#pragma once
#include <cmath>
#include <ircolib/types.hpp>

namespace ircolib {
static inline auto round_ceil(float f) {
#ifdef SIMD_SUPPORT
    __m128 t = _mm_set_ss(f);
    t = _mm_round_ss(t, t, _MM_FROUND_TO_POS_INF);
    return _mm_cvtss_f32(t);
#else
    return ceilf(f);
#endif
}

static inline auto round_ceil(double f) {
#ifdef SIMD_SUPPORT
    __m128d t = _mm_set_sd(f);
    t = _mm_round_sd(t, t, _MM_FROUND_TO_POS_INF);
    return _mm_cvtsd_f64(t);
#else
    return ceil(f);
#endif
}

static inline auto round_nearest(float f) {
#ifdef SIMD_SUPPORT
    __m128 t = _mm_set_ss(f);
    t = _mm_round_ss(t, t, _MM_FROUND_TO_NEAREST_INT);
    return _mm_cvtss_f32(t);
#else
    return roundf(f);
#endif
}

static inline auto round_nearest(double f) {
#ifdef SIMD_SUPPORT
    __m128d t = _mm_set_sd(f);
    t = _mm_round_sd(t, t, _MM_FROUND_TO_NEAREST_INT);
    return _mm_cvtsd_f64(t);
#else
    return round(f);
#endif
}

static inline auto round_current(float f) {
#ifdef SIMD_SUPPORT
    auto t = _mm_set_ss(f);
    t = _mm_round_ss(t, t, _MM_FROUND_CUR_DIRECTION);
    return _mm_cvtss_f32(t);
#else
    return rint(f);
#endif
}

static inline auto round_current(double f) {
#ifdef SIMD_SUPPORT
    auto t = _mm_set_sd(f);
    t = _mm_round_sd(t, t, _MM_FROUND_CUR_DIRECTION);
    return _mm_cvtsd_f64(t);
#else
    return rint(f);
#endif
}


static inline auto round_floor(float f) {
#ifdef SIMD_SUPPORT
    __m128 t = _mm_set_ss(f);
    t = _mm_round_ss(t, t, _MM_FROUND_TO_NEG_INF);
    return _mm_cvtss_f32(t);
#else
    return floor(f);
#endif
}

static inline auto round_floor(double f) {
#ifdef SIMD_SUPPORT
    __m128d t = _mm_set_sd(f);
    t = _mm_round_sd(t, t, _MM_FROUND_TO_NEG_INF);
    return _mm_cvtsd_f64(t);
#else
    return floor(f);
#endif
}

static inline auto round_trunc(float f) {
#ifdef SIMD_SUPPORT
    __m128 t = _mm_set_ss(f);
    t = _mm_round_ss(t, t, _MM_FROUND_TO_ZERO);
    return _mm_cvtss_f32(t);
#else
    return trunc(f);
#endif
}

static inline auto round_trunc(double f) {
#ifdef SIMD_SUPPORT
    __m128d t = _mm_set_sd(f);
    t = _mm_round_sd(t, t, _MM_FROUND_TO_ZERO);
    return _mm_cvtsd_f64(t);
#else
    return trunc(f);
#endif
}
} // namespace ircolib