Merge branch 'add_contraction_example_fp64' of...

Merge branch 'add_contraction_example_fp64' of https://github.com/ROCmSoftwarePlatform/composable_kernel into add_contraction_example_fp64

profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.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/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasesDataType,
+          typename Acc1BiasesDataType,
+          tensor_operation::device::MaskingSpecialization MaskingSpec>
+bool profile_batched_gemm_bias_softmax_gemm_permute_impl(bool do_verification,
+                                                         int init_method,
+                                                         bool do_log,
+                                                         bool time_kernel,
+                                                         int M,
+                                                         int N,
+                                                         int K,
+                                                         int O,
+                                                         int G0,
+                                                         int G1,
+                                                         float alpha = -1.f)
+
+{
+
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using ScaleAdd      = tensor_operation::element_wise::ScaleAdd;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using C0DEElementOp = ScaleAdd;
+    using Acc0ElementOp = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+    using D0DataType    = tuple_element_t<0, Acc0BiasesDataType>;
+    using tensor_operation::device::MaskingSpecialization;
+
+    // Ref Gemm0: various type in, fp32 out
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+    // Ref Softmax: fp32 in, various type out
+    using ReferenceSoftmaxInstance =
+        tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+    // Ref Gemm1: various type in, various type out
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    const int BatchCount = G0 * G1;
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<D0DataType> d0_gs_ms_ns(d0_gs_ms_ns_lengths, d0_gs_ms_ns_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    std::srand(1); // work around test flakiness
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        // Still unsure whether this kind of deterministic floating point accurary issue is expected
+        // or not. May want to try exact same approach as the GPU kernel in the host reference
+        // GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
+        // shrink the input value range as it is less likely to produce errors of around ~1e-3.
+        // a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        // b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+    d0_device_buf.ToDevice(d0_gs_ms_ns.mData.data());
+
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto c0de_element_op = C0DEElementOp{alpha};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp =
+        tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      Acc0BiasesDataType,
+                                                                      ck::Tuple<>,
+                                                                      AElementOp,
+                                                                      B0ElementOp,
+                                                                      C0DEElementOp,
+                                                                      B1ElementOp,
+                                                                      CElementOp,
+                                                                      MaskingSpec>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+        Tensor<D0DataType> d0_g_m_n({BatchCount, M, N});
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        d0_gs_ms_ns.ForEach([&](auto& self, auto idx) {
+            d0_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        acc0_g_m_n.ForEach([&](auto&, auto idx) {
+            c0de_element_op(acc0_g_m_n(idx), acc0_g_m_n(idx), d0_g_m_n(idx));
+        });
+        // mask out upper triangle
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle && idx[1] < idx[2])
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+            std::array<void*, 1>{
+                d0_device_buf.GetDeviceBuffer()}, // std::array<void*, 1> p_acc0_biases;
+            {},                                   // std::array<void*, 1> p_acc1_biases;
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // std::array<std::vector<ck::index_t>,
+                                      // 1>{acc0_biases_gs_ms_ns_strides},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+            a_element_op,
+            b0_element_op,
+            c0de_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O +
+                                     sizeof(D0DataType) * M * N) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+                // default absolute error and relative error is 0.001
+                double rtol = 1e-3;
+                double atol = 1e-3;
+
+                // when BF16 is taken, set absolute error and relative error to 0.01
+                if(std::is_same_v<ADataType, ck::bhalf_t> &&
+                   std::is_same_v<B0DataType, ck::bhalf_t> &&
+                   std::is_same_v<B1DataType, ck::bhalf_t> &&
+                   std::is_same_v<CDataType, ck::bhalf_t> &&
+                   std::is_same_v<D0DataType, ck::bhalf_t>)
+                {
+                    rtol = 1e-2;
+                    atol = 1e-2;
+                }
+
+                pass = pass & ck::utils::check_err(c_gs_ms_os_device_result,
+                                                   c_gs_ms_os_host_result,
+                                                   "Error: Incorrect results!",
+                                                   rtol,
+                                                   atol);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_gs_ms_ks: ", a_gs_ms_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_gs_ns_ks : ", b0_gs_ns_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_gs_os_ns : ", b1_gs_os_ns.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_gs_ms_os_host_result : ", c_gs_ms_os_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_gs_ms_os_device_result : ",
+                                          c_gs_ms_os_device_result.mData,
+                                          ",")
+                        << 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_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_gemm_add_relu_add_layernorm_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 "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.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/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EMeanVarDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename HElementOp>
+void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<BDataType>& b_k_n,
+                         const Tensor<D0DataType>& d0_m_n,
+                         const Tensor<D1DataType>& d1_m_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<BetaDataType>& beta_n,
+                         AElementOp a_element_op,
+                         BElementOp b_element_op,
+                         CDEElementOp cde_element_op,
+                         HElementOp h_element_op,
+                         int M,
+                         int N,
+                         AccDataType epsilon = 1e-5)
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    using ReferenceGemm = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                    BDataType,
+                                                                    AccDataType,
+                                                                    AccDataType,
+                                                                    AElementOp,
+                                                                    BElementOp,
+                                                                    PassThrough>;
+
+    using ReferenceLayernorm = ck::tensor_operation::host::ReferenceLayernorm<EMeanVarDataType,
+                                                                              GammaDataType,
+                                                                              BetaDataType,
+                                                                              HDataType,
+                                                                              AccDataType,
+                                                                              HElementOp,
+                                                                              2,
+                                                                              1>;
+
+    Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
+    Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
+
+    auto ref_gemm         = ReferenceGemm{};
+    auto ref_gemm_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_gemm_argument =
+        ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+    ref_gemm_invoker.Run(ref_gemm_argument);
+
+    for(int n = 0; n < N; ++n)
+    {
+        for(int m = 0; m < M; ++m)
+        {
+            AccDataType e  = static_cast<AccDataType>(e_m_n(m, n));
+            AccDataType d0 = static_cast<AccDataType>(d0_m_n(m, n));
+            AccDataType d1 = static_cast<AccDataType>(d1_m_n(m, n));
+            cde_element_op(e, c_m_n(m, n), d0, d1);
+            e_m_n(m, n) = static_cast<EMeanVarDataType>(e);
+        }
+    }
+
+    ReferenceLayernorm ref_layernorm;
+    auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
+
+    auto ref_layernorm_argument = ref_layernorm.MakeArgument(
+        e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
+    ref_layernorm_invoker.Run(ref_layernorm_argument);
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EMeanVarDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename HLayout>
+bool profile_gemm_add_relu_add_layernorm_impl(int do_verification,
+                                              int init_method,
+                                              bool /*do_log*/,
+                                              bool time_kernel,
+                                              int M,
+                                              int N,
+                                              int K,
+                                              int StrideA,
+                                              int StrideB,
+                                              int StrideD0,
+                                              int StrideD1,
+                                              int StrideH,
+                                              AccDataType epsilon = 1e-5)
+{
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(std::is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<D1DataType> d0_m_n(f_host_tensor_descriptor2d(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, D1Layout{}));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<HDataType> h_m_n(f_host_tensor_descriptor2d(M, N, StrideH, HLayout{}));
+    Tensor<HDataType> h_m_n_host(f_host_tensor_descriptor2d(M, N, StrideH, HLayout{}));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+        gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+        beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+        break;
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddReluAdd;
+    using HElementOp   = PassThrough;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+    const auto h_element_op   = HElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        HLayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        GammaDataType,
+        BetaDataType,
+        HDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddReluAdd,
+        ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        host_gemm_layernorm<ADataType,
+                            BDataType,
+                            AccDataType,
+                            D0DataType,
+                            D1DataType,
+                            EMeanVarDataType,
+                            GammaDataType,
+                            BetaDataType,
+                            HDataType>(h_m_n_host,
+                                       a_m_k,
+                                       b_k_n,
+                                       d0_m_n,
+                                       d1_m_n,
+                                       gamma_n,
+                                       beta_n,
+                                       a_element_op,
+                                       b_element_op,
+                                       cde_element_op,
+                                       h_element_op,
+                                       M,
+                                       N,
+                                       epsilon);
+    }
+
+    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 gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
+    DeviceMem h_device_buf(sizeof(HDataType) * h_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    bool pass      = true;
+    int num_kernel = 0;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_m_n_device_buf.GetDeviceBuffer(), d1_m_n_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            DeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+            // re-init E to zero before profiling a kernel
+            h_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                (sizeof(D0DataType) + sizeof(D1DataType) + sizeof(HDataType)) * M * N +
+                (sizeof(GammaDataType) + sizeof(BetaDataType)) * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            if(time_kernel)
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
+                          << " GB/s, " << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                h_device_buf.FromDevice(h_m_n.mData.data());
+
+                pass = pass && ck::utils::check_err(
+                                   h_m_n, h_m_n_host, "Error: Incorrect results h_m_n", 1e-2, 1e-2);
+            }
+        }
+        else
+        {
+            if(time_kernel)
+                std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        pass = false;
+    }
+    else
+    {
+        if(time_kernel)
+            std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                      << best_op_name << std::endl;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/src/CMakeLists.txt

@@ -8,6 +8,7 @@ set(PROFILER_SOURCES
     profile_gemm_add_add_fastgelu.cpp
     profile_gemm_add_multiply.cpp
     profile_gemm_add_fastgelu.cpp
+    profile_gemm_add_relu_add_layernorm.cpp
     profile_gemm_fastgelu.cpp
     profile_gemm_reduce.cpp
     profile_batched_gemm.cpp
@@ -43,6 +44,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgel
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
@@ -66,5 +68,4 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instan
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
-
 rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)

profiler/src/profile_gemm_add_relu_add_layernorm.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/profile_gemm_add_relu_add_layernorm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add_relu_add_layernorm"
+#define OP_DESC "GEMM+Add+Relu+Add+Layernorm"
+
+int profile_gemm_add_relu_add_layernorm(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN_MN, // 0
+        MK_NK_MN_MN_MN, // 1
+        KM_KN_MN_MN_MN, // 2
+        KM_NK_MN_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32,  // 0
+        F16,  // 1
+        BF16, // 2
+    };
+
+    if(argc != 16)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16)\n");
+        printf("arg3: matrix layout (0: H[m, n] = Layernorm(Relu(A[m, k] * B[k, n] + D0[m, n]) + D1[m, n]);\n");
+        printf("                     1: H[m, n] = Layernorm(Relu(A[m, k] * B[n, k] + D0[m, n]) + D1[m, n]);\n");
+        printf("                     2: H[m, n] = Layernorm(Relu(A[k, m] * B[k, n] + D0[m, n]) + D1[m, n]);\n");
+        printf("                     3: H[m, n] = Layernorm(Relu(A[k, m] * B[n, k] + D0[m, n]) + D1[m, n]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideH\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(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 M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideD1 = std::stoi(argv[14]);
+    const int StrideH  = std::stoi(argv[15]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto d1_type,
+                       auto e_mean_var_type,
+                       auto gamma_type,
+                       auto beta_type,
+                       auto h_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto d1_layout,
+                       auto h_layout) {
+        using ADataType        = decltype(a_type);
+        using BDataType        = decltype(b_type);
+        using AccDataType      = decltype(acc_type);
+        using D0DataType       = decltype(d0_type);
+        using D1DataType       = decltype(d1_type);
+        using EMeanVarDataType = decltype(e_mean_var_type);
+        using GammaDataType    = decltype(gamma_type);
+        using BetaDataType     = decltype(beta_type);
+        using HDataType        = decltype(h_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using D1Layout = decltype(d1_layout);
+        using HLayout  = decltype(h_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+        const int DefaultStrideH  = ck::is_same_v<HLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_add_relu_add_layernorm_impl<ADataType,
+                                                                           BDataType,
+                                                                           AccDataType,
+                                                                           D0DataType,
+                                                                           D1DataType,
+                                                                           EMeanVarDataType,
+                                                                           GammaDataType,
+                                                                           BetaDataType,
+                                                                           HDataType,
+                                                                           ALayout,
+                                                                           BLayout,
+                                                                           D0Layout,
+                                                                           D1Layout,
+                                                                           HLayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
+            (StrideH < 0) ? DefaultStrideH : StrideH);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
+    {
+        return profile(F16{},
+                       F16{},
+                       F32{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       Row{},
+                       Row{},
+                       Row{},
+                       Row{},
+                       Row{});
+    }
+    else if(data_type == MatrixDataType::F16 && layout == MatrixLayout::MK_NK_MN_MN_MN)
+    {
+        return profile(F16{},
+                       F16{},
+                       F32{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       Row{},
+                       Col{},
+                       Row{},
+                       Row{},
+                       Row{});
+    }
+    else if(data_type == MatrixDataType::F16 && layout == MatrixLayout::KM_KN_MN_MN_MN)
+    {
+        return profile(F16{},
+                       F16{},
+                       F32{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       Col{},
+                       Row{},
+                       Row{},
+                       Row{},
+                       Row{});
+    }
+    else if(data_type == MatrixDataType::F16 && layout == MatrixLayout::KM_NK_MN_MN_MN)
+    {
+        return profile(F16{},
+                       F16{},
+                       F32{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       Col{},
+                       Col{},
+                       Row{},
+                       Row{},
+                       Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_relu_add_layernorm);

script/process_perf_data.py

@@ -3,6 +3,7 @@ import os, io, argparse, datetime
 #import numpy as np
 import sqlalchemy
 from sqlalchemy.types import NVARCHAR, Float, Integer
+from sqlalchemy import text
 import pymysql
 import pandas as pd
 from sshtunnel import SSHTunnelForwarder
@@ -141,8 +142,8 @@ def parse_logfile(logfile):
 
 
 def get_baseline(table, connection):
-    query = '''SELECT * from '''+table+''' WHERE Datetime = (SELECT MAX(Datetime) FROM '''+table+''' where Branch_ID='develop' );'''
-    return pd.read_sql_query(query, connection)
+    query = text('''SELECT * from '''+table+''' WHERE Datetime = (SELECT MAX(Datetime) FROM '''+table+''' where Branch_ID='develop' );''')
+    return pd.read_sql(query, connection)
 
 def store_new_test_result(table_name, test_results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, connection):
     params=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(environment),str(datetime.datetime.now())]

test/CMakeLists.txt

@@ -27,7 +27,7 @@ function(add_gtest_executable TEST_NAME)
     # suppress gtest warnings
     target_compile_options(${TEST_NAME} PRIVATE -Wno-global-constructors -Wno-undef)
     target_link_libraries(${TEST_NAME} PRIVATE gtest_main)
-    add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> )
+    add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}>)
     rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
 endfunction(add_gtest_executable TEST_NAME)
 
@@ -36,6 +36,7 @@ add_subdirectory(space_filling_curve)
 add_subdirectory(conv_util)
 add_subdirectory(reference_conv_fwd)
 add_subdirectory(gemm)
+add_subdirectory(gemm_layernorm)
 add_subdirectory(gemm_split_k)
 add_subdirectory(gemm_reduce)
 add_subdirectory(batched_gemm)

test/batched_gemm_softmax_gemm_permute/CMakeLists.txt

@@ -5,4 +5,11 @@ add_gtest_executable(test_batched_gemm_softmax_gemm_permute_bf16 test_batched_ge
 target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
 target_link_libraries(test_batched_gemm_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
 add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
-add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16)
+add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16)
\ No newline at end of file
+
+add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_fp16 test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp)
+add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_bf16 test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp)
+target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
+target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
+add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_fp16)
+add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_bf16)
\ No newline at end of file

test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_bias_softmax_gemm_permute_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16
+    : public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
+{
+};
+
+using I1_t = ck::Number<1>;
+using I2_t = ck::Number<2>;
+
+using MaskDisabled_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
+using MaskOutUpperTriangle_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<BF16>, ck::Tuple<>, MaskDisabled_t>,
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<BF16>, ck::Tuple<>, MaskOutUpperTriangle_t>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Test_BF16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 3, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 2, 4},
+        {128, 128, 136, 128, 4, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 4, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 2, 3},
+        {128, 128, 129, 128, 2, 3},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Bench_BF16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
+                                                   {256, 64, 160, 64, 1, 16},
+                                                   {1024, 1024, 80, 80, 1, 16},
+                                                   {1024, 64, 80, 64, 1, 16},
+                                                   {4096, 4096, 40, 40, 1, 16},
+                                                   {4096, 64, 40, 64, 1, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Bench_BF16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 48, 16},
+        {256, 256, 128, 128, 48, 16},
+        {512, 512, 64, 64, 48, 16},
+        {512, 512, 128, 128, 48, 16},
+        {1024, 1024, 64, 64, 48, 16},
+        {1024, 1024, 128, 128, 48, 16},
+        {2048, 2048, 64, 64, 48, 16},
+        {2048, 2048, 128, 128, 48, 16},
+        {4096, 4096, 64, 64, 48, 16},
+        {4096, 4096, 128, 128, 48, 16},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 4, 6},
+        {64, 49, 64, 64, 4, 6},
+        {1020, 1020, 64, 128, 4, 6},
+        {576, 576, 64, 64, 4, 6},
+    };
+    this->Run();
+}

test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_softmax_gemm_permute_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16
+    : public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
+{
+};
+
+using I1_t = ck::Number<1>;
+using I2_t = ck::Number<2>;
+
+using MaskDisabled_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
+using MaskOutUpperTriangle_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<F16>, ck::Tuple<>, MaskDisabled_t>,
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<F16>, ck::Tuple<>, MaskOutUpperTriangle_t>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 3, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 2, 4},
+        {128, 128, 136, 128, 4, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 4, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 2, 3},
+        {128, 128, 129, 128, 2, 3},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
+                                                   {256, 64, 160, 64, 1, 16},
+                                                   {1024, 1024, 80, 80, 1, 16},
+                                                   {1024, 64, 80, 64, 1, 16},
+                                                   {4096, 4096, 40, 40, 1, 16},
+                                                   {4096, 64, 40, 64, 1, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 48, 16},
+        {256, 256, 128, 128, 48, 16},
+        {512, 512, 64, 64, 48, 16},
+        {512, 512, 128, 128, 48, 16},
+        {1024, 1024, 64, 64, 48, 16},
+        {1024, 1024, 128, 128, 48, 16},
+        {2048, 2048, 64, 64, 48, 16},
+        {2048, 2048, 128, 128, 48, 16},
+        {4096, 4096, 64, 64, 48, 16},
+        {4096, 4096, 128, 128, 48, 16},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 4, 6},
+        {64, 49, 64, 64, 4, 6},
+        {1020, 1020, 64, 128, 4, 6},
+        {576, 576, 64, 64, 4, 6},
+    };
+    this->Run();
+}

test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp"
+
+using ck::tensor_operation::device::GemmSpecialization;
+using ck::tensor_operation::device::MaskingSpecialization;
+using ck::tensor_operation::device::TensorSpecialization;
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmMaskingScaleSoftmaxGemmPermute : public ::testing::Test
+{
+    using NumDimGType      = std::tuple_element_t<0, Tuple>;
+    using NumDimMType      = std::tuple_element_t<1, Tuple>;
+    using NumDimNType      = std::tuple_element_t<2, Tuple>;
+    using NumDimKType      = std::tuple_element_t<3, Tuple>;
+    using NumDimOType      = std::tuple_element_t<4, Tuple>;
+    using ADataType        = std::tuple_element_t<5, Tuple>;
+    using B0DataType       = std::tuple_element_t<6, Tuple>;
+    using B1DataType       = std::tuple_element_t<7, Tuple>;
+    using CDataType        = std::tuple_element_t<8, Tuple>;
+    using Acc0BiasDataType = std::tuple_element_t<9, Tuple>;
+    using Acc1BiasDataType = std::tuple_element_t<10, Tuple>;
+    using MaskingType      = std::tuple_element_t<11, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {
+        {256, 256, 64, 64, 6, 4},
+        {256, 256, 128, 128, 4, 6},
+        {512, 512, 64, 64, 3, 2},
+        {512, 512, 128, 128, 2, 3},
+        {1024, 1024, 64, 64, 3, 1},
+        {1024, 1024, 128, 128, 1, 1},
+    };
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int G0, int G1)
+    {
+        bool pass =
+            ck::profiler::profile_batched_gemm_bias_softmax_gemm_permute_impl<NumDimGType::value,
+                                                                              NumDimMType::value,
+                                                                              NumDimNType::value,
+                                                                              NumDimKType::value,
+                                                                              NumDimOType::value,
+                                                                              ADataType,
+                                                                              B0DataType,
+                                                                              B1DataType,
+                                                                              CDataType,
+                                                                              Acc0BiasDataType,
+                                                                              Acc1BiasDataType,
+                                                                              MaskingType::value>(
+                verify_, 2, false, bench_, M, N, K, O, G0, G1);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M  = lengths[0];
+            int N  = lengths[1];
+            int K  = lengths[2];
+            int O  = lengths[3];
+            int G0 = lengths[4];
+            int G1 = lengths[5];
+
+            this->RunSingle(M, N, K, O, G0, G1);
+        }
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using ADataType        = F16;
+    using B0DataType       = F16;
+    using B1DataType       = F16;
+    using AccDataType      = float;
+    using CShuffleDataType = F16;
+    using CDataType        = F16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = ScaleAdd;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+            2,
+            1,
+            1,
+            1,
+            1,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            ck::Tuple<F16>,
+            ck::Tuple<>,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            TensorSpecialization::Default, // ATensorSpec
+            TensorSpecialization::Default, // B0TensorSpec
+            TensorSpecialization::Default, // B1TensorSpec
+            TensorSpecialization::Default, // CTensorSpec
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        const int G0 = 1, G1 = 1;
+
+        // A layout [G0, M, G1, K]
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+        // B0 layout [G0, N, G1, K]
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+        // B1 layout [G0, N, G1, O]
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+        // C layout [G0, M, G1, O]
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+        // D layout [G0, M, G1, N]
+        std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+        std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          std::array<void*, 1>{nullptr}, // p_acc0_biases
+                                          {},                            // p_acc1_biases
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b0_gs_ns_ks_lengths,
+                                          b0_gs_ns_ks_strides,
+                                          b1_gs_os_ns_lengths,
+                                          b1_gs_os_ns_strides,
+                                          c_gs_ms_os_lengths,
+                                          c_gs_ms_os_strides,
+                                          std::array<std::vector<ck::index_t>, 1>{
+                                              d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+                                          std::array<std::vector<ck::index_t>, 1>{
+                                              d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+                                          {},                       // acc1_biases_gs_ms_os_lengths
+                                          {},                       // acc1_biases_gs_ms_os_strides
+                                          PassThrough{},            // a_element_op
+                                          PassThrough{},            // b0_element_op
+                                          Acc0ElementOp{1.f},       // acc0_element_op
+                                          PassThrough{},            // b1_element_op
+                                          PassThrough{});           // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using ADataType        = BF16;
+    using B0DataType       = BF16;
+    using B1DataType       = BF16;
+    using AccDataType      = float;
+    using CShuffleDataType = BF16;
+    using CDataType        = BF16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = ScaleAdd;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+            2,
+            1,
+            1,
+            1,
+            1,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            ck::Tuple<BF16>,
+            ck::Tuple<>,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            TensorSpecialization::Default, // ATensorSpec
+            TensorSpecialization::Default, // B0TensorSpec
+            TensorSpecialization::Default, // B1TensorSpec
+            TensorSpecialization::Default, // CTensorSpec
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        const int G0 = 1, G1 = 1;
+
+        // A layout [G0, M, G1, K]
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+        // B0 layout [G0, N, G1, K]
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+        // B1 layout [G0, N, G1, O]
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+        // C layout [G0, M, G1, O]
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+        // D layout [G0, M, G1, N]
+        std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+        std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          std::array<void*, 1>{nullptr}, // p_acc0_biases
+                                          {},                            // p_acc1_biases
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b0_gs_ns_ks_lengths,
+                                          b0_gs_ns_ks_strides,
+                                          b1_gs_os_ns_lengths,
+                                          b1_gs_os_ns_strides,
+                                          c_gs_ms_os_lengths,
+                                          c_gs_ms_os_strides,
+                                          std::array<std::vector<ck::index_t>, 1>{
+                                              d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+                                          std::array<std::vector<ck::index_t>, 1>{
+                                              d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+                                          {},                       // acc1_biases_gs_ms_os_lengths
+                                          {},                       // acc1_biases_gs_ms_os_strides
+                                          PassThrough{},            // a_element_op
+                                          PassThrough{},            // b0_element_op
+                                          Acc0ElementOp{1.f},       // acc0_element_op
+                                          PassThrough{},            // b1_element_op
+                                          PassThrough{});           // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};

test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp

@@ -23,7 +23,7 @@ class TestElementwiseLayernorm : public ::testing::Test
     {
         // M, N
         std::vector<std::vector<ck::index_t>> lengths = {
-            {1, 1}, {25, 16}, {39, 777}, {100, 200}, {1024, 1024}, {48 * 256, 2048}};
+            {1, 1}, {25, 16}, {39, 777}, {100, 200}, {1024, 1024}, {48 * 256, 2048}, {4096, 8192}};
 
         for(auto length : lengths)
         {

test/gemm_layernorm/CMakeLists.txt

+add_custom_target(test_gemm_layernorm)
+
+add_gtest_executable(test_gemm_add_relu_add_layernorm_fp16 test_gemm_add_relu_add_layernorm_fp16.cpp)
+
+target_link_libraries(test_gemm_add_relu_add_layernorm_fp16 PRIVATE utility device_gemm_add_relu_add_layernorm_instance)
+
+add_dependencies(test_gemm_layernorm test_gemm_add_relu_add_layernorm_fp16)

test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_gemm_add_relu_add_layernorm_impl.hpp"
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestGemmAddReluAddLayernorm : public ::testing::Test
+{
+    protected:
+    using ADataType        = std::tuple_element_t<0, Tuple>;
+    using BDataType        = std::tuple_element_t<1, Tuple>;
+    using AccDataType      = std::tuple_element_t<2, Tuple>;
+    using D0DataType       = std::tuple_element_t<3, Tuple>;
+    using D1DataType       = std::tuple_element_t<4, Tuple>;
+    using EMeanVarDataType = std::tuple_element_t<5, Tuple>;
+    using GammaDataType    = std::tuple_element_t<6, Tuple>;
+    using BetaDataType     = std::tuple_element_t<7, Tuple>;
+    using HDataType        = std::tuple_element_t<8, Tuple>;
+    using ALayout          = std::tuple_element_t<9, Tuple>;
+    using BLayout          = std::tuple_element_t<10, Tuple>;
+    using D0Layout         = std::tuple_element_t<11, Tuple>;
+    using D1Layout         = std::tuple_element_t<12, Tuple>;
+    using HLayout          = std::tuple_element_t<13, Tuple>;
+
+    void Run()
+    {
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {1024, 1024, 1024}, {2048, 640, 640}, {1, 1, 1}};
+
+        for(auto length : lengths)
+        {
+            int M        = length[0];
+            int N        = length[1];
+            int K        = length[2];
+            int StrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+            int StrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+            int StrideD0 = 0;
+            int StrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+            int StrideH  = ck::is_same_v<HLayout, Row> ? N : M;
+
+            bool success = ck::profiler::profile_gemm_add_relu_add_layernorm_impl<ADataType,
+                                                                                  BDataType,
+                                                                                  AccDataType,
+                                                                                  D0DataType,
+                                                                                  D1DataType,
+                                                                                  EMeanVarDataType,
+                                                                                  GammaDataType,
+                                                                                  BetaDataType,
+                                                                                  HDataType,
+                                                                                  ALayout,
+                                                                                  BLayout,
+                                                                                  D0Layout,
+                                                                                  D1Layout,
+                                                                                  HLayout>(
+                true, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideH);
+
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    std::tuple<F16, F16, F32, F16, F16, F16, F16, F16, F16, Row, Row, Row, Row, Row>,
+    std::tuple<F16, F16, F32, F16, F16, F16, F16, F16, F16, Row, Col, Row, Row, Row>,
+    std::tuple<F16, F16, F32, F16, F16, F16, F16, F16, F16, Col, Row, Row, Row, Row>,
+    std::tuple<F16, F16, F32, F16, F16, F16, F16, F16, F16, Col, Col, Row, Row, Row>>;
+
+TYPED_TEST_SUITE(TestGemmAddReluAddLayernorm, KernelTypes);
+TYPED_TEST(TestGemmAddReluAddLayernorm, Test_FP16) { this->Run(); }

test/normalization/CMakeLists.txt

-add_custom_target(test_layernorm)
+add_custom_target(test_normalization)
 
 add_gtest_executable(test_layernorm2d_fp32 test_layernorm2d_fp32.cpp)
 add_gtest_executable(test_layernorm2d_fp16 test_layernorm2d_fp16.cpp)
 add_gtest_executable(test_groupnorm_fp16 test_groupnorm_fp16.cpp)
-add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp) 
-
+add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp)
 
 target_link_libraries(test_layernorm2d_fp32 PRIVATE utility device_normalization_instance)
 target_link_libraries(test_layernorm2d_fp16 PRIVATE utility device_normalization_instance)
 target_link_libraries(test_groupnorm_fp16 PRIVATE utility device_normalization_instance)
 target_link_libraries(test_groupnorm_fp32 PRIVATE utility device_normalization_instance)
 
-add_dependencies(test_layernorm test_layernorm2d_fp32)
-add_dependencies(test_layernorm test_layernorm2d_fp16)
-add_dependencies(test_layernorm test_groupnorm_fp16)
-add_dependencies(test_layernorm test_groupnorm_fp32)
+add_dependencies(test_normalization test_layernorm2d_fp32)
+add_dependencies(test_normalization test_layernorm2d_fp16)
+add_dependencies(test_normalization test_groupnorm_fp16)
+add_dependencies(test_normalization test_groupnorm_fp32)