merge develop

client_example/19_pool_fwd/CMakeLists.txt → client_example/19_pool/CMakeLists.txt

 add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
-target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
+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_operations)
+target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_other_operations)
\ No newline at end of file
+
+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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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_fwd/avg_pool3d_fwd.cpp → client_example/19_pool/avg_pool3d_fwd.cpp

@@ -16,6 +16,9 @@ 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
@@ -44,33 +47,41 @@ struct SimpleDeviceMem
 
 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 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;
-
-    ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Z) / window_stride_d + 1;
-    ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
-    ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
+    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_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};
 
@@ -83,13 +94,14 @@ int main(int argc, char* argv[])
 
     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>;
 
@@ -110,21 +122,22 @@ int main(int argc, char* argv[])
 
     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,
-            input_left_pads,
-            input_right_pads,
-            {2, 3, 4});
+        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();
 
@@ -170,20 +183,21 @@ int main(int argc, char* argv[])
         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,
-            input_left_pads,
-            input_right_pads,
-            {2, 3, 4});
+        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();
 

client_example/19_pool/max_pool2d_bwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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_fwd/max_pool2d_fwd.cpp → client_example/19_pool/max_pool2d_fwd.cpp

@@ -10,14 +10,18 @@
 #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/pool2d_fwd.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;
 
-constexpr ck::index_t InOutRank  = 4;
-constexpr ck::index_t WindowRank = 2;
+// 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;
@@ -42,31 +46,66 @@ struct SimpleDeviceMem
     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 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;
-
-    ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
-    ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
+    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;
@@ -75,6 +114,18 @@ int main(int argc, char* argv[])
     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);
@@ -84,6 +135,8 @@ int main(int argc, char* argv[])
                                                                  InDataType,
                                                                  OutDataType,
                                                                  IndexDataType,
+                                                                 InLayout,
+                                                                 OutLayout,
                                                                  ReduceOpId,
                                                                  OutputIndex>;
 
@@ -116,9 +169,10 @@ int main(int argc, char* argv[])
             out_tensor_stride,
             out_tensor_stride,
             window_strides,
+            window_dilations,
             input_left_pads,
             input_right_pads,
-            {2, 3});
+            pooling_dims);
 
         auto invoker_ptr = op_ptr->MakeInvokerPointer();
 
@@ -175,9 +229,10 @@ int main(int argc, char* argv[])
             out_tensor_stride,
             out_tensor_stride,
             window_strides,
+            window_dilations,
             input_left_pads,
             input_right_pads,
-            {2, 3});
+            pooling_dims);
 
         auto invoker_ptr = op_ptr->MakeInvokerPointer();
 

client_example/20_splitk_gemm/CMakeLists.txt

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

client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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, 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 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

+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)

client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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, 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, 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

+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)

client_example/22_grouped_gemm/grouped_gemm_fixed_nk_bf16.cpp

+// 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, 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<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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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, 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<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;
+}

client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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 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 ADataType  = F16;
+using BDataType  = F8;
+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 = 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, 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<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(), 16);
+
+        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_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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 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  = I8;
+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, 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<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;
+}

client_example/22_im2col_col2im/CMakeLists.txt

+add_executable(client_image_to_column image_to_column.cpp)
+target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_column_to_image column_to_image.cpp)
+target_link_libraries(client_column_to_image PRIVATE composable_kernel::device_other_operations)

client_example/22_im2col_col2im/column_to_image.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, 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/conv_tensor_rearrange.hpp"
+#include "ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+
+using ImageLayout = ck::tensor_layout::convolution::NHWGC;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 2;
+static constexpr ck::index_t N             = 32; // batch size
+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
+
+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, 2> in_spatial_lengths{Hi, Wi};
+    std::array<ck::index_t, 2> wei_spatial_lengths{Y, X};
+    std::array<ck::index_t, 2> out_spatial_lengths{Ho, Wo};
+
+    // We have NHWGC in memory space
+    // However, CK's API only accepts lengths and strides with order of GNCHW.
+    // Hence, we need to adjust the order of strides.
+    std::array<ck::index_t, 5> image_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 3> gemm_strides{Y * X * C, G * Y * X * C, 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) * G * N * Ho * Wo * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Hi * Wi * G * C);
+
+    using namespace ck::conv_tensor_rearrange_op;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvTensorRearrange<NumDimSpatial,
+                                                                             ImageLayout,
+                                                                             InDataType,
+                                                                             OutDataType,
+                                                                             ColumnToImage>;
+
+    // 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;
+
+    // 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(),
+                                                        out.GetDeviceBuffer(),
+                                                        G,
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        image_strides,
+                                                        gemm_strides,
+                                                        filter_strides,
+                                                        filter_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()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
+                                    sizeof(OutDataType) * G * N * Ho * Wo * Y * X * C;
+
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        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_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(),
+                                                        out.GetDeviceBuffer(),
+                                                        G,
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        image_strides,
+                                                        gemm_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+
+        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/22_im2col_col2im/image_to_column.cpp

+// 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 <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp"
+#include "ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+
+using ImageLayout = ck::tensor_layout::convolution::NHWGC;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 2;
+static constexpr ck::index_t N             = 32; // batch size
+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
+
+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, 2> in_spatial_lengths{Hi, Wi};
+    std::array<ck::index_t, 2> wei_spatial_lengths{Y, X};
+    std::array<ck::index_t, 2> out_spatial_lengths{Ho, Wo};
+
+    // We have NHWGC in memory space
+    // However, CK's API only accepts lengths and strides with order of GNCHW.
+    // Hence, we need to adjust the order of strides.
+    std::array<ck::index_t, 5> image_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 3> gemm_strides{Y * X * C, G * Y * X * C, 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 out(sizeof(OutDataType) * G * N * Ho * Wo * Y * X * C);
+
+    using namespace ck::conv_tensor_rearrange_op;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvTensorRearrange<NumDimSpatial,
+                                                                             ImageLayout,
+                                                                             InDataType,
+                                                                             OutDataType,
+                                                                             ImageToColumn>;
+
+    // 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;
+
+    // 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(),
+                                                        out.GetDeviceBuffer(),
+                                                        G,
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        image_strides,
+                                                        gemm_strides,
+                                                        filter_strides,
+                                                        filter_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()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
+                                    sizeof(OutDataType) * G * N * Ho * Wo * Y * X * C;
+
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        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_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(),
+                                                        out.GetDeviceBuffer(),
+                                                        G,
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        image_strides,
+                                                        gemm_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+
+        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/23_elementwise_transpose/CMakeLists.txt

+add_executable(client_elementwise_transpose3d elementwise_transpose_3d.cpp)
+target_link_libraries(client_elementwise_transpose3d PRIVATE composable_kernel::device_other_operations)

client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/transpose_3d.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    const int N = 16;
+    const int C = 8;
+    const int D = 8;
+    const int H = 8;
+    const int W = 8;
+
+    std::vector<std::size_t> ncdhw = {N, C, D, H, W};
+    std::vector<std::size_t> nchwd = {N, C, H, W, D};
+    auto size                      = N * C * D * H * W;
+
+    std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
+    std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
+    std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, C, H, W, D
+
+    SimpleDeviceMem a_dev_buf(sizeof(ADataType) * size);
+    SimpleDeviceMem b_dev_buf(sizeof(BDataType) * size);
+
+    std::array<const void*, 1> input = {a_dev_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_dev_buf.GetDeviceBuffer()};
+
+    using DeviceElementwisePermuteInstance = ck::tensor_operation::device::
+        DeviceElementwise<ck::Tuple<ADataType>, ck::Tuple<BDataType>, PassThrough, 5>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceElementwisePermuteInstance>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
+                sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/24_grouped_conv_activation/CMakeLists.txt

+# Fwd scaleadd scaleadd relu
+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 
+               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16
+               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp16 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16
+               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_bf16 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8
+               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_int8.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8 PRIVATE composable_kernel::device_conv_operations)
+# Fwd scaleadd AB
+add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp32
+               grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp32.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp32 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp16
+               grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_fp16 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_ab_bf16
+               grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_bf16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_bf16 PRIVATE composable_kernel::device_conv_operations)
+
+add_executable(client_grouped_convnd_fwd_scaleadd_ab_int8
+               grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_int8.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_int8 PRIVATE composable_kernel::device_conv_operations)
+# Fwd bilinear
+add_executable(client_grouped_convnd_fwd_bilinear_residual_fp16
+               grouped_convnd_fwd_bilinear/grouped_conv_fwd_bilinear_residual_fp16.cpp)
+target_link_libraries(client_grouped_convnd_fwd_bilinear_residual_fp16 PRIVATE composable_kernel::device_conv_operations)
+# Bwd data bilinear
+add_executable(client_grouped_convnd_bwd_data_bilinear_residual_fp16
+               grouped_convnd_bwd_data_bilinear/grouped_conv_bwd_data_bilinear_residual_fp16.cpp)
+target_link_libraries(client_grouped_convnd_bwd_data_bilinear_residual_fp16 PRIVATE composable_kernel::device_conv_operations)