[libc-commits] [libc] [libc][math] Improve the performance of sqrtf128. (PR #122578)

via libc-commits libc-commits at lists.llvm.org
Tue Feb 11 13:03:23 PST 2025


https://github.com/lntue updated https://github.com/llvm/llvm-project/pull/122578

>From c1e33345990f063625d55cc0d1dde320318e525c Mon Sep 17 00:00:00 2001
From: Tue Ly <lntue.h at gmail.com>
Date: Sat, 11 Jan 2025 05:50:43 +0000
Subject: [PATCH 1/6] [libc][math] Improve the performance of sqrtf128.

Use a combination of polynomial approximation and Newton-Raphson iterations
in 64-bit and 128-bit integers to improve the performance of sqrtf128.  The
correct rounding is provided by squaring the result and comparing it with
the argument.

Co-authored-by: Alexei Sibidanov <sibid at uvic.ca>
---
 libc/src/__support/big_int.h                  |   2 +-
 libc/src/math/generic/CMakeLists.txt          |   5 +
 libc/src/math/generic/sqrtf128.cpp            | 341 +++++++++++++++++-
 libc/test/UnitTest/FPMatcher.h                |  76 ++--
 libc/test/src/math/SqrtTest.h                 |   6 +-
 .../math/performance_testing/CMakeLists.txt   |  11 +
 .../performance_testing/sqrtf128_perf.cpp     |  20 +
 libc/test/src/math/smoke/SqrtTest.h           |   3 +-
 .../test/src/math/smoke/generic_sqrt_test.cpp |   2 +-
 .../src/math/smoke/generic_sqrtf128_test.cpp  |   2 +-
 .../src/math/smoke/generic_sqrtf_test.cpp     |   2 +-
 .../src/math/smoke/generic_sqrtl_test.cpp     |   2 +-
 libc/test/src/math/smoke/sqrt_test.cpp        |   2 +-
 libc/test/src/math/smoke/sqrtf128_test.cpp    | 126 ++++++-
 libc/test/src/math/smoke/sqrtf16_test.cpp     |   2 +-
 libc/test/src/math/smoke/sqrtf_test.cpp       |   2 +-
 libc/test/src/math/smoke/sqrtl_test.cpp       |   2 +-
 17 files changed, 568 insertions(+), 38 deletions(-)
 create mode 100644 libc/test/src/math/performance_testing/sqrtf128_perf.cpp

diff --git a/libc/src/__support/big_int.h b/libc/src/__support/big_int.h
index f591b41df037b..fb5ad99d53e7b 100644
--- a/libc/src/__support/big_int.h
+++ b/libc/src/__support/big_int.h
@@ -241,7 +241,7 @@ LIBC_INLINE constexpr void quick_mul_hi(cpp::array<word, N> &dst,
 }
 
 template <typename word, size_t N>
-LIBC_INLINE constexpr bool is_negative(cpp::array<word, N> &array) {
+LIBC_INLINE constexpr bool is_negative(const cpp::array<word, N> &array) {
   using signed_word = cpp::make_signed_t<word>;
   return cpp::bit_cast<signed_word>(array.back()) < 0;
 }
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 9faf46d491426..9353b6e7d9fb7 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -2978,6 +2978,11 @@ add_entrypoint_object(
   HDRS
     ../sqrtf128.h
   DEPENDS
+    libc.src.__support.CPP.bit
+    libc.src.__support.FPUtil.fenv_impl
+    libc.src.__support.FPUtil.fp_bits
+    libc.src.__support.FPUtil.rounding_mode
+    libc.src.__support.macros.optimization
     libc.src.__support.macros.properties.types
     libc.src.__support.FPUtil.sqrt
   )
diff --git a/libc/src/math/generic/sqrtf128.cpp b/libc/src/math/generic/sqrtf128.cpp
index f87066b6f6403..79f1681a9e871 100644
--- a/libc/src/math/generic/sqrtf128.cpp
+++ b/libc/src/math/generic/sqrtf128.cpp
@@ -1,5 +1,7 @@
 //===-- Implementation of sqrtf128 function -------------------------------===//
 //
+// Copyright (c) 2024 Alexei Sibidanov <sibid at uvic.ca>
+//
 // 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
@@ -7,14 +9,347 @@
 //===----------------------------------------------------------------------===//
 
 #include "src/math/sqrtf128.h"
-#include "src/__support/FPUtil/sqrt.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/config.h"
+#include "src/__support/macros/optimization.h"
+#include "src/__support/uint128.h"
 
 namespace LIBC_NAMESPACE_DECL {
 
+using FPBits = fputil::FPBits<float128>;
+
+namespace {
+
+template <typename T, typename U = T> 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>(uint64_t x, uint64_t y) {
+  return static_cast<uint64_t>(
+      (static_cast<UInt128>(x) * static_cast<UInt128>(y)) >> 64);
+}
+
+// Get high part of unsigned 128x64 bit multiplication.
+template <>
+inline constexpr UInt128 prod_hi<UInt128, uint64_t>(UInt128 y, uint64_t x) {
+  uint64_t y_lo = static_cast<uint64_t>(y);
+  uint64_t y_hi = static_cast<uint64_t>(y >> 64);
+  UInt128 xyl = static_cast<UInt128>(x) * static_cast<UInt128>(y_lo);
+  UInt128 xyh = static_cast<UInt128>(x) * static_cast<UInt128>(y_hi);
+  return xyh + (xyl >> 64);
+}
+
+// Get high part of signed 64x64 bit multiplication.
+template <> inline constexpr int64_t prod_hi<int64_t>(int64_t x, int64_t y) {
+  return static_cast<int64_t>(
+      (static_cast<Int128>(x) * static_cast<Int128>(y)) >> 64);
+}
+
+// Get high 128-bit part of unsigned 128x128 bit multiplication.
+template <> inline constexpr UInt128 prod_hi<UInt128>(UInt128 x, UInt128 y) {
+  uint64_t x_lo = static_cast<uint64_t>(x);
+  uint64_t x_hi = static_cast<uint64_t>(x >> 64);
+  uint64_t y_lo = static_cast<uint64_t>(y);
+  uint64_t y_hi = static_cast<uint64_t>(y >> 64);
+
+  UInt128 xh_yh = static_cast<UInt128>(x_hi) * static_cast<UInt128>(y_hi);
+  UInt128 xh_yl = static_cast<UInt128>(x_hi) * static_cast<UInt128>(y_lo);
+  UInt128 xl_yh = static_cast<UInt128>(x_lo) * static_cast<UInt128>(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, UInt128>(Int128 x, UInt128 y) {
+  UInt128 mask = static_cast<UInt128>(x >> 127);
+  UInt128 negative_part = y & mask;
+  UInt128 prod = prod_hi(static_cast<UInt128>(x), y);
+  return static_cast<Int128>(prod - negative_part);
+}
+
+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<uint64_t>(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<uint64_t>(RSQRT_COEFFS[indx][1]);
+  c1 <<= 25; // to 64 bit format
+
+  uint64_t c2 = static_cast<uint64_t>(RSQRT_COEFFS[indx][2]);
+  uint64_t c3 = static_cast<uint64_t>(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:
+  // 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.
+  uint64_t r2 = prod_hi<uint64_t>(r, r);
+  // h = r0^2*x - 1.
+  int64_t h = static_cast<int64_t>(prod_hi<uint64_t>(m, r2) + r2);
+  // hr = r * h / 2
+  int64_t hr = prod_hi<int64_t>(h, static_cast<int64_t>(r >> 1));
+  r -= hr;
+  // Adjust in the unlucky case x~1;
+  if (LIBC_UNLIKELY(!r))
+    --r;
+  return r;
+}
+
+} // anonymous namespace
+
 LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) {
-  return fputil::sqrt<float128>(x);
+  using FPBits = fputil::FPBits<float128>;
+  // 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<int>(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();
+    }
+    // x is subnormal or x=+0
+    if (x == 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<unsigned>(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^-63.
+  uint64_t r1 = rsqrt_approx(static_cast<uint64_t>(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<UInt128>(prod_hi(static_cast<Int128>(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.
+  int dd = (static_cast<int>(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;
+    Int128 t0, t1;
+    // The difference of the squared result and the argument
+    t0 = static_cast<Int128>(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.
+      t1 = t0;
+      Int128 sgn = t0 >> 127; // sign of the difference
+      t1 -= (m << 1) ^ sgn;
+      t1 += 1 + sgn;
+
+      Int128 sgn1 = t1 >> 127;
+      if (LIBC_UNLIKELY(sgn == sgn1)) {
+        t0 = t1;
+        v -= sgn << 15;
+        t1 -= (m << 1) ^ sgn;
+        t1 += 1 + sgn;
+      }
+
+      if (t1 == 0) {
+        // 1 ulp offset brings again an exact root
+        v = (m - (2 * sgn + 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 + side);
+        v |= 1; // add sticky bit since we cannot have an exact mid-point
+                // situation
+      }
+    }
+  }
+
+  unsigned frac = static_cast<unsigned>(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 if (rm == FE_DOWNWARD) {
+    rnd = 0; // round down
+  } else {
+    rnd = 0; // 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.
+  // if(frac) fputil::raise_except_if_required(FE_INEXACT);
+
+  v += static_cast<UInt128>(e2) << FPBits::FRACTION_LEN; // place exponent
+  return cpp::bit_cast<float128>(v);
 }
 
 } // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/UnitTest/FPMatcher.h b/libc/test/UnitTest/FPMatcher.h
index 53e0c16f22101..21b8a45b0726f 100644
--- a/libc/test/UnitTest/FPMatcher.h
+++ b/libc/test/UnitTest/FPMatcher.h
@@ -330,27 +330,6 @@ struct ModifyMXCSR {
     EXPECT_FP_EXCEPTION(expected_except);                                      \
   } while (0)
 
-#define EXPECT_FP_EQ_ALL_ROUNDING(expected, actual)                            \
-  do {                                                                         \
-    using namespace LIBC_NAMESPACE::fputil::testing;                           \
-    ForceRoundingMode __r1(RoundingMode::Nearest);                             \
-    if (__r1.success) {                                                        \
-      EXPECT_FP_EQ((expected), (actual));                                      \
-    }                                                                          \
-    ForceRoundingMode __r2(RoundingMode::Upward);                              \
-    if (__r2.success) {                                                        \
-      EXPECT_FP_EQ((expected), (actual));                                      \
-    }                                                                          \
-    ForceRoundingMode __r3(RoundingMode::Downward);                            \
-    if (__r3.success) {                                                        \
-      EXPECT_FP_EQ((expected), (actual));                                      \
-    }                                                                          \
-    ForceRoundingMode __r4(RoundingMode::TowardZero);                          \
-    if (__r4.success) {                                                        \
-      EXPECT_FP_EQ((expected), (actual));                                      \
-    }                                                                          \
-  } while (0)
-
 #define EXPECT_FP_EQ_ROUNDING_MODE(expected, actual, rounding_mode)            \
   do {                                                                         \
     using namespace LIBC_NAMESPACE::fputil::testing;                           \
@@ -372,6 +351,61 @@ struct ModifyMXCSR {
 #define EXPECT_FP_EQ_ROUNDING_TOWARD_ZERO(expected, actual)                    \
   EXPECT_FP_EQ_ROUNDING_MODE((expected), (actual), RoundingMode::TowardZero)
 
+#define EXPECT_FP_EQ_ALL_ROUNDING_1(expected, actual)                          \
+  do {                                                                         \
+    EXPECT_FP_EQ_ROUNDING_NEAREST((expected), (actual));                       \
+    EXPECT_FP_EQ_ROUNDING_UPWARD((expected), (actual));                        \
+    EXPECT_FP_EQ_ROUNDING_DOWNWARD((expected), (actual));                      \
+    EXPECT_FP_EQ_ROUNDING_TOWARD_ZERO((expected), (actual));                   \
+  } while (0)
+
+#define EXPECT_FP_EQ_ALL_ROUNDING_4(expected_nearest, expected_upward,         \
+                                    expected_downward, expected_toward_zero,   \
+                                    actual)                                    \
+  do {                                                                         \
+    EXPECT_FP_EQ_ROUNDING_NEAREST((expected_nearest), (actual));               \
+    EXPECT_FP_EQ_ROUNDING_UPWARD((expected_upward), (actual));                 \
+    EXPECT_FP_EQ_ROUNDING_DOWNWARD((expected_downward), (actual));             \
+    EXPECT_FP_EQ_ROUNDING_TOWARD_ZERO((expected_toward_zero), (actual));       \
+  } while (0)
+
+#define EXPECT_FP_EQ_ALL_ROUNDING_UNSUPPORTED(...)                             \
+  static_assert(false, "Unsupported number of arguments")
+
+#define EXPECT_FP_EQ_ALL_ROUNDING_GET_6TH_ARG(ARG1, ARG2, ARG3, ARG4, ARG5,    \
+                                              ARG6, ...)                       \
+  ARG6
+
+#define EXPECT_FP_EQ_ALL_ROUNDING_SELECTION(...)                               \
+  EXPECT_FP_EQ_ALL_ROUNDING_GET_6TH_ARG(                                       \
+      __VA_ARGS__, EXPECT_FP_EQ_ALL_ROUNDING_4,                                \
+      EXPECT_FP_EQ_ALL_ROUNDING_UNSUPPORTED,                                   \
+      EXPECT_FP_EQ_ALL_ROUNDING_UNSUPPORTED, EXPECT_FP_EQ_ALL_ROUNDING_1)
+
+#define EXPECT_FP_EQ_ALL_ROUNDING(...)                                         \
+  EXPECT_FP_EQ_ALL_ROUNDING_SELECTION(__VA_ARGS__)(__VA_ARGS__)
+
+#define ASSERT_FP_EQ_ROUNDING_MODE(expected, actual, rounding_mode)            \
+  do {                                                                         \
+    using namespace LIBC_NAMESPACE::fputil::testing;                           \
+    ForceRoundingMode __r((rounding_mode));                                    \
+    if (__r.success) {                                                         \
+      ASSERT_FP_EQ((expected), (actual));                                      \
+    }                                                                          \
+  } while (0)
+
+#define ASSERT_FP_EQ_ROUNDING_NEAREST(expected, actual)                        \
+  ASSERT_FP_EQ_ROUNDING_MODE((expected), (actual), RoundingMode::Nearest)
+
+#define ASSERT_FP_EQ_ROUNDING_UPWARD(expected, actual)                         \
+  ASSERT_FP_EQ_ROUNDING_MODE((expected), (actual), RoundingMode::Upward)
+
+#define ASSERT_FP_EQ_ROUNDING_DOWNWARD(expected, actual)                       \
+  ASSERT_FP_EQ_ROUNDING_MODE((expected), (actual), RoundingMode::Downward)
+
+#define ASSERT_FP_EQ_ROUNDING_TOWARD_ZERO(expected, actual)                    \
+  ASSERT_FP_EQ_ROUNDING_MODE((expected), (actual), RoundingMode::TowardZero)
+
 #define EXPECT_FP_EQ_WITH_EXCEPTION_ROUNDING_MODE(                             \
     expected, actual, expected_except, rounding_mode)                          \
   do {                                                                         \
diff --git a/libc/test/src/math/SqrtTest.h b/libc/test/src/math/SqrtTest.h
index 770cc94b3b940..fdfc4f9bb9943 100644
--- a/libc/test/src/math/SqrtTest.h
+++ b/libc/test/src/math/SqrtTest.h
@@ -29,14 +29,14 @@ class SqrtTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
       FPBits denormal(zero);
       denormal.set_mantissa(mant);
       InType x = denormal.get_val();
-      EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Sqrt, x, func(x), 0.5);
+      ASSERT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Sqrt, x, func(x), 0.5);
     }
 
     constexpr StorageType COUNT = 200'001;
     constexpr StorageType STEP = HIDDEN_BIT / COUNT;
     for (StorageType i = 0, v = 0; i <= COUNT; ++i, v += STEP) {
       InType x = FPBits(i).get_val();
-      EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Sqrt, x, func(x), 0.5);
+      ASSERT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Sqrt, x, func(x), 0.5);
     }
   }
 
@@ -48,7 +48,7 @@ class SqrtTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
       InType x = x_bits.get_val();
       if (x_bits.is_nan() || (x < 0))
         continue;
-      EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Sqrt, x, func(x), 0.5);
+      ASSERT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Sqrt, x, func(x), 0.5);
     }
   }
 };
diff --git a/libc/test/src/math/performance_testing/CMakeLists.txt b/libc/test/src/math/performance_testing/CMakeLists.txt
index 60c074a248f72..862a0f6d32386 100644
--- a/libc/test/src/math/performance_testing/CMakeLists.txt
+++ b/libc/test/src/math/performance_testing/CMakeLists.txt
@@ -500,3 +500,14 @@ add_perf_binary(
   COMPILE_OPTIONS
     -fno-builtin
 )
+
+add_perf_binary(
+  sqrtf128_perf
+  SRCS
+    sqrtf128_perf.cpp
+  DEPENDS
+    .single_input_single_output_diff
+    libc.src.math.sqrtf128
+  COMPILE_OPTIONS
+  -fno-builtin
+)
diff --git a/libc/test/src/math/performance_testing/sqrtf128_perf.cpp b/libc/test/src/math/performance_testing/sqrtf128_perf.cpp
new file mode 100644
index 0000000000000..bc04e698b2439
--- /dev/null
+++ b/libc/test/src/math/performance_testing/sqrtf128_perf.cpp
@@ -0,0 +1,20 @@
+//===-- Differential test for sqrtf128
+//----------------------------------------===//
+//
+// 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 "SingleInputSingleOutputPerf.h"
+
+#include "src/__support/FPUtil/sqrt.h"
+#include "src/math/sqrtf128.h"
+
+float128 sqrtf128_placeholder(float128 x) {
+  return LIBC_NAMESPACE::fputil::sqrt<float128>(x);
+}
+
+SINGLE_INPUT_SINGLE_OUTPUT_PERF(float128, LIBC_NAMESPACE::sqrtf128,
+                                ::sqrtf128_placeholder, "sqrtf128_perf.log")
diff --git a/libc/test/src/math/smoke/SqrtTest.h b/libc/test/src/math/smoke/SqrtTest.h
index b5eaee22fc79d..29666ad0d4e56 100644
--- a/libc/test/src/math/smoke/SqrtTest.h
+++ b/libc/test/src/math/smoke/SqrtTest.h
@@ -39,7 +39,8 @@ class SqrtTest : public LIBC_NAMESPACE::testing::FEnvSafeTest {
 
 #define LIST_SQRT_TESTS(T, func)                                               \
   using LlvmLibcSqrtTest = SqrtTest<T, T>;                                     \
-  TEST_F(LlvmLibcSqrtTest, SpecialNumbers) { test_special_numbers(&func); }
+  TEST_F(LlvmLibcSqrtTest, SpecialNumbers) { test_special_numbers(&func); }    \
+  static_assert(true, "Require semicolon.")
 
 #define LIST_NARROWING_SQRT_TESTS(OutType, InType, func)                       \
   using LlvmLibcSqrtTest = SqrtTest<OutType, InType>;                          \
diff --git a/libc/test/src/math/smoke/generic_sqrt_test.cpp b/libc/test/src/math/smoke/generic_sqrt_test.cpp
index d0ab31ffd0fe6..4451e5e82d2d4 100644
--- a/libc/test/src/math/smoke/generic_sqrt_test.cpp
+++ b/libc/test/src/math/smoke/generic_sqrt_test.cpp
@@ -10,4 +10,4 @@
 
 #include "src/__support/FPUtil/generic/sqrt.h"
 
-LIST_SQRT_TESTS(double, LIBC_NAMESPACE::fputil::sqrt<double>)
+LIST_SQRT_TESTS(double, LIBC_NAMESPACE::fputil::sqrt<double>);
diff --git a/libc/test/src/math/smoke/generic_sqrtf128_test.cpp b/libc/test/src/math/smoke/generic_sqrtf128_test.cpp
index edba114adf06c..790ff0a47bd3a 100644
--- a/libc/test/src/math/smoke/generic_sqrtf128_test.cpp
+++ b/libc/test/src/math/smoke/generic_sqrtf128_test.cpp
@@ -10,4 +10,4 @@
 
 #include "src/__support/FPUtil/generic/sqrt.h"
 
-LIST_SQRT_TESTS(float128, LIBC_NAMESPACE::fputil::sqrt<float128>)
+LIST_SQRT_TESTS(float128, LIBC_NAMESPACE::fputil::sqrt<float128>);
diff --git a/libc/test/src/math/smoke/generic_sqrtf_test.cpp b/libc/test/src/math/smoke/generic_sqrtf_test.cpp
index f22ac8829d5ac..e04d4c4f26d63 100644
--- a/libc/test/src/math/smoke/generic_sqrtf_test.cpp
+++ b/libc/test/src/math/smoke/generic_sqrtf_test.cpp
@@ -10,4 +10,4 @@
 
 #include "src/__support/FPUtil/generic/sqrt.h"
 
-LIST_SQRT_TESTS(float, LIBC_NAMESPACE::fputil::sqrt<float>)
+LIST_SQRT_TESTS(float, LIBC_NAMESPACE::fputil::sqrt<float>);
diff --git a/libc/test/src/math/smoke/generic_sqrtl_test.cpp b/libc/test/src/math/smoke/generic_sqrtl_test.cpp
index ddc6a23695be4..ccb5054296115 100644
--- a/libc/test/src/math/smoke/generic_sqrtl_test.cpp
+++ b/libc/test/src/math/smoke/generic_sqrtl_test.cpp
@@ -10,4 +10,4 @@
 
 #include "src/__support/FPUtil/generic/sqrt.h"
 
-LIST_SQRT_TESTS(long double, LIBC_NAMESPACE::fputil::sqrt<long double>)
+LIST_SQRT_TESTS(long double, LIBC_NAMESPACE::fputil::sqrt<long double>);
diff --git a/libc/test/src/math/smoke/sqrt_test.cpp b/libc/test/src/math/smoke/sqrt_test.cpp
index 1551b31d6f715..b41e06daf722e 100644
--- a/libc/test/src/math/smoke/sqrt_test.cpp
+++ b/libc/test/src/math/smoke/sqrt_test.cpp
@@ -10,4 +10,4 @@
 
 #include "src/math/sqrt.h"
 
-LIST_SQRT_TESTS(double, LIBC_NAMESPACE::sqrt)
+LIST_SQRT_TESTS(double, LIBC_NAMESPACE::sqrt);
diff --git a/libc/test/src/math/smoke/sqrtf128_test.cpp b/libc/test/src/math/smoke/sqrtf128_test.cpp
index 23397b0623ce5..d4cd1be5452d7 100644
--- a/libc/test/src/math/smoke/sqrtf128_test.cpp
+++ b/libc/test/src/math/smoke/sqrtf128_test.cpp
@@ -1,5 +1,7 @@
 //===-- Unittests for sqrtf128---------------------------------------------===//
 //
+// Copyright (c) 2024 Alexei Sibidanov <sibid at uvic.ca>
+//
 // 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
@@ -8,6 +10,128 @@
 
 #include "SqrtTest.h"
 
+#include "src/__support/uint128.h"
 #include "src/math/sqrtf128.h"
 
-LIST_SQRT_TESTS(float128, LIBC_NAMESPACE::sqrtf128)
+LIST_SQRT_TESTS(float128, LIBC_NAMESPACE::sqrtf128);
+
+TEST_F(LlvmLibcSqrtTest, HardToRound) {
+  using LIBC_NAMESPACE::fputil::testing::RoundingMode;
+  using FPBits = LIBC_NAMESPACE::fputil::FPBits<float128>;
+
+  // Since there is no exact half cases for square root I encode the
+  // round direction in the sign of the result. E.g. if the number is
+  // negative it means that the exact root is below the rounded value
+  // (the absolute value). Thus I can test not only hard to round
+  // cases for the round to nearest mode but also the directional
+  // modes.
+  float128 HARD_TO_ROUND[][2] = {
+      {0x0.000000dee2f5b6a26c8f07f05442p-16382q,
+       -0x1.ddbd8763a617cff753e2a31083p-8204q},
+      {0x0.000000c86d174c5ad8ae54a548e7p-16382q,
+       0x1.c507bb538940719890851ec1ca88p-8204q},
+      {0x0.000020ab15cfe0b8e488e128f535p-16382q,
+       -0x1.6dccb402560213bc0d62d62e910bp-8201q},
+      {0x0.0000219e97732a9970f2511989bap-16382q,
+       0x1.73163d28be706f4b5052791e28a5p-8201q},
+      {0x0.000026e477546ae99ef57066f9fdp-16382q,
+       -0x1.8f20dd0d0c570a23ea59bc2bf009p-8201q},
+      {0x0.00002d0f88d27a496b3e533f5067p-16382q,
+       0x1.ad9d4abe9f047225a7352bcc52c1p-8201q},
+      {0x1.0000000000000000000000000001p+0q, 0x1p+0q},
+      {0x1.0000000000000000000000000002p+0q,
+       -0x1.0000000000000000000000000001p+0q},
+      {0x1.0000000000000000000000000003p+0q,
+       0x1.0000000000000000000000000001p+0q},
+      {0x1.0000000000000000000000000005p+0q,
+       0x1.0000000000000000000000000002p+0q},
+      {0x1.0000000000000000000000000006p+0q,
+       -0x1.0000000000000000000000000003p+0q},
+      {0x1.1d4c381cbf3a0aa15b9aee344892p+0q,
+       0x1.0e408c3fadc5e64b449c63673f4bp+0q},
+      {0x1.2af17a4ae6f93d11310c49c11b59p+0q,
+       -0x1.14a3bdf0ea5231f12d421a5dbe33p+0q},
+      {0x1.96f893bf29fb91e0fbe19a46d0c8p+0q,
+       0x1.42c6bf6202e66f2295807dee44d9p+0q},
+      {0x1.97fb3839925b66804c429289cce8p+0q,
+       -0x1.432d4049ac1c85a241f333d326e9p+0q},
+      {0x1.be1d900eaeb1533f0f19cc15c7e6p+0q,
+       0x1.51f1715154da44f3bf11f3d96c2dp+0q},
+      {0x1.c4f5074269525063a26051a0ad27p+0q,
+       0x1.54864e9b1daa4d9135ff00663366p+0q},
+      {0x1.035cb5f298a801dc4be9b1f8cd97p+1q,
+       -0x1.6c688775bffcb3f507ba11d0abb9p+0q},
+      {0x1.274be02380427e709beab4dedeb4p+1q,
+       -0x1.84d5763281f2318422392e506b1cp+0q},
+      {0x1.64e797cfdbaa3f7e2f33279dbc6p+1q,
+       0x1.ab79b164e255b26eca00ff99cc99p+0q},
+      {0x1.693a741358c9dac44a570a7e9f6cp+1q,
+       0x1.ae0e8eaeab25bb0c40ee0c2693d3p+0q},
+      {0x1.8275db3fc4d822596047adcb71b9p+1q,
+       -0x1.bcd2bfb653e37a5dbe0ccc2cd917p+0q},
+      {0x1.83280bb98c4a7b88bd6f535899d9p+1q,
+       0x1.bd39409dfd1990dd6a7f8211bb27p+0q},
+      {0x1.d78d8352b48608b510bfd5c75315p+1q,
+       -0x1.eb5c420f15adce0ed2bde5a241cep+0q},
+      {0x1.e3e4774f564b526edff84ce46668p+1q,
+       0x1.f1bf73c0523a19b4bb639c98c0b5p+0q},
+      {0x1.fffffffffffffffffffffffffffap+1q,
+       -0x1.fffffffffffffffffffffffffffdp+0q},
+      {0x1.fffffffffffffffffffffffffffbp+1q,
+       0x1.fffffffffffffffffffffffffffdp+0q},
+      {0x1.fffffffffffffffffffffffffffdp+1q,
+       0x1.fffffffffffffffffffffffffffep+0q},
+      {0x1.fffffffffffffffffffffffffffep+1q,
+       -0x1.ffffffffffffffffffffffffffffp+0q},
+      {0x1.ffffffffffffffffffffffffffffp+1q,
+       0x1.ffffffffffffffffffffffffffffp+0q},
+  };
+
+  auto rnd = [](float128 x, RoundingMode rm) -> float128 {
+    bool is_neg = x < 0;
+    float128 y = is_neg ? -x : x;
+    FPBits ybits(y);
+
+    if (is_neg &&
+        (rm == RoundingMode::Downward || rm == RoundingMode::TowardZero))
+      return FPBits(ybits.uintval() - 1).get_val();
+    if (!is_neg && (rm == RoundingMode::Upward))
+      return FPBits(ybits.uintval() + 1).get_val();
+
+    return y;
+  };
+
+  for (auto &t : HARD_TO_ROUND) {
+    EXPECT_FP_EQ_ALL_ROUNDING(
+        rnd(t[1], RoundingMode::Nearest), rnd(t[1], RoundingMode::Upward),
+        rnd(t[1], RoundingMode::Downward), rnd(t[1], RoundingMode::TowardZero),
+        LIBC_NAMESPACE::sqrtf128(t[0]));
+  }
+
+  // Exact results for subnormal arguments
+  float128 EXACT_SUBNORMAL[][2] = {
+      {0x0.0000000000000000000000000001p-16382q, 0x1p-8247q},
+      {0x0.0000000000000000000000000004p-16382q, 0x1p-8246q},
+      {0x0.0000000000001000000000000000p-16382q, 0x1p-8217q},
+      {0x0.0000000000010000000000000000p-16382q, 0x1p-8215q},
+      {0x0.0000000000100000000000000000p-16382q, 0x1p-8213q},
+  };
+
+  for (auto t : EXACT_SUBNORMAL)
+    EXPECT_FP_EQ_ALL_ROUNDING(t[1], LIBC_NAMESPACE::sqrtf128(t[0]));
+
+  // Check exact cases starting from small numbers
+  for (unsigned k = 1; k < 100 * 100; ++k) {
+    unsigned kx = k * k;
+    float128 x = kx, y = k;
+    EXPECT_FP_EQ_ALL_ROUNDING(y, LIBC_NAMESPACE::sqrtf128(x));
+  };
+
+  // Then from the largest number.
+  uint64_t k0 = 101904826760412362ULL;
+  for (uint64_t k = k0; k > k0 - 10000; --k) {
+    UInt128 kx = static_cast<UInt128>(k) * static_cast<UInt128>(k);
+    float128 x = kx, y = k;
+    EXPECT_FP_EQ_ALL_ROUNDING(y, LIBC_NAMESPACE::sqrtf128(x));
+  }
+}
diff --git a/libc/test/src/math/smoke/sqrtf16_test.cpp b/libc/test/src/math/smoke/sqrtf16_test.cpp
index d62049661eecb..950abd28840f0 100644
--- a/libc/test/src/math/smoke/sqrtf16_test.cpp
+++ b/libc/test/src/math/smoke/sqrtf16_test.cpp
@@ -10,4 +10,4 @@
 
 #include "src/math/sqrtf16.h"
 
-LIST_SQRT_TESTS(float16, LIBC_NAMESPACE::sqrtf16)
+LIST_SQRT_TESTS(float16, LIBC_NAMESPACE::sqrtf16);
diff --git a/libc/test/src/math/smoke/sqrtf_test.cpp b/libc/test/src/math/smoke/sqrtf_test.cpp
index 3f2e973325bd0..888b6cbdd643c 100644
--- a/libc/test/src/math/smoke/sqrtf_test.cpp
+++ b/libc/test/src/math/smoke/sqrtf_test.cpp
@@ -10,4 +10,4 @@
 
 #include "src/math/sqrtf.h"
 
-LIST_SQRT_TESTS(float, LIBC_NAMESPACE::sqrtf)
+LIST_SQRT_TESTS(float, LIBC_NAMESPACE::sqrtf);
diff --git a/libc/test/src/math/smoke/sqrtl_test.cpp b/libc/test/src/math/smoke/sqrtl_test.cpp
index f80bcfb736078..4f4a64f81ab7f 100644
--- a/libc/test/src/math/smoke/sqrtl_test.cpp
+++ b/libc/test/src/math/smoke/sqrtl_test.cpp
@@ -10,4 +10,4 @@
 
 #include "src/math/sqrtl.h"
 
-LIST_SQRT_TESTS(long double, LIBC_NAMESPACE::sqrtl)
+LIST_SQRT_TESTS(long double, LIBC_NAMESPACE::sqrtl);

>From b4d2045b5042db856b7ccee56281d958f116e918 Mon Sep 17 00:00:00 2001
From: Tue Ly <lntue.h at gmail.com>
Date: Tue, 11 Feb 2025 17:07:10 +0000
Subject: [PATCH 2/6] Address comments.

---
 libc/src/math/generic/sqrtf128.cpp | 81 ++++++++++++++++++++----------
 1 file changed, 55 insertions(+), 26 deletions(-)

diff --git a/libc/src/math/generic/sqrtf128.cpp b/libc/src/math/generic/sqrtf128.cpp
index 79f1681a9e871..73b34dbb69af6 100644
--- a/libc/src/math/generic/sqrtf128.cpp
+++ b/libc/src/math/generic/sqrtf128.cpp
@@ -1,7 +1,5 @@
 //===-- Implementation of sqrtf128 function -------------------------------===//
 //
-// Copyright (c) 2024 Alexei Sibidanov <sibid at uvic.ca>
-//
 // 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
@@ -17,6 +15,35 @@
 #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 at uvic.ca):
+// 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<float128>;
@@ -35,11 +62,11 @@ inline constexpr uint64_t prod_hi<uint64_t>(uint64_t x, uint64_t y) {
 
 // Get high part of unsigned 128x64 bit multiplication.
 template <>
-inline constexpr UInt128 prod_hi<UInt128, uint64_t>(UInt128 y, uint64_t x) {
-  uint64_t y_lo = static_cast<uint64_t>(y);
-  uint64_t y_hi = static_cast<uint64_t>(y >> 64);
-  UInt128 xyl = static_cast<UInt128>(x) * static_cast<UInt128>(y_lo);
-  UInt128 xyh = static_cast<UInt128>(x) * static_cast<UInt128>(y_hi);
+inline constexpr UInt128 prod_hi<UInt128, uint64_t>(UInt128 x, uint64_t y) {
+  uint64_t x_lo = static_cast<uint64_t>(x);
+  uint64_t x_hi = static_cast<uint64_t>(x >> 64);
+  UInt128 xyl = static_cast<UInt128>(x_lo) * static_cast<UInt128>(y);
+  UInt128 xyh = static_cast<UInt128>(x_hi) * static_cast<UInt128>(y);
   return xyh + (xyl >> 64);
 }
 
@@ -178,11 +205,11 @@ LIBC_INLINE uint64_t rsqrt_approx(uint64_t m) {
   //   r1 = r0 - r0 * h / 2
   // which has error bounded by:
   //   |r1 - 1/sqrt(x)| < h^2 / 2.
-  uint64_t r2 = prod_hi<uint64_t>(r, r);
+  uint64_t r2 = prod_hi(r, r);
   // h = r0^2*x - 1.
-  int64_t h = static_cast<int64_t>(prod_hi<uint64_t>(m, r2) + r2);
+  int64_t h = static_cast<int64_t>(prod_hi(m, r2) + r2);
   // hr = r * h / 2
-  int64_t hr = prod_hi<int64_t>(h, static_cast<int64_t>(r >> 1));
+  int64_t hr = prod_hi(h, static_cast<int64_t>(r >> 1));
   r -= hr;
   // Adjust in the unlucky case x~1;
   if (LIBC_UNLIKELY(!r))
@@ -224,8 +251,10 @@ LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) {
       fputil::raise_except_if_required(FE_INVALID);
       return xbits.quiet_nan().get_val();
     }
-    // x is subnormal or x=+0
-    if (x == 0)
+    // Now x is subnormal or x = +0.
+
+    // x is +0.
+    if (x_u == 0)
       return x;
 
     // Normalize subnormal inputs.
@@ -253,7 +282,7 @@ LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) {
                                    0xb504f333f9de6484 /* 2^64/sqrt(2) */};
 
   // Approximate 1/sqrt(1 + x_frac)
-  // Error: |r_1 - 1/sqrt(x)| < 2^-63.
+  // Error: |r_1 - 1/sqrt(x)| < 2^-62.
   uint64_t r1 = rsqrt_approx(static_cast<uint64_t>(x_frac >> 64));
   // Adjust for the even/odd exponent.
   uint64_t r2 = prod_hi(r1, RSQRT_2[i]);
@@ -279,8 +308,9 @@ LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) {
   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.
+  // the rounding mode.
+  // Check the lowest 14 bits (by clearing and sign-extending the top
+  // 32 - 14 = 18 bits).
   int dd = (static_cast<int>(v) << 18) >> 18;
 
   if (LIBC_UNLIKELY(dd < 4 && dd >= -8)) { // can round correctly?
@@ -289,9 +319,8 @@ LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) {
     // compare with the initial argument.
     UInt128 m = v >> 15;
     UInt128 m2 = m * m;
-    Int128 t0, t1;
     // The difference of the squared result and the argument
-    t0 = static_cast<Int128>(m2 - (x_reduced << 98));
+    Int128 t0 = static_cast<Int128>(m2 - (x_reduced << 98));
     if (t0 == 0) {
       // the square root is exact
       v = m << 15;
@@ -299,7 +328,7 @@ LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) {
       // 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.
-      t1 = t0;
+      Int128 t1 = t0;
       Int128 sgn = t0 >> 127; // sign of the difference
       t1 -= (m << 1) ^ sgn;
       t1 += 1 + sgn;
@@ -332,20 +361,20 @@ LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) {
     rnd = frac >> 14; // round to nearest tie to even
   } else if (rm == FE_UPWARD) {
     rnd = !!frac; // round up
-  } else if (rm == FE_DOWNWARD) {
-    rnd = 0; // round down
   } else {
-    rnd = 0; // round to zero
+    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.
+  // 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<UInt128>(e2) << FPBits::FRACTION_LEN; // place exponent

>From 18a9a58ef2758b043f77973a876a570056c4c4b6 Mon Sep 17 00:00:00 2001
From: Tue Ly <lntue.h at gmail.com>
Date: Tue, 11 Feb 2025 18:13:01 +0000
Subject: [PATCH 3/6] Add small table option.

---
 libc/src/math/generic/sqrtf128.cpp | 78 ++++++++++++++++++++++++------
 1 file changed, 63 insertions(+), 15 deletions(-)

diff --git a/libc/src/math/generic/sqrtf128.cpp b/libc/src/math/generic/sqrtf128.cpp
index 73b34dbb69af6..47a349023bbad 100644
--- a/libc/src/math/generic/sqrtf128.cpp
+++ b/libc/src/math/generic/sqrtf128.cpp
@@ -101,6 +101,65 @@ inline constexpr Int128 prod_hi<Int128, UInt128>(Int128 x, UInt128 y) {
   return static_cast<Int128>(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<int64_t>(prod_hi(m, r2) + r2);
+  // hr = r * h / 2
+  int64_t hr = prod_hi(h, static_cast<int64_t>(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<uint64_t>(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<uint64_t>(z) * static_cast<uint64_t>(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<uint64_t>((x >> 34) * static_cast<int64_t>(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},
@@ -196,26 +255,15 @@ LIBC_INLINE uint64_t rsqrt_approx(uint64_t m) {
   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:
-  // 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.
-  uint64_t r2 = prod_hi(r, r);
-  // h = r0^2*x - 1.
-  int64_t h = static_cast<int64_t>(prod_hi(m, r2) + r2);
-  // hr = r * h / 2
-  int64_t hr = prod_hi(h, static_cast<int64_t>(r >> 1));
-  r -= hr;
+  // 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
 
@@ -254,7 +302,7 @@ LLVM_LIBC_FUNCTION(float128, sqrtf128, (float128 x)) {
     // Now x is subnormal or x = +0.
 
     // x is +0.
-    if (x_u == 0)
+    if (x_frac == 0)
       return x;
 
     // Normalize subnormal inputs.

>From c7243300fdbfbc8aa98e90b87b7e2822f7c50785 Mon Sep 17 00:00:00 2001
From: Tue Ly <lntue.h at gmail.com>
Date: Tue, 11 Feb 2025 18:23:45 +0000
Subject: [PATCH 4/6] Address comment on tests.

---
 libc/test/src/math/smoke/sqrtf128_test.cpp | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/libc/test/src/math/smoke/sqrtf128_test.cpp b/libc/test/src/math/smoke/sqrtf128_test.cpp
index d4cd1be5452d7..0c0456851013b 100644
--- a/libc/test/src/math/smoke/sqrtf128_test.cpp
+++ b/libc/test/src/math/smoke/sqrtf128_test.cpp
@@ -122,16 +122,18 @@ TEST_F(LlvmLibcSqrtTest, HardToRound) {
 
   // Check exact cases starting from small numbers
   for (unsigned k = 1; k < 100 * 100; ++k) {
-    unsigned kx = k * k;
-    float128 x = kx, y = k;
+    unsigned k2 = k * k;
+    float128 x = static_cast<float128>(k2);
+    float128 y = static_cast<float128>(k);
     EXPECT_FP_EQ_ALL_ROUNDING(y, LIBC_NAMESPACE::sqrtf128(x));
   };
 
   // Then from the largest number.
   uint64_t k0 = 101904826760412362ULL;
   for (uint64_t k = k0; k > k0 - 10000; --k) {
-    UInt128 kx = static_cast<UInt128>(k) * static_cast<UInt128>(k);
-    float128 x = kx, y = k;
+    UInt128 k2 = static_cast<UInt128>(k) * static_cast<UInt128>(k);
+    float128 x = static_cast<float128>(k2);
+    float128 y = static_cast<float128>(k);
     EXPECT_FP_EQ_ALL_ROUNDING(y, LIBC_NAMESPACE::sqrtf128(x));
   }
 }

>From 05cc1dc6df606e5911286db3df58909ba82a06f9 Mon Sep 17 00:00:00 2001
From: Tue Ly <lntue.h at gmail.com>
Date: Tue, 11 Feb 2025 20:24:04 +0000
Subject: [PATCH 5/6] Address comments.

---
 libc/src/math/generic/sqrtf128.cpp                    | 2 ++
 libc/test/src/math/performance_testing/CMakeLists.txt | 2 --
 libc/test/src/math/smoke/sqrtf128_test.cpp            | 2 --
 3 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/libc/src/math/generic/sqrtf128.cpp b/libc/src/math/generic/sqrtf128.cpp
index 47a349023bbad..7e473327c727c 100644
--- a/libc/src/math/generic/sqrtf128.cpp
+++ b/libc/src/math/generic/sqrtf128.cpp
@@ -17,6 +17,8 @@
 
 // Compute sqrtf128 with correct rounding for all rounding modes using integer
 // arithmetic by Alexei Sibidanov (sibid at uvic.ca):
+//   https://github.com/sibidanov/llvm-project/tree/as_sqrt_v2
+//   https://github.com/sibidanov/llvm-project/tree/as_sqrt_v3
 // Let the input be expressed as x = 2^e * m_x,
 // - Step 1: Range reduction
 //   Let x_reduced = 2^(e % 2) * m_x,
diff --git a/libc/test/src/math/performance_testing/CMakeLists.txt b/libc/test/src/math/performance_testing/CMakeLists.txt
index 862a0f6d32386..838ed9e957ca7 100644
--- a/libc/test/src/math/performance_testing/CMakeLists.txt
+++ b/libc/test/src/math/performance_testing/CMakeLists.txt
@@ -508,6 +508,4 @@ add_perf_binary(
   DEPENDS
     .single_input_single_output_diff
     libc.src.math.sqrtf128
-  COMPILE_OPTIONS
-  -fno-builtin
 )
diff --git a/libc/test/src/math/smoke/sqrtf128_test.cpp b/libc/test/src/math/smoke/sqrtf128_test.cpp
index 0c0456851013b..3b9686c4ea477 100644
--- a/libc/test/src/math/smoke/sqrtf128_test.cpp
+++ b/libc/test/src/math/smoke/sqrtf128_test.cpp
@@ -1,7 +1,5 @@
 //===-- Unittests for sqrtf128---------------------------------------------===//
 //
-// Copyright (c) 2024 Alexei Sibidanov <sibid at uvic.ca>
-//
 // 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

>From 9b17e18e993947a7a7c8c81623bc25934efdb16c Mon Sep 17 00:00:00 2001
From: Tue Ly <lntue.h at gmail.com>
Date: Tue, 11 Feb 2025 21:03:04 +0000
Subject: [PATCH 6/6] Add TODO.

---
 libc/src/math/generic/sqrtf128.cpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/libc/src/math/generic/sqrtf128.cpp b/libc/src/math/generic/sqrtf128.cpp
index 7e473327c727c..c844d3afa11c8 100644
--- a/libc/src/math/generic/sqrtf128.cpp
+++ b/libc/src/math/generic/sqrtf128.cpp
@@ -19,6 +19,9 @@
 // arithmetic by Alexei Sibidanov (sibid at 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,



More information about the libc-commits mailing list