fma_test2

Percentage Accurate: 34.4% → 99.6%
Time: 5.8s
Alternatives: 3
Speedup: 1.7×

Specification

?
\[1.9 \leq t \land t \leq 2.1\]
\[\begin{array}{l} \\ 1.7 \cdot 10^{+308} \cdot t - 1.7 \cdot 10^{+308} \end{array} \]
(FPCore (t) :precision binary64 (- (* 1.7e+308 t) 1.7e+308))
double code(double t) {
	return (1.7e+308 * t) - 1.7e+308;
}

real(8) function code(t)
    real(8), intent (in) :: t
    code = (1.7d+308 * t) - 1.7d+308
end function

public static double code(double t) {
	return (1.7e+308 * t) - 1.7e+308;
}

def code(t):
	return (1.7e+308 * t) - 1.7e+308

function code(t)
	return Float64(Float64(1.7e+308 * t) - 1.7e+308)
end

function tmp = code(t)
	tmp = (1.7e+308 * t) - 1.7e+308;
end

code[t_] := N[(N[(1.7e+308 * t), $MachinePrecision] - 1.7e+308), $MachinePrecision]

\begin{array}{l}

\\
1.7 \cdot 10^{+308} \cdot t - 1.7 \cdot 10^{+308}
\end{array}

Sampling outcomes in binary64 precision:

Local Percentage Accuracy vs ?

The average percentage accuracy by input value. Horizontal axis shows value of an input variable; the variable is choosen in the title. Vertical axis is accuracy; higher is better. Red represent the original program, while blue represents Herbie's suggestion. These can be toggled with buttons below the plot. The line is an average while dots represent individual samples.

Accuracy vs Speed?

Herbie found 3 alternatives:

AlternativeAccuracySpeedup
The accuracy (vertical axis) and speed (horizontal axis) of each alternatives. Up and to the right is better. The red square shows the initial program, and each blue circle shows an alternative.The line shows the best available speed-accuracy tradeoffs.

Initial Program: 34.4% accurate, 1.0× speedup?

\[\begin{array}{l} \\ 1.7 \cdot 10^{+308} \cdot t - 1.7 \cdot 10^{+308} \end{array} \]
(FPCore (t) :precision binary64 (- (* 1.7e+308 t) 1.7e+308))
double code(double t) {
	return (1.7e+308 * t) - 1.7e+308;
}

real(8) function code(t)
    real(8), intent (in) :: t
    code = (1.7d+308 * t) - 1.7d+308
end function

public static double code(double t) {
	return (1.7e+308 * t) - 1.7e+308;
}

def code(t):
	return (1.7e+308 * t) - 1.7e+308

function code(t)
	return Float64(Float64(1.7e+308 * t) - 1.7e+308)
end

function tmp = code(t)
	tmp = (1.7e+308 * t) - 1.7e+308;
end

code[t_] := N[(N[(1.7e+308 * t), $MachinePrecision] - 1.7e+308), $MachinePrecision]

\begin{array}{l}

\\
1.7 \cdot 10^{+308} \cdot t - 1.7 \cdot 10^{+308}
\end{array}

Alternative 1: 99.6% accurate, 0.0× speedup?

\[\begin{array}{l} \\ \mathsf{fma}\left(1.7 \cdot 10^{+308}, t, -1.7 \cdot 10^{+308}\right) \end{array} \]
(FPCore (t) :precision binary64 (fma 1.7e+308 t -1.7e+308))
double code(double t) {
	return fma(1.7e+308, t, -1.7e+308);
}

function code(t)
	return fma(1.7e+308, t, -1.7e+308)
end

code[t_] := N[(1.7e+308 * t + -1.7e+308), $MachinePrecision]

\begin{array}{l}

\\
\mathsf{fma}\left(1.7 \cdot 10^{+308}, t, -1.7 \cdot 10^{+308}\right)
\end{array}
Derivation

  1. Initial program 34.2%

    \[1.7 \cdot 10^{+308} \cdot t - 1.7 \cdot 10^{+308} \]
  2. Step-by-step derivation

    1. fma-neg99.6%

      \[\leadsto \color{blue}{\mathsf{fma}\left(1.7 \cdot 10^{+308}, t, -1.7 \cdot 10^{+308}\right)} \]

    2. metadata-eval99.6%

      \[\leadsto \mathsf{fma}\left(1.7 \cdot 10^{+308}, t, \color{blue}{-1.7 \cdot 10^{+308}}\right) \]

  3. Simplified99.6%

    \[\leadsto \color{blue}{\mathsf{fma}\left(1.7 \cdot 10^{+308}, t, -1.7 \cdot 10^{+308}\right)} \]

  4. Final simplification99.6%

    \[\leadsto \mathsf{fma}\left(1.7 \cdot 10^{+308}, t, -1.7 \cdot 10^{+308}\right) \]

Alternative 2: 34.4% accurate, 1.7× speedup?

\[\begin{array}{l} \\ 1.7 \cdot 10^{+308} \cdot t \end{array} \]
(FPCore (t) :precision binary64 (* 1.7e+308 t))
double code(double t) {
	return 1.7e+308 * t;
}

real(8) function code(t)
    real(8), intent (in) :: t
    code = 1.7d+308 * t
end function

public static double code(double t) {
	return 1.7e+308 * t;
}

def code(t):
	return 1.7e+308 * t

function code(t)
	return Float64(1.7e+308 * t)
end

function tmp = code(t)
	tmp = 1.7e+308 * t;
end

code[t_] := N[(1.7e+308 * t), $MachinePrecision]

\begin{array}{l}

\\
1.7 \cdot 10^{+308} \cdot t
\end{array}
Derivation

  1. Initial program 34.2%

    \[1.7 \cdot 10^{+308} \cdot t - 1.7 \cdot 10^{+308} \]

  2. Taylor expanded in t around inf 34.2%

    \[\leadsto \color{blue}{1.7 \cdot 10^{+308} \cdot t} \]

  3. Final simplification34.2%

    \[\leadsto 1.7 \cdot 10^{+308} \cdot t \]

Alternative 3: 0.0% accurate, 5.0× speedup?

\[\begin{array}{l} \\ -1.7 \cdot 10^{+308} \end{array} \]
(FPCore (t) :precision binary64 -1.7e+308)
double code(double t) {
	return -1.7e+308;
}

real(8) function code(t)
    real(8), intent (in) :: t
    code = -1.7d+308
end function

public static double code(double t) {
	return -1.7e+308;
}

def code(t):
	return -1.7e+308

function code(t)
	return -1.7e+308
end

function tmp = code(t)
	tmp = -1.7e+308;
end

code[t_] := -1.7e+308

\begin{array}{l}

\\
-1.7 \cdot 10^{+308}
\end{array}
Derivation

  1. Initial program 34.2%

    \[1.7 \cdot 10^{+308} \cdot t - 1.7 \cdot 10^{+308} \]

  2. Taylor expanded in t around 0 0.0%

    \[\leadsto \color{blue}{-1.7 \cdot 10^{+308}} \]

  3. Final simplification0.0%

    \[\leadsto -1.7 \cdot 10^{+308} \]

Developer target: 99.6% accurate, 0.0× speedup?

\[\begin{array}{l} \\ \mathsf{fma}\left(1.7 \cdot 10^{+308}, t, -1.7 \cdot 10^{+308}\right) \end{array} \]
(FPCore (t) :precision binary64 (fma 1.7e+308 t (- 1.7e+308)))
double code(double t) {
	return fma(1.7e+308, t, -1.7e+308);
}

function code(t)
	return fma(1.7e+308, t, Float64(-1.7e+308))
end

code[t_] := N[(1.7e+308 * t + (-1.7e+308)), $MachinePrecision]

\begin{array}{l}

\\
\mathsf{fma}\left(1.7 \cdot 10^{+308}, t, -1.7 \cdot 10^{+308}\right)
\end{array}

Reproduce

?
herbie shell --seed 2023178 
(FPCore (t)
  :name "fma_test2"
  :precision binary64
  :pre (and (<= 1.9 t) (<= t 2.1))

  :herbie-target
  (fma 1.7e+308 t (- 1.7e+308))

  (- (* 1.7e+308 t) 1.7e+308))