
(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}
dX.w_m = (fabs.f32 dX.w)
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w_m 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 d) dY.w_m))
(t_4 (* (floor d) dX.w_m))
(t_5 (* (floor h) dX.v))
(t_6 (+ (* t_0 t_0) (* t_5 t_5)))
(t_7 (+ t_6 (* t_4 t_4)))
(t_8 (* dY.w_m (floor d))))
(if (<=
(log2 (sqrt (fmax t_7 (+ (+ (* t_1 t_1) (* t_2 t_2)) (* t_3 t_3)))))
100.0)
(log2
(sqrt
(fmax
t_7
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(fma
(* (* (floor h) (floor h)) dY.v)
dY.v
(* (* t_8 dY.w_m) (floor d)))))))
(log2
(sqrt
(fmax
(+ t_6 (exp (* (log (* (- dX.w_m) (floor d))) 2.0)))
(fma
dY.w_m
(* (fabs t_8) (floor d))
(* (pow dY.u 2.0) (pow (floor w) 2.0)))))))))dX.w_m = fabs(dX_46_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_m, 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(d) * dY_46_w_m;
float t_4 = floorf(d) * dX_46_w_m;
float t_5 = floorf(h) * dX_46_v;
float t_6 = (t_0 * t_0) + (t_5 * t_5);
float t_7 = t_6 + (t_4 * t_4);
float t_8 = dY_46_w_m * floorf(d);
float tmp;
if (log2f(sqrtf(fmaxf(t_7, (((t_1 * t_1) + (t_2 * t_2)) + (t_3 * t_3))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(t_7, fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, ((t_8 * dY_46_w_m) * floorf(d)))))));
} else {
tmp = log2f(sqrtf(fmaxf((t_6 + expf((logf((-dX_46_w_m * floorf(d))) * 2.0f))), fmaf(dY_46_w_m, (fabsf(t_8) * floorf(d)), (powf(dY_46_u, 2.0f) * powf(floorf(w), 2.0f))))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, 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(d) * dY_46_w_m) t_4 = Float32(floor(d) * dX_46_w_m) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(Float32(t_0 * t_0) + Float32(t_5 * t_5)) t_7 = Float32(t_6 + Float32(t_4 * t_4)) t_8 = Float32(dY_46_w_m * floor(d)) tmp = Float32(0.0) if (log2(sqrt(fmax(t_7, Float32(Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) + Float32(t_3 * t_3))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(t_7, fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, Float32(Float32(t_8 * dY_46_w_m) * floor(d))))))); else tmp = log2(sqrt(fmax(Float32(t_6 + exp(Float32(log(Float32(Float32(-dX_46_w_m) * floor(d))) * Float32(2.0)))), fma(dY_46_w_m, Float32(abs(t_8) * floor(d)), Float32((dY_46_u ^ Float32(2.0)) * (floor(w) ^ Float32(2.0))))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
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 d\right\rfloor \cdot dY.w\_m\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\_m\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := t\_0 \cdot t\_0 + t\_5 \cdot t\_5\\
t_7 := t\_6 + t\_4 \cdot t\_4\\
t_8 := dY.w\_m \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_7, \left(t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right) + t\_3 \cdot t\_3\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_7, \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 h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, \left(t\_8 \cdot dY.w\_m\right) \cdot \left\lfloor d\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_6 + e^{\log \left(\left(-dX.w\_m\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}, \mathsf{fma}\left(dY.w\_m, \left|t\_8\right| \cdot \left\lfloor d\right\rfloor , {dY.u}^{2} \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right)\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3268.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
Applied rewrites68.0%
Taylor expanded in dY.v around 0
lower-fma.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3260.9
Applied rewrites60.9%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-unsound-exp.f32N/A
lower-unsound-*.f32N/A
lower-unsound-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3236.0
Applied rewrites36.0%
dX.w_m = (fabs.f32 dX.w)
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w_m 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 (* dY.v (floor h)))
(t_4 (* (floor h) dY.v))
(t_5 (* (floor h) dX.v))
(t_6 (* (floor d) dY.w_m))
(t_7 (* (floor d) dX.w_m))
(t_8 (+ (+ (* t_1 t_1) (* t_5 t_5)) (* t_7 t_7))))
(if (<=
(log2 (sqrt (fmax t_8 (+ (+ (* t_2 t_2) (* t_4 t_4)) (* t_6 t_6)))))
100.0)
(log2
(sqrt
(fmax
t_8
(fma
(* t_0 dY.u)
dY.u
(fma
(* (* (floor h) (floor h)) dY.v)
dY.v
(* (* (* dY.w_m (floor d)) dY.w_m) (floor d)))))))
(log2
(sqrt
(fmax
(fma
(* (* dX.w_m (floor d)) (floor d))
dX.w_m
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* dY.u dY.u)
t_0
(fma t_3 t_3 (exp (* (log (* (- dY.w_m) (floor d))) 2.0))))))))))dX.w_m = fabs(dX_46_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_m, 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 = dY_46_v * floorf(h);
float t_4 = floorf(h) * dY_46_v;
float t_5 = floorf(h) * dX_46_v;
float t_6 = floorf(d) * dY_46_w_m;
float t_7 = floorf(d) * dX_46_w_m;
float t_8 = ((t_1 * t_1) + (t_5 * t_5)) + (t_7 * t_7);
float tmp;
if (log2f(sqrtf(fmaxf(t_8, (((t_2 * t_2) + (t_4 * t_4)) + (t_6 * t_6))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(t_8, fmaf((t_0 * dY_46_u), dY_46_u, fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, (((dY_46_w_m * floorf(d)) * dY_46_w_m) * floorf(d)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w_m * floorf(d)) * floorf(d)), dX_46_w_m, (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((dY_46_u * dY_46_u), t_0, fmaf(t_3, t_3, expf((logf((-dY_46_w_m * floorf(d))) * 2.0f)))))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, 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(dY_46_v * floor(h)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(floor(d) * dY_46_w_m) t_7 = Float32(floor(d) * dX_46_w_m) t_8 = Float32(Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) + Float32(t_7 * t_7)) tmp = Float32(0.0) if (log2(sqrt(fmax(t_8, Float32(Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)) + Float32(t_6 * t_6))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(t_8, fma(Float32(t_0 * dY_46_u), dY_46_u, fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), 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(dX_46_w_m * floor(d)) * floor(d)), dX_46_w_m, Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(dY_46_u * dY_46_u), t_0, fma(t_3, t_3, exp(Float32(log(Float32(Float32(-dY_46_w_m) * floor(d))) * Float32(2.0)))))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
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 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := \left\lfloor d\right\rfloor \cdot dY.w\_m\\
t_7 := \left\lfloor d\right\rfloor \cdot dX.w\_m\\
t_8 := \left(t\_1 \cdot t\_1 + t\_5 \cdot t\_5\right) + t\_7 \cdot t\_7\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_8, \left(t\_2 \cdot t\_2 + t\_4 \cdot t\_4\right) + t\_6 \cdot t\_6\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_8, \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, 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(dX.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dX.w\_m, \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(dY.u \cdot dY.u, t\_0, \mathsf{fma}\left(t\_3, t\_3, e^{\log \left(\left(-dY.w\_m\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}\right)\right)\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
Taylor expanded in 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.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-unsound-exp.f32N/A
lower-unsound-*.f32N/A
lower-unsound-log.f32N/A
lift-*.f32N/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3236.4
Applied rewrites36.4%
dX.w_m = (fabs.f32 dX.w)
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* dX.w_m (floor d)))
(t_1 (* (floor w) dX.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_m))
(t_6 (* dY.v (floor h)))
(t_7 (* (floor w) dY.u))
(t_8 (* (* dX.u (floor w)) dX.u)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_3 t_3)) (* t_5 t_5))
(+ (+ (* t_7 t_7) (* t_2 t_2)) (* t_4 t_4)))))
100.0)
(log2
(sqrt
(fmax
(fma
(* t_0 dX.w_m)
(floor d)
(fma t_8 (floor w) (* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w_m (floor d)) dY.w_m)
(floor d)
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* t_6 dY.v) (floor h)))))))
(log2
(sqrt
(fmax
(fma (* t_0 (floor d)) dX.w_m (* t_8 (floor w)))
(fma
(* dY.u dY.u)
(* (floor w) (floor w))
(fma t_6 t_6 (exp (* (log (* (- dY.w_m) (floor d))) 2.0))))))))))dX.w_m = fabs(dX_46_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_m, float dY_46_u, float dY_46_v, float dY_46_w_m) {
float t_0 = dX_46_w_m * floorf(d);
float t_1 = floorf(w) * dX_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_m;
float t_6 = dY_46_v * floorf(h);
float t_7 = floorf(w) * dY_46_u;
float t_8 = (dX_46_u * floorf(w)) * dX_46_u;
float tmp;
if (log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_3 * t_3)) + (t_5 * t_5)), (((t_7 * t_7) + (t_2 * t_2)) + (t_4 * t_4))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_w_m), floorf(d), fmaf(t_8, 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(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), ((t_6 * dY_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * floorf(d)), dX_46_w_m, (t_8 * floorf(w))), fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), fmaf(t_6, t_6, expf((logf((-dY_46_w_m * floorf(d))) * 2.0f)))))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(dX_46_w_m * floor(d)) t_1 = Float32(floor(w) * dX_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_m) t_6 = Float32(dY_46_v * floor(h)) t_7 = Float32(floor(w) * dY_46_u) t_8 = Float32(Float32(dX_46_u * floor(w)) * dX_46_u) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_3 * t_3)) + Float32(t_5 * t_5)), Float32(Float32(Float32(t_7 * t_7) + Float32(t_2 * t_2)) + Float32(t_4 * t_4))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_w_m), floor(d), fma(t_8, floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w_m * floor(d)) * dY_46_w_m), floor(d), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(t_6 * dY_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(fma(Float32(t_0 * floor(d)), dX_46_w_m, Float32(t_8 * floor(w))), fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), fma(t_6, t_6, exp(Float32(log(Float32(Float32(-dY_46_w_m) * floor(d))) * Float32(2.0)))))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := dX.w\_m \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.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\_m\\
t_6 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_7 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_8 := \left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_3 \cdot t\_3\right) + t\_5 \cdot t\_5, \left(t\_7 \cdot t\_7 + 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(t\_0 \cdot dX.w\_m, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\_m, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_6 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot \left\lfloor d\right\rfloor , dX.w\_m, t\_8 \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \mathsf{fma}\left(t\_6, t\_6, e^{\log \left(\left(-dY.w\_m\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}\right)\right)\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
Applied rewrites68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
Taylor expanded in 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.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-unsound-exp.f32N/A
lower-unsound-*.f32N/A
lower-unsound-log.f32N/A
lift-*.f32N/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3236.4
Applied rewrites36.4%
dX.w_m = (fabs.f32 dX.w)
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (* (floor w) (floor w)) dY.u))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor d) dX.w_m))
(t_3 (* (floor h) (floor h)))
(t_4 (* (floor h) dX.v))
(t_5 (* dY.w_m (floor d))))
(if (<= dY.v 5.0)
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_4 t_4)) (* t_2 t_2))
(fma t_0 dY.u (* (* (fabs t_5) dY.w_m) (floor d))))))
(log2
(sqrt
(fmax
(fma (* dX.v dX.v) t_3 (* (* (* dX.w_m (floor d)) dX.w_m) (floor d)))
(fma
t_0
dY.u
(fma (* t_3 dY.v) dY.v (* (* t_5 dY.w_m) (floor d))))))))))dX.w_m = fabs(dX_46_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_m, float dY_46_u, float dY_46_v, float dY_46_w_m) {
float t_0 = (floorf(w) * floorf(w)) * dY_46_u;
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(d) * dX_46_w_m;
float t_3 = floorf(h) * floorf(h);
float t_4 = floorf(h) * dX_46_v;
float t_5 = dY_46_w_m * floorf(d);
float tmp;
if (dY_46_v <= 5.0f) {
tmp = log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_4 * t_4)) + (t_2 * t_2)), fmaf(t_0, dY_46_u, ((fabsf(t_5) * dY_46_w_m) * floorf(d))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_v * dX_46_v), t_3, (((dX_46_w_m * floorf(d)) * dX_46_w_m) * floorf(d))), fmaf(t_0, dY_46_u, fmaf((t_3 * dY_46_v), dY_46_v, ((t_5 * dY_46_w_m) * floorf(d)))))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(Float32(floor(w) * floor(w)) * dY_46_u) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(d) * dX_46_w_m) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(dY_46_w_m * floor(d)) tmp = Float32(0.0) if (dY_46_v <= Float32(5.0)) tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) + Float32(t_2 * t_2)), fma(t_0, dY_46_u, Float32(Float32(abs(t_5) * dY_46_w_m) * floor(d)))))); else tmp = log2(sqrt(fmax(fma(Float32(dX_46_v * dX_46_v), t_3, Float32(Float32(Float32(dX_46_w_m * floor(d)) * dX_46_w_m) * floor(d))), fma(t_0, dY_46_u, fma(Float32(t_3 * dY_46_v), dY_46_v, Float32(Float32(t_5 * dY_46_w_m) * floor(d))))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor d\right\rfloor \cdot dX.w\_m\\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := dY.w\_m \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dY.v \leq 5:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_4 \cdot t\_4\right) + t\_2 \cdot t\_2, \mathsf{fma}\left(t\_0, dY.u, \left(\left|t\_5\right| \cdot dY.w\_m\right) \cdot \left\lfloor d\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.v \cdot dX.v, t\_3, \left(\left(dX.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w\_m\right) \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(t\_0, dY.u, \mathsf{fma}\left(t\_3 \cdot dY.v, dY.v, \left(t\_5 \cdot dY.w\_m\right) \cdot \left\lfloor d\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.v < 5Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3268.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.0
Applied rewrites68.0%
Taylor expanded in dY.v around 0
lower-fma.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3260.9
Applied rewrites60.9%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-pow.f32N/A
pow2N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lower-fma.f3260.9
Applied rewrites60.9%
if 5 < dY.v Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
Taylor expanded in dX.u around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3260.9
Applied rewrites60.9%
lift-pow.f32N/A
unpow2N/A
lower-*.f3260.9
lift-pow.f32N/A
pow2N/A
lift-*.f3260.9
lift-*.f32N/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
unswap-sqrN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3260.9
Applied rewrites60.9%
dX.w_m = (fabs.f32 dX.w)
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* dY.v (floor h)))
(t_2 (* dY.w_m (floor d)))
(t_3 (* dX.w_m (floor d)))
(t_4 (* (floor w) (floor w))))
(if (<= dX.u 0.03999999910593033)
(log2
(sqrt
(fmax
(fma (* dX.v dX.v) t_0 (* (* t_3 dX.w_m) (floor d)))
(fma
(* t_4 dY.u)
dY.u
(fma (* t_0 dY.v) dY.v (* (* t_2 dY.w_m) (floor d)))))))
(log2
(sqrt
(fmax
(fma
(* t_3 (floor d))
dX.w_m
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma (* dY.u dY.u) t_4 (fma t_1 t_1 (* t_2 t_2)))))))))dX.w_m = fabs(dX_46_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_m, float dY_46_u, float dY_46_v, float dY_46_w_m) {
float t_0 = floorf(h) * floorf(h);
float t_1 = dY_46_v * floorf(h);
float t_2 = dY_46_w_m * floorf(d);
float t_3 = dX_46_w_m * floorf(d);
float t_4 = floorf(w) * floorf(w);
float tmp;
if (dX_46_u <= 0.03999999910593033f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_v * dX_46_v), t_0, ((t_3 * dX_46_w_m) * floorf(d))), fmaf((t_4 * dY_46_u), dY_46_u, fmaf((t_0 * dY_46_v), dY_46_v, ((t_2 * dY_46_w_m) * floorf(d)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_3 * floorf(d)), dX_46_w_m, (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((dY_46_u * dY_46_u), t_4, fmaf(t_1, t_1, (t_2 * t_2))))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dY_46_w_m * floor(d)) t_3 = Float32(dX_46_w_m * floor(d)) t_4 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (dX_46_u <= Float32(0.03999999910593033)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_v * dX_46_v), t_0, Float32(Float32(t_3 * dX_46_w_m) * floor(d))), fma(Float32(t_4 * dY_46_u), dY_46_u, fma(Float32(t_0 * dY_46_v), dY_46_v, Float32(Float32(t_2 * dY_46_w_m) * floor(d))))))); else tmp = log2(sqrt(fmax(fma(Float32(t_3 * floor(d)), dX_46_w_m, Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(dY_46_u * dY_46_u), t_4, fma(t_1, t_1, Float32(t_2 * t_2)))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dY.w\_m \cdot \left\lfloor d\right\rfloor \\
t_3 := dX.w\_m \cdot \left\lfloor d\right\rfloor \\
t_4 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;dX.u \leq 0.03999999910593033:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.v \cdot dX.v, t\_0, \left(t\_3 \cdot dX.w\_m\right) \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(t\_4 \cdot dY.u, dY.u, \mathsf{fma}\left(t\_0 \cdot dY.v, dY.v, \left(t\_2 \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\_3 \cdot \left\lfloor d\right\rfloor , dX.w\_m, \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(dY.u \cdot dY.u, t\_4, \mathsf{fma}\left(t\_1, t\_1, t\_2 \cdot t\_2\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 0.0399999991Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
Taylor expanded in dX.u around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3260.9
Applied rewrites60.9%
lift-pow.f32N/A
unpow2N/A
lower-*.f3260.9
lift-pow.f32N/A
pow2N/A
lift-*.f3260.9
lift-*.f32N/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
unswap-sqrN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3260.9
Applied rewrites60.9%
if 0.0399999991 < dX.u Initial program 68.0%
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.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
dX.w_m = (fabs.f32 dX.w)
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor h) (floor h)))
(t_3 (* (floor d) dX.w_m)))
(if (<= dX.u 2200000.0)
(log2
(sqrt
(fmax
(fma (* dX.v dX.v) t_2 (* (* (* dX.w_m (floor d)) dX.w_m) (floor d)))
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(fma
(* t_2 dY.v)
dY.v
(* (* (* dY.w_m (floor d)) dY.w_m) (floor d)))))))
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_0 t_0)) (* t_3 t_3))
(* dY.u (* (fabs dY.u) (* (fabs (floor w)) (floor w))))))))))dX.w_m = fabs(dX_46_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_m, float dY_46_u, float dY_46_v, float dY_46_w_m) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(h) * floorf(h);
float t_3 = floorf(d) * dX_46_w_m;
float tmp;
if (dX_46_u <= 2200000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_v * dX_46_v), t_2, (((dX_46_w_m * floorf(d)) * dX_46_w_m) * floorf(d))), fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, fmaf((t_2 * dY_46_v), dY_46_v, (((dY_46_w_m * floorf(d)) * dY_46_w_m) * floorf(d)))))));
} else {
tmp = log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_0 * t_0)) + (t_3 * t_3)), (dY_46_u * (fabsf(dY_46_u) * (fabsf(floorf(w)) * floorf(w)))))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(floor(d) * dX_46_w_m) tmp = Float32(0.0) if (dX_46_u <= Float32(2200000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_v * dX_46_v), t_2, Float32(Float32(Float32(dX_46_w_m * floor(d)) * dX_46_w_m) * floor(d))), fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, fma(Float32(t_2 * dY_46_v), dY_46_v, Float32(Float32(Float32(dY_46_w_m * floor(d)) * dY_46_w_m) * floor(d))))))); else tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) + Float32(t_3 * t_3)), Float32(dY_46_u * Float32(abs(dY_46_u) * Float32(abs(floor(w)) * floor(w))))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
dY.w_m = \left|dY.w\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 h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor d\right\rfloor \cdot dX.w\_m\\
\mathbf{if}\;dX.u \leq 2200000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.v \cdot dX.v, t\_2, \left(\left(dX.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w\_m\right) \cdot \left\lfloor d\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(t\_2 \cdot dY.v, 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(\left(t\_1 \cdot t\_1 + t\_0 \cdot t\_0\right) + t\_3 \cdot t\_3, dY.u \cdot \left(\left|dY.u\right| \cdot \left(\left|\left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 2.2e6Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
Taylor expanded in dX.u around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3260.9
Applied rewrites60.9%
lift-pow.f32N/A
unpow2N/A
lower-*.f3260.9
lift-pow.f32N/A
pow2N/A
lift-*.f3260.9
lift-*.f32N/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
unswap-sqrN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3260.9
Applied rewrites60.9%
if 2.2e6 < dX.u Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
Applied rewrites64.2%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3250.2
Applied rewrites50.2%
dX.w_m = (fabs.f32 dX.w)
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor d) dX.w_m)))
(if (<= dY.w_m 20000000000.0)
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_2 t_2))
(* dY.u (* (fabs dY.u) (* (fabs (floor w)) (floor w)))))))
(log2
(sqrt
(fmax
(fma
(* (* dX.w_m (floor d)) (floor d))
dX.w_m
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* (* (floor d) (floor d)) dY.w_m)
dY.w_m
(* (* (* dY.u (floor w)) dY.u) (floor w)))))))))dX.w_m = fabs(dX_46_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_m, 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(h) * dX_46_v;
float t_2 = floorf(d) * dX_46_w_m;
float tmp;
if (dY_46_w_m <= 20000000000.0f) {
tmp = log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_1 * t_1)) + (t_2 * t_2)), (dY_46_u * (fabsf(dY_46_u) * (fabsf(floorf(w)) * floorf(w)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w_m * floorf(d)) * floorf(d)), dX_46_w_m, (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf(((floorf(d) * floorf(d)) * dY_46_w_m), dY_46_w_m, (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w))))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(d) * dX_46_w_m) tmp = Float32(0.0) if (dY_46_w_m <= Float32(20000000000.0)) tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_2 * t_2)), Float32(dY_46_u * Float32(abs(dY_46_u) * Float32(abs(floor(w)) * floor(w))))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w_m * floor(d)) * floor(d)), dX_46_w_m, Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(Float32(floor(d) * floor(d)) * dY_46_w_m), dY_46_w_m, Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w)))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor d\right\rfloor \cdot dX.w\_m\\
\mathbf{if}\;dY.w\_m \leq 20000000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_2 \cdot t\_2, dY.u \cdot \left(\left|dY.u\right| \cdot \left(\left|\left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dX.w\_m, \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(\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\_m, dY.w\_m, \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dY.w < 2e10Initial program 68.0%
lift-+.f32N/A
lift-+.f32N/A
lift-*.f32N/A
fp-cancel-sign-sub-invN/A
associate-+l-N/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
Applied rewrites68.0%
Applied rewrites64.2%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3250.2
Applied rewrites50.2%
if 2e10 < dY.w Initial program 68.0%
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.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
Taylor expanded in dY.v around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3253.2
Applied rewrites53.2%
lift-pow.f32N/A
pow2N/A
lift-*.f3253.2
lift-*.f32N/A
lift-pow.f32N/A
lift-pow.f32N/A
pow-prod-downN/A
lift-*.f32N/A
pow2N/A
lift-*.f3253.2
lower-fma.f32N/A
*-commutativeN/A
lift-pow.f32N/A
pow2N/A
associate-*l*N/A
lift-*.f32N/A
+-commutativeN/A
Applied rewrites53.2%
dX.w_m = (fabs.f32 dX.w)
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w_m)
:precision binary32
(log2
(sqrt
(fmax
(fma
(* (* dX.w_m (floor d)) (floor d))
dX.w_m
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* (* (floor d) (floor d)) dY.w_m)
dY.w_m
(* (* (* dY.u (floor w)) dY.u) (floor w)))))))dX.w_m = fabs(dX_46_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_m, float dY_46_u, float dY_46_v, float dY_46_w_m) {
return log2f(sqrtf(fmaxf(fmaf(((dX_46_w_m * floorf(d)) * floorf(d)), dX_46_w_m, (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf(((floorf(d) * floorf(d)) * dY_46_w_m), dY_46_w_m, (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w))))));
}
dX.w_m = abs(dX_46_w) dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w_m) return log2(sqrt(fmax(fma(Float32(Float32(dX_46_w_m * floor(d)) * floor(d)), dX_46_w_m, Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(Float32(floor(d) * floor(d)) * dY_46_w_m), dY_46_w_m, Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w)))))) end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
dY.w_m = \left|dY.w\right|
\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w\_m \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dX.w\_m, \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(\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\_m, dY.w\_m, \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}\right)
\end{array}
Initial program 68.0%
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.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
Taylor expanded in dY.v around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3253.2
Applied rewrites53.2%
lift-pow.f32N/A
pow2N/A
lift-*.f3253.2
lift-*.f32N/A
lift-pow.f32N/A
lift-pow.f32N/A
pow-prod-downN/A
lift-*.f32N/A
pow2N/A
lift-*.f3253.2
lower-fma.f32N/A
*-commutativeN/A
lift-pow.f32N/A
pow2N/A
associate-*l*N/A
lift-*.f32N/A
+-commutativeN/A
Applied rewrites53.2%
herbie shell --seed 2025162
(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)))))))