Merge branch 'amd-develop' into amd-master

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_operations)
+  target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_gemm_operations)
 endif()

client_example/21_grouped_gemm_bias/CMakeLists.txt

 add_executable(client_grouped_gemm_fixed_nk_bias_fp16 grouped_gemm_fixed_nk_bias_fp16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_gemm_operations)

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_operations)
+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_operations)
+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_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_gemm_operations)

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_operations)
+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_operations)
+target_link_libraries(client_column_to_image PRIVATE composable_kernel::device_other_operations)

client_example/22_im2col_col2im/column_to_image.cpp

@@ -16,10 +16,10 @@
 using InDataType  = ck::half_t;
 using OutDataType = ck::half_t;
 
-using ImageLayout = ck::tensor_layout::convolution::GNHWC;
+using ImageLayout = ck::tensor_layout::convolution::NHWGC;
 
 static constexpr ck::index_t NumDimSpatial = 2;
-static constexpr ck::index_t G             = 1;
+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
@@ -52,18 +52,18 @@ int main()
     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 (G is dummy)
-    // However, CK's API only accept length and stride with order of GNCHW
-    // Hence, we need to adjust the order of stride
+    // 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, 2> gemm_strides{Y * X * C, 1};
+    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 * Ho * Wo * Y * X * C);
+    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;
@@ -93,6 +93,7 @@ int main()
         auto& op_ptr        = op_ptrs[i];
         auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
                                                         out.GetDeviceBuffer(),
+                                                        G,
                                                         N,
                                                         C,
                                                         in_spatial_lengths,
@@ -112,7 +113,7 @@ int main()
             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) * N * Ho * Wo * Y * X * C;
+                                    sizeof(OutDataType) * G * N * Ho * Wo * Y * X * C;
 
             float gb_per_sec = num_bytes / 1.E6 / avg_time;
 
@@ -149,6 +150,7 @@ int main()
                   << std::endl;
         auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
                                                         out.GetDeviceBuffer(),
+                                                        G,
                                                         N,
                                                         C,
                                                         in_spatial_lengths,

client_example/22_im2col_col2im/image_to_column.cpp

@@ -16,10 +16,10 @@
 using InDataType  = ck::half_t;
 using OutDataType = ck::half_t;
 
-using ImageLayout = ck::tensor_layout::convolution::GNHWC;
+using ImageLayout = ck::tensor_layout::convolution::NHWGC;
 
 static constexpr ck::index_t NumDimSpatial = 2;
-static constexpr ck::index_t G             = 1;
+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
@@ -52,11 +52,11 @@ int main()
     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 (G is dummy)
-    // However, CK's API only accept length and stride with order of GNCHW
-    // Hence, we need to adjust the order of stride
+    // 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, 2> gemm_strides{Y * X * C, 1};
+    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};
@@ -64,7 +64,7 @@ int main()
     std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
 
     SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
-    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * Y * X * C);
 
     using namespace ck::conv_tensor_rearrange_op;
 
@@ -93,6 +93,7 @@ int main()
         auto& op_ptr        = op_ptrs[i];
         auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
                                                         out.GetDeviceBuffer(),
+                                                        G,
                                                         N,
                                                         C,
                                                         in_spatial_lengths,
@@ -112,7 +113,7 @@ int main()
             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) * N * Ho * Wo * Y * X * C;
+                                    sizeof(OutDataType) * G * N * Ho * Wo * Y * X * C;
 
             float gb_per_sec = num_bytes / 1.E6 / avg_time;
 
@@ -149,6 +150,7 @@ int main()
                   << std::endl;
         auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
                                                         out.GetDeviceBuffer(),
+                                                        G,
                                                         N,
                                                         C,
                                                         in_spatial_lengths,

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
+    {
+        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/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/CMakeLists.txt

+add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 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_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_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_conv_fwd_scaleadd_scaleadd_relu_int8.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_int8 PRIVATE composable_kernel::device_conv_operations)

client_example/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu.inc

+// 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/grouped_convolution_forward_scaleadd_scaleadd_relu.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InLayout             = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout            = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout            = ck::tensor_layout::convolution::NDHWGK;
+using PassThrough          = ck::tensor_operation::element_wise::PassThrough;
+using ScaleAddScaleAddRelu = ck::tensor_operation::element_wise::ScaleAddScaleAddRelu;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 64; // batch size
+static constexpr ck::index_t K             = 64; // output channel
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Z             = 3;  // filter D
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Di            = 14; // input D
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+static constexpr ck::index_t Do            = 14; // output D
+static constexpr ck::index_t Ho            = 14; // output H
+static constexpr ck::index_t Wo            = 14; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int execute_conv_fwd_scaleadd_scaleadd_relu()
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space.
+    // However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
+    // Hence, we need to adjust the order of strides.
+    std::array<ck::index_t, 6> in_lengths{G, N, C, Di, Hi, Wi};
+    std::array<ck::index_t, 6> in_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+    std::array<ck::index_t, 6> wei_lengths{G, K, C, Z, Y, X};
+    std::array<ck::index_t, 6> wei_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+    std::array<ck::index_t, 6> out_lengths{G, N, K, Do, Ho, Wo};
+    std::array<ck::index_t, 6> out_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Di * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Do * Ho * Wo * G * K);
+    SimpleDeviceMem d0(sizeof(std::tuple_element_t<0, DDataTypes>) * N * Do * Ho * Wo * G * K);
+    SimpleDeviceMem d1(sizeof(std::tuple_element_t<1, DDataTypes>) * N * Do * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
+        NumDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<OutLayout, OutLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<std::tuple_element_t<0, DDataTypes>, std::tuple_element_t<1, DDataTypes>>,
+        OutDataType,
+        PassThrough,
+        PassThrough,
+        ScaleAddScaleAddRelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {d0.GetDeviceBuffer(), d1.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        {out_lengths, out_lengths},
+                                        {out_strides, out_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        filter_strides,
+                                        filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        ScaleAddScaleAddRelu{2.f, 2.f});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop =
+                std::size_t(2) * G * N * K * C * Ho * Wo * Y * X + 2 * N * Ho * Wo * G * K;
+            std::size_t num_bytes =
+                sizeof(InDataType) * N * Hi * Wi * G * C + sizeof(WeiDataType) * G * K * Y * X * C +
+                (sizeof(OutDataType) + sizeof(std::tuple_element_t<0, DDataTypes>) +
+                 sizeof(std::tuple_element_t<1, DDataTypes>)) *
+                    N * Ho * Wo * G * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {d0.GetDeviceBuffer(), d1.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        {out_lengths, out_lengths},
+                                        {out_strides, out_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        filter_strides,
+                                        filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        ScaleAddScaleAddRelu{2.f, 2.f});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return 0;
+}

client_example/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+using InDataType  = ck::bhalf_t;
+using WeiDataType = ck::bhalf_t;
+using OutDataType = ck::bhalf_t;
+// Use std tuple instead of ck tuple to avoid clang
+// implicit instantiation of undefined template error.
+using DDataTypes = std::tuple<ck::bhalf_t, ck::bhalf_t>;
+
+#include "grouped_conv_fwd_scaleadd_scaleadd_relu.inc"
+
+int main() { return execute_conv_fwd_scaleadd_scaleadd_relu(); }

client_example/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+// Use std tuple instead of ck tuple to avoid clang
+// implicit instantiation of undefined template error.
+using DDataTypes = std::tuple<ck::half_t, ck::half_t>;
+
+#include "grouped_conv_fwd_scaleadd_scaleadd_relu.inc"
+
+int main() { return execute_conv_fwd_scaleadd_scaleadd_relu(); }

client_example/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+using InDataType  = float;
+using WeiDataType = float;
+using OutDataType = float;
+// Use std tuple instead of ck tuple to avoid clang
+// implicit instantiation of undefined template error.
+using DDataTypes = std::tuple<float, float>;
+
+#include "grouped_conv_fwd_scaleadd_scaleadd_relu.inc"
+
+int main() { return execute_conv_fwd_scaleadd_scaleadd_relu(); }

client_example/23_grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using OutDataType = int8_t;
+// Use std tuple instead of ck tuple to avoid clang
+// implicit instantiation of undefined template error.
+using DDataTypes = std::tuple<float, float>;
+
+#include "grouped_conv_fwd_scaleadd_scaleadd_relu.inc"
+
+int main() { return execute_conv_fwd_scaleadd_scaleadd_relu(); }

client_example/24_grouped_convnd_fwd_scaleadd_ab/CMakeLists.txt

+add_executable(client_grouped_convnd_fwd_scaleadd_ab_fp32 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_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_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_conv_fwd_scaleadd_ab_int8.cpp)
+target_link_libraries(client_grouped_convnd_fwd_scaleadd_ab_int8 PRIVATE composable_kernel::device_conv_operations)

client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab.inc

+// 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/grouped_convolution_forward_scaleadd_ab.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InLayout    = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout   = ck::tensor_layout::convolution::NDHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 64; // batch size
+static constexpr ck::index_t K             = 64; // output channel
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Z             = 3;  // filter D
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Di            = 14; // input D
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+static constexpr ck::index_t Do            = 14; // output D
+static constexpr ck::index_t Ho            = 14; // output H
+static constexpr ck::index_t Wo            = 14; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int execute_conv_fwd_scaleadd_ab()
+{
+    constexpr ck::index_t NumAs = 2;
+    constexpr ck::index_t NumBs = 2;
+
+    constexpr float scale = 1.5f;
+
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space.
+    // However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
+    // Hence, we need to adjust the order of strides.
+    std::array<ck::index_t, 6> in_lengths{G, N, C, Di, Hi, Wi};
+    std::array<ck::index_t, 6> in_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+    std::array<ck::index_t, 6> wei_lengths{G, K, C, Z, Y, X};
+    std::array<ck::index_t, 6> wei_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+    std::array<ck::index_t, 6> out_lengths{G, N, K, Do, Ho, Wo};
+    std::array<ck::index_t, 6> out_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    using InputDtype      = ck::tuple_element_t<0, InDataType>;
+    using InputBiasDtype  = ck::tuple_element_t<1, InDataType>;
+    using WeightDtype     = ck::tuple_element_t<0, WeiDataType>;
+    using WeightBiasDtype = ck::tuple_element_t<1, WeiDataType>;
+
+    SimpleDeviceMem in(sizeof(InputDtype) * N * Di * Hi * Wi * G * C);
+    SimpleDeviceMem in_bias(sizeof(InputBiasDtype) * N * Di * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeightDtype) * G * K * Z * Y * X * C);
+    SimpleDeviceMem wei_bias(sizeof(WeightBiasDtype) * G * K * Z * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Do * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   ScaleAdd,
+                                                                                   ScaleAdd,
+                                                                                   PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    std::array<const void*, NumAs> as = {in.GetDeviceBuffer(), in_bias.GetDeviceBuffer()};
+    std::array<const void*, NumBs> bs = {wei.GetDeviceBuffer(), wei_bias.GetDeviceBuffer()};
+    std::array<const void*, 0> ds{};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(as,
+                                                        bs,
+                                                        ds,
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        ScaleAdd{scale},
+                                                        ScaleAdd{scale},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Z * Y * X +
+                               N * Di * Hi * Wi * G * C + G * K * Z * Y * X * C;
+            std::size_t num_bytes = 2 * sizeof(InDataType) * N * Di * Hi * Wi * G * C +
+                                    2 * sizeof(WeiDataType) * G * K * Z * Y * X * C +
+                                    sizeof(OutDataType) * N * Do * Ho * Wo * G * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(as,
+                                                        bs,
+                                                        ds,
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        ScaleAdd{scale},
+                                                        ScaleAdd{scale},
+                                                        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_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+using InDataType  = ck::Tuple<ck::bhalf_t, ck::bhalf_t>;
+using WeiDataType = ck::Tuple<ck::bhalf_t, ck::bhalf_t>;
+using OutDataType = ck::bhalf_t;
+
+#include "grouped_conv_fwd_scaleadd_ab.inc"
+
+int main() { return execute_conv_fwd_scaleadd_ab(); }

client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+using InDataType  = ck::Tuple<ck::half_t, ck::half_t>;
+using WeiDataType = ck::Tuple<ck::half_t, ck::half_t>;
+using OutDataType = ck::half_t;
+
+#include "grouped_conv_fwd_scaleadd_ab.inc"
+
+int main() { return execute_conv_fwd_scaleadd_ab(); }

client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+using InDataType  = ck::Tuple<float, float>;
+using WeiDataType = ck::Tuple<float, float>;
+using OutDataType = float;
+
+#include "grouped_conv_fwd_scaleadd_ab.inc"
+
+int main() { return execute_conv_fwd_scaleadd_ab(); }

client_example/24_grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+using InDataType  = ck::Tuple<int8_t, int8_t>;
+using WeiDataType = ck::Tuple<int8_t, int8_t>;
+using OutDataType = int8_t;
+
+#include "grouped_conv_fwd_scaleadd_ab.inc"
+
+int main() { return execute_conv_fwd_scaleadd_ab(); }