
(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 8 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.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor d) dY.w_m))
(t_5 (* (floor h) dX.v))
(t_6 (* dX.w (floor d)))
(t_7 (* (floor d) dX.w))
(t_8 (* dY.v (floor h))))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_5 t_5)) (* t_7 t_7))
(+ (+ (* t_2 t_2) (* t_3 t_3)) (* t_4 t_4)))))
100.0)
(log2
(sqrt
(fmax
(fma
(* t_0 dX.u)
dX.u
(fma
(* (* dX.v (floor h)) (floor h))
dX.v
(* (* t_6 dX.w) (floor d))))
(fma
(* t_0 dY.u)
dY.u
(fma
(* t_8 (floor h))
dY.v
(* (* (* dY.w_m (floor d)) dY.w_m) (floor d)))))))
(log2
(sqrt
(fmax
(fma (* t_6 (floor d)) dX.w (* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(exp (* 2.0 (log (- dY.w_m))))
(* (floor d) (floor d))
(fma (* dY.u dY.u) t_0 (* t_8 t_8)))))))))dY.w_m = fabs(dY_46_w);
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_m) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(d) * dY_46_w_m;
float t_5 = floorf(h) * dX_46_v;
float t_6 = dX_46_w * floorf(d);
float t_7 = floorf(d) * dX_46_w;
float t_8 = dY_46_v * floorf(h);
float tmp;
if (log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_5 * t_5)) + (t_7 * t_7)), (((t_2 * t_2) + (t_3 * t_3)) + (t_4 * t_4))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, fmaf(((dX_46_v * floorf(h)) * floorf(h)), dX_46_v, ((t_6 * dX_46_w) * floorf(d)))), fmaf((t_0 * dY_46_u), dY_46_u, fmaf((t_8 * floorf(h)), dY_46_v, (((dY_46_w_m * floorf(d)) * dY_46_w_m) * floorf(d)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_6 * floorf(d)), dX_46_w, (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf(expf((2.0f * logf(-dY_46_w_m))), (floorf(d) * floorf(d)), fmaf((dY_46_u * dY_46_u), t_0, (t_8 * t_8))))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(d) * dY_46_w_m) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(dX_46_w * floor(d)) t_7 = Float32(floor(d) * dX_46_w) t_8 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) + Float32(t_7 * t_7)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) + Float32(t_4 * t_4))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, fma(Float32(Float32(dX_46_v * floor(h)) * floor(h)), dX_46_v, Float32(Float32(t_6 * dX_46_w) * floor(d)))), fma(Float32(t_0 * dY_46_u), dY_46_u, fma(Float32(t_8 * floor(h)), dY_46_v, Float32(Float32(Float32(dY_46_w_m * floor(d)) * dY_46_w_m) * floor(d))))))); else tmp = log2(sqrt(fmax(fma(Float32(t_6 * floor(d)), dX_46_w, Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(exp(Float32(Float32(2.0) * log(Float32(-dY_46_w_m)))), Float32(floor(d) * floor(d)), fma(Float32(dY_46_u * dY_46_u), t_0, Float32(t_8 * t_8)))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\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\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\_m\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_7 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_8 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_5 \cdot t\_5\right) + t\_7 \cdot t\_7, \left(t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right) + t\_4 \cdot t\_4\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dX.v, \left(t\_6 \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor \right)\right), \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \mathsf{fma}\left(t\_8 \cdot \left\lfloor h\right\rfloor , dY.v, \left(\left(dY.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\_m\right) \cdot \left\lfloor d\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6 \cdot \left\lfloor d\right\rfloor , dX.w, \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(e^{2 \cdot \log \left(-dY.w\_m\right)}, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, t\_8 \cdot t\_8\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)))))) < 100Initial program 67.5%
lift-+.f32N/A
lift-+.f32N/A
associate-+l+N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites67.5%
lift-+.f32N/A
lift-+.f32N/A
associate-+l+N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites67.5%
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 67.5%
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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3259.9
Applied rewrites59.9%
lift-*.f32N/A
*-commutativeN/A
lift-log.f32N/A
log-pow-revN/A
pow2N/A
sqr-neg-revN/A
pow2N/A
log-powN/A
lower-*.f32N/A
lower-log.f32N/A
lower-neg.f3235.5
Applied rewrites35.5%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor d) dY.w_m))
(t_6 (* (floor d) dX.w)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_4 t_4)) (* t_6 t_6))
(+ (+ (* t_2 t_2) (* t_3 t_3)) (* t_5 t_5)))))
100.0)
(log2
(sqrt
(fmax
(fma
(* t_0 dX.u)
dX.u
(fma
(* (* dX.v (floor h)) (floor h))
dX.v
(* (* (* dX.w (floor d)) dX.w) (floor d))))
(fma
(* t_0 dY.u)
dY.u
(fma
(* (* dY.v (floor h)) (floor h))
dY.v
(* (* (* dY.w_m (floor d)) dY.w_m) (floor d)))))))
(log2
(sqrt
(fmax
(fma
(* (* (floor h) (floor h)) dX.v)
dX.v
(fma
(* t_1 (floor w))
dX.u
(* (* dX.w dX.w) (* (floor d) (floor d)))))
(- (* (* (* dY.u (floor w)) dY.u) (floor w)))))))))dY.w_m = fabs(dY_46_w);
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_m) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(d) * dY_46_w_m;
float t_6 = floorf(d) * dX_46_w;
float tmp;
if (log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_4 * t_4)) + (t_6 * t_6)), (((t_2 * t_2) + (t_3 * t_3)) + (t_5 * t_5))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, fmaf(((dX_46_v * floorf(h)) * floorf(h)), dX_46_v, (((dX_46_w * floorf(d)) * dX_46_w) * floorf(d)))), fmaf((t_0 * dY_46_u), dY_46_u, fmaf(((dY_46_v * floorf(h)) * floorf(h)), dY_46_v, (((dY_46_w_m * floorf(d)) * dY_46_w_m) * floorf(d)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, fmaf((t_1 * floorf(w)), dX_46_u, ((dX_46_w * dX_46_w) * (floorf(d) * floorf(d))))), -(((dY_46_u * floorf(w)) * dY_46_u) * floorf(w)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(d) * dY_46_w_m) t_6 = Float32(floor(d) * dX_46_w) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) + Float32(t_6 * t_6)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) + Float32(t_5 * t_5))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, fma(Float32(Float32(dX_46_v * floor(h)) * floor(h)), dX_46_v, Float32(Float32(Float32(dX_46_w * floor(d)) * dX_46_w) * floor(d)))), fma(Float32(t_0 * dY_46_u), dY_46_u, fma(Float32(Float32(dY_46_v * floor(h)) * floor(h)), dY_46_v, Float32(Float32(Float32(dY_46_w_m * floor(d)) * dY_46_w_m) * floor(d))))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, fma(Float32(t_1 * floor(w)), dX_46_u, Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d))))), Float32(-Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w)))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\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\lfloor w\right\rfloor \cdot dY.u\\
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\_m\\
t_6 := \left\lfloor d\right\rfloor \cdot dX.w\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_4 \cdot t\_4\right) + t\_6 \cdot t\_6, \left(t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right) + t\_5 \cdot t\_5\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dX.v, \left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor \right)\right), \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dY.v, \left(\left(dY.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\_m\right) \cdot \left\lfloor d\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, \mathsf{fma}\left(t\_1 \cdot \left\lfloor w\right\rfloor , dX.u, \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right)\right), -\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \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 67.5%
lift-+.f32N/A
lift-+.f32N/A
associate-+l+N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites67.5%
lift-+.f32N/A
lift-+.f32N/A
associate-+l+N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites67.5%
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 67.5%
Applied rewrites67.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.5
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
lower-sqrt.f3264.2
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around -inf
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-sqrt.f32N/A
lower-pow.f32N/A
lower-floor.f3246.2
Applied rewrites46.2%
Applied rewrites46.1%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor d) dY.w_m))
(t_5 (* (floor d) dX.w))
(t_6 (* (* dY.u (floor w)) dY.u)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_3 t_3)) (* t_5 t_5))
(+ (+ (* t_1 t_1) (* t_2 t_2)) (* t_4 t_4)))))
100.0)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w_m (floor d)) dY.w_m)
(floor d)
(fma t_6 (floor w) (* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(fma
(* (* (floor h) (floor h)) dX.v)
dX.v
(fma
(* t_0 (floor w))
dX.u
(* (* dX.w dX.w) (* (floor d) (floor d)))))
(- (* t_6 (floor w)))))))))dY.w_m = fabs(dY_46_w);
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_m) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(d) * dY_46_w_m;
float t_5 = floorf(d) * dX_46_w;
float t_6 = (dY_46_u * floorf(w)) * dY_46_u;
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_3 * t_3)) + (t_5 * t_5)), (((t_1 * t_1) + (t_2 * t_2)) + (t_4 * t_4))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(((dY_46_w_m * floorf(d)) * dY_46_w_m), floorf(d), fmaf(t_6, floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, fmaf((t_0 * floorf(w)), dX_46_u, ((dX_46_w * dX_46_w) * (floorf(d) * floorf(d))))), -(t_6 * floorf(w)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(d) * dY_46_w_m) t_5 = Float32(floor(d) * dX_46_w) t_6 = Float32(Float32(dY_46_u * floor(w)) * dY_46_u) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) + Float32(t_5 * t_5)), Float32(Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) + Float32(t_4 * t_4))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w_m * floor(d)) * dY_46_w_m), floor(d), fma(t_6, floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, fma(Float32(t_0 * floor(w)), dX_46_u, Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d))))), Float32(-Float32(t_6 * floor(w)))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\_m\\
t_5 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_6 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_3 \cdot t\_3\right) + t\_5 \cdot t\_5, \left(t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\_m, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_6, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, dX.v, \mathsf{fma}\left(t\_0 \cdot \left\lfloor w\right\rfloor , dX.u, \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right)\right), -t\_6 \cdot \left\lfloor w\right\rfloor \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 67.5%
Applied rewrites67.5%
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 67.5%
Applied rewrites67.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.5
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
lower-sqrt.f3264.2
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around -inf
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-sqrt.f32N/A
lower-pow.f32N/A
lower-floor.f3246.2
Applied rewrites46.2%
Applied rewrites46.1%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (* dX.u (floor w)) dX.u)))
(if (<= dX.v 15000000512.0)
(log2
(sqrt
(fmax
(fma (* (fabs (* (floor d) dX.w)) (floor d)) dX.w (* t_0 (floor w)))
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(fma
(* (* (floor d) (floor d)) dY.w_m)
dY.w_m
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma t_0 (floor w) (* (* (* dX.v (floor h)) dX.v) (floor h))))
(* (* (* dY.u (floor w)) dY.u) (- (floor w)))))))))dY.w_m = fabs(dY_46_w);
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_m) {
float t_0 = (dX_46_u * floorf(w)) * dX_46_u;
float tmp;
if (dX_46_v <= 15000000512.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((fabsf((floorf(d) * dX_46_w)) * floorf(d)), dX_46_w, (t_0 * floorf(w))), fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, fmaf(((floorf(d) * floorf(d)) * dY_46_w_m), dY_46_w_m, (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(t_0, floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (((dY_46_u * floorf(w)) * dY_46_u) * -floorf(w)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(Float32(dX_46_u * floor(w)) * dX_46_u) tmp = Float32(0.0) if (dX_46_v <= Float32(15000000512.0)) tmp = log2(sqrt(fmax(fma(Float32(abs(Float32(floor(d) * dX_46_w)) * floor(d)), dX_46_w, Float32(t_0 * floor(w))), fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, fma(Float32(Float32(floor(d) * floor(d)) * dY_46_w_m), dY_46_w_m, Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(t_0, floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * Float32(-floor(w)))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\\
\mathbf{if}\;dX.v \leq 15000000512:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left|\left\lfloor d\right\rfloor \cdot dX.w\right| \cdot \left\lfloor d\right\rfloor , dX.w, t\_0 \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\_m, dY.w\_m, \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_0, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left(-\left\lfloor w\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dX.v < 15000000500Initial program 67.5%
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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3256.4
Applied rewrites56.4%
Applied rewrites56.4%
if 15000000500 < dX.v Initial program 67.5%
Applied rewrites67.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.5
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
lower-sqrt.f3264.2
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around -inf
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-sqrt.f32N/A
lower-pow.f32N/A
lower-floor.f3246.2
Applied rewrites46.2%
lift-*.f32N/A
mul-1-negN/A
lift-*.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-sqrt.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
pow1/2N/A
lift-*.f32N/A
unpow-prod-downN/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
sqr-powN/A
metadata-evalN/A
unpow1N/A
Applied rewrites46.2%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* dX.w (floor d)))
(t_1 (* dY.v (floor h)))
(t_2 (* (* dX.u (floor w)) dX.u)))
(if (<= dX.v 15000000512.0)
(log2
(sqrt
(fmax
(fma (* t_0 (floor d)) dX.w (* t_2 (floor w)))
(fma
(* dY.w_m dY.w_m)
(* (floor d) (floor d))
(fma (* dY.u dY.u) (* (floor w) (floor w)) (* t_1 t_1))))))
(log2
(sqrt
(fmax
(fma
(* t_0 dX.w)
(floor d)
(fma t_2 (floor w) (* (* (* dX.v (floor h)) dX.v) (floor h))))
(* (* (* dY.u (floor w)) dY.u) (- (floor w)))))))))dY.w_m = fabs(dY_46_w);
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_m) {
float t_0 = dX_46_w * floorf(d);
float t_1 = dY_46_v * floorf(h);
float t_2 = (dX_46_u * floorf(w)) * dX_46_u;
float tmp;
if (dX_46_v <= 15000000512.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * floorf(d)), dX_46_w, (t_2 * floorf(w))), fmaf((dY_46_w_m * dY_46_w_m), (floorf(d) * floorf(d)), fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), (t_1 * t_1))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_w), floorf(d), fmaf(t_2, floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (((dY_46_u * floorf(w)) * dY_46_u) * -floorf(w)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(dX_46_w * floor(d)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(Float32(dX_46_u * floor(w)) * dX_46_u) tmp = Float32(0.0) if (dX_46_v <= Float32(15000000512.0)) tmp = log2(sqrt(fmax(fma(Float32(t_0 * floor(d)), dX_46_w, Float32(t_2 * floor(w))), fma(Float32(dY_46_w_m * dY_46_w_m), Float32(floor(d) * floor(d)), fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), Float32(t_1 * t_1)))))); else tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_w), floor(d), fma(t_2, floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * Float32(-floor(w)))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\\
\mathbf{if}\;dX.v \leq 15000000512:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot \left\lfloor d\right\rfloor , dX.w, t\_2 \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_1 \cdot t\_1\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_2, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left(-\left\lfloor w\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dX.v < 15000000500Initial program 67.5%
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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
if 15000000500 < dX.v Initial program 67.5%
Applied rewrites67.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.5
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
lower-sqrt.f3264.2
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around -inf
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-sqrt.f32N/A
lower-pow.f32N/A
lower-floor.f3246.2
Applied rewrites46.2%
lift-*.f32N/A
mul-1-negN/A
lift-*.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-sqrt.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
pow1/2N/A
lift-*.f32N/A
unpow-prod-downN/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
sqr-powN/A
metadata-evalN/A
unpow1N/A
Applied rewrites46.2%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* dX.w (floor d))) (t_1 (* dY.v (floor h))))
(if (<= dY.w_m 60.0)
(log2
(sqrt
(fmax
(fma
(* t_0 dX.w)
(floor d)
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(* dY.u (* (fabs (* dY.u (floor w))) (floor w))))))
(log2
(sqrt
(fmax
(* dX.w (* (fabs t_0) (floor d)))
(fma
(* dY.w_m dY.w_m)
(* (floor d) (floor d))
(fma (* dY.u dY.u) (* (floor w) (floor w)) (* t_1 t_1)))))))))dY.w_m = fabs(dY_46_w);
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_m) {
float t_0 = dX_46_w * floorf(d);
float t_1 = dY_46_v * floorf(h);
float tmp;
if (dY_46_w_m <= 60.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_w), floorf(d), fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (dY_46_u * (fabsf((dY_46_u * floorf(w))) * floorf(w))))));
} else {
tmp = log2f(sqrtf(fmaxf((dX_46_w * (fabsf(t_0) * floorf(d))), fmaf((dY_46_w_m * dY_46_w_m), (floorf(d) * floorf(d)), fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), (t_1 * t_1))))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(dX_46_w * floor(d)) t_1 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (dY_46_w_m <= Float32(60.0)) tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(dY_46_u * Float32(abs(Float32(dY_46_u * floor(w))) * floor(w)))))); else tmp = log2(sqrt(fmax(Float32(dX_46_w * Float32(abs(t_0) * floor(d))), fma(Float32(dY_46_w_m * dY_46_w_m), Float32(floor(d) * floor(d)), fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), Float32(t_1 * t_1)))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dY.w\_m \leq 60:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \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), dY.u \cdot \left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(dX.w \cdot \left(\left|t\_0\right| \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_1 \cdot t\_1\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.w < 60Initial program 67.5%
Applied rewrites67.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.5
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
lower-sqrt.f3264.2
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3264.2
Applied rewrites64.2%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-fma.f32N/A
+-commutativeN/A
associate-+l+N/A
Applied rewrites64.2%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.8
Applied rewrites49.8%
if 60 < dY.w Initial program 67.5%
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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3256.4
Applied rewrites56.4%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.8
Applied rewrites49.8%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* dY.v (floor h))) (t_1 (* dX.w (floor d))))
(if (<= dX.u 15000000.0)
(log2
(sqrt
(fmax
(* dX.w (* (fabs t_1) (floor d)))
(fma
(* dY.w_m dY.w_m)
(* (floor d) (floor d))
(fma (* dY.u dY.u) (* (floor w) (floor w)) (* t_0 t_0))))))
(log2
(sqrt
(fmax
(fma
(* t_1 dX.w)
(floor d)
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(* (* (* dY.u (floor w)) dY.u) (- (floor w)))))))))dY.w_m = fabs(dY_46_w);
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_m) {
float t_0 = dY_46_v * floorf(h);
float t_1 = dX_46_w * floorf(d);
float tmp;
if (dX_46_u <= 15000000.0f) {
tmp = log2f(sqrtf(fmaxf((dX_46_w * (fabsf(t_1) * floorf(d))), fmaf((dY_46_w_m * dY_46_w_m), (floorf(d) * floorf(d)), fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), (t_0 * t_0))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_1 * dX_46_w), floorf(d), fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (((dY_46_u * floorf(w)) * dY_46_u) * -floorf(w)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(dY_46_v * floor(h)) t_1 = Float32(dX_46_w * floor(d)) tmp = Float32(0.0) if (dX_46_u <= Float32(15000000.0)) tmp = log2(sqrt(fmax(Float32(dX_46_w * Float32(abs(t_1) * floor(d))), fma(Float32(dY_46_w_m * dY_46_w_m), Float32(floor(d) * floor(d)), fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), Float32(t_0 * t_0)))))); else tmp = log2(sqrt(fmax(fma(Float32(t_1 * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * Float32(-floor(w)))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dX.w \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dX.u \leq 15000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(dX.w \cdot \left(\left|t\_1\right| \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_0 \cdot t\_0\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left(-\left\lfloor w\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 1.5e7Initial program 67.5%
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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3256.4
Applied rewrites56.4%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.8
Applied rewrites49.8%
if 1.5e7 < dX.u Initial program 67.5%
Applied rewrites67.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.5
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
lower-sqrt.f3264.2
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around -inf
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-sqrt.f32N/A
lower-pow.f32N/A
lower-floor.f3246.2
Applied rewrites46.2%
lift-*.f32N/A
mul-1-negN/A
lift-*.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-sqrt.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
pow1/2N/A
lift-*.f32N/A
unpow-prod-downN/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
sqr-powN/A
metadata-evalN/A
unpow1N/A
Applied rewrites46.2%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(* (* (* dY.u (floor w)) dY.u) (- (floor w)))))))dY.w_m = fabs(dY_46_w);
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_m) {
return log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (((dY_46_u * floorf(w)) * dY_46_u) * -floorf(w)))));
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) return log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * Float32(-floor(w)))))) end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left(-\left\lfloor w\right\rfloor \right)\right)}\right)
\end{array}
Initial program 67.5%
Applied rewrites67.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.5
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
lower-sqrt.f3264.2
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around -inf
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-sqrt.f32N/A
lower-pow.f32N/A
lower-floor.f3246.2
Applied rewrites46.2%
lift-*.f32N/A
mul-1-negN/A
lift-*.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-sqrt.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
pow1/2N/A
lift-*.f32N/A
unpow-prod-downN/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
sqr-powN/A
metadata-evalN/A
unpow1N/A
Applied rewrites46.2%
herbie shell --seed 2025140
(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)))))))