType::Size() > 0)
+        {
+            return gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                     static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                     {d0_m_n_device_buf.GetDeviceBuffer()},
+                                     static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                     M,
+                                     N,
+                                     K,
+                                     StrideA,
+                                     StrideB,
+                                     {StrideD0},
+                                     StrideC,
+                                     KBatch,
+                                     a_element_op,
+                                     b_element_op,
+                                     c_element_op);
+        }
+        else
+        {
+            return gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                     static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                     {},
+                                     static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                     M,
+                                     N,
+                                     K,
+                                     StrideA,
+                                     StrideB,
+                                     {},
+                                     StrideC,
+                                     KBatch,
+                                     a_element_op,
+                                     b_element_op,
+                                     c_element_op);
+        }
+    };
+    auto argument = get_argment();
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return true;
+    }
+
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer(), StreamConfig{});
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 1});
+
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        if constexpr(DsDataType::Size() > 0)
+        {
+            c_m_n_host_result.ForEach(
+                [&](auto& self, auto idx) { c_element_op(self(idx), self(idx), d0_m_n(idx)); });
+        }
+
+        pass &= ck::utils::check_err(c_m_n_device_result,
+                                     c_m_n_host_result,
+                                     "Error: Incorrect results!",
+                                     get_rtol<CDataType>(),
+                                     get_atol<CDataType>());
+    }
+
+    if(config.time_kernel)
+    {
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+        std::size_t flop = 2_uz * M * N * K;
+        std::size_t num_btype =
+            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+    return pass;
+}
+
+bool run_gemm_splitk_example(int argc, char* argv[])
+{
+    ProblemSizeSplitK problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
+}

example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using A0DataType       = F8;
+using A1DataType       = F32;
+using B0DataType       = F8;
+using B1DataType       = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using EDataType        = F16;
+using ComputeDataType  = F8;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct Multiply
+{
+    __host__ __device__ constexpr void
+    operator()(ck::f8_t& a, const ck::f8_t& a0, const float& a1) const
+    {
+        a = ck::type_convert<ck::half_t>(ck::type_convert<float>(a0) * a1);
+    }
+};
+
+using AElementOp   = Multiply;
+using BElementOp   = Multiply;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultipleABD_Xdl_CShuffle<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ck::Tuple<A0DataType, A1DataType>,
+    ck::Tuple<B0DataType, B1DataType>,
+    AccDataType,
+    CShuffleDataType,
+    ck::Tuple<>,
+    EDataType,
+    AElementOp,
+    BElementOp,
+    CDEElementOp,
+    GemmSpec,
+    1,
+    256,
+    256,
+    128,
+    32,
+    8,
+    8,
+    32,
+    32,
+    4,
+    2,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    1,
+    8,
+    1,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    1,
+    1,
+    1,
+    S<1, 32, 1, 8>,
+    8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A0[M0, M1, K0, K1]
+    std::vector<ck::index_t> a0_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
+    // A1[M1, K1] -> A1[M0, M1, K0, K1]
+    std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 1, 0};
+    // B0[N0, N1, K0, K1]
+    std::vector<ck::index_t> b0_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b0_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
+    // B1[N0, N1, K0, K1]
+    std::vector<ck::index_t> b1_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b1_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<A0DataType> a0_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+    Tensor<A1DataType> a1_ms_ks(a1_ms_ks_lengths, a1_ms_ks_strides);
+    Tensor<B0DataType> b0_ns_ks(b0_ns_ks_lengths, b0_ns_ks_strides);
+    Tensor<B1DataType> b1_ns_ks(b1_ns_ks_lengths, b1_ns_ks_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a0_ms_ks: " << a0_ms_ks.mDesc << std::endl;
+    std::cout << "a1_ms_ks: " << a1_ms_ks.mDesc << std::endl;
+
+    std::cout << "b0_ns_ks: " << b0_ns_ks.mDesc << std::endl;
+    std::cout << "b1_ns_ks: " << b1_ns_ks.mDesc << std::endl;
+
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
+        b0_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_ns_ks.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    default:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
+        b0_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        b1_ns_ks.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(A1DataType) * a1_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_ms_ks.mData.data());
+    a1_device_buf.ToDevice(a1_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_ns_ks.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(
+        std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
+                                   a1_device_buf.GetDeviceBuffer()},
+        std::array<const void*, 2>{b0_device_buf.GetDeviceBuffer(),
+                                   b1_device_buf.GetDeviceBuffer()},
+        std::array<const void*, 0>{},
+        e_device_buf.GetDeviceBuffer(),
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_lengths, a1_ms_ks_lengths},
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_strides, a1_ms_ks_strides},
+        std::array<std::vector<ck::index_t>, 2>{b0_ns_ks_lengths, b1_ns_ks_lengths},
+        std::array<std::vector<ck::index_t>, 2>{b0_ns_ks_strides, b1_ns_ks_strides},
+        std::array<std::vector<ck::index_t>, 0>{},
+        std::array<std::vector<ck::index_t>, 0>{},
+        e_ms_ns_lengths,
+        e_ms_ns_strides,
+        a_element_op,
+        b_element_op,
+        PassThrough{});
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_contraction with the specified compilation parameters does "
+            "not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    if(time_kernel)
+    {
+        ck::index_t M =
+            ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+        ck::index_t N = ck::accumulate_n<ck::index_t>(
+            e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+        ck::index_t K = ck::accumulate_n<ck::index_t>(
+            a0_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+        std::size_t flop = std::size_t(2) * M * N * K;
+        std::size_t num_btype =
+            sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + +sizeof(EDataType) * M * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s" << std::endl;
+    }
+
+    if(do_verification)
+    {
+
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        Tensor<A0DataType> a_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+
+        for(size_t m0 = 0; m0 < a_ms_ks.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < a_ms_ks.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t k0 = 0; k0 < a_ms_ks.mDesc.GetLengths()[2]; ++k0)
+                {
+                    for(size_t k1 = 0; k1 < a_ms_ks.mDesc.GetLengths()[3]; ++k1)
+                    {
+                        a_element_op(a_ms_ks(m0, m1, k0, k1),
+                                     a0_ms_ks(m0, m1, k0, k1),
+                                     a1_ms_ks(m0, m1, k0, k1));
+                    }
+                }
+            }
+        }
+
+        Tensor<B0DataType> b_ns_ks(b0_ns_ks_lengths, b0_ns_ks_strides);
+
+        for(size_t n0 = 0; n0 < b_ns_ks.mDesc.GetLengths()[0]; ++n0)
+        {
+            for(size_t n1 = 0; n1 < b_ns_ks.mDesc.GetLengths()[1]; ++n1)
+            {
+                for(size_t k0 = 0; k0 < b_ns_ks.mDesc.GetLengths()[2]; ++k0)
+                {
+                    for(size_t k1 = 0; k1 < b_ns_ks.mDesc.GetLengths()[3]; ++k1)
+                    {
+                        b_element_op(b_ns_ks(n0, n1, k0, k1),
+                                     b0_ns_ks(n0, n1, k0, k1),
+                                     b1_ns_ks(n0, n1, k0, k1));
+                    }
+                }
+            }
+        }
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      A0DataType,
+                                                                      B0DataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      ComputeDataType,
+                                                                      PassThrough,
+                                                                      PassThrough>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
+        auto ref_argument = ref_op.MakeArgument(
+            a_ms_ks, b_ns_ks, c_ms_ns_host_result, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/62_convnd_activ/CMakeLists.txt

 add_subdirectory(binary)
 add_subdirectory(convinvscale)
 add_subdirectory(convscale)
+add_subdirectory(convscale_relu)
+add_subdirectory(convscale_add)
+add_subdirectory(convscale_reduce)
 add_subdirectory(multi_AB)
 add_subdirectory(unary)
 

example/62_convnd_activ/convscale/CMakeLists.txt

@@ -12,6 +12,9 @@ foreach(gpu IN LISTS GPU_TARGETS)
      add_example_executable(example_convnd_fwd_xdl_convscale_fp8_bf8 convnd_fwd_xdl_convscale_fp8_bf8.cpp)
      add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_fp8_bf8)
 
+     add_example_executable(example_convnd_fwd_xdl_convscale_bf8_fp8 convnd_fwd_xdl_convscale_bf8_fp8.cpp)
+     add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_bf8_fp8)
+
      set(target 1)
  endif()
 endforeach()

example/62_convnd_activ/convscale/convnd_fwd_xdl_convscale_bf8_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_convscale_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = ck::bf8_t;
+using WeiDataType      = ck::f8_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using DsDataType       = ck::Tuple<>;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::bf8_t;
+using BComputeDataType = ck::f8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ConvScale;
+
+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 DsLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        DsLayout,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // 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
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8,
+        AComputeDataType,
+        BComputeDataType>;
+
+#include "run_convnd_fwd_convscale_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }

example/62_convnd_activ/convscale_add/CMakeLists.txt

+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+     add_custom_target(example_convnd_activ_xdl_convscale_add)
+     add_example_executable(example_convnd_fwd_xdl_convscale_add_fp8 convnd_fwd_xdl_convscale_add_fp8.cpp)
+     add_example_dependencies(example_convnd_activ_xdl_convscale_add example_convnd_fwd_xdl_convscale_add_fp8 )
+
+     set(target 1)
+ endif()
+endforeach()

example/62_convnd_activ/convscale_add/convnd_fwd_convscale_add_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/utility/algorithm.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/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ConvScaleAdd = ck::tensor_operation::element_wise::ConvScaleAdd;
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetFlops(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths,
+         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths,
+         const std::size_t& ds_size)
+{
+    // G * N * C * <output spatial lengths product> * (2 * K * <filter spatial lengths product> +
+    // <number of scale factors>)
+    ck::index_t G = weights_lengths[0];
+    ck::index_t N = output_lengths[1];
+    ck::index_t K = weights_lengths[1];
+    ck::index_t C = weights_lengths[2];
+
+    return G * N * C *
+           std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           (static_cast<std::size_t>(2) * K *
+                std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
+                                std::end(weights_lengths),
+                                static_cast<std::size_t>(1),
+                                std::multiplies<>()) +
+            ds_size);
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          int init_method,
+                          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)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<DsDataType> bias(out_g_n_k_wos_desc);
+    Tensor<CShuffleDataType> c(out_g_n_k_wos_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;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-1, 1});
+        bias.GenerateTensorValue(GeneratorTensor_2<DsDataType>{-3, 3});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1.0, 1.0});
+        bias.GenerateTensorValue(GeneratorTensor_3<DsDataType>{-3.0, 3.0});
+        break;
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(DsDataType) * 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> d_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d_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 = [](const auto& x, auto& y) { ck::ranges::copy(x, 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(), d_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), d_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);
+
+    // random scale values
+    float scale_in  = float(std::rand()) / float(RAND_MAX);
+    float scale_wei = float(std::rand()) / float(RAND_MAX);
+    float scale_out = float(std::rand()) / float(RAND_MAX);
+
+    std::cout << std::endl;
+    std::cout << "scale_in: " << scale_in << std::endl;
+    std::cout << "scale_wei: " << scale_wei << std::endl;
+    std::cout << "scale_out: " << scale_out << std::endl;
+
+    // initialize out_element_op for each iteration
+    const auto out_element_op = OutElementOp{scale_in, scale_wei, scale_out};
+
+    // 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>{{d_g_n_k_wos_lengths}},
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_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 ds_size = 3 + 1; // 3 element-wise scale multipliers + 1 element-wise add
+    std::size_t flop    = GetFlops<NDimSpatial>(e_g_n_k_wos_lengths, b_g_k_c_xs_lengths, ds_size);
+    std::size_t num_btype =
+        conv_param.GetInputByte<InDataType>() + conv_param.GetWeightByte<WeiDataType>() +
+        sizeof(float) + sizeof(float) + sizeof(float) + conv_param.GetOutputByte<OutDataType>() +
+        conv_param.GetOutputByte<DsDataType>();
+
+    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;
+
+    if(do_verification)
+    {
+        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,
+                                                  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);
+
+        out_host.ForEach(
+            [&](auto&, auto idx) { out_element_op(out_host(idx), c(idx), bias(idx)); });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device,
+                                    out_host,
+                                    "Error: incorrect results!",
+                                    get_rtol<OutDataType>(),
+                                    get_atol<OutDataType>());
+    }
+
+    return true;
+}

example/62_convnd_activ/convscale_add/convnd_fwd_xdl_convscale_add_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/tuple.hpp"
+#include "convnd_fwd_convscale_add_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using DsDataType       = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ConvScaleAdd;
+
+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 DsLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<DsLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<DsDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // 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
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8,
+        AComputeDataType,
+        BComputeDataType>;
+
+#include "run_convnd_fwd_convscale_add_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }

example/62_convnd_activ/convscale_add/run_convnd_fwd_convscale_add_example.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_convnd_fwd_example(int argc, char* argv[])
+{
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    // instantiate in and wei element ops, will
+    // instantiate out_element_op below for every iteration
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+
+    const auto run =
+        [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto ds_layout, auto out_layout) {
+            constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
+
+            using InLayout  = decltype(in_layout);
+            using WeiLayout = decltype(wei_layout);
+            using DsLayout  = decltype(ds_layout);
+            using OutLayout = decltype(out_layout);
+
+            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_value,
+                                        InDataType,
+                                        WeiDataType,
+                                        CShuffleDataType,
+                                        DsDataType,
+                                        OutDataType,
+                                        InElementOp,
+                                        WeiElementOp,
+                                        OutElementOp,
+                                        DeviceGroupedConvNDFwdInstance<ndim_spatial_value,
+                                                                       InLayout,
+                                                                       WeiLayout,
+                                                                       DsLayout,
+                                                                       OutLayout>>(
+                do_verification,
+                init_method,
+                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);
+        };
+
+    namespace ctc = ck::tensor_layout::convolution;
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{}, ctc::GNWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{}, ctc::GNHWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{}, ctc::GNDHWK{});
+    }
+
+    return true;
+}

example/62_convnd_activ/convscale_reduce/CMakeLists.txt

+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+     add_custom_target(example_convnd_activ_xdl_convscale_reduce)
+     add_example_executable(example_convnd_fwd_xdl_convscale_relu_amax_fp8 convnd_fwd_xdl_convscale_relu_amax_fp8.cpp)
+     add_example_dependencies(example_convnd_activ_xdl_convscale_reduce example_convnd_fwd_xdl_convscale_relu_amax_fp8 )
+
+     set(target 1)
+ endif()
+endforeach()

example/62_convnd_activ/convscale_reduce/convnd_fwd_convscale_reduce_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/utility/algorithm.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/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ew = ck::tensor_operation::element_wise;
+
+using PassThrough   = ew::PassThrough;
+using ConvScaleRelu = ew::UnaryCombinedOp<ew::Scale, ew::Scale, ew::Relu>;
+using ConvScale     = ew::UnaryCombinedOp<ew::Scale, ew::Scale, PassThrough>;
+
+using UnaryScaleConvert = ew::Scale;
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename ConvOutDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename ConvElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          int init_method,
+                          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)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<ConvOutDataType> host_conv(out_g_n_k_wos_desc);
+    Tensor<ConvOutDataType> device_conv(out_g_n_k_wos_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;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    case 11: // used for debugging
+        in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
+        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem conv_device_buf(conv_param.GetOutputByte<ConvOutDataType>());
+    DeviceMem out_device_buf(conv_param.GetOutputByte<OutDataType>());
+
+    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 = [](const auto& x, auto& y) { ck::ranges::copy(x, 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);
+
+    // random scale values
+    float scale_in  = float(std::rand()) / float(RAND_MAX);
+    float scale_wei = float(std::rand()) / float(RAND_MAX);
+    float scale_out = float(std::rand()) / float(RAND_MAX);
+
+    std::cout << std::endl;
+    std::cout << "scale_in: " << scale_in << std::endl;
+    std::cout << "scale_wei: " << scale_wei << std::endl;
+    std::cout << "scale_out: " << scale_out << std::endl;
+
+    // convolution elementwise operation
+    auto conv_element_op = ConvElementOp{ew::Scale{scale_in}, ew::Scale{scale_wei}, {}};
+    auto scale_convert   = UnaryScaleConvert{scale_out}; // elementwise scale and type cast
+
+    // do Conv
+    auto conv         = DeviceConvNDFwdInstance{};
+    auto conv_invoker = conv.MakeInvoker();
+    auto conv_argument =
+        conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                          wei_device_buf.GetDeviceBuffer(),
+                          std::array<const void*, 0>{},
+                          conv_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,
+                          conv_element_op);
+
+    if(!conv.IsSupportedArgument(conv_argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    std::string kernels = conv.GetTypeString();
+
+    float avg_time = conv_invoker.Run(conv_argument, StreamConfig{nullptr, time_kernel});
+
+    using DeviceElementwiseScale = ck::tensor_operation::device::DeviceElementwiseImpl<
+        ck::Tuple<ConvOutDataType>, // InDataTypeTuple
+        ck::Tuple<OutDataType>,     // OutDataTypeTuple
+        UnaryScaleConvert,          // UnaryScaleConvert
+        NDimSpatial + 3,            // NumDim
+        256,                        // BlockSize
+        128,                        // M0PerBlock
+        128,                        // M1PerBlock
+        8,                          // M0PerThread
+        8,                          // M1PerThread
+        ck::Sequence<1, 0>,         // ThreadClusterArrangeOrder
+        ck::Sequence<8>,            // InScalarPerVectorSeq
+        ck::Sequence<8>>;           // OutScalarPerVectorSeq
+
+    auto device_ew_scale = DeviceElementwiseScale{};
+    auto scale_invoker   = device_ew_scale.MakeInvoker();
+    auto scale_argument  = device_ew_scale.MakeArgument(e_g_n_k_wos_lengths,
+                                                       {e_g_n_k_wos_strides},
+                                                       {e_g_n_k_wos_strides},
+                                                       {conv_device_buf.GetDeviceBuffer()},
+                                                       {out_device_buf.GetDeviceBuffer()},
+                                                       scale_convert);
+
+    if(!device_ew_scale.IsSupportedArgument(scale_argument))
+    {
+        throw std::runtime_error(
+            "wrong! DeviceElementwiseScale with the specified compilation parameters does "
+            "not support this problem");
+    }
+
+    kernels += std::string("\n\t\t ") + device_ew_scale.GetTypeString();
+
+    avg_time += scale_invoker.Run(scale_argument, StreamConfig{nullptr, time_kernel});
+
+    constexpr auto ReduceOpId = ck::ReduceTensorOp::AMAX;
+    using ReduceOperation     = typename ck::reduce_binary_operator<ReduceOpId>::opType;
+    using InElementwiseOperation =
+        typename ck::reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+    using AccElementwiseOperation =
+        typename ck::reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+    using DeviceReduceInstance =
+        ck::tensor_operation::device::DeviceReduceMultiBlock<ConvOutDataType,
+                                                             ConvOutDataType,
+                                                             ConvOutDataType,
+                                                             NDimSpatial + 3,
+                                                             NDimSpatial + 3,
+                                                             ReduceOperation,
+                                                             InElementwiseOperation,
+                                                             AccElementwiseOperation,
+                                                             ck::InMemoryDataOperationEnum::Set,
+                                                             true,  // PropagateNan
+                                                             false, // OutputIndex
+                                                             false, // HaveIndexInputIfOutputIndex
+                                                             256,   // BlockSize
+                                                             4,     // MThreadClusterSize
+                                                             64,    // KThreadClusterSize
+                                                             1,     // MThreadSliceSize
+                                                             1,     // KThreadSliceSize
+                                                             1,     // InSrcVectorDim
+                                                             1,     // InSrceVectorSize
+                                                             1>;    // OutDstVectorSize
+
+    std::vector<size_t> outLengths = {1};
+    Tensor<ConvOutDataType> amax_host(outLengths);
+    Tensor<ConvOutDataType> amax_from_device(outLengths);
+    auto amax_host_strides = amax_host.mDesc.GetStrides();
+
+    std::array<int, NDimSpatial + 3> reduce_dims;
+    std::iota(reduce_dims.begin(), reduce_dims.end(), 0); // 0,..., NDimSpatial+3-1
+
+    std::array<ck::index_t, 1> reduce_out_lengths{1};
+    std::array<ck::index_t, 1> reduce_out_strides{static_cast<ck::index_t>(amax_host_strides[0])};
+
+    DeviceMem amax_device(sizeof(ConvOutDataType) * amax_host.mDesc.GetElementSpaceSize());
+    DeviceMem index_device;
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(host_conv.mDesc.GetElementSize()));
+
+    // Hack convolution output strides for reduction as kernel expects stride 1 for the last
+    // dimension. It only works because the reduction is done on the whole tensor and result is
+    // independent of the order of elements.
+    std::array<ck::index_t, NDimSpatial + 3> reduction_strides{};
+    copy(HostTensorDescriptor(e_g_n_k_wos_lengths).GetStrides(), reduction_strides);
+
+    auto device_reduce   = DeviceReduceInstance{};
+    auto reduce_invoker  = device_reduce.MakeInvokerPointer();
+    auto reduce_argument = device_reduce.MakeArgumentPointer(e_g_n_k_wos_lengths,
+                                                             reduction_strides,
+                                                             reduce_out_lengths,
+                                                             reduce_out_strides,
+                                                             reduce_dims,
+                                                             1.0,
+                                                             0.0,
+                                                             conv_device_buf.GetDeviceBuffer(),
+                                                             nullptr,
+                                                             amax_device.GetDeviceBuffer(),
+                                                             nullptr,
+                                                             in_elementwise_op,
+                                                             acc_elementwise_op);
+
+    if(!device_reduce.IsSupportedArgument(reduce_argument.get()))
+    {
+        throw std::runtime_error(
+            "wrong! DeviceReduceInstance with the specified compilation parameters does "
+            "not support this runtime parameters!");
+    };
+
+    kernels += std::string("\n\t\t ") + device_reduce.GetTypeString();
+
+    float reduce_time =
+        reduce_invoker->Run(reduce_argument.get(), StreamConfig{nullptr, time_kernel});
+
+    if(time_kernel)
+        std::cout << "\nReduce time: " << reduce_time << " ms" << std::endl;
+
+    avg_time += reduce_time;
+
+    std::size_t flop    = conv_param.GetFlops();      // convolution FLOPs
+    auto conv_out_elems = host_conv.GetElementSize(); // number of elements in conv result tensor
+
+    // 3 element-wise scale multipliers + 1 AMAX
+    std::size_t elementwise_ops = 3 + 1;
+    if constexpr(ck::is_same_v<ConvElementOp, ConvScaleRelu>)
+    {
+        elementwise_ops += 1; // +1 element-wise relu
+    }
+
+    flop += elementwise_ops * conv_out_elems;
+
+    // convolution + elementwise scaling (in + wei + output byte count)
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, ConvOutDataType>();
+    num_btype += sizeof(float) + sizeof(float); //  + 2 scales
+
+    // elementwise scaling + F8 conversion
+    num_btype += conv_param.GetOutputByte<ConvOutDataType>() + sizeof(float) +
+                 conv_param.GetOutputByte<OutDataType>();
+
+    // AMAX
+    num_btype += conv_param.GetOutputByte<ConvOutDataType>() + sizeof(float);
+
+    if(time_kernel)
+    {
+        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, " << std::endl;
+    }
+
+    std::cout << "\nKernels: " << kernels << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     ConvOutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     ConvElementOp>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  host_conv,
+                                                  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,
+                                                  conv_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        conv_device_buf.FromDevice(device_conv.mData.data());
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        out_host.ForEach([&](auto&, auto idx) { scale_convert(out_host(idx), host_conv(idx)); });
+
+        std::cout << "\nComparing output to reference: " << std::endl;
+        auto tight_tol_check = ck::utils::check_err(out_device, out_host, "Error: ");
+        if(!tight_tol_check)
+        {
+            std::cout << "\n\tRecompare applying tolerances...\n";
+            std::cout << "\t\trtol = " << get_rtol<OutDataType>() << std::endl;
+            std::cout << "\t\tatol = " << get_atol<OutDataType>() << std::endl;
+            auto loose_tol_check = ck::utils::check_err(out_device,
+                                                        out_host,
+                                                        "Error: incorrect convolution results!",
+                                                        get_rtol<OutDataType>(),
+                                                        get_atol<OutDataType>());
+            if(!loose_tol_check)
+            {
+                return false;
+            }
+        }
+        std::cout << "Success!" << std::endl;
+
+        /// Verify AMAX
+
+        using RefReduceInstance =
+            ck::tensor_operation::host::ReferenceReduce<ConvOutDataType,
+                                                        ConvOutDataType,
+                                                        ConvOutDataType,
+                                                        NDimSpatial + 3,
+                                                        NDimSpatial + 3,
+                                                        ReduceOperation,
+                                                        InElementwiseOperation,
+                                                        AccElementwiseOperation,
+                                                        true,
+                                                        false>;
+
+        auto ref_reduce          = RefReduceInstance{};
+        auto ref_reduce_invoker  = ref_reduce.MakeInvokerPointer();
+        auto ref_reduce_argument = ref_reduce.MakeArgumentPointer(e_g_n_k_wos_lengths,
+                                                                  e_g_n_k_wos_strides,
+                                                                  reduce_out_lengths,
+                                                                  reduce_out_strides,
+                                                                  reduce_dims,
+                                                                  1.0,
+                                                                  0.0,
+                                                                  host_conv.mData.data(),
+                                                                  nullptr,
+                                                                  amax_host.mData.data(),
+                                                                  nullptr,
+                                                                  in_elementwise_op,
+                                                                  acc_elementwise_op);
+
+        if(!ref_reduce.IsSupportedArgument(ref_reduce_argument.get()))
+        {
+            throw std::runtime_error(
+                "wrong! RefReduceInstance with the specified compilation parameters does "
+                "not support this runtime parameters!");
+        };
+
+        ref_reduce_invoker->Run(ref_reduce_argument.get());
+
+        amax_device.FromDevice(amax_from_device.mData.data());
+
+        std::cout << "\namax: " << amax_from_device.mData[0] << std::endl;
+        std::cout << "amax_ref: " << amax_host.mData[0] << std::endl;
+
+        return ck::utils::check_err(amax_from_device, amax_host, "Error: incorrect AMAX results!");
+    }
+
+    return true;
+}