Merge remote-tracking branch 'origin/develop' into bwroblew/direct_loads

example/44_elementwise_permute/elementwise_permute_3d.cpp

+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp"
+
+#include "ck/library/utility/algorithm.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"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwise3dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
+                                                          ck::Tuple<BDataType>, // OutDataTypeTuple
+                                                          PassThrough,          // ElementwiseOp
+                                                          2,                    // NumDim_m, {N, C}
+                                                          2,                    // NumDim_n, {H, W}
+                                                          1,                    // NumDim_k, {D}
+                                                          8,                    // MPerThread
+                                                          8,                    // NPerThread
+                                                          8,                    // KPerThread
+                                                          ck::Sequence<8>,  // InScalarPerVectorSeq
+                                                          ck::Sequence<4>>; // OutScalarPerVectorSeq
+
+template <typename HostTensorA, typename HostTensorB, typename Functor>
+void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
+{
+    for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
+        for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
+            for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
+                for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
+                    for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
+                    {
+                        auto a_val = A_ncdhw(n, c, d, h, w);
+                        functor(B_ndhwc(n, d, h, w, c), a_val);
+                    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    const int N = 4;
+    const int C = 16;
+    const int H = 32;
+    const int W = 5;
+    const int D = 16;
+
+    std::vector<std::size_t> ncdhw = {N, C, D, H, W};
+    std::vector<std::size_t> ndhwc = {N, D, H, W, C};
+    Tensor<ADataType> a(ncdhw);
+    Tensor<BDataType> b(ndhwc);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
+    std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
+    std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, D, H, W, C
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(
+        ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (ncdhw): " << a.mDesc << std::endl;
+    std::cout << "B (ndhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    std::size_t flop = std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
+
+    std::size_t num_btype =
+        sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
+        sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        Tensor<BDataType> host_b(ndhwc);
+        host_elementwise4D(host_b, a, PassThrough{});
+
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp

@@ -19,13 +19,13 @@ using BDataType = F16;
 
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using DeviceElementwisePermuteInstance =
-    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>,
-                                                        ck::Tuple<BDataType>,
-                                                        PassThrough,
-                                                        4,
-                                                        8,
-                                                        ck::Sequence<8>,
-                                                        ck::Sequence<1>>;
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
+                                                        ck::Tuple<BDataType>, // OutDataTypeTuple
+                                                        PassThrough,          // Elementwise op
+                                                        4,                    // NumDim
+                                                        8,                    // MPerThread
+                                                        ck::Sequence<8>,  // InScalarPerVectorSeq
+                                                        ck::Sequence<1>>; // OutScalarPerVectorSeq
 
 template <typename HostTensorA, typename HostTensorB, typename Functor>
 void host_elementwise4D(HostTensorB& B_nhwc, const HostTensorA& A_nchw, Functor functor)
@@ -99,7 +99,6 @@ int main()
 
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
               << std::endl;
-
     bool pass = true;
 
     if(do_verification)

example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp

@@ -17,15 +17,15 @@ using BDataType = F16;
 
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using DeviceElementwisePermuteInstance =
-    ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>,
-                                                          ck::Tuple<BDataType>,
-                                                          PassThrough,
-                                                          3, // NumDim_M
-                                                          1, // NumDim_N
-                                                          8,
-                                                          8,
-                                                          ck::Sequence<8>,
-                                                          ck::Sequence<8>>;
+    ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
+                                                          ck::Tuple<BDataType>, // OutDataTypeTuple
+                                                          PassThrough,          // Elementwise op
+                                                          3,                    // NumDim_M
+                                                          1,                    // NumDim_N
+                                                          1,                    // MPerThread
+                                                          1,                    // NPerThread
+                                                          ck::Sequence<1>,  // InScalarPerVectorSeq
+                                                          ck::Sequence<1>>; // OutScalarPerVectorSeq
 
 template <typename HostTensorA, typename HostTensorB, typename Functor>
 void host_elementwise4D(HostTensorB& B_nhwc,
@@ -53,12 +53,6 @@ int main()
     const int H = 32;
     const int W = 1024;
 
-    /**const int N = 120;
-    const int H = 32;
-    const int W = 64;
-
-    const int C = 128;**/
-
     std::vector<std::size_t> nchw = {N, C, H, W};
     std::vector<std::size_t> nhwc = {N, H, W, C};
 
@@ -71,7 +65,6 @@ int main()
     DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
 
     a_device_buf.ToDevice(a.mData.data());
-    // LogRangeAsType<float>(std::cout << "Tensor a  : ", a.mData, ",") << std::endl;
 
     std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
     std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
@@ -115,13 +108,10 @@ int main()
     if(do_verification)
     {
         b_device_buf.FromDevice(b.mData.data());
-        // LogRangeAsType<float>(std::cout << "Tensor b  : ", b.mData, ",") << std::endl;
 
         Tensor<BDataType> host_b(nhwc);
         host_elementwise4D<Tensor<ADataType>, Tensor<BDataType>, PassThrough>(
             host_b, a, nchw, PassThrough{});
-
-        // LogRangeAsType<float>(std::cout << "Host b  : ", host_b.mData, ",") << std::endl;
         pass &=
             ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
     }

example/44_elementwise_permute/elementwise_permute_4D_fp16_col.cpp

+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
+
+#include "ck/library/utility/algorithm.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"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using UnaryOp     = ck::tensor_operation::element_wise::UnarySquare;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
+                                                        ck::Tuple<BDataType>, // OutDataTypeTuple
+                                                        PassThrough,          // ElementwiseOp
+                                                        UnaryOp,              // UnaryOp
+                                                        Scale,                // Scalar
+                                                        4,                    // NumDim
+                                                        8,                    // MPerThread
+                                                        ck::Sequence<1>,  // InScalarPerVectorSeq
+                                                        ck::Sequence<1>>; // OutScalarPerVectorSeq
+
+template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        FunctorA functor_a,
+                        FunctorB functor_b,
+                        float scale)
+{
+    std::size_t N = A_nchw.mDesc.GetLengths()[0];
+    std::size_t C = A_nchw.mDesc.GetLengths()[1];
+    std::size_t H = A_nchw.mDesc.GetLengths()[2];
+    std::size_t W = A_nchw.mDesc.GetLengths()[3];
+    for(std::size_t w = 0; w < W; ++w)
+        for(std::size_t h = 0; h < H; ++h)
+            for(std::size_t c = 0; c < C; ++c)
+                for(std::size_t n = 0; n < N; ++n)
+                {
+                    ADataType tmp_val;
+                    // auto a_val = A_nchw(n, c, h, w);
+                    auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
+                    functor_b(tmp_val, a_val);
+                    // functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
+                    functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
+                              scale * tmp_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    std::vector<std::size_t> nchw = {4, 2, 1, 8};
+    std::vector<std::size_t> nhwc = {4, 1, 8, 2};
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+    float scale = 1.f;
+    auto i      = 0;
+    for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
+        for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
+            for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
+                for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
+                {
+                    a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
+                            (h * nchw[3]) + w] = i;
+                    i++;
+                }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths;
+
+    std::array<ck::index_t, 4> a_strides = {1,
+                                            static_cast<int>(nchw[0]),
+                                            static_cast<int>(nchw[0] * nchw[1]),
+                                            static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
+
+    std::array<ck::index_t, 4> b_strides = {1,
+                                            static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
+                                            static_cast<int>(nhwc[0]),
+                                            static_cast<int>(nhwc[0] * nhwc[1])};
+    ck::ranges::copy(nchw, ab_lengths.begin());
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(ab_lengths,
+                                                         {a_strides},
+                                                         {b_strides},
+                                                         input,
+                                                         output,
+                                                         PassThrough{},
+                                                         UnaryOp{},
+                                                         Scale{scale});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
+
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/44_elementwise_permute/elementwise_permute_4D_fp16_row.cpp

+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
+
+#include "ck/library/utility/algorithm.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"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using UnaryOp     = ck::tensor_operation::element_wise::UnarySquare;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
+                                                        ck::Tuple<BDataType>, // OutDataTypeTuple
+                                                        PassThrough,          // ElementwiseOp
+                                                        UnaryOp,              // UnaryOp
+                                                        Scale,                // Scalar
+                                                        4,                    // NumDim
+                                                        8,                    // MPerThread
+                                                        ck::Sequence<8>,  // InScalarPerVectorSeq
+                                                        ck::Sequence<1>>; // OutScalarPerVectorSeq
+
+template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        FunctorA functor_a,
+                        FunctorB functor_b,
+                        float scale)
+{
+    for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
+        for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
+            for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
+                for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
+                {
+                    ADataType tmp_val;
+                    auto a_val = A_nchw(n, c, h, w);
+                    functor_b(tmp_val, a_val);
+                    functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    std::vector<std::size_t> nchw = {16, 128, 32, 64};
+    std::vector<std::size_t> nhwc = {16, 32, 64, 128};
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+    float scale = 2.f;
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths;
+    std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
+                                            static_cast<int>(nchw[2] * nchw[3]),
+                                            static_cast<int>(nchw[3]),
+                                            1};
+    std::array<ck::index_t, 4> b_strides = {static_cast<int>(nhwc[1] * nhwc[2] * nhwc[3]),
+                                            1,
+                                            static_cast<int>(nhwc[2] * nhwc[3]),
+                                            static_cast<int>(nhwc[3])};
+
+    ck::ranges::copy(nchw, ab_lengths.begin());
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(ab_lengths,
+                                                         {a_strides},
+                                                         {b_strides},
+                                                         input,
+                                                         output,
+                                                         PassThrough{},
+                                                         UnaryOp{},
+                                                         Scale{scale});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
+
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/44_elementwise_permute/elementwise_permute_4D_fp32_col.cpp

+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
+
+#include "ck/library/utility/algorithm.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"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F32;
+using BDataType = F32;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using UnaryOp     = ck::tensor_operation::element_wise::UnarySquare;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
+                                                        ck::Tuple<BDataType>, // OutDataTypeTuple
+                                                        PassThrough,          // ElementwiseOp
+                                                        UnaryOp,              // UnaryOp
+                                                        Scale,                // Scalar
+                                                        4,                    // NumDim
+                                                        1,                    // MPerThread
+                                                        ck::Sequence<1>,  // InScalarPerVectorSeq
+                                                        ck::Sequence<1>>; // OutScalarPerVectorSeq
+
+template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        FunctorA functor_a,
+                        FunctorB functor_b,
+                        float scale)
+{
+    std::size_t N = A_nchw.mDesc.GetLengths()[0];
+    std::size_t C = A_nchw.mDesc.GetLengths()[1];
+    std::size_t H = A_nchw.mDesc.GetLengths()[2];
+    std::size_t W = A_nchw.mDesc.GetLengths()[3];
+    for(std::size_t w = 0; w < W; ++w)
+        for(std::size_t h = 0; h < H; ++h)
+            for(std::size_t c = 0; c < C; ++c)
+                for(std::size_t n = 0; n < N; ++n)
+                {
+                    ADataType tmp_val;
+                    auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
+                    functor_b(tmp_val, a_val);
+                    functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
+                              scale * tmp_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    std::vector<std::size_t> nchw = {5, 4, 2, 3};
+    std::vector<std::size_t> nhwc = {5, 2, 3, 4};
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+
+    float scale = 1.f;
+    auto i      = 0;
+    for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
+        for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
+            for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
+                for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
+                {
+                    a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
+                            (h * nchw[3]) + w] = i;
+                    i++;
+                }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths;
+
+    std::array<ck::index_t, 4> a_strides = {1,
+                                            static_cast<int>(nchw[0]),
+                                            static_cast<int>(nchw[0] * nchw[1]),
+                                            static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
+
+    std::array<ck::index_t, 4> b_strides = {1,
+                                            static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
+                                            static_cast<int>(nhwc[0]),
+                                            static_cast<int>(nhwc[0] * nhwc[1])};
+    ck::ranges::copy(nchw, ab_lengths.begin());
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(ab_lengths,
+                                                         {a_strides},
+                                                         {b_strides},
+                                                         input,
+                                                         output,
+                                                         PassThrough{},
+                                                         UnaryOp{},
+                                                         Scale{scale});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
+
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/44_elementwise_permute/elementwise_permute_4D_fp32_row.cpp

+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
+
+#include "ck/library/utility/algorithm.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"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F32;
+using BDataType = F32;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using UnaryOp     = ck::tensor_operation::element_wise::UnarySquare;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
+                                                        ck::Tuple<BDataType>, // OutDataTypeTuple
+                                                        PassThrough,          // ElementwiseOp
+                                                        UnaryOp,              // UnaryOp
+                                                        Scale,                // Scalar
+                                                        4,                    // NumDim
+                                                        8,                    // MPerThread
+                                                        ck::Sequence<8>,  // InScalarPerVectorSeq
+                                                        ck::Sequence<1>>; // OutScalarPerVectorSeq
+
+template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        FunctorA functor_a,
+                        FunctorB functor_b,
+                        float scale)
+{
+    for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
+        for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
+            for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
+                for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
+                {
+                    ADataType tmp_val;
+                    auto a_val = A_nchw(n, c, h, w);
+                    functor_b(tmp_val, a_val);
+                    functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    std::vector<std::size_t> nchw = {16, 128, 32, 64};
+    std::vector<std::size_t> nhwc = {16, 32, 64, 128};
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+    float scale = 2.f;
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths;
+    std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
+                                            static_cast<int>(nchw[2] * nchw[3]),
+                                            static_cast<int>(nchw[3]),
+                                            1};
+    std::array<ck::index_t, 4> b_strides = {static_cast<int>(nhwc[1] * nhwc[2] * nhwc[3]),
+                                            1,
+                                            static_cast<int>(nhwc[2] * nhwc[3]),
+                                            static_cast<int>(nhwc[3])};
+
+    ck::ranges::copy(nchw, ab_lengths.begin());
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(ab_lengths,
+                                                         {a_strides},
+                                                         {b_strides},
+                                                         input,
+                                                         output,
+                                                         PassThrough{},
+                                                         UnaryOp{},
+                                                         Scale{scale});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
+
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/53_layernorm_bwd/CMakeLists.txt

+add_example_executable(example_layernorm2d_bwd_fp16 layernorm2d_bwd_fp16.cpp)

example/53_layernorm_bwd/layernorm2d_bwd_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#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/utility/literals.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_impl.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp"
+
+using DYDataType         = ck::half_t;
+using XDataType          = ck::half_t;
+using GammaDataType      = ck::half_t;
+using MeanInvStdDataType = float;
+using DGammaDataType     = ck::half_t;
+using DBetaDataType      = ck::half_t;
+using DXDataType         = ck::half_t;
+using ComputeDataType    = float;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+// Layernorm:
+
+// Input shape
+// dy:        [M, N]
+// x:         [M, N]
+// mean:      [M, 1]
+// inv_std:   [M, 1]
+
+// Output shape
+// dgamma: [1, N]
+// dbeta:  [1, N]
+
+// dgamma = reduce_sum(dy * (x - mean) * inv_std, axis=0)
+// dbeta = reduce_sum(dy, axis=0)
+
+// [CAUSION]
+// In DeviceNormalizationBwdGammaBetaImpl, M is invarient dimension, K is reduced dimension
+// Hence, M in this example and DeviceNormalizationBwdGammaBetaImpl is different
+using GammaBetaDeviceInstance = ck::tensor_operation::device::DeviceNormalizationBwdGammaBetaImpl<
+    DYDataType,
+    XDataType,
+    MeanInvStdDataType,
+    ComputeDataType,
+    DGammaDataType,
+    DBetaDataType,
+    Rank,
+    NumReduceDim,
+    256,   // BlockSize
+    8,     // ClusterInvarient
+    32,    // ClusterReduce
+    8,     // SliceInvarient
+    1,     // SliceReduce
+    false, // IsDYFastestDimReduced
+    8,     // DYSrcVectorSize
+    false, // IsXFastestDimReduced
+    8,     // XSrcVectorSize
+    true,  // IsMeanInvStdFastestDimReduced
+    1,     // MeanInvStdSrcVectorSize
+    1,     // DGammaDstVectorSize
+    1>;    // DBetaDstVectorSize
+
+int main()
+{
+    bool time_kernel = false;
+
+    ck::index_t M = 1024;
+    ck::index_t N = 512;
+
+    Tensor<DYDataType> dy({M, N});
+    Tensor<XDataType> x({M, N});
+    Tensor<GammaDataType> gamma({N});
+    Tensor<MeanInvStdDataType> mean({M});
+    Tensor<MeanInvStdDataType> inv_std({M});
+
+    Tensor<DGammaDataType> dgamma({N});
+    Tensor<DBetaDataType> dbeta({N});
+    Tensor<DXDataType> dx({M, N});
+
+    dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0.0, 1.0});
+    x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+    gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
+    mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
+    inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
+
+    DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
+    DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
+    DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
+    DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
+
+    dy_dev.ToDevice(dy.mData.data());
+    x_dev.ToDevice(x.mData.data());
+    mean_dev.ToDevice(mean.mData.data());
+    inv_std_dev.ToDevice(inv_std.mData.data());
+
+    auto gamma_beta_device_instance = GammaBetaDeviceInstance{};
+    auto gamma_beta_argument_ptr =
+        gamma_beta_device_instance.MakeArgumentPointer({M, N}, // inLengths
+                                                       {N, 1}, // dyStrides
+                                                       {N, 1}, // xStrides
+                                                       {1, 0}, // meanStrides
+                                                       {1, 0}, // invStdStrides
+                                                       {N},    // outLengths
+                                                       {1},    // dgammaStrides
+                                                       {1},    // dbetaStrides
+                                                       {0},    // reduceDims
+                                                       dy_dev.GetDeviceBuffer(),
+                                                       x_dev.GetDeviceBuffer(),
+                                                       mean_dev.GetDeviceBuffer(),
+                                                       inv_std_dev.GetDeviceBuffer(),
+                                                       dgamma_dev.GetDeviceBuffer(),
+                                                       dbeta_dev.GetDeviceBuffer());
+
+    if(!gamma_beta_device_instance.IsSupportedArgument(gamma_beta_argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    auto gamma_beta_invoker_ptr = gamma_beta_device_instance.MakeInvokerPointer();
+    gamma_beta_invoker_ptr->Run(gamma_beta_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+    {
+        Tensor<DGammaDataType> host_dgamma({N});
+        Tensor<DBetaDataType> host_dbeta({N});
+        Tensor<DXDataType> host_dx({M, N});
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceLayernormBwd<DYDataType,
+                                                              XDataType,
+                                                              GammaDataType,
+                                                              MeanInvStdDataType,
+                                                              DGammaDataType,
+                                                              DBetaDataType,
+                                                              DXDataType,
+                                                              ComputeDataType>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, {M, N});
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        dgamma_dev.FromDevice(dgamma.mData.data());
+        dbeta_dev.FromDevice(dbeta.mData.data());
+
+        pass &= ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
+        pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
+    }
+
+    return (pass ? 0 : 1);
+}

example/54_groupnorm_bwd/CMakeLists.txt

+add_example_executable(example_groupnorm_bwd_fp16 groupnorm_bwd_fp16.cpp)

example/54_groupnorm_bwd/groupnorm_bwd_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#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/utility/literals.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_impl.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp"
+
+using DYDataType         = ck::half_t;
+using XDataType          = ck::half_t;
+using GammaDataType      = ck::half_t;
+using MeanInvStdDataType = float;
+using DGammaDataType     = ck::half_t;
+using DBetaDataType      = ck::half_t;
+using DXDataType         = ck::half_t;
+using ComputeDataType    = float;
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+// Grouprnorm
+// kernel:                  M    , K
+// dy:     N, H, W, G, C -> G * C, N * H * W
+// x:      N, H, W, G, C -> G * C, N * H * W
+// mean:   N, 1, 1, G, 1 -> G * 1, N * 1 * 1
+// rstd:   N, 1, 1, G, 1 -> G * 1, N * 1 * 1
+
+// dgamma: 1, 1, 1, G, C -> G * C
+// dbeta:  1, 1, 1, G, C -> G * C
+
+// reduced axis: 0, 1, 2
+
+using GammaBetaDeviceInstance = ck::tensor_operation::device::DeviceNormalizationBwdGammaBetaImpl<
+    DYDataType,
+    XDataType,
+    MeanInvStdDataType,
+    ComputeDataType,
+    DGammaDataType,
+    DBetaDataType,
+    Rank,
+    NumReduceDim,
+    256,   // BlockSize
+    8,     // ClusterInvarient
+    32,    // ClusterReduce
+    8,     // SliceInvarient
+    1,     // SliceReduce
+    false, // IsDYFastestDimReduced
+    8,     // DYSrcVectorSize
+    false, // IsXFastestDimReduced
+    8,     // XSrcVectorSize
+    false, // IsMeanInvStdFastestDimReduced
+    1,     // MeanInvStdSrcVectorSize
+    1,     // DGammaDstVectorSize
+    1>;    // DBetaDstVectorSize
+
+int main()
+{
+    bool time_kernel = false;
+
+    ck::index_t N = 16;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 32;
+    ck::index_t C = 64;
+
+    Tensor<DYDataType> dy({N, H, W, G, C});
+    Tensor<XDataType> x({N, H, W, G, C});
+    Tensor<GammaDataType> gamma({G, C});
+    Tensor<MeanInvStdDataType> mean({N, G});
+    Tensor<MeanInvStdDataType> inv_std({N, G});
+
+    Tensor<DGammaDataType> dgamma({G, C});
+    Tensor<DBetaDataType> dbeta({G, C});
+    Tensor<DXDataType> dx({N, H, W, G, C});
+
+    dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0.0, 1.0});
+    x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+    gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
+    mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
+    inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
+
+    DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
+    DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
+    DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
+    DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
+
+    dy_dev.ToDevice(dy.mData.data());
+    x_dev.ToDevice(x.mData.data());
+    mean_dev.ToDevice(mean.mData.data());
+    inv_std_dev.ToDevice(inv_std.mData.data());
+
+    std::vector<ck::index_t> dyStrides{dy.mDesc.GetStrides().begin(), dy.mDesc.GetStrides().end()};
+    std::vector<ck::index_t> xStrides{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()};
+    std::vector<ck::index_t> meanStrides   = {G, 0, 0, 1, 0};
+    std::vector<ck::index_t> invStdStrides = {G, 0, 0, 1, 0};
+
+    auto gamma_beta_device_instance = GammaBetaDeviceInstance{};
+    auto gamma_beta_argument_ptr =
+        gamma_beta_device_instance.MakeArgumentPointer({N, H, W, G, C}, // inLengths
+                                                       dyStrides,       // dyStrides
+                                                       xStrides,        // xStrides
+                                                       meanStrides,     // meanStrides
+                                                       invStdStrides,   // invStdStrides
+                                                       {G, C},          // outLengths
+                                                       {C, 1},          // dgammaStrides
+                                                       {C, 1},          // dbetaStrides
+                                                       {0, 1, 2},       // reduceDims
+                                                       dy_dev.GetDeviceBuffer(),
+                                                       x_dev.GetDeviceBuffer(),
+                                                       mean_dev.GetDeviceBuffer(),
+                                                       inv_std_dev.GetDeviceBuffer(),
+                                                       dgamma_dev.GetDeviceBuffer(),
+                                                       dbeta_dev.GetDeviceBuffer());
+
+    if(!gamma_beta_device_instance.IsSupportedArgument(gamma_beta_argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    auto gamma_beta_invoker_ptr = gamma_beta_device_instance.MakeInvokerPointer();
+    gamma_beta_invoker_ptr->Run(gamma_beta_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+    {
+        Tensor<DGammaDataType> host_dgamma({G, C});
+        Tensor<DBetaDataType> host_dbeta({G, C});
+        Tensor<DXDataType> host_dx({N, H, W, G, C});
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceGroupnormBwd<DYDataType,
+                                                              XDataType,
+                                                              GammaDataType,
+                                                              MeanInvStdDataType,
+                                                              DGammaDataType,
+                                                              DBetaDataType,
+                                                              DXDataType,
+                                                              ComputeDataType>;
+
+        ReferenceInstance ref;
+        auto ref_argument = ref.MakeArgument(
+            dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, {N, H, W, G, C});
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        dgamma_dev.FromDevice(dgamma.mData.data());
+        dbeta_dev.FromDevice(dbeta.mData.data());
+
+        pass &= ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
+        pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
+    }
+
+    return (pass ? 0 : 1);
+}

example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp

@@ -34,6 +34,7 @@ using AccDataType      = F32;
 using CShuffleDataType = F32;
 using DDataType        = F16;
 using EDataType        = F16;
+using ComputeDataType  = F16;
 
 static constexpr ck::index_t NumDimM = 2;
 static constexpr ck::index_t NumDimN = 2;
@@ -291,6 +292,7 @@ int main(int argc, char* argv[])
                                                                       BDataType,
                                                                       CShuffleDataType,
                                                                       AccDataType,
+                                                                      ComputeDataType,
                                                                       PassThrough,
                                                                       BElementOp>;
 

example/62_conv_fwd_activ/CMakeLists.txt

@@ -30,6 +30,18 @@ foreach(gpu IN LISTS GPU_TARGETS)
       # Elu
       add_example_executable(example_convnd_fwd_xdl_elu_fp16 convnd_fwd_xdl_elu_fp16.cpp)
       add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_elu_fp16)
+      # ScaleAdd on A and B
+      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp16 multi_AB/conv_fwd_xdl_scaleadd_ab_fp16.cpp)
+      add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_fp16)
+      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp32 multi_AB/conv_fwd_xdl_scaleadd_ab_fp32.cpp)
+      add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_fp32)
+      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_bf16 multi_AB/conv_fwd_xdl_scaleadd_ab_bf16.cpp)
+      add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_bf16)
+      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_int8 multi_AB/conv_fwd_xdl_scaleadd_ab_int8.cpp)
+      add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_int8)
+      # ScaleAdd ScaleAdd Relu
+      add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp)
+      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16)
    set(target 1)
  endif()
 endforeach()

example/62_conv_fwd_activ/convnd_fwd_activ_common.hpp

@@ -11,7 +11,7 @@
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
 
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -47,7 +47,7 @@ static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecializatio
 
 template <typename OutElementOp>
 using DeviceGroupedConvNDFwdInstance =
-    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
         NDimSpatial,
         InLayout,
         WeiLayout,
@@ -190,9 +190,8 @@ bool run_grouped_conv_fwd(bool do_verification,
 
     if(!conv.IsSupportedArgument(argument))
     {
-        throw std::runtime_error(
-            "wrong! device_conv with the specified compilation parameters does "
-            "not support this Conv problem");
+        throw std::runtime_error("The device op with the specified compilation parameters does "
+                                 "not support this convolution problem.");
     }
 
     float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});

example/62_conv_fwd_activ/convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+constexpr ck::index_t NDimSpatial = 3;
+using InDataType                  = ck::half_t;
+using WeiDataType                 = ck::half_t;
+using AccDataType                 = float;
+using CShuffleDataType            = ck::half_t;
+using OutDataType                 = ck::half_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InLayout  = ck::tensor_layout::convolution::GNDHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::GNDHWK;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+using OutElementOp = ck::tensor_operation::element_wise::ScaleAddScaleAddRelu;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <typename OutElementOp>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<OutLayout, OutLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<OutDataType, OutDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8>;
+
+using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
+
+namespace {
+// Use custom implementation to pass two more tensors for post op
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    constexpr ck::index_t NumDs = 2;
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+    std::array<Tensor<OutDataType>, NumDs> d_tensors = {Tensor<OutDataType>(out_g_n_k_wos_desc),
+                                                        Tensor<OutDataType>(out_g_n_k_wos_desc)};
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
+        d_tensors[0].GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
+        d_tensors[1].GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.05, 0.05});
+        d_tensors[0].GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.05, 0.05});
+        d_tensors[1].GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.05, 0.05});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem d0_buf(sizeof(OutDataType) * d_tensors[0].mDesc.GetElementSpaceSize());
+    DeviceMem d1_buf(sizeof(OutDataType) * d_tensors[1].mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    d0_buf.ToDevice(d_tensors[0].mData.data());
+    d1_buf.ToDevice(d_tensors[1].mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    const std::array<const void*, NumDs> ds = {d0_buf.GetDeviceBuffer(), d1_buf.GetDeviceBuffer()};
+
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      ds,
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, NumDs>{
+                                          e_g_n_k_wos_lengths, e_g_n_k_wos_lengths},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, NumDs>{
+                                          e_g_n_k_wos_strides, e_g_n_k_wos_strides},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("The device op with the specified compilation parameters does "
+                                 "not support this convolution problem.");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop =
+        conv_param.GetFlops() + 2 * conv_param.GetOutputByte<OutDataType>() / sizeof(OutDataType);
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>() +
+                            2 * conv_param.GetOutputByte<OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv =
+            ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                         InDataType,
+                                                         WeiDataType,
+                                                         OutDataType,
+                                                         InElementOp,
+                                                         WeiElementOp,
+                                                         OutElementOp,
+                                                         0, /*Num A Elementwise Tensors*/
+                                                         0, /*Num B Elementwise Tensors*/
+                                                         NumDs>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op,
+                                                  {},
+                                                  {},
+                                                  d_tensors);
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device, out_host, "Error: incorrect results!");
+    }
+
+    return true;
+}
+
+} // namespace
+
+#include "run_convnd_fwd_activ_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }

example/62_conv_fwd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_multi_ab_common.hpp"
+
+using DataType    = ck::bhalf_t;
+using AccDataType = float;
+using InDataType  = DataType;
+using WeiDataType = DataType;
+using OutDataType = DataType;
+using ADataTypes  = ck::Tuple<DataType, DataType>;
+using BDataTypes  = ck::Tuple<DataType, DataType>;
+
+using InElementOp  = ck::tensor_operation::element_wise::ScaleAdd;
+using WeiElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+
+using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDMultiABFwdInstance<DataType,
+                                                                                  AccDataType,
+                                                                                  ADataTypes,
+                                                                                  BDataTypes,
+                                                                                  InElementOp,
+                                                                                  WeiElementOp>;
+
+#include "../run_convnd_fwd_activ_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }

example/62_conv_fwd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_multi_ab_common.hpp"
+
+using DataType    = ck::half_t;
+using AccDataType = float;
+using InDataType  = DataType;
+using WeiDataType = DataType;
+using OutDataType = DataType;
+using ADataTypes  = ck::Tuple<DataType, DataType>;
+using BDataTypes  = ck::Tuple<DataType, DataType>;
+
+using InElementOp  = ck::tensor_operation::element_wise::ScaleAdd;
+using WeiElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+
+using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDMultiABFwdInstance<DataType,
+                                                                                  AccDataType,
+                                                                                  ADataTypes,
+                                                                                  BDataTypes,
+                                                                                  InElementOp,
+                                                                                  WeiElementOp>;
+
+#include "../run_convnd_fwd_activ_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }

example/62_conv_fwd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_multi_ab_common.hpp"
+
+using DataType    = float;
+using AccDataType = float;
+using InDataType  = DataType;
+using WeiDataType = DataType;
+using OutDataType = DataType;
+using ADataTypes  = ck::Tuple<DataType, DataType>;
+using BDataTypes  = ck::Tuple<DataType, DataType>;
+
+using InElementOp  = ck::tensor_operation::element_wise::ScaleAdd;
+using WeiElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+
+using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDMultiABFwdInstance<DataType,
+                                                                                  AccDataType,
+                                                                                  ADataTypes,
+                                                                                  BDataTypes,
+                                                                                  InElementOp,
+                                                                                  WeiElementOp>;
+
+#include "../run_convnd_fwd_activ_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }

example/62_conv_fwd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_multi_ab_common.hpp"
+
+using DataType    = int8_t;
+using AccDataType = int32_t;
+using InDataType  = DataType;
+using WeiDataType = DataType;
+using OutDataType = DataType;
+using ADataTypes  = ck::Tuple<DataType, DataType>;
+using BDataTypes  = ck::Tuple<DataType, DataType>;
+
+using InElementOp  = ck::tensor_operation::element_wise::ScaleAdd;
+using WeiElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+
+using DeviceGroupedConvNDFwdActivInstance = DeviceGroupedConvNDMultiABFwdInstance<DataType,
+                                                                                  AccDataType,
+                                                                                  ADataTypes,
+                                                                                  BDataTypes,
+                                                                                  InElementOp,
+                                                                                  WeiElementOp>;
+
+#include "../run_convnd_fwd_activ_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_example(argc, argv); }

example/62_conv_fwd_activ/multi_AB/convnd_fwd_activ_multi_ab_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+constexpr ck::index_t NDimSpatial = 3;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InLayout  = ck::tensor_layout::convolution::GNDHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::GNDHWK;
+
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <typename DataType,
+          typename AccDataType,
+          typename InDataTypes,
+          typename WeiDataTypes,
+          typename InElementOp,
+          typename WeiElementOp>
+using DeviceGroupedConvNDMultiABFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataTypes,
+        WeiDataTypes,
+        AccDataType,
+        DataType,
+        ck::Tuple<>,
+        DataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8>;
+
+namespace {
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    constexpr ck::index_t NumAs = 2;
+    constexpr ck::index_t NumBs = 2;
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<InDataType> in_bias(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<WeiDataType> wei_bias(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
+        in_bias.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
+        wei_bias.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
+        in_bias.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.05, 0.05});
+        wei_bias.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1.0, 1.0});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem in_bias_device_buf(sizeof(InDataType) * in_bias.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem wei_bias_device_buf(sizeof(WeiDataType) * wei_bias.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    in_bias_device_buf.ToDevice(in_bias.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    wei_bias_device_buf.ToDevice(wei_bias.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    std::array<const void*, NumAs> as{in_device_buf.GetDeviceBuffer(),
+                                      in_bias_device_buf.GetDeviceBuffer()};
+    std::array<const void*, NumBs> bs{wei_device_buf.GetDeviceBuffer(),
+                                      wei_bias_device_buf.GetDeviceBuffer()};
+    std::array<const void*, 0> ds{};
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(as,
+                                      bs,
+                                      ds,
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      {},
+                                      {},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = conv_param.GetFlops() +
+                       2 * conv_param.GetOutputByte<InDataType>() / sizeof(InDataType) +
+                       2 * conv_param.GetOutputByte<WeiDataType>() / sizeof(WeiDataType);
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>() +
+                            conv_param.GetInputByte<InDataType>() +
+                            conv_param.GetWeightByte<WeiDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        const std::array<Tensor<InDataType>, NumAs - 1> elementwise_a_tensors  = {in_bias};
+        const std::array<Tensor<WeiDataType>, NumBs - 1> elementwise_b_tensors = {wei_bias};
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp,
+                                                                     NumAs - 1,
+                                                                     NumBs - 1>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op,
+                                                  elementwise_a_tensors,
+                                                  elementwise_b_tensors);
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device, out_host, "Error: incorrect results!");
+    }
+
+    return true;
+}
+
+} // namespace