
(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 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(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)
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u_m))
(t_1 (* (floor w) dY.u_m))
(t_2 (* dY.v (floor h)))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor d) dY.w))
(t_6 (* (floor d) dX.w))
(t_7 (* t_6 t_6))
(t_8 (* dY.w (floor d))))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_4 t_4)) t_7)
(+ (+ (* t_1 t_1) (* t_3 t_3)) (* t_5 t_5)))))
63.95000076293945)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(*
(* (sqrt (* (* dX.u_m dX.u_m) (floor w))) (sqrt (floor w)))
dX.u_m)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* t_8 dY.w)
(floor d)
(fma
(* (* dY.u_m (floor w)) dY.u_m)
(floor w)
(* (* t_2 dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(fma (* (* (floor h) (floor h)) dX.v) dX.v t_7)
(fma
t_8
t_8
(fma t_2 t_2 (exp (* (log (* (- dY.u_m) (floor w))) 2.0))))))))))dX.u_m = fabs(dX_46_u);
dY.u_m = fabs(dY_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_m, 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_m;
float t_2 = dY_46_v * floorf(h);
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(d) * dY_46_w;
float t_6 = floorf(d) * dX_46_w;
float t_7 = t_6 * t_6;
float t_8 = dY_46_w * floorf(d);
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_4 * t_4)) + t_7), (((t_1 * t_1) + (t_3 * t_3)) + (t_5 * t_5))))) <= 63.95000076293945f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((sqrtf(((dX_46_u_m * dX_46_u_m) * floorf(w))) * sqrtf(floorf(w))) * dX_46_u_m), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf((t_8 * dY_46_w), floorf(d), fmaf(((dY_46_u_m * floorf(w)) * dY_46_u_m), floorf(w), ((t_2 * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, t_7), fmaf(t_8, t_8, fmaf(t_2, t_2, expf((logf((-dY_46_u_m * floorf(w))) * 2.0f)))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u_m) t_1 = Float32(floor(w) * dY_46_u_m) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(d) * dY_46_w) t_6 = Float32(floor(d) * dX_46_w) t_7 = Float32(t_6 * t_6) t_8 = Float32(dY_46_w * floor(d)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) + t_7), Float32(Float32(Float32(t_1 * t_1) + Float32(t_3 * t_3)) + Float32(t_5 * t_5))))) <= Float32(63.95000076293945)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(sqrt(Float32(Float32(dX_46_u_m * dX_46_u_m) * floor(w))) * sqrt(floor(w))) * dX_46_u_m), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(t_8 * dY_46_w), floor(d), fma(Float32(Float32(dY_46_u_m * floor(w)) * dY_46_u_m), floor(w), Float32(Float32(t_2 * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, t_7), fma(t_8, t_8, fma(t_2, t_2, exp(Float32(log(Float32(Float32(-dY_46_u_m) * floor(w))) * Float32(2.0)))))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
dY.u_m = \left|dY.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\_m\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_6 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_7 := t\_6 \cdot t\_6\\
t_8 := dY.w \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_4 \cdot t\_4\right) + t\_7, \left(t\_1 \cdot t\_1 + t\_3 \cdot t\_3\right) + t\_5 \cdot t\_5\right)}\right) \leq 63.95000076293945:\\
\;\;\;\;\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(\sqrt{\left(dX.u\_m \cdot dX.u\_m\right) \cdot \left\lfloor w\right\rfloor } \cdot \sqrt{\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(t\_8 \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\_m, \left\lfloor w\right\rfloor , \left(t\_2 \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(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, dX.v, t\_7\right), \mathsf{fma}\left(t\_8, t\_8, \mathsf{fma}\left(t\_2, t\_2, e^{\log \left(\left(-dY.u\_m\right) \cdot \left\lfloor w\right\rfloor \right) \cdot 2}\right)\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)))))) < 63.9500008Initial program 67.1%
Applied rewrites67.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.1
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
rem-sqrt-square-revN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
sqrt-prodN/A
lower-special-*.f32N/A
Applied rewrites67.1%
if 63.9500008 < (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 67.1%
Taylor expanded in dX.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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-special-exp.f32N/A
lower-special-*.f32N/A
lower-special-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3235.9
Applied rewrites35.9%
dX.u_m = (fabs.f32 dX.u)
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u_m))
(t_1 (* (floor w) dY.u_m))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor d) dY.w))
(t_5 (* dY.w (floor d)))
(t_6 (* (floor d) dX.w))
(t_7 (* t_6 t_6))
(t_8 (* dY.v (floor h))))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_3 t_3)) t_7)
(+ (+ (* t_1 t_1) (* t_2 t_2)) (* t_4 t_4)))))
63.95000076293945)
(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
(* t_5 dY.w)
(floor d)
(fma
(* (* dY.u_m (floor w)) dY.u_m)
(floor w)
(* (* t_8 dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(fma (* (* (floor h) (floor h)) dX.v) dX.v t_7)
(fma
t_5
t_5
(fma t_8 t_8 (exp (* (log (* (- dY.u_m) (floor w))) 2.0))))))))))dX.u_m = fabs(dX_46_u);
dY.u_m = fabs(dY_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_m, 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_m;
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 = dY_46_w * floorf(d);
float t_6 = floorf(d) * dX_46_w;
float t_7 = t_6 * t_6;
float t_8 = dY_46_v * floorf(h);
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_3 * t_3)) + t_7), (((t_1 * t_1) + (t_2 * t_2)) + (t_4 * t_4))))) <= 63.95000076293945f) {
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((t_5 * dY_46_w), floorf(d), fmaf(((dY_46_u_m * floorf(w)) * dY_46_u_m), floorf(w), ((t_8 * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, t_7), fmaf(t_5, t_5, fmaf(t_8, t_8, expf((logf((-dY_46_u_m * floorf(w))) * 2.0f)))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u_m) t_1 = Float32(floor(w) * dY_46_u_m) 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(dY_46_w * floor(d)) t_6 = Float32(floor(d) * dX_46_w) t_7 = Float32(t_6 * t_6) t_8 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(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))))) <= Float32(63.95000076293945)) 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(t_5 * dY_46_w), floor(d), fma(Float32(Float32(dY_46_u_m * floor(w)) * dY_46_u_m), floor(w), Float32(Float32(t_8 * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, t_7), fma(t_5, t_5, fma(t_8, t_8, exp(Float32(log(Float32(Float32(-dY_46_u_m) * floor(w))) * Float32(2.0)))))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
dY.u_m = \left|dY.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\_m\\
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 := dY.w \cdot \left\lfloor d\right\rfloor \\
t_6 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_7 := t\_6 \cdot t\_6\\
t_8 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_3 \cdot t\_3\right) + t\_7, \left(t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4\right)}\right) \leq 63.95000076293945:\\
\;\;\;\;\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(t\_5 \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\_m, \left\lfloor w\right\rfloor , \left(t\_8 \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(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, dX.v, t\_7\right), \mathsf{fma}\left(t\_5, t\_5, \mathsf{fma}\left(t\_8, t\_8, e^{\log \left(\left(-dY.u\_m\right) \cdot \left\lfloor w\right\rfloor \right) \cdot 2}\right)\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)))))) < 63.9500008Initial program 67.1%
Applied rewrites67.1%
if 63.9500008 < (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 67.1%
Taylor expanded in dX.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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-special-exp.f32N/A
lower-special-*.f32N/A
lower-special-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3235.9
Applied rewrites35.9%
dX.u_m = (fabs.f32 dX.u)
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor w) dY.u_m))
(t_3 (* (floor d) dY.w))
(t_4 (* dY.v (floor h))))
(if (<= dX.u_m 40000.0)
(log2
(sqrt
(fmax
(fma (* (* (floor d) (floor d)) dX.w) dX.w (* t_1 t_1))
(+ (+ (* t_2 t_2) (* t_0 t_0)) (* t_3 t_3)))))
(log2
(sqrt
(fmax
(fma
(* (* (floor w) (floor w)) dX.u_m)
dX.u_m
(fma
(* (* (floor d) dX.w) (floor d))
dX.w
(* (* dX.v dX.v) (* (floor h) (floor h)))))
(fma
(fma dY.u_m (floor w) t_4)
(- (* dY.u_m (floor w)) t_4)
(* (* (* dY.w (floor d)) dY.w) (floor d)))))))))dX.u_m = fabs(dX_46_u);
dY.u_m = fabs(dY_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_m, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dY_46_v;
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u_m;
float t_3 = floorf(d) * dY_46_w;
float t_4 = dY_46_v * floorf(h);
float tmp;
if (dX_46_u_m <= 40000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * floorf(d)) * dX_46_w), dX_46_w, (t_1 * t_1)), (((t_2 * t_2) + (t_0 * t_0)) + (t_3 * t_3)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * floorf(w)) * dX_46_u_m), dX_46_u_m, fmaf(((floorf(d) * dX_46_w) * floorf(d)), dX_46_w, ((dX_46_v * dX_46_v) * (floorf(h) * floorf(h))))), fmaf(fmaf(dY_46_u_m, floorf(w), t_4), ((dY_46_u_m * floorf(w)) - t_4), (((dY_46_w * floorf(d)) * dY_46_w) * floorf(d))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dY_46_v) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u_m) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (dX_46_u_m <= Float32(40000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * floor(d)) * dX_46_w), dX_46_w, Float32(t_1 * t_1)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) + Float32(t_3 * t_3))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u_m), dX_46_u_m, fma(Float32(Float32(floor(d) * dX_46_w) * floor(d)), dX_46_w, Float32(Float32(dX_46_v * dX_46_v) * Float32(floor(h) * floor(h))))), fma(fma(dY_46_u_m, floor(w), t_4), Float32(Float32(dY_46_u_m * floor(w)) - t_4), Float32(Float32(Float32(dY_46_w * floor(d)) * dY_46_w) * floor(d)))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.u\_m \leq 40000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, dX.w, t\_1 \cdot t\_1\right), \left(t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right) + t\_3 \cdot t\_3\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\_m, dX.u\_m, \mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor , dX.w, \left(dX.v \cdot dX.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\right), \mathsf{fma}\left(\mathsf{fma}\left(dY.u\_m, \left\lfloor w\right\rfloor , t\_4\right), dY.u\_m \cdot \left\lfloor w\right\rfloor - t\_4, \left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 4e4Initial program 67.1%
Taylor expanded in dX.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.f3260.0
Applied rewrites60.0%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites60.0%
if 4e4 < dX.u Initial program 67.1%
Applied rewrites67.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.1
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.f3263.5
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.5
Applied rewrites63.5%
Applied rewrites59.7%
dX.u_m = (fabs.f32 dX.u)
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* dY.w (floor d)))
(t_1 (* dY.v (floor h)))
(t_2 (* dY.u_m (floor w)))
(t_3 (* (floor d) (floor d)))
(t_4 (* t_3 (* dX.w dX.w))))
(if (<= dX.v 3000000000.0)
(log2
(sqrt
(fmax
(fma (* (* (floor w) dX.u_m) (floor w)) dX.u_m t_4)
(fma
(* (* (floor w) (floor w)) dY.u_m)
dY.u_m
(fma (* t_3 dY.w) dY.w (* (* dY.v dY.v) (* (floor h) (floor h))))))))
(log2
(sqrt
(fmax
(fma (* (* (floor h) dX.v) (floor h)) dX.v t_4)
(fma t_0 t_0 (fma t_1 t_1 (* t_2 t_2)))))))))dX.u_m = fabs(dX_46_u);
dY.u_m = fabs(dY_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_m, float dY_46_v, float dY_46_w) {
float t_0 = dY_46_w * floorf(d);
float t_1 = dY_46_v * floorf(h);
float t_2 = dY_46_u_m * floorf(w);
float t_3 = floorf(d) * floorf(d);
float t_4 = t_3 * (dX_46_w * dX_46_w);
float tmp;
if (dX_46_v <= 3000000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u_m) * floorf(w)), dX_46_u_m, t_4), fmaf(((floorf(w) * floorf(w)) * dY_46_u_m), dY_46_u_m, fmaf((t_3 * dY_46_w), dY_46_w, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(h) * dX_46_v) * floorf(h)), dX_46_v, t_4), fmaf(t_0, t_0, fmaf(t_1, t_1, (t_2 * t_2))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(dY_46_w * floor(d)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dY_46_u_m * floor(w)) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(t_3 * Float32(dX_46_w * dX_46_w)) tmp = Float32(0.0) if (dX_46_v <= Float32(3000000000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u_m) * floor(w)), dX_46_u_m, t_4), fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u_m), dY_46_u_m, fma(Float32(t_3 * dY_46_w), dY_46_w, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(h) * dX_46_v) * floor(h)), dX_46_v, t_4), fma(t_0, t_0, fma(t_1, t_1, Float32(t_2 * t_2)))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := dY.w \cdot \left\lfloor d\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dY.u\_m \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := t\_3 \cdot \left(dX.w \cdot dX.w\right)\\
\mathbf{if}\;dX.v \leq 3000000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\_m\right) \cdot \left\lfloor w\right\rfloor , dX.u\_m, t\_4\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\_m, dY.u\_m, \mathsf{fma}\left(t\_3 \cdot dY.w, dY.w, \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , dX.v, t\_4\right), \mathsf{fma}\left(t\_0, t\_0, \mathsf{fma}\left(t\_1, t\_1, t\_2 \cdot t\_2\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.v < 3e9Initial program 67.1%
Applied rewrites67.1%
Taylor expanded in dX.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.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
if 3e9 < dX.v Initial program 67.1%
Taylor expanded in dX.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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3260.0
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3260.0
lift-pow.f32N/A
unpow2N/A
lower-*.f3260.0
lift-pow.f32N/A
unpow2N/A
lower-*.f3260.0
Applied rewrites60.0%
dX.u_m = (fabs.f32 dX.u)
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* dY.u_m (floor w)))
(t_2 (* (floor d) (floor d)))
(t_3 (* dY.w (floor d))))
(if (<= dX.u_m 5500000.0)
(log2
(sqrt
(fmax
(fma (* (* (floor h) dX.v) (floor h)) dX.v (* t_2 (* dX.w dX.w)))
(fma t_3 t_3 (fma t_0 t_0 (* t_1 t_1))))))
(log2
(sqrt
(fmax
(fma (* (* (floor w) dX.u_m) (floor w)) dX.u_m (* (* dX.w dX.w) t_2))
(fma
(fma dY.u_m (floor w) t_0)
(- t_1 t_0)
(* (* t_3 dY.w) (floor d)))))))))dX.u_m = fabs(dX_46_u);
dY.u_m = fabs(dY_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_m, float dY_46_v, float dY_46_w) {
float t_0 = dY_46_v * floorf(h);
float t_1 = dY_46_u_m * floorf(w);
float t_2 = floorf(d) * floorf(d);
float t_3 = dY_46_w * floorf(d);
float tmp;
if (dX_46_u_m <= 5500000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(h) * dX_46_v) * floorf(h)), dX_46_v, (t_2 * (dX_46_w * dX_46_w))), fmaf(t_3, t_3, fmaf(t_0, t_0, (t_1 * t_1))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u_m) * floorf(w)), dX_46_u_m, ((dX_46_w * dX_46_w) * t_2)), fmaf(fmaf(dY_46_u_m, floorf(w), t_0), (t_1 - t_0), ((t_3 * dY_46_w) * floorf(d))))));
}
return tmp;
}
dX.u_m = abs(dX_46_u) dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(dY_46_v * floor(h)) t_1 = Float32(dY_46_u_m * floor(w)) t_2 = Float32(floor(d) * floor(d)) t_3 = Float32(dY_46_w * floor(d)) tmp = Float32(0.0) if (dX_46_u_m <= Float32(5500000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(h) * dX_46_v) * floor(h)), dX_46_v, Float32(t_2 * Float32(dX_46_w * dX_46_w))), fma(t_3, t_3, fma(t_0, t_0, Float32(t_1 * t_1)))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u_m) * floor(w)), dX_46_u_m, Float32(Float32(dX_46_w * dX_46_w) * t_2)), fma(fma(dY_46_u_m, floor(w), t_0), Float32(t_1 - t_0), Float32(Float32(t_3 * dY_46_w) * floor(d)))))); end return tmp end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dY.u\_m \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_3 := dY.w \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dX.u\_m \leq 5500000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , dX.v, t\_2 \cdot \left(dX.w \cdot dX.w\right)\right), \mathsf{fma}\left(t\_3, t\_3, \mathsf{fma}\left(t\_0, t\_0, t\_1 \cdot t\_1\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\_m\right) \cdot \left\lfloor w\right\rfloor , dX.u\_m, \left(dX.w \cdot dX.w\right) \cdot t\_2\right), \mathsf{fma}\left(\mathsf{fma}\left(dY.u\_m, \left\lfloor w\right\rfloor , t\_0\right), t\_1 - t\_0, \left(t\_3 \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 5.5e6Initial program 67.1%
Taylor expanded in dX.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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3260.0
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3260.0
lift-pow.f32N/A
unpow2N/A
lower-*.f3260.0
lift-pow.f32N/A
unpow2N/A
lower-*.f3260.0
Applied rewrites60.0%
if 5.5e6 < dX.u Initial program 67.1%
Applied rewrites67.1%
Taylor expanded in dX.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.f3260.3
Applied rewrites60.3%
rem-exp-logN/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
rem-exp-logN/A
lower-fabs.f3256.1
Applied rewrites56.1%
Applied rewrites51.7%
dX.u_m = (fabs.f32 dX.u)
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u_m dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* dY.v (floor h))))
(log2
(sqrt
(fmax
(fma
(* (* (floor w) dX.u_m) (floor w))
dX.u_m
(* (* dX.w dX.w) (* (floor d) (floor d))))
(fma
(fma dY.u_m (floor w) t_0)
(- (* dY.u_m (floor w)) t_0)
(* (* (* dY.w (floor d)) dY.w) (floor d))))))))dX.u_m = fabs(dX_46_u);
dY.u_m = fabs(dY_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_m, float dY_46_v, float dY_46_w) {
float t_0 = dY_46_v * floorf(h);
return log2f(sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u_m) * floorf(w)), dX_46_u_m, ((dX_46_w * dX_46_w) * (floorf(d) * floorf(d)))), fmaf(fmaf(dY_46_u_m, floorf(w), t_0), ((dY_46_u_m * floorf(w)) - t_0), (((dY_46_w * floorf(d)) * dY_46_w) * floorf(d))))));
}
dX.u_m = abs(dX_46_u) dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u_m, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(dY_46_v * floor(h)) return log2(sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u_m) * floor(w)), dX_46_u_m, Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d)))), fma(fma(dY_46_u_m, floor(w), t_0), Float32(Float32(dY_46_u_m * floor(w)) - t_0), Float32(Float32(Float32(dY_46_w * floor(d)) * dY_46_w) * floor(d)))))) end
\begin{array}{l}
dX.u_m = \left|dX.u\right|
\\
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\_m\right) \cdot \left\lfloor w\right\rfloor , dX.u\_m, \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right), \mathsf{fma}\left(\mathsf{fma}\left(dY.u\_m, \left\lfloor w\right\rfloor , t\_0\right), dY.u\_m \cdot \left\lfloor w\right\rfloor - t\_0, \left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right)\right)}\right)
\end{array}
\end{array}
Initial program 67.1%
Applied rewrites67.1%
Taylor expanded in dX.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.f3260.3
Applied rewrites60.3%
rem-exp-logN/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
rem-exp-logN/A
lower-fabs.f3256.1
Applied rewrites56.1%
Applied rewrites51.7%
herbie shell --seed 2025151
(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)))))))