Merge pull request #64 from ROCm/merge_from_public

Merge from public

client_example/16_convnd_fwd/conv3d_fwd_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 OutDataType = ck::f8_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+using AComputeType = ck::bf8_t;
+using BComputeType = ck::f8_t;
+
+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<NumDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InLayout,
+                                WeiLayout,
+                                OutLayout,
+                                3,
+                                AComputeType,
+                                BComputeType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/16_convnd_fwd/conv3d_fwd_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 OutDataType = 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<NumDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InLayout,
+                                WeiLayout,
+                                OutLayout,
+                                3,
+                                ck::f8_t>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/16_convnd_fwd/conv3d_fwd_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 OutDataType = ck::f8_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+using AComputeType = ck::f8_t;
+using BComputeType = ck::bf8_t;
+
+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<NumDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InLayout,
+                                WeiLayout,
+                                OutLayout,
+                                3,
+                                AComputeType,
+                                BComputeType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/17_grouped_gemm_fastgelu/CMakeLists.txt

-add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
-target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
\ No newline at end of file
+    target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/20_splitk_gemm/CMakeLists.txt

-if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
+if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES))
   add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
   target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_gemm_operations)
 endif()

client_example/21_grouped_gemm_bias/CMakeLists.txt

-add_executable(client_grouped_gemm_fixed_nk_bias_fp16 grouped_gemm_fixed_nk_bias_fp16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_gemm_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_grouped_gemm_fixed_nk_bias_fp16 grouped_gemm_fixed_nk_bias_fp16.cpp)
+    target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/22_grouped_gemm/CMakeLists.txt

-add_executable(client_grouped_gemm_fixed_nk_fp16 grouped_gemm_fixed_nk_fp16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_gemm_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_grouped_gemm_fixed_nk_fp16 grouped_gemm_fixed_nk_fp16.cpp)
+    target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_gemm_operations)
 
-add_executable(client_grouped_gemm_fixed_nk_fp8 grouped_gemm_fixed_nk_fp8.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_grouped_gemm_fixed_nk_fp8 grouped_gemm_fixed_nk_fp8.cpp)
+    target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_gemm_operations)
 
-add_executable(client_grouped_gemm_fixed_nk_i8 grouped_gemm_fixed_nk_i8.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_grouped_gemm_fixed_nk_i8 grouped_gemm_fixed_nk_i8.cpp)
+    target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_gemm_operations)
 
-add_executable(client_grouped_gemm_fixed_nk_bf16 grouped_gemm_fixed_nk_bf16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_bf16 PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_grouped_gemm_fixed_nk_bf16 grouped_gemm_fixed_nk_bf16.cpp)
+    target_link_libraries(client_grouped_gemm_fixed_nk_bf16 PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp

@@ -6,7 +6,7 @@
 #include <iostream>
 
 #include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.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"

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)
@@ -38,3 +39,20 @@ target_link_libraries(client_grouped_convnd_fwd_bilinear_residual_fp16 PRIVATE c
 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_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_scale/grouped_conv_fwd_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_forward_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_fwd_scale()
+{
+    // 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,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   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(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        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{},
+                                                        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 * 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(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_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_fwd_scale(); }

client_example/25_wrapper/CMakeLists.txt

@@ -2,9 +2,7 @@ add_executable(client_tensor_transform_using_wrapper tensor_transform_using_wrap
 target_link_libraries(client_tensor_transform_using_wrapper PRIVATE composable_kernel::device_other_operations)
 add_executable(client_wrapper_img2col wrapper_img2col.cpp)
 target_link_libraries(client_wrapper_img2col PRIVATE composable_kernel::device_other_operations)
-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR
-   GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR
-   GPU_TARGETS MATCHES "gfx942")
+if(GPU_TARGETS MATCHES "gfx9")
     add_executable(client_wrapper_basic_gemm wrapper_basic_gemm.cpp)
     target_link_libraries(client_wrapper_basic_gemm PRIVATE composable_kernel::device_other_operations)
     add_executable(client_wrapper_optimized_gemm wrapper_optimized_gemm.cpp)

client_example/25_wrapper/wrapper_img2col.cpp

@@ -181,4 +181,3 @@ int main(int argc, char* argv[])
                              {1, 1, 1} /*filter_dilations*/);
     return 0;
 }
-// MI100 Perf:   0.255178 ms, 1698.9 GB/s,

client_example/30_gemm_bf16Aint8B/CMakeLists.txt

+if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf16") OR NOT DEFINED DTYPES))
+	add_executable(client_gemm_bias_fastgelu_bf16_i8_bf16 gemm_bias_fastgelu_xdl_bf16_i8.cpp)
+	target_link_libraries(client_gemm_bias_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_gemm_bias_bf16_i8_bf16 gemm_bias_xdl_bf16_i8.cpp)
+	target_link_libraries(client_gemm_bias_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_gemm_gelu_bf16_i8_bf16 gemm_xdl_gelu_bf16_i8.cpp)
+	target_link_libraries(client_gemm_gelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_gemm_bf16_i8_bf16 gemm_xdl_bf16_i8.cpp)
+	target_link_libraries(client_gemm_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_gemm_multiply_bf16_i8_bf16 gemm_xdl_multiply_bf16_i8.cpp)
+	target_link_libraries(client_gemm_multiply_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/30_gemm_bf16Aint8B/gemm_bias_fastgelu_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = AddFastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+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_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 64;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideD = N;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideD = std::stoi(argv[6]);
+        StrideE = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD, ELayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    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;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    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(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_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 flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                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_tflops << " TFlops, "
+              << 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(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_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 0;
+}

client_example/30_gemm_bf16Aint8B/gemm_bias_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+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_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 64;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = M;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = N;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideD = std::stoi(argv[6]);
+        StrideE = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD, ELayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    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;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    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(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_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 flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                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_tflops << " TFlops, "
+              << 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(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_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 0;
+}

client_example/30_gemm_bf16Aint8B/gemm_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+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_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 0;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    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;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    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(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_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 flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                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_tflops << " TFlops, "
+              << 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(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_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 0;
+}

client_example/30_gemm_bf16Aint8B/gemm_xdl_gelu_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = FastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+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_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 64;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 0;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    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;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    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(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_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 flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                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_tflops << " TFlops, "
+              << 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(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_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 0;
+}