// Copyright (c) the JPEG XL Project Authors. All rights reserved. // // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. // Fast SIMD math ops (log2, encoder only, cos, erf for splines) #if defined(LIB_JXL_BASE_FAST_MATH_INL_H_) == defined(HWY_TARGET_TOGGLE) #ifdef LIB_JXL_BASE_FAST_MATH_INL_H_ #undef LIB_JXL_BASE_FAST_MATH_INL_H_ #else #define LIB_JXL_BASE_FAST_MATH_INL_H_ #endif #include #include "lib/jxl/base/common.h" #include "lib/jxl/base/rational_polynomial-inl.h" HWY_BEFORE_NAMESPACE(); namespace jxl { namespace HWY_NAMESPACE { // These templates are not found via ADL. using hwy::HWY_NAMESPACE::Abs; using hwy::HWY_NAMESPACE::Add; using hwy::HWY_NAMESPACE::Eq; using hwy::HWY_NAMESPACE::Floor; using hwy::HWY_NAMESPACE::Ge; using hwy::HWY_NAMESPACE::GetLane; using hwy::HWY_NAMESPACE::IfThenElse; using hwy::HWY_NAMESPACE::IfThenZeroElse; using hwy::HWY_NAMESPACE::Le; using hwy::HWY_NAMESPACE::Min; using hwy::HWY_NAMESPACE::Mul; using hwy::HWY_NAMESPACE::MulAdd; using hwy::HWY_NAMESPACE::NegMulAdd; using hwy::HWY_NAMESPACE::Rebind; using hwy::HWY_NAMESPACE::ShiftLeft; using hwy::HWY_NAMESPACE::ShiftRight; using hwy::HWY_NAMESPACE::Sub; using hwy::HWY_NAMESPACE::Xor; // Computes base-2 logarithm like std::log2. Undefined if negative / NaN. // L1 error ~3.9E-6 template V FastLog2f(const DF df, V x) { // 2,2 rational polynomial approximation of std::log1p(x) / std::log(2). HWY_ALIGN const float p[4 * (2 + 1)] = {HWY_REP4(-1.8503833400518310E-06f), HWY_REP4(1.4287160470083755E+00f), HWY_REP4(7.4245873327820566E-01f)}; HWY_ALIGN const float q[4 * (2 + 1)] = {HWY_REP4(9.9032814277590719E-01f), HWY_REP4(1.0096718572241148E+00f), HWY_REP4(1.7409343003366853E-01f)}; const Rebind di; const auto x_bits = BitCast(di, x); // Range reduction to [-1/3, 1/3] - 3 integer, 2 float ops const auto exp_bits = Sub(x_bits, Set(di, 0x3f2aaaab)); // = 2/3 // Shifted exponent = log2; also used to clear mantissa. const auto exp_shifted = ShiftRight<23>(exp_bits); const auto mantissa = BitCast(df, Sub(x_bits, ShiftLeft<23>(exp_shifted))); const auto exp_val = ConvertTo(df, exp_shifted); return Add(EvalRationalPolynomial(df, Sub(mantissa, Set(df, 1.0f)), p, q), exp_val); } // max relative error ~3e-7 template V FastPow2f(const DF df, V x) { const Rebind di; auto floorx = Floor(x); auto exp = BitCast(df, ShiftLeft<23>(Add(ConvertTo(di, floorx), Set(di, 127)))); auto frac = Sub(x, floorx); auto num = Add(frac, Set(df, 1.01749063e+01)); num = MulAdd(num, frac, Set(df, 4.88687798e+01)); num = MulAdd(num, frac, Set(df, 9.85506591e+01)); num = Mul(num, exp); auto den = MulAdd(frac, Set(df, 2.10242958e-01), Set(df, -2.22328856e-02)); den = MulAdd(den, frac, Set(df, -1.94414990e+01)); den = MulAdd(den, frac, Set(df, 9.85506633e+01)); return Div(num, den); } // max relative error ~3e-5 template V FastPowf(const DF df, V base, V exponent) { return FastPow2f(df, Mul(FastLog2f(df, base), exponent)); } // Computes cosine like std::cos. // L1 error 7e-5. template V FastCosf(const DF df, V x) { // Step 1: range reduction to [0, 2pi) const auto pi2 = Set(df, kPi * 2.0f); const auto pi2_inv = Set(df, 0.5f / kPi); const auto npi2 = Mul(Floor(Mul(x, pi2_inv)), pi2); const auto xmodpi2 = Sub(x, npi2); // Step 2: range reduction to [0, pi] const auto x_pi = Min(xmodpi2, Sub(pi2, xmodpi2)); // Step 3: range reduction to [0, pi/2] const auto above_pihalf = Ge(x_pi, Set(df, kPi / 2.0f)); const auto x_pihalf = IfThenElse(above_pihalf, Sub(Set(df, kPi), x_pi), x_pi); // Step 4: Taylor-like approximation, scaled by 2**0.75 to make angle // duplication steps faster, on x/4. const auto xs = Mul(x_pihalf, Set(df, 0.25f)); const auto x2 = Mul(xs, xs); const auto x4 = Mul(x2, x2); const auto cosx_prescaling = MulAdd(x4, Set(df, 0.06960438), MulAdd(x2, Set(df, -0.84087373), Set(df, 1.68179268))); // Step 5: angle duplication. const auto cosx_scale1 = MulAdd(cosx_prescaling, cosx_prescaling, Set(df, -1.414213562)); const auto cosx_scale2 = MulAdd(cosx_scale1, cosx_scale1, Set(df, -1)); // Step 6: change sign if needed. const Rebind du; auto signbit = ShiftLeft<31>(BitCast(du, VecFromMask(df, above_pihalf))); return BitCast(df, Xor(signbit, BitCast(du, cosx_scale2))); } // Computes the error function like std::erf. // L1 error 7e-4. template V FastErff(const DF df, V x) { // Formula from // https://en.wikipedia.org/wiki/Error_function#Numerical_approximations // but constants have been recomputed. const auto xle0 = Le(x, Zero(df)); const auto absx = Abs(x); // Compute 1 - 1 / ((((x * a + b) * x + c) * x + d) * x + 1)**4 const auto denom1 = MulAdd(absx, Set(df, 7.77394369e-02), Set(df, 2.05260015e-04)); const auto denom2 = MulAdd(denom1, absx, Set(df, 2.32120216e-01)); const auto denom3 = MulAdd(denom2, absx, Set(df, 2.77820801e-01)); const auto denom4 = MulAdd(denom3, absx, Set(df, 1.0f)); const auto denom5 = Mul(denom4, denom4); const auto inv_denom5 = Div(Set(df, 1.0f), denom5); const auto result = NegMulAdd(inv_denom5, inv_denom5, Set(df, 1.0f)); // Change sign if needed. const Rebind du; auto signbit = ShiftLeft<31>(BitCast(du, VecFromMask(df, xle0))); return BitCast(df, Xor(signbit, BitCast(du, result))); } inline float FastLog2f(float f) { HWY_CAPPED(float, 1) D; return GetLane(FastLog2f(D, Set(D, f))); } inline float FastPow2f(float f) { HWY_CAPPED(float, 1) D; return GetLane(FastPow2f(D, Set(D, f))); } inline float FastPowf(float b, float e) { HWY_CAPPED(float, 1) D; return GetLane(FastPowf(D, Set(D, b), Set(D, e))); } inline float FastCosf(float f) { HWY_CAPPED(float, 1) D; return GetLane(FastCosf(D, Set(D, f))); } inline float FastErff(float f) { HWY_CAPPED(float, 1) D; return GetLane(FastErff(D, Set(D, f))); } // Returns cbrt(x) + add with 6 ulp max error. // Modified from vectormath_exp.h, Apache 2 license. // https://www.agner.org/optimize/vectorclass.zip template V CubeRootAndAdd(const V x, const V add) { const HWY_FULL(float) df; const HWY_FULL(int32_t) di; const auto kExpBias = Set(di, 0x54800000); // cast(1.) + cast(1.) / 3 const auto kExpMul = Set(di, 0x002AAAAA); // shifted 1/3 const auto k1_3 = Set(df, 1.0f / 3); const auto k4_3 = Set(df, 4.0f / 3); const auto xa = x; // assume inputs never negative const auto xa_3 = Mul(k1_3, xa); // Multiply exponent by -1/3 const auto m1 = BitCast(di, xa); // Special case for 0. 0 is represented with an exponent of 0, so the // "kExpBias - 1/3 * exp" below gives the wrong result. The IfThenZeroElse() // sets those values as 0, which prevents having NaNs in the computations // below. // TODO(eustas): use fused op const auto m2 = IfThenZeroElse( Eq(m1, Zero(di)), Sub(kExpBias, Mul((ShiftRight<23>(m1)), kExpMul))); auto r = BitCast(df, m2); // Newton-Raphson iterations for (int i = 0; i < 3; i++) { const auto r2 = Mul(r, r); r = NegMulAdd(xa_3, Mul(r2, r2), Mul(k4_3, r)); } // Final iteration auto r2 = Mul(r, r); r = MulAdd(k1_3, NegMulAdd(xa, Mul(r2, r2), r), r); r2 = Mul(r, r); r = MulAdd(r2, x, add); return r; } // NOLINTNEXTLINE(google-readability-namespace-comments) } // namespace HWY_NAMESPACE } // namespace jxl HWY_AFTER_NAMESPACE(); #endif // LIB_JXL_BASE_FAST_MATH_INL_H_ #if HWY_ONCE #ifndef LIB_JXL_BASE_FAST_MATH_ONCE #define LIB_JXL_BASE_FAST_MATH_ONCE namespace jxl { inline float FastLog2f(float f) { return HWY_STATIC_DISPATCH(FastLog2f)(f); } inline float FastPow2f(float f) { return HWY_STATIC_DISPATCH(FastPow2f)(f); } inline float FastPowf(float b, float e) { return HWY_STATIC_DISPATCH(FastPowf)(b, e); } inline float FastCosf(float f) { return HWY_STATIC_DISPATCH(FastCosf)(f); } inline float FastErff(float f) { return HWY_STATIC_DISPATCH(FastErff)(f); } } // namespace jxl #endif // LIB_JXL_BASE_FAST_MATH_ONCE #endif // HWY_ONCE