
(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}
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}
Herbie found 6 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
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}
(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) dX.v))
(t_2 (* (floor d) (fabs dY.w)))
(t_3 (* (* (floor h) (floor h)) dY.v))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u))
(t_6 (* (floor w) (floor w)))
(t_7 (* t_6 dY.u))
(t_8 (* t_1 t_1))
(t_9 (* (floor h) dY.v)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) t_8) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_9 t_9)) (* t_2 t_2)))))
100.0)
(log2
(sqrt
(fmax
(fma
(* t_6 dX.u)
dX.u
(fma
(* dX.w dX.w)
(* (floor d) (floor d))
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
t_3
dY.v
(fma
t_7
dY.u
(* (* (* (fabs dY.w) (floor d)) (fabs dY.w)) (floor d)))))))
(log2
(sqrt
(fmax
(fma (* t_4 dX.w) (floor d) t_8)
(fma
t_3
dY.v
(fma
t_7
dY.u
(exp (* (log (* (- (fabs dY.w)) (floor d))) 2.0))))))))))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) * dX_46_v;
float t_2 = floorf(d) * fabsf(dY_46_w);
float t_3 = (floorf(h) * floorf(h)) * dY_46_v;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
float t_6 = floorf(w) * floorf(w);
float t_7 = t_6 * dY_46_u;
float t_8 = t_1 * t_1;
float t_9 = floorf(h) * dY_46_v;
float tmp;
if (log2f(sqrtf(fmaxf((((t_5 * t_5) + t_8) + (t_4 * t_4)), (((t_0 * t_0) + (t_9 * t_9)) + (t_2 * t_2))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_6 * dX_46_u), dX_46_u, fmaf((dX_46_w * dX_46_w), (floorf(d) * floorf(d)), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(t_3, dY_46_v, fmaf(t_7, dY_46_u, (((fabsf(dY_46_w) * floorf(d)) * fabsf(dY_46_w)) * floorf(d)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_4 * dX_46_w), floorf(d), t_8), fmaf(t_3, dY_46_v, fmaf(t_7, dY_46_u, expf((logf((-fabsf(dY_46_w) * floorf(d))) * 2.0f)))))));
}
return tmp;
}
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) * dX_46_v) t_2 = Float32(floor(d) * abs(dY_46_w)) t_3 = Float32(Float32(floor(h) * floor(h)) * dY_46_v) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(floor(w) * floor(w)) t_7 = Float32(t_6 * dY_46_u) t_8 = Float32(t_1 * t_1) t_9 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + t_8) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_9 * t_9)) + Float32(t_2 * t_2))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(t_6 * dX_46_u), dX_46_u, fma(Float32(dX_46_w * dX_46_w), Float32(floor(d) * floor(d)), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(t_3, dY_46_v, fma(t_7, dY_46_u, Float32(Float32(Float32(abs(dY_46_w) * floor(d)) * abs(dY_46_w)) * floor(d))))))); else tmp = log2(sqrt(fmax(fma(Float32(t_4 * dX_46_w), floor(d), t_8), fma(t_3, dY_46_v, fma(t_7, dY_46_u, exp(Float32(log(Float32(Float32(-abs(dY_46_w)) * floor(d))) * Float32(2.0)))))))); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor d\right\rfloor \cdot \left|dY.w\right|\\
t_3 := \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_7 := t\_6 \cdot dY.u\\
t_8 := t\_1 \cdot t\_1\\
t_9 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_8\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_9 \cdot t\_9\right) + t\_2 \cdot t\_2\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6 \cdot dX.u, dX.u, \mathsf{fma}\left(dX.w \cdot dX.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(t\_3, dY.v, \mathsf{fma}\left(t\_7, dY.u, \left(\left(\left|dY.w\right| \cdot \left\lfloor d\right\rfloor \right) \cdot \left|dY.w\right|\right) \cdot \left\lfloor d\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot dX.w, \left\lfloor d\right\rfloor , t\_8\right), \mathsf{fma}\left(t\_3, dY.v, \mathsf{fma}\left(t\_7, dY.u, e^{\log \left(\left(-\left|dY.w\right|\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}\right)\right)\right)}\right)\\
\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 66.7%
Applied rewrites66.7%
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 rewrites66.7%
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 66.7%
Applied rewrites66.7%
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.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.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-floor.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3247.8
Applied rewrites47.8%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) (fabs dY.w)))
(t_4 (* (floor d) dX.w))
(t_5 (* t_2 t_2))
(t_6 (* (floor h) dY.v)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) t_5) (* t_4 t_4))
(+ (+ (* t_1 t_1) (* t_6 t_6)) (* t_3 t_3)))))
100.0)
(log2
(sqrt
(fmax
(fma
(* (* (fabs dY.w) (floor d)) (fabs dY.w))
(floor d)
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(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)))))))
(log2
(sqrt
(fmax
(fma (* t_4 dX.w) (floor d) t_5)
(fma
(* (* (floor h) (floor h)) dY.v)
dY.v
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(exp (* (log (* (- (fabs dY.w)) (floor d))) 2.0))))))))))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) * dX_46_u;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * fabsf(dY_46_w);
float t_4 = floorf(d) * dX_46_w;
float t_5 = t_2 * t_2;
float t_6 = floorf(h) * dY_46_v;
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + t_5) + (t_4 * t_4)), (((t_1 * t_1) + (t_6 * t_6)) + (t_3 * t_3))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((fabsf(dY_46_w) * floorf(d)) * fabsf(dY_46_w)), floorf(d), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))), 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)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_4 * dX_46_w), floorf(d), t_5), fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, expf((logf((-fabsf(dY_46_w) * floorf(d))) * 2.0f)))))));
}
return tmp;
}
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) * dX_46_u) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * abs(dY_46_w)) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(t_2 * t_2) t_6 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + t_5) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_1 * t_1) + Float32(t_6 * t_6)) + Float32(t_3 * t_3))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(abs(dY_46_w) * floor(d)) * abs(dY_46_w)), floor(d), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))), 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))))))); else tmp = log2(sqrt(fmax(fma(Float32(t_4 * dX_46_w), floor(d), t_5), fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, exp(Float32(log(Float32(Float32(-abs(dY_46_w)) * floor(d))) * Float32(2.0)))))))); end return tmp end
\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 dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot \left|dY.w\right|\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := t\_2 \cdot t\_2\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_5\right) + t\_4 \cdot t\_4, \left(t\_1 \cdot t\_1 + t\_6 \cdot t\_6\right) + t\_3 \cdot t\_3\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left|dY.w\right| \cdot \left\lfloor d\right\rfloor \right) \cdot \left|dY.w\right|, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \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)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot dX.w, \left\lfloor d\right\rfloor , t\_5\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, e^{\log \left(\left(-\left|dY.w\right|\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}\right)\right)\right)}\right)\\
\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 66.7%
Applied rewrites66.7%
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 66.7%
Applied rewrites66.7%
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.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.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-floor.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3247.8
Applied rewrites47.8%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (fabs dX.v) (floor h)))
(t_1 (* (floor d) (floor d)))
(t_2 (* (floor d) dY.w))
(t_3 (* (floor w) (floor w)))
(t_4 (* dY.v (floor h))))
(if (<= (fabs dX.v) 60.0)
(log2
(sqrt
(fmax
(fma (* dY.w dY.w) t_1 (fma (* dY.u dY.u) t_3 (* t_4 t_4)))
(fma (* dX.w dX.w) t_1 (* (* (* dX.u (floor w)) dX.u) (floor w))))))
(log2
(sqrt
(fmax
(fma (* t_3 dX.u) dX.u (fma (* t_1 dX.w) dX.w (* t_0 t_0)))
(fma (* (* (floor w) dY.u) dY.u) (floor w) (* t_2 t_2))))))))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 = fabsf(dX_46_v) * floorf(h);
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(d) * dY_46_w;
float t_3 = floorf(w) * floorf(w);
float t_4 = dY_46_v * floorf(h);
float tmp;
if (fabsf(dX_46_v) <= 60.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dY_46_w * dY_46_w), t_1, fmaf((dY_46_u * dY_46_u), t_3, (t_4 * t_4))), fmaf((dX_46_w * dX_46_w), t_1, (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_3 * dX_46_u), dX_46_u, fmaf((t_1 * dX_46_w), dX_46_w, (t_0 * t_0))), fmaf(((floorf(w) * dY_46_u) * dY_46_u), floorf(w), (t_2 * t_2)))));
}
return tmp;
}
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(abs(dX_46_v) * floor(h)) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(d) * dY_46_w) t_3 = Float32(floor(w) * floor(w)) t_4 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (abs(dX_46_v) <= Float32(60.0)) tmp = log2(sqrt(fmax(fma(Float32(dY_46_w * dY_46_w), t_1, fma(Float32(dY_46_u * dY_46_u), t_3, Float32(t_4 * t_4))), fma(Float32(dX_46_w * dX_46_w), t_1, Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)))))); else tmp = log2(sqrt(fmax(fma(Float32(t_3 * dX_46_u), dX_46_u, fma(Float32(t_1 * dX_46_w), dX_46_w, Float32(t_0 * t_0))), fma(Float32(Float32(floor(w) * dY_46_u) * dY_46_u), floor(w), Float32(t_2 * t_2))))); end return tmp end
\begin{array}{l}
t_0 := \left|dX.v\right| \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_4 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\left|dX.v\right| \leq 60:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \mathsf{fma}\left(dY.u \cdot dY.u, t\_3, t\_4 \cdot t\_4\right)\right), \mathsf{fma}\left(dX.w \cdot dX.w, t\_1, \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3 \cdot dX.u, dX.u, \mathsf{fma}\left(t\_1 \cdot dX.w, dX.w, t\_0 \cdot t\_0\right)\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u, \left\lfloor w\right\rfloor , t\_2 \cdot t\_2\right)\right)}\right)\\
\end{array}
if dX.v < 60Initial program 66.7%
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.f3259.4
Applied rewrites59.4%
Applied rewrites59.4%
if 60 < dX.v Initial program 66.7%
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.f3259.7
Applied rewrites59.7%
Applied rewrites59.7%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) (fabs dX.v)))
(t_1 (* (floor d) (floor d)))
(t_2 (* (floor w) (floor w)))
(t_3 (* dY.v (floor h))))
(if (<= (fabs dX.v) 13.5)
(log2
(sqrt
(fmax
(fma (* dY.w dY.w) t_1 (fma (* dY.u dY.u) t_2 (* t_3 t_3)))
(fma (* dX.w dX.w) t_1 (* (* (* dX.u (floor w)) dX.u) (floor w))))))
(log2
(sqrt
(fmax
(fma (* (* (floor d) dX.w) dX.w) (floor d) (* t_0 t_0))
(fma
(* (* (floor h) (floor h)) dY.v)
dY.v
(fma
(* t_2 dY.u)
dY.u
(* (* (* dY.w (floor d)) dY.w) (floor d))))))))))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(h) * fabsf(dX_46_v);
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(w) * floorf(w);
float t_3 = dY_46_v * floorf(h);
float tmp;
if (fabsf(dX_46_v) <= 13.5f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dY_46_w * dY_46_w), t_1, fmaf((dY_46_u * dY_46_u), t_2, (t_3 * t_3))), fmaf((dX_46_w * dX_46_w), t_1, (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w) * dX_46_w), floorf(d), (t_0 * t_0)), fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, fmaf((t_2 * dY_46_u), dY_46_u, (((dY_46_w * floorf(d)) * dY_46_w) * floorf(d)))))));
}
return tmp;
}
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(h) * abs(dX_46_v)) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (abs(dX_46_v) <= Float32(13.5)) tmp = log2(sqrt(fmax(fma(Float32(dY_46_w * dY_46_w), t_1, fma(Float32(dY_46_u * dY_46_u), t_2, Float32(t_3 * t_3))), fma(Float32(dX_46_w * dX_46_w), t_1, Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w) * dX_46_w), floor(d), Float32(t_0 * t_0)), fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(Float32(dY_46_w * floor(d)) * dY_46_w) * floor(d))))))); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left|dX.v\right|\\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\left|dX.v\right| \leq 13.5:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \mathsf{fma}\left(dY.u \cdot dY.u, t\_2, t\_3 \cdot t\_3\right)\right), \mathsf{fma}\left(dX.w \cdot dX.w, t\_1, \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\right) \cdot dX.w, \left\lfloor d\right\rfloor , t\_0 \cdot t\_0\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right)\right)\right)}\right)\\
\end{array}
if dX.v < 13.5Initial program 66.7%
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.f3259.4
Applied rewrites59.4%
Applied rewrites59.4%
if 13.5 < dX.v Initial program 66.7%
Applied rewrites66.7%
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.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v)))
(log2
(sqrt
(fmax
(fma (* (* (floor d) dX.w) dX.w) (floor d) (* t_0 t_0))
(fma
(* (* (floor h) (floor h)) dY.v)
dY.v
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(* (* (* dY.w (floor d)) dY.w) (floor d)))))))))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(h) * dX_46_v;
return log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w) * dX_46_w), floorf(d), (t_0 * t_0)), fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, (((dY_46_w * floorf(d)) * dY_46_w) * floorf(d)))))));
}
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(h) * dX_46_v) return log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w) * dX_46_w), floor(d), Float32(t_0 * t_0)), fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(Float32(Float32(dY_46_w * floor(d)) * dY_46_w) * floor(d))))))) end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\right) \cdot dX.w, \left\lfloor d\right\rfloor , t\_0 \cdot t\_0\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right)\right)\right)}\right)
\end{array}
Initial program 66.7%
Applied rewrites66.7%
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.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(* (fabs dY.u) (* (fabs (* (fabs dY.u) (floor w))) (floor w)))
(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))))))))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) {
return log2f(sqrtf(fmaxf((fabsf(dY_46_u) * (fabsf((fabsf(dY_46_u) * floorf(w))) * floorf(w))), 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)))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt(fmax(Float32(abs(dY_46_u) * Float32(abs(Float32(abs(dY_46_u) * floor(w))) * floor(w))), 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))))))) end
\log_{2} \left(\sqrt{\mathsf{max}\left(\left|dY.u\right| \cdot \left(\left|\left|dY.u\right| \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right), \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)\right)}\right)
Initial program 66.7%
Applied rewrites66.7%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3266.7
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.f3262.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3262.8
Applied rewrites62.8%
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.f3248.8
Applied rewrites48.8%
herbie shell --seed 2025175
(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)))))))