Merge branch 'develop' of...

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into navi3x_mD_batchedGEMM_GroupConvFwd

client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp

@@ -8,7 +8,7 @@
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp"

client_example/04_contraction/CMakeLists.txt

@@ -4,3 +4,6 @@ target_link_libraries(client_contraction_scale PRIVATE composable_kernel::device
 add_executable(client_contraction_bilinear contraction_bilinear.cpp)
 target_link_libraries(client_contraction_bilinear 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_g1m2n3k1_add_xdl_fp16.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_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 1;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 3;
+static constexpr ck::index_t NumDimK = 1;
+
+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 G0 = 1;
+
+    ck::index_t M0 = 64;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 3;
+    ck::index_t N1 = 12;
+    ck::index_t N2 = 64;
+
+    ck::index_t K0 = 768;
+
+    // A[M0, M1, M2, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[N0, N1, N2, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, N0, N1, N2, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{N0 * N1 * N2 * K0, N1 * N2 * K0, N2 * K0, K0, 1};
+
+    // D[N0, M0, N1, M1, N2]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{N0 * N1 * N2, 0, 0, N1 * N2, N2, 1};
+    // E[N0 M0 N1 N2 M1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        M0 * M1 * N0 * N1 * N2, N1 * N2 * M1, 1, M0 * N1 * N2 * M1, M1 * N2, M1};
+
+    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_gs_ms_ks_lengths, a_gs_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_gs_ns_ks_lengths, b_gs_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_gs_ms_ns_lengths, d_gs_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_gs_ms_ns_lengths, e_gs_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchedContractionMultipleD<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        DsDataType,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Add>;
+
+    // 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{};
+
+    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_gs_ms_ks_lengths,
+                                        a_gs_ms_ks_strides,
+                                        b_gs_ns_ks_lengths,
+                                        b_gs_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                        e_gs_ms_ns_lengths,
+                                        e_gs_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_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_gs_ms_ks_lengths.begin() + NumDimG + 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/06_softmax/softmax4d.cpp

@@ -47,8 +47,8 @@ int main(int argc, char* argv[])
     ck::index_t num_elements =
         std::accumulate(in_lengths.begin(), in_lengths.end(), 1, std::multiplies<ck::index_t>());
 
-    AccDataType alpha{2.0f};
-    AccDataType beta{2.0f};
+    double alpha{2.0};
+    double beta{2.0};
 
     SimpleDeviceMem in(sizeof(InDataType) * num_elements);
     SimpleDeviceMem out(sizeof(OutDataType) * num_elements);
@@ -82,8 +82,8 @@ int main(int argc, char* argv[])
         auto argument_ptr   = op_ptr->MakeArgumentPointer(in_lengths,
                                                         in_strides,
                                                         reduce_dims,
-                                                        &alpha,
-                                                        &beta,
+                                                        alpha,
+                                                        beta,
                                                         in.GetDeviceBuffer(),
                                                         out.GetDeviceBuffer(),
                                                         PassThrough{},
@@ -129,8 +129,8 @@ int main(int argc, char* argv[])
         auto argument_ptr = op_ptr->MakeArgumentPointer(in_lengths,
                                                         in_strides,
                                                         reduce_dims,
-                                                        &alpha,
-                                                        &beta,
+                                                        alpha,
+                                                        beta,
                                                         in.GetDeviceBuffer(),
                                                         out.GetDeviceBuffer(),
                                                         PassThrough{},
@@ -147,4 +147,4 @@ int main(int argc, char* argv[])
     }
 
     return 0;
-}
\ No newline at end of file
+}

client_example/13_batchnorm/CMakeLists.txt

 add_executable(client_batchnorm_fwd_nhwc batchnorm_fwd_nhwc.cpp)
 add_executable(client_batchnorm_bwd_nhwc batchnorm_bwd_nhwc.cpp)
+add_executable(client_batchnorm_infer_nhwc batchnorm_infer_nhwc.cpp)
 target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_operations)
 target_link_libraries(client_batchnorm_bwd_nhwc PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_batchnorm_infer_nhwc PRIVATE composable_kernel::device_operations)

client_example/13_batchnorm/batchnorm_infer_nhwc.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using ScaleDataType   = float;
+using BiasDataType    = float;
+using MeanVarDataType = float;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+using Normalize = ck::tensor_operation::element_wise::NormalizeInInfer;
+
+const double epsilon = std::numeric_limits<float>::epsilon();
+
+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[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+    std::array<int, Rank - NumBatchNormReduceDim> invariantDims{3};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem variance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    // values in variance need be non-negative
+    (void)hipMemset(
+        variance.GetDeviceBuffer(), 0, sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    std::array<ck::index_t, Rank> aligned_scaleBiasMeanVarStrides{0};
+
+    int i = 0;
+    for(auto dim : invariantDims)
+    {
+        assert(xyLengths[dim] == scaleBiasMeanVarLengths[i]);
+
+        aligned_scaleBiasMeanVarStrides[dim] = scaleBiasMeanVarStrides[i];
+        i++;
+    };
+
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+        ck::Tuple<YDataType>,
+        Normalize,
+        Rank>;
+
+    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(xyLengths,
+                                                        {xyStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides},
+                                                        {xyStrides},
+                                                        {x.GetDeviceBuffer(),
+                                                         mean.GetDeviceBuffer(),
+                                                         variance.GetDeviceBuffer(),
+                                                         scale.GetDeviceBuffer(),
+                                                         bias.GetDeviceBuffer()},
+                                                        {y.GetDeviceBuffer()},
+                                                        Normalize{epsilon});
+
+        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_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                numScaleBiasMeanVarElement * (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                              sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+            float gb_per_sec = num_bytes / 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;
+        }
+    }
+
+    if(found)
+    {
+        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(xyLengths,
+                                                        {xyStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides},
+                                                        {xyStrides},
+                                                        {x.GetDeviceBuffer(),
+                                                         mean.GetDeviceBuffer(),
+                                                         variance.GetDeviceBuffer(),
+                                                         scale.GetDeviceBuffer(),
+                                                         bias.GetDeviceBuffer()},
+                                                        {y.GetDeviceBuffer()},
+                                                        Normalize{epsilon});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/15_gemm_add_multiply/CMakeLists.txt

+
+add_executable(client_gemm_add_multiply gemm_add_multiply.cpp)
+target_link_libraries(client_gemm_add_multiply PRIVATE composable_kernel::device_operations)
\ No newline at end of file

client_example/15_gemm_add_multiply/gemm_add_multiply.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_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+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[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                          BLayout,
+                                                          ck::Tuple<D0Layout, D1Layout>,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          ck::Tuple<D0DataType, D1DataType>,
+                                                          EDataType,
+                                                          AElementOp,
+                                                          BElementOp,
+                                                          CDEElementOp>;
+
+    // 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{};
+
+    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*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            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});
+
+            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;
+
+    // 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(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/15_reduce/CMakeLists.txt

+add_executable(client_reduce_nhwc_c reduce_nhwc_c.cpp)
+target_link_libraries(client_reduce_nhwc_c PRIVATE composable_kernel::device_operations)

client_example/15_reduce/reduce_nhwc_c.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp"
+
+using InDataType  = float;
+using OutDataType = float;
+using AccDataType = float;
+using ReduceAdd   = ck::reduce::Add;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using UnaryDivide = ck::tensor_operation::element_wise::UnaryDivide;
+
+constexpr bool PropagateNan = false;
+constexpr bool OutputIndex  = false;
+
+constexpr int Rank         = 4;
+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[])
+{
+    std::array<ck::index_t, Rank> in_lengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> in_strides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumReduceDim> out_lengths{256};
+    std::array<ck::index_t, Rank - NumReduceDim> out_strides{1};
+    std::array<int, NumReduceDim> reduce_dims{0, 1, 2};
+
+    ck::index_t num_in_elements =
+        std::accumulate(in_lengths.begin(), in_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t num_out_elements =
+        std::accumulate(out_lengths.begin(), out_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t reduce_length = 1;
+
+    for(auto dim : reduce_dims)
+        reduce_length *= in_lengths[dim];
+
+    double alpha{1.0};
+    double beta{0.0};
+
+    SimpleDeviceMem in(sizeof(InDataType) * num_in_elements);
+    SimpleDeviceMem out(sizeof(OutDataType) * num_out_elements);
+
+    using DeviceOp     = ck::tensor_operation::device::DeviceReduce<InDataType,
+                                                                AccDataType,
+                                                                OutDataType,
+                                                                Rank,
+                                                                NumReduceDim,
+                                                                ReduceAdd,
+                                                                PassThrough,
+                                                                UnaryDivide,
+                                                                PropagateNan,
+                                                                OutputIndex>;
+    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(in_lengths,
+                                                        in_strides,
+                                                        out_lengths,
+                                                        out_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        out.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        PassThrough{},
+                                                        UnaryDivide{reduce_length});
+        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_bytes = num_in_elements * sizeof(InDataType) +
+                                    (beta == 0.0f ? 1 : 2) * num_out_elements * sizeof(OutDataType);
+
+            float gb_per_sec = num_bytes / 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
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in_lengths,
+                                                        in_strides,
+                                                        out_lengths,
+                                                        out_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        out.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        PassThrough{},
+                                                        UnaryDivide{reduce_length});
+
+        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;
+}

example/01_gemm/CMakeLists.txt

@@ -17,12 +17,14 @@ endif(USE_BITINT_EXTENSION_INT4)
 add_custom_target(example_gemm_xdl)
 
 add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
+add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
 add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
 add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
 
 add_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
 add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
 add_dependencies(example_gemm_xdl example_gemm_xdl_int8)
+add_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
 
 if(USE_BITINT_EXTENSION_INT4)
   add_example_executable(example_gemm_xdl_int4 gemm_xdl_int4.cpp)

example/01_gemm/gemm_xdl_fp16.cpp

@@ -12,6 +12,8 @@ using AccDataType      = float;
 using CShuffleDataType = float;
 using CDataType        = ck::half_t;
 
+using F16 = ck::half_t;
+
 using ALayout = Row;
 using BLayout = Col;
 using CLayout = Row;
@@ -29,7 +31,7 @@ using DeviceGemmInstance0 = ck::tensor_operation::device::DeviceGemmXdl
 // ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
 // ######|          |          |          |            |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
          < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>;
-// clang-format on
+// // clang-format on
 
 // clang-format off
 using DeviceGemmInstance1 = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
@@ -40,7 +42,7 @@ using DeviceGemmInstance1 = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffl
          < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
 // clang-format on
 
-using DeviceGemmInstance = DeviceGemmInstance0;
+using DeviceGemmInstance = DeviceGemmInstance1;
 
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;

example/01_gemm/gemm_xdl_wavelet_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using CDataType        = ck::half_t;
+
+using F16 = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_WaveletModel_CShuffle
+    // clang-format off
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     AccData|         CShuffle|     CData|           A|           B|           C|           GEMM| NumGemmK| ABBlockTransfer|       BlockGemm|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|        Type|         DataType|      Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| ThreadGroupSize| ThreadGroupSize| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |            |                 |          |   Operation|   Operation|   Operation|               |    Stage|                |                |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |            |                 |          |            |            |            |               |         |                |                |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, AccDataType,              F16, CDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,             256,             256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,                S<1, 32, 1,8>,               8>;
+// clang-format on
+
+using DeviceGemmInstance = DeviceGemmInstance;
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/12_reduce/reduce_blockwise_impl.hpp

@@ -9,6 +9,7 @@
 #include "ck/utility/reduction_enums.hpp"
 #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
 
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -16,7 +17,6 @@
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/utility/host_common_util.hpp"
-#include "ck/library/utility/host_reduction.hpp"
 
 #include "reduce_example_common.hpp"
 
@@ -236,29 +236,6 @@ int reduce_blockwise_impl(bool do_verification,
         reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
             static_cast<int32_t>(reduce_total_length));
 
-    if(do_verification)
-    {
-        ReductionHost<InOutDataType,
-                      AccDataType,
-                      InOutDataType,
-                      ReduceOperation,
-                      InElementwiseOperation,
-                      AccElementwiseOperation,
-                      Rank,
-                      NumReduceDim,
-                      PropagateNan,
-                      OutputIndex>
-            hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
-
-        hostReduce.Run(alpha,
-                       in.mData.data(),
-                       beta,
-                       out_ref.mData.data(),
-                       out_indices_ref.mData.data(),
-                       in_elementwise_op,
-                       acc_elementwise_op);
-    };
-
     std::array<index_t, Rank> arrInLengths;
     std::array<index_t, Rank> arrInStrides;
     std::array<index_t, NumOutDim> arrOutLengths;
@@ -269,6 +246,48 @@ int reduce_blockwise_impl(bool do_verification,
     ck::ranges::copy(outLengths, arrOutLengths.begin());
     ck::ranges::copy(outStrides, arrOutStrides.begin());
 
+    if(do_verification)
+    {
+        using ReferenceReduceInstance =
+            ck::tensor_operation::host::ReferenceReduce<InOutDataType,
+                                                        AccDataType,
+                                                        InOutDataType,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperation,
+                                                        AccElementwiseOperation,
+                                                        PropagateNan,
+                                                        OutputIndex>;
+
+        auto reduce_ref = ReferenceReduceInstance{};
+
+        auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths,
+                                                               arrInStrides,
+                                                               arrOutLengths,
+                                                               arrOutStrides,
+                                                               reduceDims,
+                                                               static_cast<double>(alpha),
+                                                               static_cast<double>(beta),
+                                                               in.mData.data(),
+                                                               nullptr,
+                                                               out_ref.mData.data(),
+                                                               out_indices_ref.mData.data(),
+                                                               in_elementwise_op,
+                                                               acc_elementwise_op);
+
+        if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reduce reference, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
+
+        invoker_ptr_ref->Run(argument_ptr_ref.get());
+    };
+
     auto reduce = DeviceReduceInstance{};
 
     auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
@@ -276,8 +295,8 @@ int reduce_blockwise_impl(bool do_verification,
                                                    arrOutLengths,
                                                    arrOutStrides,
                                                    reduceDims,
-                                                   alpha,
-                                                   beta,
+                                                   static_cast<double>(alpha),
+                                                   static_cast<double>(beta),
                                                    in_dev.GetDeviceBuffer(),
                                                    nullptr,
                                                    out_dev.GetDeviceBuffer(),
@@ -287,9 +306,8 @@ int reduce_blockwise_impl(bool do_verification,
 
     if(!reduce.IsSupportedArgument(argument_ptr.get()))
     {
-        std::cerr
-            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
-            << std::endl;
+        std::cerr << "The runtime parameters not supported by the DeviceReduce instance, exiting!"
+                  << std::endl;
 
         return (-2);
     };

example/12_reduce/reduce_blockwise_two_call.cpp

@@ -12,13 +12,13 @@
 #include "ck/utility/reduction_enums.hpp"
 #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.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/host_common_util.hpp"
-#include "ck/library/utility/host_reduction.hpp"
 
 using namespace ck;
 using namespace ck::tensor_operation::device;
@@ -97,8 +97,8 @@ int main(int argc, char* argv[])
     // const std::array<int, 3> invariantDims_2 = {0, 1, 2};
 
     // used by the host reduction
-    const std::array<int, 2> reduceDims    = {3, 4};
-    const std::array<int, 3> invariantDims = {0, 1, 2};
+    const std::array<int, 2> reduceDims = {3, 4};
+    // const std::array<int, 3> invariantDims = {0, 1, 2};
 
     const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
 
@@ -191,29 +191,6 @@ int main(int argc, char* argv[])
         reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
             static_cast<int32_t>(reduce_total_length));
 
-    if(do_verify)
-    {
-        ReductionHost<InOutDataType,
-                      AccDataType,
-                      InOutDataType,
-                      ReduceOperation,
-                      InElementwiseOperation,
-                      AccElementwiseOperation,
-                      5, // Rank
-                      2, // NumReduceDim
-                      PropagateNan,
-                      OutputIndex>
-            hostReduce(in_1.mDesc, out_ref.mDesc, invariantDims, reduceDims);
-
-        hostReduce.Run(alpha,
-                       in_1.mData.data(),
-                       beta,
-                       out_ref.mData.data(),
-                       nullptr,
-                       in_elementwise_op,
-                       acc_elementwise_op);
-    };
-
     std::array<index_t, 5> arrInLengths_1;
     std::array<index_t, 5> arrInStrides_1;
     std::array<index_t, 4> arrInLengths_2;
@@ -228,6 +205,48 @@ int main(int argc, char* argv[])
     ck::ranges::copy(outLengths, arrOutLengths.begin());
     ck::ranges::copy(outStrides, arrOutStrides.begin());
 
+    if(do_verify)
+    {
+        using ReferenceReduceInstance =
+            ck::tensor_operation::host::ReferenceReduce<InOutDataType,
+                                                        AccDataType,
+                                                        InOutDataType,
+                                                        5,
+                                                        2,
+                                                        ReduceOperation,
+                                                        InElementwiseOperation,
+                                                        AccElementwiseOperation,
+                                                        PropagateNan,
+                                                        OutputIndex>;
+
+        auto reduce_ref = ReferenceReduceInstance{};
+
+        auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths_1,
+                                                               arrInStrides_1,
+                                                               arrOutLengths,
+                                                               arrOutStrides,
+                                                               reduceDims,
+                                                               static_cast<double>(alpha),
+                                                               static_cast<double>(beta),
+                                                               in_1.mData.data(),
+                                                               nullptr,
+                                                               out_ref.mData.data(),
+                                                               nullptr,
+                                                               in_elementwise_op,
+                                                               acc_elementwise_op);
+
+        if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reduce reference, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
+
+        invoker_ptr_ref->Run(argument_ptr_ref.get());
+    };
+
     auto reduce_1 = DeviceReduceInstance_1{};
 
     auto argument_ptr_1 = reduce_1.MakeArgumentPointer(arrInLengths_1,
@@ -235,8 +254,8 @@ int main(int argc, char* argv[])
                                                        arrInLengths_2,
                                                        arrInStrides_2,
                                                        reduceDims_1,
-                                                       1.0f,
-                                                       0.0f,
+                                                       1.0,
+                                                       0.0,
                                                        in_1_dev.GetDeviceBuffer(),
                                                        nullptr,
                                                        in_2_dev.GetDeviceBuffer(),
@@ -246,9 +265,8 @@ int main(int argc, char* argv[])
 
     if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
     {
-        std::cout
-            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
-            << std::endl;
+        std::cout << "The runtime parameters seems supported by the DeviceReduce instance, exiting!"
+                  << std::endl;
     };
 
     auto invoker_ptr_1 = reduce_1.MakeInvokerPointer();
@@ -260,8 +278,8 @@ int main(int argc, char* argv[])
                                                        arrOutLengths,
                                                        arrOutStrides,
                                                        reduceDims_2,
-                                                       alpha,
-                                                       beta,
+                                                       static_cast<double>(alpha),
+                                                       static_cast<double>(beta),
                                                        in_2_dev.GetDeviceBuffer(),
                                                        nullptr,
                                                        out_dev.GetDeviceBuffer(),

example/12_reduce/reduce_multiblock_atomic_add_impl.hpp

@@ -9,6 +9,7 @@
 #include "ck/utility/reduction_enums.hpp"
 #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
 
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -16,7 +17,6 @@
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/utility/host_common_util.hpp"
-#include "ck/library/utility/host_reduction.hpp"
 
 #include "reduce_example_common.hpp"
 
@@ -149,29 +149,6 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
         reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
             static_cast<int32_t>(reduce_total_length));
 
-    if(do_verification)
-    {
-        ReductionHost<InOutDataType,
-                      AccDataType,
-                      InOutDataType,
-                      ReduceOperation,
-                      InElementwiseOperation,
-                      AccElementwiseOperation,
-                      Rank,
-                      NumReduceDim,
-                      PropagateNan,
-                      false>
-            hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
-
-        hostReduce.Run(alpha,
-                       in.mData.data(),
-                       beta,
-                       out_ref.mData.data(),
-                       nullptr,
-                       in_elementwise_op,
-                       acc_elementwise_op);
-    };
-
     std::array<index_t, Rank> arrInLengths;
     std::array<index_t, Rank> arrInStrides;
     std::array<index_t, NumOutDim> arrOutLengths;
@@ -182,6 +159,48 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
     ck::ranges::copy(outLengths, arrOutLengths.begin());
     ck::ranges::copy(outStrides, arrOutStrides.begin());
 
+    if(do_verification)
+    {
+        using ReferenceReduceInstance =
+            ck::tensor_operation::host::ReferenceReduce<InOutDataType,
+                                                        AccDataType,
+                                                        InOutDataType,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperation,
+                                                        AccElementwiseOperation,
+                                                        PropagateNan,
+                                                        false>;
+
+        auto reduce_ref = ReferenceReduceInstance{};
+
+        auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths,
+                                                               arrInStrides,
+                                                               arrOutLengths,
+                                                               arrOutStrides,
+                                                               reduceDims,
+                                                               static_cast<double>(alpha),
+                                                               static_cast<double>(beta),
+                                                               in.mData.data(),
+                                                               nullptr,
+                                                               out_ref.mData.data(),
+                                                               nullptr,
+                                                               in_elementwise_op,
+                                                               acc_elementwise_op);
+
+        if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reduce reference, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
+
+        invoker_ptr_ref->Run(argument_ptr_ref.get());
+    };
+
     auto reduce = DeviceReduceInstance{};
 
     auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
@@ -189,8 +208,8 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
                                                    arrOutLengths,
                                                    arrOutStrides,
                                                    reduceDims,
-                                                   alpha,
-                                                   beta,
+                                                   static_cast<double>(alpha),
+                                                   static_cast<double>(beta),
                                                    in_dev.GetDeviceBuffer(),
                                                    nullptr,
                                                    out_dev.GetDeviceBuffer(),
@@ -200,9 +219,8 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
 
     if(!reduce.IsSupportedArgument(argument_ptr.get()))
     {
-        std::cerr
-            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
-            << std::endl;
+        std::cerr << "The runtime parameters not supported by the DeviceReduce instance, exiting!"
+                  << std::endl;
 
         return (-2);
     };

example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp

@@ -6,7 +6,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -23,13 +23,13 @@ using CDataType  = F16;
 using Add = ck::tensor_operation::element_wise::Add;
 
 using DeviceElementwiseAddInstance =
-    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
-                                                    ck::Tuple<CDataType>,
-                                                    Add,
-                                                    2,
-                                                    8,
-                                                    ck::Sequence<8, 8>,
-                                                    ck::Sequence<8>>;
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
+                                                        ck::Tuple<CDataType>,
+                                                        Add,
+                                                        2,
+                                                        8,
+                                                        ck::Sequence<8, 8>,
+                                                        ck::Sequence<8>>;
 
 template <typename HostTensorA,
           typename HostTensorB,

example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp

@@ -6,7 +6,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -23,13 +23,13 @@ using CDataType  = F16;
 using Add = ck::tensor_operation::element_wise::Add;
 
 using DeviceElementwiseAddInstance =
-    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
-                                                    ck::Tuple<CDataType>,
-                                                    Add,
-                                                    3,
-                                                    8,
-                                                    ck::Sequence<1, 8>,
-                                                    ck::Sequence<8>>;
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
+                                                        ck::Tuple<CDataType>,
+                                                        Add,
+                                                        3,
+                                                        8,
+                                                        ck::Sequence<1, 8>,
+                                                        ck::Sequence<8>>;
 
 template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
 void host_broadcast3D_am_bmnk(HostTensorC& C,

example/19_binary_elementwise/elementwise_add_1d.cpp

@@ -5,7 +5,7 @@
 #include <cstdlib>
 
 #include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -21,13 +21,13 @@ using CDataType  = F16;
 using Add = ck::tensor_operation::element_wise::Add;
 
 using DeviceElementwiseAddInstance =
-    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
-                                                    ck::Tuple<CDataType>,
-                                                    Add,
-                                                    1,
-                                                    8,
-                                                    ck::Sequence<8, 8>,
-                                                    ck::Sequence<8>>;
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
+                                                        ck::Tuple<CDataType>,
+                                                        Add,
+                                                        1,
+                                                        8,
+                                                        ck::Sequence<8, 8>,
+                                                        ck::Sequence<8>>;
 
 template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
 void host_elementwise1D(

example/19_binary_elementwise/elementwise_add_4d.cpp

@@ -6,7 +6,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -23,13 +23,13 @@ using CDataType  = F16;
 using Add = ck::tensor_operation::element_wise::Add;
 
 using DeviceElementwiseAddInstance =
-    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
-                                                    ck::Tuple<CDataType>,
-                                                    Add,
-                                                    4,
-                                                    8,
-                                                    ck::Sequence<8, 8>,
-                                                    ck::Sequence<8>>;
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
+                                                        ck::Tuple<CDataType>,
+                                                        Add,
+                                                        4,
+                                                        8,
+                                                        ck::Sequence<8, 8>,
+                                                        ck::Sequence<8>>;
 
 template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
 void host_elementwise4D(HostTensorC& C,