src/opts/SkNx_neon.h

/*
 * Copyright 2015 Google Inc.
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#ifndef SkNx_neon_DEFINED
#define SkNx_neon_DEFINED

#define SKNX_IS_FAST

// ARMv8 has vrndmq_f32 to floor 4 floats.  Here we emulate it:
//   - round by adding (1<<23) with our sign, then subtracting it;
//   - if that rounded value is bigger than our input, subtract 1.
static inline float32x4_t armv7_vrndmq_f32(float32x4_t v) {
    auto sign = vandq_u32((uint32x4_t)v, vdupq_n_u32(1<<31));
    auto bias = (float32x4_t)(vorrq_u32((uint32x4_t)vdupq_n_f32(1<<23), sign));
    auto rounded = vsubq_f32(vaddq_f32(v, bias), bias);
    auto too_big = vcgtq_f32(rounded, v);
    return vsubq_f32(rounded, (float32x4_t)vandq_u32(too_big, (uint32x4_t)vdupq_n_f32(1)));
}

// Well, this is absurd.  The shifts require compile-time constant arguments.

#define SHIFT8(op, v, bits) switch(bits) { \
    case  1: return op(v,  1);  case  2: return op(v,  2);  case  3: return op(v,  3); \
    case  4: return op(v,  4);  case  5: return op(v,  5);  case  6: return op(v,  6); \
    case  7: return op(v,  7); \
    } return fVec

#define SHIFT16(op, v, bits) if (bits < 8) { SHIFT8(op, v, bits); } switch(bits) { \
                                case  8: return op(v,  8);  case  9: return op(v,  9); \
    case 10: return op(v, 10);  case 11: return op(v, 11);  case 12: return op(v, 12); \
    case 13: return op(v, 13);  case 14: return op(v, 14);  case 15: return op(v, 15); \
    } return fVec

#define SHIFT32(op, v, bits) if (bits < 16) { SHIFT16(op, v, bits); } switch(bits) { \
    case 16: return op(v, 16);  case 17: return op(v, 17);  case 18: return op(v, 18); \
    case 19: return op(v, 19);  case 20: return op(v, 20);  case 21: return op(v, 21); \
    case 22: return op(v, 22);  case 23: return op(v, 23);  case 24: return op(v, 24); \
    case 25: return op(v, 25);  case 26: return op(v, 26);  case 27: return op(v, 27); \
    case 28: return op(v, 28);  case 29: return op(v, 29);  case 30: return op(v, 30); \
    case 31: return op(v, 31); } return fVec

template <>
class SkNx<2, float> {
public:
    SkNx(float32x2_t vec) : fVec(vec) {}

    SkNx() {}
    SkNx(float val)           : fVec(vdup_n_f32(val)) {}
    static SkNx Load(const void* ptr) { return vld1_f32((const float*)ptr); }
    SkNx(float a, float b) { fVec = (float32x2_t) { a, b }; }

    void store(void* ptr) const { vst1_f32((float*)ptr, fVec); }

    SkNx approxInvert() const {
        float32x2_t est0 = vrecpe_f32(fVec),
                    est1 = vmul_f32(vrecps_f32(est0, fVec), est0);
        return est1;
    }
    SkNx invert() const {
        float32x2_t est1 = this->approxInvert().fVec,
                    est2 = vmul_f32(vrecps_f32(est1, fVec), est1);
        return est2;
    }

    SkNx operator + (const SkNx& o) const { return vadd_f32(fVec, o.fVec); }
    SkNx operator - (const SkNx& o) const { return vsub_f32(fVec, o.fVec); }
    SkNx operator * (const SkNx& o) const { return vmul_f32(fVec, o.fVec); }
    SkNx operator / (const SkNx& o) const {
    #if defined(SK_CPU_ARM64)
        return vdiv_f32(fVec, o.fVec);
    #else
        return vmul_f32(fVec, o.invert().fVec);
    #endif
    }

    SkNx operator == (const SkNx& o) const { return vreinterpret_f32_u32(vceq_f32(fVec, o.fVec)); }
    SkNx operator  < (const SkNx& o) const { return vreinterpret_f32_u32(vclt_f32(fVec, o.fVec)); }
    SkNx operator  > (const SkNx& o) const { return vreinterpret_f32_u32(vcgt_f32(fVec, o.fVec)); }
    SkNx operator <= (const SkNx& o) const { return vreinterpret_f32_u32(vcle_f32(fVec, o.fVec)); }
    SkNx operator >= (const SkNx& o) const { return vreinterpret_f32_u32(vcge_f32(fVec, o.fVec)); }
    SkNx operator != (const SkNx& o) const {
        return vreinterpret_f32_u32(vmvn_u32(vceq_f32(fVec, o.fVec)));
    }

    static SkNx Min(const SkNx& l, const SkNx& r) { return vmin_f32(l.fVec, r.fVec); }
    static SkNx Max(const SkNx& l, const SkNx& r) { return vmax_f32(l.fVec, r.fVec); }

    SkNx rsqrt0() const { return vrsqrte_f32(fVec); }
    SkNx rsqrt1() const {
        float32x2_t est0 = this->rsqrt0().fVec;
        return vmul_f32(vrsqrts_f32(fVec, vmul_f32(est0, est0)), est0);
    }
    SkNx rsqrt2() const {
        float32x2_t est1 = this->rsqrt1().fVec;
        return vmul_f32(vrsqrts_f32(fVec, vmul_f32(est1, est1)), est1);
    }

    SkNx sqrt() const {
    #if defined(SK_CPU_ARM64)
        return vsqrt_f32(fVec);
    #else
        return *this * this->rsqrt2();
    #endif
    }

    float operator[](int k) const {
        SkASSERT(0 <= k && k < 2);
        union { float32x2_t v; float fs[2]; } pun = {fVec};
        return pun.fs[k&1];
    }
    template <int k> float kth() const { return (*this)[k]; }

    bool allTrue() const {
        auto v = vreinterpret_u32_f32(fVec);
        return vget_lane_u32(v,0) && vget_lane_u32(v,1);
    }
    bool anyTrue() const {
        auto v = vreinterpret_u32_f32(fVec);
        return vget_lane_u32(v,0) || vget_lane_u32(v,1);
    }

    float32x2_t fVec;
};

template <>
class SkNx<4, float> {
public:
    SkNx(float32x4_t vec) : fVec(vec) {}

    SkNx() {}
    SkNx(float val)           : fVec(vdupq_n_f32(val)) {}
    static SkNx Load(const void* ptr) { return vld1q_f32((const float*)ptr); }
    SkNx(float a, float b, float c, float d) { fVec = (float32x4_t) { a, b, c, d }; }

    void store(void* ptr) const { vst1q_f32((float*)ptr, fVec); }
    SkNx approxInvert() const {
        float32x4_t est0 = vrecpeq_f32(fVec),
                    est1 = vmulq_f32(vrecpsq_f32(est0, fVec), est0);
        return est1;
    }
    SkNx invert() const {
        float32x4_t est1 = this->approxInvert().fVec,
                    est2 = vmulq_f32(vrecpsq_f32(est1, fVec), est1);
        return est2;
    }

    SkNx operator + (const SkNx& o) const { return vaddq_f32(fVec, o.fVec); }
    SkNx operator - (const SkNx& o) const { return vsubq_f32(fVec, o.fVec); }
    SkNx operator * (const SkNx& o) const { return vmulq_f32(fVec, o.fVec); }
    SkNx operator / (const SkNx& o) const {
    #if defined(SK_CPU_ARM64)
        return vdivq_f32(fVec, o.fVec);
    #else
        return vmulq_f32(fVec, o.invert().fVec);
    #endif
    }

    SkNx operator==(const SkNx& o) const { return vreinterpretq_f32_u32(vceqq_f32(fVec, o.fVec)); }
    SkNx operator <(const SkNx& o) const { return vreinterpretq_f32_u32(vcltq_f32(fVec, o.fVec)); }
    SkNx operator >(const SkNx& o) const { return vreinterpretq_f32_u32(vcgtq_f32(fVec, o.fVec)); }
    SkNx operator<=(const SkNx& o) const { return vreinterpretq_f32_u32(vcleq_f32(fVec, o.fVec)); }
    SkNx operator>=(const SkNx& o) const { return vreinterpretq_f32_u32(vcgeq_f32(fVec, o.fVec)); }
    SkNx operator!=(const SkNx& o) const {
        return vreinterpretq_f32_u32(vmvnq_u32(vceqq_f32(fVec, o.fVec)));
    }

    static SkNx Min(const SkNx& l, const SkNx& r) { return vminq_f32(l.fVec, r.fVec); }
    static SkNx Max(const SkNx& l, const SkNx& r) { return vmaxq_f32(l.fVec, r.fVec); }

    SkNx abs() const { return vabsq_f32(fVec); }
    SkNx floor() const {
    #if defined(SK_CPU_ARM64)
        return vrndmq_f32(fVec);
    #else
        return armv7_vrndmq_f32(fVec);
    #endif
    }


    SkNx rsqrt0() const { return vrsqrteq_f32(fVec); }
    SkNx rsqrt1() const {
        float32x4_t est0 = this->rsqrt0().fVec;
        return vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est0, est0)), est0);
    }
    SkNx rsqrt2() const {
        float32x4_t est1 = this->rsqrt1().fVec;
        return vmulq_f32(vrsqrtsq_f32(fVec, vmulq_f32(est1, est1)), est1);
    }

    SkNx sqrt() const {
    #if defined(SK_CPU_ARM64)
        return vsqrtq_f32(fVec);
    #else
        return *this * this->rsqrt2();
    #endif
    }

    float operator[](int k) const {
        SkASSERT(0 <= k && k < 4);
        union { float32x4_t v; float fs[4]; } pun = {fVec};
        return pun.fs[k&3];
    }
    template <int k> float kth() const { return (*this)[k]; }

    bool allTrue() const {
        auto v = vreinterpretq_u32_f32(fVec);
        return vgetq_lane_u32(v,0) && vgetq_lane_u32(v,1)
            && vgetq_lane_u32(v,2) && vgetq_lane_u32(v,3);
    }
    bool anyTrue() const {
        auto v = vreinterpretq_u32_f32(fVec);
        return vgetq_lane_u32(v,0) || vgetq_lane_u32(v,1)
            || vgetq_lane_u32(v,2) || vgetq_lane_u32(v,3);
    }

    SkNx thenElse(const SkNx& t, const SkNx& e) const {
        return vbslq_f32(vreinterpretq_u32_f32(fVec), t.fVec, e.fVec);
    }

    float32x4_t fVec;
};

// It's possible that for our current use cases, representing this as
// half a uint16x8_t might be better than representing it as a uint16x4_t.
// It'd make conversion to Sk4b one step simpler.
template <>
class SkNx<4, uint16_t> {
public:
    SkNx(const uint16x4_t& vec) : fVec(vec) {}

    SkNx() {}
    SkNx(uint16_t val) : fVec(vdup_n_u16(val)) {}
    static SkNx Load(const void* ptr) { return vld1_u16((const uint16_t*)ptr); }

    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) {
        fVec = (uint16x4_t) { a,b,c,d };
    }

    void store(void* ptr) const { vst1_u16((uint16_t*)ptr, fVec); }

    SkNx operator + (const SkNx& o) const { return vadd_u16(fVec, o.fVec); }
    SkNx operator - (const SkNx& o) const { return vsub_u16(fVec, o.fVec); }
    SkNx operator * (const SkNx& o) const { return vmul_u16(fVec, o.fVec); }

    SkNx operator << (int bits) const { SHIFT16(vshl_n_u16, fVec, bits); }
    SkNx operator >> (int bits) const { SHIFT16(vshr_n_u16, fVec, bits); }

    static SkNx Min(const SkNx& a, const SkNx& b) { return vmin_u16(a.fVec, b.fVec); }

    uint16_t operator[](int k) const {
        SkASSERT(0 <= k && k < 4);
        union { uint16x4_t v; uint16_t us[4]; } pun = {fVec};
        return pun.us[k&3];
    }
    template <int k> uint16_t kth() const { return (*this)[k]; }

    SkNx thenElse(const SkNx& t, const SkNx& e) const {
        return vbsl_u16(fVec, t.fVec, e.fVec);
    }

    uint16x4_t fVec;
};

template <>
class SkNx<8, uint16_t> {
public:
    SkNx(const uint16x8_t& vec) : fVec(vec) {}

    SkNx() {}
    SkNx(uint16_t val) : fVec(vdupq_n_u16(val)) {}
    static SkNx Load(const void* ptr) { return vld1q_u16((const uint16_t*)ptr); }

    SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
         uint16_t e, uint16_t f, uint16_t g, uint16_t h) {
        fVec = (uint16x8_t) { a,b,c,d, e,f,g,h };
    }

    void store(void* ptr) const { vst1q_u16((uint16_t*)ptr, fVec); }

    SkNx operator + (const SkNx& o) const { return vaddq_u16(fVec, o.fVec); }
    SkNx operator - (const SkNx& o) const { return vsubq_u16(fVec, o.fVec); }
    SkNx operator * (const SkNx& o) const { return vmulq_u16(fVec, o.fVec); }

    SkNx operator << (int bits) const { SHIFT16(vshlq_n_u16, fVec, bits); }
    SkNx operator >> (int bits) const { SHIFT16(vshrq_n_u16, fVec, bits); }

    static SkNx Min(const SkNx& a, const SkNx& b) { return vminq_u16(a.fVec, b.fVec); }

    uint16_t operator[](int k) const {
        SkASSERT(0 <= k && k < 8);
        union { uint16x8_t v; uint16_t us[8]; } pun = {fVec};
        return pun.us[k&7];
    }
    template <int k> uint16_t kth() const { return (*this)[k]; }

    SkNx thenElse(const SkNx& t, const SkNx& e) const {
        return vbslq_u16(fVec, t.fVec, e.fVec);
    }

    uint16x8_t fVec;
};

template <>
class SkNx<4, uint8_t> {
public:
    SkNx(const uint8x8_t& vec) : fVec(vec) {}

    SkNx() {}
    static SkNx Load(const void* ptr) {
        return (uint8x8_t)vld1_dup_u32((const uint32_t*)ptr);
    }
    void store(void* ptr) const {
        return vst1_lane_u32((uint32_t*)ptr, (uint32x2_t)fVec, 0);
    }

    // TODO as needed

    uint8x8_t fVec;
};

template <>
class SkNx<16, uint8_t> {
public:
    SkNx(const uint8x16_t& vec) : fVec(vec) {}

    SkNx() {}
    SkNx(uint8_t val) : fVec(vdupq_n_u8(val)) {}
    static SkNx Load(const void* ptr) { return vld1q_u8((const uint8_t*)ptr); }

    SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
         uint8_t e, uint8_t f, uint8_t g, uint8_t h,
         uint8_t i, uint8_t j, uint8_t k, uint8_t l,
         uint8_t m, uint8_t n, uint8_t o, uint8_t p) {
        fVec = (uint8x16_t) { a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p };
    }

    void store(void* ptr) const { vst1q_u8((uint8_t*)ptr, fVec); }

    SkNx saturatedAdd(const SkNx& o) const { return vqaddq_u8(fVec, o.fVec); }

    SkNx operator + (const SkNx& o) const { return vaddq_u8(fVec, o.fVec); }
    SkNx operator - (const SkNx& o) const { return vsubq_u8(fVec, o.fVec); }

    static SkNx Min(const SkNx& a, const SkNx& b) { return vminq_u8(a.fVec, b.fVec); }
    SkNx operator < (const SkNx& o) const { return vcltq_u8(fVec, o.fVec); }

    uint8_t operator[](int k) const {
        SkASSERT(0 <= k && k < 16);
        union { uint8x16_t v; uint8_t us[16]; } pun = {fVec};
        return pun.us[k&15];
    }
    template <int k> uint8_t kth() const { return (*this)[k]; }

    SkNx thenElse(const SkNx& t, const SkNx& e) const {
        return vbslq_u8(fVec, t.fVec, e.fVec);
    }

    uint8x16_t fVec;
};

#undef SHIFT32
#undef SHIFT16
#undef SHIFT8

template<> inline Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) {
    uint32x4_t _32 = vcvtq_u32_f32(src.fVec);
    uint16x4_t _16 = vqmovn_u32(_32);
    return vqmovn_u16(vcombine_u16(_16, _16));
}

template<> inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) {
    uint16x8_t _16 = vmovl_u8 (src.fVec) ;
    uint32x4_t _32 = vmovl_u16(vget_low_u16(_16));
    return vcvtq_f32_u32(_32);
}

static inline void Sk4f_ToBytes(uint8_t bytes[16],
                                const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) {
    vst1q_u8(bytes, vuzpq_u8(vuzpq_u8((uint8x16_t)vcvtq_u32_f32(a.fVec),
                                      (uint8x16_t)vcvtq_u32_f32(b.fVec)).val[0],
                             vuzpq_u8((uint8x16_t)vcvtq_u32_f32(c.fVec),
                                      (uint8x16_t)vcvtq_u32_f32(d.fVec)).val[0]).val[0]);
}

template<> inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
    return vget_low_u16(vmovl_u8(src.fVec));
}

template<> inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
    return vmovn_u16(vcombine_u16(src.fVec, src.fVec));
}

#endif//SkNx_neon_DEFINED