[libc] Improve the performance of expm1f.

[lntue]

Summary

Improve the performance of expm1f:

• Rearrange the selection logic for different cases to improve the overall

throughput.

• Use the same degree-4 polynomial for large inputs as expf

(https://reviews.llvm.org/D122418), reduced from a degree-7 polynomial.

Performance benchmark using perf tool from CORE-MATH project
(https://gitlab.inria.fr/core-math/core-math/-/tree/master):
Before this patch:

$ ./perf.sh expm1f

CORE-MATH reciprocal throughput   : 15.362
System LIBC reciprocal throughput : 53.288
LIBC reciprocal throughput        : 54.572

$ ./perf.sh expm1f --latency

CORE-MATH latency   : 57.759
System LIBC latency : 147.146
LIBC latency        : 118.057

After this patch:

$ ./perf.sh expm1f

CORE-MATH reciprocal throughput   : 15.359
System LIBC reciprocal throughput : 53.188
LIBC reciprocal throughput        : 14.600

$ ./perf.sh expm1f --latency

CORE-MATH latency   : 57.774
System LIBC latency : 147.119
LIBC latency        : 60.280

[lntue]

Improve performance for exceptional cases.

[lntue]

Final revision.

[lntue]

Remove redundant check.

[zimmermann6]

all values are still correctly rounded, and I confirm the speed improvement:

zimmerma@tomate:/tmp/core-math$ LIBM=/users/zimmerma/svn/core-math/libllvmlibc.a ./perf.sh expm1f # previous code
22.692
54.039
53.218
zimmerma@tomate:/tmp/core-math$ LIBM=/tmp/libllvmlibc.a ./perf.sh expm1f # new code
22.698
54.037
17.240

The llvm-libm results are the 3rd ones (first CORE-MATH, second GNU libc).

[michaelrj]

LGTM

[santoshn]

Looks good to me.

libc/src/math/generic/expm1f.cpp

if (xbits.uintval() < 0x7f80'0000U) {
 
       errno = ERANGE;
     }
     return x + static_cast<float>(FPBits::inf());
   }
 
   int unbiased_exponent = static_cast<int>(xbits.get_unbiased_exponent());
   // |x| < 2^-4
-  if (unbiased_exponent < 123) {
+  if (unbiased_exponent < 122) {
     // |x| < 2^-25
     if (unbiased_exponent < 102) {
       // x = -0.0f
       if (unlikely(xbits.uintval() == 0x8000'0000U))
         return x;
       // When |x| < 2^-25, the relative error:
       //   |(e^x - 1) - x| / |x| < |x^2| / |x| = |x| < 2^-25 < epsilon(1)/2.
       // So the correctly rounded values of expm1(x) are:
       //   = x + eps(x) if rounding mode = FE_UPWARD,
       //                   or (rounding mode = FE_TOWARDZERO and x is negative),
       //   = x otherwise.
       // To simplify the rounding decision and make it more efficient, we use
       //   fma(x, x, x) ~ x + x^2 instead.
       return fputil::fma(x, x, x);
     }
-    // 2^-25 <= |x| < 2^-4
+    // // 2^-25 <= |x| < 2^-4
     double xd = static_cast<double>(x);
     double xsq = xd * xd;
     // Degree-8 minimax polynomial generated by Sollya with:
     // > display = hexadecimal;
     // > P = fpminimax(expm1(x)/x, 7, [|D...|], [-2^-4, 2^-4]);
     double r =
-        fputil::polyeval(xd, 0x1p-1, 0x1.55555555559abp-3, 0x1.55555555551a7p-5,
-                         0x1.111110f70f2a4p-7, 0x1.6c16c17639e82p-10,
-                         0x1.a02526febbea6p-13, 0x1.a01dc40888fcdp-16);
+        // fputil::polyeval(xd, 0x1p-1, 0x1.55555555559abp-3,
+        // 0x1.55555555551a7p-5,
+        //                  0x1.111110f70f2a4p-7, 0x1.6c16c17639e82p-10,
+        //                  0x1.a02526febbea6p-13, 0x1.a01dc40888fcdp-16);
+        fputil::polyeval(xd, 0x1p-1, 0x1.555555555554ep-3, 0x1.5555555451452p-5,
+                         0x1.111111118667p-7, 0x1.6c18c98d68ca7p-10,
+                         0x1.a01ae76480441p-13);
+    // return static_cast<float>(xd*r);
     return static_cast<float>(fputil::fma(r, xsq, xd));
   }
 
   // For -18 < x < 89, to compute exp(x), we perform the following range
   // reduction: find hi, mid, lo such that:
   //   x = hi + mid + lo, in which
   //     hi is an integer,
   //     mid * 2^7 is an integer
   //     -2^(-8) <= lo < 2^-8.
   // In particular,
   //   hi + mid = round(x * 2^7) * 2^(-7).
   // Then,
   //   exp(x) = exp(hi + mid + lo) = exp(hi) * exp(mid) * exp(lo).
   // We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2
   // respectively.  exp(lo) is computed using a degree-7 minimax polynomial
   // generated by Sollya.
 
   // Exceptional value
-  if (xbits.uintval() == 0xbdc1'c6cbU) {
+  if (x == 0x1.6b7d8ap-3f) {
     // x = -0x1.838d96p-4f
     int round_mode = fputil::get_round();
-    if (round_mode == FE_TONEAREST || round_mode == FE_DOWNWARD)
-      return -0x1.71c884p-4f;
-    return -0x1.71c882p-4f;
+    if (round_mode == FE_TONEAREST || round_mode == FE_UPWARD)
+      return 0x1.8dbe64p-3f;
+    return 0x1.8dbe62p-3f;
+    // int round_mode = fputil::get_round();
+    // if (round_mode == FE_TONEAREST || round_mode == FE_DOWNWARD)
+    //   return -0x1.71c884p-4f;
+    // return -0x1.71c882p-4f;
   }
 
   // x_hi = hi + mid.
-  int x_hi = static_cast<int>(x * 0x1.0p7f);
+  int x_hi = static_cast<int>(x * 0x1.0p7f + (xbits.get_sign() ? -0.5f : 0.5f));
   // Subtract (hi + mid) from x to get lo.
   x -= static_cast<float>(x_hi) * 0x1.0p-7f;
   double xd = static_cast<double>(x);
-  // Make sure that -2^(-8) <= lo < 2^-8.
-  if (x >= 0x1.0p-8f) {
-    ++x_hi;
-    xd -= 0x1.0p-7;
-  }
-  if (x < -0x1.0p-8f) {
-    --x_hi;
-    xd += 0x1.0p-7;
-  }
   x_hi += 104 << 7;
   // hi = x_hi >> 7
   double exp_hi = EXP_M1[x_hi >> 7];
   // lo = x_hi & 0x0000'007fU;
   double exp_mid = EXP_M2[x_hi & 0x7f];
   double exp_hi_mid = exp_hi * exp_mid;
+  double expm1_hi_mid = fputil::fma(exp_hi, exp_mid, -1.0);
   // Degree-7 minimax polynomial generated by Sollya with the following
   // commands:
   //   > display = hexadecimal;
   //   > Q = fpminimax(expm1(x)/x, 6, [|D...|], [-2^-8, 2^-8]);
   //   > Q;
-  double exp_lo = fputil::polyeval(
-      xd, 0x1p0, 0x1p0, 0x1p-1, 0x1.5555555555555p-3, 0x1.55555555553ap-5,
-      0x1.1111111204dfcp-7, 0x1.6c16cb2da593ap-10, 0x1.9ff1648996d2ep-13);
-  return static_cast<float>(fputil::fma(exp_hi_mid, exp_lo, -1.0));
+  double expm1_lo =
+      xd * fputil::polyeval(xd, 0x1.ffffffffff777p-1, 0x1.000000000071cp-1,
+                            0x1.555566668e5e7p-3, 0x1.55555555ef243p-5);
+  return static_cast<float>(fputil::fma(exp_hi_mid, expm1_lo, expm1_hi_mid));
+  // return static_cast<float>(fputil::fma(exp_hi_mid, expm1_lo, -1.0));
 }
 
 } // namespace __llvm_libc

libc/test/src/math/exhaustive/exhaustive_test.cpp

 #include <functional>
 #include <iostream>
 #include <sstream>
 #include <string>
 #include <thread>
 #include <vector>
 
 #include "exhaustive_test.h"
+#include "src/__support/FPUtil/FPBits.h"
 #include "utils/UnitTest/Test.h"
 
 template <typename T>
 void LlvmLibcExhaustiveTest<T>::test_full_range(T start, T stop, int nthreads,
                                                 mpfr::RoundingMode rounding) {
   std::vector<std::thread> thread_list(nthreads);
   T increment = (stop - start - 1) / nthreads + 1;
   T begin = start;
   T end = start + increment - 1;
   for (int i = 0; i < nthreads; ++i) {
     thread_list.emplace_back([this, begin, end, rounding]() {
       std::stringstream msg;
       msg << "-- Testing from " << begin << " to " << end << " [0x" << std::hex
-          << begin << ", 0x" << end << ") ..." << std::endl;
+          << begin << ", 0x" << end << "), [" << std::hexfloat
+          << float(__llvm_libc::fputil::FPBits<float>(
+                 static_cast<uint32_t>(begin)))
+          << ", "
+          << float(
+                 __llvm_libc::fputil::FPBits<float>(static_cast<uint32_t>(end)))
+          << ") ..." << std::endl;
       std::cout << msg.str();
       msg.str("");
 
       bool result = check(begin, end, rounding);
 
       msg << "** Finished testing from " << std::dec << begin << " to " << end
-          << " [0x" << std::hex << begin << ", 0x" << end
+          << " [0x" << std::hex << begin << ", 0x" << end << "), ["
+          << std::hexfloat
+          << float(__llvm_libc::fputil::FPBits<float>(
+                 static_cast<uint32_t>(begin)))
+          << ", "
+          << float(
+                 __llvm_libc::fputil::FPBits<float>(static_cast<uint32_t>(end)))
           << ") : " << (result ? "PASSED" : "FAILED") << std::endl;
       std::cout << msg.str();
     });
     begin += increment;
     end += increment;
     if (end > stop)
       end = stop;
   }

libc/test/src/math/exhaustive/expm1f_test.cpp

 //===-- Exhaustive test for expm1f ----------------------------------------===//
 //
 // 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 "exhaustive_test.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/math/expm1f.h"
 #include "utils/MPFRWrapper/MPFRUtils.h"
 #include "utils/UnitTest/FPMatcher.h"
 
+#include <thread>
+
 using FPBits = __llvm_libc::fputil::FPBits<float>;
 
 namespace mpfr = __llvm_libc::testing::mpfr;
 
 struct LlvmLibcExpfExhaustiveTest : public LlvmLibcExhaustiveTest<uint32_t> {
   bool check(uint32_t start, uint32_t stop,
              mpfr::RoundingMode rounding) override {
     mpfr::ForceRoundingMode r(rounding);
-    uint32_t bits = start;
+    uint32_t bits = stop;
+    int tolerance = 2;
     bool result = true;
     do {
       FPBits xbits(bits);
       float x = float(xbits);
       result &= EXPECT_MPFR_MATCH(mpfr::Operation::Expm1, x,
                                   __llvm_libc::expm1f(x), 0.5, rounding);
-    } while (bits++ < stop);
+      if (!result)
+        --tolerance;
+      if (tolerance == 0)
+        break;
+    } while (bits-- > start);
+    return result;
   }
 };
 
-static constexpr int NUM_THREADS = 16;
+static const int NUM_THREADS = std::thread::hardware_concurrency();
 
 // Range: [0, 89];
 static constexpr uint32_t POS_START = 0x0000'0000U;
 static constexpr uint32_t POS_STOP = 0x42b2'0000U;
 
 TEST_F(LlvmLibcExpfExhaustiveTest, PostiveRangeRoundNearestTieToEven) {
   test_full_range(POS_START, POS_STOP, NUM_THREADS,
                   mpfr::RoundingMode::Nearest);