// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include <memory>
#include <string>
#include "lite/backends/opencl/cl_half.h"
#include "lite/backends/opencl/cl_image_converter.h"
#include "lite/backends/opencl/cl_include.h"
#include "lite/core/kernel.h"
#include "lite/core/op_registry.h"
#include "lite/kernels/opencl/image_helper.h"
#include "lite/operators/op_params.h"
#include "lite/utils/logging.h"
#include "lite/utils/replace_stl/stream.h"

namespace paddle {
namespace lite {
namespace kernels {
namespace opencl {
class GridSamplerImageCompute : public KernelLite<TARGET(kOpenCL),
                                                  PRECISION(kFP16),
                                                  DATALAYOUT(kImageDefault)> {
 public:
  using param_t = operators::GridSamplerParam;

  std::string doc() const override {
    return "GridSampler using cl::Image2D(ImageDefault/RGBA), kFP32";
  }

  void PrepareForRun() override {
    grid_param_ = param_.get_mutable<param_t>();

    auto& context = ctx_->As<OpenCLContext>();
    context.cl_context()->AddKernel(
        kernel_func_name_, "image/grid_sampler_kernel.cl", build_options_);
  }

  void Run() override {
    auto& context = ctx_->As<OpenCLContext>();
    CHECK(context.cl_context() != nullptr);

    auto* x = grid_param_->x;
    auto* out = grid_param_->out;
    auto* grid = grid_param_->grid;
    auto out_dims = out->dims();
    auto in_dims = x->dims();

    VLOG(4) << "x->target():" << TargetToStr(x->target());
    VLOG(4) << "out->target():" << TargetToStr(out->target());
    VLOG(4) << "x->dims():" << in_dims;
    VLOG(4) << "out->dims():" << out_dims;

    auto out_image_shape = InitImageDimInfoWith(out_dims);
    auto* x_img = x->data<half_t, cl::Image2D>();
    VLOG(4) << "x_image: " << x_img;

    auto* grid_img = x->data<half_t, cl::Image2D>();
    VLOG(4) << "grid_img: " << grid_img;

    auto* out_img = out->mutable_data<half_t, cl::Image2D>(
        out_image_shape["width"], out_image_shape["height"]);
    VLOG(4) << "out_image" << out_img;
    VLOG(4) << "out_image_shape[w,h]:" << out_image_shape["width"] << " "
            << out_image_shape["height"];

    STL::stringstream kernel_key;
    kernel_key << kernel_func_name_ << build_options_;
    auto kernel = context.cl_context()->GetKernel(kernel_key.str());

    int arg_idx = 0;
    int out_height = out_dims[2];
    int out_width = out_dims[3];
    auto default_work_size =
        DefaultWorkSize(out_dims,
                        DDim(std::vector<DDim::value_type>{
                            static_cast<int64_t>(out_image_shape["width"]),
                            static_cast<int64_t>(out_image_shape["height"])}));

    cl_int status = kernel.setArg(arg_idx++, *x_img);
    CL_CHECK_FATAL(status);
    status = kernel.setArg(arg_idx++, *grid_img);
    CL_CHECK_FATAL(status);
    status = kernel.setArg(arg_idx++, *out_img);
    CL_CHECK_FATAL(status);
    status = kernel.setArg(arg_idx++, out_height);
    CL_CHECK_FATAL(status);
    status = kernel.setArg(arg_idx++, out_width);
    CL_CHECK_FATAL(status);

    auto global_work_size =
        cl::NDRange{static_cast<cl::size_type>(default_work_size[0]),
                    static_cast<cl::size_type>(default_work_size[2]),
                    static_cast<cl::size_type>(default_work_size[3] / 4)};

    status = context.cl_context()->GetCommandQueue().enqueueNDRangeKernel(
        kernel,
        cl::NullRange,
        global_work_size,
        cl::NullRange,
        nullptr,
        event_.get());
    CL_CHECK_FATAL(status);
    context.cl_wait_list()->emplace(out_img, event_);

    VLOG(4) << "global_work_size:[2D]:" << global_work_size[0] << " "
            << global_work_size[1] << " " << global_work_size[2];
  }

 protected:
  param_t* grid_param_{nullptr};
  std::string kernel_func_name_{"grid_sampler"};
  std::string build_options_{"-DCL_DTYPE_half"};
  std::shared_ptr<cl::Event> event_{new cl::Event};
};

}  // namespace opencl
}  // namespace kernels
}  // namespace lite
}  // namespace paddle

namespace ocl = paddle::lite::kernels::opencl;
REGISTER_LITE_KERNEL(grid_sampler,
                     kOpenCL,
                     kFP16,
                     kImageDefault,
                     ocl::GridSamplerImageCompute,
                     ImageDefault)
    .BindInput("X",
               {LiteType::GetTensorTy(TARGET(kOpenCL),
                                      PRECISION(kFP16),
                                      DATALAYOUT(kImageDefault))})
    .BindInput("Grid",
               {LiteType::GetTensorTy(TARGET(kOpenCL),
                                      PRECISION(kFP16),
                                      DATALAYOUT(kImageDefault))})
    .BindOutput("Output",
                {LiteType::GetTensorTy(TARGET(kOpenCL),
                                       PRECISION(kFP16),
                                       DATALAYOUT(kImageDefault))})
    .Finalize();