Trowbridge-Reitz Sample, near normal, slope_x

Percentage Accurate: 99.0% → 99.0%
Time: 7.5s
Alternatives: 9
Speedup: 1.0×

Specification

?
\[\left(\left(cosTheta\_i > 0.9999 \land cosTheta\_i \leq 1\right) \land \left(2.328306437 \cdot 10^{-10} \leq u1 \land u1 \leq 1\right)\right) \land \left(2.328306437 \cdot 10^{-10} \leq u2 \land u2 \leq 1\right)\]
\[\begin{array}{l} \\ \sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \end{array} \]
(FPCore (cosTheta_i u1 u2)
 :precision binary32
 (* (sqrt (/ u1 (- 1.0 u1))) (cos (* 6.28318530718 u2))))
float code(float cosTheta_i, float u1, float u2) {
	return sqrtf((u1 / (1.0f - u1))) * cosf((6.28318530718f * u2));
}
real(4) function code(costheta_i, u1, u2)
    real(4), intent (in) :: costheta_i
    real(4), intent (in) :: u1
    real(4), intent (in) :: u2
    code = sqrt((u1 / (1.0e0 - u1))) * cos((6.28318530718e0 * u2))
end function
function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(u1 / Float32(Float32(1.0) - u1))) * cos(Float32(Float32(6.28318530718) * u2)))
end
function tmp = code(cosTheta_i, u1, u2)
	tmp = sqrt((u1 / (single(1.0) - u1))) * cos((single(6.28318530718) * u2));
end
\begin{array}{l}

\\
\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right)
\end{array}

Sampling outcomes in binary32 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 9 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: 99.0% accurate, 1.0× speedup?

\[\begin{array}{l} \\ \sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \end{array} \]
(FPCore (cosTheta_i u1 u2)
 :precision binary32
 (* (sqrt (/ u1 (- 1.0 u1))) (cos (* 6.28318530718 u2))))
float code(float cosTheta_i, float u1, float u2) {
	return sqrtf((u1 / (1.0f - u1))) * cosf((6.28318530718f * u2));
}
real(4) function code(costheta_i, u1, u2)
    real(4), intent (in) :: costheta_i
    real(4), intent (in) :: u1
    real(4), intent (in) :: u2
    code = sqrt((u1 / (1.0e0 - u1))) * cos((6.28318530718e0 * u2))
end function
function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(u1 / Float32(Float32(1.0) - u1))) * cos(Float32(Float32(6.28318530718) * u2)))
end
function tmp = code(cosTheta_i, u1, u2)
	tmp = sqrt((u1 / (single(1.0) - u1))) * cos((single(6.28318530718) * u2));
end
\begin{array}{l}

\\
\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right)
\end{array}

Alternative 1: 99.0% accurate, 1.0× speedup?

\[\begin{array}{l} \\ \sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \end{array} \]
(FPCore (cosTheta_i u1 u2)
 :precision binary32
 (* (sqrt (/ u1 (- 1.0 u1))) (cos (* 6.28318530718 u2))))
float code(float cosTheta_i, float u1, float u2) {
	return sqrtf((u1 / (1.0f - u1))) * cosf((6.28318530718f * u2));
}
real(4) function code(costheta_i, u1, u2)
    real(4), intent (in) :: costheta_i
    real(4), intent (in) :: u1
    real(4), intent (in) :: u2
    code = sqrt((u1 / (1.0e0 - u1))) * cos((6.28318530718e0 * u2))
end function
function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(u1 / Float32(Float32(1.0) - u1))) * cos(Float32(Float32(6.28318530718) * u2)))
end
function tmp = code(cosTheta_i, u1, u2)
	tmp = sqrt((u1 / (single(1.0) - u1))) * cos((single(6.28318530718) * u2));
end
\begin{array}{l}

\\
\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right)
\end{array}
Derivation
  1. Initial program 99.2%

    \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
  2. Add Preprocessing
  3. Add Preprocessing

Alternative 2: 93.9% accurate, 1.0× speedup?

\[\begin{array}{l} \\ \begin{array}{l} \mathbf{if}\;6.28318530718 \cdot u2 \leq 0.0003499999875202775:\\ \;\;\;\;\sqrt{\frac{u1}{1 - u1}}\\ \mathbf{else}:\\ \;\;\;\;\sqrt{u1 \cdot \left(u1 - -1\right)} \cdot \cos \left(6.28318530718 \cdot u2\right)\\ \end{array} \end{array} \]
(FPCore (cosTheta_i u1 u2)
 :precision binary32
 (if (<= (* 6.28318530718 u2) 0.0003499999875202775)
   (sqrt (/ u1 (- 1.0 u1)))
   (* (sqrt (* u1 (- u1 -1.0))) (cos (* 6.28318530718 u2)))))
float code(float cosTheta_i, float u1, float u2) {
	float tmp;
	if ((6.28318530718f * u2) <= 0.0003499999875202775f) {
		tmp = sqrtf((u1 / (1.0f - u1)));
	} else {
		tmp = sqrtf((u1 * (u1 - -1.0f))) * cosf((6.28318530718f * u2));
	}
	return tmp;
}
real(4) function code(costheta_i, u1, u2)
    real(4), intent (in) :: costheta_i
    real(4), intent (in) :: u1
    real(4), intent (in) :: u2
    real(4) :: tmp
    if ((6.28318530718e0 * u2) <= 0.0003499999875202775e0) then
        tmp = sqrt((u1 / (1.0e0 - u1)))
    else
        tmp = sqrt((u1 * (u1 - (-1.0e0)))) * cos((6.28318530718e0 * u2))
    end if
    code = tmp
end function
function code(cosTheta_i, u1, u2)
	tmp = Float32(0.0)
	if (Float32(Float32(6.28318530718) * u2) <= Float32(0.0003499999875202775))
		tmp = sqrt(Float32(u1 / Float32(Float32(1.0) - u1)));
	else
		tmp = Float32(sqrt(Float32(u1 * Float32(u1 - Float32(-1.0)))) * cos(Float32(Float32(6.28318530718) * u2)));
	end
	return tmp
end
function tmp_2 = code(cosTheta_i, u1, u2)
	tmp = single(0.0);
	if ((single(6.28318530718) * u2) <= single(0.0003499999875202775))
		tmp = sqrt((u1 / (single(1.0) - u1)));
	else
		tmp = sqrt((u1 * (u1 - single(-1.0)))) * cos((single(6.28318530718) * u2));
	end
	tmp_2 = tmp;
end
\begin{array}{l}

\\
\begin{array}{l}
\mathbf{if}\;6.28318530718 \cdot u2 \leq 0.0003499999875202775:\\
\;\;\;\;\sqrt{\frac{u1}{1 - u1}}\\

\mathbf{else}:\\
\;\;\;\;\sqrt{u1 \cdot \left(u1 - -1\right)} \cdot \cos \left(6.28318530718 \cdot u2\right)\\


\end{array}
\end{array}
Derivation
  1. Split input into 2 regimes
  2. if (*.f32 #s(literal 314159265359/50000000000 binary32) u2) < 3.49999988e-4

    1. Initial program 99.6%

      \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
    2. Add Preprocessing
    3. Taylor expanded in u2 around 0

      \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
    4. Step-by-step derivation
      1. lower-sqrt.f32N/A

        \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
      2. lower-/.f32N/A

        \[\leadsto \sqrt{\color{blue}{\frac{u1}{1 - u1}}} \]
      3. lower--.f3299.6

        \[\leadsto \sqrt{\frac{u1}{\color{blue}{1 - u1}}} \]
    5. Applied rewrites99.6%

      \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]

    if 3.49999988e-4 < (*.f32 #s(literal 314159265359/50000000000 binary32) u2)

    1. Initial program 98.5%

      \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
    2. Add Preprocessing
    3. Taylor expanded in u1 around 0

      \[\leadsto \sqrt{\color{blue}{u1 \cdot \left(1 + u1\right)}} \cdot \cos \left(\frac{314159265359}{50000000000} \cdot u2\right) \]
    4. Step-by-step derivation
      1. +-commutativeN/A

        \[\leadsto \sqrt{u1 \cdot \color{blue}{\left(u1 + 1\right)}} \cdot \cos \left(\frac{314159265359}{50000000000} \cdot u2\right) \]
      2. distribute-lft-inN/A

        \[\leadsto \sqrt{\color{blue}{u1 \cdot u1 + u1 \cdot 1}} \cdot \cos \left(\frac{314159265359}{50000000000} \cdot u2\right) \]
      3. *-rgt-identityN/A

        \[\leadsto \sqrt{u1 \cdot u1 + \color{blue}{u1}} \cdot \cos \left(\frac{314159265359}{50000000000} \cdot u2\right) \]
      4. lower-fma.f3235.1

        \[\leadsto \sqrt{\color{blue}{\mathsf{fma}\left(u1, u1, u1\right)}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
    5. Applied rewrites32.6%

      \[\leadsto \sqrt{\color{blue}{\mathsf{fma}\left(u1, u1, u1\right)}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
    6. Step-by-step derivation
      1. Applied rewrites87.0%

        \[\leadsto \sqrt{u1 \cdot \color{blue}{\left(u1 - -1\right)}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
    7. Recombined 2 regimes into one program.
    8. Add Preprocessing

    Alternative 3: 90.4% accurate, 1.0× speedup?

    \[\begin{array}{l} \\ \begin{array}{l} \mathbf{if}\;6.28318530718 \cdot u2 \leq 0.004999999888241291:\\ \;\;\;\;\sqrt{\frac{u1}{1 - u1}}\\ \mathbf{else}:\\ \;\;\;\;\sqrt{u1} \cdot \cos \left(6.28318530718 \cdot u2\right)\\ \end{array} \end{array} \]
    (FPCore (cosTheta_i u1 u2)
     :precision binary32
     (if (<= (* 6.28318530718 u2) 0.004999999888241291)
       (sqrt (/ u1 (- 1.0 u1)))
       (* (sqrt u1) (cos (* 6.28318530718 u2)))))
    float code(float cosTheta_i, float u1, float u2) {
    	float tmp;
    	if ((6.28318530718f * u2) <= 0.004999999888241291f) {
    		tmp = sqrtf((u1 / (1.0f - u1)));
    	} else {
    		tmp = sqrtf(u1) * cosf((6.28318530718f * u2));
    	}
    	return tmp;
    }
    
    real(4) function code(costheta_i, u1, u2)
        real(4), intent (in) :: costheta_i
        real(4), intent (in) :: u1
        real(4), intent (in) :: u2
        real(4) :: tmp
        if ((6.28318530718e0 * u2) <= 0.004999999888241291e0) then
            tmp = sqrt((u1 / (1.0e0 - u1)))
        else
            tmp = sqrt(u1) * cos((6.28318530718e0 * u2))
        end if
        code = tmp
    end function
    
    function code(cosTheta_i, u1, u2)
    	tmp = Float32(0.0)
    	if (Float32(Float32(6.28318530718) * u2) <= Float32(0.004999999888241291))
    		tmp = sqrt(Float32(u1 / Float32(Float32(1.0) - u1)));
    	else
    		tmp = Float32(sqrt(u1) * cos(Float32(Float32(6.28318530718) * u2)));
    	end
    	return tmp
    end
    
    function tmp_2 = code(cosTheta_i, u1, u2)
    	tmp = single(0.0);
    	if ((single(6.28318530718) * u2) <= single(0.004999999888241291))
    		tmp = sqrt((u1 / (single(1.0) - u1)));
    	else
    		tmp = sqrt(u1) * cos((single(6.28318530718) * u2));
    	end
    	tmp_2 = tmp;
    end
    
    \begin{array}{l}
    
    \\
    \begin{array}{l}
    \mathbf{if}\;6.28318530718 \cdot u2 \leq 0.004999999888241291:\\
    \;\;\;\;\sqrt{\frac{u1}{1 - u1}}\\
    
    \mathbf{else}:\\
    \;\;\;\;\sqrt{u1} \cdot \cos \left(6.28318530718 \cdot u2\right)\\
    
    
    \end{array}
    \end{array}
    
    Derivation
    1. Split input into 2 regimes
    2. if (*.f32 #s(literal 314159265359/50000000000 binary32) u2) < 0.00499999989

      1. Initial program 99.5%

        \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
      2. Add Preprocessing
      3. Taylor expanded in u2 around 0

        \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
      4. Step-by-step derivation
        1. lower-sqrt.f32N/A

          \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
        2. lower-/.f32N/A

          \[\leadsto \sqrt{\color{blue}{\frac{u1}{1 - u1}}} \]
        3. lower--.f3297.7

          \[\leadsto \sqrt{\frac{u1}{\color{blue}{1 - u1}}} \]
      5. Applied rewrites97.7%

        \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]

      if 0.00499999989 < (*.f32 #s(literal 314159265359/50000000000 binary32) u2)

      1. Initial program 98.3%

        \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
      2. Add Preprocessing
      3. Taylor expanded in u1 around 0

        \[\leadsto \color{blue}{\sqrt{u1}} \cdot \cos \left(\frac{314159265359}{50000000000} \cdot u2\right) \]
      4. Step-by-step derivation
        1. lower-sqrt.f3275.0

          \[\leadsto \color{blue}{\sqrt{u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
      5. Applied rewrites75.0%

        \[\leadsto \color{blue}{\sqrt{u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
    3. Recombined 2 regimes into one program.
    4. Add Preprocessing

    Alternative 4: 79.9% accurate, 5.4× speedup?

    \[\begin{array}{l} \\ \sqrt{\frac{u1}{1 - u1}} \end{array} \]
    (FPCore (cosTheta_i u1 u2) :precision binary32 (sqrt (/ u1 (- 1.0 u1))))
    float code(float cosTheta_i, float u1, float u2) {
    	return sqrtf((u1 / (1.0f - u1)));
    }
    
    real(4) function code(costheta_i, u1, u2)
        real(4), intent (in) :: costheta_i
        real(4), intent (in) :: u1
        real(4), intent (in) :: u2
        code = sqrt((u1 / (1.0e0 - u1)))
    end function
    
    function code(cosTheta_i, u1, u2)
    	return sqrt(Float32(u1 / Float32(Float32(1.0) - u1)))
    end
    
    function tmp = code(cosTheta_i, u1, u2)
    	tmp = sqrt((u1 / (single(1.0) - u1)));
    end
    
    \begin{array}{l}
    
    \\
    \sqrt{\frac{u1}{1 - u1}}
    \end{array}
    
    Derivation
    1. Initial program 99.2%

      \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
    2. Add Preprocessing
    3. Taylor expanded in u2 around 0

      \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
    4. Step-by-step derivation
      1. lower-sqrt.f32N/A

        \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
      2. lower-/.f32N/A

        \[\leadsto \sqrt{\color{blue}{\frac{u1}{1 - u1}}} \]
      3. lower--.f3281.4

        \[\leadsto \sqrt{\frac{u1}{\color{blue}{1 - u1}}} \]
    5. Applied rewrites81.4%

      \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
    6. Add Preprocessing

    Alternative 5: 63.3% accurate, 5.4× speedup?

    \[\begin{array}{l} \\ \sqrt{\mathsf{fma}\left(u1 - 1, u1, 1\right) \cdot u1} \end{array} \]
    (FPCore (cosTheta_i u1 u2)
     :precision binary32
     (sqrt (* (fma (- u1 1.0) u1 1.0) u1)))
    float code(float cosTheta_i, float u1, float u2) {
    	return sqrtf((fmaf((u1 - 1.0f), u1, 1.0f) * u1));
    }
    
    function code(cosTheta_i, u1, u2)
    	return sqrt(Float32(fma(Float32(u1 - Float32(1.0)), u1, Float32(1.0)) * u1))
    end
    
    \begin{array}{l}
    
    \\
    \sqrt{\mathsf{fma}\left(u1 - 1, u1, 1\right) \cdot u1}
    \end{array}
    
    Derivation
    1. Initial program 99.2%

      \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
    2. Add Preprocessing
    3. Taylor expanded in u2 around 0

      \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
    4. Step-by-step derivation
      1. lower-sqrt.f32N/A

        \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
      2. lower-/.f32N/A

        \[\leadsto \sqrt{\color{blue}{\frac{u1}{1 - u1}}} \]
      3. lower--.f3281.4

        \[\leadsto \sqrt{\frac{u1}{\color{blue}{1 - u1}}} \]
    5. Applied rewrites81.4%

      \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
    6. Step-by-step derivation
      1. Applied rewrites66.1%

        \[\leadsto \sqrt{\frac{u1}{u1 + 1}} \]
      2. Taylor expanded in u1 around 0

        \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(u1 - 1\right)\right)} \]
      3. Step-by-step derivation
        1. Applied rewrites67.2%

          \[\leadsto \sqrt{\mathsf{fma}\left(u1 - 1, u1, 1\right) \cdot u1} \]
        2. Add Preprocessing

        Alternative 6: 61.6% accurate, 7.1× speedup?

        \[\begin{array}{l} \\ \sqrt{\left(1 - u1\right) \cdot u1} \end{array} \]
        (FPCore (cosTheta_i u1 u2) :precision binary32 (sqrt (* (- 1.0 u1) u1)))
        float code(float cosTheta_i, float u1, float u2) {
        	return sqrtf(((1.0f - u1) * u1));
        }
        
        real(4) function code(costheta_i, u1, u2)
            real(4), intent (in) :: costheta_i
            real(4), intent (in) :: u1
            real(4), intent (in) :: u2
            code = sqrt(((1.0e0 - u1) * u1))
        end function
        
        function code(cosTheta_i, u1, u2)
        	return sqrt(Float32(Float32(Float32(1.0) - u1) * u1))
        end
        
        function tmp = code(cosTheta_i, u1, u2)
        	tmp = sqrt(((single(1.0) - u1) * u1));
        end
        
        \begin{array}{l}
        
        \\
        \sqrt{\left(1 - u1\right) \cdot u1}
        \end{array}
        
        Derivation
        1. Initial program 99.2%

          \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
        2. Add Preprocessing
        3. Taylor expanded in u2 around 0

          \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
        4. Step-by-step derivation
          1. lower-sqrt.f32N/A

            \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
          2. lower-/.f32N/A

            \[\leadsto \sqrt{\color{blue}{\frac{u1}{1 - u1}}} \]
          3. lower--.f3281.4

            \[\leadsto \sqrt{\frac{u1}{\color{blue}{1 - u1}}} \]
        5. Applied rewrites81.4%

          \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
        6. Step-by-step derivation
          1. Applied rewrites66.1%

            \[\leadsto \sqrt{\frac{u1}{u1 + 1}} \]
          2. Taylor expanded in u1 around 0

            \[\leadsto \sqrt{u1 \cdot \left(1 + -1 \cdot u1\right)} \]
          3. Step-by-step derivation
            1. Applied rewrites66.0%

              \[\leadsto \sqrt{\left(1 - u1\right) \cdot u1} \]
            2. Add Preprocessing

            Alternative 7: 14.3% accurate, 9.0× speedup?

            \[\begin{array}{l} \\ \frac{0.5}{u1} - 1 \end{array} \]
            (FPCore (cosTheta_i u1 u2) :precision binary32 (- (/ 0.5 u1) 1.0))
            float code(float cosTheta_i, float u1, float u2) {
            	return (0.5f / u1) - 1.0f;
            }
            
            real(4) function code(costheta_i, u1, u2)
                real(4), intent (in) :: costheta_i
                real(4), intent (in) :: u1
                real(4), intent (in) :: u2
                code = (0.5e0 / u1) - 1.0e0
            end function
            
            function code(cosTheta_i, u1, u2)
            	return Float32(Float32(Float32(0.5) / u1) - Float32(1.0))
            end
            
            function tmp = code(cosTheta_i, u1, u2)
            	tmp = (single(0.5) / u1) - single(1.0);
            end
            
            \begin{array}{l}
            
            \\
            \frac{0.5}{u1} - 1
            \end{array}
            
            Derivation
            1. Initial program 99.2%

              \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
            2. Add Preprocessing
            3. Taylor expanded in u2 around 0

              \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
            4. Step-by-step derivation
              1. lower-sqrt.f32N/A

                \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
              2. lower-/.f32N/A

                \[\leadsto \sqrt{\color{blue}{\frac{u1}{1 - u1}}} \]
              3. lower--.f3281.4

                \[\leadsto \sqrt{\frac{u1}{\color{blue}{1 - u1}}} \]
            5. Applied rewrites81.4%

              \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
            6. Step-by-step derivation
              1. Applied rewrites66.1%

                \[\leadsto \sqrt{\frac{u1}{u1 + 1}} \]
              2. Taylor expanded in u1 around -inf

                \[\leadsto \frac{1}{2} \cdot \frac{1}{u1} + \color{blue}{{\left(\sqrt{-1}\right)}^{2}} \]
              3. Step-by-step derivation
                1. Applied rewrites13.8%

                  \[\leadsto \frac{0.5}{u1} - \color{blue}{1} \]
                2. Add Preprocessing

                Alternative 8: 4.1% accurate, 9.0× speedup?

                \[\begin{array}{l} \\ 1 - \frac{0.5}{u1} \end{array} \]
                (FPCore (cosTheta_i u1 u2) :precision binary32 (- 1.0 (/ 0.5 u1)))
                float code(float cosTheta_i, float u1, float u2) {
                	return 1.0f - (0.5f / u1);
                }
                
                real(4) function code(costheta_i, u1, u2)
                    real(4), intent (in) :: costheta_i
                    real(4), intent (in) :: u1
                    real(4), intent (in) :: u2
                    code = 1.0e0 - (0.5e0 / u1)
                end function
                
                function code(cosTheta_i, u1, u2)
                	return Float32(Float32(1.0) - Float32(Float32(0.5) / u1))
                end
                
                function tmp = code(cosTheta_i, u1, u2)
                	tmp = single(1.0) - (single(0.5) / u1);
                end
                
                \begin{array}{l}
                
                \\
                1 - \frac{0.5}{u1}
                \end{array}
                
                Derivation
                1. Initial program 99.2%

                  \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
                2. Add Preprocessing
                3. Taylor expanded in u2 around 0

                  \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
                4. Step-by-step derivation
                  1. lower-sqrt.f32N/A

                    \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
                  2. lower-/.f32N/A

                    \[\leadsto \sqrt{\color{blue}{\frac{u1}{1 - u1}}} \]
                  3. lower--.f3281.4

                    \[\leadsto \sqrt{\frac{u1}{\color{blue}{1 - u1}}} \]
                5. Applied rewrites81.4%

                  \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
                6. Step-by-step derivation
                  1. Applied rewrites66.1%

                    \[\leadsto \sqrt{\frac{u1}{u1 + 1}} \]
                  2. Taylor expanded in u1 around inf

                    \[\leadsto 1 - \color{blue}{\frac{1}{2} \cdot \frac{1}{u1}} \]
                  3. Step-by-step derivation
                    1. Applied rewrites3.9%

                      \[\leadsto 1 - \color{blue}{\frac{0.5}{u1}} \]
                    2. Add Preprocessing

                    Alternative 9: 4.1% accurate, 135.0× speedup?

                    \[\begin{array}{l} \\ -1 \end{array} \]
                    (FPCore (cosTheta_i u1 u2) :precision binary32 -1.0)
                    float code(float cosTheta_i, float u1, float u2) {
                    	return -1.0f;
                    }
                    
                    real(4) function code(costheta_i, u1, u2)
                        real(4), intent (in) :: costheta_i
                        real(4), intent (in) :: u1
                        real(4), intent (in) :: u2
                        code = -1.0e0
                    end function
                    
                    function code(cosTheta_i, u1, u2)
                    	return Float32(-1.0)
                    end
                    
                    function tmp = code(cosTheta_i, u1, u2)
                    	tmp = single(-1.0);
                    end
                    
                    \begin{array}{l}
                    
                    \\
                    -1
                    \end{array}
                    
                    Derivation
                    1. Initial program 99.2%

                      \[\sqrt{\frac{u1}{1 - u1}} \cdot \cos \left(6.28318530718 \cdot u2\right) \]
                    2. Add Preprocessing
                    3. Taylor expanded in u2 around 0

                      \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
                    4. Step-by-step derivation
                      1. lower-sqrt.f32N/A

                        \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
                      2. lower-/.f32N/A

                        \[\leadsto \sqrt{\color{blue}{\frac{u1}{1 - u1}}} \]
                      3. lower--.f3281.4

                        \[\leadsto \sqrt{\frac{u1}{\color{blue}{1 - u1}}} \]
                    5. Applied rewrites81.4%

                      \[\leadsto \color{blue}{\sqrt{\frac{u1}{1 - u1}}} \]
                    6. Step-by-step derivation
                      1. Applied rewrites66.1%

                        \[\leadsto \sqrt{\frac{u1}{u1 + 1}} \]
                      2. Taylor expanded in u1 around -inf

                        \[\leadsto {\left(\sqrt{-1}\right)}^{\color{blue}{2}} \]
                      3. Step-by-step derivation
                        1. Applied rewrites3.7%

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

                        Reproduce

                        ?
                        herbie shell --seed 2024326 
                        (FPCore (cosTheta_i u1 u2)
                          :name "Trowbridge-Reitz Sample, near normal, slope_x"
                          :precision binary32
                          :pre (and (and (and (> cosTheta_i 0.9999) (<= cosTheta_i 1.0)) (and (<= 2.328306437e-10 u1) (<= u1 1.0))) (and (<= 2.328306437e-10 u2) (<= u2 1.0)))
                          (* (sqrt (/ u1 (- 1.0 u1))) (cos (* 6.28318530718 u2))))