/*
 * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved
 * 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/graph.h"

#include "Benchmark.h"

#include <cstdlib>

using namespace arm_compute::graph::frontend;

inline std::unique_ptr<arm_compute::graph::ITensorAccessor> get_input_accessor(void)
{
    return get_accessor<InputAccessor>();
}

inline std::unique_ptr<arm_compute::graph::ITensorAccessor> get_random_accessor(float lower, float upper)
{
    return get_accessor<InputAccessor>();
}

inline std::unique_ptr<arm_compute::graph::ITensorAccessor> get_weights_accessor(const std::string &path, const std::string &data_file, DataLayout file_layout = DataLayout::NCHW)
{
    return get_accessor<InputAccessor>();
}

inline std::unique_ptr<arm_compute::graph::ITensorAccessor> get_output_accessor(void)
{
    return get_accessor<OutputAccessor>();
}

/** Example demonstrating how to implement MobileNet's network using the Compute Library's graph API
 *
 * @param[in] argc Number of arguments
 * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
 */
class GraphMobilenetExample
{
public:
    void do_setup(int argc, char **argv)
    {
        std::string data_path; /* Path to the trainable data */
        std::string image;     /* Image data */
        std::string label;     /* Label data */

        // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
        const int                  target                     = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
        Target                     target_hint                = set_target_hint(target);
        ConvolutionMethod          convolution_hint           = ConvolutionMethod::GEMM;
        DepthwiseConvolutionMethod depthwise_convolution_hint = DepthwiseConvolutionMethod::Optimized3x3;
        FastMathHint               fast_math_hint             = FastMathHint::Disabled;

        // Set model to execute. 0 (MobileNetV1_1.0_224), 1 (MobileNetV1_0.75_160)
        int model_id = (argc > 2) ? std::strtol(argv[2], nullptr, 10) : 0;
        ARM_COMPUTE_ERROR_ON_MSG(model_id > 1, "Invalid model ID. Model must be 0 (MobileNetV1_1.0_224) or 1 (MobileNetV1_0.75_160)");
        int layout_id = (argc > 3) ? std::strtol(argv[3], nullptr, 10) : 0;
        ARM_COMPUTE_ERROR_ON_MSG(layout_id > 1, "Invalid layout ID. Layout must be 0 (NCHW) or 1 (NHWC)");

        float            depth_scale           = (model_id == 0) ? 1.f : 0.75;
        unsigned int     spatial_size          = (model_id == 0) ? 224 : 160;
        std::string      model_path            = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/";
        TensorDescriptor input_descriptor_nchw = TensorDescriptor(TensorShape(spatial_size, spatial_size, 3U, 1U), DataType::F32);
        TensorDescriptor input_descriptor_nhwc = TensorDescriptor(TensorShape(3U, spatial_size, spatial_size, 1U), DataType::F32).set_layout(DataLayout::NHWC);
        TensorDescriptor input_descriptor      = (layout_id == 0) ? input_descriptor_nchw : input_descriptor_nhwc;

        // Parse arguments
        if(argc < 2)
        {
            // Print help
            std::cout << "Usage: " << argv[0] << " [target] [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
            std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
            std::cout << "No data layout provided: using NCHW\n\n";
            std::cout << "No data folder provided: using random values\n\n";
        }
        else if(argc == 2)
        {
            std::cout << "Usage: " << argv[0] << " " << argv[1] << " [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
            std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
            std::cout << "No data layout provided: using NCHW\n\n";
            std::cout << "No data folder provided: using random values\n\n";
        }
        else if(argc == 3)
        {
            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
            std::cout << "No data layout provided: using NCHW\n\n";
            std::cout << "No data folder provided: using random values\n\n";
        }
        else if(argc == 4)
        {
            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
            std::cout << "No data folder provided: using random values\n\n";
        }
        else if(argc == 5)
        {
            data_path = argv[4];
            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [image] [labels] [fast_math_hint]\n\n";
            std::cout << "No image provided: using random values\n\n";
            std::cout << "No text file with labels provided: skipping output accessor\n\n";
        }
        else if(argc == 6)
        {
            data_path = argv[4];
            image     = argv[5];
            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
            std::cout << "No text file with labels provided: skipping output accessor\n\n";
        }
        else if(argc == 7)
        {
            data_path = argv[4];
            image     = argv[5];
            label     = argv[6];
            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
            std::cout << "No fast math info provided: disabling fast math\n\n";
        }
        else
        {
            data_path      = argv[4];
            image          = argv[5];
            label          = argv[6];
            fast_math_hint = (std::strtol(argv[7], nullptr, 1) == 0) ? FastMathHint::Disabled : FastMathHint::Enabled;
        }

        // Add model path to data path
        if(!data_path.empty())
        {
            data_path += model_path;
        }

        graph << target_hint
              << convolution_hint
              << depthwise_convolution_hint
              << fast_math_hint
              << InputLayer(input_descriptor,
                            get_input_accessor())
              << ConvolutionLayer(
                  3U, 3U, 32U * depth_scale,
                  get_weights_accessor(data_path, "Conv2d_0_weights.npy", DataLayout::NCHW),
                  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
                  PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))
              .set_name("Conv2d_0")
              << BatchNormalizationLayer(
                  get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_mean.npy"),
                  get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_variance.npy"),
                  get_weights_accessor(data_path, "Conv2d_0_BatchNorm_gamma.npy"),
                  get_weights_accessor(data_path, "Conv2d_0_BatchNorm_beta.npy"),
                  0.001f)
              .set_name("Conv2d_0/BatchNorm")
              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name("Conv2d_0/Relu6");
        graph << get_dwsc_node(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << get_dwsc_node(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
        graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("Logits/AvgPool_1a")
              << ConvolutionLayer(
                  1U, 1U, 1001U,
                  get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW),
                  get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"),
                  PadStrideInfo(1, 1, 0, 0))
              .set_name("Logits/Conv2d_1c_1x1")
              << ReshapeLayer(TensorShape(1001U)).set_name("Reshape")
              << SoftmaxLayer().set_name("Softmax")
              << OutputLayer(get_output_accessor());

        // Finalize graph
        GraphConfig config;
        config.use_tuner = (target == 2);
        graph.finalize(target_hint, config);
    }
    void do_run()
    {
        run_benchmark(graph);
    }

private:
    Stream graph{ 0, "MobileNetV1" };

    ConcatLayer get_dwsc_node(const std::string &data_path, std::string &&param_path,
                              unsigned int  conv_filt,
                              PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info)
    {
        std::string total_path = param_path + "_";
        SubStream   sg(graph);
        sg << DepthwiseConvolutionLayer(
               3U, 3U,
               get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW),
               std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
               dwc_pad_stride_info)
           .set_name(total_path + "depthwise/depthwise")
           << BatchNormalizationLayer(
               get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"),
               get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"),
               get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"),
               get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"),
               0.001f)
           .set_name(total_path + "depthwise/BatchNorm")
           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "depthwise/Relu6")
           << ConvolutionLayer(
               1U, 1U, conv_filt,
               get_weights_accessor(data_path, total_path + "pointwise_weights.npy", DataLayout::NCHW),
               std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
               conv_pad_stride_info)
           .set_name(total_path + "pointwise/Conv2D")
           << BatchNormalizationLayer(
               get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_mean.npy"),
               get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_variance.npy"),
               get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_gamma.npy"),
               get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_beta.npy"),
               0.001f)
           .set_name(total_path + "pointwise/BatchNorm")
           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "pointwise/Relu6");

        return ConcatLayer(std::move(sg));
    }
};

/** Main program for MobileNetV1
 *
 * @param[in] argc Number of arguments
 * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner),
 *                             [optional] Model ID (0 = MobileNetV1_1.0_224, 1 = MobileNetV1_0.75_160),
 *                             [optional] Path to the weights folder,
 *                             [optional] image,
 *                             [optional] labels,
 *                             [optional] data layout,
 *                             [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
 */
int main(int argc, char **argv)
{
    GraphMobilenetExample example;

    example.do_setup(argc, argv);
    example.do_run();

    return 0;
}