Support any dimension for elementwise operation

example/19_binary_elementwise/CMakeLists.txt

 add_example_executable(example_broadcast_add_2d broadcast_add_2d.cpp)
-add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)
+add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)
\ No newline at end of file
+add_example_executable(example_elementwise_add_4d elementwise_add_4d.cpp)
\ No newline at end of file

example/19_binary_elementwise/elementwise_add_4d.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+#include <math.h>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_reduce_util.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::binary_element_wise::Add;
+
+using DeviceElementwiseAddInstance = ck::tensor_operation::device::
+    DeviceBinaryElementwise<F16, F16, CDataType, EltwiseComputeDataType, Add, 4, 8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor>
+void host_elementwise4D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    for(std::size_t n = 0; n < shape[0]; ++n)
+        for(std::size_t c = 0; c < shape[1]; ++c)
+            for(std::size_t h = 0; h < shape[2]; ++h)
+                for(std::size_t w = 0; w < shape[3]; ++w)
+                {
+                    ComputeDataType a_val = static_cast<ComputeDataType>(A(n, c, h, w));
+                    ComputeDataType b_val = static_cast<ComputeDataType>(B(n, c, h, w));
+                    ComputeDataType c_val = 0;
+                    functor(c_val, a_val, b_val);
+                    C(n, c, h, w) = static_cast<ComputeDataType>(c_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> nchw = {4, 16, 32, 32};
+
+    Tensor<ABDataType> a_m(nchw);
+    Tensor<ABDataType> b_m(nchw);
+    Tensor<ABDataType> c_m(nchw);
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpace());
+    DeviceMem b_m_device_buf(sizeof(ABDataType) * b_m.mDesc.GetElementSpace());
+    DeviceMem c_m_device_buf(sizeof(CDataType) * c_m.mDesc.GetElementSpace());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_device_buf.ToDevice(b_m.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(a_m_device_buf.GetDeviceBuffer(),
+                                                     b_m_device_buf.GetDeviceBuffer(),
+                                                     c_m_device_buf.GetDeviceBuffer(),
+                                                     ck::to_int_vector(nchw),
+                                                     ck::to_int_vector(a_m.mDesc.GetStrides()),
+                                                     ck::to_int_vector(b_m.mDesc.GetStrides()),
+                                                     ck::to_int_vector(c_m.mDesc.GetStrides()),
+                                                     Add{},
+                                                     256);
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise_2D instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_device_buf.FromDevice(c_m.mData.data());
+        Tensor<CDataType> host_c_m(nchw);
+
+        host_elementwise4D<Tensor<ABDataType>,
+                           Tensor<ABDataType>,
+                           Tensor<CDataType>,
+                           EltwiseComputeDataType,
+                           Add>(host_c_m, a_m, b_m, nchw, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m.mData, host_c_m.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

include/ck/tensor_operation/gpu/device/device_binary_elementwise.hpp

@@ -35,54 +35,30 @@ struct DeviceBinaryElementwise : public BaseOperator
         return desc_m0_pad;
     }
 
-    static auto MakeDescriptor_M0_1d(const std::vector<int>& shape,
-                                     const std::vector<int>& stride,
-                                     index_t gridSize,
-                                     index_t threadPerBlock)
-    {
-        const auto desc_m0 =
-            make_naive_tensor_descriptor(make_tuple(shape[0]), make_tuple(stride[0]));
-
-        return PadDescriptor_M0_1d(desc_m0, gridSize, threadPerBlock);
-    }
-
-    static auto MakeDescriptor_M0_2d(const std::vector<int>& shape,
-                                     const std::vector<int>& stride,
-                                     index_t gridSize,
-                                     index_t threadPerBlock)
-    {
-        const int m = shape[0];
-        const int n = shape[1];
-
-        // 2d desc - [m, n]
-        const auto desc_m_n =
-            make_naive_tensor_descriptor(make_tuple(m, n), make_tuple(stride[0], stride[1]));
-
-        // 1d desc - [m * n]
-        const auto desc_m0 =
-            transform_tensor_descriptor(desc_m_n,
-                                        make_tuple(make_merge_transform(make_tuple(m, n))),
-                                        make_tuple(Sequence<0, 1>{}),
-                                        make_tuple(Sequence<0>{}));
-
-        return PadDescriptor_M0_1d(desc_m0, gridSize, threadPerBlock);
-    }
-
     static auto MakeDescriptor_M0(const std::vector<int>& shape,
                                   const std::vector<int>& stride,
                                   index_t gridSize,
                                   index_t threadPerBlock)
     {
-        static_assert(Dim == 1 || Dim == 2,
-                      "wrong! DeviceBinaryElementwise not support this dimension");
-
-        // TODO - 3D, 4D, 5D
-        if constexpr(Dim == 1)
-            return MakeDescriptor_M0_1d(shape, stride, gridSize, threadPerBlock);
-        else if constexpr(Dim == 2)
-            return MakeDescriptor_M0_2d(shape, stride, gridSize, threadPerBlock);
+        auto tupleOfShape  = generate_tuple([&](auto I) { return shape[I]; }, Number<Dim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<Dim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        // merge nd to 1d desc - [s0 * s1 * ...]
+        if constexpr(Dim > 1)
+        {
+            const auto desc_m0 = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(tupleOfShape)),
+                make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<Dim>{})),
+                make_tuple(Sequence<0>{}));
+
+            return PadDescriptor_M0_1d(desc_m0, gridSize, threadPerBlock);
+        }
         else
-            return make_naive_tensor_descriptor(make_tuple(0), make_tuple(0));
+            return PadDescriptor_M0_1d(desc, gridSize, threadPerBlock);
     }
 
     using GridDesc_M0        = decltype(MakeDescriptor_M0({1, 1}, {1, 1}, 1, 1));
@@ -169,7 +145,7 @@ struct DeviceBinaryElementwise : public BaseOperator
         if(pArg == nullptr)
             return false;
 
-        // m * n
+        // shape[0] * shape[1] * shape[2] * ...
         const auto m0 = pArg->c_grid_desc_m0_.GetLength(I0);
 
         if(m0 % ScalarPerVector != 0)