Merge branch 'develop' into lwpck-692

.github/dependabot.yml

+# To get started with Dependabot version updates, you'll need to specify which
+# package ecosystems to update and where the package manifests are located.
+# Please see the documentation for all configuration options:
+# https://docs.github.com/github/administering-a-repository/configuration-options-for-dependency-updates
+
+version: 2
+updates:
+  - package-ecosystem: "pip" # See documentation for possible values
+    directory: "/" # Location of package manifests
+    open-pull-requests-limit: 10
+    schedule:
+      interval: "daily"

client_example/04_contraction/CMakeLists.txt

-add_executable(client_contraction_scale contraction_scale.cpp)
-target_link_libraries(client_contraction_scale PRIVATE composable_kernel::device_operations)
+add_executable(client_contraction_scale_fp32 contraction_scale_fp32.cpp)
+target_link_libraries(client_contraction_scale_fp32 PRIVATE composable_kernel::device_operations)
 
-add_executable(client_contraction_bilinear contraction_bilinear.cpp)
-target_link_libraries(client_contraction_bilinear PRIVATE composable_kernel::device_operations)
+add_executable(client_contraction_bilinear_fp32 contraction_bilinear_fp32.cpp)
+target_link_libraries(client_contraction_bilinear_fp32 PRIVATE composable_kernel::device_operations)
+
+add_executable(client_contraction_scale_fp64 contraction_scale_fp64.cpp)
+target_link_libraries(client_contraction_scale_fp64 PRIVATE composable_kernel::device_operations)
+
+add_executable(client_contraction_bilinear_fp64 contraction_bilinear_fp64.cpp)
+target_link_libraries(client_contraction_bilinear_fp64 PRIVATE composable_kernel::device_operations)
 
 add_executable(contraction_g1m2n3k1_add_xdl_fp16 contraction_g1m2n3k1_add_xdl_fp16.cpp)
 target_link_libraries(contraction_g1m2n3k1_add_xdl_fp16 PRIVATE composable_kernel::device_operations)

client_example/04_contraction/contraction_bilinear_fp64.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Bilinear;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DDataType        = F64;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+// kknn
+#if 1
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// knnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mknn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mnnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+#endif
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 25)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21]), std::stoi(argv[22])};
+
+        alpha = std::stof(argv[23]);
+        beta  = std::stof(argv[24]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg19 to 22: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg23 to 24: alpha, beta\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_ms_ns_lengths, d_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Bilinear>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{alpha, beta};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                        e_device_buf.GetDeviceBuffer(),
+                                        a_ms_ks_lengths,
+                                        a_ms_ks_strides,
+                                        b_ns_ks_lengths,
+                                        b_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                        e_ms_ns_lengths,
+                                        e_ms_ns_strides,
+                                        a_element_op,
+                                        b_element_op,
+                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop      = std::size_t(2) * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}

client_example/04_contraction/contraction_scale_fp64.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Scale;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+// kkn
+#if 1
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// knn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mkn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+#endif
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 20)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        scale = std::stof(argv[19]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg19: scale\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Scale>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{scale};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        a_ms_ks_lengths,
+                                                        a_ms_ks_strides,
+                                                        b_ns_ks_lengths,
+                                                        b_ns_ks_strides,
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        e_ms_ns_lengths,
+                                                        e_ms_ns_strides,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}

client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp

@@ -73,7 +73,7 @@ int main(int argc, char* argv[])
     SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
     SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
     SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
-    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
     SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
 
     using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
@@ -203,4 +203,4 @@ int main(int argc, char* argv[])
     }
 
     return 0;
-}
+}
\ No newline at end of file

client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp

@@ -76,7 +76,7 @@ int main(int argc, char* argv[])
     SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
     SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
     SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
-    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
     SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
 
     using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
@@ -206,4 +206,4 @@ int main(int argc, char* argv[])
     }
 
     return 0;
-}
+}
\ No newline at end of file

client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp

@@ -69,7 +69,7 @@ int main(int argc, char* argv[])
 
     SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
     SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
-    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
     SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
 
     using DeviceOp =
@@ -196,4 +196,4 @@ int main(int argc, char* argv[])
     }
 
     return 0;
-}
+}
\ No newline at end of file

client_example/18_groupnorm/CMakeLists.txt

+add_executable(client_groupnorm_swish groupnorm_swish.cpp)
+target_link_libraries(client_groupnorm_swish PRIVATE composable_kernel::device_operations)

client_example/18_groupnorm/groupnorm_swish.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization_swish.hpp"
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using Swish           = ck::tensor_operation::element_wise::Swish;
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N = 32;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 64;
+    ck::index_t C = 128;
+
+    std::size_t xy_size         = N * H * W * G * C;
+    std::size_t gamma_beta_size = G * C;
+
+    std::vector<ck::index_t> xy_strides         = {H * W * G * C, W * G * C, G * C, C, 1};
+    std::vector<ck::index_t> gamma_beta_strides = {0, 0, 0, C, 1};
+
+    SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_beta_size);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * gamma_beta_size);
+    SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                                       GammaDataType,
+                                                                       BetaDataType,
+                                                                       ComputeDataType,
+                                                                       YDataType,
+                                                                       Swish,
+                                                                       Rank,
+                                                                       NumReduceDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C},    // lengths
+                                                        xy_strides,         // xStrides
+                                                        gamma_beta_strides, // gammaStrides
+                                                        gamma_beta_strides, // betaStrides
+                                                        xy_strides,         // yStrides
+                                                        {1, 2, 4},          // reduceDims
+                                                        1e-6,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        nullptr,
+                                                        Swish{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(XDataType) * xy_size + sizeof(GammaDataType) * gamma_beta_size +
+                sizeof(BetaDataType) * gamma_beta_size + sizeof(YDataType) * xy_size;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C},    // lengths
+                                                        xy_strides,         // xStrides
+                                                        gamma_beta_strides, // gammaStrides
+                                                        gamma_beta_strides, // betaStrides
+                                                        xy_strides,         // yStrides
+                                                        {1, 2, 4},          // reduceDims
+                                                        1e-6,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        nullptr,
+                                                        Swish{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

cmake/googletest.cmake

@@ -21,6 +21,7 @@ list(APPEND GTEST_CMAKE_CXX_FLAGS
      -Wno-comma
      -Wno-old-style-cast
      -Wno-deprecated
+     -Wno-unsafe-buffer-usage
 )
 message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLAGS}")
 

docs/.sphinx/requirements.in

-git+https://github.com/RadeonOpenCompute/rocm-docs-core.git
+rocm-docs-core==0.2.0
 sphinxcontrib-bibtex==2.5.0

docs/.sphinx/requirements.txt

@@ -2,9 +2,9 @@
 # This file is autogenerated by pip-compile with Python 3.10
 # by the following command:
 #
-#    pip-compile requirements.in
+#    pip-compile .sphinx/requirements.in
 #
-accessible-pygments==0.0.4
+accessible-pygments==0.0.3
     # via pydata-sphinx-theme
 alabaster==0.7.13
     # via sphinx
@@ -20,7 +20,7 @@ babel==2.12.1
     #   sphinx
 backcall==0.2.0
     # via ipython
-beautifulsoup4==4.12.0
+beautifulsoup4==4.11.2
     # via pydata-sphinx-theme
 breathe==4.34.0
     # via rocm-docs-core
@@ -34,7 +34,7 @@ click==8.1.3
     # via
     #   jupyter-cache
     #   sphinx-external-toc
-comm==0.1.3
+comm==0.1.2
     # via ipykernel
 debugpy==1.6.6
     # via ipykernel
@@ -65,13 +65,11 @@ idna==3.4
     # via requests
 imagesize==1.4.1
     # via sphinx
-importlib-metadata==6.1.0
+importlib-metadata==6.0.0
     # via
     #   jupyter-cache
     #   myst-nb
-importlib-resources==5.10.4
-    # via rocm-docs-core
-ipykernel==6.22.0
+ipykernel==6.21.3
     # via myst-nb
 ipython==8.11.0
     # via
@@ -87,7 +85,7 @@ jsonschema==4.17.3
     # via nbformat
 jupyter-cache==0.5.0
     # via myst-nb
-jupyter-client==8.1.0
+jupyter-client==8.0.3
     # via
     #   ipykernel
     #   nbclient
@@ -124,7 +122,7 @@ nbclient==0.5.13
     # via
     #   jupyter-cache
     #   myst-nb
-nbformat==5.8.0
+nbformat==5.7.3
     # via
     #   jupyter-cache
     #   myst-nb
@@ -187,7 +185,7 @@ pyyaml==6.0
     #   myst-parser
     #   pybtex
     #   sphinx-external-toc
-pyzmq==25.0.2
+pyzmq==25.0.1
     # via
     #   ipykernel
     #   jupyter-client
@@ -195,8 +193,8 @@ requests==2.28.2
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core @ git+https://github.com/RadeonOpenCompute/rocm-docs-core.git
-    # via -r requirements.in
+rocm-docs-core==0.2.0
+    # via -r .sphinx/requirements.in
 six==1.16.0
     # via
     #   asttokens
@@ -235,9 +233,7 @@ sphinx-notfound-page==0.8.3
 sphinxcontrib-applehelp==1.0.4
     # via sphinx
 sphinxcontrib-bibtex==2.5.0
-    # via
-    #   -r requirements.in
-    #   rocm-docs-core
+    # via -r .sphinx/requirements.in
 sphinxcontrib-devhelp==1.0.2
     # via sphinx
 sphinxcontrib-htmlhelp==2.0.1
@@ -248,7 +244,7 @@ sphinxcontrib-qthelp==1.0.3
     # via sphinx
 sphinxcontrib-serializinghtml==1.1.5
     # via sphinx
-sqlalchemy==1.4.47
+sqlalchemy==1.4.46
     # via jupyter-cache
 stack-data==0.6.2
     # via ipython

example/42_groupnorm/CMakeLists.txt

-add_example_executable(example_groupnorm_sigmoid_fp16 groupnorm_sigmoid_fp16.cpp)
+add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
+add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)

example/42_groupnorm/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.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"

example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+
+struct YElementOp
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(ck::is_same<T, float>::value || ck::is_same<T, double>::value ||
+                          ck::is_same<T, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        T a;
+
+        ck::tensor_operation::element_wise::Sigmoid{}(a, x);
+
+        y = x * a;
+    };
+};
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
+                                                          GammaDataType,
+                                                          BetaDataType,
+                                                          ComputeDataType,
+                                                          YDataType,
+                                                          YElementOp,
+                                                          Rank,
+                                                          NumReduceDim,
+                                                          1024, // BlockSize
+                                                          1,    // ClusterM
+                                                          1024, // ClusterK
+                                                          1,    // SliceM
+                                                          32,   // SliceK
+                                                          1,    // SrcVecDim (0=M, 1=K)
+                                                          2,    // SrcScalarPerVector
+                                                          1,    // GammaVecDim (0=M, 1=K)
+                                                          2,    // GammaScalarPerVector
+                                                          1,    // BetaVecDim (0=M, 1=K)
+                                                          2,    // BetaScalarPerVector
+                                                          2>;   // OutScalarPerVector
+
+#include "run_groupnorm_example.inc"
+
+int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }

example/42_groupnorm/groupnorm_swish_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using YElementOp      = ck::tensor_operation::element_wise::Swish;
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
+                                                          GammaDataType,
+                                                          BetaDataType,
+                                                          ComputeDataType,
+                                                          YDataType,
+                                                          YElementOp,
+                                                          Rank,
+                                                          NumReduceDim,
+                                                          1024, // BlockSize
+                                                          1,    // ClusterM
+                                                          1024, // ClusterK
+                                                          1,    // SliceM
+                                                          32,   // SliceK
+                                                          1,    // SrcVecDim (0=M, 1=K)
+                                                          2,    // SrcScalarPerVector
+                                                          1,    // GammaVecDim (0=M, 1=K)
+                                                          2,    // GammaScalarPerVector
+                                                          1,    // BetaVecDim (0=M, 1=K)
+                                                          2,    // BetaScalarPerVector
+                                                          2>;   // OutScalarPerVector
+
+#include "run_groupnorm_example.inc"
+
+int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }

example/42_groupnorm/groupnorm_sigmoid_fp16.cpp → example/42_groupnorm/run_groupnorm_example.inc

 // 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 <getopt.h>
-
-#include "ck/ck.hpp"
-#include "ck/utility/reduction_enums.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
-#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
-
-#include "ck/library/utility/fill.hpp"
-#include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/device_memory.hpp"
-#include "ck/library/utility/host_common_util.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"
-
-constexpr int Rank         = 5;
-constexpr int NumReduceDim = 3;
-
-using XDataType       = ck::half_t;
-using GammaDataType   = ck::half_t;
-using BetaDataType    = ck::half_t;
-using YDataType       = ck::half_t;
-using ComputeDataType = float;
-
-struct YElementOp
-{
-    template <typename T>
-    __host__ __device__ void operator()(T& y, const T& x) const
-    {
-        static_assert(ck::is_same<T, float>::value || ck::is_same<T, double>::value ||
-                          ck::is_same<T, ck::half_t>::value,
-                      "Data type is not supported by this operation!");
-
-        T a;
+#pragma once
 
-        ck::tensor_operation::element_wise::Sigmoid{}(a, x);
-
-        y = x * a;
-    };
-};
-
-using DeviceInstance =
-    ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
-                                                          GammaDataType,
-                                                          BetaDataType,
-                                                          ComputeDataType,
-                                                          YDataType,
-                                                          YElementOp,
-                                                          Rank,
-                                                          NumReduceDim,
-                                                          1024, // BlockSize
-                                                          1,    // ClusterM
-                                                          1024, // ClusterK
-                                                          1,    // SliceM
-                                                          32,   // SliceK
-                                                          1,    // SrcVecDim (0=M, 1=K)
-                                                          2,    // SrcScalarPerVector
-                                                          1,    // GammaVecDim (0=M, 1=K)
-                                                          2,    // GammaScalarPerVector
-                                                          1,    // BetaVecDim (0=M, 1=K)
-                                                          2,    // BetaScalarPerVector
-                                                          2>;   // OutScalarPerVector
-
-int main(int argc, char* argv[])
+int run_groupnorm_example(int argc, char* argv[])
 {
-    ck::index_t N = 2;
-    ck::index_t H = 32;
-    ck::index_t W = 32;
-    ck::index_t G = 32;
-    ck::index_t C = 30;
+    ck::index_t N = 32;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 64;
+    ck::index_t C = 128;
 
     if(argc == 1)
     {

include/ck/ck.hpp

@@ -36,7 +36,7 @@
 #elif defined(__gfx1030__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
 #elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
-#define CK_BUFFER_RESOURCE_3RD_DWORD 0x10020000
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
 #endif
 
 // FMA instruction
@@ -168,6 +168,11 @@
 // flag to enable (1) or disable (0) the debugging output in some kernels
 #define DEBUG_LOG 0
 
+// denorm test fix, required to work around dissue
+#ifndef CK_WORKAROUND_DENORM_FIX
+#define CK_WORKAROUND_DENORM_FIX 0
+#endif
+
 namespace ck {
 
 enum struct InMemoryDataOperationEnum