[libc] Improve hypotf performance with different algorithm correctly rounded to all rounding modes.

[lntue]

Summary

Algorithm for hypotf: compute (a*a + b*b) in double precision, then use Dekker's algorithm to find the rounding error, and then correcting it after taking its square-root.

[sivachandra]

libc/src/math/generic/hypotf.cpp

22

Incorrect variable naming style at many places in this function.

32

A call to a target independent builtin for a standard function can lead to a call back to the libc. That is, compilers are free to call the sqrt function from the libc. We can refactor our sqrt implementation so that we can replace this call to __builtin_sqrt with a call to LLVM libc's sqrt.

[lntue]

libc/src/math/generic/hypotf.cpp

32

I refactor our sqrt implementation in https://reviews.llvm.org/D118173
Will wait for that patch to be landed.

[zimmermann6]

I get some errors for rounding to nearest:

Difference for 0x1.faf49ep+25,0x1.480002p+23
llvm_hypot: 0x1.00c5bp+26
as_hypot:   0x1.00c5b2p+26
pz_hypot:   0x1.00c5b2p+26

libc/src/math/generic/hypotf.cpp

29–30

I hadn't seen that trick to compute the rounding error, do you have a reference?
By the way, I'm not sure the reference to Dekker is appropriate. For me, Dekker's algorithm splits
two floating-point numbers in two each, and computes their product (high + low part) using 4 multiplies.

[vinc17]

libc/src/math/generic/hypotf.cpp

29–30

If I understand correctly, err should get the rounding error of the sum. The algorithm is known as TwoSum. It needs 6 operations, including the sum sumSq, and this is the same number of operations as you have. But with 6 add/sub operations, I proved that there is only one algorithm (up to obvious symmetries) that works. And the above one is different. Thus it will be sometimes incorrect. I think that if xSq and ySq are close to each other and their sum is not exact, then the above algorithm will give you twice the rounding error.

[lntue]

Remove the check for non-zero error.

[lntue]

In D118157#3272333, @zimmermann6 wrote:

I get some errors for rounding to nearest:

Difference for 0x1.faf49ep+25,0x1.480002p+23
llvm_hypot: 0x1.00c5bp+26
as_hypot:   0x1.00c5b2p+26
pz_hypot:   0x1.00c5b2p+26

Thanks Paul for finding the error! For this example, the sum square is actually exact, but the rounding errors still need to be updated in order to avoid double rounding errors. By removing the check (err != 0) in line 36, I got it back correctly.

[lntue]

libc/src/math/generic/hypotf.cpp

29–30

Thanks Paul and Vincent for finding the issue with this! Actually I was trying to implement the Fast2Sum version since we are in radix-2. Moreover, since both xSq and ySq are non-negative:

max(xSq, ySq) <= sumSq <= sqrt(2) max(xSq, ySq)

and so sumSq - max(xSq, ySq) is exact. I was trying to implement it without branch and thought that (sumSq - min(xSq, ySq)) - max(xSq, ySq) = 0, which could be easily disproved by the following example: Consider single precision with xSq = 1 + 2^(-23) (I know it's not a square) and ySq = 2^(-24) with default rounding mode, then sumSq = xSq + ySq = 1 + 2^(-22). Then:

sumSq - xSq = 2^(-23)  and  sumSq - ySq = 1 + 2^(-22) = sumSq

And hence:

(sumSq - xSq) - ySq = 2^-24
(sumSq - ySq) - xSq = 2^-23
((sumSq - xSq) - ySq) + ((sumSq - ySq) - xSq) != 2^(-24)  which is the rounding error.

I change it back to the normal Fast2Sum implementation with branching now, so the rounding error computation should be correct.

[zimmermann6]

I'm still running semi-exhaustive tests, it takes some time. I wonder whether a full exhaustive test is possible, by comparing the LLVM implementation with the code from Alexei at https://core-math.gitlabpages.inria.fr/. On a 64-core machine (Intel Xeon Gold 6130 @ 2.10GHz), it takes 4.6s to check 2^33 pairs (x,y). If one tests only positive x,y and x>=y, as exhaustive comparison would have to check 2^61 pairs for each rounding mode, which would take less than 1.5 month using 10000 such machines. This would not be a proof, but the probability that both codes are wrong for the same inputs and give exactly the same wrong answer is quite small.

[lntue]

In D118157#3278638, @zimmermann6 wrote:

I'm still running semi-exhaustive tests, it takes some time. I wonder whether a full exhaustive test is possible, by comparing the LLVM implementation with the code from Alexei at https://core-math.gitlabpages.inria.fr/. On a 64-core machine (Intel Xeon Gold 6130 @ 2.10GHz), it takes 4.6s to check 2^33 pairs (x,y). If one tests only positive x,y and x>=y, as exhaustive comparison would have to check 2^61 pairs for each rounding mode, which would take less than 1.5 month using 10000 such machines. This would not be a proof, but the probability that both codes are wrong for the same inputs and give exactly the same wrong answer is quite small.

Or if you don't mind to be slower, you can compare it with the shift-and-add algorithm implemented in the LLVM-libc that this one is trying to speed up, since that can be proved mathematically to be correct.

Another option is that since the idea of this algorithm is scalable, we can have a version of it for half precision (essentially just change the data types and masks/constants), where we it can be tested exhaustively? That should at least increase the confidence with single, and maybe double precision later?

[lntue]

Use fputil::sqrt instead of __builtin_sqrtf.

[lntue]

Fix variable names and ignore compiler warnings about C++17.

[lntue]

Add a quick return when the exponent difference is at least 2 more than the mantissa length.

[zimmermann6]

the version of last Friday is fine for me: I did run exhaustive tests for 2^23 <= y < 2^24, and 2^(23+k) <= x < 2^(24+k) for 0 <= k <= 13.
However since it changed in the meantime, I don't have resources any more to review the new version.

[lntue]

In D118157#3283205, @zimmermann6 wrote:

the version of last Friday is fine for me: I did run exhaustive tests for 2^23 <= y < 2^24, and 2^(23+k) <= x < 2^(24+k) for 0 <= k <= 13.
However since it changed in the meantime, I don't have resources any more to review the new version.

Thanks Paul for checking! The new version only changes when exponent of x >= exponent of y + 25, so your verification should still hold.

[lntue]

Add exhaustive testing to test for inputs (x, y) with 2^23 <= x < 2^24, 2^(23 + 14) <= y < 2^(23 + 25).

[lntue]

Fix names in the exhaustive test.

[sivachandra]

OK for the structuring aspects.

[lntue]

In D118157#3283205, @zimmermann6 wrote:

the version of last Friday is fine for me: I did run exhaustive tests for 2^23 <= y < 2^24, and 2^(23+k) <= x < 2^(24+k) for 0 <= k <= 13.
However since it changed in the meantime, I don't have resources any more to review the new version.

@zimmermann6 : I've finished testing the remaining pairs: (x, y) with 2^23 <= y < 2^24, and 2^(23+k) <= x < 2^(24+k) for 14 <= k <= 24.

[lntue]

Sync to HEAD.

libc/src/math/generic/hypotf.cpp

 //===-- Implementation of hypotf 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/hypotf.h"
-#include "src/__support/FPUtil/Hypot.h"
+#include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/common.h"
 
 namespace __llvm_libc {
 
 LLVM_LIBC_FUNCTION(float, hypotf, (float x, float y)) {
-  return __llvm_libc::fputil::hypot(x, y);
+  using DoubleBits = fputil::FPBits<double>;
+  using FPBits = fputil::FPBits<float>;
+
+  double xd = static_cast<double>(x);
+  double yd = static_cast<double>(y);
+
+  // These squares are exact.
+  double xSq = xd * xd;
+  double ySq = yd * yd;
+
+  // Compute the sum of squares.
+  double sumSq = xSq + ySq;
+
+  // Compute the rounding error with Fast2Sum algorithm:
+  // xSq + ySq = sumSq - err
+  double err = (xSq >= ySq) ? (sumSq - xSq) - ySq : (sumSq - ySq) - xSq;
+
+  // Take sqrt in double precision.
+  DoubleBits result(__builtin_sqrt(sumSq));
+
+  if (!DoubleBits(sumSq).is_inf_or_nan()) {
+    // Correct rounding.
+    double rSq = static_cast<double>(result) * static_cast<double>(result);
+    double diff = sumSq - rSq;
+    constexpr uint64_t mask = 0x0000'0000'3FFF'FFFFULL;
+    uint64_t lrs = result.uintval() & mask;
+
+    if (lrs == 0x0000'0000'1000'0000ULL && err < diff) {
+      result.bits |= 1ULL;
+    } else if (lrs == 0x0000'0000'3000'0000ULL && err > diff) {
+      result.bits -= 1ULL;
+    }
+  } else {
+    FPBits bits_x(x), bits_y(y);
+    if (bits_x.is_inf_or_nan() || bits_y.is_inf_or_nan()) {
+      if (bits_x.is_inf() || bits_y.is_inf())
+        return static_cast<float>(FPBits::inf());
+      if (bits_x.is_nan())
+        return x;
+      return y;
+    }
+  }
+
+  return static_cast<float>(static_cast<double>(result));
 }
 
 } // namespace __llvm_libc

libc/test/src/math/hypotf_hard_to_round.h

 //===-- Hard-to-round inputs for hypotf ------------------------------C++--===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIBC_TEST_SRC_MATH_HYPOTTEST_HARD_TO_ROUND_H
 #define LLVM_LIBC_TEST_SRC_MATH_HYPOTTEST_HARD_TO_ROUND_H
 
 #include "utils/MPFRWrapper/MPFRUtils.h"
 
 namespace mpfr = __llvm_libc::testing::mpfr;
 
-constexpr int N_HARD_TO_ROUND = 1216;
+constexpr int N_HARD_TO_ROUND = 1217;
 constexpr mpfr::BinaryInput<float> HYPOTF_HARD_TO_ROUND[N_HARD_TO_ROUND] = {
+    {0x1.faf49ep+25f, 0x1.480002p+23f},
     {0x1.ffffecp-1f, 0x1.000002p+27},
     {0x1.900004p+34, 0x1.400002p+23}, /* 45 identical bits */
     {0x1.05555p+34, 0x1.bffffep+23},  /* 44 identical bits */
     {0x1.e5fffap+34, 0x1.affffep+23}, /* 45 identical bits */
     {0x1.260002p+34, 0x1.500002p+23}, /* 45 identical bits */
     {0x1.fffffap+34, 0x1.fffffep+23}, /* 45 identical bits */
     {0x1.8ffffap+34, 0x1.3ffffep+23}, /* 45 identical bits */
     {0x1.87fffcp+35, 0x1.bffffep+23}, /* 47 identical bits */