| /* |
| * 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_DEFINED |
| #define SkNx_DEFINED |
| |
| #include "SkSafe_math.h" |
| #include "SkScalar.h" |
| #include "SkTypes.h" |
| #include <limits> |
| #include <type_traits> |
| |
| // These _abi types are data-only, and so can be used to store SkNx in structs or |
| // pass them as function parameters or return values, even across compilation units. |
| template <int N, typename T> struct SkNx_abi { SkNx_abi<N/2,T> lo, hi; }; |
| template < typename T> struct SkNx_abi<1,T> { T val; }; |
| |
| // Every single SkNx method wants to be fully inlined. (We know better than MSVC). |
| #define AI SK_ALWAYS_INLINE |
| |
| namespace { |
| |
| // The default SkNx<N,T> just proxies down to a pair of SkNx<N/2, T>. |
| template <int N, typename T> |
| struct SkNx { |
| typedef SkNx<N/2, T> Half; |
| |
| Half fLo, fHi; |
| |
| AI SkNx() = default; |
| AI SkNx(const Half& lo, const Half& hi) : fLo(lo), fHi(hi) {} |
| |
| AI SkNx(T v) : fLo(v), fHi(v) {} |
| |
| AI SkNx(T a, T b) : fLo(a) , fHi(b) { static_assert(N==2, ""); } |
| AI SkNx(T a, T b, T c, T d) : fLo(a,b), fHi(c,d) { static_assert(N==4, ""); } |
| AI SkNx(T a, T b, T c, T d, T e, T f, T g, T h) : fLo(a,b,c,d), fHi(e,f,g,h) { |
| static_assert(N==8, ""); |
| } |
| AI SkNx(T a, T b, T c, T d, T e, T f, T g, T h, |
| T i, T j, T k, T l, T m, T n, T o, T p) |
| : fLo(a,b,c,d, e,f,g,h), fHi(i,j,k,l, m,n,o,p) { |
| static_assert(N==16, ""); |
| } |
| |
| AI SkNx(const SkNx_abi<N,T>& a) : fLo(a.lo), fHi(a.hi) {} |
| AI operator SkNx_abi<N,T>() const { return { (SkNx_abi<N/2,T>)fLo, (SkNx_abi<N/2,T>)fHi }; } |
| |
| AI T operator[](int k) const { |
| SkASSERT(0 <= k && k < N); |
| return k < N/2 ? fLo[k] : fHi[k-N/2]; |
| } |
| |
| AI static SkNx Load(const void* vptr) { |
| auto ptr = (const char*)vptr; |
| return { Half::Load(ptr), Half::Load(ptr + N/2*sizeof(T)) }; |
| } |
| AI void store(void* vptr) const { |
| auto ptr = (char*)vptr; |
| fLo.store(ptr); |
| fHi.store(ptr + N/2*sizeof(T)); |
| } |
| |
| AI static void Load4(const void* vptr, SkNx* a, SkNx* b, SkNx* c, SkNx* d) { |
| auto ptr = (const char*)vptr; |
| Half al, bl, cl, dl, |
| ah, bh, ch, dh; |
| Half::Load4(ptr , &al, &bl, &cl, &dl); |
| Half::Load4(ptr + 4*N/2*sizeof(T), &ah, &bh, &ch, &dh); |
| *a = SkNx{al, ah}; |
| *b = SkNx{bl, bh}; |
| *c = SkNx{cl, ch}; |
| *d = SkNx{dl, dh}; |
| } |
| AI static void Store4(void* vptr, const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) { |
| auto ptr = (char*)vptr; |
| Half::Store4(ptr, a.fLo, b.fLo, c.fLo, d.fLo); |
| Half::Store4(ptr + 4*N/2*sizeof(T), a.fHi, b.fHi, c.fHi, d.fHi); |
| } |
| |
| AI bool anyTrue() const { return fLo.anyTrue() || fHi.anyTrue(); } |
| AI bool allTrue() const { return fLo.allTrue() && fHi.allTrue(); } |
| |
| AI SkNx abs() const { return { fLo. abs(), fHi. abs() }; } |
| AI SkNx sqrt() const { return { fLo. sqrt(), fHi. sqrt() }; } |
| AI SkNx rsqrt() const { return { fLo. rsqrt(), fHi. rsqrt() }; } |
| AI SkNx floor() const { return { fLo. floor(), fHi. floor() }; } |
| AI SkNx invert() const { return { fLo.invert(), fHi.invert() }; } |
| |
| AI SkNx operator!() const { return { !fLo, !fHi }; } |
| AI SkNx operator-() const { return { -fLo, -fHi }; } |
| AI SkNx operator~() const { return { ~fLo, ~fHi }; } |
| |
| AI SkNx operator<<(int bits) const { return { fLo << bits, fHi << bits }; } |
| AI SkNx operator>>(int bits) const { return { fLo >> bits, fHi >> bits }; } |
| |
| AI SkNx operator+(const SkNx& y) const { return { fLo + y.fLo, fHi + y.fHi }; } |
| AI SkNx operator-(const SkNx& y) const { return { fLo - y.fLo, fHi - y.fHi }; } |
| AI SkNx operator*(const SkNx& y) const { return { fLo * y.fLo, fHi * y.fHi }; } |
| AI SkNx operator/(const SkNx& y) const { return { fLo / y.fLo, fHi / y.fHi }; } |
| |
| AI SkNx operator&(const SkNx& y) const { return { fLo & y.fLo, fHi & y.fHi }; } |
| AI SkNx operator|(const SkNx& y) const { return { fLo | y.fLo, fHi | y.fHi }; } |
| AI SkNx operator^(const SkNx& y) const { return { fLo ^ y.fLo, fHi ^ y.fHi }; } |
| |
| AI SkNx operator==(const SkNx& y) const { return { fLo == y.fLo, fHi == y.fHi }; } |
| AI SkNx operator!=(const SkNx& y) const { return { fLo != y.fLo, fHi != y.fHi }; } |
| AI SkNx operator<=(const SkNx& y) const { return { fLo <= y.fLo, fHi <= y.fHi }; } |
| AI SkNx operator>=(const SkNx& y) const { return { fLo >= y.fLo, fHi >= y.fHi }; } |
| AI SkNx operator< (const SkNx& y) const { return { fLo < y.fLo, fHi < y.fHi }; } |
| AI SkNx operator> (const SkNx& y) const { return { fLo > y.fLo, fHi > y.fHi }; } |
| |
| AI SkNx saturatedAdd(const SkNx& y) const { |
| return { fLo.saturatedAdd(y.fLo), fHi.saturatedAdd(y.fHi) }; |
| } |
| AI SkNx thenElse(const SkNx& t, const SkNx& e) const { |
| return { fLo.thenElse(t.fLo, e.fLo), fHi.thenElse(t.fHi, e.fHi) }; |
| } |
| |
| AI static SkNx Min(const SkNx& x, const SkNx& y) { |
| return { Half::Min(x.fLo, y.fLo), Half::Min(x.fHi, y.fHi) }; |
| } |
| AI static SkNx Max(const SkNx& x, const SkNx& y) { |
| return { Half::Max(x.fLo, y.fLo), Half::Max(x.fHi, y.fHi) }; |
| } |
| }; |
| |
| // The N -> N/2 recursion bottoms out at N == 1, a scalar value. |
| template <typename T> |
| struct SkNx<1,T> { |
| T fVal; |
| |
| AI SkNx() = default; |
| AI SkNx(T v) : fVal(v) {} |
| |
| AI SkNx(const SkNx_abi<1,T>& a) : fVal(a.val) {} |
| AI operator SkNx_abi<1,T>() const { return { fVal }; } |
| |
| // Android complains against unused parameters, so we guard it |
| AI T operator[](int SkDEBUGCODE(k)) const { |
| SkASSERT(k == 0); |
| return fVal; |
| } |
| |
| AI static SkNx Load(const void* ptr) { |
| SkNx v; |
| memcpy(&v, ptr, sizeof(T)); |
| return v; |
| } |
| AI void store(void* ptr) const { memcpy(ptr, &fVal, sizeof(T)); } |
| |
| AI static void Load4(const void* vptr, SkNx* a, SkNx* b, SkNx* c, SkNx* d) { |
| auto ptr = (const char*)vptr; |
| *a = Load(ptr + 0*sizeof(T)); |
| *b = Load(ptr + 1*sizeof(T)); |
| *c = Load(ptr + 2*sizeof(T)); |
| *d = Load(ptr + 3*sizeof(T)); |
| } |
| AI static void Store4(void* vptr, const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) { |
| auto ptr = (char*)vptr; |
| a.store(ptr + 0*sizeof(T)); |
| b.store(ptr + 1*sizeof(T)); |
| c.store(ptr + 2*sizeof(T)); |
| d.store(ptr + 3*sizeof(T)); |
| } |
| |
| AI bool anyTrue() const { return fVal != 0; } |
| AI bool allTrue() const { return fVal != 0; } |
| |
| AI SkNx abs() const { return Abs(fVal); } |
| AI SkNx sqrt() const { return Sqrt(fVal); } |
| AI SkNx rsqrt() const { return T(1) / this->sqrt(); } |
| AI SkNx floor() const { return Floor(fVal); } |
| AI SkNx invert() const { return T(1) / *this; } |
| |
| AI SkNx operator!() const { return !fVal; } |
| AI SkNx operator-() const { return -fVal; } |
| AI SkNx operator~() const { return FromBits(~ToBits(fVal)); } |
| |
| AI SkNx operator<<(int bits) const { return fVal << bits; } |
| AI SkNx operator>>(int bits) const { return fVal >> bits; } |
| |
| AI SkNx operator+(const SkNx& y) const { return fVal + y.fVal; } |
| AI SkNx operator-(const SkNx& y) const { return fVal - y.fVal; } |
| AI SkNx operator*(const SkNx& y) const { return fVal * y.fVal; } |
| AI SkNx operator/(const SkNx& y) const { return fVal / y.fVal; } |
| |
| AI SkNx operator&(const SkNx& y) const { return FromBits(ToBits(fVal) & ToBits(y.fVal)); } |
| AI SkNx operator|(const SkNx& y) const { return FromBits(ToBits(fVal) | ToBits(y.fVal)); } |
| AI SkNx operator^(const SkNx& y) const { return FromBits(ToBits(fVal) ^ ToBits(y.fVal)); } |
| |
| AI SkNx operator==(const SkNx& y) const { return FromBits(fVal == y.fVal ? ~0 : 0); } |
| AI SkNx operator!=(const SkNx& y) const { return FromBits(fVal != y.fVal ? ~0 : 0); } |
| AI SkNx operator<=(const SkNx& y) const { return FromBits(fVal <= y.fVal ? ~0 : 0); } |
| AI SkNx operator>=(const SkNx& y) const { return FromBits(fVal >= y.fVal ? ~0 : 0); } |
| AI SkNx operator< (const SkNx& y) const { return FromBits(fVal < y.fVal ? ~0 : 0); } |
| AI SkNx operator> (const SkNx& y) const { return FromBits(fVal > y.fVal ? ~0 : 0); } |
| |
| AI static SkNx Min(const SkNx& x, const SkNx& y) { return x.fVal < y.fVal ? x : y; } |
| AI static SkNx Max(const SkNx& x, const SkNx& y) { return x.fVal > y.fVal ? x : y; } |
| |
| AI SkNx saturatedAdd(const SkNx& y) const { |
| static_assert(std::is_unsigned<T>::value, ""); |
| T sum = fVal + y.fVal; |
| return sum < fVal ? std::numeric_limits<T>::max() : sum; |
| } |
| |
| AI SkNx thenElse(const SkNx& t, const SkNx& e) const { return fVal != 0 ? t : e; } |
| |
| private: |
| // Helper functions to choose the right float/double methods. (In <cmath> madness lies...) |
| AI static float Abs(float val) { return ::fabsf(val); } |
| AI static float Sqrt(float val) { return ::sqrtf(val); } |
| AI static float Floor(float val) { return ::floorf(val); } |
| |
| AI static double Abs(double val) { return ::fabs(val); } |
| AI static double Sqrt(double val) { return ::sqrt(val); } |
| AI static double Floor(double val) { return ::floor(val); } |
| |
| // Helper functions for working with floats/doubles as bit patterns. |
| template <typename U> |
| AI static U ToBits(U v) { return v; } |
| AI static int32_t ToBits(float v) { int32_t bits; memcpy(&bits, &v, sizeof(v)); return bits; } |
| AI static int64_t ToBits(double v) { int64_t bits; memcpy(&bits, &v, sizeof(v)); return bits; } |
| |
| template <typename Bits> |
| AI static T FromBits(Bits bits) { |
| static_assert(std::is_pod<T >::value && |
| std::is_pod<Bits>::value && |
| sizeof(T) <= sizeof(Bits), ""); |
| T val; |
| memcpy(&val, &bits, sizeof(T)); |
| return val; |
| } |
| }; |
| |
| // Allow scalars on the left or right of binary operators, and things like +=, &=, etc. |
| #define V template <int N, typename T> AI static SkNx<N,T> |
| V operator+ (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) + y; } |
| V operator- (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) - y; } |
| V operator* (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) * y; } |
| V operator/ (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) / y; } |
| V operator& (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) & y; } |
| V operator| (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) | y; } |
| V operator^ (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) ^ y; } |
| V operator==(T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) == y; } |
| V operator!=(T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) != y; } |
| V operator<=(T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) <= y; } |
| V operator>=(T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) >= y; } |
| V operator< (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) < y; } |
| V operator> (T x, const SkNx<N,T>& y) { return SkNx<N,T>(x) > y; } |
| |
| V operator+ (const SkNx<N,T>& x, T y) { return x + SkNx<N,T>(y); } |
| V operator- (const SkNx<N,T>& x, T y) { return x - SkNx<N,T>(y); } |
| V operator* (const SkNx<N,T>& x, T y) { return x * SkNx<N,T>(y); } |
| V operator/ (const SkNx<N,T>& x, T y) { return x / SkNx<N,T>(y); } |
| V operator& (const SkNx<N,T>& x, T y) { return x & SkNx<N,T>(y); } |
| V operator| (const SkNx<N,T>& x, T y) { return x | SkNx<N,T>(y); } |
| V operator^ (const SkNx<N,T>& x, T y) { return x ^ SkNx<N,T>(y); } |
| V operator==(const SkNx<N,T>& x, T y) { return x == SkNx<N,T>(y); } |
| V operator!=(const SkNx<N,T>& x, T y) { return x != SkNx<N,T>(y); } |
| V operator<=(const SkNx<N,T>& x, T y) { return x <= SkNx<N,T>(y); } |
| V operator>=(const SkNx<N,T>& x, T y) { return x >= SkNx<N,T>(y); } |
| V operator< (const SkNx<N,T>& x, T y) { return x < SkNx<N,T>(y); } |
| V operator> (const SkNx<N,T>& x, T y) { return x > SkNx<N,T>(y); } |
| |
| V& operator<<=(SkNx<N,T>& x, int bits) { return (x = x << bits); } |
| V& operator>>=(SkNx<N,T>& x, int bits) { return (x = x >> bits); } |
| |
| V& operator +=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x + y); } |
| V& operator -=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x - y); } |
| V& operator *=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x * y); } |
| V& operator /=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x / y); } |
| V& operator &=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x & y); } |
| V& operator |=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x | y); } |
| V& operator ^=(SkNx<N,T>& x, const SkNx<N,T>& y) { return (x = x ^ y); } |
| |
| V& operator +=(SkNx<N,T>& x, T y) { return (x = x + SkNx<N,T>(y)); } |
| V& operator -=(SkNx<N,T>& x, T y) { return (x = x - SkNx<N,T>(y)); } |
| V& operator *=(SkNx<N,T>& x, T y) { return (x = x * SkNx<N,T>(y)); } |
| V& operator /=(SkNx<N,T>& x, T y) { return (x = x / SkNx<N,T>(y)); } |
| V& operator &=(SkNx<N,T>& x, T y) { return (x = x & SkNx<N,T>(y)); } |
| V& operator |=(SkNx<N,T>& x, T y) { return (x = x | SkNx<N,T>(y)); } |
| V& operator ^=(SkNx<N,T>& x, T y) { return (x = x ^ SkNx<N,T>(y)); } |
| #undef V |
| |
| // SkNx<N,T> ~~> SkNx<N/2,T> + SkNx<N/2,T> |
| template <int N, typename T> |
| AI static void SkNx_split(const SkNx<N,T>& v, SkNx<N/2,T>* lo, SkNx<N/2,T>* hi) { |
| *lo = v.fLo; |
| *hi = v.fHi; |
| } |
| |
| // SkNx<N/2,T> + SkNx<N/2,T> ~~> SkNx<N,T> |
| template <int N, typename T> |
| AI static SkNx<N*2,T> SkNx_join(const SkNx<N,T>& lo, const SkNx<N,T>& hi) { |
| return { lo, hi }; |
| } |
| |
| // A very generic shuffle. Can reorder, duplicate, contract, expand... |
| // Sk4f v = { R,G,B,A }; |
| // SkNx_shuffle<2,1,0,3>(v) ~~> {B,G,R,A} |
| // SkNx_shuffle<2,1>(v) ~~> {B,G} |
| // SkNx_shuffle<2,1,2,1,2,1,2,1>(v) ~~> {B,G,B,G,B,G,B,G} |
| // SkNx_shuffle<3,3,3,3>(v) ~~> {A,A,A,A} |
| template <int... Ix, int N, typename T> |
| AI static SkNx<sizeof...(Ix),T> SkNx_shuffle(const SkNx<N,T>& v) { |
| return { v[Ix]... }; |
| } |
| |
| // Cast from SkNx<N, Src> to SkNx<N, Dst>, as if you called static_cast<Dst>(Src). |
| template <typename Dst, typename Src, int N> |
| AI static SkNx<N,Dst> SkNx_cast(const SkNx<N,Src>& v) { |
| return { SkNx_cast<Dst>(v.fLo), SkNx_cast<Dst>(v.fHi) }; |
| } |
| template <typename Dst, typename Src> |
| AI static SkNx<1,Dst> SkNx_cast(const SkNx<1,Src>& v) { |
| return static_cast<Dst>(v.fVal); |
| } |
| |
| template <int N, typename T> |
| AI static SkNx<N,T> SkNx_fma(const SkNx<N,T>& f, const SkNx<N,T>& m, const SkNx<N,T>& a) { |
| return f*m+a; |
| } |
| |
| } // namespace |
| |
| typedef SkNx<2, float> Sk2f; |
| typedef SkNx<4, float> Sk4f; |
| typedef SkNx<8, float> Sk8f; |
| typedef SkNx<16, float> Sk16f; |
| |
| typedef SkNx<2, SkScalar> Sk2s; |
| typedef SkNx<4, SkScalar> Sk4s; |
| typedef SkNx<8, SkScalar> Sk8s; |
| typedef SkNx<16, SkScalar> Sk16s; |
| |
| typedef SkNx<4, uint8_t> Sk4b; |
| typedef SkNx<8, uint8_t> Sk8b; |
| typedef SkNx<16, uint8_t> Sk16b; |
| |
| typedef SkNx<4, uint16_t> Sk4h; |
| typedef SkNx<8, uint16_t> Sk8h; |
| typedef SkNx<16, uint16_t> Sk16h; |
| |
| typedef SkNx<4, int32_t> Sk4i; |
| typedef SkNx<8, int32_t> Sk8i; |
| typedef SkNx<4, uint32_t> Sk4u; |
| |
| // Include platform specific specializations if available. |
| #if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
| #include "../opts/SkNx_sse.h" |
| #elif !defined(SKNX_NO_SIMD) && defined(SK_ARM_HAS_NEON) |
| #include "../opts/SkNx_neon.h" |
| #else |
| |
| AI static Sk4i Sk4f_round(const Sk4f& x) { |
| return { (int) lrintf (x[0]), |
| (int) lrintf (x[1]), |
| (int) lrintf (x[2]), |
| (int) lrintf (x[3]), }; |
| } |
| |
| #endif |
| |
| AI static void Sk4f_ToBytes(uint8_t p[16], |
| const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) { |
| SkNx_cast<uint8_t>(SkNx_join(SkNx_join(a,b), SkNx_join(c,d))).store(p); |
| } |
| |
| #undef AI |
| |
| #endif//SkNx_DEFINED |