Merge remote-tracking branch 'origin/develop' into migraphx-update

client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/transpose_3d.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    const int N = 16;
+    const int C = 8;
+    const int D = 8;
+    const int H = 8;
+    const int W = 8;
+
+    std::vector<std::size_t> ncdhw = {N, C, D, H, W};
+    std::vector<std::size_t> nchwd = {N, C, H, W, D};
+    auto size                      = N * C * D * H * W;
+
+    std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
+    std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
+    std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, C, H, W, D
+
+    SimpleDeviceMem a_dev_buf(sizeof(ADataType) * size);
+    SimpleDeviceMem b_dev_buf(sizeof(BDataType) * size);
+
+    std::array<const void*, 1> input = {a_dev_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_dev_buf.GetDeviceBuffer()};
+
+    using DeviceElementwisePermuteInstance = ck::tensor_operation::device::
+        DeviceElementwise<ck::Tuple<ADataType>, ck::Tuple<BDataType>, PassThrough, 5>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceElementwisePermuteInstance>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
+                sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/24_grouped_conv_activation/CMakeLists.txt

+if(GPU_TARGETS MATCHES "gfx9")
+# Fwd scaleadd scaleadd relu
+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32
+               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16
+               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16
+               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8
+               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_int8.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8 PRIVATE composable_kernel::device_conv_operations)
+# Fwd scaleadd AB
+add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp32
+               grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp32.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp32 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp16
+               grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp16 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_ab_bf16
+               grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_bf16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_bf16 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_ab_int8
+               grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_int8.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_int8 PRIVATE composable_kernel::device_conv_operations)
+# Fwd bilinear
+add_executable(client_grouped_convnd_fwd_bilinear_residual_fp16
+               grouped_convnd_fwd_bilinear/grouped_conv_fwd_bilinear_residual_fp16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_bilinear_residual_fp16 PRIVATE composable_kernel::device_conv_operations)
+# Fwd convinvscale
+add_executable(client_conv3d_fwd_convinvscale_fp8
+               grouped_convnd_fwd_convinvscale/conv3d_fwd_convinvscale_fp8.cpp)
+target_link_libraries(client_conv3d_fwd_convinvscale_fp8 PRIVATE composable_kernel::device_conv_operations)
+# Fwd convscale + Bias
+add_executable(client_conv3d_fwd_convscale_add_fp8
+               grouped_convnd_fwd_convscale_add/conv3d_fwd_convscale_add_fp8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_add_fp8 PRIVATE composable_kernel::device_conv_operations)
+# Fwd convscale + ReLU
+add_executable(client_conv3d_fwd_convscale_relu_fp8
+               grouped_convnd_fwd_convscale_relu/conv3d_fwd_convscale_relu_fp8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_relu_fp8 PRIVATE composable_kernel::device_conv_operations)
+# Fwd convscale + ReLU + AMAX
+add_executable(client_conv3d_fwd_convscale_relu_amax_fp8
+               grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_relu_amax_fp8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_relu_amax_fp8
+                      PRIVATE composable_kernel::device_conv_operations
+                              composable_kernel::device_other_operations
+                              composable_kernel::device_reduction_operations
+                              utility)
+# Fwd convscale + AMAX
+add_executable(client_conv3d_fwd_convscale_amax_fp8
+               grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_amax_fp8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_amax_fp8
+                      PRIVATE composable_kernel::device_conv_operations
+                              composable_kernel::device_other_operations
+                              composable_kernel::device_reduction_operations
+                              utility)
+# Fwd convscale
+add_executable(client_conv3d_fwd_convscale_fp8
+               grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_fp8 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_conv3d_fwd_convscale_bf8
+               grouped_convnd_fwd_convscale/conv3d_fwd_convscale_bf8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_bf8 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_conv3d_fwd_convscale_fp8_bf8
+               grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8_bf8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_fp8_bf8 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_conv3d_fwd_convscale_bf8_fp8
+               grouped_convnd_fwd_convscale/conv3d_fwd_convscale_bf8_fp8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
+# Bwd data bilinear
+add_executable(client_grouped_convnd_bwd_data_bilinear_residual_fp16
+               grouped_convnd_bwd_data_bilinear/grouped_conv_bwd_data_bilinear_residual_fp16.cpp)
+target_link_libraries(client_grouped_convnd_bwd_data_bilinear_residual_fp16 PRIVATE composable_kernel::device_conv_operations)
+# Bwd weight bilinear
+add_executable(client_grouped_convnd_bwd_weight_bilinear_residual_fp16
+               grouped_convnd_bwd_weight_bilinear/grouped_conv_bwd_weight_bilinear_residual_fp16.cpp)
+target_link_libraries(client_grouped_convnd_bwd_weight_bilinear_residual_fp16 PRIVATE composable_kernel::device_conv_operations)
+# Fwd scale
+add_executable(client_grouped_convnd_fwd_scale_fp16
+               grouped_convnd_fwd_scale/grouped_conv_fwd_scale_fp16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scale_fp16 PRIVATE composable_kernel::device_conv_operations)
+# Bwd data scale
+add_executable(client_grouped_convnd_bwd_data_scale_fp16
+               grouped_convnd_bwd_data_scale/grouped_conv_bwd_data_scale_fp16.cpp)
+target_link_libraries(client_grouped_convnd_bwd_data_scale_fp16 PRIVATE composable_kernel::device_conv_operations)
+# Bwd weight scale
+add_executable(client_grouped_convnd_bwd_weight_scale_fp16
+               grouped_convnd_bwd_weight_scale/grouped_conv_bwd_weight_scale_fp16.cpp)
+target_link_libraries(client_grouped_convnd_bwd_weight_scale_fp16 PRIVATE composable_kernel::device_conv_operations)
+endif()

client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_bilinear/grouped_conv_bwd_data_bilinear_residual_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_bilinear.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+// Use std tuple instead of ck tuple to avoid clang
+// implicit instantiation of undefined template error.
+using DDataTypes = std::tuple<ck::half_t>;
+
+using InLayout    = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout   = ck::tensor_layout::convolution::NDHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 64; // batch size
+static constexpr ck::index_t K             = 64; // output channel
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Z             = 3;  // filter D
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Di            = 14; // input D
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+static constexpr ck::index_t Do            = 14; // output D
+static constexpr ck::index_t Ho            = 14; // output H
+static constexpr ck::index_t Wo            = 14; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int execute_conv_bwd_data_bilinear()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                                                        OutLayout,
+                                                                        WeiLayout,
+                                                                        ck::Tuple<InLayout>,
+                                                                        InLayout,
+                                                                        OutDataType,
+                                                                        WeiDataType,
+                                                                        ck::Tuple<InDataType>,
+                                                                        InDataType,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        Bilinear>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {in.GetDeviceBuffer()},
+                                                        in.GetDeviceBuffer(),
+                                                        out_lengths,
+                                                        out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {in_lengths},
+                                                        {in_strides},
+                                                        in_lengths,
+                                                        in_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        Bilinear{2.f, 2.f});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X +
+                               3 * G * N * Di * Hi * Wi * C;
+            std::size_t num_bytes = 2 * sizeof(InDataType) * G * N * Di * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Z * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {in.GetDeviceBuffer()},
+                                                        in.GetDeviceBuffer(),
+                                                        out_lengths,
+                                                        out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {in_lengths},
+                                                        {in_strides},
+                                                        in_lengths,
+                                                        in_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        Bilinear{2.f, 2.f});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return 0;
+}
+
+int main() { return execute_conv_bwd_data_bilinear(); }

client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_scale/grouped_conv_bwd_data_scale_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_scale.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+// Use std tuple instead of ck tuple to avoid clang
+// implicit instantiation of undefined template error.
+using DDataTypes = std::tuple<ck::half_t>;
+
+using InLayout    = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout   = ck::tensor_layout::convolution::NDHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 64; // batch size
+static constexpr ck::index_t K             = 64; // output channel
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Z             = 3;  // filter D
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Di            = 14; // input D
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+static constexpr ck::index_t Do            = 14; // output D
+static constexpr ck::index_t Ho            = 14; // output H
+static constexpr ck::index_t Wo            = 14; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int execute_conv_bwd_data_scale()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                                                                     OutLayout,
+                                                                                     WeiLayout,
+                                                                                     ck::Tuple<>,
+                                                                                     InLayout,
+                                                                                     OutDataType,
+                                                                                     WeiDataType,
+                                                                                     ck::Tuple<>,
+                                                                                     InDataType,
+                                                                                     PassThrough,
+                                                                                     PassThrough,
+                                                                                     Scale>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        in.GetDeviceBuffer(),
+                                                        out_lengths,
+                                                        out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        in_lengths,
+                                                        in_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        Scale{2.f});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X +
+                               3 * G * N * Di * Hi * Wi * C;
+            std::size_t num_bytes = 2 * sizeof(InDataType) * G * N * Di * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Z * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        in.GetDeviceBuffer(),
+                                                        out_lengths,
+                                                        out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        in_lengths,
+                                                        in_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        Scale{2.f});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return 0;
+}
+
+int main() { return execute_conv_bwd_data_scale(); }

client_example/24_grouped_conv_activation/grouped_convnd_bwd_weight_bilinear/grouped_conv_bwd_weight_bilinear_residual_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_bilinear.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout   = ck::tensor_layout::convolution::NDHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 32; // batch size
+static constexpr ck::index_t K             = 32; // output channel
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Z             = 3;  // filter D
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Di            = 14; // input D
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+static constexpr ck::index_t Do            = 14; // output D
+static constexpr ck::index_t Ho            = 14; // output H
+static constexpr ck::index_t Wo            = 14; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int execute_conv_bwd_weight_bilinear()
+{
+    constexpr ck::index_t split_k = 2;
+
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvBwdWeightMultipleD<NumDimSpatial,
+                                                                          InLayout,
+                                                                          WeiLayout,
+                                                                          OutLayout,
+                                                                          ck::Tuple<WeiLayout>,
+                                                                          InDataType,
+                                                                          WeiDataType,
+                                                                          OutDataType,
+                                                                          ck::Tuple<WeiDataType>,
+                                                                          PassThrough,
+                                                                          Bilinear,
+                                                                          PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out.GetDeviceBuffer()),
+                                        {wei.GetDeviceBuffer()},
+                                        in_lengths,
+                                        in_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        out_lengths,
+                                        out_strides,
+                                        {wei_lengths},
+                                        {wei_strides},
+                                        filter_strides,
+                                        filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        PassThrough{},
+                                        Bilinear{2.f, 2.f},
+                                        PassThrough{},
+                                        split_k);
+
+        SimpleDeviceMem workspace_buf(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_buf.GetDeviceBuffer());
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop =
+                std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X + 3 * G * K * Z * Y * X * C;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Di * Hi * Wi * C +
+                                    2 * sizeof(WeiDataType) * G * K * Z * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out.GetDeviceBuffer()),
+                                        {wei.GetDeviceBuffer()},
+                                        in_lengths,
+                                        in_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        out_lengths,
+                                        out_strides,
+                                        {wei_lengths},
+                                        {wei_strides},
+                                        filter_strides,
+                                        filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        PassThrough{},
+                                        Bilinear{2.f, 2.f},
+                                        PassThrough{},
+                                        split_k);
+
+        SimpleDeviceMem workspace_buf(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_buf.GetDeviceBuffer());
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return 0;
+}
+
+int main() { return execute_conv_bwd_weight_bilinear(); }

client_example/24_grouped_conv_activation/grouped_convnd_bwd_weight_scale/grouped_conv_bwd_weight_scale_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_scale.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout   = ck::tensor_layout::convolution::NDHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 32; // batch size
+static constexpr ck::index_t K             = 32; // output channel
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Z             = 3;  // filter D
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Di            = 14; // input D
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+static constexpr ck::index_t Do            = 14; // output D
+static constexpr ck::index_t Ho            = 14; // output H
+static constexpr ck::index_t Wo            = 14; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int execute_conv_bwd_weight_scale()
+{
+    constexpr ck::index_t split_k = 2;
+
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvBwdWeightMultipleD<NumDimSpatial,
+                                                                          InLayout,
+                                                                          WeiLayout,
+                                                                          OutLayout,
+                                                                          ck::Tuple<>,
+                                                                          InDataType,
+                                                                          WeiDataType,
+                                                                          OutDataType,
+                                                                          ck::Tuple<>,
+                                                                          PassThrough,
+                                                                          Scale,
+                                                                          PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out.GetDeviceBuffer()),
+                                        {},
+                                        in_lengths,
+                                        in_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        out_lengths,
+                                        out_strides,
+                                        {},
+                                        {},
+                                        filter_strides,
+                                        filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        PassThrough{},
+                                        Scale{2.f},
+                                        PassThrough{},
+                                        split_k);
+
+        SimpleDeviceMem workspace_buf(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_buf.GetDeviceBuffer());
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop =
+                std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X + G * K * Z * Y * X * C;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Di * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Z * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out.GetDeviceBuffer()),
+                                        {},
+                                        in_lengths,
+                                        in_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        out_lengths,
+                                        out_strides,
+                                        {},
+                                        {},
+                                        filter_strides,
+                                        filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        PassThrough{},
+                                        Scale{2.f},
+                                        PassThrough{},
+                                        split_k);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+        SimpleDeviceMem workspace_buf(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_buf.GetDeviceBuffer());
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return 0;
+}
+
+int main() { return execute_conv_bwd_weight_scale(); }

client_example/24_grouped_conv_activation/grouped_convnd_fwd_bilinear/grouped_conv_fwd_bilinear_residual_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bilinear.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+// Use std tuple instead of ck tuple to avoid clang
+// implicit instantiation of undefined template error.
+using DDataTypes = std::tuple<ck::half_t>;
+
+using InLayout    = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout   = ck::tensor_layout::convolution::NDHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 64; // batch size
+static constexpr ck::index_t K             = 64; // output channel
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Z             = 3;  // filter D
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Di            = 14; // input D
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+static constexpr ck::index_t Do            = 14; // output D
+static constexpr ck::index_t Ho            = 14; // output H
+static constexpr ck::index_t Wo            = 14; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int execute_conv_fwd_bilinear()
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space.
+    // However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
+    // Hence, we need to adjust the order of strides.
+    std::array<ck::index_t, 6> in_lengths{G, N, C, Di, Hi, Wi};
+    std::array<ck::index_t, 6> in_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+    std::array<ck::index_t, 6> wei_lengths{G, K, C, Z, Y, X};
+    std::array<ck::index_t, 6> wei_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+    std::array<ck::index_t, 6> out_lengths{G, N, K, Do, Ho, Wo};
+    std::array<ck::index_t, 6> out_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+    // Logical broadcast bias (we have to pass bias lengths in the same format as output - GNKDHW)
+    std::array<ck::index_t, 6> bias_lengths{G, 1, K, 1, 1, 1};
+    std::array<ck::index_t, 6> bias_strides{K, 0, 1, 0, 0, 0};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Di * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Do * Ho * Wo * G * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                      InLayout,
+                                                                      WeiLayout,
+                                                                      ck::Tuple<OutLayout>,
+                                                                      OutLayout,
+                                                                      InDataType,
+                                                                      WeiDataType,
+                                                                      ck::Tuple<OutDataType>,
+                                                                      OutDataType,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      Bilinear>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {out.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {out_lengths},
+                                                        {out_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        Bilinear{2.f, 2.f});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop =
+                std::size_t(2) * G * N * K * C * Ho * Wo * Y * X + 3 * N * Ho * Wo * G * K;
+            std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * 2 * N * Ho * Wo * G * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {out.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {out_lengths},
+                                                        {out_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        Bilinear{2.f, 2.f});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return 0;
+}
+
+int main() { return execute_conv_fwd_bilinear(); }

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convinvscale/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convinvscale.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ConvInvscale = ck::tensor_operation::element_wise::ConvInvscale;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+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 <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
+{
+    // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * std::accumulate(std::begin(input_lengths),
+                                                std::end(input_lengths),
+                                                static_cast<std::size_t>(1),
+                                                std::multiplies<>());
+}
+
+template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * std::accumulate(std::begin(weights_lengths),
+                                                 std::end(weights_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<>());
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
+{
+    // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * std::accumulate(std::begin(output_lengths),
+                                                 std::end(output_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<std::size_t>());
+}
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumNonSpatialDim = 3,
+          typename AComputeType        = InDataType,
+          typename BComputeType        = AComputeType>
+bool run_grouped_conv_fwd_convinvscale(
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
+{
+    std::size_t in_mem_size  = GetInputByte<InDataType, NumDimSpatial>(in_lengths);
+    std::size_t wei_mem_size = GetWeightByte<WeiDataType, NumDimSpatial>(wei_lengths);
+    std::size_t out_mem_size = GetOutputByte<OutDataType, NumDimSpatial>(out_lengths);
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+
+    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::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_strides;
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_strides;
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_strides;
+    in_strides.fill(0);
+    wei_strides.fill(0);
+    out_strides.fill(0);
+    in_strides.back()  = 1;
+    wei_strides.back() = 1;
+    out_strides.back() = 1;
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    // transpose NDHWGC/KZYXGC/NDHWGK to GNDHWC/GKZYXC/GNDHWK to GNCDHW/GKCZYX/GNKDHW
+    std::rotate(std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 2), rend(in_lengths));
+    std::rotate(rbegin(in_lengths),
+                std::next(rbegin(in_lengths)),
+                std::next(rbegin(in_lengths), NumDimSpatial + 1));
+
+    std::rotate(std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 2), rend(in_strides));
+    std::rotate(rbegin(in_strides),
+                std::next(rbegin(in_strides)),
+                std::next(rbegin(in_strides), NumDimSpatial + 1));
+
+    std::rotate(rbegin(wei_lengths),
+                std::next(rbegin(wei_lengths)),
+                std::next(rbegin(wei_lengths), NumDimSpatial + 1));
+
+    std::rotate(rbegin(wei_strides),
+                std::next(rbegin(wei_strides)),
+                std::next(rbegin(wei_strides), NumDimSpatial + 1));
+
+    std::rotate(
+        std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 2), rend(out_lengths));
+    std::rotate(rbegin(out_lengths),
+                std::next(rbegin(out_lengths)),
+                std::next(rbegin(out_lengths), NumDimSpatial + 1));
+
+    std::rotate(
+        std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 2), rend(out_strides));
+    std::rotate(rbegin(out_strides),
+                std::next(rbegin(out_strides)),
+                std::next(rbegin(out_strides), NumDimSpatial + 1));
+
+    std::array<ck::index_t, NumDimSpatial> conv_filter_strides;
+    std::array<ck::index_t, NumDimSpatial> conv_filter_dilations;
+    std::array<ck::index_t, NumDimSpatial> input_left_pads;
+    std::array<ck::index_t, NumDimSpatial> input_right_pads;
+    conv_filter_strides.fill(1);
+    conv_filter_dilations.fill(1);
+    input_left_pads.fill(1);
+    input_right_pads.fill(1);
+
+    std::size_t ds_size = 3; // 3 element-wise scale multipliers
+    std::size_t flop    = GetFlops<NumDimSpatial>(out_lengths, wei_lengths, ds_size);
+    std::size_t num_bytes =
+        in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) + sizeof(float) + out_mem_size;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   ConvInvscale,
+                                                                                   AComputeType,
+                                                                                   BComputeType>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvInvscale{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvInvscale{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return true;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convinvscale/conv3d_fwd_convinvscale_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using CShuffleDataType = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convinvscale<NumDimSpatial,
+                                             InDataType,
+                                             WeiDataType,
+                                             OutDataType,
+                                             InLayout,
+                                             WeiLayout,
+                                             OutLayout,
+                                             3,
+                                             AComputeDataType,
+                                             BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ConvScale   = ck::tensor_operation::element_wise::ConvScale;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+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 <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
+{
+    // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * std::accumulate(std::begin(input_lengths),
+                                                std::end(input_lengths),
+                                                static_cast<std::size_t>(1),
+                                                std::multiplies<>());
+}
+
+template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * std::accumulate(std::begin(weights_lengths),
+                                                 std::end(weights_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<>());
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
+{
+    // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * std::accumulate(std::begin(output_lengths),
+                                                 std::end(output_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<std::size_t>());
+}
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumNonSpatialDim = 3,
+          typename AComputeType        = InDataType,
+          typename BComputeType        = AComputeType>
+bool run_grouped_conv_fwd_convscale(
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
+{
+    std::size_t in_mem_size  = GetInputByte<InDataType, NumDimSpatial>(in_lengths);
+    std::size_t wei_mem_size = GetWeightByte<WeiDataType, NumDimSpatial>(wei_lengths);
+    std::size_t out_mem_size = GetOutputByte<OutDataType, NumDimSpatial>(out_lengths);
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+
+    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::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_strides;
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_strides;
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_strides;
+    in_strides.fill(0);
+    wei_strides.fill(0);
+    out_strides.fill(0);
+    in_strides.back()  = 1;
+    wei_strides.back() = 1;
+    out_strides.back() = 1;
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    // transpose NDHWGC/KZYXGC/NDHWGK to GNDHWC/GKZYXC/GNDHWK to GNCDHW/GKCZYX/GNKDHW
+    std::rotate(std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 2), rend(in_lengths));
+    std::rotate(rbegin(in_lengths),
+                std::next(rbegin(in_lengths)),
+                std::next(rbegin(in_lengths), NumDimSpatial + 1));
+
+    std::rotate(std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 2), rend(in_strides));
+    std::rotate(rbegin(in_strides),
+                std::next(rbegin(in_strides)),
+                std::next(rbegin(in_strides), NumDimSpatial + 1));
+
+    std::rotate(rbegin(wei_lengths),
+                std::next(rbegin(wei_lengths)),
+                std::next(rbegin(wei_lengths), NumDimSpatial + 1));
+
+    std::rotate(rbegin(wei_strides),
+                std::next(rbegin(wei_strides)),
+                std::next(rbegin(wei_strides), NumDimSpatial + 1));
+
+    std::rotate(
+        std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 2), rend(out_lengths));
+    std::rotate(rbegin(out_lengths),
+                std::next(rbegin(out_lengths)),
+                std::next(rbegin(out_lengths), NumDimSpatial + 1));
+
+    std::rotate(
+        std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 2), rend(out_strides));
+    std::rotate(rbegin(out_strides),
+                std::next(rbegin(out_strides)),
+                std::next(rbegin(out_strides), NumDimSpatial + 1));
+
+    std::array<ck::index_t, NumDimSpatial> conv_filter_strides;
+    std::array<ck::index_t, NumDimSpatial> conv_filter_dilations;
+    std::array<ck::index_t, NumDimSpatial> input_left_pads;
+    std::array<ck::index_t, NumDimSpatial> input_right_pads;
+    conv_filter_strides.fill(1);
+    conv_filter_dilations.fill(1);
+    input_left_pads.fill(1);
+    input_right_pads.fill(1);
+
+    std::size_t ds_size = 3; // 3 element-wise scale multipliers
+    std::size_t flop    = GetFlops<NumDimSpatial>(out_lengths, wei_lengths, ds_size);
+    std::size_t num_bytes =
+        in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) + sizeof(float) + out_mem_size;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   ConvScale,
+                                                                                   AComputeType,
+                                                                                   BComputeType>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvScale{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvScale{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return true;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale/conv3d_fwd_convscale_bf8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::bf8_t;
+using WeiDataType      = ck::bf8_t;
+using CShuffleDataType = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = InDataType;
+using BComputeDataType = AComputeDataType;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convscale<NumDimSpatial,
+                                          InDataType,
+                                          WeiDataType,
+                                          OutDataType,
+                                          InLayout,
+                                          WeiLayout,
+                                          OutLayout,
+                                          3,
+                                          AComputeDataType,
+                                          BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale/conv3d_fwd_convscale_bf8_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::bf8_t;
+using WeiDataType      = ck::f8_t;
+using CShuffleDataType = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::bf8_t;
+using BComputeDataType = ck::f8_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convscale<NumDimSpatial,
+                                          InDataType,
+                                          WeiDataType,
+                                          OutDataType,
+                                          InLayout,
+                                          WeiLayout,
+                                          OutLayout,
+                                          3,
+                                          AComputeDataType,
+                                          BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using CShuffleDataType = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convscale<NumDimSpatial,
+                                          InDataType,
+                                          WeiDataType,
+                                          OutDataType,
+                                          InLayout,
+                                          WeiLayout,
+                                          OutLayout,
+                                          3,
+                                          AComputeDataType,
+                                          BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale/conv3d_fwd_convscale_fp8_bf8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::bf8_t;
+using CShuffleDataType = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::bf8_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convscale<NumDimSpatial,
+                                          InDataType,
+                                          WeiDataType,
+                                          OutDataType,
+                                          InLayout,
+                                          WeiLayout,
+                                          OutLayout,
+                                          3,
+                                          AComputeDataType,
+                                          BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_add/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale_add.hpp"
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ConvScaleAdd = ck::tensor_operation::element_wise::ConvScaleAdd;
+using F32          = float;
+using BiasDataType = F32;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+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 <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
+{
+    // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * std::accumulate(std::begin(input_lengths),
+                                                std::end(input_lengths),
+                                                static_cast<std::size_t>(1),
+                                                std::multiplies<>());
+}
+
+template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * std::accumulate(std::begin(weights_lengths),
+                                                 std::end(weights_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<>());
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
+{
+    // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * std::accumulate(std::begin(output_lengths),
+                                                 std::end(output_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<std::size_t>());
+}
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumNonSpatialDim = 3,
+          typename AComputeType        = InDataType,
+          typename BComputeType        = AComputeType>
+bool run_grouped_conv_fwd_convscale_add(
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
+{
+
+    namespace ctc = ck::tensor_layout::convolution;
+    static_assert(NumDimSpatial == 3 && ck::is_same_v<InLayout, ctc::NDHWGC> &&
+                      ck::is_same_v<WeiLayout, ctc::GKZYXC> &&
+                      ck::is_same_v<OutLayout, ctc::NDHWGK>,
+                  "Unsupported configuration");
+
+    const ck::index_t G  = in_lengths[4];
+    const ck::index_t N  = in_lengths[0];
+    const ck::index_t K  = wei_lengths[1];
+    const ck::index_t C  = in_lengths[5];
+    const ck::index_t Z  = wei_lengths[2];
+    const ck::index_t Y  = wei_lengths[3];
+    const ck::index_t X  = wei_lengths[4];
+    const ck::index_t Di = in_lengths[1];
+    const ck::index_t Hi = in_lengths[2];
+    const ck::index_t Wi = in_lengths[3];
+    const ck::index_t Do = out_lengths[1];
+    const ck::index_t Ho = out_lengths[2];
+    const ck::index_t Wo = out_lengths[3];
+
+    const std::size_t in_mem_size   = sizeof(InDataType) * N * Di * Hi * Wi * G * C;
+    const std::size_t wei_mem_size  = sizeof(WeiDataType) * G * K * Z * Y * X * C;
+    const std::size_t out_mem_size  = sizeof(OutDataType) * N * Do * Ho * Wo * G * K;
+    const std::size_t bias_mem_size = sizeof(BiasDataType) * N * Do * Ho * Wo * G * K;
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+    SimpleDeviceMem bias(bias_mem_size);
+
+    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);
+
+    // We have NDHWGC/GKZYXC/NDHWGK (x, weight, y) in memory space.
+    // However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
+    // Hence, we need to adjust the order of strides.
+    const std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+    const std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+    const std::array<ck::index_t, NumDimSpatial + 3> weights_lengths{G, K, C, Z, Y, X};
+    const std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+    const std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+    const std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+    const std::array<ck::index_t, NumDimSpatial + 3> bias_lengths{G, N, K, Do, Ho, Wo};
+    const std::array<ck::index_t, NumDimSpatial + 3> bias_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+    const std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    std::size_t ds_size   = 3 + 1; // 3 element-wise scale multipliers + 1 elementwise Bias
+    std::size_t flop      = GetFlops<NumDimSpatial>(output_lengths, weights_lengths, ds_size);
+    std::size_t num_bytes = in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) +
+                            sizeof(float) + out_mem_size + bias_mem_size;
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                      InLayout,
+                                                                      WeiLayout,
+                                                                      ck::Tuple<OutLayout>,
+                                                                      OutLayout,
+                                                                      InDataType,
+                                                                      WeiDataType,
+                                                                      ck::Tuple<BiasDataType>,
+                                                                      OutDataType,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      ConvScaleAdd,
+                                                                      AComputeType,
+                                                                      BComputeType>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 1>{bias.GetDeviceBuffer()},
+            out.GetDeviceBuffer(),
+            input_lengths,
+            input_strides,
+            weights_lengths,
+            weights_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 1>{
+                {bias_lengths}},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 1>{
+                {bias_strides}},
+            output_lengths,
+            output_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvScaleAdd{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 1>{bias.GetDeviceBuffer()},
+            out.GetDeviceBuffer(),
+            input_lengths,
+            input_strides,
+            weights_lengths,
+            weights_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 1>{
+                {bias_lengths}},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 1>{
+                {bias_strides}},
+            output_lengths,
+            output_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvScaleAdd{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return true;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_add/conv3d_fwd_convscale_add_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using CShuffleDataType = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convscale_add<NumDimSpatial,
+                                              InDataType,
+                                              WeiDataType,
+                                              OutDataType,
+                                              InLayout,
+                                              WeiLayout,
+                                              OutLayout,
+                                              3,
+                                              AComputeDataType,
+                                              BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_reduce/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale_relu.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/permute_scale.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+using ConvScaleRelu = ck::tensor_operation::element_wise::ScaleScaleRelu;
+using ConvScale     = ck::tensor_operation::element_wise::ScaleScalePass;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+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)
+{
+    // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product> +
+    // + ds_size * <output tensor size> =>
+    // => <output tensor size> * ( 2 * C * <filter spatial lengths product> + ds_size) =>
+    // => G * N * K * <output spatial lengths product> * (2 * C * <filter spatial lengths product> +
+    // ds_size)
+    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 * K *
+           std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           (ds_size + static_cast<std::size_t>(2) * C *
+                          std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
+                                          std::end(weights_lengths),
+                                          static_cast<std::size_t>(1),
+                                          std::multiplies<>()));
+}
+
+template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t GetTensorSize(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths)
+{
+
+    return std::accumulate(std::begin(lengths),
+                           std::end(lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<std::size_t>());
+}
+
+template <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
+{
+    // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * GetTensorSize<NumDimSpatial>(input_lengths);
+}
+
+template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * GetTensorSize<NumDimSpatial>(weights_lengths);
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
+{
+    // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * GetTensorSize<NumDimSpatial>(output_lengths);
+}
+
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename ConvElementOp,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumDimSpatial,
+          ck::index_t NumNonSpatialDim = 3,
+          typename AComputeType        = InDataType,
+          typename BComputeType        = AComputeType>
+bool ConvolutionScale(SimpleDeviceMem& in,
+                      SimpleDeviceMem& wei,
+                      SimpleDeviceMem& out,
+                      ConvElementOp elementwise_op,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_lengths,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_strides,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_lengths,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_strides,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_lengths,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_strides);
+
+template <typename InDataType,
+          typename OutDataType,
+          ck::index_t NumDimSpatial,
+          ck::index_t NumNonSpatialDim = 3>
+bool TensorScaleConvert(SimpleDeviceMem& in,
+                        SimpleDeviceMem& out,
+                        float scale_out,
+                        const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
+                        const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides);
+
+template <typename InDataType,
+          typename OutDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t NumDimSpatial,
+          ck::index_t NumNonSpatialDim = 3>
+bool TensorFullReduction(SimpleDeviceMem& tensor,
+                         SimpleDeviceMem& out_amax,
+                         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
+                         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides);
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename ConvOutDataType,
+          typename OutDataType,
+          typename ConvElementOp,
+          ck::ReduceTensorOp ReduceOp,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumNonSpatialDim = 3,
+          typename AComputeType        = InDataType,
+          typename BComputeType        = AComputeType>
+bool run_grouped_conv_fwd_convscale_reduce(
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
+{
+
+    namespace ctc = ck::tensor_layout::convolution;
+    static_assert(NumDimSpatial == 3 && ck::is_same_v<InLayout, ctc::NDHWGC> &&
+                      ck::is_same_v<WeiLayout, ctc::GKZYXC> &&
+                      ck::is_same_v<OutLayout, ctc::NDHWGK>,
+                  "Unsupported configuration");
+
+    const ck::index_t G  = in_lengths[4];
+    const ck::index_t N  = in_lengths[0];
+    const ck::index_t K  = wei_lengths[1];
+    const ck::index_t C  = in_lengths[5];
+    const ck::index_t Z  = wei_lengths[2];
+    const ck::index_t Y  = wei_lengths[3];
+    const ck::index_t X  = wei_lengths[4];
+    const ck::index_t Di = in_lengths[1];
+    const ck::index_t Hi = in_lengths[2];
+    const ck::index_t Wi = in_lengths[3];
+    const ck::index_t Do = out_lengths[1];
+    const ck::index_t Ho = out_lengths[2];
+    const ck::index_t Wo = out_lengths[3];
+
+    const std::size_t in_mem_size       = sizeof(InDataType) * N * Di * Hi * Wi * G * C;
+    const std::size_t wei_mem_size      = sizeof(WeiDataType) * G * K * Z * Y * X * C;
+    const std::size_t conv_out_mem_size = sizeof(ConvOutDataType) * N * Do * Ho * Wo * G * K;
+    const std::size_t out_mem_size      = sizeof(OutDataType) * N * Do * Ho * Wo * G * K;
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem conv_out(conv_out_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+
+    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);
+
+    // We have NDHWGC/GKZYXC/NDHWGK (x, weight, y) in memory space.
+    // However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
+    // Hence, we need to adjust the order of strides.
+    const std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+    const std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+    const std::array<ck::index_t, NumDimSpatial + 3> weights_lengths{G, K, C, Z, Y, X};
+    const std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+    const std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+    const std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+
+    /*
+     * FP8 Convolution with Scaling
+     */
+    std::cout << "\n\nConvolution with scale Benchmarking:" << std::endl;
+    auto elementwise_op = ConvElementOp{ck::tensor_operation::element_wise::Scale{scale_in},
+                                        ck::tensor_operation::element_wise::Scale{scale_wei},
+                                        {}};
+    auto conv_ok        = ConvolutionScale<InDataType,
+                                    WeiDataType,
+                                    ConvOutDataType,
+                                    ConvElementOp,
+                                    InLayout,
+                                    WeiLayout,
+                                    OutLayout,
+                                    NumDimSpatial>(in,
+                                                   wei,
+                                                   conv_out,
+                                                   elementwise_op,
+                                                   input_lengths,
+                                                   input_strides,
+                                                   weights_lengths,
+                                                   weights_strides,
+                                                   output_lengths,
+                                                   output_strides);
+
+    if(!conv_ok)
+        return false;
+
+    /*
+     *  Scale with output weight and convert to FP8
+     */
+    std::cout << "\n\nElement-wise scale + convert Benchmarking:" << std::endl;
+    auto elem_wise_ok = TensorScaleConvert<ConvOutDataType, OutDataType, NumDimSpatial>(
+        conv_out, out, scale_out, output_lengths, output_strides);
+
+    if(!elem_wise_ok)
+        return false;
+
+    /*
+     *  Compute AMAX
+     */
+    std::cout << "\n\nAMAX Benchmarking:" << std::endl;
+    SimpleDeviceMem amax_device(sizeof(ConvOutDataType));
+    auto reduction_ok =
+        TensorFullReduction<ConvOutDataType,
+                            ConvOutDataType,
+                            ck::ReduceTensorOp::AMAX,
+                            NumDimSpatial>(conv_out, amax_device, output_lengths, output_strides);
+
+    if(!reduction_ok)
+        return false;
+
+    return true;
+}
+
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename ConvElementOp,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumDimSpatial,
+          ck::index_t NumNonSpatialDim,
+          typename AComputeType,
+          typename BComputeType>
+bool ConvolutionScale(SimpleDeviceMem& in,
+                      SimpleDeviceMem& wei,
+                      SimpleDeviceMem& out,
+                      ConvElementOp elementwise_op,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_lengths,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& in_strides,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_lengths,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& wei_strides,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_lengths,
+                      const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& out_strides)
+{
+
+    const std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    const auto in_mem_size  = GetInputByte<InDataType, NumDimSpatial>(in_lengths);
+    const auto wei_mem_size = GetWeightByte<WeiDataType, NumDimSpatial>(wei_lengths);
+    const auto out_mem_size = GetOutputByte<OutDataType, NumDimSpatial>(out_lengths);
+
+    std::size_t ds_size = 2; // 2 element-wise scale multipliers
+    if constexpr(ck::is_same_v<ConvElementOp, ConvScaleRelu>)
+    {
+        ds_size += 1; // +1 element-wise relu
+    }
+    std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths, ds_size);
+    std::size_t num_bytes =
+        in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) + out_mem_size;
+
+    using ConvDeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                      InLayout,
+                                                                      WeiLayout,
+                                                                      ck::Tuple<>,
+                                                                      OutLayout,
+                                                                      InDataType,
+                                                                      WeiDataType,
+                                                                      ck::Tuple<>,
+                                                                      OutDataType,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      ConvElementOp,
+                                                                      AComputeType,
+                                                                      BComputeType>;
+    // get device op instances
+    const auto conv_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        ConvDeviceOp>::GetInstances();
+
+    std::cout << "found " << conv_ptrs.size() << " instances" << std::endl;
+
+    std::string conv_best_op_name;
+    int conv_best_op_id        = -1;
+    float conv_best_avg_time   = std::numeric_limits<float>::max();
+    float conv_best_gb_per_sec = 0;
+    float conv_best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all convolution instances and do timing" << std::endl;
+
+    for(int i = 0; i < conv_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = conv_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            elementwise_op);
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > conv_best_tflops)
+            {
+                conv_best_op_id      = i;
+                conv_best_op_name    = op_name;
+                conv_best_avg_time   = avg_time;
+                conv_best_gb_per_sec = gb_per_sec;
+                conv_best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(conv_best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << conv_best_avg_time << " ms, " << conv_best_tflops
+              << " TFlops, " << conv_best_gb_per_sec << " GB/s, " << conv_best_op_name << std::endl;
+
+    // run the best instance
+    {
+        auto& op_ptr = conv_ptrs[conv_best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            elementwise_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return true;
+}
+
+template <typename InDataType,
+          typename OutDataType,
+          ck::index_t NumDimSpatial,
+          ck::index_t NumNonSpatialDim>
+bool TensorScaleConvert(SimpleDeviceMem& in,
+                        SimpleDeviceMem& out,
+                        float scale_out,
+                        const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
+                        const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides)
+{
+
+    const auto tensor_size = GetTensorSize<NumDimSpatial>(lengths);
+
+    const std::size_t in_mem_size  = sizeof(InDataType) * tensor_size;
+    const std::size_t out_mem_size = sizeof(OutDataType) * tensor_size;
+
+    std::size_t flop = 2 * tensor_size; // element-wise scale + convert
+
+    std::size_t bytes =
+        in_mem_size + sizeof(float) + out_mem_size; // read from in, scale, write to out
+
+    using DeviceScaleConvert =
+        ck::tensor_operation::device::DeviceElementwise<ck::Tuple<InDataType>,
+                                                        ck::Tuple<OutDataType>,
+                                                        ck::tensor_operation::element_wise::Scale,
+                                                        NumDimSpatial + NumNonSpatialDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceScaleConvert>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all DeviceScaleConvert instances and do timing" << std::endl;
+
+    auto scale_convert = ck::tensor_operation::element_wise::Scale{scale_out};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(lengths,
+                                                        {strides},
+                                                        {strides},
+                                                        {in.GetDeviceBuffer()},
+                                                        {out.GetDeviceBuffer()},
+                                                        scale_convert);
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance found." << std::endl;
+        return false;
+    }
+    else
+    {
+        std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+                  << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+        // run the best intance
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(lengths,
+                                                        {strides},
+                                                        {strides},
+                                                        {in.GetDeviceBuffer()},
+                                                        {out.GetDeviceBuffer()},
+                                                        scale_convert);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return true;
+}
+
+template <typename InDataType,
+          typename OutDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t NumDimSpatial,
+          ck::index_t NumNonSpatialDim>
+bool TensorFullReduction(SimpleDeviceMem& tensor,
+                         SimpleDeviceMem& out_amax,
+                         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& lengths,
+                         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& strides)
+{
+    const auto spatial_dim_size = std::accumulate(std::next(std::begin(lengths), NumNonSpatialDim),
+                                                  std::end(lengths),
+                                                  static_cast<std::size_t>(1),
+                                                  std::multiplies<>());
+    const auto tensor_size      = GetTensorSize<NumDimSpatial>(lengths);
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    // Get the reduction operation
+    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;
+
+    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>(tensor_size));
+
+    std::array<ck::index_t, 1> reduce_out_lengths{1};
+    std::array<ck::index_t, 1> reduce_out_strides{1};
+
+    SimpleDeviceMem partial_reduce_tensor(sizeof(OutDataType) * spatial_dim_size);
+    std::array<ck::index_t, NumDimSpatial> reduce_part_lengths;
+    std::copy(std::next(std::begin(lengths), NumNonSpatialDim),
+              std::end(lengths),
+              std::begin(reduce_part_lengths));
+    std::array<ck::index_t, NumDimSpatial> reduce_part_strides;
+    copy(HostTensorDescriptor(reduce_part_lengths).GetStrides(), reduce_part_strides);
+
+    {
+        std::cout << "\nReduction of nonspatial dimensions:" << std::endl;
+        using DeviceOp =
+            ck::tensor_operation::device::DeviceReduce<InDataType,
+                                                       OutDataType,
+                                                       OutDataType,
+                                                       NumDimSpatial + NumNonSpatialDim,
+                                                       NumNonSpatialDim,
+                                                       ReduceOperation,
+                                                       InElementwiseOperation,
+                                                       PassThrough,
+                                                       true,   // PropagateNan
+                                                       false>; // OutputIndex
+        const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+        std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+        std::string best_op_name;
+        int best_op_id        = -1;
+        float best_ave_time   = std::numeric_limits<float>::max();
+        float best_gb_per_sec = 0;
+
+        std::array<int, NumNonSpatialDim> reduce_dims;
+        std::iota(reduce_dims.begin(), reduce_dims.end(), 0); // 0,..., NumNonSpatialDim-1
+
+        ck::index_t num_in_elements  = tensor_size;
+        ck::index_t num_out_elements = spatial_dim_size;
+
+        // profile device operation instances
+        std::cout << "Run partial reduction and do timing" << std::endl;
+
+        for(int i = 0; i < op_ptrs.size(); ++i)
+        {
+            auto& op_ptr = op_ptrs[i];
+
+            auto argument_ptr   = op_ptr->MakeArgumentPointer(lengths,
+                                                            strides,
+                                                            reduce_part_lengths,
+                                                            reduce_part_strides,
+                                                            reduce_dims,
+                                                            1.0,
+                                                            0.0,
+                                                            tensor.GetDeviceBuffer(),
+                                                            nullptr,
+                                                            partial_reduce_tensor.GetDeviceBuffer(),
+                                                            nullptr,
+                                                            in_elementwise_op,
+                                                            PassThrough{});
+            auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+            std::string op_name = op_ptr->GetTypeString();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+                std::size_t num_bytes =
+                    num_in_elements * sizeof(InDataType) + num_out_elements * sizeof(OutDataType);
+
+                float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
+                          << " GB/s, " << op_name << std::endl;
+
+                if(ave_time < best_ave_time)
+                {
+                    best_op_id      = i;
+                    best_op_name    = op_name;
+                    best_ave_time   = ave_time;
+                    best_gb_per_sec = gb_per_sec;
+                }
+            }
+            else
+            {
+                std::cout << op_name << " does not support this problem" << std::endl;
+            }
+        }
+
+        if(best_op_id < 0)
+        {
+            std::cerr << "no suitable instance found." << std::endl;
+            return false;
+        }
+        else
+        {
+            std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                      << best_op_name << std::endl;
+
+            // run the best instance
+            auto& op_ptr = op_ptrs[best_op_id];
+            std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                      << std::endl;
+            auto argument_ptr = op_ptr->MakeArgumentPointer(lengths,
+                                                            strides,
+                                                            reduce_part_lengths,
+                                                            reduce_part_strides,
+                                                            reduce_dims,
+                                                            1.0,
+                                                            0.0,
+                                                            tensor.GetDeviceBuffer(),
+                                                            nullptr,
+                                                            partial_reduce_tensor.GetDeviceBuffer(),
+                                                            nullptr,
+                                                            in_elementwise_op,
+                                                            PassThrough{});
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+            }
+
+            std::cout << "Done" << std::endl;
+        }
+    }
+
+    {
+        std::cout << "\nReduction of spatial dimensions:" << std::endl;
+        using DeviceOp     = ck::tensor_operation::device::DeviceReduce<OutDataType,
+                                                                    OutDataType,
+                                                                    OutDataType,
+                                                                    NumDimSpatial,
+                                                                    NumDimSpatial,
+                                                                    ReduceOperation,
+                                                                    PassThrough,
+                                                                    AccElementwiseOperation,
+                                                                    true,   // PropagateNan
+                                                                    false>; // OutputIndex
+        const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+        std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+        std::string best_op_name;
+        int best_op_id        = -1;
+        float best_ave_time   = std::numeric_limits<float>::max();
+        float best_gb_per_sec = 0;
+
+        std::array<int, NumDimSpatial> reduce_dims;
+        std::iota(reduce_dims.begin(), reduce_dims.end(), 0); // 0,..., NumDimSpatial-1
+
+        ck::index_t num_in_elements  = spatial_dim_size;
+        ck::index_t num_out_elements = 1;
+
+        // profile device operation instances
+        std::cout << "Run final reduction and do timing" << std::endl;
+
+        for(int i = 0; i < op_ptrs.size(); ++i)
+        {
+            auto& op_ptr = op_ptrs[i];
+
+            auto argument_ptr   = op_ptr->MakeArgumentPointer(reduce_part_lengths,
+                                                            reduce_part_strides,
+                                                            reduce_out_lengths,
+                                                            reduce_out_strides,
+                                                            reduce_dims,
+                                                            1.0,
+                                                            0.0,
+                                                            partial_reduce_tensor.GetDeviceBuffer(),
+                                                            nullptr,
+                                                            out_amax.GetDeviceBuffer(),
+                                                            nullptr,
+                                                            PassThrough{},
+                                                            acc_elementwise_op);
+            auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+            std::string op_name = op_ptr->GetTypeString();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+                std::size_t num_bytes =
+                    num_in_elements * sizeof(OutDataType) + num_out_elements * sizeof(OutDataType);
+
+                float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
+                          << " GB/s, " << op_name << std::endl;
+
+                if(ave_time < best_ave_time)
+                {
+                    best_op_id      = i;
+                    best_op_name    = op_name;
+                    best_ave_time   = ave_time;
+                    best_gb_per_sec = gb_per_sec;
+                }
+            }
+            else
+            {
+                std::cout << op_name << " does not support this problem" << std::endl;
+            }
+        }
+
+        if(best_op_id < 0)
+        {
+            std::cerr << "no suitable instance found." << std::endl;
+            return false;
+        }
+        else
+        {
+            std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                      << best_op_name << std::endl;
+
+            // run the best instance
+            auto& op_ptr = op_ptrs[best_op_id];
+            std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                      << std::endl;
+            auto argument_ptr = op_ptr->MakeArgumentPointer(reduce_part_lengths,
+                                                            reduce_part_strides,
+                                                            reduce_out_lengths,
+                                                            reduce_out_strides,
+                                                            reduce_dims,
+                                                            1.0,
+                                                            0.0,
+                                                            partial_reduce_tensor.GetDeviceBuffer(),
+                                                            nullptr,
+                                                            out_amax.GetDeviceBuffer(),
+                                                            nullptr,
+                                                            PassThrough{},
+                                                            acc_elementwise_op);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+            }
+
+            std::cout << "Done" << std::endl;
+        }
+    }
+
+    return true;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_amax_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using CShuffleDataType = float;
+using ConvOutDataType  = float;    // data type of convolution result
+using OutDataType      = ck::f8_t; // data type of final result
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+using ConvElementOp = ConvScale;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+constexpr auto ReduceOpId = ck::ReduceTensorOp::AMAX;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convscale_reduce<NumDimSpatial,
+                                                 InDataType,
+                                                 WeiDataType,
+                                                 ConvOutDataType,
+                                                 OutDataType,
+                                                 ConvElementOp,
+                                                 ReduceOpId,
+                                                 InLayout,
+                                                 WeiLayout,
+                                                 OutLayout,
+                                                 3,
+                                                 AComputeDataType,
+                                                 BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_reduce/conv3d_fwd_convscale_relu_amax_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using CShuffleDataType = float;
+using ConvOutDataType  = float;    // data type of convolution result
+using OutDataType      = ck::f8_t; // data type of final result
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+using ConvElementOp = ConvScaleRelu;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+constexpr auto ReduceOpId = ck::ReduceTensorOp::AMAX;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convscale_reduce<NumDimSpatial,
+                                                 InDataType,
+                                                 WeiDataType,
+                                                 ConvOutDataType,
+                                                 OutDataType,
+                                                 ConvElementOp,
+                                                 ReduceOpId,
+                                                 InLayout,
+                                                 WeiLayout,
+                                                 OutLayout,
+                                                 3,
+                                                 AComputeDataType,
+                                                 BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_relu/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale_relu.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+using ConvScaleRelu = ck::tensor_operation::element_wise::ConvScaleRelu;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+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 <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
+{
+    // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * std::accumulate(std::begin(input_lengths),
+                                                std::end(input_lengths),
+                                                static_cast<std::size_t>(1),
+                                                std::multiplies<>());
+}
+
+template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * std::accumulate(std::begin(weights_lengths),
+                                                 std::end(weights_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<>());
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
+{
+    // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * std::accumulate(std::begin(output_lengths),
+                                                 std::end(output_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<std::size_t>());
+}
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumNonSpatialDim = 3,
+          typename AComputeType        = InDataType,
+          typename BComputeType        = AComputeType>
+bool run_grouped_conv_fwd_convscale_relu(
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
+{
+    std::size_t in_mem_size  = GetInputByte<InDataType, NumDimSpatial>(in_lengths);
+    std::size_t wei_mem_size = GetWeightByte<WeiDataType, NumDimSpatial>(wei_lengths);
+    std::size_t out_mem_size = GetOutputByte<OutDataType, NumDimSpatial>(out_lengths);
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+
+    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::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_strides;
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_strides;
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_strides;
+    in_strides.fill(0);
+    wei_strides.fill(0);
+    out_strides.fill(0);
+    in_strides.back()  = 1;
+    wei_strides.back() = 1;
+    out_strides.back() = 1;
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    // transpose NDHWGC/KZYXGC/NDHWGK to GNDHWC/GKZYXC/GNDHWK to GNCDHW/GKCZYX/GNKDHW
+    std::rotate(std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 2), rend(in_lengths));
+    std::rotate(rbegin(in_lengths),
+                std::next(rbegin(in_lengths)),
+                std::next(rbegin(in_lengths), NumDimSpatial + 1));
+
+    std::rotate(std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 2), rend(in_strides));
+    std::rotate(rbegin(in_strides),
+                std::next(rbegin(in_strides)),
+                std::next(rbegin(in_strides), NumDimSpatial + 1));
+
+    std::rotate(rbegin(wei_lengths),
+                std::next(rbegin(wei_lengths)),
+                std::next(rbegin(wei_lengths), NumDimSpatial + 1));
+
+    std::rotate(rbegin(wei_strides),
+                std::next(rbegin(wei_strides)),
+                std::next(rbegin(wei_strides), NumDimSpatial + 1));
+
+    std::rotate(
+        std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 2), rend(out_lengths));
+    std::rotate(rbegin(out_lengths),
+                std::next(rbegin(out_lengths)),
+                std::next(rbegin(out_lengths), NumDimSpatial + 1));
+
+    std::rotate(
+        std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 2), rend(out_strides));
+    std::rotate(rbegin(out_strides),
+                std::next(rbegin(out_strides)),
+                std::next(rbegin(out_strides), NumDimSpatial + 1));
+
+    std::array<ck::index_t, NumDimSpatial> conv_filter_strides;
+    std::array<ck::index_t, NumDimSpatial> conv_filter_dilations;
+    std::array<ck::index_t, NumDimSpatial> input_left_pads;
+    std::array<ck::index_t, NumDimSpatial> input_right_pads;
+    conv_filter_strides.fill(1);
+    conv_filter_dilations.fill(1);
+    input_left_pads.fill(1);
+    input_right_pads.fill(1);
+
+    std::size_t ds_size = 3 + 1; // 3 element-wise scale multipliers + 1 elementwise Relu
+    std::size_t flop    = GetFlops<NumDimSpatial>(out_lengths, wei_lengths, ds_size);
+    std::size_t num_bytes =
+        in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) + sizeof(float) + out_mem_size;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   ConvScaleRelu,
+                                                                                   AComputeType,
+                                                                                   BComputeType>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvScaleRelu{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvScaleRelu{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return true;
+}