//===-- Implementation of sqrtf128 function -------------------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// #include "src/math/sqrtf128.h" #include "src/__support/CPP/bit.h" #include "src/__support/FPUtil/FEnvImpl.h" #include "src/__support/FPUtil/FPBits.h" #include "src/__support/FPUtil/rounding_mode.h" #include "src/__support/common.h" #include "src/__support/macros/optimization.h" #include "src/__support/uint128.h" // Compute sqrtf128 with correct rounding for all rounding modes using integer // arithmetic by Alexei Sibidanov (sibid@uvic.ca): // https://github.com/sibidanov/llvm-project/tree/as_sqrt_v2 // https://github.com/sibidanov/llvm-project/tree/as_sqrt_v3 // TODO: Update the reference once Alexei's implementation is in the CORE-MATH // project. https://github.com/llvm/llvm-project/issues/126794 // Let the input be expressed as x = 2^e * m_x, // - Step 1: Range reduction // Let x_reduced = 2^(e % 2) * m_x, // Then sqrt(x) = 2^(e / 2) * sqrt(x_reduced), with // 1 <= x_reduced < 4. // - Step 2: Polynomial approximation // Approximate 1/sqrt(x_reduced) using polynomial approximation with the // result errors bounded by: // |r0 - 1/sqrt(x_reduced)| < 2^-32. // The computations are done in uint64_t. // - Step 3: First Newton iteration // Let the scaled error defined by: // h0 = r0^2 * x_reduced - 1. // Then we compute the first Newton iteration: // r1 = r0 - r0 * h0 / 2. // The result is then bounded by: // |r1 - 1 / sqrt(x_reduced)| < 2^-62. // - Step 4: Second Newton iteration // We calculate the scaled error from Step 3: // h1 = r1^2 * x_reduced - 1. // Then the second Newton iteration is computed by: // r2 = x_reduced * (r1 - r1 * h0 / 2) // ~ x_reduced * (1/sqrt(x_reduced)) = sqrt(x_reduced) // - Step 5: Perform rounding test and correction if needed. // Rounding correction is done by computing the exact rounding errors: // x_reduced - r2^2. namespace LIBC_NAMESPACE_DECL { using FPBits = fputil::FPBits; namespace { template static inline constexpr T prod_hi(T, U); // Get high part of integer multiplications. // Use template to prevent implicit conversion. template <> inline constexpr uint64_t prod_hi(uint64_t x, uint64_t y) { return static_cast( (static_cast(x) * static_cast(y)) >> 64); } // Get high part of unsigned 128x64 bit multiplication. template <> inline constexpr UInt128 prod_hi(UInt128 x, uint64_t y) { uint64_t x_lo = static_cast(x); uint64_t x_hi = static_cast(x >> 64); UInt128 xyl = static_cast(x_lo) * static_cast(y); UInt128 xyh = static_cast(x_hi) * static_cast(y); return xyh + (xyl >> 64); } // Get high part of signed 64x64 bit multiplication. template <> inline constexpr int64_t prod_hi(int64_t x, int64_t y) { return static_cast( (static_cast(x) * static_cast(y)) >> 64); } // Get high 128-bit part of unsigned 128x128 bit multiplication. template <> inline constexpr UInt128 prod_hi(UInt128 x, UInt128 y) { uint64_t x_lo = static_cast(x); uint64_t x_hi = static_cast(x >> 64); uint64_t y_lo = static_cast(y); uint64_t y_hi = static_cast(y >> 64); UInt128 xh_yh = static_cast(x_hi) * static_cast(y_hi); UInt128 xh_yl = static_cast(x_hi) * static_cast(y_lo); UInt128 xl_yh = static_cast(x_lo) * static_cast(y_hi); xh_yh += xh_yl >> 64; return xh_yh + (xl_yh >> 64); } // Get high 128-bit part of mixed sign 128x128 bit multiplication. template <> inline constexpr Int128 prod_hi(Int128 x, UInt128 y) { UInt128 mask = static_cast(x >> 127); UInt128 negative_part = y & mask; UInt128 prod = prod_hi(static_cast(x), y); return static_cast(prod - negative_part); } // Newton-Raphson first order step to improve accuracy of the result. // For the initial approximation r0 ~ 1/sqrt(x), let // h = r0^2 * x - 1 // be its scaled error. Then the first-order Newton-Raphson iteration is: // r1 = r0 - r0 * h / 2 // which has error bounded by: // |r1 - 1/sqrt(x)| < h^2 / 2. LIBC_INLINE uint64_t rsqrt_newton_raphson(uint64_t m, uint64_t r) { uint64_t r2 = prod_hi(r, r); // h = r0^2*x - 1. int64_t h = static_cast(prod_hi(m, r2) + r2); // hr = r * h / 2 int64_t hr = prod_hi(h, static_cast(r >> 1)); return r - hr; } #ifdef LIBC_MATH_HAS_SMALL_TABLES // Degree-12 minimax polynomials for 1/sqrt(x) on [1, 2]. constexpr uint32_t RSQRT_COEFFS[12] = { 0xb5947a4a, 0x2d651e32, 0x9ad50532, 0x2d28d093, 0x0d8be653, 0x04239014, 0x01492449, 0x0066ff7d, 0x001e74a1, 0x000984cc, 0x00049abc, 0x00018340, }; LIBC_INLINE uint64_t rsqrt_approx(uint64_t m) { int64_t x = static_cast(m) ^ (uint64_t(1) << 63); int64_t x_26 = x >> 2; int64_t z = x >> 31; if (LIBC_UNLIKELY(z <= -4294967296)) return ~(m >> 1); uint64_t x2 = static_cast(z) * static_cast(z); uint64_t x2_26 = x2 >> 5; x2 >>= 32; // Calculate the odd part of the polynomial using Horner's method. uint64_t c0 = RSQRT_COEFFS[8] + ((x2 * RSQRT_COEFFS[10]) >> 32); uint64_t c1 = RSQRT_COEFFS[6] + ((x2 * c0) >> 32); uint64_t c2 = RSQRT_COEFFS[4] + ((x2 * c1) >> 32); uint64_t c3 = RSQRT_COEFFS[2] + ((x2 * c2) >> 32); uint64_t c4 = RSQRT_COEFFS[0] + ((x2 * c3) >> 32); uint64_t odd = static_cast((x >> 34) * static_cast(c4 >> 3)) + x_26; // Calculate the even part of the polynomial using Horner's method. uint64_t d0 = RSQRT_COEFFS[9] + ((x2 * RSQRT_COEFFS[11]) >> 32); uint64_t d1 = RSQRT_COEFFS[7] + ((x2 * d0) >> 32); uint64_t d2 = RSQRT_COEFFS[5] + ((x2 * d1) >> 32); uint64_t d3 = RSQRT_COEFFS[3] + ((x2 * d2) >> 32); uint64_t d4 = RSQRT_COEFFS[1] + ((x2 * d3) >> 32); uint64_t even = 0xd105eb806655d608ul + ((x2 * d4) >> 6) + x2_26; uint64_t r = even - odd; // error < 1.5e-10 // Newton-Raphson first order step to improve accuracy of the result to almost // 64 bits. return rsqrt_newton_raphson(m, r); } #else // Cubic minimax polynomials for 1/sqrt(x) on [1 + k/64, 1 + (k + 1)/64] // for k = 0..63. constexpr uint32_t RSQRT_COEFFS[64][4] = { {0xffffffff, 0xfffff780, 0xbff55815, 0x9bb5b6e7}, {0xfc0bd889, 0xfa1d6e7d, 0xb8a95a89, 0x938bf8f0}, {0xf82ec882, 0xf473bea9, 0xb1bf4705, 0x8bed0079}, {0xf467f280, 0xeefff2a1, 0xab309d4a, 0x84cdb431}, {0xf0b6848c, 0xe9bf46f4, 0xa4f76232, 0x7e24037b}, {0xed19b75e, 0xe4af2628, 0x9f0e1340, 0x77e6ca62}, {0xe990cdad, 0xdfcd2521, 0x996f9b96, 0x720db8df}, {0xe61b138e, 0xdb16ffde, 0x94174a00, 0x6c913cff}, {0xe2b7dddf, 0xd68a967b, 0x8f00c812, 0x676a6f92}, {0xdf6689b7, 0xd225ea80, 0x8a281226, 0x62930308}, {0xdc267bea, 0xcde71c63, 0x8589702c, 0x5e05343e}, {0xd8f7208e, 0xc9cc6948, 0x81216f2e, 0x59bbbcf8}, {0xd5d7ea91, 0xc5d428ee, 0x7cecdb76, 0x55b1c7d6}, {0xd2c8534e, 0xc1fccbc9, 0x78e8bb45, 0x51e2e592}, {0xcfc7da32, 0xbe44d94a, 0x75124a0a, 0x4e4b0369}, {0xccd6045f, 0xbaaaee41, 0x7166f40f, 0x4ae66284}, {0xc9f25c5c, 0xb72dbb69, 0x6de45288, 0x47b19045}, {0xc71c71c7, 0xb3cc040f, 0x6a882804, 0x44a95f5f}, {0xc453d90f, 0xb0849cd4, 0x67505d2a, 0x41cae1a0}, {0xc1982b2e, 0xad566a85, 0x643afdc8, 0x3f13625c}, {0xbee9056f, 0xaa406113, 0x6146361f, 0x3c806169}, {0xbc46092e, 0xa7418293, 0x5e70506d, 0x3a0f8e8e}, {0xb9aedba5, 0xa458de58, 0x5bb7b2b1, 0x37bec572}, {0xb72325b7, 0xa1859022, 0x591adc9a, 0x358c09e2}, {0xb4a293c2, 0x9ec6bf52, 0x569865a7, 0x33758476}, {0xb22cd56d, 0x9c1b9e36, 0x542efb6a, 0x31797f8a}, {0xafc19d86, 0x9983695c, 0x51dd5ffb, 0x2f96647a}, {0xad60a1d1, 0x96fd66f7, 0x4fa2687c, 0x2dcab91f}, {0xab099ae9, 0x9488e64b, 0x4d7cfbc9, 0x2c151d8a}, {0xa8bc441a, 0x92253f20, 0x4b6c1139, 0x2a7449ef}, {0xa6785b42, 0x8fd1d14a, 0x496eaf82, 0x28e70cc3}, {0xa43da0ae, 0x8d8e042a, 0x4783eba7, 0x276c4900}, {0xa20bd701, 0x8b594648, 0x45aae80a, 0x2602f493}, {0x9fe2c315, 0x89330ce4, 0x43e2d382, 0x24aa16ec}, {0x9dc22be4, 0x871ad399, 0x422ae88c, 0x2360c7af}, {0x9ba9da6c, 0x85101c05, 0x40826c88, 0x22262d7b}, {0x99999999, 0x83126d70, 0x3ee8af07, 0x20f97cd2}, {0x97913630, 0x81215480, 0x3d5d0922, 0x1fd9f714}, {0x95907eb8, 0x7f3c62ef, 0x3bdedce0, 0x1ec6e994}, {0x93974369, 0x7d632f45, 0x3a6d94a9, 0x1dbfacbb}, {0x91a55615, 0x7b955498, 0x3908a2be, 0x1cc3a33b}, {0x8fba8a1c, 0x79d2724e, 0x37af80bf, 0x1bd23960}, {0x8dd6b456, 0x781a2be4, 0x3661af39, 0x1aeae458}, {0x8bf9ab07, 0x766c28ba, 0x351eb539, 0x1a0d21a2}, {0x8a2345cc, 0x74c813dd, 0x33e61feb, 0x19387676}, {0x88535d90, 0x732d9bdc, 0x32b7823a, 0x186c6f3e}, {0x8689cc7e, 0x719c7297, 0x3192747d, 0x17a89f21}, {0x84c66df1, 0x70144d19, 0x30769424, 0x16ec9f89}, {0x83091e6a, 0x6e94e36c, 0x2f63836f, 0x16380fbf}, {0x8151bb87, 0x6d1df079, 0x2e58e925, 0x158a9484}, {0x7fa023f1, 0x6baf31de, 0x2d567053, 0x14e3d7ba}, {0x7df43758, 0x6a4867d3, 0x2c5bc811, 0x1443880e}, {0x7c4dd664, 0x68e95508, 0x2b68a346, 0x13a958ab}, {0x7aace2b0, 0x6791be86, 0x2a7cb871, 0x131500ee}, {0x79113ebc, 0x66416b95, 0x2997c17a, 0x12863c29}, {0x777acde8, 0x64f825a1, 0x28b97b82, 0x11fcc95c}, {0x75e9746a, 0x63b5b822, 0x27e1a6b4, 0x11786b03}, {0x745d1746, 0x6279f081, 0x2710061d, 0x10f8e6da}, {0x72d59c46, 0x61449e06, 0x26445f86, 0x107e05ac}, {0x7152e9f4, 0x601591be, 0x257e7b4d, 0x10079327}, {0x6fd4e793, 0x5eec9e6b, 0x24be2445, 0x0f955da9}, {0x6e5b7d16, 0x5dc9986e, 0x24032795, 0x0f273620}, {0x6ce6931d, 0x5cac55b7, 0x234d5496, 0x0ebcefdb}, {0x6b7612ec, 0x5b94adb2, 0x229c7cbc, 0x0e56606e}, }; // Approximate rsqrt with cubic polynomials. // The range [1,2] is splitted into 64 equal sub-ranges and the reciprocal // square root is approximated by a cubic polynomial by the minimax method in // each subrange. The approximation accuracy fits into 32-33 bits and thus it is // natural to round coefficients into 32 bit. The constant coefficient can be // rounded to 33 bits since the most significant bit is always 1 and implicitly // assumed in the table. LIBC_INLINE uint64_t rsqrt_approx(uint64_t m) { // ULP(m) = 2^-64. // Use the top 6 bits as index for looking up polynomial coeffs. uint64_t indx = m >> 58; uint64_t c0 = static_cast(RSQRT_COEFFS[indx][0]); c0 <<= 31; // to 64 bit with the space for the implicit bit c0 |= 1ull << 63; // add implicit bit uint64_t c1 = static_cast(RSQRT_COEFFS[indx][1]); c1 <<= 25; // to 64 bit format uint64_t c2 = static_cast(RSQRT_COEFFS[indx][2]); uint64_t c3 = static_cast(RSQRT_COEFFS[indx][3]); uint64_t d = (m << 6) >> 32; // local coordinate in the subrange [0, 2^32] uint64_t d2 = (d * d) >> 32; // square of the local coordinate uint64_t re = c0 + (d2 * c2 >> 13); // even part of the polynomial (positive) uint64_t ro = d * ((c1 + ((d2 * c3) >> 19)) >> 26) >> 6; // odd part of the polynomial (negative) uint64_t r = re - ro; // maximal error < 1.55e-10 and it is less than 2^-32 // Newton-Raphson first order step to improve accuracy of the result to almost // 64 bits. r = rsqrt_newton_raphson(m, r); // Adjust in the unlucky case x~1; if (LIBC_UNLIKELY(!r)) --r; return r; } #endif // LIBC_MATH_HAS_SMALL_TABLES } // anonymous namespace LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) { using FPBits = fputil::FPBits; // Get rounding mode. uint32_t rm = fputil::get_round(); FPBits xbits(x); UInt128 x_u = xbits.uintval(); // Bring leading bit of the mantissa to the highest bit. // ulp(x_frac) = 2^-128. UInt128 x_frac = xbits.get_mantissa() << (FPBits::EXP_LEN + 1); int sign_exp = static_cast(x_u >> FPBits::FRACTION_LEN); if (LIBC_UNLIKELY(sign_exp == 0 || sign_exp >= 0x7fff)) { // Special cases: NAN, inf, negative numbers if (sign_exp >= 0x7fff) { // x = -0 or x = inf if (xbits.is_zero() || xbits == xbits.inf()) return x; // x is nan if (xbits.is_nan()) { // pass through quiet nan if (xbits.is_quiet_nan()) return x; // transform signaling nan to quiet and return return xbits.quiet_nan().get_val(); } // x < 0 or x = -inf fputil::set_errno_if_required(EDOM); fputil::raise_except_if_required(FE_INVALID); return xbits.quiet_nan().get_val(); } // Now x is subnormal or x = +0. // x is +0. if (x_frac == 0) return x; // Normalize subnormal inputs. sign_exp = -cpp::countl_zero(x_frac); int normal_shifts = 1 - sign_exp; x_frac <<= normal_shifts; } // For sign_exp = biased exponent of x = real_exponent + 16383, // let f be the real exponent of the output: // f = floor(real_exponent / 2) // Then: // floor((sign_exp + 1) / 2) = f + 8192 // Hence, the biased exponent of the final result is: // f + 16383 = floor((sign_exp + 1) / 2) + 8191. // Since the output mantissa will include the hidden bit, we can define the // output exponent part: // e2 = floor((sign_exp + 1) / 2) + 8190 unsigned i = static_cast(1 - (sign_exp & 1)); uint32_t q2 = (sign_exp + 1) >> 1; // Exponent of the final result uint32_t e2 = q2 + 8190; constexpr uint64_t RSQRT_2[2] = {~0ull, 0xb504f333f9de6484 /* 2^64/sqrt(2) */}; // Approximate 1/sqrt(1 + x_frac) // Error: |r_1 - 1/sqrt(x)| < 2^-62. uint64_t r1 = rsqrt_approx(static_cast(x_frac >> 64)); // Adjust for the even/odd exponent. uint64_t r2 = prod_hi(r1, RSQRT_2[i]); unsigned shift = 2 - i; // Normalized input: // 1 <= x_reduced < 4 UInt128 x_reduced = (x_frac >> shift) | (UInt128(1) << (126 + i)); // With r2 ~ 1/sqrt(x) up to 2^-63, we perform another round of Newton-Raphson // iteration: // r3 = r2 - r2 * h / 2, // for h = r2^2 * x - 1. // Then: // sqrt(x) = x * (1 / sqrt(x)) // ~ x * r3 // = x * (r2 - r2 * h / 2) // = (x * r2) - (x * r2) * h / 2 UInt128 sx = prod_hi(x_reduced, r2); UInt128 h = prod_hi(sx, r2) << 2; UInt128 ds = static_cast(prod_hi(static_cast(h), sx)); UInt128 v = (sx << 1) - ds; uint32_t nrst = rm == FE_TONEAREST; // The result lies within (-2,5) of true square root so we now // test that we can correctly round the result taking into account // the rounding mode. // Check the lowest 14 bits (by clearing and sign-extending the top // 32 - 14 = 18 bits). int dd = (static_cast(v) << 18) >> 18; if (LIBC_UNLIKELY(dd < 4 && dd >= -8)) { // can round correctly? // m is almost the final result it can be only 1 ulp off so we // just need to test both possibilities. We square it and // compare with the initial argument. UInt128 m = v >> 15; UInt128 m2 = m * m; // The difference of the squared result and the argument Int128 t0 = static_cast(m2 - (x_reduced << 98)); if (t0 == 0) { // the square root is exact v = m << 15; } else { // Add +-1 ulp to m depend on the sign of the difference. Here // we do not need to square again since (m+1)^2 = m^2 + 2*m + // 1 so just need to add shifted m and 1. Int128 t1 = t0; Int128 sgn = t0 >> 127; // sign of the difference Int128 m_xor_sgn = static_cast(m << 1) ^ sgn; t1 -= m_xor_sgn; t1 += Int128(1) + sgn; Int128 sgn1 = t1 >> 127; if (LIBC_UNLIKELY(sgn == sgn1)) { t0 = t1; v -= sgn << 15; t1 -= m_xor_sgn; t1 += Int128(1) + sgn; } if (t1 == 0) { // 1 ulp offset brings again an exact root v = (m - static_cast((sgn << 1) + 1)) << 15; } else { t1 += t0; Int128 side = t1 >> 127; // select what is closer m or m+-1 v &= ~UInt128(0) << 15; // wipe the fractional bits v -= ((sgn & side) | (~sgn & 1)) << (15 + static_cast(side)); v |= 1; // add sticky bit since we cannot have an exact mid-point // situation } } } unsigned frac = static_cast(v) & 0x7fff; // fractional part unsigned rnd; // round bit if (LIBC_LIKELY(nrst != 0)) { rnd = frac >> 14; // round to nearest tie to even } else if (rm == FE_UPWARD) { rnd = !!frac; // round up } else { rnd = 0; // round down or round to zero } v >>= 15; // position mantissa v += rnd; // round // Set inexact flag only if square root is inexact // TODO: We will have to raise FE_INEXACT most of the time, but this // operation is very costly, especially in x86-64, since technically, it // needs to synchronize both SSE and x87 flags. Need to investigate // further to see how we can make this performant. // https://github.com/llvm/llvm-project/issues/126753 // if(frac) fputil::raise_except_if_required(FE_INEXACT); v += static_cast(e2) << FPBits::FRACTION_LEN; // place exponent return cpp::bit_cast(v); } } // namespace LIBC_NAMESPACE_DECL