// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <algorithm>
#include <array>
#include <cmath>
#include <functional>
#include <limits>
#include <random>
#include <vector>

#include <xnnpack.h>

#include <benchmark/benchmark.h>
#include "bench/utils.h"
#ifdef BENCHMARK_TENSORFLOW_LITE
#include "flatbuffers/include/flatbuffers/flatbuffers.h"
#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/schema/schema_generated.h"
#include "tensorflow/lite/version.h"
#endif  // BENCHMARK_TENSORFLOW_LITE


static void xnnpack_square_root_f32(benchmark::State& state) {
  const size_t batch_size = state.range(0);
  const size_t channels = state.range(1);

  std::random_device random_device;
  auto rng = std::mt19937(random_device());
  auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 5.0f), std::ref(rng));

  std::vector<float> input(batch_size * channels);
  std::vector<float> output(batch_size * channels);
  std::generate(input.begin(), input.end(), std::ref(f32rng));
  std::fill(output.begin(), output.end(), std::nanf(""));

  xnn_status status = xnn_initialize(nullptr /* allocator */);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to initialize XNNPACK");
    return;
  }

  xnn_operator_t sqrt_op = nullptr;
  status = xnn_create_square_root_nc_f32(
    channels, channels /* input stride */, channels /* output stride */,
    0 /* flags */, &sqrt_op);
  if (status != xnn_status_success || sqrt_op == nullptr) {
    state.SkipWithError("failed to create Square Root operator");
    return;
  }

  status = xnn_setup_square_root_nc_f32(
    sqrt_op,
    batch_size,
    input.data(), output.data(),
    nullptr /* thread pool */);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to setup Square Root operator");
    return;
  }

  for (auto _ : state) {
    status = xnn_run_operator(sqrt_op, nullptr /* thread pool */);
    if (status != xnn_status_success) {
      state.SkipWithError("failed to run Square Root operator");
      return;
    }
  }

  status = xnn_delete_operator(sqrt_op);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to delete Square Root operator");
    return;
  }

  const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
  if (cpu_frequency != 0) {
    state.counters["cpufreq"] = cpu_frequency;
  }

  const size_t elements_per_iteration = batch_size * channels;
  state.counters["elements"] =
    benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);

  const size_t bytes_per_iteration = 2 * elements_per_iteration * sizeof(float);
  state.counters["bytes"] =
    benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
}

#ifdef BENCHMARK_TENSORFLOW_LITE
static void tflite_square_root_f32(benchmark::State& state) {
  const size_t batch_size = state.range(0);
  const size_t channels = state.range(1);

  std::random_device random_device;
  auto rng = std::mt19937(random_device());
  auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 5.0f), std::ref(rng));

  flatbuffers::FlatBufferBuilder builder;
  const flatbuffers::Offset<tflite::OperatorCode> operator_code =
      CreateOperatorCode(builder, tflite::BuiltinOperator_SQRT);

  const std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers{{
    tflite::CreateBuffer(builder, builder.CreateVector({})),
  }};

  const std::array<int32_t, 4> input_shape{{
    static_cast<int32_t>(batch_size),
    static_cast<int32_t>(1 /* height */),
    static_cast<int32_t>(1 /* width */),
    static_cast<int32_t>(channels)
  }};
  const std::array<int32_t, 4> output_shape{{
    static_cast<int32_t>(batch_size),
    static_cast<int32_t>(1 /* height */),
    static_cast<int32_t>(1 /* width */),
    static_cast<int32_t>(channels)
  }};

  const std::array<flatbuffers::Offset<tflite::Tensor>, 2> tensors{{
    tflite::CreateTensor(builder,
                         builder.CreateVector<int32_t>(input_shape.data(), input_shape.size()),
                         tflite::TensorType_FLOAT32),
    tflite::CreateTensor(builder,
                         builder.CreateVector<int32_t>(output_shape.data(), output_shape.size()),
                         tflite::TensorType_FLOAT32),
  }};

  const std::array<int32_t, 1> op_inputs{{ 0 }};
  const std::array<int32_t, 1> op_outputs{{ 1 }};
  flatbuffers::Offset<tflite::Operator> op = tflite::CreateOperator(
      builder,
      0 /* opcode_index */,
      builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()),
      builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size()));

  const std::array<int32_t, 1> graph_inputs{{ 0 }};
  const std::array<int32_t, 1> graph_outputs{{ 1 }};
  const flatbuffers::Offset<tflite::SubGraph> subgraph = tflite::CreateSubGraph(
      builder,
      builder.CreateVector(tensors.data(), tensors.size()),
      builder.CreateVector<int32_t>(graph_inputs.data(), graph_inputs.size()),
      builder.CreateVector<int32_t>(graph_outputs.data(), graph_outputs.size()),
      builder.CreateVector(&op, 1));

  const flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder,
      TFLITE_SCHEMA_VERSION,
      builder.CreateVector(&operator_code, 1),
      builder.CreateVector(&subgraph, 1),
      builder.CreateString("Square Root model"),
      builder.CreateVector(buffers.data(), buffers.size()));

  builder.Finish(model_buffer);

  const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer());
  tflite::ops::builtin::BuiltinOpResolver resolver;
  tflite::InterpreterBuilder interpreterBuilder(model, resolver);
  std::unique_ptr<tflite::Interpreter> interpreter;
  if (interpreterBuilder(&interpreter) != kTfLiteOk) {
    state.SkipWithError("failed to create TFLite interpreter");
    return;
  }
  if (interpreter == nullptr) {
    state.SkipWithError("TFLite interpreter is null");
    return;
  }
  interpreter->SetNumThreads(1);

  if (interpreter->AllocateTensors() != kTfLiteOk) {
    state.SkipWithError("failed to allocate tensors");
    return;
  }

  std::generate(
    interpreter->typed_tensor<float>(0),
    interpreter->typed_tensor<float>(0) + batch_size * channels,
    std::ref(f32rng));

  for (auto _ : state) {
    if (interpreter->Invoke() != kTfLiteOk) {
      state.SkipWithError("failed to invoke TFLite interpreter");
      return;
    }
  }

  const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
  if (cpu_frequency != 0) {
    state.counters["cpufreq"] = cpu_frequency;
  }

  const size_t elements_per_iteration = batch_size * channels;
  state.counters["elements"] =
    benchmark::Counter(uint64_t(state.iterations()) * elements_per_iteration, benchmark::Counter::kIsRate);

  const size_t bytes_per_iteration = 2 * elements_per_iteration * sizeof(float);
  state.counters["bytes"] =
    benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);

  interpreter.reset();
}
#endif  // BENCHMARK_TENSORFLOW_LITE

static void CharacteristicArguments(benchmark::internal::Benchmark* b)
{
  b->ArgNames({"N", "C"});

  int32_t c = 16;
  for (int32_t n = 224; n >= 7; n /= 2) {
    b->Args({n * n, c});
    c *= 2;
  }
}

BENCHMARK(xnnpack_square_root_f32)->Apply(CharacteristicArguments)->UseRealTime();

#ifdef BENCHMARK_TENSORFLOW_LITE
  BENCHMARK(tflite_square_root_f32)->Apply(CharacteristicArguments)->UseRealTime();
#endif  // BENCHMARK_TENSORFLOW_LITE

#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif