/*
 * Copyright (c) 2017 ARM Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#include "arm_compute/core/CL/kernels/CLNormalizationLayerKernel.h"

#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/FixedPoint.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"

using namespace arm_compute;

namespace
{
Error validate_arguments(const ITensorInfo *input, const ITensorInfo *output, NormalizationLayerInfo norm_info)
{
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);

    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd");

    if(is_data_type_fixed_point(input->data_type()))
    {
        ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.beta(), input);
        ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.kappa(), input);
        ARM_COMPUTE_RETURN_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.scale_coeff(), input);
    }

    // Checks performed when output is configured
    if(output->total_size() != 0)
    {
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
    }

    return Error{};
}

std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, NormalizationLayerInfo norm_info)
{
    // Output tensor auto initialization if not yet initialized
    auto_init_if_empty(*output, *input->clone());

    const unsigned int norm_size = norm_info.norm_size();
    bool               is_in_map = norm_info.is_in_map();

    const unsigned int border_width = is_in_map ? std::min(norm_size / 2, 3U) : 0;
    const BorderSize   border_size  = BorderSize(0, border_width);

    const unsigned int num_elems_processed_per_iteration = (is_data_type_fixed_point(input->data_type())) ? 16 : 4;
    const unsigned int num_elems_read_per_iteration      = is_in_map ? (num_elems_processed_per_iteration + 2 * (norm_size / 2)) : num_elems_processed_per_iteration;

    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));

    // We do not use a Rectangle window for IN_MAP_2D as we clamp the top and bottom accesses inside the kernel, avoiding padding
    AccessWindowHorizontal input_access(input, -border_size.left, num_elems_read_per_iteration);
    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);

    bool window_changed = update_window_and_padding(win, input_access, output_access);

    output_access.set_valid_region(win, input->valid_region());

    Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
    return std::make_pair(err, win);
}
} // namespace

CLNormalizationLayerKernel::CLNormalizationLayerKernel()
    : _input(nullptr), _output(nullptr), _border_size(0), _is_in_map(false)
{
}

BorderSize CLNormalizationLayerKernel::border_size() const
{
    return _border_size;
}

void CLNormalizationLayerKernel::configure(const ICLTensor *input, ICLTensor *output, NormalizationLayerInfo norm_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);

    // Output tensor auto initialization if not yet initialized
    auto_init_if_empty(*output->info(), *input->info()->clone());

    // Perform validation step
    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), norm_info));

    _input  = input;
    _output = output;

    _is_in_map                      = norm_info.is_in_map();
    const unsigned int border_width = _is_in_map ? std::min(norm_info.norm_size() / 2, 3U) : 0;
    _border_size                    = BorderSize(0, border_width);

    const unsigned int num_elems_processed_per_iteration = (is_data_type_fixed_point(input->info()->data_type())) ? 16 : 4;
    const bool         is_in_map_2D                      = (norm_info.type() == NormType::IN_MAP_2D);

    // Set build options
    CLBuildOptions build_opts;
    build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
    build_opts.add_option_if(is_data_type_fixed_point(input->info()->data_type()),
                             "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
    build_opts.add_option(("-DCOEFF=" + float_to_string_with_full_precision(norm_info.scale_coeff())));
    build_opts.add_option(("-DBETA=" + float_to_string_with_full_precision(norm_info.beta())));
    build_opts.add_option(("-DKAPPA=" + float_to_string_with_full_precision(norm_info.kappa())));
    build_opts.add_option(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
    build_opts.add_option(("-DRADIUS=" + support::cpp11::to_string(norm_info.norm_size() / 2)));
    build_opts.add_option(("-DNUM_SLICES=" + support::cpp11::to_string(input->info()->dimension(2))));
    build_opts.add_option_if(is_in_map_2D, "-DIN_MAP_2D");

    // Create kernel
    std::string kernel_name = _is_in_map ? "normalization_layer_in_map" : "normalization_layer_cross_map";
    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));

    // Configure kernel window
    auto win_config = validate_and_configure_window(input->info(), output->info(), norm_info);
    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
    ICLKernel::configure(win_config.second);

    // Set config_id for enabling LWS tuning
    _config_id = "normalization_layer_";
    _config_id += lower_string(string_from_data_type(input->info()->data_type()));
    _config_id += "_";
    _config_id += support::cpp11::to_string(static_cast<std::underlying_type<NormType>::type>(norm_info.type()));
    _config_id += "_";
    _config_id += support::cpp11::to_string(norm_info.norm_size());
    _config_id += "_";
    _config_id += support::cpp11::to_string(input->info()->dimension(0));
    _config_id += "_";
    _config_id += support::cpp11::to_string(input->info()->dimension(1));
}

Error CLNormalizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, NormalizationLayerInfo norm_info)
{
    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, norm_info));
    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), norm_info).first);

    return Error{};
}

void CLNormalizationLayerKernel::run(const Window &window, cl::CommandQueue &queue)
{
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);

    const int collapsed_dimension = _is_in_map ? Window::DimZ : 4;
    Window    window_collapsed    = window.collapse_if_possible(ICLKernel::window(), collapsed_dimension);
    Window    slice               = window_collapsed.first_slice_window_3D();

    do
    {
        unsigned int idx = 0;
        add_3D_tensor_argument(idx, _input, slice);
        add_3D_tensor_argument(idx, _output, slice);
        enqueue(queue, *this, slice);
    }
    while(window_collapsed.slide_window_slice_3D(slice));
}