UniformSampleCone, z

Percentage Accurate: 99.9% → 99.9%
Time: 1.8s
Alternatives: 4
Speedup: 1.1×

Specification

?
\[\left(\left(2.328306437 \cdot 10^{-10} \leq ux \land ux \leq 1\right) \land \left(2.328306437 \cdot 10^{-10} \leq uy \land uy \leq 1\right)\right) \land \left(0 \leq maxCos \land maxCos \leq 1\right)\]
\[\left(1 - ux\right) + ux \cdot maxCos \]
(FPCore (ux uy maxCos)
  :precision binary32
  :pre (and (and (and (<= 2.328306437e-10 ux) (<= ux 1.0))
          (and (<= 2.328306437e-10 uy) (<= uy 1.0)))
     (and (<= 0.0 maxCos) (<= maxCos 1.0)))
  (+ (- 1.0 ux) (* ux maxCos)))
float code(float ux, float uy, float maxCos) {
	return (1.0f - ux) + (ux * maxCos);
}

real(4) function code(ux, uy, maxcos)
use fmin_fmax_functions
    real(4), intent (in) :: ux
    real(4), intent (in) :: uy
    real(4), intent (in) :: maxcos
    code = (1.0e0 - ux) + (ux * maxcos)
end function

function code(ux, uy, maxCos)
	return Float32(Float32(Float32(1.0) - ux) + Float32(ux * maxCos))
end

function tmp = code(ux, uy, maxCos)
	tmp = (single(1.0) - ux) + (ux * maxCos);
end

\left(1 - ux\right) + ux \cdot maxCos

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 4 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: 99.9% accurate, 1.0× speedup?

\[\left(\left(2.328306437 \cdot 10^{-10} \leq ux \land ux \leq 1\right) \land \left(2.328306437 \cdot 10^{-10} \leq uy \land uy \leq 1\right)\right) \land \left(0 \leq maxCos \land maxCos \leq 1\right)\]
\[\left(1 - ux\right) + ux \cdot maxCos \]
(FPCore (ux uy maxCos)
  :precision binary32
  :pre (and (and (and (<= 2.328306437e-10 ux) (<= ux 1.0))
          (and (<= 2.328306437e-10 uy) (<= uy 1.0)))
     (and (<= 0.0 maxCos) (<= maxCos 1.0)))
  (+ (- 1.0 ux) (* ux maxCos)))
float code(float ux, float uy, float maxCos) {
	return (1.0f - ux) + (ux * maxCos);
}

real(4) function code(ux, uy, maxcos)
use fmin_fmax_functions
    real(4), intent (in) :: ux
    real(4), intent (in) :: uy
    real(4), intent (in) :: maxcos
    code = (1.0e0 - ux) + (ux * maxcos)
end function

function code(ux, uy, maxCos)
	return Float32(Float32(Float32(1.0) - ux) + Float32(ux * maxCos))
end

function tmp = code(ux, uy, maxCos)
	tmp = (single(1.0) - ux) + (ux * maxCos);
end

\left(1 - ux\right) + ux \cdot maxCos

Alternative 1: 99.9% accurate, 1.1× speedup?

\[\left(\left(2.328306437 \cdot 10^{-10} \leq ux \land ux \leq 1\right) \land \left(2.328306437 \cdot 10^{-10} \leq uy \land uy \leq 1\right)\right) \land \left(0 \leq maxCos \land maxCos \leq 1\right)\]
\[\mathsf{fma}\left(ux, maxCos, 1 - ux\right) \]
(FPCore (ux uy maxCos)
  :precision binary32
  :pre (and (and (and (<= 2.328306437e-10 ux) (<= ux 1.0))
          (and (<= 2.328306437e-10 uy) (<= uy 1.0)))
     (and (<= 0.0 maxCos) (<= maxCos 1.0)))
  (fma ux maxCos (- 1.0 ux)))
float code(float ux, float uy, float maxCos) {
	return fmaf(ux, maxCos, (1.0f - ux));
}

function code(ux, uy, maxCos)
	return fma(ux, maxCos, Float32(Float32(1.0) - ux))
end

\mathsf{fma}\left(ux, maxCos, 1 - ux\right)
Derivation

  1. Initial program 99.9%

    \[\left(1 - ux\right) + ux \cdot maxCos \]
  2. Step-by-step derivation

    1. Applied rewrites99.9%

      \[\leadsto \mathsf{fma}\left(ux, maxCos, 1 - ux\right) \]
    2. Add Preprocessing

    Alternative 2: 99.8% accurate, 1.1× speedup?

    \[\left(\left(2.328306437 \cdot 10^{-10} \leq ux \land ux \leq 1\right) \land \left(2.328306437 \cdot 10^{-10} \leq uy \land uy \leq 1\right)\right) \land \left(0 \leq maxCos \land maxCos \leq 1\right)\]
    \[\mathsf{fma}\left(maxCos, ux, 1\right) - ux \]
    (FPCore (ux uy maxCos)
  :precision binary32
  :pre (and (and (and (<= 2.328306437e-10 ux) (<= ux 1.0))
          (and (<= 2.328306437e-10 uy) (<= uy 1.0)))
     (and (<= 0.0 maxCos) (<= maxCos 1.0)))
  (- (fma maxCos ux 1.0) ux))
    float code(float ux, float uy, float maxCos) {
	return fmaf(maxCos, ux, 1.0f) - ux;
}

    function code(ux, uy, maxCos)
	return Float32(fma(maxCos, ux, Float32(1.0)) - ux)
end

    \mathsf{fma}\left(maxCos, ux, 1\right) - ux
    Derivation

    1. Initial program 99.9%

      \[\left(1 - ux\right) + ux \cdot maxCos \]
    2. Step-by-step derivation

      1. Applied rewrites99.8%

        \[\leadsto \mathsf{fma}\left(maxCos, ux, 1\right) - ux \]
      2. Add Preprocessing

      Alternative 3: 98.1% accurate, 2.6× speedup?

      \[\left(\left(2.328306437 \cdot 10^{-10} \leq ux \land ux \leq 1\right) \land \left(2.328306437 \cdot 10^{-10} \leq uy \land uy \leq 1\right)\right) \land \left(0 \leq maxCos \land maxCos \leq 1\right)\]
      \[1 - ux \]
      (FPCore (ux uy maxCos)
  :precision binary32
  :pre (and (and (and (<= 2.328306437e-10 ux) (<= ux 1.0))
          (and (<= 2.328306437e-10 uy) (<= uy 1.0)))
     (and (<= 0.0 maxCos) (<= maxCos 1.0)))
  (- 1.0 ux))
      float code(float ux, float uy, float maxCos) {
	return 1.0f - ux;
}

      real(4) function code(ux, uy, maxcos)
use fmin_fmax_functions
    real(4), intent (in) :: ux
    real(4), intent (in) :: uy
    real(4), intent (in) :: maxcos
    code = 1.0e0 - ux
end function

      function code(ux, uy, maxCos)
	return Float32(Float32(1.0) - ux)
end

      function tmp = code(ux, uy, maxCos)
	tmp = single(1.0) - ux;
end

      1 - ux
      Derivation

      1. Initial program 99.9%

        \[\left(1 - ux\right) + ux \cdot maxCos \]

      2. Taylor expanded in maxCos around 0

        \[\leadsto 1 - ux \]
      3. Step-by-step derivation

        1. Applied rewrites98.1%

          \[\leadsto 1 - ux \]
        2. Add Preprocessing

        Alternative 4: 71.3% accurate, 9.2× speedup?

        \[\left(\left(2.328306437 \cdot 10^{-10} \leq ux \land ux \leq 1\right) \land \left(2.328306437 \cdot 10^{-10} \leq uy \land uy \leq 1\right)\right) \land \left(0 \leq maxCos \land maxCos \leq 1\right)\]
        \[1 \]
        (FPCore (ux uy maxCos)
  :precision binary32
  :pre (and (and (and (<= 2.328306437e-10 ux) (<= ux 1.0))
          (and (<= 2.328306437e-10 uy) (<= uy 1.0)))
     (and (<= 0.0 maxCos) (<= maxCos 1.0)))
  1.0)
        float code(float ux, float uy, float maxCos) {
	return 1.0f;
}

        real(4) function code(ux, uy, maxcos)
use fmin_fmax_functions
    real(4), intent (in) :: ux
    real(4), intent (in) :: uy
    real(4), intent (in) :: maxcos
    code = 1.0e0
end function

        function code(ux, uy, maxCos)
	return Float32(1.0)
end

        function tmp = code(ux, uy, maxCos)
	tmp = single(1.0);
end

        1
        Derivation

        1. Initial program 99.9%

          \[\left(1 - ux\right) + ux \cdot maxCos \]

        2. Taylor expanded in maxCos around 0

          \[\leadsto 1 - ux \]
        3. Step-by-step derivation

          1. Applied rewrites98.1%

            \[\leadsto 1 - ux \]

          2. Taylor expanded in ux around 0

            \[\leadsto 1 \]
          3. Step-by-step derivation

            1. Applied rewrites71.3%

              \[\leadsto 1 \]
            2. Add Preprocessing

            Reproduce

            ?
            herbie shell --seed 2025359 
(FPCore (ux uy maxCos)
  :name "UniformSampleCone, z"
  :precision binary32
  :pre (and (and (and (<= 2.328306437e-10 ux) (<= ux 1.0)) (and (<= 2.328306437e-10 uy) (<= uy 1.0))) (and (<= 0.0 maxCos) (<= maxCos 1.0)))
  (+ (- 1.0 ux) (* ux maxCos)))