// Copyright 2018-2022 David Robillard <d@drobilla.net>
// SPDX-License-Identifier: GPL-2.0-or-later

#include "bench_utils.hpp"
#include "test_utils.hpp"

#include "chilbert/BoundedBitVec.hpp" // IWYU pragma: keep
#include "chilbert/DynamicBitVec.hpp" // IWYU pragma: keep
#include "chilbert/SmallBitVec.hpp"
#include "chilbert/StaticBitVec.hpp" // IWYU pragma: keep

#include <array>
#include <cstddef>
#include <fstream>
#include <string>
#include <vector>

template<class T, size_t N>
struct BenchAnd {
  Duration operator()(Context& ctx)
  {
    const T v = make_random_bitvec<T, N>(ctx);
    T       r = make_random_bitvec<T, N>(ctx);

    return run_bench([&](auto) { r &= v; });
  }
};

template<class T, size_t N>
struct BenchOr {
  Duration operator()(Context& ctx)
  {
    const T v = make_random_bitvec<T, N>(ctx);
    T       r = make_random_bitvec<T, N>(ctx);

    return run_bench([&](auto) { r |= v; });
  }
};

template<class T, size_t N>
struct BenchXor {
  Duration operator()(Context& ctx)
  {
    const T v = make_random_bitvec<T, N>(ctx);
    T       r = make_random_bitvec<T, N>(ctx);

    return run_bench([&](auto) { r ^= v; });
  }
};

template<class T, size_t N>
struct BenchSetOne {
  Duration operator()(Context& ctx)
  {
    T v = make_random_bitvec<T, N>(ctx);
    return run_bench([&](const auto i) { v.set(i % N); });
  }
};

template<class T, size_t N>
struct BenchSetAll {
  Duration operator()(Context& ctx)
  {
    T v = make_random_bitvec<T, N>(ctx);
    return run_bench([&](auto) { v.set(); });
  }
};

template<class T, size_t N>
struct BenchResetOne {
  Duration operator()(Context& ctx)
  {
    T v = make_random_bitvec<T, N>(ctx);
    return run_bench([&](const auto i) { v.reset(i % N); });
  }
};

template<class T, size_t N>
struct BenchResetAll {
  Duration operator()(Context& ctx)
  {
    T v = make_random_bitvec<T, N>(ctx);
    return run_bench([&](auto) { v.reset(); });
  }
};

template<class T, size_t N>
struct BenchFlipOne {
  Duration operator()(Context& ctx)
  {
    T v = make_random_bitvec<T, N>(ctx);
    return run_bench([&](const auto i) { v.flip(i % N); });
  }
};

template<class T, size_t N>
struct BenchFlipAll {
  Duration operator()(Context& ctx)
  {
    T v = make_random_bitvec<T, N>(ctx);
    return run_bench([&](auto) { v.flip(); });
  }
};

template<class T, size_t N>
struct BenchNone {
  Duration operator()(Context& ctx)
  {
    T    v = make_random_bitvec<T, N>(ctx);
    bool r = false;
    (void)r;

    return run_bench([&](auto) { r = v.none(); });
  }
};

template<class T, size_t N>
struct BenchCount {
  Duration operator()(Context& ctx)
  {
    T               v = make_random_bitvec<T, N>(ctx);
    volatile size_t r = 0;
    (void)r;

    return run_bench([&](auto) { r = v.count(); });
  }
};

template<class T, size_t N>
struct BenchLeftShift {
  Duration operator()(Context& ctx)
  {
    const T v = make_random_bitvec<T, N>(ctx);
    T       r = v;
    return run_bench([&](const auto i) { r <<= (i % N); });
  }
};

template<class T, size_t N>
struct BenchRightShift {
  Duration operator()(Context& ctx)
  {
    const T v = make_random_bitvec<T, N>(ctx);
    T       r = v;
    return run_bench([&](const auto i) { r >>= (i % N); });
  }
};

template<class T, size_t N>
struct BenchLeftRotate {
  Duration operator()(Context& ctx)
  {
    T v = make_random_bitvec<T, N>(ctx);
    return run_bench([&](const auto i) { v.rotl(i % N); });
  }
};

template<class T, size_t N>
struct BenchRightRotate {
  Duration operator()(Context& ctx)
  {
    T             v = make_random_bitvec<T, N>(ctx);
    volatile bool r = false;
    (void)r;

    return run_bench([&](const auto i) {
      v.rotr(i % N);
      r = v.test(0);
    });
  }
};

template<class T, size_t N>
struct BenchFindFirst {
  Duration operator()(Context& ctx)
  {
    T               v = make_random_bitvec<T, N>(ctx);
    volatile size_t r = 0;
    (void)r;

    return run_bench([&](auto) { r = v.find_first(); });
  }
};

template<class T, size_t N>
struct BenchGrayCode {
  Duration operator()(Context& ctx)
  {
    const T v = make_random_bitvec<T, N>(ctx);
    T       r = v;

    return run_bench([&](auto) { chilbert::detail::gray_code(r); });
  }
};

template<class T, size_t N>
struct BenchGrayCodeInv {
  Duration operator()(Context& ctx)
  {
    const T v = make_random_bitvec<T, N>(ctx);
    T       r = v;

    return run_bench([&](auto) { chilbert::detail::gray_code_inv(r); });
  }
};

template<class T, size_t N>
struct BenchComparison {
  Duration operator()(Context& ctx)
  {
    std::vector<T> vecs;
    for (size_t i = 0; i < 32; ++i) {
      vecs.emplace_back(make_random_bitvec<T, N>(ctx));
    }

    volatile bool r = false;

    return run_bench([&](const auto i) {
      r |= vecs[i % vecs.size()] < vecs[(i + 1) % vecs.size()];
    });
  }
};

template<class T, size_t N>
struct BenchIteration {
  Duration operator()(Context& ctx)
  {
    T             v = make_random_bitvec<T, N>(ctx);
    auto          i = v.begin();
    volatile bool r = false;
    return run_bench([&](const auto) {
      r = (++i == v.end());
      if (r) {
        i = v.begin();
      }
    });
  }
};

/// Run benchmark for size N
template<template<class U, size_t M> class Bench, size_t N>
void
bench_row(Context& ctx, std::ostream& os)
{
  const std::array<Duration, 4> results = {
    ((N <= chilbert::SmallBitVec::bits_per_rack)
       ? Bench<chilbert::SmallBitVec, N>{}(ctx)
       : Duration{}),
    Bench<chilbert::StaticBitVec<N>, N>{}(ctx),
    Bench<chilbert::BoundedBitVec<N>, N>{}(ctx),
    Bench<chilbert::DynamicBitVec, N>{}(ctx),
  };

  write_row(os, N, results);
}

/// Terminate recursion
template<template<class U, size_t M> class Bench>
void
bench_rec(Context&, std::ostream&)
{}

/// Run benchmark for sizes N, Ns... (recursive helper)
template<template<class U, size_t M> class Bench, size_t N, size_t... Ns>
void
bench_rec(Context& ctx, std::ostream& os)
{
  bench_row<Bench, N>(ctx, os);
  bench_rec<Bench, Ns...>(ctx, os);
}

/// Run benchmark
template<template<class U, size_t M> class Bench>
void
bench(Context& ctx, const std::string& name)
{
  std::ofstream out("bitvec_" + name + ".txt");
  out << "n\tsmall\tstatic\tbounded\tdynamic\n";
  bench_rec<Bench, 16, 32, 64, 128, 256, 512>(ctx, out);
}

int
main()
{
  Context ctx;

  bench<BenchAnd>(ctx, "and");
  bench<BenchOr>(ctx, "or");
  bench<BenchXor>(ctx, "xor");
  bench<BenchNone>(ctx, "none");

  bench<BenchSetOne>(ctx, "set_one");
  bench<BenchSetAll>(ctx, "set_all");

  bench<BenchResetOne>(ctx, "reset_one");
  bench<BenchResetAll>(ctx, "reset_all");

  bench<BenchFlipOne>(ctx, "flip_one");
  bench<BenchFlipAll>(ctx, "flip_all");

  bench<BenchCount>(ctx, "count");
  bench<BenchFindFirst>(ctx, "find_first");

  bench<BenchLeftShift>(ctx, "left_shift");
  bench<BenchRightShift>(ctx, "right_shift");

  bench<BenchLeftRotate>(ctx, "left_rotate");
  bench<BenchRightRotate>(ctx, "right_rotate");

  bench<BenchGrayCode>(ctx, "gray_code");
  bench<BenchGrayCodeInv>(ctx, "gray_code_inv");

  bench<BenchComparison>(ctx, "comparison");
  bench<BenchIteration>(ctx, "iteration");

  return 0;
}