
(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 10 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.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m 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 (* (floor h) dX.v_m))
(t_4 (* (floor d) dY.w))
(t_5 (* (floor d) (floor d)))
(t_6 (* (floor d) dX.w))
(t_7
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_3 t_3)) (* t_6 t_6))
(+ (+ (* t_1 t_1) (* t_2 t_2)) (* t_4 t_4)))))))
(if (<= t_7 100.0)
t_7
(log2
(sqrt
(fmax
(fma
t_5
(* dX.w dX.w)
(*
(* (floor h) (floor h))
(exp (+ 0.0 (* 2.0 (- (- (log (- dX.v_m)))))))))
(fma (* dY.w dY.w) t_5 (* (* dY.u dY.u) (* (floor w) (floor w))))))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, 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 = floorf(h) * dX_46_v_m;
float t_4 = floorf(d) * dY_46_w;
float t_5 = floorf(d) * floorf(d);
float t_6 = floorf(d) * dX_46_w;
float t_7 = log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_3 * t_3)) + (t_6 * t_6)), (((t_1 * t_1) + (t_2 * t_2)) + (t_4 * t_4)))));
float tmp;
if (t_7 <= 100.0f) {
tmp = t_7;
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_5, (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * expf((0.0f + (2.0f * -(-logf(-dX_46_v_m))))))), fmaf((dY_46_w * dY_46_w), t_5, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
return tmp;
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, 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(floor(h) * dX_46_v_m) t_4 = Float32(floor(d) * dY_46_w) t_5 = Float32(floor(d) * floor(d)) t_6 = Float32(floor(d) * dX_46_w) t_7 = log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) + Float32(t_6 * t_6)), Float32(Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) + Float32(t_4 * t_4))))) tmp = Float32(0.0) if (t_7 <= Float32(100.0)) tmp = t_7; else tmp = log2(sqrt(fmax(fma(t_5, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(h) * floor(h)) * exp(Float32(Float32(0.0) + Float32(Float32(2.0) * Float32(-Float32(-log(Float32(-dX_46_v_m))))))))), fma(Float32(dY_46_w * dY_46_w), t_5, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))); end return tmp end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\_m\\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_5 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_6 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_7 := \log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_3 \cdot t\_3\right) + t\_6 \cdot t\_6, \left(t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4\right)}\right)\\
\mathbf{if}\;t\_7 \leq 100:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5, dX.w \cdot dX.w, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot e^{0 + 2 \cdot \left(-\left(-\log \left(-dX.v\_m\right)\right)\right)}\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_5, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 67.5%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3252.6
Applied rewrites52.6%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3252.5
Applied rewrites52.5%
Taylor expanded in dX.v around -inf
distribute-lft-inN/A
log-pow-revN/A
metadata-evalN/A
metadata-evalN/A
lower-+.f32N/A
lower-*.f32N/A
mul-1-negN/A
lower-neg.f32N/A
frac-2negN/A
metadata-evalN/A
log-recN/A
lower-neg.f32N/A
lower-log.f32N/A
lower-neg.f3235.7
Applied rewrites35.7%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor d) (floor d)))
(t_2 (* (floor h) dX.v_m))
(t_3 (* (floor d) dX.w)))
(if (<= dX.u 70000.0)
(log2
(sqrt
(fmax
(fma t_1 (* dX.w dX.w) (* (* (floor h) (floor h)) (* dX.v_m dX.v_m)))
(fma
(* (* (floor w) dY.u) (floor w))
dY.u
(fma
(* (* dY.v dY.v) (floor h))
(floor h)
(* (* (* (floor d) dY.w) (floor d)) dY.w))))))
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_2 t_2)) (* t_3 t_3))
(fma (* dY.w dY.w) t_1 (* (* dY.u dY.u) (* (floor w) (floor w))))))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, 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(d) * floorf(d);
float t_2 = floorf(h) * dX_46_v_m;
float t_3 = floorf(d) * dX_46_w;
float tmp;
if (dX_46_u <= 70000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * (dX_46_v_m * dX_46_v_m))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, fmaf(((dY_46_v * dY_46_v) * floorf(h)), floorf(h), (((floorf(d) * dY_46_w) * floorf(d)) * dY_46_w))))));
} else {
tmp = log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_2 * t_2)) + (t_3 * t_3)), fmaf((dY_46_w * dY_46_w), t_1, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
return tmp;
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(h) * dX_46_v_m) t_3 = Float32(floor(d) * dX_46_w) tmp = Float32(0.0) if (dX_46_u <= Float32(70000.0)) tmp = log2(sqrt(fmax(fma(t_1, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v_m * dX_46_v_m))), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, fma(Float32(Float32(dY_46_v * dY_46_v) * floor(h)), floor(h), Float32(Float32(Float32(floor(d) * dY_46_w) * floor(d)) * dY_46_w)))))); else tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) + Float32(t_3 * t_3)), fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))); end return tmp end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\_m\\
t_3 := \left\lfloor d\right\rfloor \cdot dX.w\\
\mathbf{if}\;dX.u \leq 70000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, dX.w \cdot dX.w, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v\_m \cdot dX.v\_m\right)\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, \mathsf{fma}\left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_2 \cdot t\_2\right) + t\_3 \cdot t\_3, \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 7e4Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Applied rewrites60.2%
if 7e4 < dX.u Initial program 67.5%
Taylor expanded in dY.v around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.5
Applied rewrites60.5%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m 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 d) dY.w))
(t_3 (* (floor d) (floor d))))
(if (<= dY.u 50000000.0)
(log2
(sqrt
(fmax
(fma
(* (* (floor w) dX.u) (floor w))
dX.u
(fma (* dX.v_m dX.v_m) (* (floor h) (floor h)) (* (* dX.w dX.w) t_3)))
(fma
(* (* dY.v dY.v) (floor h))
(floor h)
(* (* t_2 (floor d)) dY.w)))))
(log2
(sqrt
(fmax
(fma t_3 (* dX.w dX.w) (* (* (floor w) (floor w)) (* dX.u dX.u)))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_2 t_2))))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, 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(d) * dY_46_w;
float t_3 = floorf(d) * floorf(d);
float tmp;
if (dY_46_u <= 50000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, fmaf((dX_46_v_m * dX_46_v_m), (floorf(h) * floorf(h)), ((dX_46_w * dX_46_w) * t_3))), fmaf(((dY_46_v * dY_46_v) * floorf(h)), floorf(h), ((t_2 * floorf(d)) * dY_46_w)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_3, (dX_46_w * dX_46_w), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u))), (((t_0 * t_0) + (t_1 * t_1)) + (t_2 * t_2)))));
}
return tmp;
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, 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(d) * dY_46_w) t_3 = Float32(floor(d) * floor(d)) tmp = Float32(0.0) if (dY_46_u <= Float32(50000000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, fma(Float32(dX_46_v_m * dX_46_v_m), Float32(floor(h) * floor(h)), Float32(Float32(dX_46_w * dX_46_w) * t_3))), fma(Float32(Float32(dY_46_v * dY_46_v) * floor(h)), floor(h), Float32(Float32(t_2 * floor(d)) * dY_46_w))))); else tmp = log2(sqrt(fmax(fma(t_3, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_2 * t_2))))); end return tmp end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\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 d\right\rfloor \cdot dY.w\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dY.u \leq 50000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, \mathsf{fma}\left(dX.v\_m \cdot dX.v\_m, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , \left(dX.w \cdot dX.w\right) \cdot t\_3\right)\right), \mathsf{fma}\left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(t\_2 \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.w \cdot dX.w, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right), \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_2 \cdot t\_2\right)}\right)\\
\end{array}
\end{array}
if dY.u < 5e7Initial program 67.5%
Taylor expanded in dY.u around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.4
Applied rewrites60.4%
Applied rewrites60.4%
if 5e7 < dY.u Initial program 67.5%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.4
Applied rewrites60.4%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1
(fma
(* (* dY.v dY.v) (floor h))
(floor h)
(* (* (* (floor d) dY.w) (floor d)) dY.w)))
(t_2 (* (floor h) (floor h))))
(if (<= dX.u 40000.0)
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* t_2 (* dX.v_m dX.v_m)))
(fma (* (* (floor w) dY.u) (floor w)) dY.u t_1))))
(log2
(sqrt
(fmax
(fma
(* (* (floor w) dX.u) (floor w))
dX.u
(fma (* dX.v_m dX.v_m) t_2 (* (* dX.w dX.w) t_0)))
t_1))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
float t_1 = fmaf(((dY_46_v * dY_46_v) * floorf(h)), floorf(h), (((floorf(d) * dY_46_w) * floorf(d)) * dY_46_w));
float t_2 = floorf(h) * floorf(h);
float tmp;
if (dX_46_u <= 40000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_2 * (dX_46_v_m * dX_46_v_m))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, t_1))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, fmaf((dX_46_v_m * dX_46_v_m), t_2, ((dX_46_w * dX_46_w) * t_0))), t_1)));
}
return tmp;
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = fma(Float32(Float32(dY_46_v * dY_46_v) * floor(h)), floor(h), Float32(Float32(Float32(floor(d) * dY_46_w) * floor(d)) * dY_46_w)) t_2 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dX_46_u <= Float32(40000.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_2 * Float32(dX_46_v_m * dX_46_v_m))), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, t_1)))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, fma(Float32(dX_46_v_m * dX_46_v_m), t_2, Float32(Float32(dX_46_w * dX_46_w) * t_0))), t_1))); end return tmp end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \mathsf{fma}\left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right)\\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.u \leq 40000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_2 \cdot \left(dX.v\_m \cdot dX.v\_m\right)\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, \mathsf{fma}\left(dX.v\_m \cdot dX.v\_m, t\_2, \left(dX.w \cdot dX.w\right) \cdot t\_0\right)\right), t\_1\right)}\right)\\
\end{array}
\end{array}
if dX.u < 4e4Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Applied rewrites60.2%
if 4e4 < dX.u Initial program 67.5%
Taylor expanded in dY.u around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.4
Applied rewrites60.4%
Applied rewrites60.4%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(fma
(* (floor d) (floor d))
(* dX.w dX.w)
(* (* (floor h) (floor h)) (* dX.v_m dX.v_m)))
(fma
(* (* (floor w) dY.u) (floor w))
dY.u
(fma
(* (* dY.v dY.v) (floor h))
(floor h)
(* (* (* (floor d) dY.w) (floor d)) dY.w)))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
return log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * (dX_46_v_m * dX_46_v_m))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, fmaf(((dY_46_v * dY_46_v) * floorf(h)), floorf(h), (((floorf(d) * dY_46_w) * floorf(d)) * dY_46_w))))));
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, dX_46_w, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt(fmax(fma(Float32(floor(d) * floor(d)), Float32(dX_46_w * dX_46_w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v_m * dX_46_v_m))), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, fma(Float32(Float32(dY_46_v * dY_46_v) * floor(h)), floor(h), Float32(Float32(Float32(floor(d) * dY_46_w) * floor(d)) * dY_46_w)))))) end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , dX.w \cdot dX.w, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v\_m \cdot dX.v\_m\right)\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, \mathsf{fma}\left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right)\right)\right)}\right)
\end{array}
Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Applied rewrites60.2%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dY.w (floor d))))
(log2
(sqrt
(fmax
(fma
(* (floor d) (floor d))
(* dX.w dX.w)
(* (* (floor h) (floor h)) (* dX.v_m dX.v_m)))
(fma (* (* (floor w) dY.u) (floor w)) dY.u (* t_0 t_0)))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = dY_46_w * floorf(d);
return log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * (dX_46_v_m * dX_46_v_m))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, (t_0 * t_0)))));
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(dY_46_w * floor(d)) return log2(sqrt(fmax(fma(Float32(floor(d) * floor(d)), Float32(dX_46_w * dX_46_w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v_m * dX_46_v_m))), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, Float32(t_0 * t_0))))) end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := dY.w \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , dX.w \cdot dX.w, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v\_m \cdot dX.v\_m\right)\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, t\_0 \cdot t\_0\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3252.6
Applied rewrites52.6%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lift-*.f32N/A
pow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites52.6%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* (* (floor h) (floor h)) (* dX.v_m dX.v_m)))
(fma (* dY.w dY.w) t_0 (* (* dY.u dY.u) (* (floor w) (floor w)))))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * (dX_46_v_m * dX_46_v_m))), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v_m * dX_46_v_m))), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))) end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v\_m \cdot dX.v\_m\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3252.6
Applied rewrites52.6%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* (* (floor h) (floor h)) (* dX.v_m dX.v_m)))
(* (/ (* (* dY.w dY.w) t_0) (* dY.u dY.u)) (* dY.u dY.u)))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * (dX_46_v_m * dX_46_v_m))), ((((dY_46_w * dY_46_w) * t_0) / (dY_46_u * dY_46_u)) * (dY_46_u * dY_46_u)))));
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v_m * dX_46_v_m))), Float32(Float32(Float32(Float32(dY_46_w * dY_46_w) * t_0) / Float32(dY_46_u * dY_46_u)) * Float32(dY_46_u * dY_46_u))))) end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v\_m \cdot dX.v\_m\right)\right), \frac{\left(dY.w \cdot dY.w\right) \cdot t\_0}{dY.u \cdot dY.u} \cdot \left(dY.u \cdot dY.u\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Taylor expanded in dY.u around inf
Applied rewrites46.2%
Taylor expanded in dY.v around inf
lower-/.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f3237.1
Applied rewrites37.1%
Taylor expanded in dY.w around inf
pow2N/A
lift-floor.f32N/A
pow2N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
*-commutativeN/A
lift-floor.f32N/A
pow2N/A
lower-/.f32N/A
Applied rewrites37.0%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) (floor h))))
(log2
(sqrt
(fmax
(fma (* (floor d) (floor d)) (* dX.w dX.w) (* t_0 (* dX.v_m dX.v_m)))
(* (/ (* (* dY.v dY.v) t_0) (* dY.u dY.u)) (* dY.u dY.u)))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * floorf(h);
return log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w * dX_46_w), (t_0 * (dX_46_v_m * dX_46_v_m))), ((((dY_46_v * dY_46_v) * t_0) / (dY_46_u * dY_46_u)) * (dY_46_u * dY_46_u)))));
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * floor(h)) return log2(sqrt(fmax(fma(Float32(floor(d) * floor(d)), Float32(dX_46_w * dX_46_w), Float32(t_0 * Float32(dX_46_v_m * dX_46_v_m))), Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * t_0) / Float32(dY_46_u * dY_46_u)) * Float32(dY_46_u * dY_46_u))))) end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , dX.w \cdot dX.w, t\_0 \cdot \left(dX.v\_m \cdot dX.v\_m\right)\right), \frac{\left(dY.v \cdot dY.v\right) \cdot t\_0}{dY.u \cdot dY.u} \cdot \left(dY.u \cdot dY.u\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Taylor expanded in dY.u around inf
Applied rewrites46.2%
Taylor expanded in dY.v around inf
lower-/.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f3237.1
Applied rewrites37.1%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h d dX.u dX.v_m dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) (floor h))))
(log2
(sqrt
(fmax
(fma (* (floor d) (floor d)) (* dX.w dX.w) (* t_0 (* dX.v_m dX.v_m)))
(* (* (* dY.v dY.v) (/ t_0 (* dY.u dY.u))) (* dY.u dY.u)))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float d, float dX_46_u, float dX_46_v_m, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * floorf(h);
return log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w * dX_46_w), (t_0 * (dX_46_v_m * dX_46_v_m))), (((dY_46_v * dY_46_v) * (t_0 / (dY_46_u * dY_46_u))) * (dY_46_u * dY_46_u)))));
}
dX.v_m = abs(dX_46_v) function code(w, h, d, dX_46_u, dX_46_v_m, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * floor(h)) return log2(sqrt(fmax(fma(Float32(floor(d) * floor(d)), Float32(dX_46_w * dX_46_w), Float32(t_0 * Float32(dX_46_v_m * dX_46_v_m))), Float32(Float32(Float32(dY_46_v * dY_46_v) * Float32(t_0 / Float32(dY_46_u * dY_46_u))) * Float32(dY_46_u * dY_46_u))))) end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , dX.w \cdot dX.w, t\_0 \cdot \left(dX.v\_m \cdot dX.v\_m\right)\right), \left(\left(dY.v \cdot dY.v\right) \cdot \frac{t\_0}{dY.u \cdot dY.u}\right) \cdot \left(dY.u \cdot dY.u\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.5%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.2
Applied rewrites60.2%
Taylor expanded in dY.u around inf
Applied rewrites46.2%
Taylor expanded in dY.v around inf
lower-/.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f3237.1
Applied rewrites37.1%
lift-*.f32N/A
lift-/.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
*-commutativeN/A
associate-/l*N/A
Applied rewrites35.7%
herbie shell --seed 2025131
(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)))))))