Merge remote-tracking branch 'origin/develop' into lwpck-367

profiler/include/profile_batched_gemm_reduce_impl.hpp

@@ -10,10 +10,10 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/conv_util.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 
 namespace ck {
@@ -193,13 +193,13 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
         }
     }
 
-    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
-    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
-    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
     DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
-                                 d0_g_m_device_result.mDesc.GetElementSpace());
+                                 d0_g_m_device_result.mDesc.GetElementSpaceSize());
     DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
-                                 d1_g_m_device_result.mDesc.GetElementSpace());
+                                 d1_g_m_device_result.mDesc.GetElementSpaceSize());
 
     std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
                                       reduce1_device_buf.GetDeviceBuffer()};
@@ -319,11 +319,11 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
                 reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
 
                 bool c_error =
-                    ck::utils::check_err(c_g_m_n_host_result.mData, c_g_m_n_device_result.mData);
+                    ck::utils::check_err(c_g_m_n_device_result.mData, c_g_m_n_host_result.mData);
                 bool d0_error =
-                    ck::utils::check_err(d0_g_m_host_result.mData, d0_g_m_device_result.mData);
+                    ck::utils::check_err(d0_g_m_device_result.mData, d0_g_m_host_result.mData);
                 bool d1_error =
-                    ck::utils::check_err(d1_g_m_host_result.mData, d1_g_m_device_result.mData);
+                    ck::utils::check_err(d1_g_m_device_result.mData, d1_g_m_host_result.mData);
 
                 pass = pass && (c_error == true);
                 pass = pass && (d0_error == true);

profiler/include/profile_conv_bwd_data_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename DataType>
+void show_data_nhwc_layout(Tensor<DataType>& nhwc)
+{
+    std::cout << "[";
+    for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
+    {
+        std::cout << "[";
+        for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
+        {
+            std::cout << "[";
+            for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
+            {
+                std::cout << "[";
+                for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
+                {
+                    std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << "  ";
+                }
+                std::cout << "]";
+            }
+            std::cout << "]";
+        }
+        std::cout << "]";
+    }
+    std::cout << "]";
+}
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_conv_bwd_data_impl(int do_verification,
+                                int init_method,
+                                bool do_log,
+                                bool time_kernel,
+                                const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input_host_result(in_g_n_c_wis_desc);
+    Tensor<InDataType> input_device_result(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input_host_result.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        output.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(output.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>{};
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(input_host_result,
+                                                  weight,
+                                                  output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  OutElementOp{});
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvBwdData<NDimSpatial,
+                                                                     InLayout,
+                                                                     WeiLayout,
+                                                                     OutLayout,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    bool pass = true;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        conv_param.N_,
+                                        conv_param.K_,
+                                        conv_param.C_,
+                                        conv_param.input_spatial_lengths_,
+                                        conv_param.filter_spatial_lengths_,
+                                        conv_param.output_spatial_lengths_,
+                                        conv_param.conv_filter_strides_,
+                                        conv_param.conv_filter_dilations_,
+                                        conv_param.input_left_pads_,
+                                        conv_param.input_right_pads_,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // for conv bwd data, some input tensor element are zero, but not written by kernel,
+            // need to set zero
+            in_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s" << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                in_device_buf.FromDevice(input_device_result.mData.data());
+
+                pass =
+                    pass & ck::utils::check_err(input_device_result.mData, input_host_result.mData);
+
+                if(do_log)
+                {
+                    std::cout << "in : ";
+                    show_data_nhwc_layout(output);
+                    std::cout << std::endl;
+
+                    std::cout << "wei: ";
+                    show_data_nhwc_layout(weight);
+                    std::cout << std::endl;
+
+                    std::cout << "out_host  : ";
+                    show_data_nhwc_layout(input_host_result);
+                    std::cout << std::endl;
+
+                    std::cout << "out_device: ";
+                    show_data_nhwc_layout(input_device_result);
+                    std::cout << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profile_conv_bwd_weight_impl.hpp

@@ -3,141 +3,134 @@
 
 #pragma once
 
+#include "ck/ck.hpp"
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/device_conv_backward_weight.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
-#include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/conv_util.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
-#include "ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-using DeviceConvBwdWeightNoOpPtr =
-    DeviceConvBwdWeightPtr<ck::tensor_operation::element_wise::PassThrough,
-                           ck::tensor_operation::element_wise::PassThrough,
-                           ck::tensor_operation::element_wise::PassThrough>;
-
-void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(
-    std::vector<DeviceConvBwdWeightNoOpPtr>&);
-
-void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(
-    std::vector<DeviceConvBwdWeightNoOpPtr>&);
+#include "ck/library/tensor_operation_instance/gpu/convolution_backward_weight.hpp"
 
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
 
 namespace ck {
 namespace profiler {
 
-template <int NDimSpatial,
-          typename InDataType,
-          typename WeiDataType,
-          typename OutDataType,
+template <typename DataType>
+void show_data_nhwc_layout(Tensor<DataType>& nhwc)
+{
+    std::cout << "[";
+    for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
+    {
+        std::cout << "[";
+        for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
+        {
+            std::cout << "[";
+            for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
+            {
+                std::cout << "[";
+                for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
+                {
+                    std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << "  ";
+                }
+                std::cout << "]";
+            }
+            std::cout << "]";
+        }
+        std::cout << "]";
+    }
+    std::cout << "]";
+}
+
+template <ck::index_t NDimSpatial,
           typename InLayout,
           typename WeiLayout,
-          typename OutLayout>
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
 bool profile_conv_bwd_weight_impl(int do_verification,
                                   int init_method,
                                   bool do_log,
                                   bool time_kernel,
-                                  ck::index_t N,
-                                  ck::index_t K,
-                                  ck::index_t C,
-                                  std::vector<ck::index_t> input_spatial_lengths,
-                                  std::vector<ck::index_t> filter_spatial_lengths,
-                                  std::vector<ck::index_t> output_spatial_lengths,
-                                  std::vector<ck::index_t> conv_filter_strides,
-                                  std::vector<ck::index_t> conv_filter_dilations,
-                                  std::vector<ck::index_t> input_left_pads,
-                                  std::vector<ck::index_t> input_right_pads,
+                                  const ck::utils::conv::ConvParam& conv_param,
                                   ck::index_t split_k)
 {
-    const ck::index_t Y = filter_spatial_lengths[0];
-    const ck::index_t X = filter_spatial_lengths[1];
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
 
-    const ck::index_t Hi = input_spatial_lengths[0];
-    const ck::index_t Wi = input_spatial_lengths[1];
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
 
-    const ck::index_t Ho = output_spatial_lengths[0];
-    const ck::index_t Wo = output_spatial_lengths[1];
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
 
-    auto f_host_tensor_descriptor =
-        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
-            if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
-                         is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
-                         is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
-            {
-                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
-                                            std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
-            }
-            else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
-                              is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
-                              is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
-            {
-                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
-                                            std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
-            }
-        };
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
 
-    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
-    Tensor<WeiDataType> wei_k_c_y_x_host_result(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
-    Tensor<WeiDataType> wei_k_c_y_x_device_result(
-        f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
-    Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight_host_result(wei_g_k_c_xs_desc);
+    Tensor<WeiDataType> weight_device_result(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> output(out_g_n_k_wos_desc);
 
-    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
-    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
-    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight_host_result.mDesc << std::endl;
+    std::cout << "output: " << output.mDesc << std::endl;
 
     switch(init_method)
     {
     case 0: break;
     case 1:
-        out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
-        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
         break;
     default:
-        out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
-        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        output.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
     }
 
-    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
-    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
-    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) *
+                             weight_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpaceSize());
 
-    const auto in_element_op  = InElementOp{};
-    const auto wei_element_op = WeiElementOp{};
-    const auto out_element_op = OutElementOp{};
+    in_device_buf.ToDevice(input.mData.data());
+    out_device_buf.ToDevice(output.mData.data());
 
     if(do_verification)
     {
-        using ReferenceConvBwdWeightInstance =
-            ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                           InDataType,
                                                                            WeiDataType,
                                                                            OutDataType,
                                                                            InElementOp,
                                                                            WeiElementOp,
-                                                               OutElementOp>;
+                                                                           OutElementOp>{};
 
-        auto ref_conv     = ReferenceConvBwdWeightInstance{};
         auto ref_invoker = ref_conv.MakeInvoker();
-        auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
-                                                  wei_k_c_y_x_host_result,
-                                                  out_n_k_ho_wo,
-                                                  conv_filter_strides,
-                                                  conv_filter_dilations,
-                                                  input_left_pads,
-                                                  input_right_pads,
+
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight_host_result,
+                                                  output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
                                                   in_element_op,
                                                   wei_element_op,
                                                   out_element_op);
@@ -145,140 +138,126 @@ bool profile_conv_bwd_weight_impl(int do_verification,
         ref_invoker.Run(ref_argument);
     }
 
-    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
-    DeviceMem wei_device_buf(sizeof(WeiDataType) *
-                             wei_k_c_y_x_device_result.mDesc.GetElementSpace());
-    DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
-
-    out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
-    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
-
-    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using DeviceOp = ck::tensor_operation::device::DeviceConvBwdWeight<NDimSpatial,
+                                                                       InLayout,
+                                                                       WeiLayout,
+                                                                       OutLayout,
+                                                                       InDataType,
+                                                                       WeiDataType,
+                                                                       OutDataType,
+                                                                       InElementOp,
+                                                                       WeiElementOp,
+                                                                       OutElementOp>;
 
-    using DeviceConvBwdWeightNoOpPtr =
-        ck::tensor_operation::device::DeviceConvBwdWeightPtr<PassThrough, PassThrough, PassThrough>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
 
-    // add device Conv instances
-    std::vector<DeviceConvBwdWeightNoOpPtr> conv_ptrs;
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
 
-    if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
-                 ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
-                 ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
-    {
-        ck::tensor_operation::device::instance::
-            add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
-    }
-    else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
-                      ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
-                      ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
-    {
-        ck::tensor_operation::device::instance::
-            add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
-    }
-
-    if(conv_ptrs.size() <= 0)
-    {
-        throw std::runtime_error("wrong! no device Conv instance found");
-    }
-
-    std::string best_conv_name;
-    float best_ave_time   = 0;
+    std::string best_op_name;
+    float best_avg_time   = 0;
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
 
     // profile device Conv instances
-    bool pass = true;
+    bool all_pass = true;
 
-    for(auto& conv_ptr : conv_ptrs)
-    {
-        // using atomic, so need to reset input
-        if(split_k > 1)
+    for(auto& op_ptr : op_ptrs)
     {
-            wei_device_buf.SetZero();
-        }
-
-        auto argument_ptr = conv_ptr->MakeArgumentPointer(
-            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                         static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                         static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-            N,
-            K,
-            C,
-            input_spatial_lengths,
-            filter_spatial_lengths,
-            output_spatial_lengths,
-            conv_filter_strides,
-            conv_filter_dilations,
-            input_left_pads,
-            input_right_pads,
+                                        conv_param.N_,
+                                        conv_param.K_,
+                                        conv_param.C_,
+                                        conv_param.input_spatial_lengths_,
+                                        conv_param.filter_spatial_lengths_,
+                                        conv_param.output_spatial_lengths_,
+                                        conv_param.conv_filter_strides_,
+                                        conv_param.conv_filter_dilations_,
+                                        conv_param.input_left_pads_,
+                                        conv_param.input_right_pads_,
                                         in_element_op,
                                         wei_element_op,
                                         out_element_op,
                                         split_k);
 
-        auto invoker_ptr = conv_ptr->MakeInvokerPointer();
-
-        if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
         {
-            std::string conv_name = conv_ptr->GetTypeString();
+            // using atomic add, so need to reset input
+            wei_device_buf.SetZero();
 
-            float ave_time =
-                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+            std::string op_name = op_ptr->GetTypeString();
 
-            std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
 
-            std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
-                                    sizeof(WeiDataType) * (K * C * Y * X) +
-                                    sizeof(OutDataType) * (N * K * Ho * Wo);
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
-            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
 
-            float gb_per_sec = num_btype / 1.E6 / ave_time;
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
 
-            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
-                      << " GB/s, " << conv_name << std::endl;
+            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_conv_name  = conv_name;
+                best_op_name    = op_name;
                 best_tflops     = tflops;
-                best_ave_time   = ave_time;
+                best_avg_time   = avg_time;
                 best_gb_per_sec = gb_per_sec;
             }
 
             if(do_verification)
             {
-                wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
+                wei_device_buf.FromDevice(weight_device_result.mData.data());
 
-                pass = ck::utils::check_err(wei_k_c_y_x_host_result.mData,
-                                            wei_k_c_y_x_device_result.mData);
+                bool pass =
+                    ck::utils::check_err(weight_host_result.mData, weight_device_result.mData);
 
-                if(pass == false)
+                if(!pass)
                 {
-                    std::cout << "Fail info:" << conv_ptr->GetTypeString() << std::endl;
+                    std::cout << "Fail info:" << op_ptr->GetTypeString() << std::endl;
                 }
 
+                all_pass &= pass;
+
                 if(do_log)
                 {
-                    LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",")
-                        << std::endl;
-                    LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",")
-                        << std::endl;
-                    LogRangeAsType<float>(
-                        std::cout << "wei_host  : ", wei_k_c_y_x_host_result.mData, ",")
-                        << std::endl;
-                    LogRangeAsType<float>(
-                        std::cout << "wei_device: ", wei_k_c_y_x_device_result.mData, ",")
-                        << std::endl;
+                    std::cout << "in : ";
+                    show_data_nhwc_layout(output);
+                    std::cout << std::endl;
+
+                    std::cout << "wei: ";
+                    show_data_nhwc_layout(weight_host_result);
+                    std::cout << std::endl;
+
+                    std::cout << "out  : ";
+                    show_data_nhwc_layout(input);
+                    std::cout << std::endl;
+
+                    std::cout << "wei_device: ";
+                    show_data_nhwc_layout(weight_device_result);
+                    std::cout << std::endl;
                 }
             }
         }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " 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_conv_name << std::endl;
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
 
-    return pass;
+    return all_pass;
 }
 
 } // namespace profiler

profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp

@@ -9,9 +9,9 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp"
 
 namespace ck {
@@ -157,12 +157,12 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
         ref_invoker.Run(ref_argument);
     }
 
-    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
-    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpaceSize());
     DeviceMem out_device_buf(sizeof(OutDataType) *
-                             out_n_k_ho_wo_device_result.mDesc.GetElementSpace());
-    DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpace());
-    DeviceMem resi_device_buf(sizeof(OutDataType) * resi_n_k_ho_wo.mDesc.GetElementSpace());
+                             out_n_k_ho_wo_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpaceSize());
+    DeviceMem resi_device_buf(sizeof(OutDataType) * resi_n_k_ho_wo.mDesc.GetElementSpaceSize());
 
     in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
     wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());

profiler/include/profile_conv_fwd_bias_relu_impl.hpp

@@ -9,9 +9,9 @@
 #include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp"
 
 namespace ck {
@@ -149,11 +149,11 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
         ref_invoker.Run(ref_argument);
     }
 
-    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
-    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpaceSize());
     DeviceMem out_device_buf(sizeof(OutDataType) *
-                             out_n_k_ho_wo_device_result.mDesc.GetElementSpace());
-    DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpace());
+                             out_n_k_ho_wo_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpaceSize());
 
     in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
     wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());

profiler/include/profile_conv_fwd_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/convolution_forward.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_conv_fwd_impl(int do_verification,
+                           int init_method,
+                           bool do_log,
+                           bool time_kernel,
+                           const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
+    Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    // run reference op
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>{};
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight,
+                                                  host_output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        // init host output to zero
+        host_output.SetZero();
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvFwd<NDimSpatial,
+                                                                 InLayout,
+                                                                 WeiLayout,
+                                                                 OutLayout,
+                                                                 InDataType,
+                                                                 WeiDataType,
+                                                                 OutDataType,
+                                                                 InElementOp,
+                                                                 WeiElementOp,
+                                                                 OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass = true;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        conv_param.N_,
+                                        conv_param.K_,
+                                        conv_param.C_,
+                                        conv_param.input_spatial_lengths_,
+                                        conv_param.filter_spatial_lengths_,
+                                        conv_param.GetOutputSpatialLengths(),
+                                        conv_param.conv_filter_strides_,
+                                        conv_param.conv_filter_dilations_,
+                                        conv_param.input_left_pads_,
+                                        conv_param.input_right_pads_,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init output to zero before profiling next kernel
+            out_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 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_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(device_output.mData.data());
+
+                pass = pass & ck::utils::check_err(device_output.mData, host_output.mData);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "weight: ", weight.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "host_output  : ", host_output.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profile_gemm_add_add_fastgelu_impl.hpp

@@ -13,9 +13,9 @@
 #include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -29,7 +29,9 @@ template <typename ADataType,
           typename EDataType,
           typename ALayout,
           typename BLayout,
-          typename DELayout> // assume Ds and E have same layout
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
 bool profile_gemm_add_add_fastgelu_impl(int do_verification,
                                         int init_method,
                                         bool /*do_log*/,
@@ -59,10 +61,10 @@ bool profile_gemm_add_add_fastgelu_impl(int do_verification,
 
     Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
     Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
-    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, DELayout{}));
-    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, DELayout{}));
-    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
-    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
 
     std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
     std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
@@ -100,7 +102,8 @@ bool profile_gemm_add_add_fastgelu_impl(int do_verification,
     using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
         ALayout,
         BLayout,
-        DELayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
         ADataType,
         BDataType,
         ck::Tuple<D0DataType, D1DataType>,
@@ -146,11 +149,11 @@ bool profile_gemm_add_add_fastgelu_impl(int do_verification,
         }
     }
 
-    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
-    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
-    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
-    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpace());
-    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
 
     a_device_buf.ToDevice(a_m_k.mData.data());
     b_device_buf.ToDevice(b_k_n.mData.data());

profiler/include/profile_gemm_bias_add_reduce_impl.hpp

@@ -10,10 +10,10 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/conv_util.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -217,15 +217,15 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
         }
     }
 
-    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
-    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
-    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
-    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpace());
-    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
     DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
-                                 reduce0_m_device_result.mDesc.GetElementSpace());
+                                 reduce0_m_device_result.mDesc.GetElementSpaceSize());
     DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
-                                 reduce1_m_device_result.mDesc.GetElementSpace());
+                                 reduce1_m_device_result.mDesc.GetElementSpaceSize());
 
     std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
                                       reduce1_device_buf.GetDeviceBuffer()};

profiler/include/profile_gemm_bilinear_impl.hpp

@@ -13,9 +13,9 @@
 #include "ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -28,7 +28,8 @@ template <typename ADataType,
           typename EDataType,
           typename ALayout,
           typename BLayout,
-          typename DELayout> // assume Ds and E have same layout
+          typename DLayout,
+          typename ELayout>
 bool profile_gemm_bilinear_impl(int do_verification,
                                 int init_method,
                                 bool /*do_log*/,
@@ -59,9 +60,9 @@ bool profile_gemm_bilinear_impl(int do_verification,
 
     Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
     Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
-    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DELayout{}));
-    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
-    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
 
     std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
     std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
@@ -96,7 +97,8 @@ bool profile_gemm_bilinear_impl(int do_verification,
     using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
         ALayout,
         BLayout,
-        DELayout,
+        ck::Tuple<DLayout>,
+        ELayout,
         ADataType,
         BDataType,
         ck::Tuple<DDataType>,
@@ -142,10 +144,10 @@ bool profile_gemm_bilinear_impl(int do_verification,
         }
     }
 
-    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
-    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
-    DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpace());
-    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
 
     a_device_buf.ToDevice(a_m_k.mData.data());
     b_device_buf.ToDevice(b_k_n.mData.data());

profiler/include/profile_gemm_impl.hpp

@@ -15,21 +15,21 @@
 #include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
 namespace profiler {
 
-template <typename ADataType,
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
           typename BDataType,
           typename AccDataType,
-          typename CDataType,
-          typename ALayout,
-          typename BLayout,
-          typename CLayout>
+          typename CDataType>
 int profile_gemm_impl(int do_verification,
                       int init_method,
                       bool do_log,
@@ -86,13 +86,12 @@ int profile_gemm_impl(int do_verification,
     const auto b_element_op = BElementOp{};
     const auto c_element_op = CElementOp{};
 
-    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
-    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
-    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
 
     a_device_buf.ToDevice(a_m_k.mData.data());
     b_device_buf.ToDevice(b_k_n.mData.data());
-    c_device_buf.ToDevice(c_m_n_device_result.mData.data());
 
     using DeviceOp = ck::tensor_operation::device::DeviceGemm<ALayout,
                                                               BLayout,
@@ -110,7 +109,7 @@ int profile_gemm_impl(int do_verification,
 
     std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
 
-    // Run reference GEMM
+    // Run reference op
     if(do_verification)
     {
         using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
@@ -131,11 +130,11 @@ int profile_gemm_impl(int do_verification,
     }
 
     std::string best_op_name;
-    float best_ave_time   = 0;
+    float best_avg_time   = 0;
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
 
-    // profile device GEMM instances
+    // profile device op instances
     for(auto& op_ptr : op_ptrs)
     {
         auto argument_ptr =
@@ -161,7 +160,7 @@ int profile_gemm_impl(int do_verification,
 
             std::string op_name = op_ptr->GetTypeString();
 
-            float ave_time =
+            float avg_time =
                 invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * M * N * K;
@@ -169,18 +168,18 @@ int profile_gemm_impl(int do_verification,
             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 tflops = static_cast<float>(flop) / 1.E9 / avg_time;
 
-            float gb_per_sec = num_btype / 1.E6 / ave_time;
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
 
-            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+            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_name    = op_name;
                 best_tflops     = tflops;
-                best_ave_time   = ave_time;
+                best_avg_time   = avg_time;
                 best_gb_per_sec = gb_per_sec;
             }
 
@@ -244,7 +243,7 @@ int profile_gemm_impl(int do_verification,
     }
 
     std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
-              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_ave_time
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_avg_time
               << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
               << best_op_name << std::endl;
 

profiler/include/profile_gemm_reduce_impl.hpp

@@ -10,10 +10,10 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/conv_util.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -189,13 +189,13 @@ bool profile_gemm_reduce_impl(int do_verification,
         }
     }
 
-    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
-    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
-    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
     DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
-                                 reduce0_m_device_result.mDesc.GetElementSpace());
+                                 reduce0_m_device_result.mDesc.GetElementSpaceSize());
     DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
-                                 reduce1_m_device_result.mDesc.GetElementSpace());
+                                 reduce1_m_device_result.mDesc.GetElementSpaceSize());
 
     std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
                                       reduce1_device_buf.GetDeviceBuffer()};

profiler/include/profile_gemm_splitk_impl.hpp

@@ -15,9 +15,9 @@
 #include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -87,9 +87,9 @@ bool profile_gemm_splitk_impl(int do_verification,
     const auto b_element_op = BElementOp{};
     const auto c_element_op = CElementOp{};
 
-    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
-    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
-    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
 
     a_device_buf.ToDevice(a_m_k.mData.data());
     b_device_buf.ToDevice(b_k_n.mData.data());

profiler/include/profile_grouped_conv_fwd_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_grouped_conv_fwd_impl(int do_verification,
+                                   int init_method,
+                                   bool do_log,
+                                   bool time_kernel,
+                                   const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
+    Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    // run reference op
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>{};
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight,
+                                                  host_output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        // init host output to zero
+        host_output.SetZero();
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 InElementOp,
+                                                                                 WeiElementOp,
+                                                                                 OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass = true;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
+                                        wei_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 0>{},
+                                        out_device_buf.GetDeviceBuffer(),
+                                        a_g_n_c_wis_lengths,
+                                        a_g_n_c_wis_strides,
+                                        b_g_k_c_xs_lengths,
+                                        b_g_k_c_xs_strides,
+                                        std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                        std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                        e_g_n_k_wos_lengths,
+                                        e_g_n_k_wos_strides,
+                                        conv_filter_strides,
+                                        conv_filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init output to zero before profiling next kernel
+            out_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 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_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(device_output.mData.data());
+
+                pass = pass & ck::utils::check_err(device_output.mData, host_output.mData);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "weight: ", weight.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "host_output  : ", host_output.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profile_grouped_gemm_impl.hpp

@@ -13,10 +13,10 @@
 #include "ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/conv_util.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -24,7 +24,7 @@ namespace profiler {
 
 template <typename ADataType,
           typename BDataType,
-          typename EDataType,
+          typename CDataType,
           typename AccDataType,
           typename ALayout,
           typename BLayout,
@@ -67,7 +67,7 @@ bool profile_grouped_gemm_impl(int do_verification,
 
     std::vector<Tensor<ADataType>> a_m_k;
     std::vector<Tensor<BDataType>> b_k_n;
-    std::vector<Tensor<EDataType>> c_m_n_device_results;
+    std::vector<Tensor<CDataType>> c_m_n_device_results;
 
     for(std::size_t i = 0; i < group_count; i++)
     {
@@ -77,7 +77,7 @@ bool profile_grouped_gemm_impl(int do_verification,
             Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
 
         c_m_n_device_results.push_back(
-            Tensor<EDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
 
         std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
                   << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
@@ -96,7 +96,7 @@ bool profile_grouped_gemm_impl(int do_verification,
             b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
         }
 
-        c_m_n_device_results[i].GenerateTensorValue(GeneratorTensor_0<EDataType>{}, num_thread);
+        c_m_n_device_results[i].GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
     }
 
     using AElementOp = ck::tensor_operation::element_wise::PassThrough;
@@ -133,12 +133,12 @@ bool profile_grouped_gemm_impl(int do_verification,
     for(std::size_t i = 0; i < group_count; i++)
     {
         a_device_buf.emplace_back(
-            std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpace()));
+            std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
         b_device_buf.emplace_back(
-            std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpace()));
+            std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
 
         c_device_buf.emplace_back(std::make_unique<DeviceMem>(
-            sizeof(EDataType) * c_m_n_device_results[i].mDesc.GetElementSpace()));
+            sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
 
         a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
         b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
@@ -153,11 +153,12 @@ bool profile_grouped_gemm_impl(int do_verification,
 
     using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
                                                                      BLayout,
+                                                                     ck::Tuple<>,
                                                                      CLayout,
                                                                      ADataType,
                                                                      BDataType,
                                                                      ck::Tuple<>,
-                                                                     EDataType,
+                                                                     CDataType,
                                                                      AElementOp,
                                                                      BElementOp,
                                                                      CElementOp>;
@@ -209,7 +210,7 @@ bool profile_grouped_gemm_impl(int do_verification,
                 flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
 
                 num_btype += sizeof(ADataType) * Ms[i] * Ks[i] + sizeof(BDataType) * Ks[i] * Ns[i] +
-                             sizeof(EDataType) * Ms[i] * Ns[i];
+                             sizeof(CDataType) * Ms[i] * Ns[i];
             }
 
             float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -233,13 +234,13 @@ bool profile_grouped_gemm_impl(int do_verification,
 
                     c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
 
-                    Tensor<EDataType> c_m_n_host_result(
+                    Tensor<CDataType> c_m_n_host_result(
                         f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{}));
 
                     using ReferenceGemmInstance =
                         ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                   BDataType,
-                                                                  EDataType,
+                                                                  CDataType,
                                                                   AccDataType,
                                                                   AElementOp,
                                                                   BElementOp,

profiler/include/profile_normalization_impl.hpp

@@ -9,10 +9,10 @@
 #include "ck/tensor_operation/gpu/device/device_softmax.hpp"
 
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/conv_util.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_tensor.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
 
 namespace ck {
@@ -92,8 +92,8 @@ void profile_normalization_impl(int do_verification,
 
     Tensor<OutDataType> out_ref(out);
 
-    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
-    DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
+    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
     in_dev.ToDevice(in.mData.data());
     out_dev.ToDevice(out.mData.data());
 

profiler/include/profile_reduce_impl.hpp

@@ -8,10 +8,10 @@
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp"
-#include "ck/library/host_tensor/device_memory.hpp"
-#include "ck/library/host_tensor/host_reduction.hpp"
-#include "ck/library/host_tensor/host_common_util.hpp"
-#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_reduction.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -245,13 +245,13 @@ bool profile_reduce_impl_impl(bool do_verification,
             }
 
             if(beta != 0.0f)
-                for(size_t i = 0; i < out_ref.mDesc.GetElementSpace(); i++)
+                for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
                     out.mData[i] = out_ref.mData[i];
         };
 
         // these buffers are usually provided by the user application
-        DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
-        DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
+        DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+        DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
 
         in_dev.ToDevice(in.mData.data());
 

profiler/src/profile_batched_gemm_reduce.cpp

@@ -24,9 +24,9 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
         F16_F16_F16_F32_F32, // 1
     };
 
-    if(!(argc == 15 || argc == 16))
+    if(argc != 15)
     {
-        printf("arg1: tensor operation (batched_gemm: BatchedGEMM+Reduce)\n");
+        printf("arg1: tensor operation (batched_gemm_reduce: BatchedGEMM+Reduce)\n");
         printf("arg2: data type (0: fp32; 1: fp16)\n");
         printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
         printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
@@ -37,7 +37,6 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
         printf("arg6: print tensor value (0: no; 1: yes)\n");
         printf("arg7: time kernel (0=n0, 1=yes)\n");
         printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
-        printf("arg15: split k into  mulitiple batch\n");
         exit(1);
     }
 

profiler/src/profile_conv_bwd_data.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/include/profile_conv_bwd_data_impl.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    NCHW_KCYX_NKHW, // 0
+    NHWC_KYXC_NHWK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+static void print_helper_msg()
+{
+    std::cout
+        << "arg1: tensor operation (conv_bwd_data: Convolution Backward Data)\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
+        << "                     1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, "
+           "K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+} // namespace
+
+int profile_conv_bwd_data(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using NWC   = ck::tensor_layout::convolution::NWC;
+    using NHWC  = ck::tensor_layout::convolution::NHWC;
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    using KXC   = ck::tensor_layout::convolution::KXC;
+    using KYXC  = ck::tensor_layout::convolution::KYXC;
+    using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+    using NWK   = ck::tensor_layout::convolution::NWK;
+    using NHWK  = ck::tensor_layout::convolution::NHWK;
+    using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_conv_bwd_data_impl<NDimSpatial,
+                                                             InLayout,
+                                                             WeiLayout,
+                                                             OutLayout,
+                                                             InDataType,
+                                                             WeiDataType,
+                                                             OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}

profiler/src/profile_conv_bwd_weight.cpp

@@ -8,141 +8,168 @@
 
 #include "profiler/include/profile_conv_bwd_weight_impl.hpp"
 
-enum struct ConvDataType
-{
-    F32_F32_F32,    // 0
-    F16_F16_F16,    // 1
-    BF16_BF16_BF16, // 2
-    INT8_INT8_INT8, // 3
-};
+namespace {
 
-enum struct ConvInputLayout
+enum struct ConvLayout
 {
-    NCHW, // 0
-    NHWC, // 1
+    NCHW_KCYX_NKHW, // 0
+    NHWC_KYXC_NHWK, // 1
 };
 
-enum struct ConvWeightLayout
+enum struct ConvDataType
 {
-    KCYX, // 0
-    KYXC, // 1
+    F32_F32_F32,   // 0
+    F16_F16_F16,   // 1
+    BF16_F32_BF16, // 2
 };
 
-enum struct ConvOutputLayout
+static void print_helper_msg()
 {
-    NKHW, // 0
-    NHWK, // 1
-};
+    std::cout
+        << "arg1: tensor operation (conv_bwd_weight: Convolution Backward Weight\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight fp32, Output bf16)\n"
+        << "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
+        << "                     1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, K]\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << " SplitK\n"
+        << std::endl;
+}
+
+} // namespace
 
 int profile_conv_bwd_weight(int argc, char* argv[])
 {
-    if(argc != 26)
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
     {
-        printf("arg1: tensor operation (conv_fwd: ForwardConvolution)\n");
-        printf("arg2: data type (0: fp32; 1: fp16)\n");
-        printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
-        printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
-        printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
-        printf("arg6: verification (0: no; 1: yes)\n");
-        printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg8: print tensor value (0: no; 1: yes)\n");
-        printf("arg9: run kernel # of times (>1)\n");
-        printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
-               "RightPx\n");
-        printf("arg25: split k (>=1)\n");
-        exit(1);
+        print_helper_msg();
+        return 1;
     }
 
     const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
-    const auto in_layout       = static_cast<ConvInputLayout>(std::stoi(argv[3]));
-    const auto wei_layout      = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
-    const auto out_layout      = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
-    const bool do_verification = std::stoi(argv[6]);
-    const int init_method      = std::stoi(argv[7]);
-    const bool do_log          = std::stoi(argv[8]);
-    const bool time_kernel     = std::stoi(argv[9]);
-
-    const ck::index_t N  = std::stoi(argv[10]);
-    const ck::index_t K  = std::stoi(argv[11]);
-    const ck::index_t C  = std::stoi(argv[12]);
-    const ck::index_t Y  = std::stoi(argv[13]);
-    const ck::index_t X  = std::stoi(argv[14]);
-    const ck::index_t Hi = std::stoi(argv[15]);
-    const ck::index_t Wi = std::stoi(argv[16]);
-
-    const ck::index_t conv_stride_h   = std::stoi(argv[17]);
-    const ck::index_t conv_stride_w   = std::stoi(argv[18]);
-    const ck::index_t conv_dilation_h = std::stoi(argv[19]);
-    const ck::index_t conv_dilation_w = std::stoi(argv[20]);
-    const ck::index_t in_left_pad_h   = std::stoi(argv[21]);
-    const ck::index_t in_left_pad_w   = std::stoi(argv[22]);
-    const ck::index_t in_right_pad_h  = std::stoi(argv[23]);
-    const ck::index_t in_right_pad_w  = std::stoi(argv[24]);
-    ck::index_t split_k               = std::stoi(argv[25]);
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial, 1 for split-K
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial + 1)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    ck::index_t split_k = std::stoi(argv[8 + 1 + 4 + 6 * num_dim_spatial]);
     split_k             = std::max(1, split_k);
 
-    const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
-    const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+
+    using NWC   = ck::tensor_layout::convolution::NWC;
+    using NHWC  = ck::tensor_layout::convolution::NHWC;
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    using KXC   = ck::tensor_layout::convolution::KXC;
+    using KYXC  = ck::tensor_layout::convolution::KYXC;
+    using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+    using NWK   = ck::tensor_layout::convolution::NWK;
+    using NHWK  = ck::tensor_layout::convolution::NHWK;
+    using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
 
-    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
-    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
 
-    if(data_type == ConvDataType::F32_F32_F32 && in_layout == ConvInputLayout::NHWC &&
-       wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
+        bool pass = ck::profiler::profile_conv_bwd_weight_impl<NDimSpatial,
+                                                               InLayout,
+                                                               WeiLayout,
+                                                               OutLayout,
+                                                               InDataType,
+                                                               WeiDataType,
+                                                               OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params, split_k);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I1, NWC{}, KXC{}, NWK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
         {
-        ck::profiler::profile_conv_bwd_weight_impl<2,
-                                                   float,
-                                                   float,
-                                                   float,
-                                                   ck::tensor_layout::convolution::NHWC,
-                                                   ck::tensor_layout::convolution::KYXC,
-                                                   ck::tensor_layout::convolution::NHWK>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            N,
-            K,
-            C,
-            std::vector<ck::index_t>{Hi, Wi},
-            std::vector<ck::index_t>{Y, X},
-            std::vector<ck::index_t>{Ho, Wo},
-            std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
-            std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
-            std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
-            std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w},
-            split_k);
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F32{}, F32{}, F32{});
         }
-    else if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
-            wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
+        else if(data_type == ConvDataType::F16_F16_F16)
         {
-        ck::profiler::profile_conv_bwd_weight_impl<2,
-                                                   ck::half_t,
-                                                   ck::half_t,
-                                                   ck::half_t,
-                                                   ck::tensor_layout::convolution::NHWC,
-                                                   ck::tensor_layout::convolution::KYXC,
-                                                   ck::tensor_layout::convolution::NHWK>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            N,
-            K,
-            C,
-            std::vector<ck::index_t>{Hi, Wi},
-            std::vector<ck::index_t>{Y, X},
-            std::vector<ck::index_t>{Ho, Wo},
-            std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
-            std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
-            std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
-            std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w},
-            split_k);
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F16{}, F16{}, F16{});
         }
-    else
+        else if(data_type == ConvDataType::BF16_F32_BF16)
         {
-        throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, BF16{}, F32{}, BF16{});
         }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
 
-    return 0;
+    return 1;
 }

profiler/src/profile_conv_fwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/include/profile_conv_fwd_impl.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    NCHW_KCYX_NKHW, // 0
+    NHWC_KYXC_NHWK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+static void print_helper_msg()
+{
+    std::cout
+        // clang-format-off
+        << "arg1: tensor operation (conv_fwd: Convolution Forward)\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
+        << "                     1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, "
+           "K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+    // clang-format-on
+}
+
+} // namespace
+
+int profile_conv_fwd(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using NWC   = ck::tensor_layout::convolution::NWC;
+    using NHWC  = ck::tensor_layout::convolution::NHWC;
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    using KXC   = ck::tensor_layout::convolution::KXC;
+    using KYXC  = ck::tensor_layout::convolution::KYXC;
+    using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+    using NWK   = ck::tensor_layout::convolution::NWK;
+    using NHWK  = ck::tensor_layout::convolution::NHWK;
+    using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_conv_fwd_impl<NDimSpatial,
+                                                        InLayout,
+                                                        WeiLayout,
+                                                        OutLayout,
+                                                        InDataType,
+                                                        WeiDataType,
+                                                        OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}