
(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 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}
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) dX.u))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (+ (* t_1 t_1) (* t_2 t_2)))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor d) (fabs dY.w)))
(t_6 (* (floor d) dX.w)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_4 t_4)) (* t_6 t_6))
(+ t_3 (* t_5 t_5)))))
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
(pow dX.v 2.0)
(pow (floor h) 2.0)
(* (pow dX.w 2.0) (pow (floor d) 2.0)))
(+ t_3 (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) * dY_46_v;
float t_3 = (t_1 * t_1) + (t_2 * t_2);
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(d) * fabsf(dY_46_w);
float t_6 = floorf(d) * dX_46_w;
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_4 * t_4)) + (t_6 * t_6)), (t_3 + (t_5 * t_5))))) <= 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(powf(dX_46_v, 2.0f), powf(floorf(h), 2.0f), (powf(dX_46_w, 2.0f) * powf(floorf(d), 2.0f))), (t_3 + 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) * dY_46_v) t_3 = Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(d) * abs(dY_46_w)) t_6 = Float32(floor(d) * dX_46_w) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) + Float32(t_6 * t_6)), Float32(t_3 + Float32(t_5 * t_5))))) <= 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((dX_46_v ^ Float32(2.0)), (floor(h) ^ Float32(2.0)), Float32((dX_46_w ^ Float32(2.0)) * (floor(d) ^ Float32(2.0)))), Float32(t_3 + 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 dY.v\\
t_3 := t\_1 \cdot t\_1 + t\_2 \cdot t\_2\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor d\right\rfloor \cdot \left|dY.w\right|\\
t_6 := \left\lfloor d\right\rfloor \cdot dX.w\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_4 \cdot t\_4\right) + t\_6 \cdot t\_6, t\_3 + t\_5 \cdot t\_5\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({dX.v}^{2}, {\left(\left\lfloor h\right\rfloor \right)}^{2}, {dX.w}^{2} \cdot {\left(\left\lfloor d\right\rfloor \right)}^{2}\right), t\_3 + e^{\log \left(\left(-\left|dY.w\right|\right) \cdot \left\lfloor d\right\rfloor \right) \cdot 2}\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 68.5%
Applied rewrites68.5%
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 68.5%
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.f3261.3%
Applied rewrites61.3%
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
lift-*.f32N/A
lift-*.f32N/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3248.8%
Applied rewrites48.8%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) (fabs dY.u)))
(t_3 (* (floor d) dY.w))
(t_4 (* dX.w (floor d)))
(t_5 (* (floor d) dX.w))
(t_6 (* (floor h) dX.v))
(t_7 (* (floor h) dY.v)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_6 t_6)) (* t_5 t_5))
(+ (+ (* t_2 t_2) (* t_7 t_7)) (* t_3 t_3)))))
100.0)
(log2
(sqrt
(fmax
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* (fabs dY.u) (floor w)) (fabs dY.u))
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(fma
(* t_4 dX.w)
(floor d)
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* t_0 dX.v) (floor h)))))))
(log2
(sqrt
(fmax
(fma
(* t_3 dY.w)
(floor d)
(fma t_7 t_7 (exp (* (log (* (- (fabs dY.u)) (floor w))) 2.0))))
(fma t_4 t_4 (* t_0 t_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 = dX_46_v * floorf(h);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * fabsf(dY_46_u);
float t_3 = floorf(d) * dY_46_w;
float t_4 = dX_46_w * floorf(d);
float t_5 = floorf(d) * dX_46_w;
float t_6 = floorf(h) * dX_46_v;
float t_7 = floorf(h) * dY_46_v;
float tmp;
if (log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_6 * t_6)) + (t_5 * t_5)), (((t_2 * t_2) + (t_7 * t_7)) + (t_3 * t_3))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((fabsf(dY_46_u) * floorf(w)) * fabsf(dY_46_u)), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))), fmaf((t_4 * dX_46_w), floorf(d), fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), ((t_0 * dX_46_v) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_3 * dY_46_w), floorf(d), fmaf(t_7, t_7, expf((logf((-fabsf(dY_46_u) * floorf(w))) * 2.0f)))), fmaf(t_4, t_4, (t_0 * t_0)))));
}
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(dX_46_v * floor(h)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * abs(dY_46_u)) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(dX_46_w * floor(d)) t_5 = Float32(floor(d) * dX_46_w) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_6 * t_6)) + Float32(t_5 * t_5)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_7 * t_7)) + Float32(t_3 * t_3))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(abs(dY_46_u) * floor(w)) * abs(dY_46_u)), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))), fma(Float32(t_4 * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(t_0 * dX_46_v) * floor(h))))))); else tmp = log2(sqrt(fmax(fma(Float32(t_3 * dY_46_w), floor(d), fma(t_7, t_7, exp(Float32(log(Float32(Float32(-abs(dY_46_u)) * floor(w))) * Float32(2.0))))), fma(t_4, t_4, Float32(t_0 * t_0))))); end return tmp end
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left|dY.u\right|\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_5 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_6 \cdot t\_6\right) + t\_5 \cdot t\_5, \left(t\_2 \cdot t\_2 + t\_7 \cdot t\_7\right) + t\_3 \cdot t\_3\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(\left|dY.u\right| \cdot \left\lfloor w\right\rfloor \right) \cdot \left|dY.u\right|, \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(t\_4 \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(t\_0 \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\_3 \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_7, t\_7, e^{\log \left(\left(-\left|dY.u\right|\right) \cdot \left\lfloor w\right\rfloor \right) \cdot 2}\right)\right), \mathsf{fma}\left(t\_4, t\_4, t\_0 \cdot t\_0\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 68.5%
Applied rewrites68.5%
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 68.5%
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.f3261.3%
Applied rewrites61.3%
Applied rewrites61.3%
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.f3249.2%
Applied rewrites49.2%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dX.w (floor d)))
(t_1 (* dX.v (floor h)))
(t_2 (* (* (floor w) (floor w)) (* (fabs dX.u) (fabs dX.u))))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor w) dY.u)))
(if (<= (fabs dX.u) 5000000.0)
(log2
(sqrt
(fmax
(fma (* (* (floor d) dY.w) dY.w) (floor d) (fma t_3 t_3 (* t_4 t_4)))
(fma t_0 t_0 (* t_1 t_1)))))
(log2
(sqrt
(fmax
(* dY.w (* (fabs (* dY.w (floor d))) (floor d)))
(fma
(* t_0 dX.w)
(floor d)
(* (+ 1.0 (/ (* (* t_1 dX.v) (floor h)) 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 = dX_46_w * floorf(d);
float t_1 = dX_46_v * floorf(h);
float t_2 = (floorf(w) * floorf(w)) * (fabsf(dX_46_u) * fabsf(dX_46_u));
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(w) * dY_46_u;
float tmp;
if (fabsf(dX_46_u) <= 5000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dY_46_w) * dY_46_w), floorf(d), fmaf(t_3, t_3, (t_4 * t_4))), fmaf(t_0, t_0, (t_1 * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf((dY_46_w * (fabsf((dY_46_w * floorf(d))) * floorf(d))), fmaf((t_0 * dX_46_w), floorf(d), ((1.0f + (((t_1 * dX_46_v) * floorf(h)) / 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(dX_46_w * floor(d)) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(Float32(floor(w) * floor(w)) * Float32(abs(dX_46_u) * abs(dX_46_u))) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(w) * dY_46_u) tmp = Float32(0.0) if (abs(dX_46_u) <= Float32(5000000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dY_46_w) * dY_46_w), floor(d), fma(t_3, t_3, Float32(t_4 * t_4))), fma(t_0, t_0, Float32(t_1 * t_1))))); else tmp = log2(sqrt(fmax(Float32(dY_46_w * Float32(abs(Float32(dY_46_w * floor(d))) * floor(d))), fma(Float32(t_0 * dX_46_w), floor(d), Float32(Float32(Float32(1.0) + Float32(Float32(Float32(t_1 * dX_46_v) * floor(h)) / t_2)) * t_2))))); end return tmp end
\begin{array}{l}
t_0 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(\left|dX.u\right| \cdot \left|dX.u\right|\right)\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\\
\mathbf{if}\;\left|dX.u\right| \leq 5000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_3, t\_3, t\_4 \cdot t\_4\right)\right), \mathsf{fma}\left(t\_0, t\_0, t\_1 \cdot t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(dY.w \cdot \left(\left|dY.w \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(t\_0 \cdot dX.w, \left\lfloor d\right\rfloor , \left(1 + \frac{\left(t\_1 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor }{t\_2}\right) \cdot t\_2\right)\right)}\right)\\
\end{array}
if dX.u < 5e6Initial program 68.5%
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.f3261.3%
Applied rewrites61.3%
Applied rewrites61.3%
if 5e6 < dX.u Initial program 68.5%
Applied rewrites68.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.5%
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
lower-fabs.f3264.4%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.4%
Applied rewrites64.4%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3250.6%
Applied rewrites50.6%
Applied rewrites44.4%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dX.w (floor d)))
(t_1 (* dX.v (floor h)))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dY.u)))
(if (<= (fabs dX.u) 5000000.0)
(log2
(sqrt
(fmax
(fma (* (* (floor d) dY.w) dY.w) (floor d) (fma t_2 t_2 (* t_3 t_3)))
(fma t_0 t_0 (* t_1 t_1)))))
(log2
(sqrt
(fmax
(* dY.w (* (fabs (* dY.w (floor d))) (floor d)))
(fma
(* t_0 dX.w)
(floor d)
(fma
t_1
t_1
(* (* (floor w) (floor w)) (* (fabs dX.u) (fabs dX.u)))))))))))float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = dX_46_w * floorf(d);
float t_1 = dX_46_v * floorf(h);
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dY_46_u;
float tmp;
if (fabsf(dX_46_u) <= 5000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dY_46_w) * dY_46_w), floorf(d), fmaf(t_2, t_2, (t_3 * t_3))), fmaf(t_0, t_0, (t_1 * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf((dY_46_w * (fabsf((dY_46_w * floorf(d))) * floorf(d))), fmaf((t_0 * dX_46_w), floorf(d), fmaf(t_1, t_1, ((floorf(w) * floorf(w)) * (fabsf(dX_46_u) * fabsf(dX_46_u))))))));
}
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(dX_46_w * floor(d)) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dY_46_u) tmp = Float32(0.0) if (abs(dX_46_u) <= Float32(5000000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dY_46_w) * dY_46_w), floor(d), fma(t_2, t_2, Float32(t_3 * t_3))), fma(t_0, t_0, Float32(t_1 * t_1))))); else tmp = log2(sqrt(fmax(Float32(dY_46_w * Float32(abs(Float32(dY_46_w * floor(d))) * floor(d))), fma(Float32(t_0 * dX_46_w), floor(d), fma(t_1, t_1, Float32(Float32(floor(w) * floor(w)) * Float32(abs(dX_46_u) * abs(dX_46_u)))))))); end return tmp end
\begin{array}{l}
t_0 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
\mathbf{if}\;\left|dX.u\right| \leq 5000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_2, t\_2, t\_3 \cdot t\_3\right)\right), \mathsf{fma}\left(t\_0, t\_0, t\_1 \cdot t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(dY.w \cdot \left(\left|dY.w \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(t\_0 \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_1, t\_1, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(\left|dX.u\right| \cdot \left|dX.u\right|\right)\right)\right)\right)}\right)\\
\end{array}
if dX.u < 5e6Initial program 68.5%
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.f3261.3%
Applied rewrites61.3%
Applied rewrites61.3%
if 5e6 < dX.u Initial program 68.5%
Applied rewrites68.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.5%
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
lower-fabs.f3264.4%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.4%
Applied rewrites64.4%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3250.6%
Applied rewrites50.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3250.6%
lower-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
sqr-abs-revN/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
sqr-abs-revN/A
Applied rewrites50.6%
(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 w) (floor w)) (* dX.u dX.u)))
(t_2 (* (floor d) (floor d)))
(t_3 (* (floor w) dY.u)))
(if (<= (fabs dX.v) 600000.0)
(log2
(sqrt
(fmax
(fma
(* dY.v dY.v)
(* (floor h) (floor h))
(fma (* (fabs dY.w) (fabs dY.w)) t_2 (* t_3 t_3)))
(fma (* t_2 dX.w) dX.w t_1))))
(log2
(sqrt
(fmax
(* (fabs dY.w) (* (fabs (* (fabs dY.w) (floor d))) (floor d)))
(fma (* (* dX.w (floor d)) dX.w) (floor d) (fma t_0 t_0 t_1))))))))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(w) * floorf(w)) * (dX_46_u * dX_46_u);
float t_2 = floorf(d) * floorf(d);
float t_3 = floorf(w) * dY_46_u;
float tmp;
if (fabsf(dX_46_v) <= 600000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dY_46_v * dY_46_v), (floorf(h) * floorf(h)), fmaf((fabsf(dY_46_w) * fabsf(dY_46_w)), t_2, (t_3 * t_3))), fmaf((t_2 * dX_46_w), dX_46_w, t_1))));
} else {
tmp = log2f(sqrtf(fmaxf((fabsf(dY_46_w) * (fabsf((fabsf(dY_46_w) * floorf(d))) * floorf(d))), fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(t_0, t_0, t_1)))));
}
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(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u)) t_2 = Float32(floor(d) * floor(d)) t_3 = Float32(floor(w) * dY_46_u) tmp = Float32(0.0) if (abs(dX_46_v) <= Float32(600000.0)) tmp = log2(sqrt(fmax(fma(Float32(dY_46_v * dY_46_v), Float32(floor(h) * floor(h)), fma(Float32(abs(dY_46_w) * abs(dY_46_w)), t_2, Float32(t_3 * t_3))), fma(Float32(t_2 * dX_46_w), dX_46_w, t_1)))); else tmp = log2(sqrt(fmax(Float32(abs(dY_46_w) * Float32(abs(Float32(abs(dY_46_w) * floor(d))) * floor(d))), fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(t_0, t_0, t_1))))); end return tmp end
\begin{array}{l}
t_0 := \left|dX.v\right| \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\\
t_2 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
\mathbf{if}\;\left|dX.v\right| \leq 600000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.v \cdot dY.v, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , \mathsf{fma}\left(\left|dY.w\right| \cdot \left|dY.w\right|, t\_2, t\_3 \cdot t\_3\right)\right), \mathsf{fma}\left(t\_2 \cdot dX.w, dX.w, t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left|dY.w\right| \cdot \left(\left|\left|dY.w\right| \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\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(t\_0, t\_0, t\_1\right)\right)\right)}\right)\\
\end{array}
if dX.v < 6e5Initial program 68.5%
lift-+.f32N/A
lift-+.f32N/A
+-commutativeN/A
associate-+l+N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
+-commutativeN/A
lower-fma.f32N/A
Applied rewrites68.5%
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.f3261.2%
Applied rewrites61.2%
Applied rewrites61.2%
if 6e5 < dX.v Initial program 68.5%
Applied rewrites68.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.5%
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
lower-fabs.f3264.4%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.4%
Applied rewrites64.4%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3250.6%
Applied rewrites50.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3250.6%
lower-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
sqr-abs-revN/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
sqr-abs-revN/A
Applied rewrites50.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) (floor w))) (t_1 (* (fabs dX.v) (floor h))))
(if (<= (fabs dX.v) 10000.0)
(log2
(sqrt
(fmax
(* 1.0 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* (floor h) (floor h))
(* dY.v dY.v)
(fma
(* dY.u dY.u)
t_0
(* (* (* (floor d) (fabs dY.w)) (fabs dY.w)) (floor d)))))))
(log2
(sqrt
(fmax
(* (fabs dY.w) (* (fabs (* (fabs dY.w) (floor d))) (floor d)))
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma t_1 t_1 (* t_0 (* dX.u dX.u))))))))))float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * floorf(w);
float t_1 = fabsf(dX_46_v) * floorf(h);
float tmp;
if (fabsf(dX_46_v) <= 10000.0f) {
tmp = log2f(sqrtf(fmaxf((1.0f * (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((floorf(h) * floorf(h)), (dY_46_v * dY_46_v), fmaf((dY_46_u * dY_46_u), t_0, (((floorf(d) * fabsf(dY_46_w)) * fabsf(dY_46_w)) * floorf(d)))))));
} else {
tmp = log2f(sqrtf(fmaxf((fabsf(dY_46_w) * (fabsf((fabsf(dY_46_w) * floorf(d))) * floorf(d))), fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(t_1, t_1, (t_0 * (dX_46_u * dX_46_u)))))));
}
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) * floor(w)) t_1 = Float32(abs(dX_46_v) * floor(h)) tmp = Float32(0.0) if (abs(dX_46_v) <= Float32(10000.0)) tmp = log2(sqrt(fmax(Float32(Float32(1.0) * Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(floor(h) * floor(h)), Float32(dY_46_v * dY_46_v), fma(Float32(dY_46_u * dY_46_u), t_0, Float32(Float32(Float32(floor(d) * abs(dY_46_w)) * abs(dY_46_w)) * floor(d))))))); else tmp = log2(sqrt(fmax(Float32(abs(dY_46_w) * Float32(abs(Float32(abs(dY_46_w) * floor(d))) * floor(d))), fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(t_1, t_1, Float32(t_0 * Float32(dX_46_u * dX_46_u))))))); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left|dX.v\right| \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\left|dX.v\right| \leq 10000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(1 \cdot \left(\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\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , dY.v \cdot dY.v, \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, \left(\left(\left\lfloor d\right\rfloor \cdot \left|dY.w\right|\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(\left|dY.w\right| \cdot \left(\left|\left|dY.w\right| \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\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(t\_1, t\_1, t\_0 \cdot \left(dX.u \cdot dX.u\right)\right)\right)\right)}\right)\\
\end{array}
if dX.v < 1e4Initial program 68.5%
lift-+.f32N/A
lift-+.f32N/A
associate-+l+N/A
sum-to-multN/A
lower-unsound-*.f32N/A
Applied rewrites56.7%
Taylor expanded in dX.u around inf
Applied rewrites54.1%
Applied rewrites54.1%
if 1e4 < dX.v Initial program 68.5%
Applied rewrites68.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.5%
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
lower-fabs.f3264.4%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.4%
Applied rewrites64.4%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3250.6%
Applied rewrites50.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3250.6%
lower-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
sqr-abs-revN/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
sqr-abs-revN/A
Applied rewrites50.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dX.v (floor h))))
(log2
(sqrt
(fmax
(* (fabs dY.w) (* (fabs (* (fabs dY.w) (floor d))) (floor d)))
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma t_0 t_0 (* (* (floor w) (floor w)) (* dX.u dX.u)))))))))float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = dX_46_v * floorf(h);
return log2f(sqrtf(fmaxf((fabsf(dY_46_w) * (fabsf((fabsf(dY_46_w) * floorf(d))) * floorf(d))), fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(t_0, t_0, ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u)))))));
}
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(dX_46_v * floor(h)) return log2(sqrt(fmax(Float32(abs(dY_46_w) * Float32(abs(Float32(abs(dY_46_w) * floor(d))) * floor(d))), fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(t_0, t_0, Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))))))) end
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left|dY.w\right| \cdot \left(\left|\left|dY.w\right| \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\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(t\_0, t\_0, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right)\right)\right)}\right)
\end{array}
Initial program 68.5%
Applied rewrites68.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.5%
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
lower-fabs.f3264.4%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.4%
Applied rewrites64.4%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3250.6%
Applied rewrites50.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3250.6%
lower-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
sqr-abs-revN/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
sqr-abs-revN/A
Applied rewrites50.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(* (fabs dY.w) (* (fabs (* (fabs dY.w) (floor d))) (floor d)))
(* (pow dX.u 2.0) (pow (floor w) 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) {
return log2f(sqrtf(fmaxf((fabsf(dY_46_w) * (fabsf((fabsf(dY_46_w) * floorf(d))) * floorf(d))), (powf(dX_46_u, 2.0f) * powf(floorf(w), 2.0f)))));
}
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_w) * Float32(abs(Float32(abs(dY_46_w) * floor(d))) * floor(d))), Float32((dX_46_u ^ Float32(2.0)) * (floor(w) ^ Float32(2.0)))))) 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) tmp = log2(sqrt(max((abs(dY_46_w) * (abs((abs(dY_46_w) * floor(d))) * floor(d))), ((dX_46_u ^ single(2.0)) * (floor(w) ^ single(2.0)))))); end
\log_{2} \left(\sqrt{\mathsf{max}\left(\left|dY.w\right| \cdot \left(\left|\left|dY.w\right| \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\right\rfloor \right), {dX.u}^{2} \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right)}\right)
Initial program 68.5%
Applied rewrites68.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.5%
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
lower-fabs.f3264.4%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.4%
Applied rewrites64.4%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3250.6%
Applied rewrites50.6%
Taylor expanded in dX.u around inf
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3229.6%
Applied rewrites29.6%
herbie shell --seed 2025189
(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)))))))