Reorganize project folders (#6)

client_example/01_gemm/CMakeLists.txt

@@ -0,0 +1,2 @@
+add_executable(client_gemm gemm.cpp)
+target_link_libraries(client_gemm PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)

client_example/01_gemm/README.md

@@ -0,0 +1,126 @@
+[Back to supported operations](../../../include/ck/README.md)
+# Composable Kernel GEMM
+
+## GEMM
+General matrix multiplications operation. In CK GEMM operation is called as `DeviceGemm` and requires following types as template parameters:
+
+* **ALayout** - A matrix layout (RowMajor/ColumnMajor).
+* **BLayout** - B matrix layout (RowMajor/ColumnMajor).
+* **CLayout** - B matrix layout (RowMajor/ColumnMajor).
+* **ADataType** - A matrix data type.
+* **BDataType** - B matrix data type.
+* **CDataType** - B matrix data type.
+* **AElementwiseOperation** - Fused operation on tensor A before GEMM.
+* **BElementwiseOperation** - Fused operation on tensor B before GEMM.
+* **CElementwiseOperation** - Fused operation on tensor C after GEMM.
+
+For matrices with large K dimension `DeviceGemmSplitK` implementation is available. This implementation allows user to split K dimension between work groups. This implementation uses `AtomicAdd` operation on global memory, thus need to zero-out output buffer for correct results.
+
+For fused operations with additional tensor there are `DeviceGemmMultipleABD` or `DeviceGemmMultipleD` operation which require following parameters:
+* **DsLayout** - layouts for additional tensors for fused operations.
+* **DsDataType** - data types for additional tensors for fused operations.
+
+For `DeviceGemmMultipleABD` **ALayout**, **BLayout**, **ADataType** and **BDataType** user should pass a tuple.
+
+List of the device operations in CK:
+
+* **DeviceGemmDl** - Device operation with DL instructions.
+* **DeviceGemmDpp** - Device operation with DL instructions with DPP instructions during data load.
+* **DeviceGemmWmma_CShuffle** - Device operation with WMMA instructions with CShuffle optimization for more optimized data store.
+* **DeviceGemm_Xdl_CShuffle_LdsDirectLoad** - Device operation with XDL instructions and CShuffle optimization for more optimized data store and direct load from global memory to shared memory.
+* **DeviceGemm_Xdl_CShuffle** - Device operation with XDL instructions with CShuffle optimization for more optimized data store.
+* **DeviceGemm_Xdl_CShuffleV2** - Device operation with XDL instructions with CShuffle optimization for more optimized data store. GEMM pipeline has been optimized compared to **DeviceGemm_Xdl_CShuffle**.
+* **DeviceGemmXdlSkipBLds** - Device operation with XDL instructions. Load to shared memory has been skiped for B matrix.
+* **DeviceGemm_Xdl_WaveletModel_CShuffle** - Device operation with XDL instructions with CShuffle optimization for more optimized data store. Producer and consumer scheme cooperation between waves in workgroup.
+* **DeviceGemmXdl** - Device operation with XDL instructions.
+
+Table of supported cases by instance factory with XDL instruction for Row/Row/Row, Row/Column/Row, Column/Row/Row or Column/Column/Row:
+
+|       |Is supported|
+|-------|---|
+|bf16|✓|
+|fp16|✓|
+|fp32|✓|
+|int8|✓|
+|fp8 |✓|
+
+Table of supported cases by instance factory with WMMA instruction for Row/Row/Row, Row/Column/Row, Column/Row/Row or Column/Column/Row:
+
+|       |Is supported|
+|-------|---|
+|bf16|✓|
+|fp16|✓|
+|fp32|✗|
+|int8|✓|
+|fp8 |✗|
+
+Table of supported cases by instance factory with DL instruction for Row/Row/Row, Row/Column/Row, Column/Row/Row or Column/Column/Row:
+
+|       |Is supported|
+|-------|---|
+|bf16|✗|
+|fp16|✓|
+|fp32|✓|
+|int8|✓|
+|fp8 |✗|
+
+Table of supported cases by instance factory with fused output elementwise operation:
+
+* **B Matrix Multiply + Add + Gelu** - bf16 (int8 for B matrix)
+* **B Matrix Multiply + Add** - bf16 (int8 for B matrix)
+* **B Matrix Multiply + Gelu** - bf16 (int8 for B matrix)
+* **B Matrix Multiply** - bf16 (int8 for B matrix)
+
+* **Add + Add + Gelu** - fp16
+* **Add + Gelu** - fp16, bf16 (int8 for B matrix) for Row/Column/Row
+* **Multiply** - fp16
+* **Add + Multiply** - fp16
+* **Add + Relu** - fp16 (int8 for B matrix)  for Row/Column/Row, bf16 (int8 for B matrix) for Row/Column/Row
+* **Add + Silu** - fp16 (int8 for B matrix)  for Row/Column/Row, bf16 (int8 for B matrix) for Row/Column/Row
+* **Add** - fp16 (int8 for B matrix)  for Row/Column/Row, bf16 (int8 for B matrix) for Row/Column/Row
+* **Bilinear** - fp16, int8
+* **Gelu** - fp16
+* **Multiply + Add** - fp16 for Row/Column/Row and Row/Row/Row, fp16 (int8 for B matrix, fp32 for Bias) for Row/Column/Row and Row/Row/Row, 
+* **Quantization** - int8
+
+## GEMM V2 (Universal GEMM)
+General matrix multiplications operation optimized for MI300 series. Operation is called as `DeviceGemmV2` and requires following types as template parameters:
+
+* **ALayout** - A matrix layout (RowMajor/ColumnMajor).
+* **BLayout** - B matrix layout (RowMajor/ColumnMajor).
+* **CLayout** - B matrix layout (RowMajor/ColumnMajor).
+* **ADataType** - A matrix data type.
+* **BDataType** - B matrix data type.
+* **CDataType** - B matrix data type.
+* **AElementwiseOperation** - Fused operation on tensor A before GEMM.
+* **BElementwiseOperation** - Fused operation on tensor B before GEMM.
+* **CElementwiseOperation** - Fused operation on tensor C after GEMM.
+
+This implementation allows user to split K dimension between work groups. This implementation requires AtomicAdd operation on global memory (output buffer must be set to zeroes if splitK parameter is larger than one).
+
+List of the device operations for in CK:
+
+* **DeviceGemm_Xdl_CShuffleV3** - Device operation with XDL instructions with CShuffle optimization for more optimized data store.
+* **DeviceGemm_Xdl_CShuffleV3R1** - Device operation with XDL instructions with CShuffle optimization for more optimized data store. This implementation perform reduction on splitted K dimension after GEMM instead of AtomicAdd instruction. 
+
+Table of supported cases by instance factory with XDL instruction for Row/Row/Row, Row/Column/Row, Column/Row/Row or Column/Column/Row:
+
+|       |Is supported|
+|-------|---|
+|bf16|✓|
+|fp16|✓|
+|fp32|✗|
+|int8|✗|
+|fp8 (C bf16)|✓|
+|fp16 (A fp8)|✓|
+|fp16 (B fp8)|✓|
+
+## Others
+
+* **DeviceGemm_dequantB** - GEMM with dequantization (implemented with WMMA instructions).
+* **DeviceGemmMultipleD_ABScale** - GEMM with scale for A and B matrix.
+* **DeviceGemmMultipleDLayernorm** - GEMM fused with layernorm.
+* **DeviceGemmMultipleDMultipleR** - GEMM fused with reductions and custom global reductions operators.
+* **DeviceGemmReduce** - GEMM fused with reduction.
+* **DeviceGemm_Streamk_V2** - GEMM stream K implementation. Implementation allows to use reduction instead of AtomicAdd.
+* **DeviceGemmStreamK** - GEMM stream K implementation using AtomicAdd.

client_example/01_gemm/gemm.cpp

@@ -0,0 +1,219 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+using ADataType = F16;
+using BDataType = F16;
+using CDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    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]);
+        StrideC = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideC\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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemm<ALayout,
+                                                 BLayout,
+                                                 CLayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 CDataType,
+                                                 ck::tensor_operation::element_wise::PassThrough,
+                                                 ck::tensor_operation::element_wise::PassThrough,
+                                                 ck::tensor_operation::element_wise::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;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_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(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << 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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_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/02_gemm_add_add_fastgelu/CMakeLists.txt

@@ -0,0 +1,29 @@
+if(GPU_TARGETS MATCHES "gfx9")
+    add_custom_target(client_gemm_fastgelu_examples)
+
+    add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
+    target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+
+    add_executable(client_gemm_add_fastgelu gemm_add_fastgelu.cpp)
+    target_link_libraries(client_gemm_add_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+
+    add_executable(client_gemm_fastgelu gemm_fastgelu.cpp)
+    target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+
+    add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu
+                 client_gemm_fastgelu)
+
+    add_custom_target(client_gemm_fastgelu_generic_examples)
+
+    add_executable(client_gemm_add_add_fastgelu_generic gemm_add_add_fastgelu_generic.cpp)
+    target_link_libraries(client_gemm_add_add_fastgelu_generic composable_kernel::device_gemm_operations)
+
+    add_executable(client_gemm_add_fastgelu_generic gemm_add_fastgelu_generic.cpp)
+    target_link_libraries(client_gemm_add_fastgelu_generic PRIVATE composable_kernel::device_gemm_operations)
+
+    add_executable(client_gemm_fastgelu_generic gemm_fastgelu_generic.cpp)
+    target_link_libraries(client_gemm_fastgelu_generic PRIVATE composable_kernel::device_gemm_operations)
+
+    add_dependencies(client_gemm_fastgelu_generic_examples client_gemm_add_add_fastgelu_generic 
+                 client_gemm_add_fastgelu_generic client_gemm_fastgelu_generic)
+endif()

client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp

@@ -0,0 +1,242 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        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]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, 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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            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(ADataType) * M * K + sizeof(BDataType) * 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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            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/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp

@@ -0,0 +1,176 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        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]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, 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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    auto argument_ptr =
+        op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                                               d1_m_n_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    M,
+                                    N,
+                                    K,
+                                    StrideA,
+                                    StrideB,
+                                    std::array<ck::index_t, 2>{StrideD0, StrideD1},
+                                    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});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}

client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp

@@ -0,0 +1,234 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using ELayout  = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideE  = 4096;
+
+    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]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideE  = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, 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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddFastGelu>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            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(ADataType) * M * K + sizeof(BDataType) * 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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            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/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp

@@ -0,0 +1,169 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using ELayout  = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideE  = 4096;
+
+    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]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideE  = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, 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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddFastGelu>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    auto argument_ptr =
+        op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    M,
+                                    N,
+                                    K,
+                                    StrideA,
+                                    StrideB,
+                                    std::array<ck::index_t, 1>{StrideD0},
+                                    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});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}

client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp

@@ -0,0 +1,226 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = FastGelu;
+
+using ADataType = F16;
+using BDataType = F16;
+using EDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideE = 4096;
+
+    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 6: 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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::FastGelu>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        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(ADataType) * M * K + sizeof(BDataType) * 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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        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/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp

@@ -0,0 +1,162 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = FastGelu;
+
+using ADataType = F16;
+using BDataType = F16;
+using EDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideE = 4096;
+
+    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 6: 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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::FastGelu>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                    b_device_buf.GetDeviceBuffer(),
+                                                    {},
+                                                    e_device_buf.GetDeviceBuffer(),
+                                                    M,
+                                                    N,
+                                                    K,
+                                                    StrideA,
+                                                    StrideB,
+                                                    {},
+                                                    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});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}

client_example/03_gemm_layernorm/CMakeLists.txt

@@ -0,0 +1,7 @@
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_gemm_add_add_layernorm_naive gemm_add_add_layernorm_naive.cpp)
+    target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_gemm_operations composable_kernel::device_other_operations)
+
+    add_executable(client_gemm_add_relu_add_layernorm_welford gemm_add_relu_add_layernorm_welford.cpp)
+    target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_gemm_operations composable_kernel::device_other_operations)
+endif()

client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp

@@ -0,0 +1,277 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.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/device_elementwise_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+
+using ADataType            = F16;
+using BDataType            = F16;
+using BiasDataType         = F32;
+using CDataType            = F16;
+using D0DataType           = F16;
+using ReduceDataType       = F32;
+using GammaDataType        = F16;
+using BetaDataType         = F16;
+using LayerNormOutDataType = F16;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+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 <typename gemm_reduce_op_ptr>
+bool RunDeviceGemmMeanSquareMean(gemm_reduce_op_ptr& p_op,
+                                 const void* p_a,
+                                 const void* p_b,
+                                 const void* p_bias,
+                                 const void* p_d0,
+                                 void* p_c,
+                                 void* p_mean,
+                                 void* p_square_mean,
+                                 int M,
+                                 int N,
+                                 int K,
+                                 int StrideA,
+                                 int StrideB,
+                                 int StrideC,
+                                 int StrideD0,
+                                 bool time_kernel)
+{
+    using PassThrough          = ck::tensor_operation::element_wise::PassThrough;
+    using UnaryDivElementOp    = ck::tensor_operation::element_wise::UnaryDivide;
+    using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
+
+    auto passOp   = PassThrough{};
+    auto squareOp = UnarySquareElementOp{};
+    auto divOp    = UnaryDivElementOp{N};
+
+    auto argument_ptr =
+        p_op->MakeArgumentPointer(p_a,
+                                  p_b,
+                                  p_bias,
+                                  {p_d0},
+                                  p_c,
+                                  {p_mean, p_square_mean},
+                                  M,
+                                  N,
+                                  K,
+                                  StrideA,
+                                  StrideB,
+                                  StrideC,
+                                  {StrideD0},
+                                  {&passOp, &passOp, &passOp}, // functor for a, b, c
+                                  {&passOp},                   // functor for d0
+                                  {&passOp, &squareOp},        // functor for inputs of reduction
+                                  {&divOp, &divOp});           // functor for outputs of reduction
+
+    if(p_op->IsSupportedArgument(argument_ptr.get()))
+    {
+        auto invoker_ptr = p_op->MakeInvokerPointer();
+
+        // If we evaluate running time of gemm_reduce. The output may wrong.
+        // Because we need to initialize the reduction tensor before runing the kernel.
+        // However we run kernel many times for time_kernel = trie without reinitialize the out
+        // of reduction tensor.
+        float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            std::cout << "Gemm + reduce Perf: " << std::setw(10) << ave_time << " ms" << std::endl;
+
+        return true;
+    }
+
+    return false;
+}
+
+template <typename normalize_op_ptr>
+bool RunDeviceNormalize2D(normalize_op_ptr& p_op,
+                          const void* p_x,
+                          const void* p_mean,
+                          const void* p_square_mean,
+                          const void* p_gamma,
+                          const void* p_beta,
+                          void* p_y,
+                          int M,
+                          int N,
+                          int StrideX,
+                          bool time_kernel)
+{
+    std::array<const void*, 5> input = {p_x, p_mean, p_square_mean, p_gamma, p_beta};
+    std::array<void*, 1> output      = {p_y};
+    auto normalize_functor           = ck::tensor_operation::element_wise::Normalize{};
+
+    std::array<ck::index_t, 2> xyLengths = {M, N};
+    std::array<ck::index_t, 2> xyStrides = {StrideX, 1};
+
+    auto argument_ptr = p_op->MakeArgumentPointer(xyLengths,
+                                                  {xyStrides, {1, 0}, {1, 0}, {0, 1}, {0, 1}},
+                                                  {xyStrides},
+                                                  input,
+                                                  output,
+                                                  ck::tensor_operation::element_wise::Normalize{});
+
+    if(p_op->IsSupportedArgument(argument_ptr.get()))
+    {
+        auto invoker_ptr = p_op->MakeInvokerPointer();
+        float ave_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            std::cout << "Normalize Perf: " << std::setw(10) << ave_time << " ms" << std::endl;
+
+        return true;
+    }
+
+    return false;
+}
+
+int main()
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = 1024;
+    ck::index_t StrideB  = 1024;
+    ck::index_t StrideC  = 1024;
+    ck::index_t StrideD0 = 1024;
+
+    const auto gemm_reduce_ptrs =
+        ck::tensor_operation::device::instance::get_device_gemm_add_add_mean_squaremean_instances<
+            ADataType,
+            BDataType,
+            CDataType,
+            ALayout,
+            BLayout,
+            CLayout>();
+
+    std::cout << "found " << gemm_reduce_ptrs.size()
+              << " gemm_reduceMean_reduceSquareMean instances" << std::endl;
+
+    using NormalizeDeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<CDataType, ReduceDataType, ReduceDataType, GammaDataType, BetaDataType>,
+        ck::Tuple<LayerNormOutDataType>,
+        ck::tensor_operation::element_wise::Normalize,
+        2>;
+
+    const auto normalize_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            NormalizeDeviceOp>::GetInstances();
+
+    std::cout << "found " << normalize_ptrs.size() << " normalize instances" << std::endl;
+
+    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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem bias_device_buf(sizeof(BiasDataType) * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD0, CLayout{}));
+    SimpleDeviceMem reduceMean_device_buf(sizeof(ReduceDataType) * M);
+    SimpleDeviceMem reduceMeanSquare_device_buf(sizeof(ReduceDataType) * M);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) * M * N);
+
+    bool b_time_kernel           = true;
+    bool b_only_run_first_kernel = true;
+
+    // layernorm => (1) + (2)
+    // (1). c = gemm(a, b), reduce_mean(c), reduce_square_mean(c)
+    // (2). normalize(c, mean, square_mean, gamma, beta)
+    for(auto& gemm_reduce_ptr : gemm_reduce_ptrs)
+    {
+        // run first available kernel
+        if(RunDeviceGemmMeanSquareMean(gemm_reduce_ptr,
+                                       a_device_buf.GetDeviceBuffer(),
+                                       b_device_buf.GetDeviceBuffer(),
+                                       bias_device_buf.GetDeviceBuffer(),
+                                       d0_device_buf.GetDeviceBuffer(),
+                                       c_device_buf.GetDeviceBuffer(),
+                                       reduceMean_device_buf.GetDeviceBuffer(),
+                                       reduceMeanSquare_device_buf.GetDeviceBuffer(),
+                                       M,
+                                       N,
+                                       K,
+                                       StrideA,
+                                       StrideB,
+                                       StrideC,
+                                       StrideD0,
+                                       b_time_kernel))
+        {
+            if(b_only_run_first_kernel)
+                break;
+        }
+        else
+        {
+            std::cout << gemm_reduce_ptr->GetTypeString() << " does not support this problem"
+                      << std::endl;
+        }
+    }
+
+    for(auto& normalize_ptr : normalize_ptrs)
+    {
+        if(RunDeviceNormalize2D(normalize_ptr,
+                                c_device_buf.GetDeviceBuffer(),
+                                reduceMean_device_buf.GetDeviceBuffer(),
+                                reduceMeanSquare_device_buf.GetDeviceBuffer(),
+                                gamma_device_buf.GetDeviceBuffer(),
+                                beta_device_buf.GetDeviceBuffer(),
+                                layerNorm_device_buf.GetDeviceBuffer(),
+                                M,
+                                N,
+                                StrideC,
+                                b_time_kernel))
+        {
+            if(b_only_run_first_kernel)
+                break;
+        }
+        else
+        {
+            std::cout << normalize_ptr->GetTypeString() << " does not support this problem"
+                      << std::endl;
+        }
+    }
+}

client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp

@@ -0,0 +1,245 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+// DataType
+using ADataType     = F16;
+using BDataType     = F16;
+using D0DataType    = F16;
+using D1DataType    = F16;
+using GammaDataType = F16;
+using BetaDataType  = F16;
+using HDataType     = F16;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using HLayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddReluAdd;
+using HElementOp   = PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+    std::size_t mMemSize_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = K;
+    ck::index_t StrideB  = K;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = N;
+    ck::index_t StrideH  = N;
+
+    float epsilon = 1e-5;
+
+    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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_device_buf(sizeof(D1DataType) *
+                                  f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem h_device_buf(sizeof(HDataType) * f_matrix_space_size(M, N, StrideH, HLayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        HLayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        GammaDataType,
+        BetaDataType,
+        HDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddReluAdd,
+        ck::tensor_operation::element_wise::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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+    const auto h_element_op   = HElementOp{};
+
+    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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+            h_device_buf.SetZero();
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                (sizeof(D0DataType) + sizeof(D1DataType) + sizeof(HDataType)) * M * N +
+                (sizeof(GammaDataType) + sizeof(BetaDataType)) * N;
+
+            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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+            h_device_buf.SetZero();
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/04_contraction/CMakeLists.txt

@@ -0,0 +1,16 @@
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_contraction_scale_fp32 contraction_scale_fp32.cpp)
+    target_link_libraries(client_contraction_scale_fp32 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_contraction_bilinear_fp32 contraction_bilinear_fp32.cpp)
+    target_link_libraries(client_contraction_bilinear_fp32 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_contraction_scale_fp64 contraction_scale_fp64.cpp)
+    target_link_libraries(client_contraction_scale_fp64 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_contraction_bilinear_fp64 contraction_bilinear_fp64.cpp)
+    target_link_libraries(client_contraction_bilinear_fp64 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+
+    add_executable(contraction_g1m2n3k1_add_xdl_fp16 contraction_g1m2n3k1_add_xdl_fp16.cpp)
+    target_link_libraries(contraction_g1m2n3k1_add_xdl_fp16 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+endif()

client_example/04_contraction/contraction_bilinear_fp32.cpp

@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Bilinear;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F32;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+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(int argc, char* argv[])
+{
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 25)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21]), std::stoi(argv[22])};
+
+        alpha = std::stof(argv[23]);
+        beta  = std::stof(argv[24]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg19 to 22: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg23 to 24: alpha, beta\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_ms_ns_lengths, d_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{alpha, beta};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                        e_device_buf.GetDeviceBuffer(),
+                                        a_ms_ks_lengths,
+                                        a_ms_ks_strides,
+                                        b_ns_ks_lengths,
+                                        b_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                        e_ms_ns_lengths,
+                                        e_ms_ns_strides,
+                                        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});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop      = std::size_t(2) * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(DDataType) * M * 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;
+
+    return 0;
+}

client_example/04_contraction/contraction_bilinear_fp64.cpp

@@ -0,0 +1,281 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Bilinear;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DDataType        = F64;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+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(int argc, char* argv[])
+{
+// kknn
+#if 1
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// knnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mknn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mnnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+#endif
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 25)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21]), std::stoi(argv[22])};
+
+        alpha = std::stof(argv[23]);
+        beta  = std::stof(argv[24]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg19 to 22: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg23 to 24: alpha, beta\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_ms_ns_lengths, d_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{alpha, beta};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                        e_device_buf.GetDeviceBuffer(),
+                                        a_ms_ks_lengths,
+                                        a_ms_ks_strides,
+                                        b_ns_ks_lengths,
+                                        b_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                        e_ms_ns_lengths,
+                                        e_ms_ns_strides,
+                                        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});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop      = std::size_t(2) * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(DDataType) * M * 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;
+
+    return 0;
+}

client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp

@@ -0,0 +1,204 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 1;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 3;
+static constexpr ck::index_t NumDimK = 1;
+
+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(int argc, char* argv[])
+{
+    ck::index_t G0 = 1;
+
+    ck::index_t M0 = 64;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 3;
+    ck::index_t N1 = 12;
+    ck::index_t N2 = 64;
+
+    ck::index_t K0 = 768;
+
+    // A[M0, M1, M2, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[N0, N1, N2, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, N0, N1, N2, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{N0 * N1 * N2 * K0, N1 * N2 * K0, N2 * K0, K0, 1};
+
+    // D[N0, M0, N1, M1, N2]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{N0 * N1 * N2, 0, 0, N1 * N2, N2, 1};
+    // E[N0 M0 N1 N2 M1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        M0 * M1 * N0 * N1 * N2, N1 * N2 * M1, 1, M0 * N1 * N2 * M1, M1 * N2, M1};
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_gs_ms_ks_lengths, a_gs_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_gs_ns_ks_lengths, b_gs_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_gs_ms_ns_lengths, d_gs_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_gs_ms_ns_lengths, e_gs_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchedContractionMultipleD<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        DsDataType,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Add>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                        e_device_buf.GetDeviceBuffer(),
+                                        a_gs_ms_ks_lengths,
+                                        a_gs_ms_ks_strides,
+                                        b_gs_ns_ks_lengths,
+                                        b_gs_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                        e_gs_ms_ns_lengths,
+                                        e_gs_ms_ns_strides,
+                                        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});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop      = std::size_t(2) * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(DDataType) * M * 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;
+
+    return 0;
+}

client_example/04_contraction/contraction_scale_fp32.cpp

@@ -0,0 +1,222 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Scale;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+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(int argc, char* argv[])
+{
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 20)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        scale = std::stof(argv[19]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg19: scale\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{scale};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        a_ms_ks_lengths,
+                                                        a_ms_ks_strides,
+                                                        b_ns_ks_lengths,
+                                                        b_ns_ks_strides,
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        e_ms_ns_lengths,
+                                                        e_ms_ns_strides,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        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});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * 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;
+
+    return 0;
+}

client_example/04_contraction/contraction_scale_fp64.cpp

@@ -0,0 +1,270 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Scale;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+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(int argc, char* argv[])
+{
+// kkn
+#if 1
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// knn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mkn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+#endif
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 20)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        scale = std::stof(argv[19]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg19: scale\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{scale};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        a_ms_ks_lengths,
+                                                        a_ms_ks_strides,
+                                                        b_ns_ks_lengths,
+                                                        b_ns_ks_strides,
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        e_ms_ns_lengths,
+                                                        e_ms_ns_strides,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        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});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * 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;
+
+    return 0;
+}

client_example/05_layernorm/CMakeLists.txt

@@ -0,0 +1,11 @@
+add_executable(client_layernorm2d_bwd_data layernorm2d_bwd_data.cpp)
+target_link_libraries(client_layernorm2d_bwd_data PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_layernorm2d_bwd_gamma_beta layernorm2d_bwd_gamma_beta.cpp)
+target_link_libraries(client_layernorm2d_bwd_gamma_beta PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_layernorm2d_fwd layernorm2d_fwd.cpp)
+target_link_libraries(client_layernorm2d_fwd PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_layernorm4d_fwd layernorm4d_fwd.cpp)
+target_link_libraries(client_layernorm4d_fwd PRIVATE composable_kernel::device_other_operations)

client_example/05_layernorm/layernorm2d_bwd_data.cpp

@@ -0,0 +1,170 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/layernorm_bwd_data.hpp"
+
+using DYDataType         = float;
+using XDataType          = float;
+using GammaDataType      = float;
+using MeanInvStdDataType = float;
+using DXDataType         = float;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+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(int argc, char* argv[])
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+
+    SimpleDeviceMem dy_dev(sizeof(DYDataType) * M * N);
+    SimpleDeviceMem x_dev(sizeof(XDataType) * M * N);
+    SimpleDeviceMem gamma_dev(sizeof(GammaDataType) * N);
+    SimpleDeviceMem mean_dev(sizeof(MeanInvStdDataType) * M);
+    SimpleDeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * M);
+    SimpleDeviceMem dx_dev(sizeof(DXDataType) * M * N);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationBwdData<DYDataType,
+                                                                              XDataType,
+                                                                              GammaDataType,
+                                                                              MeanInvStdDataType,
+                                                                              DXDataType,
+                                                                              Rank,
+                                                                              NumReduceDim>;
+
+    // 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;
+    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({M, N}, // lengths
+                                                        {N, 1}, // dyStrides
+                                                        {N, 1}, // xStrides
+                                                        {0, 1}, // gammaStrides
+                                                        {1, 0}, // meanStrides
+                                                        {1, 0}, // invStdStrides
+                                                        {N, 1}, // dxStrides
+                                                        {1},    // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        gamma_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dx_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte = sizeof(DYDataType) * M * N + sizeof(XDataType) * M * N +
+                                   sizeof(GammaDataType) * N + sizeof(MeanInvStdDataType) * M * 2 +
+                                   sizeof(DXDataType) * M * N;
+
+            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({M, N}, // lengths
+                                                        {N, 1}, // dyStrides
+                                                        {N, 1}, // xStrides
+                                                        {0, 1}, // gammaStrides
+                                                        {1, 0}, // meanStrides
+                                                        {1, 0}, // invStdStrides
+                                                        {N, 1}, // dxStrides
+                                                        {1},    // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        gamma_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dx_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/05_layernorm/layernorm2d_bwd_gamma_beta.cpp

@@ -0,0 +1,171 @@
+// 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/device_normalization_bwd_gamma_beta.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/layernorm_bwd_gamma_beta.hpp"
+
+using DYDataType         = float;
+using XDataType          = float;
+using GammaDataType      = float;
+using MeanInvStdDataType = float;
+using DGammaDataType     = float;
+using DBetaDataType      = float;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+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(int argc, char* argv[])
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+
+    SimpleDeviceMem dy_dev(sizeof(DYDataType) * M * N);
+    SimpleDeviceMem x_dev(sizeof(XDataType) * M * N);
+    SimpleDeviceMem mean_dev(sizeof(MeanInvStdDataType) * M);
+    SimpleDeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * M);
+    SimpleDeviceMem dgamma_dev(sizeof(DGammaDataType) * N);
+    SimpleDeviceMem dbeta_dev(sizeof(DBetaDataType) * N);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceNormalizationBwdGammaBeta<DYDataType,
+                                                                      XDataType,
+                                                                      MeanInvStdDataType,
+                                                                      DGammaDataType,
+                                                                      DBetaDataType,
+                                                                      Rank,
+                                                                      NumReduceDim>;
+
+    // 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;
+    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;
+
+    std::size_t num_bytes = sizeof(DYDataType) * M * N + sizeof(XDataType) * M * N +
+                            sizeof(MeanInvStdDataType) * M * 2 + sizeof(DGammaDataType) * N +
+                            sizeof(DBetaDataType) * N;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // inLengths
+                                                        {N, 1}, // dyStrides
+                                                        {N, 1}, // xStrides
+                                                        {1, 0}, // meanStrides
+                                                        {1, 0}, // invStdStrides
+                                                        {N},    // outLengths
+                                                        {1},    // dgammaStrides
+                                                        {1},    // dbetaStrides
+                                                        {0},    // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dgamma_dev.GetDeviceBuffer(),
+                                                        dbeta_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+            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)
+            {
+                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({M, N}, // inLengths
+                                                        {N, 1}, // dyStrides
+                                                        {N, 1}, // xStrides
+                                                        {1, 0}, // meanStrides
+                                                        {1, 0}, // invStdStrides
+                                                        {N},    // outLengths
+                                                        {1},    // dgammaStrides
+                                                        {1},    // dbetaStrides
+                                                        {0},    // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dgamma_dev.GetDeviceBuffer(),
+                                                        dbeta_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/05_layernorm/layernorm2d_fwd.cpp

@@ -0,0 +1,196 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp"
+
+using XDataType              = ck::half_t;
+using GammaDataType          = ck::half_t;
+using BetaDataType           = ck::half_t;
+using YDataType              = ck::half_t;
+using SaveMeanInvStdDataType = ck::half_t;
+using PassThrough            = ck::tensor_operation::element_wise::PassThrough;
+
+#define SAVE_MEAN_INV_STD
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+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(int argc, char* argv[])
+{
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = 1024;
+
+    auto xy_size = (M - 1) * Stride + N;
+
+    SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
+#ifdef SAVE_MEAN_INV_STD
+    SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * M);
+    SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * M);
+#endif
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
+                                                                          GammaDataType,
+                                                                          BetaDataType,
+                                                                          YDataType,
+                                                                          SaveMeanInvStdDataType,
+                                                                          PassThrough,
+                                                                          Rank,
+                                                                          NumReduceDim>;
+
+    // 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;
+    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({M, N},      // lengths
+                                                        {Stride, 1}, // xStrides
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // save_mean Strides
+                                                        {1},         // save_inv_std Strides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                                        save_mean_device_buf.GetDeviceBuffer(),
+                                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                                        nullptr,
+                                                        nullptr,
+#endif
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
+                                   sizeof(BetaDataType) * N + sizeof(YDataType) * M * N;
+
+#ifdef SAVE_MEAN_INV_STD
+            num_byte += sizeof(SaveMeanInvStdDataType) * M * 2;
+#endif
+
+            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({M, N},      // lengths
+                                                        {Stride, 1}, // xStrides
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // save_mean Strides
+                                                        {1},         // save_inv_std Strides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                                        save_mean_device_buf.GetDeviceBuffer(),
+                                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                                        nullptr,
+                                                        nullptr,
+#endif
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/05_layernorm/layernorm4d_fwd.cpp

@@ -0,0 +1,202 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp"
+
+using XDataType              = ck::half_t;
+using GammaDataType          = ck::half_t;
+using BetaDataType           = ck::half_t;
+using YDataType              = ck::half_t;
+using SaveMeanInvStdDataType = ck::half_t;
+using PassThrough            = ck::tensor_operation::element_wise::PassThrough;
+
+#define SAVE_MEAN_INV_STD
+
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+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(int argc, char* argv[])
+{
+    ck::index_t N = 256;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t C = 8;
+
+    std::vector<ck::index_t> strideXY             = {H * W * C, W * C, C, 1};
+    std::vector<ck::index_t> strideGammaBeta      = {0, W * C, C, 1};
+    std::vector<ck::index_t> strideSaveMeanInvStd = {1};
+
+    SimpleDeviceMem x_device_buf(sizeof(XDataType) * N * H * W * C);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * H * W * C);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * H * W * C);
+    SimpleDeviceMem y_device_buf(sizeof(YDataType) * N * H * W * C);
+#ifdef SAVE_MEAN_INV_STD
+    SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * N);
+    SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N);
+#endif
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
+                                                                          GammaDataType,
+                                                                          BetaDataType,
+                                                                          YDataType,
+                                                                          SaveMeanInvStdDataType,
+                                                                          PassThrough,
+                                                                          Rank,
+                                                                          NumReduceDim>;
+
+    // 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;
+    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({N, H, W, C},         // lengths
+                                        strideXY,             // xStrides
+                                        strideGammaBeta,      // gammaStrides
+                                        strideGammaBeta,      // betaStrides
+                                        strideXY,             // yStrides
+                                        strideSaveMeanInvStd, // save_mean Strides
+                                        strideSaveMeanInvStd, // save_inv_std Strides
+                                        {1, 2, 3},            // reduceDims
+                                        1e-4,
+                                        x_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer(),
+                                        y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                        save_mean_device_buf.GetDeviceBuffer(),
+                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                        nullptr,
+                                        nullptr,
+#endif
+                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(XDataType) * N * H * W * C + sizeof(GammaDataType) * H * W * C +
+                sizeof(BetaDataType) * H * W * C + sizeof(YDataType) * N * H * W * C;
+
+#ifdef SAVE_MEAN_INV_STD
+            num_byte += sizeof(SaveMeanInvStdDataType) * N * 2;
+#endif
+
+            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({N, H, W, C},         // lengths
+                                        strideXY,             // xStrides
+                                        strideGammaBeta,      // gammaStrides
+                                        strideGammaBeta,      // betaStrides
+                                        strideXY,             // yStrides
+                                        strideSaveMeanInvStd, // save_mean Strides
+                                        strideSaveMeanInvStd, // save_inv_std Strides
+                                        {1, 2, 3},            // reduceDims
+                                        1e-4,
+                                        x_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer(),
+                                        y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                        save_mean_device_buf.GetDeviceBuffer(),
+                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                        nullptr,
+                                        nullptr,
+#endif
+                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/06_softmax/CMakeLists.txt

@@ -0,0 +1,2 @@
+add_executable(client_softmax4d softmax4d.cpp)
+target_link_libraries(client_softmax4d PRIVATE composable_kernel::device_other_operations composable_kernel::device_reduction_operations)

client_example/06_softmax/softmax4d.cpp

@@ -0,0 +1,169 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/softmax.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 2;
+
+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(int argc, char* argv[])
+{
+    std::vector<ck::index_t> in_lengths{2, 8, 128, 1024};
+    std::vector<ck::index_t> in_strides{8 * 128 * 1024, 128 * 1024, 1024, 1};
+    std::vector<ck::index_t> reduce_dims{2, 3};
+
+    ck::index_t num_elements =
+        std::accumulate(in_lengths.begin(), in_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    double alpha{2.0};
+    double beta{2.0};
+
+    SimpleDeviceMem in(sizeof(InDataType) * num_elements);
+    SimpleDeviceMem out(sizeof(OutDataType) * num_elements);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceSoftmax<InDataType,
+                                                                 AccDataType,
+                                                                 OutDataType,
+                                                                 PassThrough,
+                                                                 PassThrough,
+                                                                 Rank,
+                                                                 NumReduceDim>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    auto& generic_op_ptr = op_ptrs[0];
+
+    auto generic_argument_ptr = generic_op_ptr->MakeArgumentPointer(in_lengths,
+                                                                    in_strides,
+                                                                    reduce_dims,
+                                                                    alpha,
+                                                                    beta,
+                                                                    in.GetDeviceBuffer(),
+                                                                    out.GetDeviceBuffer(),
+                                                                    PassThrough{},
+                                                                    PassThrough{});
+
+    if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The generic kernel instance should be able to support any input shapes");
+    };
+
+    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(in_lengths,
+                                                        in_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        PassThrough{},
+                                                        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_elements * sizeof(InDataType) +
+                                    (beta == 0.0f ? 1 : 2) * num_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)
+            {
+                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(in_lengths,
+                                                        in_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        PassThrough{},
+                                                        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/07_grouped_convnd_fwd/CMakeLists.txt

@@ -0,0 +1,28 @@
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp)
+    target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_conv_operations)
+
+    add_executable(client_grouped_conv1d_fwd grouped_conv1d_fwd.cpp)
+    target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_conv_operations)
+
+    if((DTYPES MATCHES "fp8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
+        add_executable(client_grouped_conv3d_fwd_fp8 grouped_conv3d_fwd_fp8.cpp)
+        target_link_libraries(client_grouped_conv3d_fwd_fp8 PRIVATE composable_kernel::device_conv_operations)
+    endif()
+
+    if((DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
+        add_executable(client_grouped_conv3d_fwd_bf8 grouped_conv3d_fwd_bf8.cpp)
+        target_link_libraries(client_grouped_conv3d_fwd_bf8 PRIVATE composable_kernel::device_conv_operations)
+    endif()
+
+    if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
+        add_executable(client_grouped_conv3d_fwd_fp8_bf8 grouped_conv3d_fwd_fp8_bf8.cpp)
+        target_link_libraries(client_grouped_conv3d_fwd_fp8_bf8 PRIVATE composable_kernel::device_conv_operations)
+
+        add_executable(client_grouped_conv3d_fwd_bf8_fp8 grouped_conv3d_fwd_bf8_fp8.cpp)
+        target_link_libraries(client_grouped_conv3d_fwd_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
+    endif()
+
+    add_executable(grouped_conv2d_fwd_ngchw grouped_conv2d_fwd_ngchw.cpp)
+    target_link_libraries(grouped_conv2d_fwd_ngchw PRIVATE composable_kernel::device_conv_operations)
+endif()

client_example/07_grouped_convnd_fwd/README.md

@@ -0,0 +1,68 @@
+[Back to supported operations](../../../include/ck/README.md)
+# Composable Kernel Grouped Convolution
+
+## Grouped Convolution Forward
+Grouped convolution operation for 1D, 2D or 3D spatial dimensions. Convolution utilizes GEMM kernel after tensor coordinate transform. In CK Grouped Convolution Forward operation is called as `DeviceGroupedConvFwdMultipleABD` and requires following types as template parameters:
+
+* **NumDimSpatial** - number of spatial dimensions (1D, 2D, 3D).
+* **InLayout** - input layout (NHWGC, GNHWC, NGCHW).
+* **WeiLayout** - weight layout (GKYXC).
+* **DsLayout** - layouts for additional tensors for fused operations.
+* **OutLayout** - output layout (NHWGK, GNHWK, NGKHW).
+* **ADataType** - input data type. Pass tuple if there is fused operation with input.
+* **BDataType** - weight data type. Pass tuple if there is fused operation with weight.
+* **DsDataType** - data types for additional tensors for fused operations.
+* **EDataType** - Output data type.
+* **AElementwiseOperation** - fused operation on tensor A (input).
+* **BElementwiseOperation** - fused operation on tensor B (weight).
+* **CDEElementwiseOperation** - fused operation on tensor C (output).
+* **AComputeType** - compute data type of tensor A for mfma instruction (ADataType by default).
+* **BComputeType** - compute data type of tensor B for mfma instruction (AComputeType by default).
+
+Grouped convolution forward support tensors larger than 2GB.
+
+List of the device operations for grouped convolution forward in CK:
+
+* **DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3** - Device operation with XDL instructions. Optimized for AMD Instinct MI300 series.
+* **DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle** - Device operation with XDL instructions and support of fused operations to input, weight and output.
+* **DeviceGroupedConvFwdMultipleD_Wmma_CShuffle** - Device operation with WMMA instructions.
+* **DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK** - Device operation with DL instructions.
+
+Table of supported cases by instance factory with XDL instruction:
+
+|       |NHWGC/GKYXC/NHWGK|NGCHW/GKYXC/NGKHW|NGCHW/GKCYX/NGKHW|GNHWC/GKYXC/GNHWK|
+|-------|---|---|---|---|
+|bf16 |2D, 3D|2D|2D|1D, 2D, 3D|
+|fp16 |2D, 3D|2D|2D|1D, 2D, 3D|
+|fp32 |2D, 3D|2D|2D|1D, 2D, 3D|
+|int8 |2D, 3D|2D|2D|1D, 3D|
+|fp8  |3D|✗|✗|✗|
+|bf8  |3D|✗|✗|✗|
+
+Table of supported cases by instance factory with WMMA instruction:
+
+|       |NHWGC/GKYXC/NHWGK|NGCHW/GKYXC/NGKHW|GNHWC/GKYXC/GNHWK|
+|-------|---|---|---|
+|fp16 |2D, 3D|✗|2D, 3D|
+|int8 |2D, 3D|✗|2D, 3D|
+
+Table of supported cases by instance factory with DL instruction:
+
+|       |NHWGC/GKYXC/NHWGK|NGCHW/GKYXC/NGKHW|GNHWC/GKYXC/GNHWK|
+|-------|---|---|---|
+|bf16 |✗|✗|2D|
+|fp16 |✗|✗|2D|
+|fp32 |✗|✗|2D|
+|int8 |✗|✗|2D|
+
+Table of supported cases by instance factory with fused elementwise operation:
+
+* **Dynamic elementwise operation** - 2D/3D, NHWGC, bf16/fp16/fp32/int8
+* **Bilinear** - 3D, NHWGC, bf16/fp16/fp32/int8
+* **ConvInvScale** - 3D, NHWGC, fp8
+* **ConvScale** - 3D, NHWGC, fp8/bf8
+* **ConvScale + Add** - 3D, NHWGC, fp8
+* **ConvScale + Relu** - 3D, NHWGC, fp8
+* **Scale** - 3D, NHWGC, bf16/fp16/fp32/int8
+* **Scale + Add (for A and B)** - 3D, NHWGC, bf16/fp16/fp32/int8
+* **Scale + Add + Scale + Add + Relu** - 3D, NHWGC, bf16/fp16/fp32/int8

client_example/07_grouped_convnd_fwd/common.hpp

@@ -0,0 +1,304 @@
+// 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.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;
+
+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)
+{
+    // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+    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 static_cast<std::size_t>(2) * G * N * K * C *
+           std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
+                           std::end(weights_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+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(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);
+
+    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 flop      = GetFlops<NumDimSpatial>(out_lengths, wei_lengths);
+    std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   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{},
+            PassThrough{});
+
+        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{},
+            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 true;
+}

client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp

@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::GNWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKXC;
+using OutLayout   = ck::tensor_layout::convolution::GNWK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 1;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+int main()
+{
+    return run_grouped_conv_fwd<NumDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InLayout,
+                                WeiLayout,
+                                OutLayout,
+                                3>({N, Wi, G, C}, {G, K, X, C}, {N, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp

@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::NHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256; // 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 Y             = 3;   // filter H
+static constexpr ck::index_t X             = 3;   // filter W
+static constexpr ck::index_t Hi            = 28;  // input H
+static constexpr ck::index_t Wi            = 28;  // input W
+static constexpr ck::index_t Ho            = 28;  // output H
+static constexpr ck::index_t Wo            = 28;  // output W
+
+int main()
+{
+    return run_grouped_conv_fwd<NumDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InLayout,
+                                WeiLayout,
+                                OutLayout,
+                                3>({N, Hi, Wi, G, C}, {G, K, Y, X, C}, {N, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd_ngchw.cpp

@@ -0,0 +1,216 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, 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.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::NGCHW;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::NGKHW;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+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 Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+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()
+{
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C * Hi * Wi, G * C * Hi * Wi, Hi * Wi, Wi, 1};
+    std::array<ck::index_t, 5> wei_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> wei_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{K * Ho * Wo, G * K * Ho * Wo, Ho * Wo, Wo, 1};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   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(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{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        // workspace_sz will be equal to 0 for other layout than NGCHW
+        const std::size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+        SimpleDeviceMem workspace_dev(workspace_sz);
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        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{},
+                                                        PassThrough{});
+
+        const std::size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+        SimpleDeviceMem workspace_dev(workspace_sz);
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.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_fwd(); }

client_example/07_grouped_convnd_fwd/grouped_conv3d_fwd_bf8.cpp

@@ -0,0 +1,46 @@
+// 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 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::bf8_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/07_grouped_convnd_fwd/grouped_conv3d_fwd_bf8_fp8.cpp

@@ -0,0 +1,50 @@
+// 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/07_grouped_convnd_fwd/grouped_conv3d_fwd_fp8.cpp

@@ -0,0 +1,46 @@
+// 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/07_grouped_convnd_fwd/grouped_conv3d_fwd_fp8_bf8.cpp

@@ -0,0 +1,50 @@
+// 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/08_fused_attention/CMakeLists.txt

@@ -0,0 +1,7 @@
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_fused_attention fused_attention.cpp)
+    target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_fused_attention_bias fused_attention_bias.cpp)
+    target_link_libraries(client_fused_attention_bias PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+endif()

client_example/08_fused_attention/fused_attention.cpp

@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+using ADataType   = ck::half_t;
+using B0DataType  = ck::half_t;
+using B1DataType  = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+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(int argc, char* argv[])
+{
+    int G0 = 48;
+    int G1 = 16;
+    int M  = 1024;
+    int N  = 1024;
+    int K  = 64;
+    int O  = 64;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          ADataType,
+                                                                          B0DataType,
+                                                                          B1DataType,
+                                                                          CDataType,
+                                                                          ck::Tuple<>,
+                                                                          ck::Tuple<>,
+                                                                          AElementOp,
+                                                                          B0ElementOp,
+                                                                          Acc0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CElementOp,
+                                                                          MaskingSpec>;
+
+    // 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_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op 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(a_device_buf.GetDeviceBuffer(),
+                                                        b0_device_buf.GetDeviceBuffer(),
+                                                        b1_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        {}, // p_acc0_biases
+                                                        {}, // p_acc1_biases
+                                                        a_gs_ms_ks_lengths,
+                                                        a_gs_ms_ks_strides,
+                                                        b0_gs_ns_ks_lengths,
+                                                        b0_gs_ns_ks_strides,
+                                                        b1_gs_os_ns_lengths,
+                                                        b1_gs_os_ns_strides,
+                                                        c_gs_ms_os_lengths,
+                                                        c_gs_ms_os_strides,
+                                                        {}, // acc0_biases_gs_ms_ns_lengths
+                                                        {}, // acc0_biases_gs_ms_ns_strides
+                                                        {}, // acc1_biases_gs_ms_os_lengths
+                                                        {}, // acc1_biases_gs_ms_os_strides
+                                                        AElementOp{},
+                                                        B0ElementOp{},
+                                                        Acc0ElementOp{1 / sqrtf(K)},
+                                                        B1ElementOp{},
+                                                        CElementOp{});
+
+        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      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    G0 * G1;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                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 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(a_device_buf.GetDeviceBuffer(),
+                                                        b0_device_buf.GetDeviceBuffer(),
+                                                        b1_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        {}, // p_acc0_biases
+                                                        {}, // p_acc1_biases
+                                                        a_gs_ms_ks_lengths,
+                                                        a_gs_ms_ks_strides,
+                                                        b0_gs_ns_ks_lengths,
+                                                        b0_gs_ns_ks_strides,
+                                                        b1_gs_os_ns_lengths,
+                                                        b1_gs_os_ns_strides,
+                                                        c_gs_ms_os_lengths,
+                                                        c_gs_ms_os_strides,
+                                                        {}, // acc0_biases_gs_ms_ns_lengths
+                                                        {}, // acc0_biases_gs_ms_ns_strides
+                                                        {}, // acc1_biases_gs_ms_os_lengths
+                                                        {}, // acc1_biases_gs_ms_os_strides
+                                                        AElementOp{},
+                                                        B0ElementOp{},
+                                                        Acc0ElementOp{1 / sqrtf(K)},
+                                                        B1ElementOp{},
+                                                        CElementOp{});
+
+        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/08_fused_attention/fused_attention_bias.cpp

@@ -0,0 +1,226 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+using ADataType   = ck::half_t;
+using B0DataType  = ck::half_t;
+using B1DataType  = ck::half_t;
+using CDataType   = ck::half_t;
+using D0DataType  = ck::half_t;
+using AccDataType = float;
+
+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(int argc, char* argv[])
+{
+    int G0 = 48;
+    int G1 = 16;
+    int M  = 1024;
+    int N  = 1024;
+    int K  = 64;
+    int O  = 64;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) * G0 * G1 * M * N);
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          ADataType,
+                                                                          B0DataType,
+                                                                          B1DataType,
+                                                                          CDataType,
+                                                                          ck::Tuple<D0DataType>,
+                                                                          ck::Tuple<>,
+                                                                          AElementOp,
+                                                                          B0ElementOp,
+                                                                          Acc0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CElementOp,
+                                                                          MaskingSpec>;
+
+    // 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_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op 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(
+            a_device_buf.GetDeviceBuffer(),
+            b0_device_buf.GetDeviceBuffer(),
+            b1_device_buf.GetDeviceBuffer(),
+            c_device_buf.GetDeviceBuffer(),
+            std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+            {},                                                    // p_acc1_biases
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+            {},                       // acc1_biases_gs_ms_os_lengths
+            {},                       // acc1_biases_gs_ms_os_strides
+            AElementOp{},
+            B0ElementOp{},
+            Acc0ElementOp{1 / sqrtf(K)},
+            B1ElementOp{},
+            CElementOp{});
+
+        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      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O +
+                                     sizeof(D0DataType) * M * N) *
+                                    G0 * G1;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                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 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(
+            a_device_buf.GetDeviceBuffer(),
+            b0_device_buf.GetDeviceBuffer(),
+            b1_device_buf.GetDeviceBuffer(),
+            c_device_buf.GetDeviceBuffer(),
+            std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+            {},                                                    // p_acc1_biases
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+            {},                       // acc1_biases_gs_ms_os_lengths
+            {},                       // acc1_biases_gs_ms_os_strides
+            AElementOp{},
+            B0ElementOp{},
+            Acc0ElementOp{1 / sqrtf(K)},
+            B1ElementOp{},
+            CElementOp{});
+
+        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/09_quantization/CMakeLists.txt

@@ -0,0 +1,22 @@
+if(GPU_TARGETS MATCHES "gfx9" AND (DTYPES MATCHES "int8" OR NOT DEFINED DTYPES))
+    add_executable(client_conv2d_fwd_bias_tanh_perchannel_quantization conv2d_fwd_bias_tanh_perchannel_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_bias_tanh_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_conv2d_fwd_bias_tanh_perlayer_quantization conv2d_fwd_bias_tanh_perlayer_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_bias_tanh_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+
+    add_executable(client_gemm_quantization gemm_quantization.cpp)
+    target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+endif()

client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp

@@ -0,0 +1,210 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout         = ck::tensor_layout::convolution::G_K;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::NHWGK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;  // batch size
+static constexpr ck::index_t K             = 32; // output channel
+static constexpr ck::index_t C             = 64; // input channel (per group)
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Hi            = 71; // input H
+static constexpr ck::index_t Wi            = 71; // input W
+static constexpr ck::index_t Ho            = 36; // output H
+static constexpr ck::index_t Wo            = 36; // 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 main(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
+        NumDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        PassThrough,
+        PassThrough,
+        OutElementOp>;
+    // 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(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{ActivationOp{}});
+
+        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 = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(BiasDataType) * K + G * sizeof(RequantScaleDataType) * K +
+                G * sizeof(OutDataType) * N * 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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        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;
+
+        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(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{ActivationOp{}});
+
+        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/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp

@@ -0,0 +1,206 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType   = int8_t;
+using WeiDataType  = int8_t;
+using BiasDataType = int32_t;
+using OutDataType  = int8_t;
+
+using InLayout     = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout    = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout   = ck::tensor_layout::convolution::G_K;
+using OutLayout    = ck::tensor_layout::convolution::NHWGK;
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;    // batch size
+static constexpr ck::index_t K             = 32;   // output channel
+static constexpr ck::index_t C             = 64;   // input channel (per group)
+static constexpr ck::index_t Y             = 3;    // filter H
+static constexpr ck::index_t X             = 3;    // filter W
+static constexpr ck::index_t Hi            = 71;   // input H
+static constexpr ck::index_t Wi            = 71;   // input W
+static constexpr ck::index_t Ho            = 36;   // output H
+static constexpr ck::index_t Wo            = 36;   // output W
+static constexpr float requant_scale       = 0.5f; // requantize qAcc to qz
+
+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(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                      InLayout,
+                                                                      WeiLayout,
+                                                                      ck::Tuple<BiasLayout>,
+                                                                      OutLayout,
+                                                                      InDataType,
+                                                                      WeiDataType,
+                                                                      ck::Tuple<BiasDataType>,
+                                                                      OutDataType,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      OutElementOp>;
+    // 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(),
+                                        {bias.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths},
+                                        {bias_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{requant_scale, ActivationOp{}});
+
+        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 = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(BiasDataType) * K + G * sizeof(OutDataType) * N * 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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        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;
+
+        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(),
+                                        {bias.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths},
+                                        {bias_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{requant_scale, ActivationOp{}});
+
+        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/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp

@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout         = ck::tensor_layout::convolution::G_K;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::NHWGK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = ck::tensor_operation::element_wise::TanH;
+using OutElementOp =
+    ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;    // batch size
+static constexpr ck::index_t K             = 32;   // output channel
+static constexpr ck::index_t C             = 64;   // input channel (per group)
+static constexpr ck::index_t Y             = 3;    // filter H
+static constexpr ck::index_t X             = 3;    // filter W
+static constexpr ck::index_t Hi            = 71;   // input H
+static constexpr ck::index_t Wi            = 71;   // input W
+static constexpr ck::index_t Ho            = 36;   // output H
+static constexpr ck::index_t Wo            = 36;   // output W
+static constexpr float sz_inv              = 0.5f; // inverse of scale_z
+
+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(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
+        NumDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        PassThrough,
+        PassThrough,
+        OutElementOp>;
+    // 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(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{sz_inv, ActivationOp{}});
+
+        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 = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(BiasDataType) * K + G * sizeof(RequantScaleDataType) * K +
+                G * sizeof(OutDataType) * N * 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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        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;
+
+        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(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{sz_inv, ActivationOp{}});
+
+        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/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp

@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType   = int8_t;
+using WeiDataType  = int8_t;
+using BiasDataType = int32_t;
+using OutDataType  = int8_t;
+
+using InLayout     = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout    = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout   = ck::tensor_layout::convolution::G_K;
+using OutLayout    = ck::tensor_layout::convolution::NHWGK;
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::TanH;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;    // batch size
+static constexpr ck::index_t K             = 32;   // output channel
+static constexpr ck::index_t C             = 64;   // input channel (per group)
+static constexpr ck::index_t Y             = 3;    // filter H
+static constexpr ck::index_t X             = 3;    // filter W
+static constexpr ck::index_t Hi            = 71;   // input H
+static constexpr ck::index_t Wi            = 71;   // input W
+static constexpr ck::index_t Ho            = 36;   // output H
+static constexpr ck::index_t Wo            = 36;   // output W
+static constexpr float sacc                = 0.5f; //  scale of acc
+static constexpr float sz_inv              = 0.5f; // inverse of scale_z
+
+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(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                      InLayout,
+                                                                      WeiLayout,
+                                                                      ck::Tuple<BiasLayout>,
+                                                                      OutLayout,
+                                                                      InDataType,
+                                                                      WeiDataType,
+                                                                      ck::Tuple<BiasDataType>,
+                                                                      OutDataType,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      OutElementOp>;
+    // 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(),
+                                                        {bias.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {bias_lengths},
+                                                        {bias_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{sacc, sz_inv, ActivationOp{}});
+
+        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 = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(BiasDataType) * K + G * sizeof(OutDataType) * N * 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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        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;
+
+        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(),
+                                                        {bias.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {bias_lengths},
+                                                        {bias_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{sacc, sz_inv, ActivationOp{}});
+
+        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/09_quantization/conv2d_fwd_perchannel_quantization.cpp

@@ -0,0 +1,203 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::NHWGK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = PassThrough;
+using OutElementOp       = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;  // batch size
+static constexpr ck::index_t K             = 32; // output channel
+static constexpr ck::index_t C             = 64; // input channel (per group)
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Hi            = 71; // input H
+static constexpr ck::index_t Wi            = 71; // input W
+static constexpr ck::index_t Ho            = 36; // output H
+static constexpr ck::index_t Wo            = 36; // 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 main(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
+        NumDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<RequantScaleDataType>,
+        OutDataType,
+        PassThrough,
+        PassThrough,
+        OutElementOp>;
+    // 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(),
+                                                        {requant_scale.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {requant_scale_lengths},
+                                                        {requant_scale_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{ActivationOp{}});
+
+        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 = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(RequantScaleDataType) * K + G * sizeof(OutDataType) * N * 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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        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;
+
+        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,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{ActivationOp{}});
+
+        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/09_quantization/conv2d_fwd_perlayer_quantization.cpp

@@ -0,0 +1,199 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using OutDataType = int8_t;
+
+using InLayout     = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout    = ck::tensor_layout::convolution::GKYXC;
+using OutLayout    = ck::tensor_layout::convolution::NHWGK;
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;    // batch size
+static constexpr ck::index_t K             = 32;   // output channel
+static constexpr ck::index_t C             = 64;   // input channel (per group)
+static constexpr ck::index_t Y             = 3;    // filter H
+static constexpr ck::index_t X             = 3;    // filter W
+static constexpr ck::index_t Hi            = 71;   // input H
+static constexpr ck::index_t Wi            = 71;   // input W
+static constexpr ck::index_t Ho            = 36;   // output H
+static constexpr ck::index_t Wo            = 36;   // output W
+static constexpr float requant_scale       = 0.5f; // requantize qAcc to qY
+
+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(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   OutElementOp>;
+    // 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,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {},
+                                        {},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{requant_scale, ActivationOp{}});
+
+        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      = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
+                                    G * sizeof(WeiDataType) * K * Y * X * C +
+                                    G * sizeof(OutDataType) * N * 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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id != -1)
+    {
+        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;
+
+        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,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {},
+                                        {},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{requant_scale, ActivationOp{}});
+
+        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/09_quantization/gemm_quantization.cpp

@@ -0,0 +1,193 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp"
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using ActivationOp = PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+using ADataType = int8_t;
+using BDataType = int8_t;
+using EDataType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+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(int argc, char* argv[])
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    float requant_scale = 0.03;
+
+    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, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                                       BLayout,
+                                                                       ck::Tuple<>,
+                                                                       ELayout,
+                                                                       ADataType,
+                                                                       BDataType,
+                                                                       ck::Tuple<>,
+                                                                       EDataType,
+                                                                       AElementOp,
+                                                                       BElementOp,
+                                                                       CDEElementOp>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        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 avg_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_bytes =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id != -1)
+    {
+        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;
+
+        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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        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/10_grouped_convnd_bwd_data/CMakeLists.txt

@@ -0,0 +1,13 @@
+add_executable(client_grouped_conv2d_bwd_data grouped_conv2d_bwd_data.cpp)
+target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_conv2d_bwd_data_ngchw grouped_conv2d_bwd_data_ngchw.cpp)
+target_link_libraries(client_grouped_conv2d_bwd_data_ngchw PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_conv3d_bwd_data grouped_conv3d_bwd_data.cpp)
+target_link_libraries(client_grouped_conv3d_bwd_data PRIVATE composable_kernel::device_conv_operations)
+
+if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
+    add_executable(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp)
+    target_link_libraries(client_grouped_conv3d_bwd_data_input_fp16_comp_bf8f8 PRIVATE composable_kernel::device_conv_operations)
+endif()

client_example/10_grouped_convnd_bwd_data/README.md

@@ -0,0 +1,48 @@
+[Back to supported operations](../../../include/ck/README.md)
+# Composable Kernel Grouped Convolution
+
+## Grouped Convolution Backward Data
+
+Grouped convolution operation for 1D, 2D or 3D spatial dimensions. Convolution utilizes GEMM kernel after tensor coordinate transform. In CK Grouped Convolution Backward Data operation is called as `DeviceGroupedConvBwdDataMultipleD` and requires following types as template parameters:
+
+* **NumDimSpatial** - number of spatial dimensions (1D, 2D, 3D).
+* **ALayout** - output layout (NHWGK, GNHWK, NGKHW).
+* **BLayout** - weight layout (GKYXC).
+* **DsLayout** - layouts for additional tensors for fused operations.
+* **ELayout** - input layout (NHWGC, GNHWC, NGCHW).
+* **ADataType** - output data type.
+* **BDataType** - weight data type.
+* **DsDataType** - data types for additional tensors for fused operations.
+* **EDataType** - input data type.
+* **AElementwiseOperation** - fused operation on tensor A (output).
+* **BElementwiseOperation** - fused operation on tensor B (weight).
+* **CDEElementwiseOperation** - fused operation on tensor C (input).
+* **AComputeType** - compute data type of tensor A for mfma instruction (ADataType by default).
+* **BComputeType** - compute data type of tensor B for mfma instruction (AComputeType by default).
+
+Grouped convolution backward data supports tensors larger than 2GB (except when image is larger than 2GB).
+
+List of the device operations for grouped convolution backward data in CK:
+
+* **DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1** - Device operation with XDL instructions and support of fused operations to input.
+* **DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle** - Device operation with WMMA instructions.
+
+Table of supported cases by instance factory with XDL instruction:
+
+|       |NHWGC/GKYXC/NHWGK|NGCHW/GKYXC/NGKHW|GNHWC/GKYXC/GNHWK|
+|-------|---|---|---|
+|bf16|2D, 3D|2D, 3D|2D, 3D|
+|fp16 |2D, 3D|2D, 3D|2D, 3D|
+|fp32  |2D, 3D|2D, 3D|2D, 3D|
+
+Table of supported cases by instance factory with WMMA instruction:
+
+|       |NHWGC/GKYXC/NHWGK|NGCHW/GKYXC/NGKHW|GNHWC/GKYXC/GNHWK|
+|-------|---|---|---|
+|fp16 |2D, 3D|✗|2D, 3D|
+|int8 |2D, 3D|✗|2D, 3D|
+
+Table of supported cases by instance factory with fused elementwise operation:
+
+* **Bilinear** - 3D, NHWGC, bf16/fp16/fp32
+* **Scale** - 3D, NHWGC, bf16/fp16/fp32

client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp

@@ -0,0 +1,226 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.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::GNHWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::GNHWK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+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()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 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 GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
+    std::rotate(
+        rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
+    std::rotate(
+        rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
+    std::rotate(
+        rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
+    std::rotate(
+        rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
+    std::rotate(
+        rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
+    std::rotate(
+        rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                                                                     OutLayout,
+                                                                                     WeiLayout,
+                                                                                     ck::Tuple<>,
+                                                                                     InLayout,
+                                                                                     OutDataType,
+                                                                                     WeiDataType,
+                                                                                     ck::Tuple<>,
+                                                                                     InDataType,
+                                                                                     PassThrough,
+                                                                                     PassThrough,
+                                                                                     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(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{},
+                                                        PassThrough{});
+        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;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * G * N * 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{},
+                                                        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;
+    }
+}

client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data_ngchw.cpp

@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.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::NGCHW;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::NGKHW;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+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()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{
+        C * Hi * Wi, G * C * Hi * Wi, Wi, 1, Hi * Wi};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{K * Y * X * C, Y * X * C, X * C, C, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{
+        K * Ho * Wo, G * K * Ho * Wo, Wo, 1, Ho * Wo};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                                                                     OutLayout,
+                                                                                     WeiLayout,
+                                                                                     ck::Tuple<>,
+                                                                                     InLayout,
+                                                                                     OutDataType,
+                                                                                     WeiDataType,
+                                                                                     ck::Tuple<>,
+                                                                                     InDataType,
+                                                                                     PassThrough,
+                                                                                     PassThrough,
+                                                                                     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(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{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        const std::size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+        SimpleDeviceMem workspace_dev(workspace_sz);
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        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;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * G * N * 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{},
+                                                        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;
+    }
+}

client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp

@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.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;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 2;
+static constexpr ck::index_t N             = 16;
+static constexpr ck::index_t K             = 16;
+static constexpr ck::index_t C             = 16;
+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            = 14;
+static constexpr ck::index_t Hi            = 14;
+static constexpr ck::index_t Wi            = 14;
+static constexpr ck::index_t Do            = 14;
+static constexpr ck::index_t Ho            = 14;
+static constexpr ck::index_t Wo            = 14;
+
+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()
+{
+    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,
+                                                                                     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(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{},
+                                                        PassThrough{});
+        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;
+            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(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{},
+                                                        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;
+    }
+}

client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp

@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.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;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 2;
+static constexpr ck::index_t N             = 16;
+static constexpr ck::index_t K             = 16;
+static constexpr ck::index_t C             = 16;
+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            = 14;
+static constexpr ck::index_t Hi            = 14;
+static constexpr ck::index_t Wi            = 14;
+static constexpr ck::index_t Do            = 14;
+static constexpr ck::index_t Ho            = 14;
+static constexpr ck::index_t Wo            = 14;
+
+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()
+{
+    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,
+                                                                                     PassThrough,
+                                                                                     ck::bf8_t,
+                                                                                     ck::f8_t>;
+    // 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{},
+                                                        PassThrough{});
+        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;
+            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(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{},
+                                                        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;
+    }
+}

client_example/11_grouped_conv_bwd_weight/CMakeLists.txt

@@ -0,0 +1,14 @@
+add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_fp16.cpp)
+add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
+add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
+add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
+
+target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
+target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
+target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
+target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_conv_operations)
+
+if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
+    add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
+    target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)
+endif()

client_example/11_grouped_conv_bwd_weight/README.md

@@ -0,0 +1,62 @@
+[Back to supported operations](../../../include/ck/README.md)
+# Composable Kernel Grouped Convolution
+
+## Grouped Convolution Backward Weight
+
+Grouped convolution operation for 1D, 2D or 3D spatial dimensions. Convolution utilizes GEMM kernel after tensor coordinate transform. Backward weight version uses splitK feature (due to large GEMM K dimension). In CK Grouped Convolution Backward Weight operation is called as `DeviceGroupedConvBwdWeight` and requires following types as template parameters:
+
+* **NumDimSpatial** - number of spatial dimensions (1D, 2D, 3D).
+* **InLayout** - input layout (NHWGC, GNHWC, NGCHW).
+* **WeiLayout** - weight layout (GKYXC).
+* **OutLayout** - output layout (NHWGK, GNHWK, NGKHW).
+* **InDataType** - input data type.
+* **WeiDataType** - weight data type.
+* **OutDataType** - output data type.
+* **InElementwiseOperation** - fused operation on tensor input.
+* **WeiElementwiseOperation** - fused operation on tensor weight.
+* **OutElementwiseOperation** - fused operation on tensor output.
+* **ComputeTypeA** - compute data type of tensor A for mfma instruction (ADataType by default).
+* **ComputeTypeB** - compute data type of tensor B for mfma instruction (ComputeTypeA by default).
+
+For fused operations with additional tensor there is `DeviceGroupedConvBwdWeightMultipleD` operation which requires following parameters:
+* **DsLayout** - layouts for additional tensors for fused operations.
+* **DsDataType** - data types for additional tensors for fused operations.
+
+Grouped convolution backward weight doesn't supports tensors larger than 2GB.
+
+List of the device operations for grouped convolution backward weight in CK:
+
+* **DeviceGroupedConvBwdWeight_Xdl_CShuffle** - Device operation with XDL instructions.
+* **DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle** - Device operation with XDL instructions. Optimized for small C or K.
+* **DeviceGroupedConvBwdWeight_Wmma_CShuffle** - Device operation with WMMA instructions.
+* **DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle** - Device operation with XDL instructions and support of fused operations to output.
+* **DeviceGroupedConvBwdWeight_Dl** - Device operation with DL instructions.
+
+Table of supported cases by instance factory with XDL instruction:
+
+|       |NHWGC/GKYXC/NHWGK|NGCHW/GKYXC/NGKHW|NGCHW/GKCYX/NGKHW|GNHWC/GKYXC/GNHWK|
+|-------|---|---|---|---|
+|bf16|2D, 3D|2D, 3D|2D, 3D|✗|
+|bf16(fp32 for weight)|2D, 3D|✗|✗|1D, 2D, 3D|
+|fp16 |2D, 3D|2D, 3D|2D, 3D|1D, 2D, 3D|
+|fp32  |2D, 3D|2D, 3D|2D, 3D|1D, 2D, 3D|
+
+Table of supported cases by instance factory with WMMA instruction:
+
+|       |NHWGC/GKYXC/NHWGK|NGCHW/GKYXC/NGKHW|GNHWC/GKYXC/GNHWK|
+|-------|---|---|---|
+|fp16 |3D|✗|3D|
+|int8 |3D|✗|3D|
+
+Table of supported cases by instance factory with DL instruction:
+
+|       |NHWGC/GKYXC/NHWGK|NGCHW/GKYXC/NGKHW|GNHWC/GKYXC/GNHWK|
+|-------|---|---|---|
+|bf16(fp32 for weight)|1D, 2D, 3D|✗|1D, 2D, 3D|
+|fp16 |1D, 2D, 3D|✗|1D, 2D, 3D|
+|fp32  |1D, 2D, 3D|✗|1D, 2D, 3D|
+
+Table of supported cases by instance factory with fused elementwise operation:
+
+* **Bilinear** - 3D, NHWGC, bf16(fp32 for weight)/fp16/fp32
+* **Scale** - 3D, NHWGC, bf16(fp32 for weight)/fp16/fp32

client_example/11_grouped_conv_bwd_weight/common.hpp

@@ -0,0 +1,239 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+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_;
+};
+
+template <ck::index_t NumDimSpatial>
+std::size_t GetFlops(const std::array<ck::index_t, NumDimSpatial>& output_lengths,
+                     const std::array<ck::index_t, NumDimSpatial>& filter_lengths)
+{
+    constexpr ck::index_t spatial_offset = 3;
+    const auto C                         = filter_lengths[2];
+    // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+    return static_cast<std::size_t>(2) * C *
+           std::accumulate(std::begin(output_lengths),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           std::accumulate(std::begin(filter_lengths) + spatial_offset,
+                           std::end(filter_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename InDataType, ck::index_t NumDimSpatial>
+std::size_t GetInputByte(const std::array<ck::index_t, NumDimSpatial>& 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>
+std::size_t GetWeightByte(const std::array<ck::index_t, NumDimSpatial>& filter_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * (std::accumulate(std::begin(filter_lengths),
+                                                  std::end(filter_lengths),
+                                                  static_cast<std::size_t>(1),
+                                                  std::multiplies<>()));
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial>
+std::size_t GetOutputByte(const std::array<ck::index_t, NumDimSpatial>& 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,
+          typename AComputeType = InDataType,
+          typename BComputeType = AComputeType>
+bool run_grouped_conv_bwd_weight(
+    const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths,
+    const std::array<ck::index_t, NumDimSpatial + 3>& input_strides,
+    const std::array<ck::index_t, NumDimSpatial + 3>& filter_lengths,
+    const std::array<ck::index_t, NumDimSpatial + 3>& weights_strides,
+    const std::array<ck::index_t, NumDimSpatial + 3>& output_lengths,
+    const std::array<ck::index_t, NumDimSpatial + 3>& output_strides,
+    const std::array<ck::index_t, NumDimSpatial>& conv_filter_strides,
+    const std::array<ck::index_t, NumDimSpatial>& conv_filter_dilations,
+    const std::array<ck::index_t, NumDimSpatial>& input_left_pads,
+    const std::array<ck::index_t, NumDimSpatial>& input_right_pads)
+{
+
+    ck::index_t split_k = 2;
+    SimpleDeviceMem in(GetInputByte<InDataType, NumDimSpatial + 3>(input_lengths));
+    SimpleDeviceMem wei(GetWeightByte<WeiDataType, NumDimSpatial + 3>(filter_lengths));
+    SimpleDeviceMem out(GetOutputByte<OutDataType, NumDimSpatial + 3>(output_lengths));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NumDimSpatial,
+                                                                              InLayout,
+                                                                              WeiLayout,
+                                                                              OutLayout,
+                                                                              InDataType,
+                                                                              WeiDataType,
+                                                                              OutDataType,
+                                                                              PassThrough,
+                                                                              PassThrough,
+                                                                              PassThrough,
+                                                                              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;
+
+    std::array<ck::index_t, NumDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NumDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NumDimSpatial + 3> b_g_k_c_xs_lengths{};
+
+    // 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(),
+                                                        input_lengths,
+                                                        input_strides,
+                                                        filter_lengths,
+                                                        weights_strides,
+                                                        output_lengths,
+                                                        output_strides,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        split_k);
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        const std::size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+        SimpleDeviceMem workspace_dev(workspace_sz);
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = GetFlops<NumDimSpatial + 3>(output_lengths, filter_lengths);
+            std::size_t num_bytes = GetInputByte<InDataType, NumDimSpatial + 3>(input_lengths) +
+                                    GetWeightByte<WeiDataType, NumDimSpatial + 3>(filter_lengths) +
+                                    GetOutputByte<OutDataType, NumDimSpatial + 3>(output_lengths);
+
+            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(),
+                                                        out.GetDeviceBuffer(),
+                                                        input_lengths,
+                                                        input_strides,
+                                                        filter_lengths,
+                                                        weights_strides,
+                                                        output_lengths,
+                                                        output_strides,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        split_k);
+        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/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp

@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::GNWC;
+using WeiLayout = ck::tensor_layout::convolution::GKXC;
+using OutLayout = ck::tensor_layout::convolution::GNWK;
+
+static constexpr ck::index_t NumDimSpatial = 1;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Wo            = 28;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{N * Wi * C, Wi* C, 1, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{K * X * C, X* C, 1, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{N * Wo * K, Wo* K, 1, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout>(input_lengths,
+                                                  input_strides,
+                                                  filter_lengths,
+                                                  weights_strides,
+                                                  output_lengths,
+                                                  output_strides,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp

@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+using OutLayout = ck::tensor_layout::convolution::GNHWK;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Hi, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Y, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Ho, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+    N * Hi * Wi * C, Hi* Wi* C, 1, Wi* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
+    K * Y * X * C, Y* X* C, 1, X* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+    N * Ho * Wo * K, Ho* Wo* K, 1, Wo* K, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout>(input_lengths,
+                                                  input_strides,
+                                                  filter_lengths,
+                                                  weights_strides,
+                                                  output_lengths,
+                                                  output_strides,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp

@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::GNDHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::GNDHWK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 8;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 128;
+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;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+    N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
+static constexpr 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};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+    N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout>(input_lengths,
+                                                  input_strides,
+                                                  filter_lengths,
+                                                  weights_strides,
+                                                  output_lengths,
+                                                  output_strides,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp

@@ -0,0 +1,70 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::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 AComputeType = ck::bf8_t;
+using BComputeType = ck::f8_t;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 8;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 128;
+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;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+    N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
+static constexpr 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};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+    N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout,
+                                       AComputeType,
+                                       BComputeType>(input_lengths,
+                                                     input_strides,
+                                                     filter_lengths,
+                                                     weights_strides,
+                                                     output_lengths,
+                                                     output_strides,
+                                                     conv_filter_strides,
+                                                     conv_filter_dilations,
+                                                     input_left_pads,
+                                                     input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp

@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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  = float;
+using WeiDataType = float;
+using OutDataType = float;
+
+using InLayout  = ck::tensor_layout::convolution::GNDHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::GNDHWK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 8;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 128;
+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;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+    N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
+static constexpr 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};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+    N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout>(input_lengths,
+                                                  input_strides,
+                                                  filter_lengths,
+                                                  weights_strides,
+                                                  output_lengths,
+                                                  output_strides,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/12_elementwise_normalization/CMakeLists.txt

@@ -0,0 +1,2 @@
+add_executable(client_elementwise_layernorm2d elementwise_layernorm2d.cpp)
+target_link_libraries(client_elementwise_layernorm2d PRIVATE composable_kernel::device_other_operations)

client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp

@@ -0,0 +1,176 @@
+// 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/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp"
+
+using ADataType             = ck::half_t; // Input 1
+using BDataType             = ck::half_t; // Input 2
+using XDataType             = ck::half_t;
+using GammaDataType         = ck::half_t;
+using BetaDataType          = ck::half_t;
+using YDataType             = ck::half_t;
+using AccDataType           = float;
+using XElementwiseOperation = ck::tensor_operation::element_wise::Add;
+using YElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+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()
+{
+    bool time_kernel = true;
+
+    ck::index_t M      = 48 * 256;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    auto mn_size = (M - 1) * Stride + N;
+
+    SimpleDeviceMem a_dev_buf(sizeof(ADataType) * mn_size);
+    SimpleDeviceMem b_dev_buf(sizeof(BDataType) * mn_size);
+    SimpleDeviceMem gamma_dev_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_dev_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem y_dev_buf(sizeof(YDataType) * mn_size);
+
+    std::array<const void*, 2> ab_input               = {a_dev_buf.GetDeviceBuffer(),
+                                           b_dev_buf.GetDeviceBuffer()};
+    std::vector<ck::index_t> abStride                 = {Stride, 1};
+    std::array<std::vector<ck::index_t>, 2> abStrides = {abStride, abStride};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwiseNormalization<
+        ck::Tuple<ADataType, BDataType>,
+        GammaDataType,
+        BetaDataType,
+        AccDataType,
+        YDataType,
+        XElementwiseOperation,
+        YElementwiseOperation,
+        Rank,
+        NumReduceDim>;
+
+    // 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;
+    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({M, N}, // lengths
+                                                        abStrides,
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        ab_input,
+                                                        gamma_dev_buf.GetDeviceBuffer(),
+                                                        beta_dev_buf.GetDeviceBuffer(),
+                                                        y_dev_buf.GetDeviceBuffer(),
+                                                        XElementwiseOperation{},
+                                                        YElementwiseOperation{});
+
+        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) * M * N + sizeof(BDataType) * M * N +
+                                   sizeof(GammaDataType) * N + sizeof(BetaDataType) * N +
+                                   sizeof(YDataType) * M * N;
+
+            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({M, N}, // lengths
+                                                        abStrides,
+                                                        {1},         // gammaStrides
+                                                        {1},         // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        ab_input,
+                                                        gamma_dev_buf.GetDeviceBuffer(),
+                                                        beta_dev_buf.GetDeviceBuffer(),
+                                                        y_dev_buf.GetDeviceBuffer(),
+                                                        XElementwiseOperation{},
+                                                        YElementwiseOperation{});
+
+        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/13_batchnorm/CMakeLists.txt

@@ -0,0 +1,6 @@
+add_executable(client_batchnorm_fwd_nhwc batchnorm_fwd_nhwc.cpp)
+add_executable(client_batchnorm_bwd_nhwc batchnorm_bwd_nhwc.cpp)
+add_executable(client_batchnorm_infer_nhwc batchnorm_infer_nhwc.cpp)
+target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_other_operations)
+target_link_libraries(client_batchnorm_bwd_nhwc PRIVATE composable_kernel::device_other_operations)
+target_link_libraries(client_batchnorm_infer_nhwc PRIVATE composable_kernel::device_other_operations)

client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp

@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp"
+
+using XDataType           = ck::half_t;
+using DxDataType          = float;
+using DyDataType          = float;
+using AccDataType         = float;
+using ScaleDataType       = ck::half_t;
+using DscaleDbiasDataType = float;
+using MeanVarDataType     = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon = std::numeric_limits<float>::epsilon();
+
+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(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem dy(sizeof(DyDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem dx(sizeof(DxDataType) * numXYElement);
+    SimpleDeviceMem dscale(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem dbias(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                                      DxDataType,
+                                                                      DyDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      DscaleDbiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    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   = 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(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        dy.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        dx.GetDeviceBuffer(),
+                                                        dscale.GetDeviceBuffer(),
+                                                        dbias.GetDeviceBuffer());
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(DyDataType) + sizeof(DxDataType)) +
+                numScaleBiasMeanVarElement *
+                    (sizeof(ScaleDataType) + sizeof(DscaleDbiasDataType) * 2 +
+                     sizeof(MeanVarDataType) * 2);
+
+            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)
+            {
+                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;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << 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(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        dy.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        dx.GetDeviceBuffer(),
+                                                        dscale.GetDeviceBuffer(),
+                                                        dbias.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp

@@ -0,0 +1,203 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using AccDataType     = float;
+using ScaleDataType   = AccDataType;
+using BiasDataType    = AccDataType;
+using MeanVarDataType = AccDataType;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon       = std::numeric_limits<float>::epsilon();
+const double averageFactor = 0.1;
+
+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(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    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   = 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(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                numScaleBiasMeanVarElement * (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                              sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+            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)
+            {
+                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;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << 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(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/13_batchnorm/batchnorm_infer_nhwc.cpp

@@ -0,0 +1,189 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using ScaleDataType   = float;
+using BiasDataType    = float;
+using MeanVarDataType = float;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+using Normalize = ck::tensor_operation::element_wise::NormalizeInInfer;
+
+const double epsilon = std::numeric_limits<float>::epsilon();
+
+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(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+    std::array<int, Rank - NumBatchNormReduceDim> invariantDims{3};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem variance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    // values in variance need be non-negative
+    (void)hipMemset(
+        variance.GetDeviceBuffer(), 0, sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    std::array<ck::index_t, Rank> aligned_scaleBiasMeanVarStrides{0};
+
+    int i = 0;
+    for(auto dim : invariantDims)
+    {
+        assert(xyLengths[dim] == scaleBiasMeanVarLengths[i]);
+
+        aligned_scaleBiasMeanVarStrides[dim] = scaleBiasMeanVarStrides[i];
+        i++;
+    };
+
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+        ck::Tuple<YDataType>,
+        Normalize,
+        Rank>;
+
+    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   = 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(xyLengths,
+                                                        {xyStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides},
+                                                        {xyStrides},
+                                                        {x.GetDeviceBuffer(),
+                                                         mean.GetDeviceBuffer(),
+                                                         variance.GetDeviceBuffer(),
+                                                         scale.GetDeviceBuffer(),
+                                                         bias.GetDeviceBuffer()},
+                                                        {y.GetDeviceBuffer()},
+                                                        Normalize{epsilon});
+
+        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 =
+                numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                numScaleBiasMeanVarElement * (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                              sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+            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)
+            {
+                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;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << 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(xyLengths,
+                                                        {xyStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides},
+                                                        {xyStrides},
+                                                        {x.GetDeviceBuffer(),
+                                                         mean.GetDeviceBuffer(),
+                                                         variance.GetDeviceBuffer(),
+                                                         scale.GetDeviceBuffer(),
+                                                         bias.GetDeviceBuffer()},
+                                                        {y.GetDeviceBuffer()},
+                                                        Normalize{epsilon});
+
+        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/14_instance_id/CMakeLists.txt

@@ -0,0 +1,2 @@
+add_executable(client_batchnorm_fwd_instance_id batchnorm_fwd_instance_id.cpp)
+target_link_libraries(client_batchnorm_fwd_instance_id PRIVATE composable_kernel::device_other_operations)

client_example/14_instance_id/batchnorm_fwd_instance_id.cpp

@@ -0,0 +1,206 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using AccDataType     = float;
+using ScaleDataType   = AccDataType;
+using BiasDataType    = AccDataType;
+using MeanVarDataType = AccDataType;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon       = std::numeric_limits<float>::epsilon();
+const double averageFactor = 0.1;
+
+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_;
+};
+
+// In the actual application, the instance index and name are usually from the perf db
+static int instance_index = -1;
+static std::string instance_name;
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    bool found          = false;
+    int best_op_index   = -1;
+    float best_ave_time = std::numeric_limits<float>::max();
+
+    // profile device operation instances and save the best performant instance index and instance
+    // name
+    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(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            if(ave_time < best_ave_time)
+            {
+                found         = true;
+                best_op_index = i;
+                best_ave_time = ave_time;
+            }
+        }
+    }
+
+    if(found)
+    {
+        instance_index = best_op_index;
+        instance_name  = op_ptrs[instance_index]->GetTypeIdHashCode();
+    };
+
+    // simulate the execution of the operation when the instance index and name are available
+    const auto op_ptrs_2 = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(instance_index >= 0 && instance_index < op_ptrs_2.size())
+    {
+        auto& op_ptr = op_ptrs_2[instance_index];
+
+        if(op_ptr->GetTypeIdHashCode() == instance_name)
+        {
+
+            auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                            xyStrides,
+                                                            xyStrides,
+                                                            reduceDims,
+                                                            scaleBiasMeanVarLengths,
+                                                            scaleBiasMeanVarStrides,
+                                                            scaleBiasMeanVarStrides,
+                                                            scaleBiasMeanVarStrides,
+                                                            x.GetDeviceBuffer(),
+                                                            scale.GetDeviceBuffer(),
+                                                            bias.GetDeviceBuffer(),
+                                                            epsilon,
+                                                            PassThrough{},
+                                                            y.GetDeviceBuffer(),
+                                                            mean.GetDeviceBuffer(),
+                                                            invVariance.GetDeviceBuffer(),
+                                                            averageFactor,
+                                                            nullptr,
+                                                            nullptr);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+                SimpleDeviceMem workspace(workspace_sz);
+
+                op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+                float exec_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+                size_t num_bytes = numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                                   numScaleBiasMeanVarElement *
+                                       (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                        sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+                float gb_per_sec = num_bytes / 1.E6 / exec_time;
+
+                std::cout << "Kernel execution time: " << std::setw(10) << exec_time
+                          << " ms,  effective data transfer bandwidth: " << gb_per_sec << " GB/s"
+                          << std::endl;
+            }
+        };
+    }
+
+    return 0;
+}

client_example/15_convnd_bwd_data/CMakeLists.txt

@@ -0,0 +1,7 @@
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_conv3d_bwd_data_fp16 conv3d_bwd_data_fp16.cpp)
+    add_executable(client_conv3d_bwd_data_fp32 conv3d_bwd_data_fp32.cpp)
+
+    target_link_libraries(client_conv3d_bwd_data_fp16 PRIVATE composable_kernel::device_conv_operations)
+    target_link_libraries(client_conv3d_bwd_data_fp32 PRIVATE composable_kernel::device_conv_operations)
+endif()

client_example/15_convnd_bwd_data/common.hpp

@@ -0,0 +1,233 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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/convolution_backward_data.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+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_;
+};
+
+std::size_t GetFlops(ck::index_t N,
+                     ck::index_t K,
+                     ck::index_t C,
+                     const std::vector<ck::index_t>& output_spatial_lengths,
+                     const std::vector<ck::index_t>& weights_spatial_lengths)
+{
+    // 2 * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+
+    return static_cast<std::size_t>(2) * N * K * C *
+           std::accumulate(std::begin(output_spatial_lengths),
+                           std::end(output_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           std::accumulate(std::begin(weights_spatial_lengths),
+                           std::end(weights_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename InDataType>
+std::size_t
+GetInputByte(ck::index_t N, ck::index_t C, const std::vector<ck::index_t>& input_spatial_lengths)
+{
+    // sizeof(InDataType) * (N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * N * C *
+           std::accumulate(std::begin(input_spatial_lengths),
+                           std::end(input_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename WeiDataType>
+std::size_t
+GetWeightByte(ck::index_t K, ck::index_t C, const std::vector<ck::index_t>& weights_spatial_lengths)
+{
+    // sizeof(WeiDataType) * (K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * K * C *
+           std::accumulate(std::begin(weights_spatial_lengths),
+                           std::end(weights_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename OutDataType>
+std::size_t
+GetOutputByte(ck::index_t N, ck::index_t K, const std::vector<ck::index_t>& output_spatial_lengths)
+{
+    // sizeof(OutDataType) * (N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * N * K *
+           std::accumulate(std::begin(output_spatial_lengths),
+                           std::end(output_spatial_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>
+bool run_conv_bwd_data(ck::index_t N,
+                       ck::index_t K,
+                       ck::index_t C,
+                       const std::vector<ck::index_t>& in_spatial_lengths,
+                       const std::vector<ck::index_t>& wei_spatial_lengths,
+                       const std::vector<ck::index_t>& out_spatial_lengths)
+{
+    std::size_t in_mem_size  = GetInputByte<InDataType>(N, C, in_spatial_lengths);
+    std::size_t wei_mem_size = GetWeightByte<WeiDataType>(K, C, wei_spatial_lengths);
+    std::size_t out_mem_size = GetOutputByte<OutDataType>(N, K, out_spatial_lengths);
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+
+    std::vector<ck::index_t> filter_strides(NumDimSpatial, 1);
+    std::vector<ck::index_t> filter_dilations(NumDimSpatial, 1);
+    std::vector<ck::index_t> input_left_pads(NumDimSpatial, 1);
+    std::vector<ck::index_t> input_right_pads(NumDimSpatial, 1);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvBwdData<NumDimSpatial,
+                                                                     InLayout,
+                                                                     WeiLayout,
+                                                                     OutLayout,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     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;
+
+    std::size_t flop      = GetFlops(N, K, C, out_spatial_lengths, wei_spatial_lengths);
+    std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
+
+    // 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(),
+                                                        N,
+                                                        K,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+        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(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        K,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        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 true;
+}

client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp

@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using WeiLayout = ck::tensor_layout::convolution::KZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+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            = 28;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+int main()
+{
+    return run_conv_bwd_data<NumDimSpatial,
+                             InDataType,
+                             WeiDataType,
+                             OutDataType,
+                             InLayout,
+                             WeiLayout,
+                             OutLayout>(N, K, C, {Di, Hi, Wi}, {Z, Y, X}, {Do, Ho, Wo})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp

@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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  = float;
+using WeiDataType = float;
+using OutDataType = float;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using WeiLayout = ck::tensor_layout::convolution::KZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+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            = 28;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+int main()
+{
+    return run_conv_bwd_data<NumDimSpatial,
+                             InDataType,
+                             WeiDataType,
+                             OutDataType,
+                             InLayout,
+                             WeiLayout,
+                             OutLayout>(N, K, C, {Di, Hi, Wi}, {Z, Y, X}, {Do, Ho, Wo})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/15_gemm_add_multiply/CMakeLists.txt

@@ -0,0 +1,4 @@
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_gemm_add_multiply gemm_add_multiply.cpp)
+    target_link_libraries(client_gemm_add_multiply PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/15_gemm_add_multiply/gemm_add_multiply.cpp

@@ -0,0 +1,242 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        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]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, 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, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                          BLayout,
+                                                          ck::Tuple<D0Layout, D1Layout>,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          ck::Tuple<D0DataType, D1DataType>,
+                                                          EDataType,
+                                                          AElementOp,
+                                                          BElementOp,
+                                                          CDEElementOp>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            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(ADataType) * M * K + sizeof(BDataType) * 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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            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/15_reduce/CMakeLists.txt

@@ -0,0 +1,2 @@
+add_executable(client_reduce_nhwc_c reduce_nhwc_c.cpp)
+target_link_libraries(client_reduce_nhwc_c PRIVATE composable_kernel::device_reduction_operations)

client_example/15_reduce/reduce_nhwc_c.cpp

@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp"
+
+using InDataType  = float;
+using OutDataType = float;
+using AccDataType = float;
+using ReduceAdd   = ck::reduce::Add;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using UnaryDivide = ck::tensor_operation::element_wise::UnaryDivide;
+
+constexpr bool PropagateNan = false;
+constexpr bool OutputIndex  = false;
+
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+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(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> in_lengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> in_strides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumReduceDim> out_lengths{256};
+    std::array<ck::index_t, Rank - NumReduceDim> out_strides{1};
+    std::array<int, NumReduceDim> reduce_dims{0, 1, 2};
+
+    ck::index_t num_in_elements =
+        std::accumulate(in_lengths.begin(), in_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t num_out_elements =
+        std::accumulate(out_lengths.begin(), out_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t reduce_length = 1;
+
+    for(auto dim : reduce_dims)
+        reduce_length *= in_lengths[dim];
+
+    double alpha{1.0};
+    double beta{0.0};
+
+    SimpleDeviceMem in(sizeof(InDataType) * num_in_elements);
+    SimpleDeviceMem out(sizeof(OutDataType) * num_out_elements);
+
+    using DeviceOp     = ck::tensor_operation::device::DeviceReduce<InDataType,
+                                                                AccDataType,
+                                                                OutDataType,
+                                                                Rank,
+                                                                NumReduceDim,
+                                                                ReduceAdd,
+                                                                PassThrough,
+                                                                UnaryDivide,
+                                                                PropagateNan,
+                                                                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;
+    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(in_lengths,
+                                                        in_strides,
+                                                        out_lengths,
+                                                        out_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        out.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        PassThrough{},
+                                                        UnaryDivide{reduce_length});
+        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) +
+                                    (beta == 0.0f ? 1 : 2) * 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)
+            {
+                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(in_lengths,
+                                                        in_strides,
+                                                        out_lengths,
+                                                        out_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        out.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        PassThrough{},
+                                                        UnaryDivide{reduce_length});
+
+        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/16_convnd_fwd/CMakeLists.txt

@@ -0,0 +1,15 @@
+if((DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
+    add_executable(client_conv3d_fwd_fp16 conv3d_fwd_fp16.cpp)
+    target_link_libraries(client_conv3d_fwd_fp16 PRIVATE composable_kernel::device_conv_operations)
+
+endif()
+
+if((DTYPES MATCHES "fp8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
+    add_executable(client_conv3d_fwd_fp16_comp_fp8 conv3d_fwd_fp16_comp_fp8.cpp)
+    target_link_libraries(client_conv3d_fwd_fp16_comp_fp8 PRIVATE composable_kernel::device_conv_operations)
+endif()
+
+if((DTYPES MATCHES "fp32") OR NOT DEFINED DTYPES)
+    add_executable(client_conv3d_fwd_fp32 conv3d_fwd_fp32.cpp)
+    target_link_libraries(client_conv3d_fwd_fp32 PRIVATE composable_kernel::device_conv_operations)
+endif()

client_example/16_convnd_fwd/common.hpp

@@ -0,0 +1,304 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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.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;
+
+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)
+{
+    // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+    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 static_cast<std::size_t>(2) * G * N * K * C *
+           std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
+                           std::end(weights_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+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(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);
+
+    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 flop      = GetFlops<NumDimSpatial>(out_lengths, wei_lengths);
+    std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   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{},
+            PassThrough{});
+
+        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{},
+            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 true;
+}

client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp

@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::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;
+
+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>(
+               {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_fp16_comp_fp8.cpp

@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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::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;
+
+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_fp32.cpp

@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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  = float;
+using WeiDataType = float;
+using OutDataType = float;
+
+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>(
+               {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

@@ -0,0 +1,4 @@
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
+    target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp

@@ -0,0 +1,232 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using DsDataType = ck::Tuple<>;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = FastGelu;
+
+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()
+{
+    std::mt19937 gen(19391);
+    std::uniform_int_distribution<> distrib(1, 10);
+    int group_count = distrib(gen);
+
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ms.push_back(256 + 256 * distrib(gen));
+        Ns.push_back(256 + 256 * distrib(gen));
+        Ks.push_back(128 + 128 * distrib(gen));
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+    }
+
+    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, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_e;
+
+    p_a.reserve(group_count);
+    p_b.reserve(group_count);
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideEs[i], {}});
+
+        p_a.push_back(a_dev_bufs[i].GetDeviceBuffer());
+        p_b.push_back(b_dev_bufs[i].GetDeviceBuffer());
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
+                                                                     BLayout,
+                                                                     DsLayout,
+                                                                     ELayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     DsDataType,
+                                                                     EDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CDEElementOp>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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;
+
+    auto p_ds = std::vector<std::array<const void*, 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(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+        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 = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            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(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+
+        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/18_groupnorm/CMakeLists.txt

@@ -0,0 +1,8 @@
+add_executable(client_groupnorm_bwd_data groupnorm_bwd_data.cpp)
+target_link_libraries(client_groupnorm_bwd_data PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_groupnorm_bwd_gamma_beta groupnorm_bwd_gamma_beta.cpp)
+target_link_libraries(client_groupnorm_bwd_gamma_beta PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_groupnorm_swish_fwd groupnorm_swish_fwd.cpp)
+target_link_libraries(client_groupnorm_swish_fwd PRIVATE composable_kernel::device_other_operations)

client_example/18_groupnorm/groupnorm_bwd_data.cpp

@@ -0,0 +1,182 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/groupnorm_bwd_data.hpp"
+
+using DYDataType         = float;
+using XDataType          = float;
+using GammaDataType      = float;
+using MeanInvStdDataType = float;
+using DXDataType         = float;
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+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(int argc, char* argv[])
+{
+    ck::index_t N = 32;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 64;
+    ck::index_t C = 128;
+
+    std::size_t length = N * H * W * G * C;
+
+    std::vector<ck::index_t> strideDy = {H * W * G * C, W * G * C, G * C, C, 1};
+    std::vector<ck::index_t> strideX  = strideDy;
+    std::vector<ck::index_t> strideDx = strideDy;
+
+    std::vector<ck::index_t> strideGamma      = {0, 0, 0, C, 1};
+    std::vector<ck::index_t> strideMeanInvStd = {G, 0, 0, 1, 0};
+
+    SimpleDeviceMem dy_dev(sizeof(DYDataType) * length);
+    SimpleDeviceMem x_dev(sizeof(XDataType) * length);
+    SimpleDeviceMem gamma_dev(sizeof(GammaDataType) * G * C);
+    SimpleDeviceMem mean_dev(sizeof(MeanInvStdDataType) * N * G);
+    SimpleDeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * N * G);
+    SimpleDeviceMem dx_dev(sizeof(DXDataType) * length);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationBwdData<DYDataType,
+                                                                              XDataType,
+                                                                              GammaDataType,
+                                                                              MeanInvStdDataType,
+                                                                              DXDataType,
+                                                                              Rank,
+                                                                              NumReduceDim>;
+
+    // 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;
+    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({N, H, W, G, C},
+                                                        strideDy,
+                                                        strideX,
+                                                        strideGamma,
+                                                        strideMeanInvStd,
+                                                        strideMeanInvStd,
+                                                        strideDx,
+                                                        {1, 2, 4}, // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        gamma_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dx_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte = sizeof(DYDataType) * length + sizeof(XDataType) * length +
+                                   sizeof(GammaDataType) * G * C +
+                                   sizeof(MeanInvStdDataType) * N * G * 2 +
+                                   sizeof(DXDataType) * length;
+
+            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;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        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({N, H, W, G, C},
+                                                        strideDy,
+                                                        strideX,
+                                                        strideGamma,
+                                                        strideMeanInvStd,
+                                                        strideMeanInvStd,
+                                                        strideDx,
+                                                        {1, 2, 4}, // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        gamma_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dx_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/18_groupnorm/groupnorm_bwd_gamma_beta.cpp

@@ -0,0 +1,180 @@
+// 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/device_normalization_bwd_gamma_beta.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/groupnorm_bwd_gamma_beta.hpp"
+
+using DYDataType         = float;
+using XDataType          = float;
+using GammaDataType      = float;
+using MeanInvStdDataType = float;
+using DGammaDataType     = float;
+using DBetaDataType      = float;
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+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(int argc, char* argv[])
+{
+    ck::index_t N = 32;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 64;
+    ck::index_t C = 128;
+
+    std::size_t length = N * H * W * G * C;
+
+    std::vector<ck::index_t> strideDy         = {H * W * G * C, W * G * C, G * C, C, 1};
+    std::vector<ck::index_t> strideX          = strideDy;
+    std::vector<ck::index_t> strideMeanInvStd = {G, 0, 0, 1, 0};
+    std::vector<ck::index_t> strideDGammaBeta = {C, 1};
+
+    SimpleDeviceMem dy_dev(sizeof(DYDataType) * length);
+    SimpleDeviceMem x_dev(sizeof(XDataType) * length);
+    SimpleDeviceMem mean_dev(sizeof(MeanInvStdDataType) * N * G);
+    SimpleDeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * N * G);
+    SimpleDeviceMem dgamma_dev(sizeof(DGammaDataType) * G * C);
+    SimpleDeviceMem dbeta_dev(sizeof(DBetaDataType) * G * C);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceNormalizationBwdGammaBeta<DYDataType,
+                                                                      XDataType,
+                                                                      MeanInvStdDataType,
+                                                                      DGammaDataType,
+                                                                      DBetaDataType,
+                                                                      Rank,
+                                                                      NumReduceDim>;
+
+    // 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;
+    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;
+
+    std::size_t num_bytes = sizeof(DYDataType) * length + sizeof(XDataType) * length +
+                            sizeof(GammaDataType) * G * C + sizeof(MeanInvStdDataType) * N * G * 2 +
+                            sizeof(DGammaDataType) * G * C + sizeof(DBetaDataType) * G * C;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C},
+                                                        strideDy,
+                                                        strideX,
+                                                        strideMeanInvStd,
+                                                        strideMeanInvStd,
+                                                        {G, C},
+                                                        strideDGammaBeta,
+                                                        strideDGammaBeta,
+                                                        {0, 1, 2}, // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dgamma_dev.GetDeviceBuffer(),
+                                                        dbeta_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+            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)
+            {
+                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;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        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({N, H, W, G, C},
+                                                        strideDy,
+                                                        strideX,
+                                                        strideMeanInvStd,
+                                                        strideMeanInvStd,
+                                                        {G, C},
+                                                        strideDGammaBeta,
+                                                        strideDGammaBeta,
+                                                        {0, 1, 2}, // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dgamma_dev.GetDeviceBuffer(),
+                                                        dbeta_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/18_groupnorm/groupnorm_swish_fwd.cpp

@@ -0,0 +1,236 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization_fwd_swish.hpp"
+
+using XDataType              = ck::half_t;
+using GammaDataType          = float;
+using BetaDataType           = float;
+using YDataType              = ck::half_t;
+using SaveMeanInvStdDataType = float;
+using Swish                  = ck::tensor_operation::element_wise::Swish;
+
+#define SAVE_MEAN_INV_STD
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+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(int argc, char* argv[])
+{
+    ck::index_t N = 32;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 64;
+    ck::index_t C = 128;
+
+    std::size_t xy_size         = N * H * W * G * C;
+    std::size_t gamma_beta_size = G * C;
+
+    std::vector<ck::index_t> xy_strides                = {H * W * G * C, W * G * C, G * C, C, 1};
+    std::vector<ck::index_t> gamma_beta_strides        = {0, 0, 0, C, 1};
+    std::vector<ck::index_t> save_mean_inv_std_strides = {G, 1};
+
+    SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_beta_size);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * gamma_beta_size);
+    SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
+#ifdef SAVE_MEAN_INV_STD
+    SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
+    SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
+#endif
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
+                                                                          GammaDataType,
+                                                                          BetaDataType,
+                                                                          YDataType,
+                                                                          SaveMeanInvStdDataType,
+                                                                          Swish,
+                                                                          Rank,
+                                                                          NumReduceDim>;
+
+    // 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;
+
+    const auto& generic_op_ptr = op_ptrs[0];
+
+    auto generic_argument_ptr =
+        generic_op_ptr->MakeArgumentPointer({N, H, W, G, C},           // lengths
+                                            xy_strides,                // xStrides
+                                            gamma_beta_strides,        // gammaStrides
+                                            gamma_beta_strides,        // betaStrides
+                                            xy_strides,                // yStrides
+                                            save_mean_inv_std_strides, // save_mean Strides
+                                            save_mean_inv_std_strides, // save_inv_std Strides
+                                            {1, 2, 4},                 // reduceDims
+                                            1e-6,
+                                            x_device_buf.GetDeviceBuffer(),
+                                            gamma_device_buf.GetDeviceBuffer(),
+                                            beta_device_buf.GetDeviceBuffer(),
+                                            y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                            save_mean_device_buf.GetDeviceBuffer(),
+                                            save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                            nullptr,
+                                            nullptr,
+#endif
+                                            Swish{});
+
+    if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The generic kernel instance should be able to support any input shapes");
+    };
+
+    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({N, H, W, G, C},           // lengths
+                                        xy_strides,                // xStrides
+                                        gamma_beta_strides,        // gammaStrides
+                                        gamma_beta_strides,        // betaStrides
+                                        xy_strides,                // yStrides
+                                        save_mean_inv_std_strides, // save_mean Strides
+                                        save_mean_inv_std_strides, // save_inv_std Strides
+                                        {1, 2, 4},                 // reduceDims
+                                        1e-6,
+                                        x_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer(),
+                                        y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                        save_mean_device_buf.GetDeviceBuffer(),
+                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                        nullptr,
+                                        nullptr,
+#endif
+                                        Swish{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(XDataType) * xy_size + sizeof(GammaDataType) * gamma_beta_size +
+                sizeof(BetaDataType) * gamma_beta_size + sizeof(YDataType) * xy_size;
+
+#ifdef SAVE_MEAN_INV_STD
+            num_byte += sizeof(SaveMeanInvStdDataType) * N * G * 2;
+#endif
+
+            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;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        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({N, H, W, G, C},           // lengths
+                                        xy_strides,                // xStrides
+                                        gamma_beta_strides,        // gammaStrides
+                                        gamma_beta_strides,        // betaStrides
+                                        xy_strides,                // yStrides
+                                        save_mean_inv_std_strides, // save_mean Strides
+                                        save_mean_inv_std_strides, // save_inv_std Strides
+                                        {1, 2, 4},                 // reduceDims
+                                        1e-6,
+                                        x_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer(),
+                                        y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                        save_mean_device_buf.GetDeviceBuffer(),
+                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                        nullptr,
+                                        nullptr,
+#endif
+                                        Swish{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/19_pool/CMakeLists.txt

@@ -0,0 +1,11 @@
+add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
+target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_max_pool2d_bwd max_pool2d_bwd.cpp)
+target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
+target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_avg_pool3d_bwd avg_pool3d_bwd.cpp)
+target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_other_operations)

client_example/19_pool/avg_pool3d_bwd.cpp

@@ -0,0 +1,191 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp"
+
+using DOutDataType = ck::half_t;
+using DInDataType  = ck::half_t;
+
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+    std::size_t mMemSize_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Z                 = 2;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Di                = 30;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_d   = 2;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_d = 1;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_d     = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_d    = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    ck::index_t Do       = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
+    ck::index_t Ho       = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    ck::index_t Wo       = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    // Pool API only support the order of NCDHW
+    std::vector<ck::index_t> in_length              = {N, C, Di, Hi, Wi};
+    std::vector<ck::index_t> out_length             = {N, C, Do, Ho, Wo};
+    std::vector<ck::index_t> window_spatial_lengths = {Z, Y, X};
+    std::vector<ck::index_t> window_strides = {window_stride_d, window_stride_h, window_stride_w};
+    std::vector<ck::index_t> window_dilations{
+        window_dilation_d, window_dilation_h, window_dilation_w};
+    std::vector<ck::index_t> input_left_pads  = {in_left_pad_d, in_left_pad_h, in_left_pad_w};
+    std::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
+
+    std::size_t in_tensor_size  = N * C * Di * Hi * Wi;
+    std::size_t out_tensor_size = N * C * Do * Ho * Wo;
+
+    // tensor layout = NDHWC
+    std::vector<ck::index_t> in_tensor_stride  = {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C};
+    std::vector<ck::index_t> out_tensor_stride = {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C};
+
+    SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
+    SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceAvgPoolBwd<3, DOutDataType, DInDataType, DOutLayout, DInLayout>;
+
+    // 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;
+    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(
+            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            out_length,
+            in_length,
+            out_tensor_stride,
+            in_tensor_stride,
+            window_spatial_lengths,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            din_device_buf.SetZero();
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                in_tensor_size * sizeof(DInDataType) + out_tensor_size * sizeof(DOutDataType);
+
+            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)
+            {
+                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;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        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<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            out_length,
+            in_length,
+            out_tensor_stride,
+            in_tensor_stride,
+            window_spatial_lengths,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            din_device_buf.SetZero();
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/19_pool/avg_pool3d_fwd.cpp

@@ -0,0 +1,213 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+
+using InDataType    = ck::half_t;
+using OutDataType   = ck::half_t;
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using OutLayout = ck::tensor_layout::convolution::NDHWC;
+
+constexpr ck::index_t InOutRank  = 5;
+constexpr ck::index_t WindowRank = 3;
+#if 0
+constexpr auto ReduceOpId  = ck::ReduceTensorOp::MAX;
+constexpr bool OutputIndex = false;
+#else
+constexpr auto ReduceOpId  = ck::ReduceTensorOp::AVG;
+constexpr bool OutputIndex = false;
+#endif
+
+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(int argc, char* argv[])
+{
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Z                 = 2;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Di                = 30;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_d   = 2;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_d = 1;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_d     = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_d    = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    ck::index_t Do       = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
+    ck::index_t Ho       = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    ck::index_t Wo       = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    // Pool API only support the order of NCDHW
+    std::vector<ck::index_t> in_length              = {N, C, Di, Hi, Wi};
+    std::vector<ck::index_t> out_length             = {N, C, Do, Ho, Wo};
+    std::vector<ck::index_t> window_spatial_lengths = {Z, Y, X};
+    std::vector<ck::index_t> window_strides = {window_stride_d, window_stride_h, window_stride_w};
+    std::vector<ck::index_t> window_dilations{
+        window_dilation_d, window_dilation_h, window_dilation_w};
+    std::vector<ck::index_t> input_left_pads  = {in_left_pad_d, in_left_pad_h, in_left_pad_w};
+    std::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
+
+    std::size_t in_tensor_size  = N * C * Di * Hi * Wi;
+    std::size_t out_tensor_size = N * C * Do * Ho * Wo;
+
+    // tensor layout = NDHWC
+    std::vector<ck::index_t> in_tensor_stride  = {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C};
+    std::vector<ck::index_t> out_tensor_stride = {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C};
+
+    SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
+    SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
+
+    using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                 WindowRank,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 IndexDataType,
+                                                                 InLayout,
+                                                                 OutLayout,
+                                                                 ReduceOpId,
+                                                                 OutputIndex>;
+
+    // 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;
+    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(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        nullptr,
+                                        in_length,
+                                        window_spatial_lengths,
+                                        out_length,
+                                        in_tensor_stride,
+                                        out_tensor_stride,
+                                        out_tensor_stride,
+                                        window_strides,
+                                        window_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        {2, 3, 4});
+
+        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 =
+                in_tensor_size * sizeof(InDataType) + out_tensor_size * sizeof(OutDataType);
+
+            if constexpr(OutputIndex)
+                num_bytes += out_tensor_size * sizeof(IndexDataType);
+
+            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)
+            {
+                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;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        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_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        nullptr,
+                                        in_length,
+                                        window_spatial_lengths,
+                                        out_length,
+                                        in_tensor_stride,
+                                        out_tensor_stride,
+                                        out_tensor_stride,
+                                        window_strides,
+                                        window_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        {2, 3, 4});
+
+        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/19_pool/max_pool2d_bwd.cpp

@@ -0,0 +1,280 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp"
+
+using InDataType    = ck::half_t;
+using OutDataType   = ck::half_t;
+using DOutDataType  = ck::half_t;
+using DInDataType   = ck::half_t;
+using IndexDataType = int32_t;
+
+// We use pool3d to implement pool2d in this example
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using OutLayout = ck::tensor_layout::convolution::NDHWC;
+
+constexpr ck::index_t InOutRank  = 5;
+constexpr ck::index_t WindowRank = 3;
+
+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_;
+};
+
+void TransformPool2dparamToPool3d(std::vector<ck::index_t>& input_lengths,
+                                  std::vector<ck::index_t>& window_lengths,
+                                  std::vector<ck::index_t>& output_lengths,
+                                  std::vector<ck::index_t>& input_stride,
+                                  std::vector<ck::index_t>& output_stride,
+                                  std::vector<ck::index_t>& indices_stride,
+                                  std::vector<ck::index_t>& window_strides,
+                                  std::vector<ck::index_t>& window_dilations,
+                                  std::vector<ck::index_t>& input_left_pads,
+                                  std::vector<ck::index_t>& input_right_pads,
+                                  std::vector<ck::index_t>& pooling_dims)
+{
+    // NCHW to NCDHW
+    input_lengths.insert(input_lengths.begin() + 2, 1);
+    output_lengths.insert(output_lengths.begin() + 2, 1);
+    input_stride.insert(input_stride.begin() + 2, 0);
+    output_stride.insert(output_stride.begin() + 2, 0);
+    indices_stride.insert(indices_stride.begin() + 2, 0);
+
+    // YX to ZYX
+    window_lengths.insert(window_lengths.begin(), 1);
+    window_strides.insert(window_strides.begin(), 0);
+    window_dilations.insert(window_dilations.begin(), 0);
+    input_left_pads.insert(input_left_pads.begin(), 0);
+    input_right_pads.insert(input_right_pads.begin(), 0);
+
+    pooling_dims = {2, 3, 4};
+}
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    ck::index_t Ho       = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    ck::index_t Wo       = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    // Pool API only support the order of NCHW
+    std::vector<ck::index_t> in_length              = {N, C, Hi, Wi};
+    std::vector<ck::index_t> out_length             = {N, C, Ho, Wo};
+    std::vector<ck::index_t> window_spatial_lengths = {Y, X};
+    std::vector<ck::index_t> window_strides         = {window_stride_h, window_stride_w};
+    std::vector<ck::index_t> window_dilations       = {window_dilation_h, window_dilation_w};
+    std::vector<ck::index_t> input_left_pads        = {in_left_pad_h, in_left_pad_w};
+    std::vector<ck::index_t> input_right_pads       = {in_right_pad_h, in_right_pad_w};
+    std::vector<ck::index_t> pooling_dims           = {2, 3};
+
+    std::size_t in_tensor_size  = N * C * Hi * Wi;
+    std::size_t out_tensor_size = N * C * Ho * Wo;
+
+    // tensor layout = NHWC
+    std::vector<ck::index_t> in_tensor_stride  = {C * Hi * Wi, 1, Wi * C, C};
+    std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
+
+    TransformPool2dparamToPool3d(in_length,
+                                 window_spatial_lengths,
+                                 out_length,
+                                 in_tensor_stride,
+                                 out_tensor_stride,
+                                 out_tensor_stride,
+                                 window_strides,
+                                 window_dilations,
+                                 input_left_pads,
+                                 input_right_pads,
+                                 pooling_dims);
+
+    SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
+    SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
+    SimpleDeviceMem indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
+    SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
+    SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
+
+    // Generate index data from max pool forward
+    {
+        using MaxPoolFwdDeviceOp =
+            ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                        WindowRank,
+                                                        InDataType,
+                                                        OutDataType,
+                                                        IndexDataType,
+                                                        InLayout,
+                                                        OutLayout,
+                                                        ck::ReduceTensorOp::MAX,
+                                                        true>;
+
+        const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            MaxPoolFwdDeviceOp>::GetInstances();
+
+        auto& op_ptr = op_ptrs[0];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            in_tensor_stride,
+            out_tensor_stride,
+            out_tensor_stride,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            pooling_dims);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+    }
+
+    // Run MaxPool bwd
+    using MaxPoolBwdDeviceOp =
+        ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        MaxPoolBwdDeviceOp>::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(
+            static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            out_tensor_size,
+            in_tensor_size,
+            window_spatial_lengths,
+            window_strides,
+            window_dilations);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = in_tensor_size * sizeof(DInDataType) +
+                                    out_tensor_size * sizeof(IndexDataType) +
+                                    out_tensor_size * sizeof(DOutDataType);
+
+            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)
+            {
+                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;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        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*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            out_tensor_size,
+            in_tensor_size,
+            window_spatial_lengths,
+            window_strides,
+            window_dilations);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/19_pool/max_pool2d_fwd.cpp

@@ -0,0 +1,248 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+
+using InDataType    = ck::half_t;
+using OutDataType   = ck::half_t;
+using IndexDataType = int32_t;
+
+// We use pool3d to implement pool2d in this example
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using OutLayout = ck::tensor_layout::convolution::NDHWC;
+
+constexpr ck::index_t InOutRank  = 5;
+constexpr ck::index_t WindowRank = 3;
+#if 1
+constexpr auto ReduceOpId  = ck::ReduceTensorOp::MAX;
+constexpr bool OutputIndex = true;
+#else
+constexpr auto ReduceOpId  = ck::ReduceTensorOp::AVG;
+constexpr bool OutputIndex = false;
+#endif
+
+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_;
+};
+
+void TransformPool2dparamToPool3d(std::vector<ck::index_t>& input_lengths,
+                                  std::vector<ck::index_t>& window_lengths,
+                                  std::vector<ck::index_t>& output_lengths,
+                                  std::vector<ck::index_t>& input_stride,
+                                  std::vector<ck::index_t>& output_stride,
+                                  std::vector<ck::index_t>& indices_stride,
+                                  std::vector<ck::index_t>& window_strides,
+                                  std::vector<ck::index_t>& window_dilations,
+                                  std::vector<ck::index_t>& input_left_pads,
+                                  std::vector<ck::index_t>& input_right_pads,
+                                  std::vector<ck::index_t>& pooling_dims)
+{
+    // NCHW to NCDHW
+    input_lengths.insert(input_lengths.begin() + 2, 1);
+    output_lengths.insert(output_lengths.begin() + 2, 1);
+    input_stride.insert(input_stride.begin() + 2, 0);
+    output_stride.insert(output_stride.begin() + 2, 0);
+    indices_stride.insert(indices_stride.begin() + 2, 0);
+
+    // YX to ZYX
+    window_lengths.insert(window_lengths.begin(), 1);
+    window_strides.insert(window_strides.begin(), 0);
+    window_dilations.insert(window_dilations.begin(), 0);
+    input_left_pads.insert(input_left_pads.begin(), 0);
+    input_right_pads.insert(input_right_pads.begin(), 0);
+
+    pooling_dims = {2, 3, 4};
+}
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    ck::index_t Ho       = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    ck::index_t Wo       = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    // Pool API only support the order of NCHW
+    std::vector<ck::index_t> in_length              = {N, C, Hi, Wi};
+    std::vector<ck::index_t> out_length             = {N, C, Ho, Wo};
+    std::vector<ck::index_t> window_spatial_lengths = {Y, X};
+    std::vector<ck::index_t> window_strides         = {window_stride_h, window_stride_w};
+    std::vector<ck::index_t> window_dilations       = {window_dilation_h, window_dilation_w};
+    std::vector<ck::index_t> input_left_pads        = {in_left_pad_h, in_left_pad_w};
+    std::vector<ck::index_t> input_right_pads       = {in_right_pad_h, in_right_pad_w};
+    std::vector<ck::index_t> pooling_dims           = {2, 3};
+
+    std::size_t in_tensor_size  = N * C * Hi * Wi;
+    std::size_t out_tensor_size = N * C * Ho * Wo;
+
+    // tensor layout = NHWC
+    std::vector<ck::index_t> in_tensor_stride  = {C * Hi * Wi, 1, Wi * C, C};
+    std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
+
+    TransformPool2dparamToPool3d(in_length,
+                                 window_spatial_lengths,
+                                 out_length,
+                                 in_tensor_stride,
+                                 out_tensor_stride,
+                                 out_tensor_stride,
+                                 window_strides,
+                                 window_dilations,
+                                 input_left_pads,
+                                 input_right_pads,
+                                 pooling_dims);
+
+    SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
+    SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
+    SimpleDeviceMem out_indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
+
+    using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                 WindowRank,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 IndexDataType,
+                                                                 InLayout,
+                                                                 OutLayout,
+                                                                 ReduceOpId,
+                                                                 OutputIndex>;
+
+    // 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;
+    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(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            in_tensor_stride,
+            out_tensor_stride,
+            out_tensor_stride,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            pooling_dims);
+
+        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 =
+                in_tensor_size * sizeof(InDataType) + out_tensor_size * sizeof(OutDataType);
+
+            if constexpr(OutputIndex)
+                num_bytes += out_tensor_size * sizeof(IndexDataType);
+
+            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)
+            {
+                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;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        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_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            in_tensor_stride,
+            out_tensor_stride,
+            out_tensor_stride,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            pooling_dims);
+
+        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/20_splitk_gemm/CMakeLists.txt

@@ -0,0 +1,4 @@
+if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
+  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/20_splitk_gemm/splitK_gemm_fp16_f8.cpp

@@ -0,0 +1,226 @@
+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+using ADataType = F8;
+using BDataType = F16;
+using CDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+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(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    ck::index_t KBatch = 1;
+
+    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]);
+        StrideC = std::stoi(argv[6]);
+
+        KBatch = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideC, KBatch\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 a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<
+        ALayout,
+        BLayout,
+        CLayout,
+        ADataType,
+        BDataType,
+        CDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::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;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_element_op,
+                                                        KBatch);
+
+        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(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << 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(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_element_op,
+                                                        KBatch);
+
+        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/21_grouped_gemm_bias/CMakeLists.txt

@@ -0,0 +1,4 @@
+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/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp

@@ -0,0 +1,243 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F32;
+using DsDataType = ck::Tuple<D0DataType>;
+using EDataType  = F32;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+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()
+{
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    int sum_of_m = 0;
+
+    const int group_count = 16;
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ms.push_back(256 + 256 * i);
+        Ns.push_back(128 + 128 * i);
+        Ks.push_back(128 + 64 * i);
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+
+        sum_of_m += Ms[i];
+    }
+
+    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;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, d0_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    d0_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<void*> p_e;
+
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        d0_dev_bufs.emplace_back(sizeof(D0DataType) *
+                                 f_matrix_space_size(Ms[i], Ns[i], 0, D0Layout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
+
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back(
+            {a_dev_bufs[i].GetDeviceBuffer(),
+             b_dev_bufs[i].GetDeviceBuffer(),
+             std::array<const void*, 1>{d0_dev_bufs[i].GetDeviceBuffer()},
+             e_dev_bufs[i].GetDeviceBuffer(),
+             Ms[i],
+             Ns[i],
+             Ks[i],
+             StrideAs[i],
+             StrideBs[i],
+             std::array<ck::index_t, 1>{0},
+             StrideEs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            DsLayout,
+                                                                            ELayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            DsDataType,
+                                                                            EDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CDEElementOp>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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;
+
+    std::vector<const void*> p_a = {}, p_b = {};
+    std::vector<std::array<const void*, 1>> p_ds = {};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem grouped_gemm_kernel_args_dev(
+            op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                    grouped_gemm_workspace_dev.GetDeviceBuffer());
+
+        op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                    grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        op_ptr->SetKBatch(argument_ptr.get(), 2);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            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;
+
+    return 0;
+}

client_example/22_grouped_gemm/CMakeLists.txt

@@ -0,0 +1,13 @@
+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_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)
+endif()

client_example/22_grouped_gemm/grouped_gemm_fixed_nk_bf16.cpp

@@ -0,0 +1,237 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
+
+using I8   = int8_t;
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType  = BF16;
+using BDataType  = I8;
+using DsDataType = ck::Tuple<>;
+using EDataType  = BF16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = 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()
+{
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    int sum_of_m = 0;
+
+    const int group_count = 16;
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ms.push_back(256 + 256 * i);
+        Ns.push_back(128 + 128 * i);
+        Ks.push_back(128 + 64 * i);
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+
+        sum_of_m += Ms[i];
+    }
+
+    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;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<void*> p_e;
+
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
+
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
+                                             b_dev_bufs[i].GetDeviceBuffer(),
+                                             {},
+                                             e_dev_bufs[i].GetDeviceBuffer(),
+                                             Ms[i],
+                                             Ns[i],
+                                             Ks[i],
+                                             StrideAs[i],
+                                             StrideBs[i],
+                                             {},
+                                             StrideEs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            DsLayout,
+                                                                            ELayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            DsDataType,
+                                                                            EDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CDEElementOp>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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;
+
+    std::vector<const void*> p_a = {}, p_b = {};
+    std::vector<std::array<const void*, 0>> p_ds = {};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem grouped_gemm_kernel_args_dev(
+            op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                    grouped_gemm_workspace_dev.GetDeviceBuffer());
+
+        op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                    grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        op_ptr->SetKBatch(argument_ptr.get(), 1);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            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;
+
+    return 0;
+}

client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp

@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using DsDataType = ck::Tuple<>;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = 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()
+{
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    int sum_of_m = 0;
+
+    const int group_count = 16;
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ms.push_back(256 + 256 * i);
+        Ns.push_back(128 + 128 * i);
+        Ks.push_back(128 + 64 * i);
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+
+        sum_of_m += Ms[i];
+    }
+
+    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;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<void*> p_e;
+
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
+
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
+                                             b_dev_bufs[i].GetDeviceBuffer(),
+                                             {},
+                                             e_dev_bufs[i].GetDeviceBuffer(),
+                                             Ms[i],
+                                             Ns[i],
+                                             Ks[i],
+                                             StrideAs[i],
+                                             StrideBs[i],
+                                             {},
+                                             StrideEs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            DsLayout,
+                                                                            ELayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            DsDataType,
+                                                                            EDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CDEElementOp>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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;
+
+    std::vector<const void*> p_a = {}, p_b = {};
+    std::vector<std::array<const void*, 0>> p_ds = {};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem grouped_gemm_kernel_args_dev(
+            op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                    grouped_gemm_workspace_dev.GetDeviceBuffer());
+
+        op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                    grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        op_ptr->SetKBatch(argument_ptr.get(), 32);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            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;
+
+    return 0;
+}