
(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 16 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}
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(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 (* (floor d) dX.w))
(t_6 (* t_5 t_5))
(t_7 (+ (+ (* t_1 t_1) (* t_2 t_2)) (* t_4 t_4))))
(if (<= (log2 (sqrt (fmax (+ (+ (* t_0 t_0) (* t_3 t_3)) t_6) t_7))) 100.0)
(log2
(sqrt
(fmax
(+
(fma
(* (* (floor w) (floor w)) dX.u)
dX.u
(* (* (floor h) (floor h)) (* dX.v dX.v)))
t_6)
t_7)))
(log2
(sqrt
(fmax
(* (* (floor d) (floor d)) (* dX.w dX.w))
(exp (* (log (* dY.u_m (floor w))) 2.0))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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;
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 = floorf(d) * dX_46_w;
float t_6 = t_5 * t_5;
float t_7 = ((t_1 * t_1) + (t_2 * t_2)) + (t_4 * t_4);
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_3 * t_3)) + t_6), t_7))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf((fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))) + t_6), t_7)));
} else {
tmp = log2f(sqrtf(fmaxf(((floorf(d) * floorf(d)) * (dX_46_w * dX_46_w)), expf((logf((dY_46_u_m * floorf(w))) * 2.0f)))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u) 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(floor(d) * dX_46_w) t_6 = Float32(t_5 * t_5) 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(Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) + t_6), t_7))) <= Float32(100.0)) tmp = log2(sqrt(fmax(Float32(fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))) + t_6), t_7))); else tmp = log2(sqrt(fmax(Float32(Float32(floor(d) * floor(d)) * Float32(dX_46_w * dX_46_w)), exp(Float32(log(Float32(dY_46_u_m * floor(w))) * Float32(2.0)))))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
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 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_6 := t\_5 \cdot t\_5\\
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(\left(t\_0 \cdot t\_0 + t\_3 \cdot t\_3\right) + t\_6, t\_7\right)}\right) \leq 100:\\
\;\;\;\;\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, dX.u, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right) + t\_6, t\_7\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right) \cdot \left(dX.w \cdot dX.w\right), e^{\log \left(dY.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot 2}\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 100.0%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites100.0%
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 6.5%
Taylor expanded in dX.w around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3213.2
Applied rewrites13.2%
Taylor expanded in dY.u around inf
pow-prod-downN/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
unpow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
unpow2N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f3216.4
Applied rewrites16.4%
lift-*.f32N/A
pow2N/A
lower-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
sqr-neg-revN/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
Applied rewrites16.4%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dY.u_m))
(t_6 (* (floor d) (floor d)))
(t_7 (* (floor h) dY.v))
(t_8 (* (floor w) (floor w))))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_5 t_5) (* t_7 t_7)) (* t_3 t_3)))))
100.0)
(log2
(sqrt
(fmax
(fma (* t_0 dX.v) dX.v (fma t_6 (* dX.w dX.w) (* t_8 (* dX.u dX.u))))
(fma
(* t_8 dY.u_m)
dY.u_m
(fma (* dY.w dY.w) t_6 (* (* dY.v dY.v) t_0))))))
(log2
(sqrt
(fmax
(* t_6 (* dX.w dX.w))
(exp (* (log (* dY.u_m (floor w))) 2.0))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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) * floorf(h);
float t_1 = floorf(w) * dX_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 = floorf(w) * dY_46_u_m;
float t_6 = floorf(d) * floorf(d);
float t_7 = floorf(h) * dY_46_v;
float t_8 = floorf(w) * floorf(w);
float tmp;
if (log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_2 * t_2)) + (t_4 * t_4)), (((t_5 * t_5) + (t_7 * t_7)) + (t_3 * t_3))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_v), dX_46_v, fmaf(t_6, (dX_46_w * dX_46_w), (t_8 * (dX_46_u * dX_46_u)))), fmaf((t_8 * dY_46_u_m), dY_46_u_m, fmaf((dY_46_w * dY_46_w), t_6, ((dY_46_v * dY_46_v) * t_0))))));
} else {
tmp = log2f(sqrtf(fmaxf((t_6 * (dX_46_w * dX_46_w)), expf((logf((dY_46_u_m * floorf(w))) * 2.0f)))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * dX_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(floor(w) * dY_46_u_m) t_6 = Float32(floor(d) * floor(d)) t_7 = Float32(floor(h) * dY_46_v) t_8 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_5 * t_5) + Float32(t_7 * t_7)) + Float32(t_3 * t_3))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_v), dX_46_v, fma(t_6, Float32(dX_46_w * dX_46_w), Float32(t_8 * Float32(dX_46_u * dX_46_u)))), fma(Float32(t_8 * dY_46_u_m), dY_46_u_m, fma(Float32(dY_46_w * dY_46_w), t_6, Float32(Float32(dY_46_v * dY_46_v) * t_0)))))); else tmp = log2(sqrt(fmax(Float32(t_6 * Float32(dX_46_w * dX_46_w)), exp(Float32(log(Float32(dY_46_u_m * floor(w))) * Float32(2.0)))))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.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 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_6 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_8 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_5 \cdot t\_5 + 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(t\_0 \cdot dX.v, dX.v, \mathsf{fma}\left(t\_6, dX.w \cdot dX.w, t\_8 \cdot \left(dX.u \cdot dX.u\right)\right)\right), \mathsf{fma}\left(t\_8 \cdot dY.u\_m, dY.u\_m, \mathsf{fma}\left(dY.w \cdot dY.w, t\_6, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_6 \cdot \left(dX.w \cdot dX.w\right), e^{\log \left(dY.u\_m \cdot \left\lfloor w\right\rfloor \right) \cdot 2}\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 100.0%
Applied rewrites100.0%
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 6.5%
Taylor expanded in dX.w around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3213.2
Applied rewrites13.2%
Taylor expanded in dY.u around inf
pow-prod-downN/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
unpow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
unpow2N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f3216.4
Applied rewrites16.4%
lift-*.f32N/A
pow2N/A
lower-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
sqr-neg-revN/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
Applied rewrites16.4%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor d) dY.w))
(t_3 (* (floor d) dX.w))
(t_4 (* (floor w) dX.u))
(t_5 (* (floor w) dY.u_m))
(t_6 (* (floor h) dX.v))
(t_7 (* (floor w) (floor w))))
(if (<= dY.v 100000.0)
(log2
(sqrt
(fmax
(+ (+ (* t_4 t_4) (* t_6 t_6)) (* t_3 t_3))
(fma (* dY.w dY.w) t_0 (* (* dY.u_m dY.u_m) t_7)))))
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* t_7 (* dX.u dX.u)))
(+ (+ (* t_5 t_5) (* t_1 t_1)) (* t_2 t_2))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * floorf(d);
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(d) * dY_46_w;
float t_3 = floorf(d) * dX_46_w;
float t_4 = floorf(w) * dX_46_u;
float t_5 = floorf(w) * dY_46_u_m;
float t_6 = floorf(h) * dX_46_v;
float t_7 = floorf(w) * floorf(w);
float tmp;
if (dY_46_v <= 100000.0f) {
tmp = log2f(sqrtf(fmaxf((((t_4 * t_4) + (t_6 * t_6)) + (t_3 * t_3)), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_u_m * dY_46_u_m) * t_7)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_7 * (dX_46_u * dX_46_u))), (((t_5 * t_5) + (t_1 * t_1)) + (t_2 * t_2)))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(d) * dY_46_w) t_3 = Float32(floor(d) * dX_46_w) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(floor(w) * dY_46_u_m) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (dY_46_v <= Float32(100000.0)) tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_4 * t_4) + Float32(t_6 * t_6)) + Float32(t_3 * t_3)), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_u_m * dY_46_u_m) * t_7))))); else tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_7 * Float32(dX_46_u * dX_46_u))), Float32(Float32(Float32(t_5 * t_5) + Float32(t_1 * t_1)) + Float32(t_2 * t_2))))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_3 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;dY.v \leq 100000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_4 \cdot t\_4 + t\_6 \cdot t\_6\right) + t\_3 \cdot t\_3, \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.u\_m \cdot dY.u\_m\right) \cdot t\_7\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_7 \cdot \left(dX.u \cdot dX.u\right)\right), \left(t\_5 \cdot t\_5 + t\_1 \cdot t\_1\right) + t\_2 \cdot t\_2\right)}\right)\\
\end{array}
\end{array}
if dY.v < 1e5Initial program 70.1%
Taylor expanded in dY.v around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3265.1
Applied rewrites65.1%
if 1e5 < dY.v Initial program 61.7%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3259.0
Applied rewrites59.0%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor d) dX.w))
(t_2 (* (floor d) dY.w))
(t_3 (* (floor d) (floor d)))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor w) dY.u_m))
(t_6 (* (floor w) dX.u)))
(if (<= dX.v 28000.0)
(log2
(sqrt
(fmax
(fma t_3 (* dX.w dX.w) (* (* (floor w) (floor w)) (* dX.u dX.u)))
(+ (+ (* t_5 t_5) (* t_0 t_0)) (* t_2 t_2)))))
(log2
(sqrt
(fmax
(+ (+ (* t_6 t_6) (* t_4 t_4)) (* t_1 t_1))
(* (* dY.w dY.w) t_3)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * dX_46_w;
float t_2 = floorf(d) * dY_46_w;
float t_3 = floorf(d) * floorf(d);
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(w) * dY_46_u_m;
float t_6 = floorf(w) * dX_46_u;
float tmp;
if (dX_46_v <= 28000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_3, (dX_46_w * dX_46_w), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u))), (((t_5 * t_5) + (t_0 * t_0)) + (t_2 * t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf((((t_6 * t_6) + (t_4 * t_4)) + (t_1 * t_1)), ((dY_46_w * dY_46_w) * t_3))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, 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(d) * dX_46_w) t_2 = Float32(floor(d) * dY_46_w) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(w) * dY_46_u_m) t_6 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (dX_46_v <= Float32(28000.0)) tmp = log2(sqrt(fmax(fma(t_3, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))), Float32(Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)) + Float32(t_2 * t_2))))); else tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_6 * t_6) + Float32(t_4 * t_4)) + Float32(t_1 * t_1)), Float32(Float32(dY_46_w * dY_46_w) * t_3)))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_2 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_6 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;dX.v \leq 28000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.w \cdot dX.w, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right), \left(t\_5 \cdot t\_5 + t\_0 \cdot t\_0\right) + t\_2 \cdot t\_2\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_6 \cdot t\_6 + t\_4 \cdot t\_4\right) + t\_1 \cdot t\_1, \left(dY.w \cdot dY.w\right) \cdot t\_3\right)}\right)\\
\end{array}
\end{array}
if dX.v < 28000Initial program 70.6%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3265.7
Applied rewrites65.7%
if 28000 < dX.v Initial program 60.2%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3254.2
Applied rewrites54.2%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dY.w))
(t_1 (* (* (floor h) (floor h)) (* dX.v dX.v)))
(t_2 (* (floor w) dY.u_m))
(t_3 (* (floor d) dX.w))
(t_4 (* (floor h) dY.v)))
(if (<= dY.v 6.599999904632568)
(log2
(sqrt
(fmax
(+ (fma (* (* (floor w) (floor w)) dX.u) dX.u t_1) (* t_3 t_3))
(* (* dY.w dY.w) (* (floor d) (floor d))))))
(log2 (sqrt (fmax t_1 (+ (+ (* t_2 t_2) (* t_4 t_4)) (* t_0 t_0))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * dY_46_w;
float t_1 = (floorf(h) * floorf(h)) * (dX_46_v * dX_46_v);
float t_2 = floorf(w) * dY_46_u_m;
float t_3 = floorf(d) * dX_46_w;
float t_4 = floorf(h) * dY_46_v;
float tmp;
if (dY_46_v <= 6.599999904632568f) {
tmp = log2f(sqrtf(fmaxf((fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, t_1) + (t_3 * t_3)), ((dY_46_w * dY_46_w) * (floorf(d) * floorf(d))))));
} else {
tmp = log2f(sqrtf(fmaxf(t_1, (((t_2 * t_2) + (t_4 * t_4)) + (t_0 * t_0)))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * dY_46_w) t_1 = Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v)) t_2 = Float32(floor(w) * dY_46_u_m) t_3 = Float32(floor(d) * dX_46_w) t_4 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (dY_46_v <= Float32(6.599999904632568)) tmp = log2(sqrt(fmax(Float32(fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, t_1) + Float32(t_3 * t_3)), Float32(Float32(dY_46_w * dY_46_w) * Float32(floor(d) * floor(d)))))); else tmp = log2(sqrt(fmax(t_1, Float32(Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)) + Float32(t_0 * t_0))))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_1 := \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_3 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;dY.v \leq 6.599999904632568:\\
\;\;\;\;\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, dX.u, t\_1\right) + t\_3 \cdot t\_3, \left(dY.w \cdot dY.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_1, \left(t\_2 \cdot t\_2 + t\_4 \cdot t\_4\right) + t\_0 \cdot t\_0\right)}\right)\\
\end{array}
\end{array}
if dY.v < 6.5999999Initial program 69.9%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites69.9%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3257.5
Applied rewrites57.5%
if 6.5999999 < dY.v Initial program 64.0%
Taylor expanded in dX.v around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3255.8
Applied rewrites55.8%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor d) dY.w))
(t_2 (* (floor d) (floor d)))
(t_3 (* (floor d) dX.w))
(t_4 (* (floor w) dY.u_m)))
(if (<= dY.v 100.0)
(log2
(sqrt
(fmax
(+
(fma
(* (* (floor w) (floor w)) dX.u)
dX.u
(* (* (floor h) (floor h)) (* dX.v dX.v)))
(* t_3 t_3))
(* (* dY.w dY.w) t_2))))
(log2
(sqrt
(fmax
(* t_2 (* dX.w dX.w))
(+ (+ (* t_4 t_4) (* t_0 t_0)) (* t_1 t_1))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * dY_46_w;
float t_2 = floorf(d) * floorf(d);
float t_3 = floorf(d) * dX_46_w;
float t_4 = floorf(w) * dY_46_u_m;
float tmp;
if (dY_46_v <= 100.0f) {
tmp = log2f(sqrtf(fmaxf((fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))) + (t_3 * t_3)), ((dY_46_w * dY_46_w) * t_2))));
} else {
tmp = log2f(sqrtf(fmaxf((t_2 * (dX_46_w * dX_46_w)), (((t_4 * t_4) + (t_0 * t_0)) + (t_1 * t_1)))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, 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(d) * dY_46_w) t_2 = Float32(floor(d) * floor(d)) t_3 = Float32(floor(d) * dX_46_w) t_4 = Float32(floor(w) * dY_46_u_m) tmp = Float32(0.0) if (dY_46_v <= Float32(100.0)) tmp = log2(sqrt(fmax(Float32(fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))) + Float32(t_3 * t_3)), Float32(Float32(dY_46_w * dY_46_w) * t_2)))); else tmp = log2(sqrt(fmax(Float32(t_2 * Float32(dX_46_w * dX_46_w)), Float32(Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) + Float32(t_1 * t_1))))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_2 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_3 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
\mathbf{if}\;dY.v \leq 100:\\
\;\;\;\;\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, dX.u, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right) + t\_3 \cdot t\_3, \left(dY.w \cdot dY.w\right) \cdot t\_2\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_2 \cdot \left(dX.w \cdot dX.w\right), \left(t\_4 \cdot t\_4 + t\_0 \cdot t\_0\right) + t\_1 \cdot t\_1\right)}\right)\\
\end{array}
\end{array}
if dY.v < 100Initial program 70.0%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites70.0%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3257.6
Applied rewrites57.6%
if 100 < dY.v Initial program 63.5%
Taylor expanded in dX.w around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3257.1
Applied rewrites57.1%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (* dY.w dY.w) (* (floor d) (floor d))))
(t_1 (* (floor w) (floor w)))
(t_2 (* (floor d) dX.w))
(t_3 (* (floor h) dY.v)))
(if (<= dY.v 100.0)
(log2
(sqrt
(fmax
(+
(fma (* t_1 dX.u) dX.u (* (* (floor h) (floor h)) (* dX.v dX.v)))
(* t_2 t_2))
t_0)))
(log2
(sqrt
(fmax
(* t_1 (* dX.u dX.u))
(fma (* t_1 dY.u_m) dY.u_m (fma t_3 t_3 t_0))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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 * dY_46_w) * (floorf(d) * floorf(d));
float t_1 = floorf(w) * floorf(w);
float t_2 = floorf(d) * dX_46_w;
float t_3 = floorf(h) * dY_46_v;
float tmp;
if (dY_46_v <= 100.0f) {
tmp = log2f(sqrtf(fmaxf((fmaf((t_1 * dX_46_u), dX_46_u, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))) + (t_2 * t_2)), t_0)));
} else {
tmp = log2f(sqrtf(fmaxf((t_1 * (dX_46_u * dX_46_u)), fmaf((t_1 * dY_46_u_m), dY_46_u_m, fmaf(t_3, t_3, t_0)))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(Float32(dY_46_w * dY_46_w) * Float32(floor(d) * floor(d))) t_1 = Float32(floor(w) * floor(w)) t_2 = Float32(floor(d) * dX_46_w) t_3 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (dY_46_v <= Float32(100.0)) tmp = log2(sqrt(fmax(Float32(fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))) + Float32(t_2 * t_2)), t_0))); else tmp = log2(sqrt(fmax(Float32(t_1 * Float32(dX_46_u * dX_46_u)), fma(Float32(t_1 * dY_46_u_m), dY_46_u_m, fma(t_3, t_3, t_0))))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left(dY.w \cdot dY.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;dY.v \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right) + t\_2 \cdot t\_2, t\_0\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_1 \cdot \left(dX.u \cdot dX.u\right), \mathsf{fma}\left(t\_1 \cdot dY.u\_m, dY.u\_m, \mathsf{fma}\left(t\_3, t\_3, t\_0\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.v < 100Initial program 70.0%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites70.0%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3257.6
Applied rewrites57.6%
if 100 < dY.v Initial program 63.5%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3255.9
Applied rewrites55.9%
Applied rewrites55.9%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dX.w))
(t_4 (* (floor w) (floor w))))
(if (<= dY.w 40000.0)
(log2 (sqrt (fmax (fma t_1 t_1 (fma t_3 t_3 (* t_2 t_2))) (* t_0 t_0))))
(log2
(sqrt
(fmax
(* t_4 (* dX.u dX.u))
(fma
(* t_4 dY.u_m)
dY.u_m
(fma t_0 t_0 (* (* dY.w dY.w) (* (floor d) (floor d)))))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(w) * dX_46_u;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dX_46_w;
float t_4 = floorf(w) * floorf(w);
float tmp;
if (dY_46_w <= 40000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, t_1, fmaf(t_3, t_3, (t_2 * t_2))), (t_0 * t_0))));
} else {
tmp = log2f(sqrtf(fmaxf((t_4 * (dX_46_u * dX_46_u)), fmaf((t_4 * dY_46_u_m), dY_46_u_m, fmaf(t_0, t_0, ((dY_46_w * dY_46_w) * (floorf(d) * floorf(d))))))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, 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(w) * dX_46_u) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dX_46_w) t_4 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (dY_46_w <= Float32(40000.0)) tmp = log2(sqrt(fmax(fma(t_1, t_1, fma(t_3, t_3, Float32(t_2 * t_2))), Float32(t_0 * t_0)))); else tmp = log2(sqrt(fmax(Float32(t_4 * Float32(dX_46_u * dX_46_u)), fma(Float32(t_4 * dY_46_u_m), dY_46_u_m, fma(t_0, t_0, Float32(Float32(dY_46_w * dY_46_w) * Float32(floor(d) * floor(d)))))))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_4 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;dY.w \leq 40000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, \mathsf{fma}\left(t\_3, t\_3, t\_2 \cdot t\_2\right)\right), t\_0 \cdot t\_0\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_4 \cdot \left(dX.u \cdot dX.u\right), \mathsf{fma}\left(t\_4 \cdot dY.u\_m, dY.u\_m, \mathsf{fma}\left(t\_0, t\_0, \left(dY.w \cdot dY.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.w < 4e4Initial program 70.0%
Taylor expanded in dY.v around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3257.5
Applied rewrites57.5%
Applied rewrites57.5%
if 4e4 < dY.w Initial program 62.4%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3255.0
Applied rewrites55.0%
Applied rewrites55.0%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* (* dY.w dY.w) (* (floor d) (floor d))))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor h) dX.v)))
(if (<= dX.v 28000.0)
(log2
(sqrt
(fmax
(* t_0 (* dX.u dX.u))
(fma (* t_0 dY.u_m) dY.u_m (fma t_3 t_3 t_2)))))
(log2 (sqrt (fmax (fma t_1 t_1 (* t_4 t_4)) t_2))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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) * floorf(w);
float t_1 = floorf(w) * dX_46_u;
float t_2 = (dY_46_w * dY_46_w) * (floorf(d) * floorf(d));
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(h) * dX_46_v;
float tmp;
if (dX_46_v <= 28000.0f) {
tmp = log2f(sqrtf(fmaxf((t_0 * (dX_46_u * dX_46_u)), fmaf((t_0 * dY_46_u_m), dY_46_u_m, fmaf(t_3, t_3, t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, t_1, (t_4 * t_4)), t_2)));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(Float32(dY_46_w * dY_46_w) * Float32(floor(d) * floor(d))) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if (dX_46_v <= Float32(28000.0)) tmp = log2(sqrt(fmax(Float32(t_0 * Float32(dX_46_u * dX_46_u)), fma(Float32(t_0 * dY_46_u_m), dY_46_u_m, fma(t_3, t_3, t_2))))); else tmp = log2(sqrt(fmax(fma(t_1, t_1, Float32(t_4 * t_4)), t_2))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left(dY.w \cdot dY.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;dX.v \leq 28000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_0 \cdot \left(dX.u \cdot dX.u\right), \mathsf{fma}\left(t\_0 \cdot dY.u\_m, dY.u\_m, \mathsf{fma}\left(t\_3, t\_3, t\_2\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, t\_4 \cdot t\_4\right), t\_2\right)}\right)\\
\end{array}
\end{array}
if dX.v < 28000Initial program 70.6%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3257.6
Applied rewrites57.6%
Applied rewrites57.6%
if 28000 < dX.v Initial program 60.2%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites60.2%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3254.2
Applied rewrites54.2%
Taylor expanded in dX.w around 0
Applied rewrites49.8%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor d) dX.w))
(t_3 (* (floor d) (floor d))))
(if (<= dY.v 100000.0)
(log2 (sqrt (fmax (fma t_0 t_0 (* t_2 t_2)) (* (* dY.w dY.w) t_3))))
(log2
(sqrt
(fmax
(* t_3 (* dX.w dX.w))
(fma t_1 t_1 (* (* dY.u_m dY.u_m) (* (floor w) (floor w))))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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) * dX_46_v;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(d) * dX_46_w;
float t_3 = floorf(d) * floorf(d);
float tmp;
if (dY_46_v <= 100000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, t_0, (t_2 * t_2)), ((dY_46_w * dY_46_w) * t_3))));
} else {
tmp = log2f(sqrtf(fmaxf((t_3 * (dX_46_w * dX_46_w)), fmaf(t_1, t_1, ((dY_46_u_m * dY_46_u_m) * (floorf(w) * floorf(w)))))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(d) * dX_46_w) t_3 = Float32(floor(d) * floor(d)) tmp = Float32(0.0) if (dY_46_v <= Float32(100000.0)) tmp = log2(sqrt(fmax(fma(t_0, t_0, Float32(t_2 * t_2)), Float32(Float32(dY_46_w * dY_46_w) * t_3)))); else tmp = log2(sqrt(fmax(Float32(t_3 * Float32(dX_46_w * dX_46_w)), fma(t_1, t_1, Float32(Float32(dY_46_u_m * dY_46_u_m) * Float32(floor(w) * floor(w))))))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dY.v \leq 100000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_2 \cdot t\_2\right), \left(dY.w \cdot dY.w\right) \cdot t\_3\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_3 \cdot \left(dX.w \cdot dX.w\right), \mathsf{fma}\left(t\_1, t\_1, \left(dY.u\_m \cdot dY.u\_m\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.v < 1e5Initial program 70.1%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites70.1%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3257.4
Applied rewrites57.4%
Taylor expanded in dX.u around 0
Applied rewrites47.9%
if 1e5 < dY.v Initial program 61.7%
Taylor expanded in dX.w around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3256.7
Applied rewrites56.7%
Taylor expanded in dY.w around 0
Applied rewrites52.6%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor d) dX.w))
(t_3 (* (* dY.w dY.w) (* (floor d) (floor d)))))
(if (<= dX.u 30000.0)
(log2 (sqrt (fmax (fma t_0 t_0 (* t_2 t_2)) t_3)))
(log2 (sqrt (fmax (fma t_1 t_1 (* t_0 t_0)) t_3))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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) * dX_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(d) * dX_46_w;
float t_3 = (dY_46_w * dY_46_w) * (floorf(d) * floorf(d));
float tmp;
if (dX_46_u <= 30000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, t_0, (t_2 * t_2)), t_3)));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, t_1, (t_0 * t_0)), t_3)));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(d) * dX_46_w) t_3 = Float32(Float32(dY_46_w * dY_46_w) * Float32(floor(d) * floor(d))) tmp = Float32(0.0) if (dX_46_u <= Float32(30000.0)) tmp = log2(sqrt(fmax(fma(t_0, t_0, Float32(t_2 * t_2)), t_3))); else tmp = log2(sqrt(fmax(fma(t_1, t_1, Float32(t_0 * t_0)), t_3))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_3 := \left(dY.w \cdot dY.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\\
\mathbf{if}\;dX.u \leq 30000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_2 \cdot t\_2\right), t\_3\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, t\_0 \cdot t\_0\right), t\_3\right)}\right)\\
\end{array}
\end{array}
if dX.u < 3e4Initial program 70.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites70.7%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3254.3
Applied rewrites54.3%
Taylor expanded in dX.u around 0
Applied rewrites48.1%
if 3e4 < dX.u Initial program 59.4%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites59.4%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3253.8
Applied rewrites53.8%
Taylor expanded in dX.w around 0
Applied rewrites49.4%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor w) (floor w)))
(t_3 (* (floor d) dX.w)))
(if (<= dX.v 2.0)
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* t_2 (* dX.u dX.u)))
(* (* dY.u_m dY.u_m) t_2))))
(log2 (sqrt (fmax (fma t_1 t_1 (* t_3 t_3)) (* (* dY.w dY.w) t_0)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * floorf(d);
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * floorf(w);
float t_3 = floorf(d) * dX_46_w;
float tmp;
if (dX_46_v <= 2.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_2 * (dX_46_u * dX_46_u))), ((dY_46_u_m * dY_46_u_m) * t_2))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, t_1, (t_3 * t_3)), ((dY_46_w * dY_46_w) * t_0))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(floor(d) * dX_46_w) tmp = Float32(0.0) if (dX_46_v <= Float32(2.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_2 * Float32(dX_46_u * dX_46_u))), Float32(Float32(dY_46_u_m * dY_46_u_m) * t_2)))); else tmp = log2(sqrt(fmax(fma(t_1, t_1, Float32(t_3 * t_3)), Float32(Float32(dY_46_w * dY_46_w) * t_0)))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor d\right\rfloor \cdot dX.w\\
\mathbf{if}\;dX.v \leq 2:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_2 \cdot \left(dX.u \cdot dX.u\right)\right), \left(dY.u\_m \cdot dY.u\_m\right) \cdot t\_2\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, t\_3 \cdot t\_3\right), \left(dY.w \cdot dY.w\right) \cdot t\_0\right)}\right)\\
\end{array}
\end{array}
if dX.v < 2Initial program 70.4%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3265.7
Applied rewrites65.7%
Taylor expanded in dY.u around inf
pow-prod-downN/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
unpow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
unpow2N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f3248.3
Applied rewrites48.3%
if 2 < dX.v Initial program 62.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites62.7%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3254.4
Applied rewrites54.4%
Taylor expanded in dX.u around 0
Applied rewrites49.5%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor w) (floor w))))
(if (<= dX.v 7500.0)
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* t_2 (* dX.u dX.u)))
(* (* dY.u_m dY.u_m) t_2))))
(log2 (sqrt (fmax (* t_1 t_1) (* (* dY.w dY.w) t_0)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * floorf(d);
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * floorf(w);
float tmp;
if (dX_46_v <= 7500.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_2 * (dX_46_u * dX_46_u))), ((dY_46_u_m * dY_46_u_m) * t_2))));
} else {
tmp = log2f(sqrtf(fmaxf((t_1 * t_1), ((dY_46_w * dY_46_w) * t_0))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (dX_46_v <= Float32(7500.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_2 * Float32(dX_46_u * dX_46_u))), Float32(Float32(dY_46_u_m * dY_46_u_m) * t_2)))); else tmp = log2(sqrt(fmax(Float32(t_1 * t_1), Float32(Float32(dY_46_w * dY_46_w) * t_0)))); end return tmp end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;dX.v \leq 7500:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_2 \cdot \left(dX.u \cdot dX.u\right)\right), \left(dY.u\_m \cdot dY.u\_m\right) \cdot t\_2\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_1 \cdot t\_1, \left(dY.w \cdot dY.w\right) \cdot t\_0\right)}\right)\\
\end{array}
\end{array}
if dX.v < 7500Initial program 70.6%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3265.8
Applied rewrites65.8%
Taylor expanded in dY.u around inf
pow-prod-downN/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
unpow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
unpow2N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f3248.4
Applied rewrites48.4%
if 7500 < dX.v Initial program 60.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites60.3%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3254.0
Applied rewrites54.0%
Taylor expanded in dX.v around inf
Applied rewrites46.2%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (* dY.w dY.w) t_0))
(t_2 (* (floor h) dX.v)))
(if (<= dX.v 0.10000000149011612)
(log2 (sqrt (fmax (* t_0 (* dX.w dX.w)) t_1)))
(log2 (sqrt (fmax (* t_2 t_2) t_1))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * floorf(d);
float t_1 = (dY_46_w * dY_46_w) * t_0;
float t_2 = floorf(h) * dX_46_v;
float tmp;
if (dX_46_v <= 0.10000000149011612f) {
tmp = log2f(sqrtf(fmaxf((t_0 * (dX_46_w * dX_46_w)), t_1)));
} else {
tmp = log2f(sqrtf(fmaxf((t_2 * t_2), t_1)));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(Float32(dY_46_w * dY_46_w) * t_0) t_2 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if (dX_46_v <= Float32(0.10000000149011612)) tmp = log2(sqrt(fmax(Float32(t_0 * Float32(dX_46_w * dX_46_w)), t_1))); else tmp = log2(sqrt(fmax(Float32(t_2 * t_2), t_1))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = floor(d) * floor(d); t_1 = (dY_46_w * dY_46_w) * t_0; t_2 = floor(h) * dX_46_v; tmp = single(0.0); if (dX_46_v <= single(0.10000000149011612)) tmp = log2(sqrt(max((t_0 * (dX_46_w * dX_46_w)), t_1))); else tmp = log2(sqrt(max((t_2 * t_2), t_1))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left(dY.w \cdot dY.w\right) \cdot t\_0\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;dX.v \leq 0.10000000149011612:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_0 \cdot \left(dX.w \cdot dX.w\right), t\_1\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_2 \cdot t\_2, t\_1\right)}\right)\\
\end{array}
\end{array}
if dX.v < 0.100000001Initial program 70.4%
Taylor expanded in dX.w around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3257.7
Applied rewrites57.7%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3237.6
Applied rewrites37.6%
if 0.100000001 < dX.v Initial program 63.1%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites63.1%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3254.6
Applied rewrites54.6%
Taylor expanded in dX.v around inf
Applied rewrites44.4%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dX.w)) (t_1 (* (floor d) (floor d))))
(if (<= dY.w 200000.0)
(log2
(sqrt (fmax (* t_0 t_0) (* (* dY.u_m dY.u_m) (* (floor w) (floor w))))))
(log2 (sqrt (fmax (* t_1 (* dX.w dX.w)) (* (* dY.w dY.w) t_1)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * dX_46_w;
float t_1 = floorf(d) * floorf(d);
float tmp;
if (dY_46_w <= 200000.0f) {
tmp = log2f(sqrtf(fmaxf((t_0 * t_0), ((dY_46_u_m * dY_46_u_m) * (floorf(w) * floorf(w))))));
} else {
tmp = log2f(sqrtf(fmaxf((t_1 * (dX_46_w * dX_46_w)), ((dY_46_w * dY_46_w) * t_1))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * dX_46_w) t_1 = Float32(floor(d) * floor(d)) tmp = Float32(0.0) if (dY_46_w <= Float32(200000.0)) tmp = log2(sqrt(fmax(Float32(t_0 * t_0), Float32(Float32(dY_46_u_m * dY_46_u_m) * Float32(floor(w) * floor(w)))))); else tmp = log2(sqrt(fmax(Float32(t_1 * Float32(dX_46_w * dX_46_w)), Float32(Float32(dY_46_w * dY_46_w) * t_1)))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = floor(d) * dX_46_w; t_1 = floor(d) * floor(d); tmp = single(0.0); if (dY_46_w <= single(200000.0)) tmp = log2(sqrt(max((t_0 * t_0), ((dY_46_u_m * dY_46_u_m) * (floor(w) * floor(w)))))); else tmp = log2(sqrt(max((t_1 * (dX_46_w * dX_46_w)), ((dY_46_w * dY_46_w) * t_1)))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dY.w \leq 200000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0, \left(dY.u\_m \cdot dY.u\_m\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_1 \cdot \left(dX.w \cdot dX.w\right), \left(dY.w \cdot dY.w\right) \cdot t\_1\right)}\right)\\
\end{array}
\end{array}
if dY.w < 2e5Initial program 70.0%
Taylor expanded in dX.w around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3254.0
Applied rewrites54.0%
Taylor expanded in dY.u around inf
pow-prod-downN/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
unpow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
unpow2N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f3237.6
Applied rewrites37.6%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3237.6
Applied rewrites37.6%
if 2e5 < dY.w Initial program 62.0%
Taylor expanded in dX.w around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3255.8
Applied rewrites55.8%
Taylor expanded in dY.w around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f3248.4
Applied rewrites48.4%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h d dX.u dX.v dX.w dY.u_m dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dX.w)))
(log2
(sqrt (fmax (* t_0 t_0) (* (* dY.u_m dY.u_m) (* (floor w) (floor w))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float d, float dX_46_u, 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(d) * dX_46_w;
return log2f(sqrtf(fmaxf((t_0 * t_0), ((dY_46_u_m * dY_46_u_m) * (floorf(w) * floorf(w))))));
}
dY.u_m = abs(dY_46_u) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * dX_46_w) return log2(sqrt(fmax(Float32(t_0 * t_0), Float32(Float32(dY_46_u_m * dY_46_u_m) * Float32(floor(w) * floor(w)))))) end
dY.u_m = abs(dY_46_u); function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u_m, dY_46_v, dY_46_w) t_0 = floor(d) * dX_46_w; tmp = log2(sqrt(max((t_0 * t_0), ((dY_46_u_m * dY_46_u_m) * (floor(w) * floor(w)))))); end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dX.w\\
\log_{2} \left(\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0, \left(dY.u\_m \cdot dY.u\_m\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)}\right)
\end{array}
\end{array}
Initial program 68.5%
Taylor expanded in dX.w around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3254.3
Applied rewrites54.3%
Taylor expanded in dY.u around inf
pow-prod-downN/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
unpow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
unpow2N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f3235.9
Applied rewrites35.9%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3235.9
Applied rewrites35.9%
herbie shell --seed 2025130
(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)))))))