// Copyright (c) Facebook, Inc. and its affiliates.
// All rights reserved.
//
// 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.

#pragma once

#include <gtest/gtest.h>

#include <algorithm>
#include <cassert>
#include <cstddef>
#include <cstdlib>
#include <functional>
#include <limits>
#include <random>
#include <vector>

#include <xnnpack.h>
#include <xnnpack/params-init.h>
#include <xnnpack/params.h>
#include <xnnpack/requantization.h>


class VAddCMicrokernelTester {
 public:
  enum class Variant {
    Native,
    Scalar,
  };

  inline VAddCMicrokernelTester& batch_size(size_t batch_size) {
    assert(batch_size != 0);
    this->batch_size_ = batch_size;
    return *this;
  }

  inline size_t batch_size() const {
    return this->batch_size_;
  }

  inline VAddCMicrokernelTester& inplace(bool inplace) {
    this->inplace_ = inplace;
    return *this;
  }

  inline bool inplace() const {
    return this->inplace_;
  }

  inline VAddCMicrokernelTester& a_scale(float a_scale) {
    assert(a_scale > 0.0f);
    assert(std::isnormal(a_scale));
    this->a_scale_ = a_scale;
    return *this;
  }

  inline float a_scale() const {
    return this->a_scale_;
  }

  inline VAddCMicrokernelTester& a_zero_point(uint8_t a_zero_point) {
    this->a_zero_point_ = a_zero_point;
    return *this;
  }

  inline uint8_t a_zero_point() const {
    return this->a_zero_point_;
  }

  inline VAddCMicrokernelTester& b_scale(float b_scale) {
    assert(b_scale > 0.0f);
    assert(std::isnormal(b_scale));
    this->b_scale_ = b_scale;
    return *this;
  }

  inline float b_scale() const {
    return this->b_scale_;
  }

  inline VAddCMicrokernelTester& b_zero_point(uint8_t b_zero_point) {
    this->b_zero_point_ = b_zero_point;
    return *this;
  }

  inline uint8_t b_zero_point() const {
    return this->b_zero_point_;
  }

  inline VAddCMicrokernelTester& y_scale(float y_scale) {
    assert(y_scale > 0.0f);
    assert(std::isnormal(y_scale));
    this->y_scale_ = y_scale;
    return *this;
  }

  inline float y_scale() const {
    return this->y_scale_;
  }

  inline VAddCMicrokernelTester& y_zero_point(uint8_t y_zero_point) {
    this->y_zero_point_ = y_zero_point;
    return *this;
  }

  inline uint8_t y_zero_point() const {
    return this->y_zero_point_;
  }

  inline VAddCMicrokernelTester& qmin(uint8_t qmin) {
    this->qmin_ = qmin;
    return *this;
  }

  inline uint8_t qmin() const {
    return this->qmin_;
  }

  inline VAddCMicrokernelTester& qmax(uint8_t qmax) {
    this->qmax_ = qmax;
    return *this;
  }

  inline uint8_t qmax() const {
    return this->qmax_;
  }

  inline VAddCMicrokernelTester& iterations(size_t iterations) {
    this->iterations_ = iterations;
    return *this;
  }

  inline size_t iterations() const {
    return this->iterations_;
  }

  void Test(xnn_qs8_vadd_minmax_ukernel_function vadd_minmax, Variant variant = Variant::Native) const {
    std::random_device random_device;
    auto rng = std::mt19937(random_device());
    auto i8rng = std::bind(
      std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), rng);

    std::vector<int8_t> a(batch_size() + XNN_EXTRA_BYTES / sizeof(int8_t));
    std::vector<int8_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(int8_t) : 0));
    std::vector<float> y_fp(batch_size());
    std::vector<int8_t> y_ref(batch_size());
    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      std::generate(a.begin(), a.end(), std::ref(i8rng));
      if (inplace()) {
        std::generate(y.begin(), y.end(), std::ref(i8rng));
      } else {
        std::fill(y.begin(), y.end(), 0xA5);
      }
      const int8_t* a_data = inplace() ? y.data() : a.data();
      const int8_t b = i8rng();

      // Prepare parameters.
      xnn_qs8_add_params quantization_params = { };
      switch (variant) {
        case Variant::Native:
          quantization_params = xnn_init_qs8_add_params(
            int8_t(a_zero_point() - 0x80), int8_t(b_zero_point() - 0x80), int8_t(y_zero_point() - 0x80),
            a_scale() / y_scale(), b_scale() / y_scale(),
            int8_t(qmin() - 0x80), int8_t(qmax() - 0x80));
          break;
        case Variant::Scalar:
          quantization_params = xnn_init_scalar_qs8_add_params(
            int8_t(a_zero_point() - 0x80), int8_t(b_zero_point() - 0x80), int8_t(y_zero_point() - 0x80),
            a_scale() / y_scale(), b_scale() / y_scale(),
            int8_t(qmin() - 0x80), int8_t(qmax() - 0x80));
          break;
      }
      const xnn_qs8_add_params scalar_quantization_params =
          xnn_init_scalar_qs8_add_params(
            int8_t(a_zero_point() - 0x80), int8_t(b_zero_point() - 0x80), int8_t(y_zero_point() - 0x80),
            a_scale() / y_scale(), b_scale() / y_scale(),
            int8_t(qmin() - 0x80), int8_t(qmax() - 0x80));

      // Compute reference results.
      for (size_t i = 0; i < batch_size(); i++) {
        y_fp[i] = float(int32_t(y_zero_point() - 0x80)) +
          float(int32_t(a_data[i]) - int32_t(a_zero_point() - 0x80)) * (a_scale() / y_scale()) +
          float(int32_t(b) - int32_t(b_zero_point() - 0x80)) * (b_scale() / y_scale());
        y_fp[i] = std::min<float>(y_fp[i], float(int32_t(qmax() - 0x80)));
        y_fp[i] = std::max<float>(y_fp[i], float(int32_t(qmin() - 0x80)));
        y_ref[i] = xnn_qs8_quantize_add(a_data[i], b, scalar_quantization_params);
      }

      // Call optimized micro-kernel.
      vadd_minmax(batch_size(), a_data, &b, y.data(), &quantization_params);

      // Verify results.
      for (size_t i = 0; i < batch_size(); i++) {
        ASSERT_LE(int32_t(y[i]), int32_t(qmax() - 0x80))
          << "at element " << i << " / " << batch_size();
        ASSERT_GE(int32_t(y[i]), int32_t(qmin() - 0x80))
          << "at element " << i << " / " << batch_size();
        ASSERT_EQ(int32_t(y_ref[i]), int32_t(y[i]))
          << "at element " << i << " / " << batch_size();
        ASSERT_NEAR(float(int32_t(y[i])), y_fp[i], 0.6f)
          << "at element " << i << " / " << batch_size();
      }
    }
  }

 private:
  size_t batch_size_{1};
  bool inplace_{false};
  float a_scale_{0.75f};
  float b_scale_{1.25f};
  float y_scale_{0.96875f};
  uint8_t a_zero_point_{121};
  uint8_t b_zero_point_{127};
  uint8_t y_zero_point_{133};
  uint8_t qmin_{0};
  uint8_t qmax_{255};
  size_t iterations_{15};
};