
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(((Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)) != Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4))) ? Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3)) : ((Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3)) != Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))) ? Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)) : max(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 6 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(((Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)) != Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4))) ? Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3)) : ((Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3)) != Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))) ? Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)) : max(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(+
(pow (* (floor d) dX.w) 2.0)
(+ (pow (* (floor w) dX.u) 2.0) (pow (* (floor h) dX.v) 2.0)))
(+
(pow (* (floor w) dY.u) 2.0)
(+ (pow (* (floor d) dY.w) 2.0) (pow (* (floor h) dY.v) 2.0)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
return log2f(sqrtf(fmaxf((powf((floorf(d) * dX_46_w), 2.0f) + (powf((floorf(w) * dX_46_u), 2.0f) + powf((floorf(h) * dX_46_v), 2.0f))), (powf((floorf(w) * dY_46_u), 2.0f) + (powf((floorf(d) * dY_46_w), 2.0f) + powf((floorf(h) * dY_46_v), 2.0f))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt(((Float32((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) + Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0)))) != Float32((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) + Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))))) ? Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) : ((Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) != Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))) ? Float32((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) + Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0)))) : max(Float32((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) + Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0)))), Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) tmp = log2(sqrt(max((((floor(d) * dX_46_w) ^ single(2.0)) + (((floor(w) * dX_46_u) ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0)))), (((floor(w) * dY_46_u) ^ single(2.0)) + (((floor(d) * dY_46_w) ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0))))))); end
\begin{array}{l}
\\
\log_{2} \left(\sqrt{\mathsf{max}\left({\left(\left\lfloor d\right\rfloor \cdot dX.w\right)}^{2} + \left({\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\right), {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + \left({\left(\left\lfloor d\right\rfloor \cdot dY.w\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\right)\right)}\right)
\end{array}
Initial program 66.4%
Simplified66.4%
Applied egg-rr66.4%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0
(+
(pow (* (floor w) dY.u) 2.0)
(+ (pow (* (floor d) dY.w) 2.0) (pow (* (floor h) dY.v) 2.0)))))
(if (<= dX.v 0.800000011920929)
(log2
(sqrt
(fmax
(+ (pow (* (floor d) dX.w) 2.0) (pow (* (floor w) dX.u) 2.0))
t_0)))
(log2 (sqrt (fmax (pow (* (floor h) dX.v) 2.0) t_0))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = powf((floorf(w) * dY_46_u), 2.0f) + (powf((floorf(d) * dY_46_w), 2.0f) + powf((floorf(h) * dY_46_v), 2.0f));
float tmp;
if (dX_46_v <= 0.800000011920929f) {
tmp = log2f(sqrtf(fmaxf((powf((floorf(d) * dX_46_w), 2.0f) + powf((floorf(w) * dX_46_u), 2.0f)), t_0)));
} else {
tmp = log2f(sqrtf(fmaxf(powf((floorf(h) * dX_46_v), 2.0f), t_0)));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) tmp = Float32(0.0) if (dX_46_v <= Float32(0.800000011920929)) tmp = log2(sqrt(((Float32((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) != Float32((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0)))) ? t_0 : ((t_0 != t_0) ? Float32((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) : max(Float32((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))), t_0))))); else tmp = log2(sqrt((((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) != (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) ? t_0 : ((t_0 != t_0) ? (Float32(floor(h) * dX_46_v) ^ Float32(2.0)) : max((Float32(floor(h) * dX_46_v) ^ Float32(2.0)), t_0))))); end return tmp end
function tmp_2 = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = ((floor(w) * dY_46_u) ^ single(2.0)) + (((floor(d) * dY_46_w) ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0))); tmp = single(0.0); if (dX_46_v <= single(0.800000011920929)) tmp = log2(sqrt(max((((floor(d) * dX_46_w) ^ single(2.0)) + ((floor(w) * dX_46_u) ^ single(2.0))), t_0))); else tmp = log2(sqrt(max(((floor(h) * dX_46_v) ^ single(2.0)), t_0))); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + \left({\left(\left\lfloor d\right\rfloor \cdot dY.w\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\right)\\
\mathbf{if}\;dX.v \leq 0.800000011920929:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left({\left(\left\lfloor d\right\rfloor \cdot dX.w\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, t\_0\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}, t\_0\right)}\right)\\
\end{array}
\end{array}
if dX.v < 0.800000012Initial program 67.3%
Simplified67.3%
Applied egg-rr67.3%
Taylor expanded in dX.v around 0
*-rgt-identityN/A
*-inversesN/A
associate-/l*N/A
associate-*l/N/A
+-lowering-+.f32N/A
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f32N/A
associate-*l/N/A
associate-/l*N/A
*-inversesN/A
*-rgt-identityN/A
Simplified62.0%
sqrt-lowering-sqrt.f32N/A
+-commutativeN/A
fmax-lowering-fmax.f32N/A
Applied egg-rr62.0%
if 0.800000012 < dX.v Initial program 64.1%
Taylor expanded in dX.v around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3258.4%
Simplified58.4%
sqrt-lowering-sqrt.f32N/A
associate-*r*N/A
pow2N/A
unpow-prod-downN/A
*-commutativeN/A
fmax-lowering-fmax.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
Applied egg-rr58.4%
Final simplification61.0%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(+ (pow (* (floor w) dX.u) 2.0) (pow (* (floor h) dX.v) 2.0))
(+
(pow (* (floor w) dY.u) 2.0)
(+ (pow (* (floor d) dY.w) 2.0) (pow (* (floor h) dY.v) 2.0)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
return log2f(sqrtf(fmaxf((powf((floorf(w) * dX_46_u), 2.0f) + powf((floorf(h) * dX_46_v), 2.0f)), (powf((floorf(w) * dY_46_u), 2.0f) + (powf((floorf(d) * dY_46_w), 2.0f) + powf((floorf(h) * dY_46_v), 2.0f))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt(((Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) != Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0)))) ? Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) : ((Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) != Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))) ? Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) : max(Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))), Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) tmp = log2(sqrt(max((((floor(w) * dX_46_u) ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0))), (((floor(w) * dY_46_u) ^ single(2.0)) + (((floor(d) * dY_46_w) ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0))))))); end
\begin{array}{l}
\\
\log_{2} \left(\sqrt{\mathsf{max}\left({\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}, {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + \left({\left(\left\lfloor d\right\rfloor \cdot dY.w\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\right)\right)}\right)
\end{array}
Initial program 66.4%
Taylor expanded in dX.w around 0
*-lft-identityN/A
*-inversesN/A
*-commutativeN/A
associate-/l*N/A
associate-*l/N/A
*-commutativeN/A
+-lowering-+.f32N/A
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f32N/A
associate-*r/N/A
Simplified62.8%
Applied egg-rr62.8%
Final simplification62.8%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0
(+
(pow (* (floor w) dY.u) 2.0)
(+ (pow (* (floor d) dY.w) 2.0) (pow (* (floor h) dY.v) 2.0)))))
(if (<= dX.v 2.0)
(log2 (sqrt (fmax (pow (* (floor w) dX.u) 2.0) t_0)))
(log2 (sqrt (fmax (pow (* (floor h) dX.v) 2.0) t_0))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = powf((floorf(w) * dY_46_u), 2.0f) + (powf((floorf(d) * dY_46_w), 2.0f) + powf((floorf(h) * dY_46_v), 2.0f));
float tmp;
if (dX_46_v <= 2.0f) {
tmp = log2f(sqrtf(fmaxf(powf((floorf(w) * dX_46_u), 2.0f), t_0)));
} else {
tmp = log2f(sqrtf(fmaxf(powf((floorf(h) * dX_46_v), 2.0f), t_0)));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) tmp = Float32(0.0) if (dX_46_v <= Float32(2.0)) tmp = log2(sqrt((((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) != (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) ? t_0 : ((t_0 != t_0) ? (Float32(floor(w) * dX_46_u) ^ Float32(2.0)) : max((Float32(floor(w) * dX_46_u) ^ Float32(2.0)), t_0))))); else tmp = log2(sqrt((((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) != (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) ? t_0 : ((t_0 != t_0) ? (Float32(floor(h) * dX_46_v) ^ Float32(2.0)) : max((Float32(floor(h) * dX_46_v) ^ Float32(2.0)), t_0))))); end return tmp end
function tmp_2 = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = ((floor(w) * dY_46_u) ^ single(2.0)) + (((floor(d) * dY_46_w) ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0))); tmp = single(0.0); if (dX_46_v <= single(2.0)) tmp = log2(sqrt(max(((floor(w) * dX_46_u) ^ single(2.0)), t_0))); else tmp = log2(sqrt(max(((floor(h) * dX_46_v) ^ single(2.0)), t_0))); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + \left({\left(\left\lfloor d\right\rfloor \cdot dY.w\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\right)\\
\mathbf{if}\;dX.v \leq 2:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left({\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, t\_0\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}, t\_0\right)}\right)\\
\end{array}
\end{array}
if dX.v < 2Initial program 67.0%
Simplified66.9%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3257.1%
Simplified57.1%
sqrt-lowering-sqrt.f32N/A
associate-*r*N/A
pow2N/A
*-commutativeN/A
unpow-prod-downN/A
associate-*r*N/A
swap-sqrN/A
associate-*r*N/A
swap-sqrN/A
Applied egg-rr57.2%
if 2 < dX.v Initial program 64.9%
Taylor expanded in dX.v around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3258.9%
Simplified58.9%
sqrt-lowering-sqrt.f32N/A
associate-*r*N/A
pow2N/A
unpow-prod-downN/A
*-commutativeN/A
fmax-lowering-fmax.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
Applied egg-rr58.9%
Final simplification57.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(pow (* (floor h) dX.v) 2.0)
(+
(pow (* (floor w) dY.u) 2.0)
(+ (pow (* (floor d) dY.w) 2.0) (pow (* (floor h) dY.v) 2.0)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
return log2f(sqrtf(fmaxf(powf((floorf(h) * dX_46_v), 2.0f), (powf((floorf(w) * dY_46_u), 2.0f) + (powf((floorf(d) * dY_46_w), 2.0f) + powf((floorf(h) * dY_46_v), 2.0f))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt((((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) != (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) ? Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) : ((Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) != Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))) ? (Float32(floor(h) * dX_46_v) ^ Float32(2.0)) : max((Float32(floor(h) * dX_46_v) ^ Float32(2.0)), Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) tmp = log2(sqrt(max(((floor(h) * dX_46_v) ^ single(2.0)), (((floor(w) * dY_46_u) ^ single(2.0)) + (((floor(d) * dY_46_w) ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0))))))); end
\begin{array}{l}
\\
\log_{2} \left(\sqrt{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}, {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + \left({\left(\left\lfloor d\right\rfloor \cdot dY.w\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\right)\right)}\right)
\end{array}
Initial program 66.4%
Taylor expanded in dX.v around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3254.3%
Simplified54.3%
sqrt-lowering-sqrt.f32N/A
associate-*r*N/A
pow2N/A
unpow-prod-downN/A
*-commutativeN/A
fmax-lowering-fmax.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
Applied egg-rr54.3%
Final simplification54.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(pow (* (floor d) dX.w) 2.0)
(+
(pow (* (floor w) dY.u) 2.0)
(+ (pow (* (floor d) dY.w) 2.0) (pow (* (floor h) dY.v) 2.0)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
return log2f(sqrtf(fmaxf(powf((floorf(d) * dX_46_w), 2.0f), (powf((floorf(w) * dY_46_u), 2.0f) + (powf((floorf(d) * dY_46_w), 2.0f) + powf((floorf(h) * dY_46_v), 2.0f))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt((((Float32(floor(d) * dX_46_w) ^ Float32(2.0)) != (Float32(floor(d) * dX_46_w) ^ Float32(2.0))) ? Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) : ((Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) != Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))) ? (Float32(floor(d) * dX_46_w) ^ Float32(2.0)) : max((Float32(floor(d) * dX_46_w) ^ Float32(2.0)), Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + Float32((Float32(floor(d) * dY_46_w) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) tmp = log2(sqrt(max(((floor(d) * dX_46_w) ^ single(2.0)), (((floor(w) * dY_46_u) ^ single(2.0)) + (((floor(d) * dY_46_w) ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0))))))); end
\begin{array}{l}
\\
\log_{2} \left(\sqrt{\mathsf{max}\left({\left(\left\lfloor d\right\rfloor \cdot dX.w\right)}^{2}, {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2} + \left({\left(\left\lfloor d\right\rfloor \cdot dY.w\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\right)\right)}\right)
\end{array}
Initial program 66.4%
Taylor expanded in dX.w around inf
*-lowering-*.f32N/A
unpow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3250.7%
Simplified50.7%
sqrt-lowering-sqrt.f32N/A
*-commutativeN/A
pow2N/A
unpow-prod-downN/A
fmax-lowering-fmax.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
associate-+l+N/A
Applied egg-rr50.7%
Final simplification50.7%
herbie shell --seed 2024161
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:name "Isotropic LOD (LOD)"
:precision binary32
:pre (and (and (and (and (and (and (and (and (and (<= 1.0 w) (<= w 16384.0)) (and (<= 1.0 h) (<= h 16384.0))) (and (<= 1.0 d) (<= d 4096.0))) (and (<= 1e-20 (fabs dX.u)) (<= (fabs dX.u) 1e+20))) (and (<= 1e-20 (fabs dX.v)) (<= (fabs dX.v) 1e+20))) (and (<= 1e-20 (fabs dX.w)) (<= (fabs dX.w) 1e+20))) (and (<= 1e-20 (fabs dY.u)) (<= (fabs dY.u) 1e+20))) (and (<= 1e-20 (fabs dY.v)) (<= (fabs dY.v) 1e+20))) (and (<= 1e-20 (fabs dY.w)) (<= (fabs dY.w) 1e+20)))
(log2 (sqrt (fmax (+ (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (* (* (floor d) dX.w) (* (floor d) dX.w))) (+ (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v))) (* (* (floor d) dY.w) (* (floor d) dY.w)))))))