Support 1 dimension

example/19_binary_elementwise/CMakeLists.txt

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

example/19_binary_elementwise/broadcast_add_2d.cpp

@@ -67,6 +67,11 @@ int main()
     ck::index_t N      = 1024;
     ck::index_t Stride = 1024;
 
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
     auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
         return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
                                     std::vector<std::size_t>({stride, 1}));
@@ -74,8 +79,7 @@ int main()
 
     Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride));
 
-    Tensor<ABDataType> b_n(std::vector<std::size_t>({static_cast<std::size_t>(N)}),
-                           std::vector<std::size_t>({1}));
+    Tensor<ABDataType> b_n(f_host_tensor_descriptor1d(N, 1));
 
     Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
 

example/19_binary_elementwise/elementwise_add_1d.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_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, 1, 8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor,
+          int broadcastDim>
+void host_elementwise1D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, Functor functor)
+{
+    for(int m = 0; m < M; ++m)
+    {
+        ComputeDataType Am = static_cast<ComputeDataType>(A(m));
+        ComputeDataType Bm = static_cast<ComputeDataType>(B(m));
+        ComputeDataType Cm = 0;
+        functor(Cm, Am, Bm);
+        C(m) = static_cast<ComputeDataType>(Cm);
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1));
+
+    Tensor<ABDataType> b_m(f_host_tensor_descriptor1d(M, 1));
+
+    Tensor<ABDataType> c_m(f_host_tensor_descriptor1d(M, 1));
+
+    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(),
+                                                     {M},
+                                                     {1},
+                                                     {1},
+                                                     {1},
+                                                     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(f_host_tensor_descriptor1d(M, 1));
+
+        host_elementwise1D<Tensor<ABDataType>,
+                           Tensor<ABDataType>,
+                           Tensor<CDataType>,
+                           EltwiseComputeDataType,
+                           Add,
+                           0>(host_c_m, a_m, b_m, M, 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

@@ -21,6 +21,27 @@ struct DeviceBinaryElementwise : public BaseOperator
 {
     static constexpr auto I0 = Number<0>{};
 
+    static auto MakeDescriptor_M0_1d(const std::vector<int>& shape,
+                                     const std::vector<int>& stride,
+                                     index_t gridSize,
+                                     index_t threadPerBlock)
+    {
+        // 1d desc - [m]
+        const auto desc_m0 =
+            make_naive_tensor_descriptor(make_tuple(shape[0]), make_tuple(stride[0]));
+
+        // pad
+        const auto m0           = desc_m0.GetLength(I0);
+        const index_t loop_step = gridSize * threadPerBlock * ScalarPerVector;
+        const auto pad          = math::integer_least_multiple(m0, loop_step) - m0;
+        const auto desc_m0_pad =
+            transform_tensor_descriptor(desc_m0,
+                                        make_tuple(make_right_pad_transform(m0, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m0_pad;
+    }
+
     static auto MakeDescriptor_M0_2d(const std::vector<int>& shape,
                                      const std::vector<int>& stride,
                                      index_t gridSize,
@@ -57,7 +78,9 @@ struct DeviceBinaryElementwise : public BaseOperator
                                   index_t gridSize,
                                   index_t threadPerBlock)
     {
-        if constexpr(Dim == 2)
+        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);
         else
             return make_naive_tensor_descriptor(make_tuple(0), make_tuple(0));