Add groupnorm ckProfiler

library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f16_instance.cpp

@@ -15,36 +15,43 @@ namespace instance {
 using F16 = ck::half_t;
 using F32 = float;
 
-using Pass = ck::tensor_operation::element_wise::PassThrough;
+using Pass    = ck::tensor_operation::element_wise::PassThrough;
+using Sigmoid = ck::tensor_operation::element_wise::Sigmoid;
 
-template <index_t Rank, index_t Reduce>
+template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_layernorm_f16_instances = std::tuple<
     // clang-format off
         // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>
     // clang-format on
     >;
 
 void add_device_layernorm_f16_rank2_instances(
     std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Pass, 2, 1>>& instances)
 {
-    add_device_operation_instances(instances, device_layernorm_f16_instances<2, 1>{});
+    add_device_operation_instances(instances, device_layernorm_f16_instances<Pass, 2, 1>{});
 }
 
 void add_device_layernorm_f16_rank4_instances(
     std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Pass, 4, 3>>& instances)
 {
-    add_device_operation_instances(instances, device_layernorm_f16_instances<4, 3>{});
+    add_device_operation_instances(instances, device_layernorm_f16_instances<Pass, 4, 3>{});
+}
+
+void add_device_groupnorm_f16_instances(
+    std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Sigmoid, 5, 3>>& instances)
+{
+    add_device_operation_instances(instances, device_layernorm_f16_instances<Sigmoid, 5, 3>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f32_instance.cpp

@@ -14,35 +14,42 @@ namespace instance {
 
 using F32 = float;
 
-using Pass = ck::tensor_operation::element_wise::PassThrough;
+using Pass    = ck::tensor_operation::element_wise::PassThrough;
+using Sigmoid = ck::tensor_operation::element_wise::Sigmoid;
 
-template <index_t Rank, index_t Reduce>
+template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_layernorm_f32_instances = std::tuple<
     // clang-format off
         // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>
     // clang-format on
     >;
 
 void add_device_layernorm_f32_rank2_instances(
     std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Pass, 2, 1>>& instances)
 {
-    add_device_operation_instances(instances, device_layernorm_f32_instances<2, 1>{});
+    add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 2, 1>{});
 }
 
 void add_device_layernorm_f32_rank4_instances(
     std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Pass, 4, 3>>& instances)
 {
-    add_device_operation_instances(instances, device_layernorm_f32_instances<4, 3>{});
+    add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 4, 3>{});
+}
+
+void add_device_groupnorm_f32_instances(
+    std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Sigmoid, 5, 3>>& instances)
+{
+    add_device_operation_instances(instances, device_layernorm_f32_instances<Sigmoid, 5, 3>{});
 }
 
 } // namespace instance

profiler/CMakeLists.txt

@@ -23,6 +23,7 @@ set(PROFILER_SOURCE
     src/profile_conv_bwd_weight.cpp
     src/profile_grouped_conv_fwd.cpp
     src/profile_reduce.cpp
+    src/profile_groupnorm.cpp
     src/profile_layernorm.cpp
     src/profile_normalization.cpp
 )
@@ -55,3 +56,4 @@ target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
 target_link_libraries(ckProfiler PRIVATE device_normalization_instance)
 target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
+

profiler/include/profile_groupnorm_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "profiler/include/data_type_enum.hpp"
+#include "ck/tensor_operation/gpu/device/device_layernorm_impl.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/reference_tensor_operation/cpu/reference_groupnorm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16     = ck::half_t;
+using F32     = float;
+using Sigmoid = ck::tensor_operation::element_wise::Sigmoid;
+
+void add_device_groupnorm_f16_instances(
+    std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Sigmoid, 5, 3>>&);
+
+void add_device_groupnorm_f32_instances(
+    std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Sigmoid, 5, 3>>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+enum struct ElementwiseOpEnum
+{
+    ePassthrough = 0,
+    eSigmoid     = 1
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType>
+void profile_groupnorm_impl(int do_verification,
+                            int init_method,
+                            bool do_log,
+                            bool time_kernel,
+                            std::vector<index_t> length,
+                            ElementwiseOpEnum OutelementwiseOp)
+{
+    using F16     = ck::half_t;
+    using F32     = float;
+    using Sigmoid = ck::tensor_operation::element_wise::Sigmoid;
+
+    if(length.size() != 5)
+        return;
+
+    index_t G = length[3];
+    index_t C = length[4];
+
+    std::vector<index_t> reduce_dim      = {1, 2, 4};
+    std::vector<index_t> gammaBetaLength = {G, C};
+    std::vector<index_t> gammaBetaStride = {0, 0, 0, C, 1};
+
+    Tensor<XDataType> x(length);
+    Tensor<GammaDataType> gamma(gammaBetaLength);
+    Tensor<BetaDataType> beta(gammaBetaLength);
+    Tensor<YDataType> y(length);
+    Tensor<YDataType> host_y(length);
+
+    switch(init_method)
+    {
+    case 0:
+        x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
+        beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
+        break;
+    case 1:
+        x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
+        gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+        beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+        break;
+    default:
+        x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    // add device normalization instances
+    std::vector<tensor_operation::device::DeviceLayernormPtr<XDataType,
+                                                             GammaDataType,
+                                                             BetaDataType,
+                                                             AccDataType,
+                                                             YDataType,
+                                                             Sigmoid,
+                                                             5,
+                                                             3>>
+        instances;
+
+    if constexpr(is_same<XDataType, F16>::value && is_same<GammaDataType, F16>::value &&
+                 is_same<BetaDataType, F16>::value && is_same<YDataType, F16>::value &&
+                 is_same<AccDataType, F32>::value)
+    {
+        if(OutelementwiseOp == ElementwiseOpEnum::eSigmoid)
+            tensor_operation::device::instance::add_device_groupnorm_f16_instances(instances);
+    }
+    else if constexpr(is_same<XDataType, F32>::value && is_same<GammaDataType, F32>::value &&
+                      is_same<BetaDataType, F32>::value && is_same<YDataType, F32>::value &&
+                      is_same<AccDataType, F32>::value)
+    {
+        if(OutelementwiseOp == ElementwiseOpEnum::eSigmoid)
+            tensor_operation::device::instance::add_device_groupnorm_f32_instances(instances);
+    }
+
+    if(instances.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device normalization instance found");
+    }
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        if(OutelementwiseOp == ElementwiseOpEnum::eSigmoid)
+        {
+            using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
+                                                                                     GammaDataType,
+                                                                                     BetaDataType,
+                                                                                     YDataType,
+                                                                                     AccDataType,
+                                                                                     Sigmoid>;
+
+            ReferenceInstance ref;
+            auto ref_argument = ref.MakeArgument(x, gamma, beta, host_y, Sigmoid{}, length, 1e-6);
+            auto ref_invoker  = ref.MakeInvoker();
+            ref_invoker.Run(ref_argument);
+        }
+    }
+
+    for(auto& inst_ptr : instances)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            length,
+            std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
+            gammaBetaStride,
+            gammaBetaStride,
+            std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+            reduce_dim,
+            1e-6,
+            x_dev.GetDeviceBuffer(),
+            gamma_dev.GetDeviceBuffer(),
+            beta_dev.GetDeviceBuffer(),
+            y_dev.GetDeviceBuffer(),
+            Sigmoid{});
+
+        if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+            LogRange(std::cout << "input lengths = [", length, "], ") << std::endl;
+
+            return;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = x.mDesc.GetElementSize() * sizeof(XDataType) +
+                                gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
+                                beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
+                                y.mDesc.GetElementSize() * sizeof(YDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                  << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            y_dev.FromDevice(y.mData.data());
+
+            bool pass =
+                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "x  : ", x.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "host_y  : ", host_y.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "y  : ", y.mData, ",") << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
+                return;
+            }
+            else
+            {
+                std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    LogRange(std::cout << "length = ", length, ",") << ", ";
+    std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_instance_name << std::endl;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/src/profile_groupnorm.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/include/profile_groupnorm_impl.hpp"
+
+using ck::index_t;
+using ck::profiler::ElementwiseOpEnum;
+
+struct GroupnormArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_groupnorm()
+{
+    std::cout << "arg1: tensor operation (groupnorm: Group normalization)\n"
+              << "arg2: data type (0: fp16; 1: fp32)\n"
+              << "arg3: verification (0: no; 1: yes)\n"
+              << "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg5: print tensor value (0: no; 1: yes)\n"
+              << "arg6: time kernel (0=n0, 1=yes)\n"
+              << "arg7: out elementwise op (0=passthrough, 1=sigmoid)\n"
+              << "--length: tensor extents (e.g, --length 1024 1024) \n"
+              << std::endl;
+}
+
+int profile_groupnorm(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type         = ck::DataTypeEnum::Half;
+    bool do_verification               = false;
+    int init_method                    = 0;
+    bool do_log                        = 0;
+    bool time_kernel                   = 1;
+    ElementwiseOpEnum outElementwiseOp = ElementwiseOpEnum::eSigmoid;
+    std::vector<index_t> length        = {1, 16, 16, 32, 40};
+
+    if(argc != 1 && argc != 14)
+    {
+        print_help_groupnorm();
+        return 0;
+    }
+
+    if(argc == 14)
+    {
+        data_type        = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification  = std::stoi(argv[3]);
+        init_method      = std::stoi(argv[4]);
+        do_log           = std::stoi(argv[5]);
+        time_kernel      = std::stoi(argv[6]);
+        outElementwiseOp = static_cast<ElementwiseOpEnum>(std::stoi(argv[7]));
+
+        // parse the long options
+        GroupnormArgParser arg_parser;
+        arg_parser(argc, argv);
+        length = arg_parser.long_opts["length"];
+    }
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    if(data_type == ck::DataTypeEnum::Half && outElementwiseOp == ElementwiseOpEnum::eSigmoid)
+    {
+        ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16>(
+            do_verification, init_method, do_log, time_kernel, length, outElementwiseOp);
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+// hijack main() for quick debugging
+// int main(int argc, char* argv[])
+// {
+//     profile_groupnorm(argc, argv);
+//     return 0;
+// }

profiler/src/profiler.cpp

@@ -136,6 +136,10 @@ int main(int argc, char* argv[])
     {
         return profile_layernorm(argc, argv);
     }
+    else if(strcmp(argv[1], "groupnorm") == 0)
+    {
+        return profile_groupnorm(argc, argv);
+    }
     else
     {
         print_helper_message();