MNN/source/math/Vec.hpp

//
//  Vec.hpp
//  MNN
//
//  Created by MNN on 2019/04/01.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#ifndef Vec_hpp
#define Vec_hpp
#include "core/Macro.h"
#include "core/SimdHeader.h"
#include <array>
#include <algorithm>  // supply std::max and std::min
#include <math.h>
namespace MNN {
namespace Math {

template <typename T, int N>
struct Vec {
    using VecType = Vec<T, N>;
    std::array<T, N> value;
    VecType operator+(const VecType& lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            dst.value[i] = value[i] + lr.value[i];
        }
        return dst;
    }
    VecType operator-(const VecType& lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            dst.value[i] = value[i] - lr.value[i];
        }
        return dst;
    }
    Vec(std::array<T, N>&& v) {
        value = std::move(v);
    }
    VecType operator==(const VecType& lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            if (value[i] == lr.value[i]) {
                dst.value[i] = 1;
            } else {
                dst.value[i] = 0;
            }
        }
        return dst;
    }
    VecType operator<(const VecType& lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            if (value[i] < lr.value[i]) {
                dst.value[i] = 1;
            } else {
                dst.value[i] = 0;
            }
        }
        return dst;
    }
    VecType operator<=(const VecType& lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            if (value[i] <= lr.value[i]) {
                dst.value[i] = 1;
            } else {
                dst.value[i] = 0;
            }
        }
        return dst;
    }
    VecType operator>(const VecType& lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            if (value[i] > lr.value[i]) {
                dst.value[i] = 1;
            } else {
                dst.value[i] = 0;
            }
        }
        return dst;
    }
    VecType operator>=(const VecType& lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            if (value[i] >= lr.value[i]) {
                dst.value[i] = 1;
            } else {
                dst.value[i] = 0;
            }
        }
        return dst;
    }
    VecType operator*(const VecType& lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            dst.value[i] = value[i] * lr.value[i];
        }
        return dst;
    }
    VecType operator*(T lr) const {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            dst.value[i] = value[i] * lr;
        }
        return dst;
    }
    VecType operator+=(const VecType& lr) {
        for (int i = 0; i < N; ++i) {
            value[i] = value[i] + lr.value[i];
        }
        return *this;
    }
    VecType operator-=(const VecType& lr) {
        for (int i = 0; i < N; ++i) {
            value[i] = value[i] - lr.value[i];
        }
        return *this;
    }

    VecType& operator=(const VecType& lr) {
        for (int i = 0; i < N; ++i) {
            value[i] = lr.value[i];
        }
        return *this;
    }
    VecType operator-() {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            dst.value[i] = -value[i];
        }
        return dst;
    }
    Vec() {
    }
    Vec(const T v) {
        for (int i = 0; i < N; ++i) {
            value[i] = v;
        }
    }

    Vec(const VecType& lr) {
        for (int i = 0; i < N; ++i) {
            value[i] = lr.value[i];
        }
    }
    T operator[](size_t i) {
        return value[i];
    }
    template<typename U>
    static VecType load(const U* addr) {
        VecType v;
        for (int i = 0; i < N; ++i) {
            v.value[i] = static_cast<T>(addr[i]);
        }
        return v;
    }
    template<typename U>
    static VecType broadcast(const U* addr) {
        VecType v;
        v.value[0] = static_cast<T>(addr[0]);
        for (int i = 1; i < N; ++i) {
            v.value[i] = v.value[0];
        }
        return v;
    }
    template<typename U>
    static void save(U* addr, const VecType& v) {
        for (int i = 0; i < N; ++i) {
            addr[i] = static_cast<U>(v.value[i]);
        }
    }
    static VecType max(const VecType& v1, const VecType& v2) {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            dst.value[i] = std::max(v1.value[i], v2.value[i]);
        }
        return dst;
    }
    static VecType min(const VecType& v1, const VecType& v2) {
        VecType dst;
        for (int i = 0; i < N; ++i) {
            dst.value[i] = std::min(v1.value[i], v2.value[i]);
        }
        return dst;
    }
    static VecType fma(const VecType& v1, const VecType& v2, const VecType& v3) {
         return v1 + v2 * v3;
    }
    static VecType fms(const VecType& v1, const VecType& v2, const VecType& v3) {
        return v1 - v2 * v3;
    }
    static inline void transpose4(VecType& vec0, VecType& vec1, VecType& vec2, VecType& vec3) {
        VecType source[4] = {vec0, vec1, vec2, vec3};
        for (int i = 0; i < N; ++i) {
            vec0.value[i] = source[i % 4].value[i >> 2];
            vec1.value[i] = source[i % 4].value[(i + N)>> 2];
            vec2.value[i] = source[i % 4].value[(i + 2 * N)>> 2];
            vec3.value[i] = source[i % 4].value[(i + 3 * N)>> 2];
        }
    }
};

#ifdef MNN_USE_NEON
template<>
struct Vec<int32_t, 4> {
    using VecType = Vec<int32_t, 4>;
    int32x4_t value;
    Vec() {
    }
    Vec(const int32_t v) {
        value = vdupq_n_s32(v);
    }
    Vec(const float v) {
        value = vdupq_n_s32((int32_t)v);
    }
    Vec(const int32x4_t v) {
        value = v;
    }
    Vec(const VecType& lr) {
        value = lr.value;
    }
    Vec(const VecType&& lr) {
        value = std::move(lr.value);
    }
    float operator[](size_t i) {
        return value[i];
    }
    static VecType load(const float* addr) {
        VecType v = { (int32x4_t)(vld1q_f32(addr)) };
        return v;
    }
    static VecType broadcast(const float* addr) {
        VecType dst = { (int32x4_t)(vld1q_dup_f32(addr)) };
        return dst;
    }
    static VecType broadcast(const int32_t* addr) {
        VecType dst = { vld1q_dup_s32(addr) };
        return dst;
    }
    static VecType load(const int32_t* addr) {
        VecType v = { vld1q_s32(addr) };
        return v;
    }
    static void save(float* addr, const VecType& v) {
        vst1q_f32(addr, (float32x4_t)(v.value));
    }
    static void save(int32_t* addr, const VecType& v) {
        vst1q_s32(addr, v.value);
    }
    static VecType max(const VecType& v1, const VecType& v2) {
        VecType dst = { vmaxq_s32(v1.value, v2.value) };
        return dst;
    }
    static VecType min(const VecType& v1, const VecType& v2) {
        VecType dst = { vminq_s32(v1.value, v2.value) };
        return dst;
    }
    static VecType fma(const VecType& v1, const VecType& v2, const VecType& v3) {
        VecType dst = {vmlaq_s32(v1.value, v2.value, v3.value)};
        return dst;
    }
    static VecType fms(const VecType& v1, const VecType& v2, const VecType& v3) {
        VecType dst = {vmlsq_s32(v1.value, v2.value, v3.value)};
        return dst;
    }
    static inline void transpose4(VecType& vec0, VecType& vec1, VecType& vec2, VecType& vec3) {
#ifdef __aarch64__
        auto m0 = vtrn1q_s32(vec0.value, vec1.value);
        auto m1 = vtrn2q_s32(vec0.value, vec1.value);
        auto m2 = vtrn1q_s32(vec2.value, vec3.value);
        auto m3 = vtrn2q_s32(vec2.value, vec3.value);
        vec0.value = (int32x4_t)(vtrn1q_s64((int64x2_t)(m0), (int64x2_t)(m2)));
        vec1.value = (int32x4_t)(vtrn1q_s64((int64x2_t)(m1), (int64x2_t)(m3)));
        vec2.value = (int32x4_t)(vtrn2q_s64((int64x2_t)(m0), (int64x2_t)(m2)));
        vec3.value = (int32x4_t)(vtrn2q_s64((int64x2_t)(m1), (int64x2_t)(m3)));
#else

        auto m0m1 = vtrnq_s32(vec0.value, vec1.value);
        auto m2m3 = vtrnq_s32(vec2.value, vec3.value);
        vec0.value = m0m1.val[0];
        vec1.value = m0m1.val[1];
        vec2.value = m2m3.val[0];
        vec3.value = m2m3.val[1];
        vec0.value = (int32x4_t)(vsetq_lane_s64(vgetq_lane_s64((int64x2_t)m2m3.val[0], 0), (int64x2_t)vec0.value, 1));
        vec1.value = (int32x4_t)(vsetq_lane_s64(vgetq_lane_s64((int64x2_t)m2m3.val[1], 0), (int64x2_t)vec1.value, 1));
        vec2.value = (int32x4_t)(vsetq_lane_s64(vgetq_lane_s64((int64x2_t)m0m1.val[0], 1), (int64x2_t)vec2.value, 0));
        vec3.value = (int32x4_t)(vsetq_lane_s64(vgetq_lane_s64((int64x2_t)m0m1.val[1], 1), (int64x2_t)vec3.value, 0));
#endif
    }

    VecType operator+(const VecType& lr) const {
        VecType dst = { vaddq_s32(value, lr.value) };
        return dst;
    }
    VecType operator-(const VecType& lr) const {
        VecType dst = { vsubq_s32(value, lr.value) };
        return dst;
    }
    VecType operator+=(const VecType& lr) {
        value = vaddq_s32(value, lr.value);
        return *this;
    }
    VecType operator-=(const VecType& lr) {
        value = vsubq_s32(value, lr.value);
        return *this;
    }
    VecType operator*(int32_t lr) const {
        VecType dst = { vmulq_n_s32(value, lr) };
        return dst;
    }
    VecType operator*(float lr) const {
        VecType dst = { vmulq_n_s32(value, (int32_t)lr) };
        return dst;
    }
    VecType operator*(const VecType& lr) const {
        VecType dst = { vmulq_s32(value, lr.value) };
        return dst;
    }
    VecType& operator=(const VecType& lr) {
        value = lr.value;
        return *this;
    }
    VecType& operator=(const VecType&& lr) {
        value = std::move(lr.value);
        return *this;
    }
    VecType operator-() {
        VecType dst = { vnegq_s32(value) };
        return dst;
    }
    VecType operator<(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vcltq_s32(value, lr.value);
        VecType dst = { vbslq_s32(res, one, zero) };
        return dst;
    }
    VecType operator>(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vcgtq_s32(value, lr.value);
        VecType dst = { vbslq_s32(res, one, zero) };
        return dst;
    }
    VecType operator<=(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vcleq_s32(value, lr.value);
        VecType dst = { vbslq_s32(res, one, zero) };
        return dst;
    }
    VecType operator>=(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vcgeq_s32(value, lr.value);
        VecType dst = { vbslq_s32(res, one, zero) };
        return dst;
    }
    VecType operator==(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vceqq_s32(value, lr.value);
        VecType dst = { vbslq_s32(res, one, zero) };
        return dst;
    }
};

template<>
struct Vec<float, 4> {
    using VecType = Vec<float, 4>;
    using VecTypeInt32 = Vec<int32_t, 4>;
    float32x4_t value;
    Vec() {
    }
    Vec(const float v) {
        value = vdupq_n_f32(v);
    }
    Vec(const float32x4_t v) {
        value = v;
    }
    Vec(const VecType& lr) {
        value = lr.value;
    }
    Vec(const VecType&& lr) {
        value = std::move(lr.value);
    }
    float operator[](size_t i) {
        return value[i];
    }
    static VecType load(const float* addr) {
        VecType v = { vld1q_f32(addr) };
        return v;
    }
    static VecType broadcast(const float* addr) {
        VecType dst = { vld1q_dup_f32(addr) };
        return dst;
    }
    static VecType load(const int32_t* addr) {
        VecType v = { vcvtq_f32_s32(vld1q_s32(addr)) };
        return v;
    }
    static void save(float* addr, const VecType& v) {
        vst1q_f32(addr, v.value);
    }
    static void save(float* addr, const VecTypeInt32& v) {
        vst1q_f32(addr, (float32x4_t)(v.value));
    }
    static void save(int32_t* addr, const VecType& v) {
        vst1q_s32(addr, (int32x4_t)(v.value));
    }
    static VecType max(const VecType& v1, const VecType& v2) {
        VecType dst = { vmaxq_f32(v1.value, v2.value) };
        return dst;
    }
    static VecType min(const VecType& v1, const VecType& v2) {
        VecType dst = { vminq_f32(v1.value, v2.value) };
        return dst;
    }
    static VecType fma(const VecType& v1, const VecType& v2, const VecType& v3) {
        VecType dst = {vmlaq_f32(v1.value, v2.value, v3.value)};
        return dst;
    }
    static VecType fms(const VecType& v1, const VecType& v2, const VecType& v3) {
        VecType dst = {vmlsq_f32(v1.value, v2.value, v3.value)};
        return dst;
    }
    static inline void transpose4(VecType& vec0, VecType& vec1, VecType& vec2, VecType& vec3) {
#ifdef __aarch64__
        auto m0 = vtrn1q_s32((int32x4_t)(vec0.value), (int32x4_t)(vec1.value));
        auto m1 = vtrn2q_s32((int32x4_t)(vec0.value), (int32x4_t)(vec1.value));
        auto m2 = vtrn1q_s32((int32x4_t)(vec2.value), (int32x4_t)(vec3.value));
        auto m3 = vtrn2q_s32((int32x4_t)(vec2.value), (int32x4_t)(vec3.value));
        vec0.value = (float32x4_t)(vtrn1q_s64((int64x2_t)(m0), (int64x2_t)(m2)));
        vec1.value = (float32x4_t)(vtrn1q_s64((int64x2_t)(m1), (int64x2_t)(m3)));
        vec2.value = (float32x4_t)(vtrn2q_s64((int64x2_t)(m0), (int64x2_t)(m2)));
        vec3.value = (float32x4_t)(vtrn2q_s64((int64x2_t)(m1), (int64x2_t)(m3)));
#else

        auto m0m1 = vtrnq_s32((int32x4_t)(vec0.value), (int32x4_t)(vec1.value));
        auto m2m3 = vtrnq_s32((int32x4_t)(vec2.value), (int32x4_t)(vec3.value));
        vec0.value = (float32x4_t)(m0m1.val[0]);
        vec1.value = (float32x4_t)(m0m1.val[1]);
        vec2.value = (float32x4_t)(m2m3.val[0]);
        vec3.value = (float32x4_t)(m2m3.val[1]);
        vec0.value = (float32x4_t)(vsetq_lane_s64(vgetq_lane_s64((int64x2_t)(m2m3.val[0]), 0), (int64x2_t)(vec0.value), 1));
        vec1.value = (float32x4_t)(vsetq_lane_s64(vgetq_lane_s64((int64x2_t)(m2m3.val[1]), 0), (int64x2_t)(vec1.value), 1));
        vec2.value = (float32x4_t)(vsetq_lane_s64(vgetq_lane_s64((int64x2_t)(m0m1.val[0]), 1), (int64x2_t)(vec2.value), 0));
        vec3.value = (float32x4_t)(vsetq_lane_s64(vgetq_lane_s64((int64x2_t)(m0m1.val[1]), 1), (int64x2_t)(vec3.value), 0));
        /*
        generated arm32 assembly code is almost the same as:
            vtrn.32 d0, d2
            vtrn.32 d1, d3
            vtrn.32 d4, d6
            vtrn.32 d5, d7
            vswp d1, d4
            vswp d3, d6
        */

#endif
    }

    VecType operator+(const VecType& lr) const {
        VecType dst = { vaddq_f32(value, lr.value) };
        return dst;
    }
    VecType operator-(const VecType& lr) const {
        VecType dst = { vsubq_f32(value, lr.value) };
        return dst;
    }
    VecType operator+=(const VecType& lr) {
        value = vaddq_f32(value, lr.value);
        return *this;
    }
    VecType operator-=(const VecType& lr) {
        value = vsubq_f32(value, lr.value);
        return *this;
    }
    VecType operator*(float lr) const {
        VecType dst = { vmulq_n_f32(value, lr) };
        return dst;
    }
    VecType operator*(const VecType& lr) const {
        VecType dst = { vmulq_f32(value, lr.value) };
        return dst;
    }
    VecType& operator=(const VecType& lr) {
        value = lr.value;
        return *this;
    }
    VecType& operator=(const VecType&& lr) {
        value = std::move(lr.value);
        return *this;
    }
    VecType operator-() {
        VecType dst = { vnegq_f32(value) };
        return dst;
    }
    VecType operator<(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vcltq_f32(value, lr.value);
        VecType dst = { (float32x4_t)(vbslq_s32(res, one, zero)) };
        return dst;
    }
    VecType operator>(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vcgtq_f32(value, lr.value);
        VecType dst = { (float32x4_t)(vbslq_s32(res, one, zero)) };
        return dst;
    }
    VecType operator<=(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vcleq_f32(value, lr.value);
        VecType dst = { (float32x4_t)(vbslq_s32(res, one, zero)) };
        return dst;
    }
    VecType operator>=(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vcgeq_f32(value, lr.value);
        VecType dst = { (float32x4_t)(vbslq_s32(res, one, zero)) };
        return dst;
    }
    VecType operator==(const VecType& lr) const {
        int32x4_t one = vdupq_n_s32(1);
        int32x4_t zero = vdupq_n_s32(0);
        uint32x4_t res = vceqq_f32(value, lr.value);
        VecType dst = { (float32x4_t)(vbslq_s32(res, one, zero)) };
        return dst;
    }
};

#elif defined(MNN_USE_SSE)
template<>
struct Vec<int32_t, 4> {
    using VecType = Vec<int32_t, 4>;
    using VecTypeArray = std::array<VecType, 4>;
    __m128i value;
    VecType operator+(const VecType& lr) const {
        VecType dst = { _mm_add_epi32(value, lr.value) };
        return dst;
    }
    VecType operator-(const VecType& lr) const {
        VecType dst = { _mm_sub_epi32(value, lr.value) };
        return dst;
    }
    VecType operator+=(const VecType& lr) {
        value = _mm_add_epi32(value, lr.value);
        return *this;
    }
    VecType operator-=(const VecType& lr) {
        value = _mm_sub_epi32(value, lr.value);
        return *this;
    }
    VecType operator*(const VecType& lr) const {
        VecType dst = {_mm_cvtps_epi32(_mm_mul_ps(_mm_cvtepi32_ps(value), _mm_cvtepi32_ps(lr.value)))};
        return dst;
    }

    VecType& operator=(const VecType& lr) {
        value = lr.value;
        return *this;
    }
    VecType operator==(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128i mask = _mm_cmpeq_epi32(value, lr.value);
        VecType dst = { _mm_and_si128(one, mask) };
        return dst;
    }
    VecType operator<(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128i mask = _mm_cmplt_epi32(value, lr.value);
        VecType dst = { _mm_and_si128(one, mask) };
        return dst;
    }
    VecType operator<=(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128i mask = _mm_cmpgt_epi32(value, lr.value);
        VecType dst = { _mm_andnot_si128(mask, one) };
        return dst;
    }
    VecType operator>(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128i mask = _mm_cmpgt_epi32(value, lr.value);
        VecType dst = { _mm_and_si128(one, mask) };
        return dst;
    }
    VecType operator>=(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128i mask = _mm_cmplt_epi32(value, lr.value);
        VecType dst = { _mm_andnot_si128(mask, one) };
        return dst;
    }
    VecType operator-() {
        VecType dst;
#if defined(_MSC_VER)
        dst.value = _mm_cvtps_epi32(_mm_xor_ps(_mm_cvtepi32_ps(value), _mm_set1_ps(-0.f))); // Using unary operation to SSE vec is GCC extension. We can not do this directly in MSVC.
#else
        dst.value = -value;
#endif
        return dst;
    }
    Vec() {
    }
    Vec(const float v) {
        int u = static_cast<int32_t>(v);
        value = _mm_set_epi32(u, u, u, u);
    }
    Vec(const int32_t v) {
        value = _mm_set_epi32(v, v, v, v);
    }
    Vec(__m128i&& v) {
        value = v;
    }
    Vec(__m128&& v) {
        value = _mm_castps_si128(v);
    }
    Vec(const VecType& lr) {
        value = lr.value;
    }
    float operator[](size_t i) {
#if defined(_MSC_VER)  // X64 native only mandatory support SSE and SSE2 extension, and we can not find intrinsic function to extract element directly by index in SSE and SSE2 extension.
        int32_t temp[4];
        _mm_storeu_si128((__m128i*)temp, value);
        return temp[i];
#else
        return value[i];
#endif
    }
    static VecType load(const int32_t* addr) {
        VecType v = {_mm_loadu_si128((__m128i const*)(addr))};
        return v;
    }
    static VecType broadcast(const int32_t* addr) {
        int32_t arr[4] = {*addr, 0, 0, 0};
        VecType dst = { _mm_loadu_si128((__m128i const*)(arr)) };
        return dst;
    }
    static void save(int32_t* addr, const VecType& v) {
        _mm_storeu_si128((__m128i*)addr, v.value);
    }
    static VecType max(const VecType& v1, const VecType& v2) {
        VecType dst = {_mm_cvtps_epi32(_mm_max_ps(_mm_cvtepi32_ps(v1.value), _mm_cvtepi32_ps(v2.value)))};
        return dst;
    }
    static VecType min(const VecType& v1, const VecType& v2) {
        VecType dst = {_mm_cvtps_epi32(_mm_min_ps(_mm_cvtepi32_ps(v1.value), _mm_cvtepi32_ps(v2.value)))};
        return dst;
    }
    static VecType fma(const VecType& v1, const VecType& v2, const VecType& v3) {
        return v1 + v2 * v3; // TODO: use fma instruction
    }
    static VecType fms(const VecType& v1, const VecType& v2, const VecType& v3) {
        return v1 - v2 * v3; // TODO: use fma instruction
    }
    static inline void transpose4(VecType& vec0, VecType& vec1, VecType& vec2, VecType& vec3) {
        __m128 tmp3, tmp2, tmp1, tmp0;
        tmp0   = _mm_unpacklo_ps(_mm_castsi128_ps(vec0.value), _mm_castsi128_ps(vec1.value));
        tmp2   = _mm_unpacklo_ps(_mm_castsi128_ps(vec2.value), _mm_castsi128_ps(vec3.value));
        tmp1   = _mm_unpackhi_ps(_mm_castsi128_ps(vec0.value), _mm_castsi128_ps(vec1.value));
        tmp3   = _mm_unpackhi_ps(_mm_castsi128_ps(vec2.value), _mm_castsi128_ps(vec3.value));
        vec0.value = _mm_castps_si128(_mm_movelh_ps(tmp0, tmp2));
        vec1.value = _mm_castps_si128(_mm_movehl_ps(tmp2, tmp0));
        vec2.value = _mm_castps_si128(_mm_movelh_ps(tmp1, tmp3));
        vec3.value = _mm_castps_si128(_mm_movehl_ps(tmp3, tmp1));
    }
};

template<>
struct Vec<float, 4> {
    using VecType = Vec<float, 4>;
    using VecTypeInt32 = Vec<int32_t, 4>;
    using VecTypeArray = std::array<VecType, 4>;
    __m128 value;
    VecType operator+(const VecType& lr) const {
        VecType dst = { _mm_add_ps(value, lr.value) };
        return dst;
    }
    VecType operator-(const VecType& lr) const {
        VecType dst = { _mm_sub_ps(value, lr.value) };
        return dst;
    }
    VecType operator+=(const VecType& lr) {
        value = _mm_add_ps(value, lr.value);
        return *this;
    }
    VecType operator-=(const VecType& lr) {
        value = _mm_sub_ps(value, lr.value);
        return *this;
    }
    VecType operator*(const VecType& lr) const {
        VecType dst = { _mm_mul_ps(value, lr.value) };
        return dst;
    }
    VecType operator*(float lr) const {
        VecType dst = { _mm_mul_ps(value, _mm_set1_ps(lr)) };
        return dst;
    }

    VecType& operator=(const VecType& lr) {
        value = lr.value;
        return *this;
    }
    VecType operator-() {
        VecType dst;
#if defined(_MSC_VER)
        dst.value = _mm_xor_ps(value, _mm_set1_ps(-0.f)); // Using unary operation to SSE vec is GCC extension. We can not do this directly in MSVC.
#else
        dst.value = -value;
#endif
        return dst;
    }
    VecType operator==(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128i mask = _mm_cmpeq_epi32(_mm_castps_si128(value), _mm_castps_si128(lr.value));
        VecType dst = { _mm_castsi128_ps(_mm_and_si128(one, mask)) };
        return dst;
    }
    VecType operator<(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128i mask = _mm_cmplt_epi32(_mm_castps_si128(value), _mm_castps_si128(lr.value));
        VecType dst = { _mm_castsi128_ps(_mm_and_si128(one, mask)) };
        return dst;
    }
    VecType operator<=(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128 mask = _mm_cmple_ps(value, lr.value);
        VecType dst = { _mm_castsi128_ps(_mm_and_si128(one, _mm_castps_si128(mask))) };
        return dst;
    }
    VecType operator>(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128 mask = _mm_cmpgt_ps(value, lr.value);
        VecType dst = { _mm_castsi128_ps(_mm_and_si128(one, _mm_castps_si128(mask))) };
        return dst;
    }
    VecType operator>=(const VecType& lr) const {
        __m128i one = _mm_set1_epi32(1);
        __m128 mask = _mm_cmpge_ps(value, lr.value);
        VecType dst = { _mm_castsi128_ps(_mm_and_si128(one, _mm_castps_si128(mask))) };
        return dst;
    }
    Vec() {
    }
    Vec(const float v) {
        value = _mm_set1_ps(v);
    }
    Vec(__m128&& v) {
        value = v;
    }
    Vec(const VecType& lr) {
        value = lr.value;
    }
    float operator[](size_t i) {
#if defined(_MSC_VER)  // X64 native only mandatory support SSE and SSE2 extension, and we can not find intrinsic function to extract element directly by index in SSE and SSE2 extension.
        float temp[4];
        _mm_storeu_ps(temp, value);
        return temp[i];
#else
        return value[i];
#endif
    }
    static VecType load(const float* addr) {
        VecType v = { _mm_loadu_ps(addr) };
        return v;
    }
    static VecType broadcast(const float* addr) {
        VecType dst = { _mm_load_ss(addr) };
        return dst;
    }
    static void save(float* addr, const VecType& v) {
        _mm_storeu_ps(addr, v.value);
    }
    static void save(float* addr, const VecTypeInt32& v) {
        _mm_storeu_ps(addr, _mm_castsi128_ps(v.value));
    }
    static void save(int32_t* addr, const VecType& v) {
        _mm_storeu_si128((__m128i*)addr, _mm_castps_si128(v.value));
    }
    static VecType max(const VecType& v1, const VecType& v2) {
        VecType dst = { _mm_max_ps(v1.value, v2.value) };
        return dst;
    }
    static VecType min(const VecType& v1, const VecType& v2) {
        VecType dst = { _mm_min_ps(v1.value, v2.value) };
        return dst;
    }
    static VecType fma(const VecType& v1, const VecType& v2, const VecType& v3) {
        return v1 + v2 * v3; // TODO: use fma instruction
    }
    static VecType fms(const VecType& v1, const VecType& v2, const VecType& v3) {
        return v1 - v2 * v3; // TODO: use fma instruction
    }
    static inline void transpose4(VecType& vec0, VecType& vec1, VecType& vec2, VecType& vec3) {
        __m128 tmp3, tmp2, tmp1, tmp0;
        tmp0   = _mm_unpacklo_ps((vec0.value), (vec1.value));
        tmp2   = _mm_unpacklo_ps((vec2.value), (vec3.value));
        tmp1   = _mm_unpackhi_ps((vec0.value), (vec1.value));
        tmp3   = _mm_unpackhi_ps((vec2.value), (vec3.value));
        vec0.value = _mm_movelh_ps(tmp0, tmp2);
        vec1.value = _mm_movehl_ps(tmp2, tmp0);
        vec2.value = _mm_movelh_ps(tmp1, tmp3);
        vec3.value = _mm_movehl_ps(tmp3, tmp1);
    }
};

#endif
} // namespace Math
} // namespace MNN

#endif /* Vec_hpp */