
(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(fmax(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}
Herbie found 14 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(fmax(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}
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u_m))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* t_4 t_4))
(t_6 (* t_2 t_2))
(t_7 (* (floor h) dY.v)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) t_6) t_5)
(+ (+ (* t_1 t_1) (* t_7 t_7)) (* t_3 t_3)))))
100.0)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u_m (floor w)) dX.u_m)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(+ (+ (exp (* (log (* (- dX.u_m) (floor w))) 2.0)) t_6) t_5)
(fma
(pow dY.u 2.0)
(pow (floor w) 2.0)
(* (pow dY.w 2.0) (pow (floor d) 2.0)))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, 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) * dX_46_u_m;
float t_1 = floorf(w) * dY_46_u;
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 = t_4 * t_4;
float t_6 = t_2 * t_2;
float t_7 = floorf(h) * dY_46_v;
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + t_6) + t_5), (((t_1 * t_1) + (t_7 * t_7)) + (t_3 * t_3))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u_m * floorf(w)) * dX_46_u_m), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(((expf((logf((-dX_46_u_m * floorf(w))) * 2.0f)) + t_6) + t_5), fmaf(powf(dY_46_u, 2.0f), powf(floorf(w), 2.0f), (powf(dY_46_w, 2.0f) * powf(floorf(d), 2.0f))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u_m) t_1 = Float32(floor(w) * dY_46_u) 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(t_4 * t_4) t_6 = Float32(t_2 * t_2) t_7 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + t_6) + t_5), Float32(Float32(Float32(t_1 * t_1) + Float32(t_7 * t_7)) + Float32(t_3 * t_3))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u_m * floor(w)) * dX_46_u_m), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(Float32(Float32(exp(Float32(log(Float32(Float32(-dX_46_u_m) * floor(w))) * Float32(2.0))) + t_6) + t_5), fma((dY_46_u ^ Float32(2.0)), (floor(w) ^ Float32(2.0)), Float32((dY_46_w ^ Float32(2.0)) * (floor(d) ^ Float32(2.0))))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\_m\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
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 := t\_4 \cdot t\_4\\
t_6 := t\_2 \cdot t\_2\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_6\right) + t\_5, \left(t\_1 \cdot t\_1 + t\_7 \cdot t\_7\right) + t\_3 \cdot t\_3\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(e^{\log \left(\left(-dX.u\_m\right) \cdot \left\lfloor w\right\rfloor \right) \cdot 2} + t\_6\right) + t\_5, \mathsf{fma}\left({dY.u}^{2}, {\left(\left\lfloor w\right\rfloor \right)}^{2}, {dY.w}^{2} \cdot {\left(\left\lfloor d\right\rfloor \right)}^{2}\right)\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
Taylor expanded in dY.v around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3236.6
Applied rewrites36.6%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u_m))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* t_2 t_2))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor d) dY.w))
(t_6 (* t_5 t_5))
(t_7 (* (floor d) dX.w))
(t_8 (* dY.u (floor w)))
(t_9 (+ (* t_0 t_0) (* t_4 t_4))))
(if (<=
(log2 (sqrt (fmax (+ t_9 (* t_7 t_7)) (+ (+ (* t_1 t_1) t_3) t_6))))
100.0)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u_m (floor w)) dX.u_m)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* t_8 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(+ t_9 (exp (* (log (* (- dX.w) (floor d))) 2.0)))
(+ (+ (pow (exp 2.0) (log t_8)) t_3) t_6)))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, 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) * dX_46_u_m;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = t_2 * t_2;
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(d) * dY_46_w;
float t_6 = t_5 * t_5;
float t_7 = floorf(d) * dX_46_w;
float t_8 = dY_46_u * floorf(w);
float t_9 = (t_0 * t_0) + (t_4 * t_4);
float tmp;
if (log2f(sqrtf(fmaxf((t_9 + (t_7 * t_7)), (((t_1 * t_1) + t_3) + t_6)))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u_m * floorf(w)) * dX_46_u_m), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf((t_8 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf((t_9 + expf((logf((-dX_46_w * floorf(d))) * 2.0f))), ((powf(expf(2.0f), logf(t_8)) + t_3) + t_6))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u_m) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(t_2 * t_2) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(d) * dY_46_w) t_6 = Float32(t_5 * t_5) t_7 = Float32(floor(d) * dX_46_w) t_8 = Float32(dY_46_u * floor(w)) t_9 = Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(t_9 + Float32(t_7 * t_7)), Float32(Float32(Float32(t_1 * t_1) + t_3) + t_6)))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u_m * floor(w)) * dX_46_u_m), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(t_8 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(Float32(t_9 + exp(Float32(log(Float32(Float32(-dX_46_w) * floor(d))) * Float32(2.0)))), Float32(Float32((exp(Float32(2.0)) ^ log(t_8)) + t_3) + t_6)))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\_m\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := t\_2 \cdot t\_2\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_6 := t\_5 \cdot t\_5\\
t_7 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_8 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_9 := t\_0 \cdot t\_0 + t\_4 \cdot t\_4\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_9 + t\_7 \cdot t\_7, \left(t\_1 \cdot t\_1 + t\_3\right) + t\_6\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_8 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_9 + e^{\log \left(\left(-dX.w\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}, \left({\left(e^{2}\right)}^{\log t\_8} + t\_3\right) + t\_6\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3257.4
Applied rewrites57.4%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
exp-to-powN/A
lift-log.f32N/A
*-commutativeN/A
exp-prodN/A
lower-pow.f32N/A
lower-exp.f3240.0
Applied rewrites40.0%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u_m))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* t_2 t_2))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor d) dY.w))
(t_6 (* t_5 t_5))
(t_7 (* (floor d) dX.w))
(t_8 (* dY.u (floor w)))
(t_9 (+ (* t_0 t_0) (* t_4 t_4))))
(if (<=
(log2 (sqrt (fmax (+ t_9 (* t_7 t_7)) (+ (+ (* t_1 t_1) t_3) t_6))))
100.0)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u_m (floor w)) dX.u_m)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* t_8 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(+ t_9 (exp (* (log (* (- dX.w) (floor d))) 2.0)))
(+ (+ (exp (* 2.0 (log t_8))) t_3) t_6)))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, 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) * dX_46_u_m;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = t_2 * t_2;
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(d) * dY_46_w;
float t_6 = t_5 * t_5;
float t_7 = floorf(d) * dX_46_w;
float t_8 = dY_46_u * floorf(w);
float t_9 = (t_0 * t_0) + (t_4 * t_4);
float tmp;
if (log2f(sqrtf(fmaxf((t_9 + (t_7 * t_7)), (((t_1 * t_1) + t_3) + t_6)))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u_m * floorf(w)) * dX_46_u_m), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf((t_8 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf((t_9 + expf((logf((-dX_46_w * floorf(d))) * 2.0f))), ((expf((2.0f * logf(t_8))) + t_3) + t_6))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u_m) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(t_2 * t_2) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(d) * dY_46_w) t_6 = Float32(t_5 * t_5) t_7 = Float32(floor(d) * dX_46_w) t_8 = Float32(dY_46_u * floor(w)) t_9 = Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(t_9 + Float32(t_7 * t_7)), Float32(Float32(Float32(t_1 * t_1) + t_3) + t_6)))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u_m * floor(w)) * dX_46_u_m), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(t_8 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(Float32(t_9 + exp(Float32(log(Float32(Float32(-dX_46_w) * floor(d))) * Float32(2.0)))), Float32(Float32(exp(Float32(Float32(2.0) * log(t_8))) + t_3) + t_6)))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\_m\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := t\_2 \cdot t\_2\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_6 := t\_5 \cdot t\_5\\
t_7 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_8 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_9 := t\_0 \cdot t\_0 + t\_4 \cdot t\_4\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_9 + t\_7 \cdot t\_7, \left(t\_1 \cdot t\_1 + t\_3\right) + t\_6\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_8 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_9 + e^{\log \left(\left(-dX.w\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}, \left(e^{2 \cdot \log t\_8} + t\_3\right) + t\_6\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3257.4
Applied rewrites57.4%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3240.0
lift-*.f32N/A
*-commutativeN/A
lower-*.f3240.0
Applied rewrites40.0%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u_m))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))
(t_4 (* (floor d) dY.w))
(t_5 (* (floor d) dX.w))
(t_6 (* (* dX.w (floor d)) dX.w))
(t_7 (* (floor w) dY.u))
(t_8 (* (* dX.u_m (floor w)) dX.u_m)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_2 t_2)) (* t_5 t_5))
(+ (+ (* t_7 t_7) (* t_1 t_1)) (* t_4 t_4)))))
100.0)
(log2
(sqrt
(fmax
(fma
t_6
(floor d)
(fma t_8 (floor w) (* (* (* dX.v (floor h)) dX.v) (floor h))))
t_3)))
(log2
(sqrt
(fmax
(fma
t_6
(floor d)
(fma t_8 (floor w) (exp (* (log (* (- dX.v) (floor h))) 2.0))))
t_3))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, 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) * dX_46_u_m;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h))));
float t_4 = floorf(d) * dY_46_w;
float t_5 = floorf(d) * dX_46_w;
float t_6 = (dX_46_w * floorf(d)) * dX_46_w;
float t_7 = floorf(w) * dY_46_u;
float t_8 = (dX_46_u_m * floorf(w)) * dX_46_u_m;
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_2 * t_2)) + (t_5 * t_5)), (((t_7 * t_7) + (t_1 * t_1)) + (t_4 * t_4))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_6, floorf(d), fmaf(t_8, floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), t_3)));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_6, floorf(d), fmaf(t_8, floorf(w), expf((logf((-dX_46_v * floorf(h))) * 2.0f)))), t_3)));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u_m) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))) t_4 = Float32(floor(d) * dY_46_w) t_5 = Float32(floor(d) * dX_46_w) t_6 = Float32(Float32(dX_46_w * floor(d)) * dX_46_w) t_7 = Float32(floor(w) * dY_46_u) t_8 = Float32(Float32(dX_46_u_m * floor(w)) * dX_46_u_m) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) + Float32(t_5 * t_5)), Float32(Float32(Float32(t_7 * t_7) + Float32(t_1 * t_1)) + Float32(t_4 * t_4))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(t_6, floor(d), fma(t_8, floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), t_3))); else tmp = log2(sqrt(fmax(fma(t_6, floor(d), fma(t_8, floor(w), exp(Float32(log(Float32(Float32(-dX_46_v) * floor(h))) * Float32(2.0))))), t_3))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\_m\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_5 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_6 := \left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w\\
t_7 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_8 := \left(dX.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_2 \cdot t\_2\right) + t\_5 \cdot t\_5, \left(t\_7 \cdot t\_7 + t\_1 \cdot t\_1\right) + t\_4 \cdot t\_4\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), t\_3\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , e^{\log \left(\left(-dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot 2}\right)\right), t\_3\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3257.4
Applied rewrites57.4%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u_m))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor d) dY.w))
(t_5 (* (floor d) dX.w))
(t_6 (+ (* t_0 t_0) (* t_3 t_3)))
(t_7 (+ (+ (* t_1 t_1) (* t_2 t_2)) (* t_4 t_4))))
(if (<= (log2 (sqrt (fmax (+ t_6 (* t_5 t_5)) t_7))) 100.0)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u_m (floor w)) dX.u_m)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(log2
(sqrt (fmax (+ t_6 (exp (* (log (* (- dX.w) (floor d))) 2.0))) t_7))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, 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) * dX_46_u_m;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(d) * dY_46_w;
float t_5 = floorf(d) * dX_46_w;
float t_6 = (t_0 * t_0) + (t_3 * t_3);
float t_7 = ((t_1 * t_1) + (t_2 * t_2)) + (t_4 * t_4);
float tmp;
if (log2f(sqrtf(fmaxf((t_6 + (t_5 * t_5)), t_7))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u_m * floorf(w)) * dX_46_u_m), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf((t_6 + expf((logf((-dX_46_w * floorf(d))) * 2.0f))), t_7)));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u_m) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(d) * dY_46_w) t_5 = Float32(floor(d) * dX_46_w) t_6 = Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) t_7 = Float32(Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(t_6 + Float32(t_5 * t_5)), t_7))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u_m * floor(w)) * dX_46_u_m), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(Float32(t_6 + exp(Float32(log(Float32(Float32(-dX_46_w) * floor(d))) * Float32(2.0)))), t_7))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\_m\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_5 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_6 := t\_0 \cdot t\_0 + t\_3 \cdot t\_3\\
t_7 := \left(t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_6 + t\_5 \cdot t\_5, t\_7\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_6 + e^{\log \left(\left(-dX.w\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}, t\_7\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3257.4
Applied rewrites57.4%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u_m))
(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 (* dY.u (floor w)))
(t_6 (* (* dX.w (floor d)) dX.w))
(t_7 (* (floor w) dY.u))
(t_8 (* (* dX.u_m (floor w)) dX.u_m)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_7 t_7) (* t_1 t_1)) (* t_3 t_3)))))
100.0)
(log2
(sqrt
(fmax
(fma
t_6
(floor d)
(fma t_8 (floor w) (* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* t_5 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(fma
t_6
(floor d)
(fma
t_8
(floor w)
(*
(* (* (exp (log (- dX.v))) (exp (log (floor h)))) dX.v)
(floor h))))
(* dY.u (* (fabs t_5) (floor w)))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, 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) * dX_46_u_m;
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 = dY_46_u * floorf(w);
float t_6 = (dX_46_w * floorf(d)) * dX_46_w;
float t_7 = floorf(w) * dY_46_u;
float t_8 = (dX_46_u_m * floorf(w)) * dX_46_u_m;
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_2 * t_2)) + (t_4 * t_4)), (((t_7 * t_7) + (t_1 * t_1)) + (t_3 * t_3))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_6, floorf(d), fmaf(t_8, floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf((t_5 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_6, floorf(d), fmaf(t_8, floorf(w), (((expf(logf(-dX_46_v)) * expf(logf(floorf(h)))) * dX_46_v) * floorf(h)))), (dY_46_u * (fabsf(t_5) * floorf(w))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u_m) 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(dY_46_u * floor(w)) t_6 = Float32(Float32(dX_46_w * floor(d)) * dX_46_w) t_7 = Float32(floor(w) * dY_46_u) t_8 = Float32(Float32(dX_46_u_m * floor(w)) * dX_46_u_m) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_7 * t_7) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(t_6, floor(d), fma(t_8, floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(t_5 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(fma(t_6, floor(d), fma(t_8, floor(w), Float32(Float32(Float32(exp(log(Float32(-dX_46_v))) * exp(log(floor(h)))) * dX_46_v) * floor(h)))), Float32(dY_46_u * Float32(abs(t_5) * floor(w)))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\_m\\
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 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_6 := \left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w\\
t_7 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_8 := \left(dX.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_7 \cdot t\_7 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_5 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , \left(\left(e^{\log \left(-dX.v\right)} \cdot e^{\log \left(\left\lfloor h\right\rfloor \right)}\right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), dY.u \cdot \left(\left|t\_5\right| \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.5
Applied rewrites49.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3249.5
rem-exp-logN/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
rem-exp-logN/A
neg-fabsN/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-outN/A
lift-neg.f32N/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
log-prodN/A
exp-sumN/A
lower-*.f32N/A
Applied rewrites23.7%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u_m (floor w)) dX.u_m)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
return log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u_m * floorf(w)) * dX_46_u_m), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u_m * floor(w)) * dX_46_u_m), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))) end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)
\end{array}
Initial program 68.0%
Applied rewrites68.0%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dY.w (floor d)))
(t_2 (* (floor h) (floor h))))
(if (<= dY.u 60000002048.0)
(log2
(sqrt
(fmax
(fma
(* (floor d) (* (floor d) dX.w))
dX.w
(fma (* dX.v dX.v) t_2 (* (* dX.u_m dX.u_m) (* (floor w) (floor w)))))
(fma (* (* dY.v (floor h)) (floor h)) dY.v (* t_1 t_1)))))
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) (floor d))
dX.w
(* (fma (floor w) dX.u_m t_0) (- (* (floor w) dX.u_m) t_0)))
(fma
(* t_2 dY.v)
dY.v
(fma
(* t_1 (floor d))
dY.w
(* (* (* dY.u (floor w)) dY.u) (floor w))))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = dY_46_w * floorf(d);
float t_2 = floorf(h) * floorf(h);
float tmp;
if (dY_46_u <= 60000002048.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((floorf(d) * (floorf(d) * dX_46_w)), dX_46_w, fmaf((dX_46_v * dX_46_v), t_2, ((dX_46_u_m * dX_46_u_m) * (floorf(w) * floorf(w))))), fmaf(((dY_46_v * floorf(h)) * floorf(h)), dY_46_v, (t_1 * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * floorf(d)), dX_46_w, (fmaf(floorf(w), dX_46_u_m, t_0) * ((floorf(w) * dX_46_u_m) - t_0))), fmaf((t_2 * dY_46_v), dY_46_v, fmaf((t_1 * floorf(d)), dY_46_w, (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w)))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(dY_46_w * floor(d)) t_2 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dY_46_u <= Float32(60000002048.0)) tmp = log2(sqrt(fmax(fma(Float32(floor(d) * Float32(floor(d) * dX_46_w)), dX_46_w, fma(Float32(dX_46_v * dX_46_v), t_2, Float32(Float32(dX_46_u_m * dX_46_u_m) * Float32(floor(w) * floor(w))))), fma(Float32(Float32(dY_46_v * floor(h)) * floor(h)), dY_46_v, Float32(t_1 * t_1))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * floor(d)), dX_46_w, Float32(fma(floor(w), dX_46_u_m, t_0) * Float32(Float32(floor(w) * dX_46_u_m) - t_0))), fma(Float32(t_2 * dY_46_v), dY_46_v, fma(Float32(t_1 * floor(d)), dY_46_w, Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w))))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dY.w \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dY.u \leq 60000002048:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left(\left\lfloor d\right\rfloor \cdot dX.w\right), dX.w, \mathsf{fma}\left(dX.v \cdot dX.v, t\_2, \left(dX.u\_m \cdot dX.u\_m\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right), \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dY.v, t\_1 \cdot t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dX.w, \mathsf{fma}\left(\left\lfloor w\right\rfloor , dX.u\_m, t\_0\right) \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\_m - t\_0\right)\right), \mathsf{fma}\left(t\_2 \cdot dY.v, dY.v, \mathsf{fma}\left(t\_1 \cdot \left\lfloor d\right\rfloor , dY.w, \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 60000002000Initial program 68.0%
Applied rewrites68.0%
Taylor expanded in dY.u around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3260.9
Applied rewrites60.9%
Applied rewrites60.9%
if 60000002000 < dY.u Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3257.4
Applied rewrites57.4%
Applied rewrites60.6%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dY.v (floor h))) (t_1 (* dY.w (floor d))))
(if (<= dY.u 60000002048.0)
(log2
(sqrt
(fmax
(fma
(* (floor d) (* (floor d) dX.w))
dX.w
(fma
(* dX.v dX.v)
(* (floor h) (floor h))
(* (* dX.u_m dX.u_m) (* (floor w) (floor w)))))
(fma (* t_0 (floor h)) dY.v (* t_1 t_1)))))
(log2
(sqrt
(fmax
(* dX.w (* (fabs (* dX.w (floor d))) (floor d)))
(fma
(* t_1 dY.w)
(floor d)
(fma
(* (fabs (* dY.u (floor w))) dY.u)
(floor w)
(* (* t_0 dY.v) (floor h))))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = dY_46_v * floorf(h);
float t_1 = dY_46_w * floorf(d);
float tmp;
if (dY_46_u <= 60000002048.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((floorf(d) * (floorf(d) * dX_46_w)), dX_46_w, fmaf((dX_46_v * dX_46_v), (floorf(h) * floorf(h)), ((dX_46_u_m * dX_46_u_m) * (floorf(w) * floorf(w))))), fmaf((t_0 * floorf(h)), dY_46_v, (t_1 * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf((dX_46_w * (fabsf((dX_46_w * floorf(d))) * floorf(d))), fmaf((t_1 * dY_46_w), floorf(d), fmaf((fabsf((dY_46_u * floorf(w))) * dY_46_u), floorf(w), ((t_0 * dY_46_v) * floorf(h)))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(dY_46_v * floor(h)) t_1 = Float32(dY_46_w * floor(d)) tmp = Float32(0.0) if (dY_46_u <= Float32(60000002048.0)) tmp = log2(sqrt(fmax(fma(Float32(floor(d) * Float32(floor(d) * dX_46_w)), dX_46_w, fma(Float32(dX_46_v * dX_46_v), Float32(floor(h) * floor(h)), Float32(Float32(dX_46_u_m * dX_46_u_m) * Float32(floor(w) * floor(w))))), fma(Float32(t_0 * floor(h)), dY_46_v, Float32(t_1 * t_1))))); else tmp = log2(sqrt(fmax(Float32(dX_46_w * Float32(abs(Float32(dX_46_w * floor(d))) * floor(d))), fma(Float32(t_1 * dY_46_w), floor(d), fma(Float32(abs(Float32(dY_46_u * floor(w))) * dY_46_u), floor(w), Float32(Float32(t_0 * dY_46_v) * floor(h))))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dY.w \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dY.u \leq 60000002048:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left(\left\lfloor d\right\rfloor \cdot dX.w\right), dX.w, \mathsf{fma}\left(dX.v \cdot dX.v, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , \left(dX.u\_m \cdot dX.u\_m\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right), \mathsf{fma}\left(t\_0 \cdot \left\lfloor h\right\rfloor , dY.v, t\_1 \cdot t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(dX.w \cdot \left(\left|dX.w \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(t\_1 \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_0 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 60000002000Initial program 68.0%
Applied rewrites68.0%
Taylor expanded in dY.u around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3260.9
Applied rewrites60.9%
Applied rewrites60.9%
if 60000002000 < dY.u Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3259.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3259.1
Applied rewrites59.1%
Taylor expanded in dX.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3243.4
Applied rewrites43.4%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dX.w (floor d)))
(t_1 (fabs (* dY.u (floor w))))
(t_2 (* (floor h) dX.v)))
(if (<= dY.v 3.5)
(log2
(sqrt
(fmax
(fma
(* t_0 dX.w)
(floor d)
(fma (* (* dX.u_m dX.u_m) (floor w)) (floor w) (* t_2 t_2)))
(* dY.u (* t_1 (floor w))))))
(log2
(sqrt
(fmax
(* dX.w (* (fabs t_0) (floor d)))
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* t_1 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = dX_46_w * floorf(d);
float t_1 = fabsf((dY_46_u * floorf(w)));
float t_2 = floorf(h) * dX_46_v;
float tmp;
if (dY_46_v <= 3.5f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_w), floorf(d), fmaf(((dX_46_u_m * dX_46_u_m) * floorf(w)), floorf(w), (t_2 * t_2))), (dY_46_u * (t_1 * floorf(w))))));
} else {
tmp = log2f(sqrtf(fmaxf((dX_46_w * (fabsf(t_0) * floorf(d))), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf((t_1 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(dX_46_w * floor(d)) t_1 = abs(Float32(dY_46_u * floor(w))) t_2 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if (dY_46_v <= Float32(3.5)) tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u_m * dX_46_u_m) * floor(w)), floor(w), Float32(t_2 * t_2))), Float32(dY_46_u * Float32(t_1 * floor(w)))))); else tmp = log2(sqrt(fmax(Float32(dX_46_w * Float32(abs(t_0) * floor(d))), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(t_1 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_1 := \left|dY.u \cdot \left\lfloor w\right\rfloor \right|\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;dY.v \leq 3.5:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u\_m \cdot dX.u\_m\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_2 \cdot t\_2\right)\right), dY.u \cdot \left(t\_1 \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(dX.w \cdot \left(\left|t\_0\right| \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_1 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.v < 3.5Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.5
Applied rewrites49.5%
Applied rewrites49.5%
if 3.5 < dY.v Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3259.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3259.1
Applied rewrites59.1%
Taylor expanded in dX.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3243.4
Applied rewrites43.4%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v)))
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma (* (* dX.u_m dX.u_m) (floor w)) (floor w) (* t_0 t_0)))
(* dY.u (* (fabs (* dY.u (floor w))) (floor w))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
return log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u_m * dX_46_u_m) * floorf(w)), floorf(w), (t_0 * t_0))), (dY_46_u * (fabsf((dY_46_u * floorf(w))) * floorf(w))))));
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) return log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u_m * dX_46_u_m) * floor(w)), floor(w), Float32(t_0 * t_0))), Float32(dY_46_u * Float32(abs(Float32(dY_46_u * floor(w))) * floor(w)))))) end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u\_m \cdot dX.u\_m\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_0 \cdot t\_0\right)\right), dY.u \cdot \left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)}\right)
\end{array}
\end{array}
Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.5
Applied rewrites49.5%
Applied rewrites49.5%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v)))
(if (<= dY.u 60000002048.0)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u_m (floor w)) dX.u_m)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(* (* (* (- dY.u) (floor w)) (floor w)) dY.u))))
(log2
(sqrt
(fmax
(fma
(fma (floor w) dX.u_m t_0)
(- (* (floor w) dX.u_m) t_0)
(* (* dX.w dX.w) (* (floor d) (floor d))))
(* dY.u (* (fabs (* dY.u (floor w))) (floor w)))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
float tmp;
if (dY_46_u <= 60000002048.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u_m * floorf(w)) * dX_46_u_m), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (((-dY_46_u * floorf(w)) * floorf(w)) * dY_46_u))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(fmaf(floorf(w), dX_46_u_m, t_0), ((floorf(w) * dX_46_u_m) - t_0), ((dX_46_w * dX_46_w) * (floorf(d) * floorf(d)))), (dY_46_u * (fabsf((dY_46_u * floorf(w))) * floorf(w))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if (dY_46_u <= Float32(60000002048.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u_m * floor(w)) * dX_46_u_m), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(Float32(Float32(Float32(-dY_46_u) * floor(w)) * floor(w)) * dY_46_u)))); else tmp = log2(sqrt(fmax(fma(fma(floor(w), dX_46_u_m, t_0), Float32(Float32(floor(w) * dX_46_u_m) - t_0), Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d)))), Float32(dY_46_u * Float32(abs(Float32(dY_46_u * floor(w))) * floor(w)))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;dY.u \leq 60000002048:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \left(\left(\left(-dY.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\mathsf{fma}\left(\left\lfloor w\right\rfloor , dX.u\_m, t\_0\right), \left\lfloor w\right\rfloor \cdot dX.u\_m - t\_0, \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right), dY.u \cdot \left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 60000002000Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.5
Applied rewrites49.5%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3249.5
Applied rewrites45.9%
if 60000002000 < dY.u Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.5
Applied rewrites49.5%
Applied rewrites36.3%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v)))
(log2
(sqrt
(fmax
(fma
(fma (floor w) dX.u_m t_0)
(- (* (floor w) dX.u_m) t_0)
(* (* dX.w dX.w) (* (floor d) (floor d))))
(* dY.u (* (fabs (* dY.u (floor w))) (floor w))))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
return log2f(sqrtf(fmaxf(fmaf(fmaf(floorf(w), dX_46_u_m, t_0), ((floorf(w) * dX_46_u_m) - t_0), ((dX_46_w * dX_46_w) * (floorf(d) * floorf(d)))), (dY_46_u * (fabsf((dY_46_u * floorf(w))) * floorf(w))))));
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) return log2(sqrt(fmax(fma(fma(floor(w), dX_46_u_m, t_0), Float32(Float32(floor(w) * dX_46_u_m) - t_0), Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d)))), Float32(dY_46_u * Float32(abs(Float32(dY_46_u * floor(w))) * floor(w)))))) end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\mathsf{fma}\left(\left\lfloor w\right\rfloor , dX.u\_m, t\_0\right), \left\lfloor w\right\rfloor \cdot dX.u\_m - t\_0, \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right), dY.u \cdot \left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)}\right)
\end{array}
\end{array}
Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.5
Applied rewrites49.5%
Applied rewrites36.3%
dX.u_m = (fabs.f32 dX.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v)))
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) (floor d))
dX.w
(* (fma (floor w) dX.u_m t_0) (- (* (floor w) dX.u_m) t_0)))
(* (* (* (- dY.u) (floor w)) (floor w)) dY.u))))))dX.u_m = fabs(dX_46_u);
float code(float w, float h, float d, float dX_46_u_m, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
return log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * floorf(d)), dX_46_w, (fmaf(floorf(w), dX_46_u_m, t_0) * ((floorf(w) * dX_46_u_m) - t_0))), (((-dY_46_u * floorf(w)) * floorf(w)) * dY_46_u))));
}
dX.u_m = abs(dX_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) return log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * floor(d)), dX_46_w, Float32(fma(floor(w), dX_46_u_m, t_0) * Float32(Float32(floor(w) * dX_46_u_m) - t_0))), Float32(Float32(Float32(Float32(-dY_46_u) * floor(w)) * floor(w)) * dY_46_u)))) end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dX.w, \mathsf{fma}\left(\left\lfloor w\right\rfloor , dX.u\_m, t\_0\right) \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\_m - t\_0\right)\right), \left(\left(\left(-dY.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right)}\right)
\end{array}
\end{array}
Initial program 68.0%
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.5
Applied rewrites49.5%
Applied rewrites30.1%
herbie shell --seed 2025156
(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)))))))