Beckmann Sample, near normal, slope_x

Percentage Accurate: 57.9% → 99.2%

Time: 4.6s

Alternatives: 12

Speedup: 1.0×

Specification

\[\left(\left(cosTheta\_i > 0.9999 \land cosTheta\_i \leq 1\right) \land \left(2.328306437 \cdot 10^{-10} \leq u1 \land u1 \leq 1\right)\right) \land \left(2.328306437 \cdot 10^{-10} \leq u2 \land u2 \leq 1\right)\]

Math

\[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (* (sqrt (- (log (- 1.0 u1)))) (cos (* (* 2.0 PI) u2))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-logf((1.0f - u1))) * cosf(((2.0f * ((float) M_PI)) * u2));
}

Julia

function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(-log(Float32(Float32(1.0) - u1)))) * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2)))
end

MATLAB

function tmp = code(cosTheta_i, u1, u2)
	tmp = sqrt(-log((single(1.0) - u1))) * cos(((single(2.0) * single(pi)) * u2));
end

Initial Program: 57.9% accurate, 1.0× speedup

Math

\[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (* (sqrt (- (log (- 1.0 u1)))) (cos (* (* 2.0 PI) u2))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-logf((1.0f - u1))) * cosf(((2.0f * ((float) M_PI)) * u2));
}

Julia

function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(-log(Float32(Float32(1.0) - u1)))) * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2)))
end

MATLAB

function tmp = code(cosTheta_i, u1, u2)
	tmp = sqrt(-log((single(1.0) - u1))) * cos(((single(2.0) * single(pi)) * u2));
end

Alternative 1: 99.2% accurate, 0.8× speedup

Math

\[\sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(u2 \cdot \mathsf{fma}\left(-2, \pi, 0.5 \cdot \frac{\pi}{u2}\right)\right) \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (*
 (sqrt (- (log1p (- u1))))
 (sin (* u2 (fma -2.0 PI (* 0.5 (/ PI u2)))))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-log1pf(-u1)) * sinf((u2 * fmaf(-2.0f, ((float) M_PI), (0.5f * (((float) M_PI) / u2)))));
}

Julia

function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(-log1p(Float32(-u1)))) * sin(Float32(u2 * fma(Float32(-2.0), Float32(pi), Float32(Float32(0.5) * Float32(Float32(pi) / u2))))))
end

Derivation

1. Initial program 57.9%

   \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
2. Step-by-step derivation

   1. lift-log.f32 N/A

      \[\leadsto \sqrt{-\color{blue}{\log \left(1 - u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. lift--.f32 N/A

      \[\leadsto \sqrt{-\log \color{blue}{\left(1 - u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   3. sub-flip N/A

      \[\leadsto \sqrt{-\log \color{blue}{\left(1 + \left(\mathsf{neg}\left(u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   4. lower-log1p.f32 N/A

      \[\leadsto \sqrt{-\color{blue}{\mathsf{log1p}\left(\mathsf{neg}\left(u1\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   5. lower-neg.f32 99.1%

      \[\leadsto \sqrt{-\mathsf{log1p}\left(\color{blue}{-u1}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

3. Applied rewrites 99.1%

   \[\leadsto \sqrt{-\color{blue}{\mathsf{log1p}\left(-u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
4. Step-by-step derivation

   1. lift-cos.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\cos \left(\left(2 \cdot \pi\right) \cdot u2\right)} \]

   2. cos-neg-rev N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\cos \left(\mathsf{neg}\left(\left(2 \cdot \pi\right) \cdot u2\right)\right)} \]

   3. sin-+PI/2-rev N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\sin \left(\left(\mathsf{neg}\left(\left(2 \cdot \pi\right) \cdot u2\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right)} \]

   4. lower-sin.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\sin \left(\left(\mathsf{neg}\left(\left(2 \cdot \pi\right) \cdot u2\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right)} \]

   5. lift-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\left(\mathsf{neg}\left(\color{blue}{\left(2 \cdot \pi\right) \cdot u2}\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right) \]

   6. lift-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\left(\mathsf{neg}\left(\color{blue}{\left(2 \cdot \pi\right)} \cdot u2\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right) \]

   7. associate-*l* N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\left(\mathsf{neg}\left(\color{blue}{2 \cdot \left(\pi \cdot u2\right)}\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right) \]

   8. distribute-lft-neg-in N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\color{blue}{\left(\mathsf{neg}\left(2\right)\right) \cdot \left(\pi \cdot u2\right)} + \frac{\mathsf{PI}\left(\right)}{2}\right) \]

   9. lower-fma.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(\mathsf{fma}\left(\mathsf{neg}\left(2\right), \pi \cdot u2, \frac{\mathsf{PI}\left(\right)}{2}\right)\right)} \]

   10. metadata-eval N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(\color{blue}{-2}, \pi \cdot u2, \frac{\mathsf{PI}\left(\right)}{2}\right)\right) \]

   11. *-commutative N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, \color{blue}{u2 \cdot \pi}, \frac{\mathsf{PI}\left(\right)}{2}\right)\right) \]

   12. lower-*.f32 N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, \color{blue}{u2 \cdot \pi}, \frac{\mathsf{PI}\left(\right)}{2}\right)\right) \]

   13. lift-PI.f32 N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \frac{\color{blue}{\pi}}{2}\right)\right) \]

   14. mult-flip N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \color{blue}{\pi \cdot \frac{1}{2}}\right)\right) \]

   15. metadata-eval N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \pi \cdot \color{blue}{\frac{1}{2}}\right)\right) \]

   16. *-commutative N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \color{blue}{\frac{1}{2} \cdot \pi}\right)\right) \]

   17. lower-*.f32 99.1%

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \color{blue}{0.5 \cdot \pi}\right)\right) \]

5. Applied rewrites 99.1%

   \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, 0.5 \cdot \pi\right)\right)} \]

6. Taylor expanded in u2 around inf

   \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(u2 \cdot \left(-2 \cdot \pi + \frac{1}{2} \cdot \frac{\pi}{u2}\right)\right)} \]
7. Step-by-step derivation

   1. lower-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(u2 \cdot \color{blue}{\left(-2 \cdot \mathsf{PI}\left(\right) + \frac{1}{2} \cdot \frac{\mathsf{PI}\left(\right)}{u2}\right)}\right) \]

   2. lower-fma.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(u2 \cdot \mathsf{fma}\left(-2, \color{blue}{\mathsf{PI}\left(\right)}, \frac{1}{2} \cdot \frac{\mathsf{PI}\left(\right)}{u2}\right)\right) \]

   3. lower-PI.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(u2 \cdot \mathsf{fma}\left(-2, \pi, \frac{1}{2} \cdot \frac{\mathsf{PI}\left(\right)}{u2}\right)\right) \]

   4. lower-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(u2 \cdot \mathsf{fma}\left(-2, \pi, \frac{1}{2} \cdot \frac{\mathsf{PI}\left(\right)}{u2}\right)\right) \]

   5. lower-/.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(u2 \cdot \mathsf{fma}\left(-2, \pi, \frac{1}{2} \cdot \frac{\mathsf{PI}\left(\right)}{u2}\right)\right) \]

   6. lower-PI.f32 99.1%

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(u2 \cdot \mathsf{fma}\left(-2, \pi, 0.5 \cdot \frac{\pi}{u2}\right)\right) \]

8. Applied rewrites 99.1%

   \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(u2 \cdot \mathsf{fma}\left(-2, \pi, 0.5 \cdot \frac{\pi}{u2}\right)\right)} \]
9. Add Preprocessing

Alternative 2: 99.1% accurate, 0.9× speedup

Math

\[\sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\pi \cdot \mathsf{fma}\left(u2, -2, 0.5\right)\right) \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (* (sqrt (- (log1p (- u1)))) (sin (* PI (fma u2 -2.0 0.5)))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-log1pf(-u1)) * sinf((((float) M_PI) * fmaf(u2, -2.0f, 0.5f)));
}

Julia

function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(-log1p(Float32(-u1)))) * sin(Float32(Float32(pi) * fma(u2, Float32(-2.0), Float32(0.5)))))
end

Derivation

1. Initial program 57.9%

   \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
2. Step-by-step derivation

   1. lift-log.f32 N/A

      \[\leadsto \sqrt{-\color{blue}{\log \left(1 - u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. lift--.f32 N/A

      \[\leadsto \sqrt{-\log \color{blue}{\left(1 - u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   3. sub-flip N/A

      \[\leadsto \sqrt{-\log \color{blue}{\left(1 + \left(\mathsf{neg}\left(u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   4. lower-log1p.f32 N/A

      \[\leadsto \sqrt{-\color{blue}{\mathsf{log1p}\left(\mathsf{neg}\left(u1\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   5. lower-neg.f32 99.1%

      \[\leadsto \sqrt{-\mathsf{log1p}\left(\color{blue}{-u1}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

3. Applied rewrites 99.1%

   \[\leadsto \sqrt{-\color{blue}{\mathsf{log1p}\left(-u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
4. Step-by-step derivation

   1. lift-cos.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\cos \left(\left(2 \cdot \pi\right) \cdot u2\right)} \]

   2. cos-neg-rev N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\cos \left(\mathsf{neg}\left(\left(2 \cdot \pi\right) \cdot u2\right)\right)} \]

   3. sin-+PI/2-rev N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\sin \left(\left(\mathsf{neg}\left(\left(2 \cdot \pi\right) \cdot u2\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right)} \]

   4. lower-sin.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\sin \left(\left(\mathsf{neg}\left(\left(2 \cdot \pi\right) \cdot u2\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right)} \]

   5. lift-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\left(\mathsf{neg}\left(\color{blue}{\left(2 \cdot \pi\right) \cdot u2}\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right) \]

   6. lift-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\left(\mathsf{neg}\left(\color{blue}{\left(2 \cdot \pi\right)} \cdot u2\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right) \]

   7. associate-*l* N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\left(\mathsf{neg}\left(\color{blue}{2 \cdot \left(\pi \cdot u2\right)}\right)\right) + \frac{\mathsf{PI}\left(\right)}{2}\right) \]

   8. distribute-lft-neg-in N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\color{blue}{\left(\mathsf{neg}\left(2\right)\right) \cdot \left(\pi \cdot u2\right)} + \frac{\mathsf{PI}\left(\right)}{2}\right) \]

   9. lower-fma.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(\mathsf{fma}\left(\mathsf{neg}\left(2\right), \pi \cdot u2, \frac{\mathsf{PI}\left(\right)}{2}\right)\right)} \]

   10. metadata-eval N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(\color{blue}{-2}, \pi \cdot u2, \frac{\mathsf{PI}\left(\right)}{2}\right)\right) \]

   11. *-commutative N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, \color{blue}{u2 \cdot \pi}, \frac{\mathsf{PI}\left(\right)}{2}\right)\right) \]

   12. lower-*.f32 N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, \color{blue}{u2 \cdot \pi}, \frac{\mathsf{PI}\left(\right)}{2}\right)\right) \]

   13. lift-PI.f32 N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \frac{\color{blue}{\pi}}{2}\right)\right) \]

   14. mult-flip N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \color{blue}{\pi \cdot \frac{1}{2}}\right)\right) \]

   15. metadata-eval N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \pi \cdot \color{blue}{\frac{1}{2}}\right)\right) \]

   16. *-commutative N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \color{blue}{\frac{1}{2} \cdot \pi}\right)\right) \]

   17. lower-*.f32 99.1%

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, \color{blue}{0.5 \cdot \pi}\right)\right) \]

5. Applied rewrites 99.1%

   \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \color{blue}{\sin \left(\mathsf{fma}\left(-2, u2 \cdot \pi, 0.5 \cdot \pi\right)\right)} \]
6. Step-by-step derivation

   1. lift-fma.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(-2 \cdot \left(u2 \cdot \pi\right) + \frac{1}{2} \cdot \pi\right)} \]

   2. lift-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(-2 \cdot \left(u2 \cdot \pi\right) + \color{blue}{\frac{1}{2} \cdot \pi}\right) \]

   3. fp-cancel-sign-sub-inv N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(-2 \cdot \left(u2 \cdot \pi\right) - \left(\mathsf{neg}\left(\frac{1}{2}\right)\right) \cdot \pi\right)} \]

   4. metadata-eval N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(-2 \cdot \left(u2 \cdot \pi\right) - \color{blue}{\frac{-1}{2}} \cdot \pi\right) \]

   5. lift-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(-2 \cdot \color{blue}{\left(u2 \cdot \pi\right)} - \frac{-1}{2} \cdot \pi\right) \]

   6. associate-*r* N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\color{blue}{\left(-2 \cdot u2\right) \cdot \pi} - \frac{-1}{2} \cdot \pi\right) \]

   7. distribute-rgt-out-- N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(\pi \cdot \left(-2 \cdot u2 - \frac{-1}{2}\right)\right)} \]

   8. lower-*.f32 N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(\pi \cdot \left(-2 \cdot u2 - \frac{-1}{2}\right)\right)} \]

   9. metadata-eval N/A

      \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\pi \cdot \left(-2 \cdot u2 - \color{blue}{\left(\mathsf{neg}\left(\frac{1}{2}\right)\right)}\right)\right) \]

   10. add-flip-rev N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\pi \cdot \color{blue}{\left(-2 \cdot u2 + \frac{1}{2}\right)}\right) \]

   11. *-commutative N/A

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\pi \cdot \left(\color{blue}{u2 \cdot -2} + \frac{1}{2}\right)\right) \]

   12. lower-fma.f32 99.2%

       \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \left(\pi \cdot \color{blue}{\mathsf{fma}\left(u2, -2, 0.5\right)}\right) \]

7. Applied rewrites 99.2%

   \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \sin \color{blue}{\left(\pi \cdot \mathsf{fma}\left(u2, -2, 0.5\right)\right)} \]
8. Add Preprocessing

Alternative 3: 99.1% accurate, 1.0× speedup

Math

\[\sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \cos \left(6.2831854820251465 \cdot u2\right) \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (* (sqrt (- (log1p (- u1)))) (cos (* 6.2831854820251465 u2))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-log1pf(-u1)) * cosf((6.2831854820251465f * u2));
}

Julia

function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(-log1p(Float32(-u1)))) * cos(Float32(Float32(6.2831854820251465) * u2)))
end

Derivation

1. Initial program 57.9%

   \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
2. Step-by-step derivation

   1. lift-log.f32 N/A

      \[\leadsto \sqrt{-\color{blue}{\log \left(1 - u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. lift--.f32 N/A

      \[\leadsto \sqrt{-\log \color{blue}{\left(1 - u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   3. sub-flip N/A

      \[\leadsto \sqrt{-\log \color{blue}{\left(1 + \left(\mathsf{neg}\left(u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   4. lower-log1p.f32 N/A

      \[\leadsto \sqrt{-\color{blue}{\mathsf{log1p}\left(\mathsf{neg}\left(u1\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   5. lower-neg.f32 99.1%

      \[\leadsto \sqrt{-\mathsf{log1p}\left(\color{blue}{-u1}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

3. Applied rewrites 99.1%

   \[\leadsto \sqrt{-\color{blue}{\mathsf{log1p}\left(-u1\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

4. Evaluated real constant 99.1%

   \[\leadsto \sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]
5. Add Preprocessing

Alternative 4: 97.4% accurate, 0.8× speedup

Math

\[\begin{array}{l} t_0 := \log \left(1 - u1\right)\\ \mathbf{if}\;t\_0 \leq -0.004000000189989805:\\ \;\;\;\;\sqrt{-t\_0} \cdot \cos \left(6.2831854820251465 \cdot u2\right)\\ \mathbf{else}:\\ \;\;\;\;\sqrt{\mathsf{fma}\left(0.5 \cdot u1, u1, u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right)\\ \end{array} \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (let* ((t_0 (log (- 1.0 u1))))
  (if (<= t_0 -0.004000000189989805)
    (* (sqrt (- t_0)) (cos (* 6.2831854820251465 u2)))
    (* (sqrt (fma (* 0.5 u1) u1 u1)) (cos (* (* 2.0 PI) u2))))))

C

float code(float cosTheta_i, float u1, float u2) {
	float t_0 = logf((1.0f - u1));
	float tmp;
	if (t_0 <= -0.004000000189989805f) {
		tmp = sqrtf(-t_0) * cosf((6.2831854820251465f * u2));
	} else {
		tmp = sqrtf(fmaf((0.5f * u1), u1, u1)) * cosf(((2.0f * ((float) M_PI)) * u2));
	}
	return tmp;
}

Julia

function code(cosTheta_i, u1, u2)
	t_0 = log(Float32(Float32(1.0) - u1))
	tmp = Float32(0.0)
	if (t_0 <= Float32(-0.004000000189989805))
		tmp = Float32(sqrt(Float32(-t_0)) * cos(Float32(Float32(6.2831854820251465) * u2)));
	else
		tmp = Float32(sqrt(fma(Float32(Float32(0.5) * u1), u1, u1)) * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2)));
	end
	return tmp
end

Derivation

1. Split input into 2 regimes

2. if (log.f32 (-.f32 #s(literal 1 binary32) u1)) < -0.00400000019

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Evaluated real constant 57.9%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

   if -0.00400000019 < (log.f32 (-.f32 #s(literal 1 binary32) u1))

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Taylor expanded in u1 around 0

      \[\leadsto \sqrt{\color{blue}{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
   3. Step-by-step derivation

      1. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \color{blue}{\left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      2. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + \color{blue}{u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      3. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \color{blue}{\left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      4. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + \color{blue}{u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)}\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      5. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \color{blue}{\left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)}\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      6. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \color{blue}{\frac{1}{4} \cdot u1}\right)\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      7. lower-*.f32 93.9%

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(0.5 + u1 \cdot \left(0.3333333333333333 + 0.25 \cdot \color{blue}{u1}\right)\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   4. Applied rewrites 93.9%

      \[\leadsto \sqrt{\color{blue}{u1 \cdot \left(1 + u1 \cdot \left(0.5 + u1 \cdot \left(0.3333333333333333 + 0.25 \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   5. Taylor expanded in u1 around 0

      \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \frac{1}{2}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
   6. Step-by-step derivation

   7. Applied rewrites 88.2%

      \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot 0.5\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
   8. Step-by-step derivation

      1. lift-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \color{blue}{\left(1 + u1 \cdot \frac{1}{2}\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      2. lift-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + \color{blue}{u1 \cdot \frac{1}{2}}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      3. +-commutative N/A

         \[\leadsto \sqrt{u1 \cdot \left(u1 \cdot \frac{1}{2} + \color{blue}{1}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      4. distribute-rgt-in N/A

         \[\leadsto \sqrt{\left(u1 \cdot \frac{1}{2}\right) \cdot u1 + \color{blue}{1 \cdot u1}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      5. *-lft-identity N/A

         \[\leadsto \sqrt{\left(u1 \cdot \frac{1}{2}\right) \cdot u1 + u1} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      6. lower-fma.f32 88.3%

         \[\leadsto \sqrt{\mathsf{fma}\left(u1 \cdot 0.5, \color{blue}{u1}, u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      7. lift-*.f32 N/A

         \[\leadsto \sqrt{\mathsf{fma}\left(u1 \cdot \frac{1}{2}, u1, u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      8. *-commutative N/A

         \[\leadsto \sqrt{\mathsf{fma}\left(\frac{1}{2} \cdot u1, u1, u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      9. lower-*.f32 88.3%

         \[\leadsto \sqrt{\mathsf{fma}\left(0.5 \cdot u1, u1, u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   9. Applied rewrites 88.3%

      \[\leadsto \sqrt{\mathsf{fma}\left(0.5 \cdot u1, \color{blue}{u1}, u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 5: 97.3% accurate, 0.9× speedup

Math

\[\begin{array}{l} t_0 := \log \left(1 - u1\right)\\ t_1 := \cos \left(6.2831854820251465 \cdot u2\right)\\ \mathbf{if}\;t\_0 \leq -0.004000000189989805:\\ \;\;\;\;\sqrt{-t\_0} \cdot t\_1\\ \mathbf{else}:\\ \;\;\;\;\sqrt{u1 \cdot \left(1 + u1 \cdot 0.5\right)} \cdot t\_1\\ \end{array} \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (let* ((t_0 (log (- 1.0 u1))) (t_1 (cos (* 6.2831854820251465 u2))))
  (if (<= t_0 -0.004000000189989805)
    (* (sqrt (- t_0)) t_1)
    (* (sqrt (* u1 (+ 1.0 (* u1 0.5)))) t_1))))

C

float code(float cosTheta_i, float u1, float u2) {
	float t_0 = logf((1.0f - u1));
	float t_1 = cosf((6.2831854820251465f * u2));
	float tmp;
	if (t_0 <= -0.004000000189989805f) {
		tmp = sqrtf(-t_0) * t_1;
	} else {
		tmp = sqrtf((u1 * (1.0f + (u1 * 0.5f)))) * t_1;
	}
	return tmp;
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(4) function code(costheta_i, u1, u2)
use fmin_fmax_functions
    real(4), intent (in) :: costheta_i
    real(4), intent (in) :: u1
    real(4), intent (in) :: u2
    real(4) :: t_0
    real(4) :: t_1
    real(4) :: tmp
    t_0 = log((1.0e0 - u1))
    t_1 = cos((6.2831854820251465e0 * u2))
    if (t_0 <= (-0.004000000189989805e0)) then
        tmp = sqrt(-t_0) * t_1
    else
        tmp = sqrt((u1 * (1.0e0 + (u1 * 0.5e0)))) * t_1
    end if
    code = tmp
end function

Julia

function code(cosTheta_i, u1, u2)
	t_0 = log(Float32(Float32(1.0) - u1))
	t_1 = cos(Float32(Float32(6.2831854820251465) * u2))
	tmp = Float32(0.0)
	if (t_0 <= Float32(-0.004000000189989805))
		tmp = Float32(sqrt(Float32(-t_0)) * t_1);
	else
		tmp = Float32(sqrt(Float32(u1 * Float32(Float32(1.0) + Float32(u1 * Float32(0.5))))) * t_1);
	end
	return tmp
end

MATLAB

function tmp_2 = code(cosTheta_i, u1, u2)
	t_0 = log((single(1.0) - u1));
	t_1 = cos((single(6.2831854820251465) * u2));
	tmp = single(0.0);
	if (t_0 <= single(-0.004000000189989805))
		tmp = sqrt(-t_0) * t_1;
	else
		tmp = sqrt((u1 * (single(1.0) + (u1 * single(0.5))))) * t_1;
	end
	tmp_2 = tmp;
end

Derivation

1. Split input into 2 regimes

2. if (log.f32 (-.f32 #s(literal 1 binary32) u1)) < -0.00400000019

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Evaluated real constant 57.9%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

   if -0.00400000019 < (log.f32 (-.f32 #s(literal 1 binary32) u1))

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Taylor expanded in u1 around 0

      \[\leadsto \sqrt{\color{blue}{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
   3. Step-by-step derivation

      1. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \color{blue}{\left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      2. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + \color{blue}{u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      3. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \color{blue}{\left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      4. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + \color{blue}{u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)}\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      5. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \color{blue}{\left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)}\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      6. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \color{blue}{\frac{1}{4} \cdot u1}\right)\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      7. lower-*.f32 93.9%

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(0.5 + u1 \cdot \left(0.3333333333333333 + 0.25 \cdot \color{blue}{u1}\right)\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   4. Applied rewrites 93.9%

      \[\leadsto \sqrt{\color{blue}{u1 \cdot \left(1 + u1 \cdot \left(0.5 + u1 \cdot \left(0.3333333333333333 + 0.25 \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   5. Taylor expanded in u1 around 0

      \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \frac{1}{2}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
   6. Step-by-step derivation

   7. Applied rewrites 88.2%

      \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot 0.5\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   8. Evaluated real constant 88.2%

      \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot 0.5\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 6: 95.4% accurate, 0.5× speedup

Math

\[\begin{array}{l} t_0 := \sqrt{-\log \left(1 - u1\right)}\\ \mathbf{if}\;t\_0 \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \leq 0.05999999865889549:\\ \;\;\;\;\sqrt{u1 \cdot \left(1 + u1 \cdot 0.5\right)} \cdot \cos \left(6.2831854820251465 \cdot u2\right)\\ \mathbf{else}:\\ \;\;\;\;t\_0 + -19.739209900765786 \cdot \left({u2}^{2} \cdot t\_0\right)\\ \end{array} \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (let* ((t_0 (sqrt (- (log (- 1.0 u1))))))
  (if (<= (* t_0 (cos (* (* 2.0 PI) u2))) 0.05999999865889549)
    (*
     (sqrt (* u1 (+ 1.0 (* u1 0.5))))
     (cos (* 6.2831854820251465 u2)))
    (+ t_0 (* -19.739209900765786 (* (pow u2 2.0) t_0))))))

C

float code(float cosTheta_i, float u1, float u2) {
	float t_0 = sqrtf(-logf((1.0f - u1)));
	float tmp;
	if ((t_0 * cosf(((2.0f * ((float) M_PI)) * u2))) <= 0.05999999865889549f) {
		tmp = sqrtf((u1 * (1.0f + (u1 * 0.5f)))) * cosf((6.2831854820251465f * u2));
	} else {
		tmp = t_0 + (-19.739209900765786f * (powf(u2, 2.0f) * t_0));
	}
	return tmp;
}

Julia

function code(cosTheta_i, u1, u2)
	t_0 = sqrt(Float32(-log(Float32(Float32(1.0) - u1))))
	tmp = Float32(0.0)
	if (Float32(t_0 * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2))) <= Float32(0.05999999865889549))
		tmp = Float32(sqrt(Float32(u1 * Float32(Float32(1.0) + Float32(u1 * Float32(0.5))))) * cos(Float32(Float32(6.2831854820251465) * u2)));
	else
		tmp = Float32(t_0 + Float32(Float32(-19.739209900765786) * Float32((u2 ^ Float32(2.0)) * t_0)));
	end
	return tmp
end

MATLAB

function tmp_2 = code(cosTheta_i, u1, u2)
	t_0 = sqrt(-log((single(1.0) - u1)));
	tmp = single(0.0);
	if ((t_0 * cos(((single(2.0) * single(pi)) * u2))) <= single(0.05999999865889549))
		tmp = sqrt((u1 * (single(1.0) + (u1 * single(0.5))))) * cos((single(6.2831854820251465) * u2));
	else
		tmp = t_0 + (single(-19.739209900765786) * ((u2 ^ single(2.0)) * t_0));
	end
	tmp_2 = tmp;
end

Derivation

1. Split input into 2 regimes

2. if (*.f32 (sqrt.f32 (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1)))) (cos.f32 (*.f32 (*.f32 #s(literal 2 binary32) (PI.f32)) u2))) < 0.0599999987

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Taylor expanded in u1 around 0

      \[\leadsto \sqrt{\color{blue}{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
   3. Step-by-step derivation

      1. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \color{blue}{\left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      2. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + \color{blue}{u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      3. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \color{blue}{\left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)\right)}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      4. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + \color{blue}{u1 \cdot \left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)}\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      5. lower-*.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \color{blue}{\left(\frac{1}{3} + \frac{1}{4} \cdot u1\right)}\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      6. lower-+.f32 N/A

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(\frac{1}{2} + u1 \cdot \left(\frac{1}{3} + \color{blue}{\frac{1}{4} \cdot u1}\right)\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

      7. lower-*.f32 93.9%

         \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \left(0.5 + u1 \cdot \left(0.3333333333333333 + 0.25 \cdot \color{blue}{u1}\right)\right)\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   4. Applied rewrites 93.9%

      \[\leadsto \sqrt{\color{blue}{u1 \cdot \left(1 + u1 \cdot \left(0.5 + u1 \cdot \left(0.3333333333333333 + 0.25 \cdot u1\right)\right)\right)}} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   5. Taylor expanded in u1 around 0

      \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot \frac{1}{2}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]
   6. Step-by-step derivation

   7. Applied rewrites 88.2%

      \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot 0.5\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   8. Evaluated real constant 88.2%

      \[\leadsto \sqrt{u1 \cdot \left(1 + u1 \cdot 0.5\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

   if 0.0599999987 < (*.f32 (sqrt.f32 (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1)))) (cos.f32 (*.f32 (*.f32 #s(literal 2 binary32) (PI.f32)) u2)))

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Evaluated real constant 57.9%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

   3. Taylor expanded in u2 around 0

      \[\leadsto \color{blue}{\sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right)} \]
   4. Step-by-step derivation

      1. lower-+.f32 N/A

         \[\leadsto \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} + \color{blue}{\frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right)} \]

      2. lower-sqrt.f32 N/A

         \[\leadsto \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} + \color{blue}{\frac{-173627926472025}{8796093022208}} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      3. lower-neg.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      4. lower-log.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      5. lower--.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      6. lower-*.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \color{blue}{\left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right)} \]

      7. lower-*.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \color{blue}{\sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}}\right) \]

      8. lower-pow.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\color{blue}{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}}\right) \]

      9. lower-sqrt.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      10. lower-neg.f32 N/A

          \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{-\log \left(1 - u1\right)}\right) \]

      11. lower-log.f32 N/A

          \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{-\log \left(1 - u1\right)}\right) \]

      12. lower--.f32 53.3%

          \[\leadsto \sqrt{-\log \left(1 - u1\right)} + -19.739209900765786 \cdot \left({u2}^{2} \cdot \sqrt{-\log \left(1 - u1\right)}\right) \]

   5. Applied rewrites 53.3%

      \[\leadsto \color{blue}{\sqrt{-\log \left(1 - u1\right)} + -19.739209900765786 \cdot \left({u2}^{2} \cdot \sqrt{-\log \left(1 - u1\right)}\right)} \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 7: 88.7% accurate, 0.5× speedup

Math

\[\begin{array}{l} t_0 := \sqrt{-\log \left(1 - u1\right)}\\ \mathbf{if}\;t\_0 \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \leq 0.014999999664723873:\\ \;\;\;\;\cos \left(6.2831854820251465 \cdot u2\right) \cdot \sqrt{u1}\\ \mathbf{else}:\\ \;\;\;\;t\_0 + -19.739209900765786 \cdot \left({u2}^{2} \cdot t\_0\right)\\ \end{array} \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (let* ((t_0 (sqrt (- (log (- 1.0 u1))))))
  (if (<= (* t_0 (cos (* (* 2.0 PI) u2))) 0.014999999664723873)
    (* (cos (* 6.2831854820251465 u2)) (sqrt u1))
    (+ t_0 (* -19.739209900765786 (* (pow u2 2.0) t_0))))))

C

float code(float cosTheta_i, float u1, float u2) {
	float t_0 = sqrtf(-logf((1.0f - u1)));
	float tmp;
	if ((t_0 * cosf(((2.0f * ((float) M_PI)) * u2))) <= 0.014999999664723873f) {
		tmp = cosf((6.2831854820251465f * u2)) * sqrtf(u1);
	} else {
		tmp = t_0 + (-19.739209900765786f * (powf(u2, 2.0f) * t_0));
	}
	return tmp;
}

Julia

function code(cosTheta_i, u1, u2)
	t_0 = sqrt(Float32(-log(Float32(Float32(1.0) - u1))))
	tmp = Float32(0.0)
	if (Float32(t_0 * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2))) <= Float32(0.014999999664723873))
		tmp = Float32(cos(Float32(Float32(6.2831854820251465) * u2)) * sqrt(u1));
	else
		tmp = Float32(t_0 + Float32(Float32(-19.739209900765786) * Float32((u2 ^ Float32(2.0)) * t_0)));
	end
	return tmp
end

MATLAB

function tmp_2 = code(cosTheta_i, u1, u2)
	t_0 = sqrt(-log((single(1.0) - u1)));
	tmp = single(0.0);
	if ((t_0 * cos(((single(2.0) * single(pi)) * u2))) <= single(0.014999999664723873))
		tmp = cos((single(6.2831854820251465) * u2)) * sqrt(u1);
	else
		tmp = t_0 + (single(-19.739209900765786) * ((u2 ^ single(2.0)) * t_0));
	end
	tmp_2 = tmp;
end

Derivation

1. Split input into 2 regimes

2. if (*.f32 (sqrt.f32 (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1)))) (cos.f32 (*.f32 (*.f32 #s(literal 2 binary32) (PI.f32)) u2))) < 0.0149999997

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Evaluated real constant 57.9%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

   3. Taylor expanded in u1 around 0

      \[\leadsto \color{blue}{\cos \left(\frac{13176795}{2097152} \cdot u2\right) \cdot \sqrt{u1}} \]
   4. Step-by-step derivation

      1. lower-*.f32 N/A

         \[\leadsto \cos \left(\frac{13176795}{2097152} \cdot u2\right) \cdot \color{blue}{\sqrt{u1}} \]

      2. lower-cos.f32 N/A

         \[\leadsto \cos \left(\frac{13176795}{2097152} \cdot u2\right) \cdot \sqrt{\color{blue}{u1}} \]

      3. lower-*.f32 N/A

         \[\leadsto \cos \left(\frac{13176795}{2097152} \cdot u2\right) \cdot \sqrt{u1} \]

      4. lower-sqrt.f32 76.4%

         \[\leadsto \cos \left(6.2831854820251465 \cdot u2\right) \cdot \sqrt{u1} \]

   5. Applied rewrites 76.4%

      \[\leadsto \color{blue}{\cos \left(6.2831854820251465 \cdot u2\right) \cdot \sqrt{u1}} \]

   if 0.0149999997 < (*.f32 (sqrt.f32 (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1)))) (cos.f32 (*.f32 (*.f32 #s(literal 2 binary32) (PI.f32)) u2)))

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Evaluated real constant 57.9%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

   3. Taylor expanded in u2 around 0

      \[\leadsto \color{blue}{\sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right)} \]
   4. Step-by-step derivation

      1. lower-+.f32 N/A

         \[\leadsto \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} + \color{blue}{\frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right)} \]

      2. lower-sqrt.f32 N/A

         \[\leadsto \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} + \color{blue}{\frac{-173627926472025}{8796093022208}} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      3. lower-neg.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      4. lower-log.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      5. lower--.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      6. lower-*.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \color{blue}{\left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right)} \]

      7. lower-*.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \color{blue}{\sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}}\right) \]

      8. lower-pow.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\color{blue}{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}}\right) \]

      9. lower-sqrt.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}\right) \]

      10. lower-neg.f32 N/A

          \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{-\log \left(1 - u1\right)}\right) \]

      11. lower-log.f32 N/A

          \[\leadsto \sqrt{-\log \left(1 - u1\right)} + \frac{-173627926472025}{8796093022208} \cdot \left({u2}^{2} \cdot \sqrt{-\log \left(1 - u1\right)}\right) \]

      12. lower--.f32 53.3%

          \[\leadsto \sqrt{-\log \left(1 - u1\right)} + -19.739209900765786 \cdot \left({u2}^{2} \cdot \sqrt{-\log \left(1 - u1\right)}\right) \]

   5. Applied rewrites 53.3%

      \[\leadsto \color{blue}{\sqrt{-\log \left(1 - u1\right)} + -19.739209900765786 \cdot \left({u2}^{2} \cdot \sqrt{-\log \left(1 - u1\right)}\right)} \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 8: 88.7% accurate, 0.5× speedup

Math

\[\begin{array}{l} t_0 := \sqrt{-\log \left(1 - u1\right)}\\ \mathbf{if}\;t\_0 \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \leq 0.014999999664723873:\\ \;\;\;\;\cos \left(6.2831854820251465 \cdot u2\right) \cdot \sqrt{u1}\\ \mathbf{else}:\\ \;\;\;\;t\_0 \cdot \left(1 + -19.739209900765786 \cdot {u2}^{2}\right)\\ \end{array} \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (let* ((t_0 (sqrt (- (log (- 1.0 u1))))))
  (if (<= (* t_0 (cos (* (* 2.0 PI) u2))) 0.014999999664723873)
    (* (cos (* 6.2831854820251465 u2)) (sqrt u1))
    (* t_0 (+ 1.0 (* -19.739209900765786 (pow u2 2.0)))))))

C

float code(float cosTheta_i, float u1, float u2) {
	float t_0 = sqrtf(-logf((1.0f - u1)));
	float tmp;
	if ((t_0 * cosf(((2.0f * ((float) M_PI)) * u2))) <= 0.014999999664723873f) {
		tmp = cosf((6.2831854820251465f * u2)) * sqrtf(u1);
	} else {
		tmp = t_0 * (1.0f + (-19.739209900765786f * powf(u2, 2.0f)));
	}
	return tmp;
}

Julia

function code(cosTheta_i, u1, u2)
	t_0 = sqrt(Float32(-log(Float32(Float32(1.0) - u1))))
	tmp = Float32(0.0)
	if (Float32(t_0 * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2))) <= Float32(0.014999999664723873))
		tmp = Float32(cos(Float32(Float32(6.2831854820251465) * u2)) * sqrt(u1));
	else
		tmp = Float32(t_0 * Float32(Float32(1.0) + Float32(Float32(-19.739209900765786) * (u2 ^ Float32(2.0)))));
	end
	return tmp
end

MATLAB

function tmp_2 = code(cosTheta_i, u1, u2)
	t_0 = sqrt(-log((single(1.0) - u1)));
	tmp = single(0.0);
	if ((t_0 * cos(((single(2.0) * single(pi)) * u2))) <= single(0.014999999664723873))
		tmp = cos((single(6.2831854820251465) * u2)) * sqrt(u1);
	else
		tmp = t_0 * (single(1.0) + (single(-19.739209900765786) * (u2 ^ single(2.0))));
	end
	tmp_2 = tmp;
end

Derivation

1. Split input into 2 regimes

2. if (*.f32 (sqrt.f32 (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1)))) (cos.f32 (*.f32 (*.f32 #s(literal 2 binary32) (PI.f32)) u2))) < 0.0149999997

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Evaluated real constant 57.9%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

   3. Taylor expanded in u1 around 0

      \[\leadsto \color{blue}{\cos \left(\frac{13176795}{2097152} \cdot u2\right) \cdot \sqrt{u1}} \]
   4. Step-by-step derivation

      1. lower-*.f32 N/A

         \[\leadsto \cos \left(\frac{13176795}{2097152} \cdot u2\right) \cdot \color{blue}{\sqrt{u1}} \]

      2. lower-cos.f32 N/A

         \[\leadsto \cos \left(\frac{13176795}{2097152} \cdot u2\right) \cdot \sqrt{\color{blue}{u1}} \]

      3. lower-*.f32 N/A

         \[\leadsto \cos \left(\frac{13176795}{2097152} \cdot u2\right) \cdot \sqrt{u1} \]

      4. lower-sqrt.f32 76.4%

         \[\leadsto \cos \left(6.2831854820251465 \cdot u2\right) \cdot \sqrt{u1} \]

   5. Applied rewrites 76.4%

      \[\leadsto \color{blue}{\cos \left(6.2831854820251465 \cdot u2\right) \cdot \sqrt{u1}} \]

   if 0.0149999997 < (*.f32 (sqrt.f32 (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1)))) (cos.f32 (*.f32 (*.f32 #s(literal 2 binary32) (PI.f32)) u2)))

   1. Initial program 57.9%

      \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

   2. Evaluated real constant 57.9%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

   3. Taylor expanded in u2 around 0

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \color{blue}{\left(1 + \frac{-173627926472025}{8796093022208} \cdot {u2}^{2}\right)} \]
   4. Step-by-step derivation

      1. lower-+.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \left(1 + \color{blue}{\frac{-173627926472025}{8796093022208} \cdot {u2}^{2}}\right) \]

      2. lower-*.f32 N/A

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \left(1 + \frac{-173627926472025}{8796093022208} \cdot \color{blue}{{u2}^{2}}\right) \]

      3. lower-pow.f32 53.3%

         \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \left(1 + -19.739209900765786 \cdot {u2}^{\color{blue}{2}}\right) \]

   5. Applied rewrites 53.3%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \color{blue}{\left(1 + -19.739209900765786 \cdot {u2}^{2}\right)} \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 9: 53.3% accurate, 1.5× speedup

Math

\[\sqrt{-\log \left(1 - u1\right)} \cdot \left(1 + -19.739209900765786 \cdot {u2}^{2}\right) \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (*
 (sqrt (- (log (- 1.0 u1))))
 (+ 1.0 (* -19.739209900765786 (pow u2 2.0)))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-logf((1.0f - u1))) * (1.0f + (-19.739209900765786f * powf(u2, 2.0f)));
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(4) function code(costheta_i, u1, u2)
use fmin_fmax_functions
    real(4), intent (in) :: costheta_i
    real(4), intent (in) :: u1
    real(4), intent (in) :: u2
    code = sqrt(-log((1.0e0 - u1))) * (1.0e0 + ((-19.739209900765786e0) * (u2 ** 2.0e0)))
end function

Julia

function code(cosTheta_i, u1, u2)
	return Float32(sqrt(Float32(-log(Float32(Float32(1.0) - u1)))) * Float32(Float32(1.0) + Float32(Float32(-19.739209900765786) * (u2 ^ Float32(2.0)))))
end

MATLAB

function tmp = code(cosTheta_i, u1, u2)
	tmp = sqrt(-log((single(1.0) - u1))) * (single(1.0) + (single(-19.739209900765786) * (u2 ^ single(2.0))));
end

Derivation

1. Initial program 57.9%

   \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

2. Evaluated real constant 57.9%

   \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\color{blue}{6.2831854820251465} \cdot u2\right) \]

3. Taylor expanded in u2 around 0

   \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \color{blue}{\left(1 + \frac{-173627926472025}{8796093022208} \cdot {u2}^{2}\right)} \]
4. Step-by-step derivation

   1. lower-+.f32 N/A

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \left(1 + \color{blue}{\frac{-173627926472025}{8796093022208} \cdot {u2}^{2}}\right) \]

   2. lower-*.f32 N/A

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \left(1 + \frac{-173627926472025}{8796093022208} \cdot \color{blue}{{u2}^{2}}\right) \]

   3. lower-pow.f32 53.3%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \left(1 + -19.739209900765786 \cdot {u2}^{\color{blue}{2}}\right) \]

5. Applied rewrites 53.3%

   \[\leadsto \sqrt{-\log \left(1 - u1\right)} \cdot \color{blue}{\left(1 + -19.739209900765786 \cdot {u2}^{2}\right)} \]
6. Add Preprocessing

Alternative 10: 50.1% accurate, 2.6× speedup

Math

\[\sqrt{-\log \left(\mathsf{fma}\left(u1, \frac{1}{u1}, u1 \cdot -1\right)\right)} \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (sqrt (- (log (fma u1 (/ 1.0 u1) (* u1 -1.0))))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-logf(fmaf(u1, (1.0f / u1), (u1 * -1.0f))));
}

Julia

function code(cosTheta_i, u1, u2)
	return sqrt(Float32(-log(fma(u1, Float32(Float32(1.0) / u1), Float32(u1 * Float32(-1.0))))))
end

Derivation

1. Initial program 57.9%

   \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

2. Taylor expanded in u2 around 0

   \[\leadsto \color{blue}{\sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}} \]
3. Step-by-step derivation

   1. lower-sqrt.f32 N/A

      \[\leadsto \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} \]

   2. lower-neg.f32 N/A

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

   3. lower-log.f32 N/A

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

   4. lower--.f32 50.1%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

4. Applied rewrites 50.1%

   \[\leadsto \color{blue}{\sqrt{-\log \left(1 - u1\right)}} \]

5. Taylor expanded in u1 around inf

   \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} - 1\right)\right)} \]
6. Step-by-step derivation

   1. lower-*.f32 N/A

      \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} - 1\right)\right)} \]

   2. lower--.f32 N/A

      \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} - 1\right)\right)} \]

   3. lower-/.f32 49.2%

      \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} - 1\right)\right)} \]

7. Applied rewrites 49.2%

   \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} - 1\right)\right)} \]
8. Step-by-step derivation

   1. lift-*.f32 N/A

      \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} - 1\right)\right)} \]

   2. lift--.f32 N/A

      \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} - 1\right)\right)} \]

   3. sub-flip N/A

      \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} + \left(\mathsf{neg}\left(1\right)\right)\right)\right)} \]

   4. metadata-eval N/A

      \[\leadsto \sqrt{-\log \left(u1 \cdot \left(\frac{1}{u1} + -1\right)\right)} \]

   5. distribute-lft-in N/A

      \[\leadsto \sqrt{-\log \left(u1 \cdot \frac{1}{u1} + u1 \cdot -1\right)} \]

   6. lower-fma.f32 N/A

      \[\leadsto \sqrt{-\log \left(\mathsf{fma}\left(u1, \frac{1}{u1}, u1 \cdot -1\right)\right)} \]

   7. lower-*.f32 49.6%

      \[\leadsto \sqrt{-\log \left(\mathsf{fma}\left(u1, \frac{1}{u1}, u1 \cdot -1\right)\right)} \]

9. Applied rewrites 49.6%

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

Alternative 11: 49.6% accurate, 4.4× speedup

Math

\[\sqrt{-\log \left(1 - u1\right)} \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (sqrt (- (log (- 1.0 u1)))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-logf((1.0f - u1)));
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(4) function code(costheta_i, u1, u2)
use fmin_fmax_functions
    real(4), intent (in) :: costheta_i
    real(4), intent (in) :: u1
    real(4), intent (in) :: u2
    code = sqrt(-log((1.0e0 - u1)))
end function

Julia

function code(cosTheta_i, u1, u2)
	return sqrt(Float32(-log(Float32(Float32(1.0) - u1))))
end

MATLAB

function tmp = code(cosTheta_i, u1, u2)
	tmp = sqrt(-log((single(1.0) - u1)));
end

Derivation

1. Initial program 57.9%

   \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

2. Taylor expanded in u2 around 0

   \[\leadsto \color{blue}{\sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}} \]
3. Step-by-step derivation

   1. lower-sqrt.f32 N/A

      \[\leadsto \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} \]

   2. lower-neg.f32 N/A

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

   3. lower-log.f32 N/A

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

   4. lower--.f32 50.1%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

4. Applied rewrites 50.1%

   \[\leadsto \color{blue}{\sqrt{-\log \left(1 - u1\right)}} \]
5. Add Preprocessing

Alternative 12: 6.6% accurate, 5.6× speedup

Math

\[\sqrt{-\log 1} \]

FPCore

(FPCore (cosTheta_i u1 u2)
  :precision binary32
  (sqrt (- (log 1.0))))

C

float code(float cosTheta_i, float u1, float u2) {
	return sqrtf(-logf(1.0f));
}

Fortran

module fmin_fmax_functions
    implicit none
    private
    public fmax
    public fmin

    interface fmax
        module procedure fmax88
        module procedure fmax44
        module procedure fmax84
        module procedure fmax48
    end interface
    interface fmin
        module procedure fmin88
        module procedure fmin44
        module procedure fmin84
        module procedure fmin48
    end interface
contains
    real(8) function fmax88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(4) function fmax44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, max(x, y), y /= y), x /= x)
    end function
    real(8) function fmax84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmax48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
    end function
    real(8) function fmin88(x, y) result (res)
        real(8), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(4) function fmin44(x, y) result (res)
        real(4), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(y, merge(x, min(x, y), y /= y), x /= x)
    end function
    real(8) function fmin84(x, y) result(res)
        real(8), intent (in) :: x
        real(4), intent (in) :: y
        res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
    end function
    real(8) function fmin48(x, y) result(res)
        real(4), intent (in) :: x
        real(8), intent (in) :: y
        res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
    end function
end module

real(4) function code(costheta_i, u1, u2)
use fmin_fmax_functions
    real(4), intent (in) :: costheta_i
    real(4), intent (in) :: u1
    real(4), intent (in) :: u2
    code = sqrt(-log(1.0e0))
end function

Julia

function code(cosTheta_i, u1, u2)
	return sqrt(Float32(-log(Float32(1.0))))
end

MATLAB

function tmp = code(cosTheta_i, u1, u2)
	tmp = sqrt(-log(single(1.0)));
end

Derivation

1. Initial program 57.9%

   \[\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right) \]

2. Taylor expanded in u2 around 0

   \[\leadsto \color{blue}{\sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)}} \]
3. Step-by-step derivation

   1. lower-sqrt.f32 N/A

      \[\leadsto \sqrt{\mathsf{neg}\left(\log \left(1 - u1\right)\right)} \]

   2. lower-neg.f32 N/A

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

   3. lower-log.f32 N/A

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

   4. lower--.f32 50.1%

      \[\leadsto \sqrt{-\log \left(1 - u1\right)} \]

4. Applied rewrites 50.1%

   \[\leadsto \color{blue}{\sqrt{-\log \left(1 - u1\right)}} \]

5. Taylor expanded in u1 around 0

   \[\leadsto \sqrt{-\log 1} \]
6. Step-by-step derivation

7. Applied rewrites 6.6%

   \[\leadsto \sqrt{-\log 1} \]
8. Add Preprocessing

Reproduce

herbie shell --seed 2025212 
(FPCore (cosTheta_i u1 u2)
  :name "Beckmann 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 (- (log (- 1.0 u1)))) (cos (* (* 2.0 PI) u2))))