Recalculate lerp if we got infinity. Eliminates some overflows.

[Igor Zhukov (fsb4000)]

Fixes #1917

[Igor Zhukov (fsb4000)]

statementreply

It breaks monotonicity.

Thank you for finding some issues. I fixed some but I have no idea how to restore monotonicity for lerp :(

[statementreply]

I fixed some but I have no idea how to restore monotonicity for lerp :(

Never mind. It can be tricky to implement floating point functionalities when there are requirements on precision, handling of extreme values, and/or preservation of math properties.

I'm going to open a PR (maybe on the weekend) with an alternative approach to fix the false overflow. We need to investigate whether there's a simple-ish fix for the original issue.

[Alexander Bolz (abolz)]

FWIW, this could be fixed with a fused multiply-add

--const auto _Candidate = _ArgA + _ArgT * (_ArgB - _ArgA);
++const auto _Candidate = std::fma(_ArgT, _ArgB - _ArgA, _ArgA);

But fma is not constexpr.

[Igor Zhukov (fsb4000)]

statementreply

Well, at least we can reuse new test code :(

What do you think about fma? It seems it doesn't overflow but we can get different results for runtime/compiletime invocation of lerp...

https://gcc.godbolt.org/z/T564z58qo

[Matt Stephanson (MattStephanson)]

The condition we have to handle here is limited to |t*(b-a)| ≤ 2*MEOW_MAX, right? Otherwise we'd have a true, not spurious, overflow. If that's the case, maybe we can simply scale the calculation by a fixed factor.

auto _Smaller     = _ArgT;
auto _Larger      = _ArgB - _ArgA;
auto _Abs_smaller = _Float_abs(_Smaller);
auto _Abs_larger  = _Float_abs(_Larger);
if (_Abs_larger < _Abs_smaller) {
    _STD swap(_Smaller, _Larger);
    _STD swap(_Abs_smaller, _Abs_larger);
}

if (_Abs_larger > _STD sqrt(numeric_limits<_Ty>::max()) && _Abs_smaller > 1) {
    return 2 * (_Ty{0.5} * _ArgA + _Smaller * (_Ty{0.5} * _Larger));
} else {
    return _ArgA + _Smaller * _Larger;
}

_Larger is too large to be subnormal, so scaling by 0.5 is exact, and the product _Smaller * _Larger is large enough that if _ArgA is subnormal, it will be too small to contribute anyway. So I think the two expressions should be equivalent, apart from internal overflow, so there won't be a loss of monotonicity when switching from one to the other.

[statementreply]

What do you think about fma? It seems it doesn't overflow but we can get different results for runtime/compiletime invocation of lerp...

• Accuracy

  fma(t, b - a, a) is (on average) more accurate than a + t * (b - a), and doesn't suffer from t * (b - a) overflowing as long as the final result doesn't overflow. (1 - t) * a + t * b in the other branch could also be fma(t, b, fma(-t, a, a)).

• Performance

  Recent CPUs (AMD since ~2012, Intel since ~2014, and all ARM that Windows arm64 supports) have hardware fma instructions. Older CPUs don't support hardware fma , so std::fma needs to be implemented in software.

  On the current version of MSVC, if not compiled with /arch:AVX2 or alike, std::fma isn't inlined and generates a call into UCRT (https://godbolt.org/z/q4vbe7xqn). It then checks whether the CPU supports fma. If so, it calls the CPU instruction directly (however the function call overhead is non-trivial). If not, UCRT uses another algorithm to compute the result, which is unfortunately neither accurate (DevCom-242309) nor fast (hundreds of nanoseconds for some input values) as of ucrtbase.dll 10.0.19041.789.

• Different results between compile time and runtime

  There are already situations where results could differ.

  functions (exp, sin, etc.) use fma powered implementation on CPUs with hardware fma, and a different implementation without fma on CPUs without hardware fma. Code using these math functions could generate different results on different CPUs.

  GCC performs constant folding (not limited to constexpr) on math functions. The compile time results could be different from the results from the runtime math library. (https://godbolt.org/z/r49WEax6E)

  Also, the compiler is allowed to compile x * y + z into fma(x, y, z) (https://godbolt.org/z/zaY6veW9j), which may also lead to different results between compile time and runtime.

  However, should the results differ between compile time and runtime, it's desirable that the compile time results are correctly rounded (as if computed with infinite precision and then rounded), or at least more accurate than the runtime results.

[statementreply]

The condition we have to handle here is limited to |t*(b-a)| ≤ 2*MEOW_MAX, right? Otherwise we'd have a true, not spurious, overflow. If that's the case, maybe we can simply scale the calculation by a fixed factor.

auto _Smaller     = _ArgT;
auto _Larger      = _ArgB - _ArgA;
auto _Abs_smaller = _Float_abs(_Smaller);
auto _Abs_larger  = _Float_abs(_Larger);
if (_Abs_larger < _Abs_smaller) {
    _STD swap(_Smaller, _Larger);
    _STD swap(_Abs_smaller, _Abs_larger);
}

if (_Abs_larger > _STD sqrt(numeric_limits<_Ty>::max()) && _Abs_smaller > 1) {
    return 2 * (_Ty{0.5} * _ArgA + _Smaller * (_Ty{0.5} * _Larger));
} else {
    return _ArgA + _Smaller * _Larger;
}

_Larger is too large to be subnormal, so scaling by 0.5 is exact, and the product _Smaller * _Larger is large enough that if _ArgA is subnormal, it will be too small to contribute anyway. So I think the two expressions should be equivalent, apart from internal overflow, so there won't be a loss of monotonicity when switching from one to the other.

Note that we have |b - a| <= MEOW_MAX in this branch, so it is limited to |t| > 1. I think we could do the following to save a few branches (untested, needs benchmark):

// precondition: T{0.5} * t is exact
T linear_for_lerp(T intercept, T slope, T t) {
    T half_prod = slope * (T{0.5} * t);
    if (abs(half_prod) <= numeric_limits::max() * T{0.5}) {
        return intercept + slope * t;
    } else {
        return T{2} * (T{0.5} * intercept + half_prod);
    }
}

T lerp(T a, T b, T t) {
    // [...]
    {
        if (abs(t) < T{1}) {
            T candidate = a + t * (b - a);
            // [...] fix monotonicity
        }

        if (t >= T{1}) {
            return linear_for_lerp(b, b - a, t - T{1});
        }

        // t <= -1
        return linear_for_lerp(a, b - a, t);
    }
    // [[...]]
}

[Stephan T. Lavavej (StephanTLavavej)]

I'm speculatively mirroring this to the MSVC-internal repo. Further changes can be pushed, but please notify me.

[Casey Carter (CaseyCarter)]

We need to find that Twitter thread mocking our huge implementation of lerp to let them know we've added another 25 lines. (No change requested.)

[Stephan T. Lavavej (StephanTLavavej)]

Thanks for infinitely improving this implementation! ♾️ 🚀 😹