#pragma once

// DO NOT DEFINE STATIC DATA IN THIS HEADER!
// See Note [Do not compile initializers with AVX]

#include <ATen/cpu/vec256/intrinsics.h>
#include <ATen/cpu/vec256/vec256_base.h>
#include <c10/macros/Macros.h>

namespace at {
namespace vec256 {
namespace {

#ifdef CPU_CAPABILITY_AVX2

struct Vec256i {
protected:
  __m256i values;

  static inline __m256i invert(const __m256i& v) {
    const auto ones = _mm256_set1_epi64x(-1);
    return _mm256_xor_si256(ones, v);
  }
public:
  Vec256i() {}
  Vec256i(__m256i v) : values(v) {}
  operator __m256i() const {
    return values;
  }
};

#else

struct Vec256i {};  // dummy definition to make Vec256i always defined

#endif // CPU_CAPABILITY_AVX2

#ifdef CPU_CAPABILITY_AVX2

template <>
class Vec256<int64_t> : public Vec256i {
private:
  static const Vec256<int64_t> ones;
public:
  using value_type = int64_t;
  static constexpr int size() {
    return 4;
  }
  using Vec256i::Vec256i;
  Vec256() {}
  Vec256(int64_t v) { values = _mm256_set1_epi64x(v); }
  Vec256(int64_t val1, int64_t val2, int64_t val3, int64_t val4) {
    values = _mm256_setr_epi64x(val1, val2, val3, val4);
  }
  template <int64_t mask>
  static Vec256<int64_t> blend(Vec256<int64_t> a, Vec256<int64_t> b) {
    __at_align32__ int64_t tmp_values[size()];
    a.store(tmp_values);
    if (mask & 0x01)
      tmp_values[0] = _mm256_extract_epi64(b.values, 0);
    if (mask & 0x02)
      tmp_values[1] = _mm256_extract_epi64(b.values, 1);
    if (mask & 0x04)
      tmp_values[2] = _mm256_extract_epi64(b.values, 2);
    if (mask & 0x08)
      tmp_values[3] = _mm256_extract_epi64(b.values, 3);
    return loadu(tmp_values);
  }
  static Vec256<int64_t> blendv(const Vec256<int64_t>& a, const Vec256<int64_t>& b,
                                const Vec256<int64_t>& mask) {
    return _mm256_blendv_epi8(a.values, b.values, mask.values);
  }
  template <typename step_t>
  static Vec256<int64_t> arange(int64_t base = 0, step_t step = static_cast<step_t>(1)) {
    return Vec256<int64_t>(base, base + step, base + 2 * step, base + 3 * step);
  }
  static Vec256<int64_t>
  set(Vec256<int64_t> a, Vec256<int64_t> b, int64_t count = size()) {
    switch (count) {
      case 0:
        return a;
      case 1:
        return blend<1>(a, b);
      case 2:
        return blend<3>(a, b);
      case 3:
        return blend<7>(a, b);
    }
    return b;
  }
  static Vec256<int64_t> loadu(const void* ptr) {
    return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
  }
  static Vec256<int64_t> loadu(const void* ptr, int64_t count) {
    __at_align32__ int64_t tmp_values[size()];
    // Ensure uninitialized memory does not change the output value See https://github.com/pytorch/pytorch/issues/32502
    // for more details. We do not initialize arrays to zero using "={0}" because gcc would compile it to two
    // instructions while a loop would be compiled to one instruction.
    for (auto i = 0; i < size(); ++i) {
      tmp_values[i] = 0;
    }
    std::memcpy(tmp_values, ptr, count * sizeof(int64_t));
    return loadu(tmp_values);
  }
  void store(void* ptr, int count = size()) const {
    if (count == size()) {
      // ptr need not to be aligned here. See
      // https://software.intel.com/content/www/us/en/develop/documentation/cpp-compiler-developer-guide-and-reference/top/compiler-reference/intrinsics/intrinsics-for-intel-advanced-vector-extensions/intrinsics-for-load-and-store-operations-1/mm256-storeu-si256.html
      _mm256_storeu_si256(reinterpret_cast<__m256i*>(ptr), values);
    } else if (count > 0) {
      __at_align32__ int64_t tmp_values[size()];
      _mm256_storeu_si256(reinterpret_cast<__m256i*>(tmp_values), values);
      std::memcpy(ptr, tmp_values, count * sizeof(int64_t));
    }
  }
  const int64_t& operator[](int idx) const  = delete;
  int64_t& operator[](int idx)  = delete;
  Vec256<int64_t> abs() const {
    auto zero = _mm256_set1_epi64x(0);
    auto is_larger = _mm256_cmpgt_epi64(zero, values);
    auto inverse = _mm256_xor_si256(values, is_larger);
    return _mm256_sub_epi64(inverse, is_larger);
  }
  Vec256<int64_t> real() const {
    return *this;
  }
  Vec256<int64_t> imag() const {
    return _mm256_set1_epi64x(0);
  }
  Vec256<int64_t> conj() const {
    return *this;
  }
  Vec256<int64_t> frac() const;
  Vec256<int64_t> neg() const;
  Vec256<int64_t> operator==(const Vec256<int64_t>& other) const {
    return _mm256_cmpeq_epi64(values, other.values);
  }
  Vec256<int64_t> operator!=(const Vec256<int64_t>& other) const {
    return invert(_mm256_cmpeq_epi64(values, other.values));
  }
  Vec256<int64_t> operator<(const Vec256<int64_t>& other) const {
    return _mm256_cmpgt_epi64(other.values, values);
  }
  Vec256<int64_t> operator<=(const Vec256<int64_t>& other) const {
    return invert(_mm256_cmpgt_epi64(values, other.values));
  }
  Vec256<int64_t> operator>(const Vec256<int64_t>& other) const {
    return _mm256_cmpgt_epi64(values, other.values);
  }
  Vec256<int64_t> operator>=(const Vec256<int64_t>& other) const {
    return invert(_mm256_cmpgt_epi64(other.values, values));
  }

  Vec256<int64_t> eq(const Vec256<int64_t>& other) const;
  Vec256<int64_t> ne(const Vec256<int64_t>& other) const;
  Vec256<int64_t> gt(const Vec256<int64_t>& other) const;
  Vec256<int64_t> ge(const Vec256<int64_t>& other) const;
  Vec256<int64_t> lt(const Vec256<int64_t>& other) const;
  Vec256<int64_t> le(const Vec256<int64_t>& other) const;
};

template <>
class Vec256<int32_t> : public Vec256i {
private:
  static const Vec256<int32_t> ones;
public:
  using value_type = int32_t;
  static constexpr int size() {
    return 8;
  }
  using Vec256i::Vec256i;
  Vec256() {}
  Vec256(int32_t v) { values = _mm256_set1_epi32(v); }
  Vec256(int32_t val1, int32_t val2, int32_t val3, int32_t val4,
         int32_t val5, int32_t val6, int32_t val7, int32_t val8) {
    values = _mm256_setr_epi32(val1, val2, val3, val4, val5, val6, val7, val8);
  }
  template <int64_t mask>
  static Vec256<int32_t> blend(Vec256<int32_t> a, Vec256<int32_t> b) {
    return _mm256_blend_epi32(a, b, mask);
  }
  static Vec256<int32_t> blendv(const Vec256<int32_t>& a, const Vec256<int32_t>& b,
                                const Vec256<int32_t>& mask) {
    return _mm256_blendv_epi8(a.values, b.values, mask.values);
  }
  template <typename step_t>
  static Vec256<int32_t> arange(int32_t base = 0, step_t step = static_cast<step_t>(1)) {
    return Vec256<int32_t>(
      base,            base +     step, base + 2 * step, base + 3 * step,
      base + 4 * step, base + 5 * step, base + 6 * step, base + 7 * step);
  }
  static Vec256<int32_t>
  set(Vec256<int32_t> a, Vec256<int32_t> b, int32_t count = size()) {
    switch (count) {
      case 0:
        return a;
      case 1:
        return blend<1>(a, b);
      case 2:
        return blend<3>(a, b);
      case 3:
        return blend<7>(a, b);
      case 4:
        return blend<15>(a, b);
      case 5:
        return blend<31>(a, b);
      case 6:
        return blend<63>(a, b);
      case 7:
        return blend<127>(a, b);
    }
    return b;
  }
  static Vec256<int32_t> loadu(const void* ptr) {
    return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
  }
  static Vec256<int32_t> loadu(const void* ptr, int32_t count) {
    __at_align32__ int32_t tmp_values[size()];
    // Ensure uninitialized memory does not change the output value See https://github.com/pytorch/pytorch/issues/32502
    // for more details. We do not initialize arrays to zero using "={0}" because gcc would compile it to two
    // instructions while a loop would be compiled to one instruction.
    for (auto i = 0; i < size(); ++i) {
      tmp_values[i] = 0;
    }
    std::memcpy(tmp_values, ptr, count * sizeof(int32_t));
    return loadu(tmp_values);
  }
  void store(void* ptr, int count = size()) const {
    if (count == size()) {
      // ptr need not to be aligned here. See
      // https://software.intel.com/content/www/us/en/develop/documentation/cpp-compiler-developer-guide-and-reference/top/compiler-reference/intrinsics/intrinsics-for-intel-advanced-vector-extensions/intrinsics-for-load-and-store-operations-1/mm256-storeu-si256.html
      _mm256_storeu_si256(reinterpret_cast<__m256i*>(ptr), values);
    } else if (count > 0) {
      __at_align32__ int32_t tmp_values[size()];
      _mm256_storeu_si256(reinterpret_cast<__m256i*>(tmp_values), values);
      std::memcpy(ptr, tmp_values, count * sizeof(int32_t));
    }
  }
  void dump() const {
      for (size_t i = 0; i < size(); ++i) {
          std::cout << (int)((value_type*)&values)[i] << " ";
      }
      std::cout << std::endl;
  }
  const int32_t& operator[](int idx) const  = delete;
  int32_t& operator[](int idx)  = delete;
  Vec256<int32_t> abs() const {
    return _mm256_abs_epi32(values);
  }
  Vec256<int32_t> real() const {
    return *this;
  }
  Vec256<int32_t> imag() const {
    return _mm256_set1_epi32(0);
  }
  Vec256<int32_t> conj() const {
    return *this;
  }
  Vec256<int32_t> frac() const;
  Vec256<int32_t> neg() const;
  Vec256<int32_t> operator==(const Vec256<int32_t>& other) const {
    return _mm256_cmpeq_epi32(values, other.values);
  }
  Vec256<int32_t> operator!=(const Vec256<int32_t>& other) const {
    return invert(_mm256_cmpeq_epi32(values, other.values));
  }
  Vec256<int32_t> operator<(const Vec256<int32_t>& other) const {
    return _mm256_cmpgt_epi32(other.values, values);
  }
  Vec256<int32_t> operator<=(const Vec256<int32_t>& other) const {
    return invert(_mm256_cmpgt_epi32(values, other.values));
  }
  Vec256<int32_t> operator>(const Vec256<int32_t>& other) const {
    return _mm256_cmpgt_epi32(values, other.values);
  }
  Vec256<int32_t> operator>=(const Vec256<int32_t>& other) const {
    return invert(_mm256_cmpgt_epi32(other.values, values));
  }
  Vec256<int32_t> eq(const Vec256<int32_t>& other) const;
  Vec256<int32_t> ne(const Vec256<int32_t>& other) const;
  Vec256<int32_t> gt(const Vec256<int32_t>& other) const;
  Vec256<int32_t> ge(const Vec256<int32_t>& other) const;
  Vec256<int32_t> lt(const Vec256<int32_t>& other) const;
  Vec256<int32_t> le(const Vec256<int32_t>& other) const;
};

template <>
inline void convert(const int32_t *src, float *dst, int64_t n) {
  int64_t i;
  // int32_t and float have same size
#ifndef _MSC_VER
# pragma unroll
#endif
  for (i = 0; i <= (n - Vec256<int32_t>::size()); i += Vec256<int32_t>::size()) {
    auto input_vec = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(src + i));
    auto output_vec = _mm256_cvtepi32_ps(input_vec);
    _mm256_storeu_ps(reinterpret_cast<float*>(dst + i), output_vec);
  }
#ifndef _MSC_VER
# pragma unroll
#endif
  for (; i < n; i++) {
    dst[i] = static_cast<float>(src[i]);
  }
}

template <>
inline void convert(const int32_t *src, double *dst, int64_t n) {
  int64_t i;
  // int32_t has half the size of double
#ifndef _MSC_VER
# pragma unroll
#endif
  for (i = 0; i <= (n - Vec256<double>::size()); i += Vec256<double>::size()) {
    auto input_128_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src + i));
    auto output_vec = _mm256_cvtepi32_pd(input_128_vec);
    _mm256_storeu_pd(reinterpret_cast<double*>(dst + i), output_vec);
  }
#ifndef _MSC_VER
# pragma unroll
#endif
  for (; i < n; i++) {
    dst[i] = static_cast<double>(src[i]);
  }
}

template <>
class Vec256<int16_t> : public Vec256i {
private:
  static const Vec256<int16_t> ones;
public:
  using value_type = int16_t;
  static constexpr int size() {
    return 16;
  }
  using Vec256i::Vec256i;
  Vec256() {}
  Vec256(int16_t v) { values = _mm256_set1_epi16(v); }
  Vec256(int16_t val1, int16_t val2, int16_t val3, int16_t val4,
         int16_t val5, int16_t val6, int16_t val7, int16_t val8,
         int16_t val9, int16_t val10, int16_t val11, int16_t val12,
         int16_t val13, int16_t val14, int16_t val15, int16_t val16) {
    values = _mm256_setr_epi16(val1, val2, val3, val4, val5, val6, val7, val8,
                               val9, val10, val11, val12, val13, val14, val15, val16);
  }