michael@0: /* -*- Mode: C++; tab-width: 20; indent-tabs-mode: nil; c-basic-offset: 2 -*- michael@0: * This Source Code Form is subject to the terms of the Mozilla Public michael@0: * License, v. 2.0. If a copy of the MPL was not distributed with this michael@0: * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ michael@0: michael@0: #ifndef _MOZILLA_GFX_SIMD_H_ michael@0: #define _MOZILLA_GFX_SIMD_H_ michael@0: michael@0: /** michael@0: * Consumers of this file need to #define SIMD_COMPILE_SSE2 before including it michael@0: * if they want access to the SSE2 functions. michael@0: */ michael@0: michael@0: #ifdef SIMD_COMPILE_SSE2 michael@0: #include michael@0: #endif michael@0: michael@0: namespace mozilla { michael@0: namespace gfx { michael@0: michael@0: namespace simd { michael@0: michael@0: template michael@0: u8x16_t Load8(const uint8_t* aSource); michael@0: michael@0: template michael@0: u8x16_t From8(uint8_t a, uint8_t b, uint8_t c, uint8_t d, uint8_t e, uint8_t f, uint8_t g, uint8_t h, michael@0: uint8_t i, uint8_t j, uint8_t k, uint8_t l, uint8_t m, uint8_t n, uint8_t o, uint8_t p); michael@0: michael@0: template michael@0: u8x16_t FromZero8(); michael@0: michael@0: template michael@0: i16x8_t FromI16(int16_t a, int16_t b, int16_t c, int16_t d, int16_t e, int16_t f, int16_t g, int16_t h); michael@0: michael@0: template michael@0: u16x8_t FromU16(uint16_t a, uint16_t b, uint16_t c, uint16_t d, uint16_t e, uint16_t f, uint16_t g, uint16_t h); michael@0: michael@0: template michael@0: i16x8_t FromI16(int16_t a); michael@0: michael@0: template michael@0: u16x8_t FromU16(uint16_t a); michael@0: michael@0: template michael@0: i32x4_t From32(int32_t a, int32_t b, int32_t c, int32_t d); michael@0: michael@0: template michael@0: i32x4_t From32(int32_t a); michael@0: michael@0: template michael@0: f32x4_t FromF32(float a, float b, float c, float d); michael@0: michael@0: template michael@0: f32x4_t FromF32(float a); michael@0: michael@0: // All SIMD backends overload these functions for their SIMD types: michael@0: michael@0: #if 0 michael@0: michael@0: // Store 16 bytes to a 16-byte aligned address michael@0: void Store8(uint8_t* aTarget, u8x16_t aM); michael@0: michael@0: // Fixed shifts michael@0: template i16x8_t ShiftRight16(i16x8_t aM); michael@0: template i32x4_t ShiftRight32(i32x4_t aM); michael@0: michael@0: i16x8_t Add16(i16x8_t aM1, i16x8_t aM2); michael@0: i32x4_t Add32(i32x4_t aM1, i32x4_t aM2); michael@0: i16x8_t Sub16(i16x8_t aM1, i16x8_t aM2); michael@0: i32x4_t Sub32(i32x4_t aM1, i32x4_t aM2); michael@0: u8x16_t Min8(u8x16_t aM1, iu8x16_t aM2); michael@0: u8x16_t Max8(u8x16_t aM1, iu8x16_t aM2); michael@0: i32x4_t Min32(i32x4_t aM1, i32x4_t aM2); michael@0: i32x4_t Max32(i32x4_t aM1, i32x4_t aM2); michael@0: michael@0: // Truncating i16 -> i16 multiplication michael@0: i16x8_t Mul16(i16x8_t aM1, i16x8_t aM2); michael@0: michael@0: // Long multiplication i16 -> i32 michael@0: // aFactorsA1B1 = (a1[4] b1[4]) michael@0: // aFactorsA2B2 = (a2[4] b2[4]) michael@0: // aProductA = a1 * a2, aProductB = b1 * b2 michael@0: void Mul16x4x2x2To32x4x2(i16x8_t aFactorsA1B1, i16x8_t aFactorsA2B2, michael@0: i32x4_t& aProductA, i32x4_t& aProductB); michael@0: michael@0: // Long multiplication + pairwise addition i16 -> i32 michael@0: // See the scalar implementation for specifics. michael@0: i32x4_t MulAdd16x8x2To32x4(i16x8_t aFactorsA, i16x8_t aFactorsB); michael@0: i32x4_t MulAdd16x8x2To32x4(u16x8_t aFactorsA, u16x8_t aFactorsB); michael@0: michael@0: // Set all four 32-bit components to the value of the component at aIndex. michael@0: template michael@0: i32x4_t Splat32(i32x4_t aM); michael@0: michael@0: // Interpret the input as four 32-bit values, apply Splat32 on them, michael@0: // re-interpret the result as sixteen 8-bit values. michael@0: template michael@0: u8x16_t Splat32On8(u8x16_t aM); michael@0: michael@0: template i32x4 Shuffle32(i32x4 aM); michael@0: template i16x8 ShuffleLo16(i16x8 aM); michael@0: template i16x8 ShuffleHi16(i16x8 aM); michael@0: michael@0: u8x16_t InterleaveLo8(u8x16_t m1, u8x16_t m2); michael@0: u8x16_t InterleaveHi8(u8x16_t m1, u8x16_t m2); michael@0: i16x8_t InterleaveLo16(i16x8_t m1, i16x8_t m2); michael@0: i16x8_t InterleaveHi16(i16x8_t m1, i16x8_t m2); michael@0: i32x4_t InterleaveLo32(i32x4_t m1, i32x4_t m2); michael@0: michael@0: i16x8_t UnpackLo8x8ToI16x8(u8x16_t m); michael@0: i16x8_t UnpackHi8x8ToI16x8(u8x16_t m); michael@0: u16x8_t UnpackLo8x8ToU16x8(u8x16_t m); michael@0: u16x8_t UnpackHi8x8ToU16x8(u8x16_t m); michael@0: michael@0: i16x8_t PackAndSaturate32To16(i32x4_t m1, i32x4_t m2); michael@0: u8x16_t PackAndSaturate16To8(i16x8_t m1, i16x8_t m2); michael@0: u8x16_t PackAndSaturate32To8(i32x4_t m1, i32x4_t m2, i32x4_t m3, const i32x4_t& m4); michael@0: michael@0: i32x4 FastDivideBy255(i32x4 m); michael@0: i16x8 FastDivideBy255_16(i16x8 m); michael@0: michael@0: #endif michael@0: michael@0: // Scalar michael@0: michael@0: struct Scalaru8x16_t { michael@0: uint8_t u8[16]; michael@0: }; michael@0: michael@0: union Scalari16x8_t { michael@0: int16_t i16[8]; michael@0: uint16_t u16[8]; michael@0: }; michael@0: michael@0: typedef Scalari16x8_t Scalaru16x8_t; michael@0: michael@0: struct Scalari32x4_t { michael@0: int32_t i32[4]; michael@0: }; michael@0: michael@0: struct Scalarf32x4_t { michael@0: float f32[4]; michael@0: }; michael@0: michael@0: template<> michael@0: inline Scalaru8x16_t michael@0: Load8(const uint8_t* aSource) michael@0: { michael@0: return *(Scalaru8x16_t*)aSource; michael@0: } michael@0: michael@0: inline void Store8(uint8_t* aTarget, Scalaru8x16_t aM) michael@0: { michael@0: *(Scalaru8x16_t*)aTarget = aM; michael@0: } michael@0: michael@0: template<> michael@0: inline Scalaru8x16_t From8(uint8_t a, uint8_t b, uint8_t c, uint8_t d, uint8_t e, uint8_t f, uint8_t g, uint8_t h, michael@0: uint8_t i, uint8_t j, uint8_t k, uint8_t l, uint8_t m, uint8_t n, uint8_t o, uint8_t p) michael@0: { michael@0: Scalaru8x16_t _m; michael@0: _m.u8[0] = a; michael@0: _m.u8[1] = b; michael@0: _m.u8[2] = c; michael@0: _m.u8[3] = d; michael@0: _m.u8[4] = e; michael@0: _m.u8[5] = f; michael@0: _m.u8[6] = g; michael@0: _m.u8[7] = h; michael@0: _m.u8[8+0] = i; michael@0: _m.u8[8+1] = j; michael@0: _m.u8[8+2] = k; michael@0: _m.u8[8+3] = l; michael@0: _m.u8[8+4] = m; michael@0: _m.u8[8+5] = n; michael@0: _m.u8[8+6] = o; michael@0: _m.u8[8+7] = p; michael@0: return _m; michael@0: } michael@0: michael@0: template<> michael@0: inline Scalaru8x16_t FromZero8() michael@0: { michael@0: return From8(0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0); michael@0: } michael@0: michael@0: template<> michael@0: inline Scalari16x8_t FromI16(int16_t a, int16_t b, int16_t c, int16_t d, int16_t e, int16_t f, int16_t g, int16_t h) michael@0: { michael@0: Scalari16x8_t m; michael@0: m.i16[0] = a; michael@0: m.i16[1] = b; michael@0: m.i16[2] = c; michael@0: m.i16[3] = d; michael@0: m.i16[4] = e; michael@0: m.i16[5] = f; michael@0: m.i16[6] = g; michael@0: m.i16[7] = h; michael@0: return m; michael@0: } michael@0: michael@0: template<> michael@0: inline Scalaru16x8_t FromU16(uint16_t a, uint16_t b, uint16_t c, uint16_t d, uint16_t e, uint16_t f, uint16_t g, uint16_t h) michael@0: { michael@0: Scalaru16x8_t m; michael@0: m.u16[0] = a; michael@0: m.u16[1] = b; michael@0: m.u16[2] = c; michael@0: m.u16[3] = d; michael@0: m.u16[4] = e; michael@0: m.u16[5] = f; michael@0: m.u16[6] = g; michael@0: m.u16[7] = h; michael@0: return m; michael@0: } michael@0: michael@0: template<> michael@0: inline Scalari16x8_t FromI16(int16_t a) michael@0: { michael@0: return FromI16(a, a, a, a, a, a, a, a); michael@0: } michael@0: michael@0: template<> michael@0: inline Scalaru16x8_t FromU16(uint16_t a) michael@0: { michael@0: return FromU16(a, a, a, a, a, a, a, a); michael@0: } michael@0: michael@0: template<> michael@0: inline Scalari32x4_t From32(int32_t a, int32_t b, int32_t c, int32_t d) michael@0: { michael@0: Scalari32x4_t m; michael@0: m.i32[0] = a; michael@0: m.i32[1] = b; michael@0: m.i32[2] = c; michael@0: m.i32[3] = d; michael@0: return m; michael@0: } michael@0: michael@0: template<> michael@0: inline Scalarf32x4_t FromF32(float a, float b, float c, float d) michael@0: { michael@0: Scalarf32x4_t m; michael@0: m.f32[0] = a; michael@0: m.f32[1] = b; michael@0: m.f32[2] = c; michael@0: m.f32[3] = d; michael@0: return m; michael@0: } michael@0: michael@0: template<> michael@0: inline Scalarf32x4_t FromF32(float a) michael@0: { michael@0: return FromF32(a, a, a, a); michael@0: } michael@0: michael@0: template<> michael@0: inline Scalari32x4_t From32(int32_t a) michael@0: { michael@0: return From32(a, a, a, a); michael@0: } michael@0: michael@0: template michael@0: inline Scalari16x8_t ShiftRight16(Scalari16x8_t aM) michael@0: { michael@0: return FromI16(uint16_t(aM.i16[0]) >> aNumberOfBits, uint16_t(aM.i16[1]) >> aNumberOfBits, michael@0: uint16_t(aM.i16[2]) >> aNumberOfBits, uint16_t(aM.i16[3]) >> aNumberOfBits, michael@0: uint16_t(aM.i16[4]) >> aNumberOfBits, uint16_t(aM.i16[5]) >> aNumberOfBits, michael@0: uint16_t(aM.i16[6]) >> aNumberOfBits, uint16_t(aM.i16[7]) >> aNumberOfBits); michael@0: } michael@0: michael@0: template michael@0: inline Scalari32x4_t ShiftRight32(Scalari32x4_t aM) michael@0: { michael@0: return From32(aM.i32[0] >> aNumberOfBits, aM.i32[1] >> aNumberOfBits, michael@0: aM.i32[2] >> aNumberOfBits, aM.i32[3] >> aNumberOfBits); michael@0: } michael@0: michael@0: inline Scalaru16x8_t Add16(Scalaru16x8_t aM1, Scalaru16x8_t aM2) michael@0: { michael@0: return FromU16(aM1.u16[0] + aM2.u16[0], aM1.u16[1] + aM2.u16[1], michael@0: aM1.u16[2] + aM2.u16[2], aM1.u16[3] + aM2.u16[3], michael@0: aM1.u16[4] + aM2.u16[4], aM1.u16[5] + aM2.u16[5], michael@0: aM1.u16[6] + aM2.u16[6], aM1.u16[7] + aM2.u16[7]); michael@0: } michael@0: michael@0: inline Scalari32x4_t Add32(Scalari32x4_t aM1, Scalari32x4_t aM2) michael@0: { michael@0: return From32(aM1.i32[0] + aM2.i32[0], aM1.i32[1] + aM2.i32[1], michael@0: aM1.i32[2] + aM2.i32[2], aM1.i32[3] + aM2.i32[3]); michael@0: } michael@0: michael@0: inline Scalaru16x8_t Sub16(Scalaru16x8_t aM1, Scalaru16x8_t aM2) michael@0: { michael@0: return FromU16(aM1.u16[0] - aM2.u16[0], aM1.u16[1] - aM2.u16[1], michael@0: aM1.u16[2] - aM2.u16[2], aM1.u16[3] - aM2.u16[3], michael@0: aM1.u16[4] - aM2.u16[4], aM1.u16[5] - aM2.u16[5], michael@0: aM1.u16[6] - aM2.u16[6], aM1.u16[7] - aM2.u16[7]); michael@0: } michael@0: michael@0: inline Scalari32x4_t Sub32(Scalari32x4_t aM1, Scalari32x4_t aM2) michael@0: { michael@0: return From32(aM1.i32[0] - aM2.i32[0], aM1.i32[1] - aM2.i32[1], michael@0: aM1.i32[2] - aM2.i32[2], aM1.i32[3] - aM2.i32[3]); michael@0: } michael@0: michael@0: inline int32_t michael@0: umin(int32_t a, int32_t b) michael@0: { michael@0: return a - ((a - b) & -(a > b)); michael@0: } michael@0: michael@0: inline int32_t michael@0: umax(int32_t a, int32_t b) michael@0: { michael@0: return a - ((a - b) & -(a < b)); michael@0: } michael@0: michael@0: inline Scalaru8x16_t Min8(Scalaru8x16_t aM1, Scalaru8x16_t aM2) michael@0: { michael@0: return From8(umin(aM1.u8[0], aM2.u8[0]), umin(aM1.u8[1], aM2.u8[1]), michael@0: umin(aM1.u8[2], aM2.u8[2]), umin(aM1.u8[3], aM2.u8[3]), michael@0: umin(aM1.u8[4], aM2.u8[4]), umin(aM1.u8[5], aM2.u8[5]), michael@0: umin(aM1.u8[6], aM2.u8[6]), umin(aM1.u8[7], aM2.u8[7]), michael@0: umin(aM1.u8[8+0], aM2.u8[8+0]), umin(aM1.u8[8+1], aM2.u8[8+1]), michael@0: umin(aM1.u8[8+2], aM2.u8[8+2]), umin(aM1.u8[8+3], aM2.u8[8+3]), michael@0: umin(aM1.u8[8+4], aM2.u8[8+4]), umin(aM1.u8[8+5], aM2.u8[8+5]), michael@0: umin(aM1.u8[8+6], aM2.u8[8+6]), umin(aM1.u8[8+7], aM2.u8[8+7])); michael@0: } michael@0: michael@0: inline Scalaru8x16_t Max8(Scalaru8x16_t aM1, Scalaru8x16_t aM2) michael@0: { michael@0: return From8(umax(aM1.u8[0], aM2.u8[0]), umax(aM1.u8[1], aM2.u8[1]), michael@0: umax(aM1.u8[2], aM2.u8[2]), umax(aM1.u8[3], aM2.u8[3]), michael@0: umax(aM1.u8[4], aM2.u8[4]), umax(aM1.u8[5], aM2.u8[5]), michael@0: umax(aM1.u8[6], aM2.u8[6]), umax(aM1.u8[7], aM2.u8[7]), michael@0: umax(aM1.u8[8+0], aM2.u8[8+0]), umax(aM1.u8[8+1], aM2.u8[8+1]), michael@0: umax(aM1.u8[8+2], aM2.u8[8+2]), umax(aM1.u8[8+3], aM2.u8[8+3]), michael@0: umax(aM1.u8[8+4], aM2.u8[8+4]), umax(aM1.u8[8+5], aM2.u8[8+5]), michael@0: umax(aM1.u8[8+6], aM2.u8[8+6]), umax(aM1.u8[8+7], aM2.u8[8+7])); michael@0: } michael@0: michael@0: inline Scalari32x4_t Min32(Scalari32x4_t aM1, Scalari32x4_t aM2) michael@0: { michael@0: return From32(umin(aM1.i32[0], aM2.i32[0]), umin(aM1.i32[1], aM2.i32[1]), michael@0: umin(aM1.i32[2], aM2.i32[2]), umin(aM1.i32[3], aM2.i32[3])); michael@0: } michael@0: michael@0: inline Scalari32x4_t Max32(Scalari32x4_t aM1, Scalari32x4_t aM2) michael@0: { michael@0: return From32(umax(aM1.i32[0], aM2.i32[0]), umax(aM1.i32[1], aM2.i32[1]), michael@0: umax(aM1.i32[2], aM2.i32[2]), umax(aM1.i32[3], aM2.i32[3])); michael@0: } michael@0: michael@0: inline Scalaru16x8_t Mul16(Scalaru16x8_t aM1, Scalaru16x8_t aM2) michael@0: { michael@0: return FromU16(uint16_t(int32_t(aM1.u16[0]) * int32_t(aM2.u16[0])), uint16_t(int32_t(aM1.u16[1]) * int32_t(aM2.u16[1])), michael@0: uint16_t(int32_t(aM1.u16[2]) * int32_t(aM2.u16[2])), uint16_t(int32_t(aM1.u16[3]) * int32_t(aM2.u16[3])), michael@0: uint16_t(int32_t(aM1.u16[4]) * int32_t(aM2.u16[4])), uint16_t(int32_t(aM1.u16[5]) * int32_t(aM2.u16[5])), michael@0: uint16_t(int32_t(aM1.u16[6]) * int32_t(aM2.u16[6])), uint16_t(int32_t(aM1.u16[7]) * int32_t(aM2.u16[7]))); michael@0: } michael@0: michael@0: inline void Mul16x4x2x2To32x4x2(Scalari16x8_t aFactorsA1B1, michael@0: Scalari16x8_t aFactorsA2B2, michael@0: Scalari32x4_t& aProductA, michael@0: Scalari32x4_t& aProductB) michael@0: { michael@0: aProductA = From32(aFactorsA1B1.i16[0] * aFactorsA2B2.i16[0], michael@0: aFactorsA1B1.i16[1] * aFactorsA2B2.i16[1], michael@0: aFactorsA1B1.i16[2] * aFactorsA2B2.i16[2], michael@0: aFactorsA1B1.i16[3] * aFactorsA2B2.i16[3]); michael@0: aProductB = From32(aFactorsA1B1.i16[4] * aFactorsA2B2.i16[4], michael@0: aFactorsA1B1.i16[5] * aFactorsA2B2.i16[5], michael@0: aFactorsA1B1.i16[6] * aFactorsA2B2.i16[6], michael@0: aFactorsA1B1.i16[7] * aFactorsA2B2.i16[7]); michael@0: } michael@0: michael@0: inline Scalari32x4_t MulAdd16x8x2To32x4(Scalari16x8_t aFactorsA, michael@0: Scalari16x8_t aFactorsB) michael@0: { michael@0: return From32(aFactorsA.i16[0] * aFactorsB.i16[0] + aFactorsA.i16[1] * aFactorsB.i16[1], michael@0: aFactorsA.i16[2] * aFactorsB.i16[2] + aFactorsA.i16[3] * aFactorsB.i16[3], michael@0: aFactorsA.i16[4] * aFactorsB.i16[4] + aFactorsA.i16[5] * aFactorsB.i16[5], michael@0: aFactorsA.i16[6] * aFactorsB.i16[6] + aFactorsA.i16[7] * aFactorsB.i16[7]); michael@0: } michael@0: michael@0: template michael@0: inline void AssertIndex() michael@0: { michael@0: static_assert(aIndex == 0 || aIndex == 1 || aIndex == 2 || aIndex == 3, michael@0: "Invalid splat index"); michael@0: } michael@0: michael@0: template michael@0: inline Scalari32x4_t Splat32(Scalari32x4_t aM) michael@0: { michael@0: AssertIndex(); michael@0: return From32(aM.i32[aIndex], aM.i32[aIndex], michael@0: aM.i32[aIndex], aM.i32[aIndex]); michael@0: } michael@0: michael@0: template michael@0: inline Scalaru8x16_t Splat32On8(Scalaru8x16_t aM) michael@0: { michael@0: AssertIndex(); michael@0: return From8(aM.u8[i*4], aM.u8[i*4+1], aM.u8[i*4+2], aM.u8[i*4+3], michael@0: aM.u8[i*4], aM.u8[i*4+1], aM.u8[i*4+2], aM.u8[i*4+3], michael@0: aM.u8[i*4], aM.u8[i*4+1], aM.u8[i*4+2], aM.u8[i*4+3], michael@0: aM.u8[i*4], aM.u8[i*4+1], aM.u8[i*4+2], aM.u8[i*4+3]); michael@0: } michael@0: michael@0: template michael@0: inline Scalari32x4_t Shuffle32(Scalari32x4_t aM) michael@0: { michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: Scalari32x4_t m = aM; michael@0: m.i32[0] = aM.i32[i3]; michael@0: m.i32[1] = aM.i32[i2]; michael@0: m.i32[2] = aM.i32[i1]; michael@0: m.i32[3] = aM.i32[i0]; michael@0: return m; michael@0: } michael@0: michael@0: template michael@0: inline Scalari16x8_t ShuffleLo16(Scalari16x8_t aM) michael@0: { michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: Scalari16x8_t m = aM; michael@0: m.i16[0] = aM.i16[i3]; michael@0: m.i16[1] = aM.i16[i2]; michael@0: m.i16[2] = aM.i16[i1]; michael@0: m.i16[3] = aM.i16[i0]; michael@0: return m; michael@0: } michael@0: michael@0: template michael@0: inline Scalari16x8_t ShuffleHi16(Scalari16x8_t aM) michael@0: { michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: Scalari16x8_t m = aM; michael@0: m.i16[4 + 0] = aM.i16[4 + i3]; michael@0: m.i16[4 + 1] = aM.i16[4 + i2]; michael@0: m.i16[4 + 2] = aM.i16[4 + i1]; michael@0: m.i16[4 + 3] = aM.i16[4 + i0]; michael@0: return m; michael@0: } michael@0: michael@0: template michael@0: inline Scalaru16x8_t Splat16(Scalaru16x8_t aM) michael@0: { michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: Scalaru16x8_t m; michael@0: int16_t chosenValueLo = aM.u16[aIndexLo]; michael@0: m.u16[0] = chosenValueLo; michael@0: m.u16[1] = chosenValueLo; michael@0: m.u16[2] = chosenValueLo; michael@0: m.u16[3] = chosenValueLo; michael@0: int16_t chosenValueHi = aM.u16[4 + aIndexHi]; michael@0: m.u16[4] = chosenValueHi; michael@0: m.u16[5] = chosenValueHi; michael@0: m.u16[6] = chosenValueHi; michael@0: m.u16[7] = chosenValueHi; michael@0: return m; michael@0: } michael@0: michael@0: inline Scalaru8x16_t michael@0: InterleaveLo8(Scalaru8x16_t m1, Scalaru8x16_t m2) michael@0: { michael@0: return From8(m1.u8[0], m2.u8[0], m1.u8[1], m2.u8[1], michael@0: m1.u8[2], m2.u8[2], m1.u8[3], m2.u8[3], michael@0: m1.u8[4], m2.u8[4], m1.u8[5], m2.u8[5], michael@0: m1.u8[6], m2.u8[6], m1.u8[7], m2.u8[7]); michael@0: } michael@0: michael@0: inline Scalaru8x16_t michael@0: InterleaveHi8(Scalaru8x16_t m1, Scalaru8x16_t m2) michael@0: { michael@0: return From8(m1.u8[8+0], m2.u8[8+0], m1.u8[8+1], m2.u8[8+1], michael@0: m1.u8[8+2], m2.u8[8+2], m1.u8[8+3], m2.u8[8+3], michael@0: m1.u8[8+4], m2.u8[8+4], m1.u8[8+5], m2.u8[8+5], michael@0: m1.u8[8+6], m2.u8[8+6], m1.u8[8+7], m2.u8[8+7]); michael@0: } michael@0: michael@0: inline Scalaru16x8_t michael@0: InterleaveLo16(Scalaru16x8_t m1, Scalaru16x8_t m2) michael@0: { michael@0: return FromU16(m1.u16[0], m2.u16[0], m1.u16[1], m2.u16[1], michael@0: m1.u16[2], m2.u16[2], m1.u16[3], m2.u16[3]); michael@0: } michael@0: michael@0: inline Scalaru16x8_t michael@0: InterleaveHi16(Scalaru16x8_t m1, Scalaru16x8_t m2) michael@0: { michael@0: return FromU16(m1.u16[4], m2.u16[4], m1.u16[5], m2.u16[5], michael@0: m1.u16[6], m2.u16[6], m1.u16[7], m2.u16[7]); michael@0: } michael@0: michael@0: inline Scalari32x4_t michael@0: InterleaveLo32(Scalari32x4_t m1, Scalari32x4_t m2) michael@0: { michael@0: return From32(m1.i32[0], m2.i32[0], m1.i32[1], m2.i32[1]); michael@0: } michael@0: michael@0: inline Scalari16x8_t michael@0: UnpackLo8x8ToI16x8(Scalaru8x16_t aM) michael@0: { michael@0: Scalari16x8_t m; michael@0: m.i16[0] = aM.u8[0]; michael@0: m.i16[1] = aM.u8[1]; michael@0: m.i16[2] = aM.u8[2]; michael@0: m.i16[3] = aM.u8[3]; michael@0: m.i16[4] = aM.u8[4]; michael@0: m.i16[5] = aM.u8[5]; michael@0: m.i16[6] = aM.u8[6]; michael@0: m.i16[7] = aM.u8[7]; michael@0: return m; michael@0: } michael@0: michael@0: inline Scalari16x8_t michael@0: UnpackHi8x8ToI16x8(Scalaru8x16_t aM) michael@0: { michael@0: Scalari16x8_t m; michael@0: m.i16[0] = aM.u8[8+0]; michael@0: m.i16[1] = aM.u8[8+1]; michael@0: m.i16[2] = aM.u8[8+2]; michael@0: m.i16[3] = aM.u8[8+3]; michael@0: m.i16[4] = aM.u8[8+4]; michael@0: m.i16[5] = aM.u8[8+5]; michael@0: m.i16[6] = aM.u8[8+6]; michael@0: m.i16[7] = aM.u8[8+7]; michael@0: return m; michael@0: } michael@0: michael@0: inline Scalaru16x8_t michael@0: UnpackLo8x8ToU16x8(Scalaru8x16_t aM) michael@0: { michael@0: return FromU16(uint16_t(aM.u8[0]), uint16_t(aM.u8[1]), uint16_t(aM.u8[2]), uint16_t(aM.u8[3]), michael@0: uint16_t(aM.u8[4]), uint16_t(aM.u8[5]), uint16_t(aM.u8[6]), uint16_t(aM.u8[7])); michael@0: } michael@0: michael@0: inline Scalaru16x8_t michael@0: UnpackHi8x8ToU16x8(Scalaru8x16_t aM) michael@0: { michael@0: return FromU16(aM.u8[8+0], aM.u8[8+1], aM.u8[8+2], aM.u8[8+3], michael@0: aM.u8[8+4], aM.u8[8+5], aM.u8[8+6], aM.u8[8+7]); michael@0: } michael@0: michael@0: template michael@0: inline Scalaru8x16_t michael@0: Rotate8(Scalaru8x16_t a1234, Scalaru8x16_t a5678) michael@0: { michael@0: Scalaru8x16_t m; michael@0: for (uint8_t i = 0; i < 16; i++) { michael@0: uint8_t sourceByte = i + aNumBytes; michael@0: m.u8[i] = sourceByte < 16 ? a1234.u8[sourceByte] : a5678.u8[sourceByte - 16]; michael@0: } michael@0: return m; michael@0: } michael@0: michael@0: template michael@0: inline int16_t michael@0: SaturateTo16(T a) michael@0: { michael@0: return int16_t(a >= INT16_MIN ? (a <= INT16_MAX ? a : INT16_MAX) : INT16_MIN); michael@0: } michael@0: michael@0: inline Scalari16x8_t michael@0: PackAndSaturate32To16(Scalari32x4_t m1, Scalari32x4_t m2) michael@0: { michael@0: Scalari16x8_t m; michael@0: m.i16[0] = SaturateTo16(m1.i32[0]); michael@0: m.i16[1] = SaturateTo16(m1.i32[1]); michael@0: m.i16[2] = SaturateTo16(m1.i32[2]); michael@0: m.i16[3] = SaturateTo16(m1.i32[3]); michael@0: m.i16[4] = SaturateTo16(m2.i32[0]); michael@0: m.i16[5] = SaturateTo16(m2.i32[1]); michael@0: m.i16[6] = SaturateTo16(m2.i32[2]); michael@0: m.i16[7] = SaturateTo16(m2.i32[3]); michael@0: return m; michael@0: } michael@0: michael@0: template michael@0: inline uint16_t michael@0: SaturateToU16(T a) michael@0: { michael@0: return uint16_t(umin(a & -(a >= 0), INT16_MAX)); michael@0: } michael@0: michael@0: inline Scalaru16x8_t michael@0: PackAndSaturate32ToU16(Scalari32x4_t m1, Scalari32x4_t m2) michael@0: { michael@0: Scalaru16x8_t m; michael@0: m.u16[0] = SaturateToU16(m1.i32[0]); michael@0: m.u16[1] = SaturateToU16(m1.i32[1]); michael@0: m.u16[2] = SaturateToU16(m1.i32[2]); michael@0: m.u16[3] = SaturateToU16(m1.i32[3]); michael@0: m.u16[4] = SaturateToU16(m2.i32[0]); michael@0: m.u16[5] = SaturateToU16(m2.i32[1]); michael@0: m.u16[6] = SaturateToU16(m2.i32[2]); michael@0: m.u16[7] = SaturateToU16(m2.i32[3]); michael@0: return m; michael@0: } michael@0: michael@0: template michael@0: inline uint8_t michael@0: SaturateTo8(T a) michael@0: { michael@0: return uint8_t(umin(a & -(a >= 0), 255)); michael@0: } michael@0: michael@0: inline Scalaru8x16_t michael@0: PackAndSaturate32To8(Scalari32x4_t m1, Scalari32x4_t m2, Scalari32x4_t m3, const Scalari32x4_t& m4) michael@0: { michael@0: Scalaru8x16_t m; michael@0: m.u8[0] = SaturateTo8(m1.i32[0]); michael@0: m.u8[1] = SaturateTo8(m1.i32[1]); michael@0: m.u8[2] = SaturateTo8(m1.i32[2]); michael@0: m.u8[3] = SaturateTo8(m1.i32[3]); michael@0: m.u8[4] = SaturateTo8(m2.i32[0]); michael@0: m.u8[5] = SaturateTo8(m2.i32[1]); michael@0: m.u8[6] = SaturateTo8(m2.i32[2]); michael@0: m.u8[7] = SaturateTo8(m2.i32[3]); michael@0: m.u8[8] = SaturateTo8(m3.i32[0]); michael@0: m.u8[9] = SaturateTo8(m3.i32[1]); michael@0: m.u8[10] = SaturateTo8(m3.i32[2]); michael@0: m.u8[11] = SaturateTo8(m3.i32[3]); michael@0: m.u8[12] = SaturateTo8(m4.i32[0]); michael@0: m.u8[13] = SaturateTo8(m4.i32[1]); michael@0: m.u8[14] = SaturateTo8(m4.i32[2]); michael@0: m.u8[15] = SaturateTo8(m4.i32[3]); michael@0: return m; michael@0: } michael@0: michael@0: inline Scalaru8x16_t michael@0: PackAndSaturate16To8(Scalari16x8_t m1, Scalari16x8_t m2) michael@0: { michael@0: Scalaru8x16_t m; michael@0: m.u8[0] = SaturateTo8(m1.i16[0]); michael@0: m.u8[1] = SaturateTo8(m1.i16[1]); michael@0: m.u8[2] = SaturateTo8(m1.i16[2]); michael@0: m.u8[3] = SaturateTo8(m1.i16[3]); michael@0: m.u8[4] = SaturateTo8(m1.i16[4]); michael@0: m.u8[5] = SaturateTo8(m1.i16[5]); michael@0: m.u8[6] = SaturateTo8(m1.i16[6]); michael@0: m.u8[7] = SaturateTo8(m1.i16[7]); michael@0: m.u8[8] = SaturateTo8(m2.i16[0]); michael@0: m.u8[9] = SaturateTo8(m2.i16[1]); michael@0: m.u8[10] = SaturateTo8(m2.i16[2]); michael@0: m.u8[11] = SaturateTo8(m2.i16[3]); michael@0: m.u8[12] = SaturateTo8(m2.i16[4]); michael@0: m.u8[13] = SaturateTo8(m2.i16[5]); michael@0: m.u8[14] = SaturateTo8(m2.i16[6]); michael@0: m.u8[15] = SaturateTo8(m2.i16[7]); michael@0: return m; michael@0: } michael@0: michael@0: // Fast approximate division by 255. It has the property that michael@0: // for all 0 <= n <= 255*255, FAST_DIVIDE_BY_255(n) == n/255. michael@0: // But it only uses two adds and two shifts instead of an michael@0: // integer division (which is expensive on many processors). michael@0: // michael@0: // equivalent to v/255 michael@0: template michael@0: inline B FastDivideBy255(A v) michael@0: { michael@0: return ((v << 8) + v + 255) >> 16; michael@0: } michael@0: michael@0: inline Scalaru16x8_t michael@0: FastDivideBy255_16(Scalaru16x8_t m) michael@0: { michael@0: return FromU16(FastDivideBy255(int32_t(m.u16[0])), michael@0: FastDivideBy255(int32_t(m.u16[1])), michael@0: FastDivideBy255(int32_t(m.u16[2])), michael@0: FastDivideBy255(int32_t(m.u16[3])), michael@0: FastDivideBy255(int32_t(m.u16[4])), michael@0: FastDivideBy255(int32_t(m.u16[5])), michael@0: FastDivideBy255(int32_t(m.u16[6])), michael@0: FastDivideBy255(int32_t(m.u16[7]))); michael@0: } michael@0: michael@0: inline Scalari32x4_t michael@0: FastDivideBy255(Scalari32x4_t m) michael@0: { michael@0: return From32(FastDivideBy255(m.i32[0]), michael@0: FastDivideBy255(m.i32[1]), michael@0: FastDivideBy255(m.i32[2]), michael@0: FastDivideBy255(m.i32[3])); michael@0: } michael@0: michael@0: inline Scalaru8x16_t michael@0: Pick(Scalaru8x16_t mask, Scalaru8x16_t a, Scalaru8x16_t b) michael@0: { michael@0: return From8((a.u8[0] & (~mask.u8[0])) | (b.u8[0] & mask.u8[0]), michael@0: (a.u8[1] & (~mask.u8[1])) | (b.u8[1] & mask.u8[1]), michael@0: (a.u8[2] & (~mask.u8[2])) | (b.u8[2] & mask.u8[2]), michael@0: (a.u8[3] & (~mask.u8[3])) | (b.u8[3] & mask.u8[3]), michael@0: (a.u8[4] & (~mask.u8[4])) | (b.u8[4] & mask.u8[4]), michael@0: (a.u8[5] & (~mask.u8[5])) | (b.u8[5] & mask.u8[5]), michael@0: (a.u8[6] & (~mask.u8[6])) | (b.u8[6] & mask.u8[6]), michael@0: (a.u8[7] & (~mask.u8[7])) | (b.u8[7] & mask.u8[7]), michael@0: (a.u8[8+0] & (~mask.u8[8+0])) | (b.u8[8+0] & mask.u8[8+0]), michael@0: (a.u8[8+1] & (~mask.u8[8+1])) | (b.u8[8+1] & mask.u8[8+1]), michael@0: (a.u8[8+2] & (~mask.u8[8+2])) | (b.u8[8+2] & mask.u8[8+2]), michael@0: (a.u8[8+3] & (~mask.u8[8+3])) | (b.u8[8+3] & mask.u8[8+3]), michael@0: (a.u8[8+4] & (~mask.u8[8+4])) | (b.u8[8+4] & mask.u8[8+4]), michael@0: (a.u8[8+5] & (~mask.u8[8+5])) | (b.u8[8+5] & mask.u8[8+5]), michael@0: (a.u8[8+6] & (~mask.u8[8+6])) | (b.u8[8+6] & mask.u8[8+6]), michael@0: (a.u8[8+7] & (~mask.u8[8+7])) | (b.u8[8+7] & mask.u8[8+7])); michael@0: } michael@0: michael@0: inline Scalari32x4_t michael@0: Pick(Scalari32x4_t mask, Scalari32x4_t a, Scalari32x4_t b) michael@0: { michael@0: return From32((a.i32[0] & (~mask.i32[0])) | (b.i32[0] & mask.i32[0]), michael@0: (a.i32[1] & (~mask.i32[1])) | (b.i32[1] & mask.i32[1]), michael@0: (a.i32[2] & (~mask.i32[2])) | (b.i32[2] & mask.i32[2]), michael@0: (a.i32[3] & (~mask.i32[3])) | (b.i32[3] & mask.i32[3])); michael@0: } michael@0: michael@0: inline Scalarf32x4_t MixF32(Scalarf32x4_t a, Scalarf32x4_t b, float t) michael@0: { michael@0: return FromF32(a.f32[0] + (b.f32[0] - a.f32[0]) * t, michael@0: a.f32[1] + (b.f32[1] - a.f32[1]) * t, michael@0: a.f32[2] + (b.f32[2] - a.f32[2]) * t, michael@0: a.f32[3] + (b.f32[3] - a.f32[3]) * t); michael@0: } michael@0: michael@0: inline Scalarf32x4_t WSumF32(Scalarf32x4_t a, Scalarf32x4_t b, float wa, float wb) michael@0: { michael@0: return FromF32(a.f32[0] * wa + b.f32[0] * wb, michael@0: a.f32[1] * wa + b.f32[1] * wb, michael@0: a.f32[2] * wa + b.f32[2] * wb, michael@0: a.f32[3] * wa + b.f32[3] * wb); michael@0: } michael@0: michael@0: inline Scalarf32x4_t AbsF32(Scalarf32x4_t a) michael@0: { michael@0: return FromF32(fabs(a.f32[0]), michael@0: fabs(a.f32[1]), michael@0: fabs(a.f32[2]), michael@0: fabs(a.f32[3])); michael@0: } michael@0: michael@0: inline Scalarf32x4_t AddF32(Scalarf32x4_t a, Scalarf32x4_t b) michael@0: { michael@0: return FromF32(a.f32[0] + b.f32[0], michael@0: a.f32[1] + b.f32[1], michael@0: a.f32[2] + b.f32[2], michael@0: a.f32[3] + b.f32[3]); michael@0: } michael@0: michael@0: inline Scalarf32x4_t MulF32(Scalarf32x4_t a, Scalarf32x4_t b) michael@0: { michael@0: return FromF32(a.f32[0] * b.f32[0], michael@0: a.f32[1] * b.f32[1], michael@0: a.f32[2] * b.f32[2], michael@0: a.f32[3] * b.f32[3]); michael@0: } michael@0: michael@0: inline Scalarf32x4_t DivF32(Scalarf32x4_t a, Scalarf32x4_t b) michael@0: { michael@0: return FromF32(a.f32[0] / b.f32[0], michael@0: a.f32[1] / b.f32[1], michael@0: a.f32[2] / b.f32[2], michael@0: a.f32[3] / b.f32[3]); michael@0: } michael@0: michael@0: template michael@0: inline Scalarf32x4_t SplatF32(Scalarf32x4_t m) michael@0: { michael@0: AssertIndex(); michael@0: return FromF32(m.f32[aIndex], michael@0: m.f32[aIndex], michael@0: m.f32[aIndex], michael@0: m.f32[aIndex]); michael@0: } michael@0: michael@0: inline Scalari32x4_t F32ToI32(Scalarf32x4_t m) michael@0: { michael@0: return From32(int32_t(floor(m.f32[0] + 0.5f)), michael@0: int32_t(floor(m.f32[1] + 0.5f)), michael@0: int32_t(floor(m.f32[2] + 0.5f)), michael@0: int32_t(floor(m.f32[3] + 0.5f))); michael@0: } michael@0: michael@0: #ifdef SIMD_COMPILE_SSE2 michael@0: michael@0: // SSE2 michael@0: michael@0: template<> michael@0: inline __m128i michael@0: Load8<__m128i>(const uint8_t* aSource) michael@0: { michael@0: return _mm_load_si128((const __m128i*)aSource); michael@0: } michael@0: michael@0: inline void Store8(uint8_t* aTarget, __m128i aM) michael@0: { michael@0: _mm_store_si128((__m128i*)aTarget, aM); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128i FromZero8<__m128i>() michael@0: { michael@0: return _mm_setzero_si128(); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128i From8<__m128i>(uint8_t a, uint8_t b, uint8_t c, uint8_t d, uint8_t e, uint8_t f, uint8_t g, uint8_t h, michael@0: uint8_t i, uint8_t j, uint8_t k, uint8_t l, uint8_t m, uint8_t n, uint8_t o, uint8_t p) michael@0: { michael@0: return _mm_setr_epi16((b << 8) + a, (d << 8) + c, (e << 8) + f, (h << 8) + g, michael@0: (j << 8) + i, (l << 8) + k, (m << 8) + n, (p << 8) + o); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128i FromI16<__m128i>(int16_t a, int16_t b, int16_t c, int16_t d, int16_t e, int16_t f, int16_t g, int16_t h) michael@0: { michael@0: return _mm_setr_epi16(a, b, c, d, e, f, g, h); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128i FromU16<__m128i>(uint16_t a, uint16_t b, uint16_t c, uint16_t d, uint16_t e, uint16_t f, uint16_t g, uint16_t h) michael@0: { michael@0: return _mm_setr_epi16(a, b, c, d, e, f, g, h); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128i FromI16<__m128i>(int16_t a) michael@0: { michael@0: return _mm_set1_epi16(a); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128i FromU16<__m128i>(uint16_t a) michael@0: { michael@0: return _mm_set1_epi16((int16_t)a); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128i From32<__m128i>(int32_t a, int32_t b, int32_t c, int32_t d) michael@0: { michael@0: return _mm_setr_epi32(a, b, c, d); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128i From32<__m128i>(int32_t a) michael@0: { michael@0: return _mm_set1_epi32(a); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128 FromF32<__m128>(float a, float b, float c, float d) michael@0: { michael@0: return _mm_setr_ps(a, b, c, d); michael@0: } michael@0: michael@0: template<> michael@0: inline __m128 FromF32<__m128>(float a) michael@0: { michael@0: return _mm_set1_ps(a); michael@0: } michael@0: michael@0: template michael@0: inline __m128i ShiftRight16(__m128i aM) michael@0: { michael@0: return _mm_srli_epi16(aM, aNumberOfBits); michael@0: } michael@0: michael@0: template michael@0: inline __m128i ShiftRight32(__m128i aM) michael@0: { michael@0: return _mm_srai_epi32(aM, aNumberOfBits); michael@0: } michael@0: michael@0: inline __m128i Add16(__m128i aM1, __m128i aM2) michael@0: { michael@0: return _mm_add_epi16(aM1, aM2); michael@0: } michael@0: michael@0: inline __m128i Add32(__m128i aM1, __m128i aM2) michael@0: { michael@0: return _mm_add_epi32(aM1, aM2); michael@0: } michael@0: michael@0: inline __m128i Sub16(__m128i aM1, __m128i aM2) michael@0: { michael@0: return _mm_sub_epi16(aM1, aM2); michael@0: } michael@0: michael@0: inline __m128i Sub32(__m128i aM1, __m128i aM2) michael@0: { michael@0: return _mm_sub_epi32(aM1, aM2); michael@0: } michael@0: michael@0: inline __m128i Min8(__m128i aM1, __m128i aM2) michael@0: { michael@0: return _mm_min_epu8(aM1, aM2); michael@0: } michael@0: michael@0: inline __m128i Max8(__m128i aM1, __m128i aM2) michael@0: { michael@0: return _mm_max_epu8(aM1, aM2); michael@0: } michael@0: michael@0: inline __m128i Min32(__m128i aM1, __m128i aM2) michael@0: { michael@0: __m128i m1_minus_m2 = _mm_sub_epi32(aM1, aM2); michael@0: __m128i m1_greater_than_m2 = _mm_cmpgt_epi32(aM1, aM2); michael@0: return _mm_sub_epi32(aM1, _mm_and_si128(m1_minus_m2, m1_greater_than_m2)); michael@0: } michael@0: michael@0: inline __m128i Max32(__m128i aM1, __m128i aM2) michael@0: { michael@0: __m128i m1_minus_m2 = _mm_sub_epi32(aM1, aM2); michael@0: __m128i m2_greater_than_m1 = _mm_cmpgt_epi32(aM2, aM1); michael@0: return _mm_sub_epi32(aM1, _mm_and_si128(m1_minus_m2, m2_greater_than_m1)); michael@0: } michael@0: michael@0: inline __m128i Mul16(__m128i aM1, __m128i aM2) michael@0: { michael@0: return _mm_mullo_epi16(aM1, aM2); michael@0: } michael@0: michael@0: inline __m128i MulU16(__m128i aM1, __m128i aM2) michael@0: { michael@0: return _mm_mullo_epi16(aM1, aM2); michael@0: } michael@0: michael@0: inline void Mul16x4x2x2To32x4x2(__m128i aFactorsA1B1, michael@0: __m128i aFactorsA2B2, michael@0: __m128i& aProductA, michael@0: __m128i& aProductB) michael@0: { michael@0: __m128i prodAB_lo = _mm_mullo_epi16(aFactorsA1B1, aFactorsA2B2); michael@0: __m128i prodAB_hi = _mm_mulhi_epi16(aFactorsA1B1, aFactorsA2B2); michael@0: aProductA = _mm_unpacklo_epi16(prodAB_lo, prodAB_hi); michael@0: aProductB = _mm_unpackhi_epi16(prodAB_lo, prodAB_hi); michael@0: } michael@0: michael@0: inline __m128i MulAdd16x8x2To32x4(__m128i aFactorsA, michael@0: __m128i aFactorsB) michael@0: { michael@0: return _mm_madd_epi16(aFactorsA, aFactorsB); michael@0: } michael@0: michael@0: template michael@0: inline __m128i Shuffle32(__m128i aM) michael@0: { michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: return _mm_shuffle_epi32(aM, _MM_SHUFFLE(i0, i1, i2, i3)); michael@0: } michael@0: michael@0: template michael@0: inline __m128i ShuffleLo16(__m128i aM) michael@0: { michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: return _mm_shufflelo_epi16(aM, _MM_SHUFFLE(i0, i1, i2, i3)); michael@0: } michael@0: michael@0: template michael@0: inline __m128i ShuffleHi16(__m128i aM) michael@0: { michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: return _mm_shufflehi_epi16(aM, _MM_SHUFFLE(i0, i1, i2, i3)); michael@0: } michael@0: michael@0: template michael@0: inline __m128i Splat32(__m128i aM) michael@0: { michael@0: return Shuffle32(aM); michael@0: } michael@0: michael@0: template michael@0: inline __m128i Splat32On8(__m128i aM) michael@0: { michael@0: return Shuffle32(aM); michael@0: } michael@0: michael@0: template michael@0: inline __m128i Splat16(__m128i aM) michael@0: { michael@0: AssertIndex(); michael@0: AssertIndex(); michael@0: return ShuffleHi16( michael@0: ShuffleLo16(aM)); michael@0: } michael@0: michael@0: inline __m128i michael@0: UnpackLo8x8ToI16x8(__m128i m) michael@0: { michael@0: __m128i zero = _mm_set1_epi8(0); michael@0: return _mm_unpacklo_epi8(m, zero); michael@0: } michael@0: michael@0: inline __m128i michael@0: UnpackHi8x8ToI16x8(__m128i m) michael@0: { michael@0: __m128i zero = _mm_set1_epi8(0); michael@0: return _mm_unpackhi_epi8(m, zero); michael@0: } michael@0: michael@0: inline __m128i michael@0: UnpackLo8x8ToU16x8(__m128i m) michael@0: { michael@0: __m128i zero = _mm_set1_epi8(0); michael@0: return _mm_unpacklo_epi8(m, zero); michael@0: } michael@0: michael@0: inline __m128i michael@0: UnpackHi8x8ToU16x8(__m128i m) michael@0: { michael@0: __m128i zero = _mm_set1_epi8(0); michael@0: return _mm_unpackhi_epi8(m, zero); michael@0: } michael@0: michael@0: inline __m128i michael@0: InterleaveLo8(__m128i m1, __m128i m2) michael@0: { michael@0: return _mm_unpacklo_epi8(m1, m2); michael@0: } michael@0: michael@0: inline __m128i michael@0: InterleaveHi8(__m128i m1, __m128i m2) michael@0: { michael@0: return _mm_unpackhi_epi8(m1, m2); michael@0: } michael@0: michael@0: inline __m128i michael@0: InterleaveLo16(__m128i m1, __m128i m2) michael@0: { michael@0: return _mm_unpacklo_epi16(m1, m2); michael@0: } michael@0: michael@0: inline __m128i michael@0: InterleaveHi16(__m128i m1, __m128i m2) michael@0: { michael@0: return _mm_unpackhi_epi16(m1, m2); michael@0: } michael@0: michael@0: inline __m128i michael@0: InterleaveLo32(__m128i m1, __m128i m2) michael@0: { michael@0: return _mm_unpacklo_epi32(m1, m2); michael@0: } michael@0: michael@0: template michael@0: inline __m128i michael@0: Rotate8(__m128i a1234, __m128i a5678) michael@0: { michael@0: return _mm_or_si128(_mm_srli_si128(a1234, aNumBytes), _mm_slli_si128(a5678, 16 - aNumBytes)); michael@0: } michael@0: michael@0: inline __m128i michael@0: PackAndSaturate32To16(__m128i m1, __m128i m2) michael@0: { michael@0: return _mm_packs_epi32(m1, m2); michael@0: } michael@0: michael@0: inline __m128i michael@0: PackAndSaturate32ToU16(__m128i m1, __m128i m2) michael@0: { michael@0: return _mm_packs_epi32(m1, m2); michael@0: } michael@0: michael@0: inline __m128i michael@0: PackAndSaturate32To8(__m128i m1, __m128i m2, __m128i m3, const __m128i& m4) michael@0: { michael@0: // Pack into 8 16bit signed integers (saturating). michael@0: __m128i m12 = _mm_packs_epi32(m1, m2); michael@0: __m128i m34 = _mm_packs_epi32(m3, m4); michael@0: michael@0: // Pack into 16 8bit unsigned integers (saturating). michael@0: return _mm_packus_epi16(m12, m34); michael@0: } michael@0: michael@0: inline __m128i michael@0: PackAndSaturate16To8(__m128i m1, __m128i m2) michael@0: { michael@0: // Pack into 16 8bit unsigned integers (saturating). michael@0: return _mm_packus_epi16(m1, m2); michael@0: } michael@0: michael@0: inline __m128i michael@0: FastDivideBy255(__m128i m) michael@0: { michael@0: // v = m << 8 michael@0: __m128i v = _mm_slli_epi32(m, 8); michael@0: // v = v + (m + (255,255,255,255)) michael@0: v = _mm_add_epi32(v, _mm_add_epi32(m, _mm_set1_epi32(255))); michael@0: // v = v >> 16 michael@0: return _mm_srai_epi32(v, 16); michael@0: } michael@0: michael@0: inline __m128i michael@0: FastDivideBy255_16(__m128i m) michael@0: { michael@0: __m128i zero = _mm_set1_epi16(0); michael@0: __m128i lo = _mm_unpacklo_epi16(m, zero); michael@0: __m128i hi = _mm_unpackhi_epi16(m, zero); michael@0: return _mm_packs_epi32(FastDivideBy255(lo), FastDivideBy255(hi)); michael@0: } michael@0: michael@0: inline __m128i michael@0: Pick(__m128i mask, __m128i a, __m128i b) michael@0: { michael@0: return _mm_or_si128(_mm_andnot_si128(mask, a), _mm_and_si128(mask, b)); michael@0: } michael@0: michael@0: inline __m128 MixF32(__m128 a, __m128 b, float t) michael@0: { michael@0: return _mm_add_ps(a, _mm_mul_ps(_mm_sub_ps(b, a), _mm_set1_ps(t))); michael@0: } michael@0: michael@0: inline __m128 WSumF32(__m128 a, __m128 b, float wa, float wb) michael@0: { michael@0: return _mm_add_ps(_mm_mul_ps(a, _mm_set1_ps(wa)), _mm_mul_ps(b, _mm_set1_ps(wb))); michael@0: } michael@0: michael@0: inline __m128 AbsF32(__m128 a) michael@0: { michael@0: return _mm_max_ps(_mm_sub_ps(_mm_setzero_ps(), a), a); michael@0: } michael@0: michael@0: inline __m128 AddF32(__m128 a, __m128 b) michael@0: { michael@0: return _mm_add_ps(a, b); michael@0: } michael@0: michael@0: inline __m128 MulF32(__m128 a, __m128 b) michael@0: { michael@0: return _mm_mul_ps(a, b); michael@0: } michael@0: michael@0: inline __m128 DivF32(__m128 a, __m128 b) michael@0: { michael@0: return _mm_div_ps(a, b); michael@0: } michael@0: michael@0: template michael@0: inline __m128 SplatF32(__m128 m) michael@0: { michael@0: AssertIndex(); michael@0: return _mm_shuffle_ps(m, m, _MM_SHUFFLE(aIndex, aIndex, aIndex, aIndex)); michael@0: } michael@0: michael@0: inline __m128i F32ToI32(__m128 m) michael@0: { michael@0: return _mm_cvtps_epi32(m); michael@0: } michael@0: michael@0: #endif // SIMD_COMPILE_SSE2 michael@0: michael@0: } // namespace simd michael@0: michael@0: } // namespace gfx michael@0: } // namespace mozilla michael@0: michael@0: #endif // _MOZILLA_GFX_SIMD_H_