
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
Herbie found 8 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
(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) dY.w))
(t_6 (* (floor d) dX.w))
(t_7
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_4 t_4)) (* t_6 t_6))
(+ t_3 (* t_5 t_5)))))))
(if (<= t_7 100.0)
t_7
(log2
(sqrt
(fmax
(fma
(* (floor d) (floor d))
(* dX.w dX.w)
(* (* (floor h) (floor h)) (* dX.v dX.v)))
(+ t_3 (exp (fma 2.0 (log (- (floor d))) (* (log dY.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) {
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) * dY_46_w;
float t_6 = floorf(d) * dX_46_w;
float t_7 = log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_4 * t_4)) + (t_6 * t_6)), (t_3 + (t_5 * t_5)))));
float tmp;
if (t_7 <= 100.0f) {
tmp = t_7;
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), (t_3 + expf(fmaf(2.0f, logf(-floorf(d)), (logf(dY_46_w) * 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) * dY_46_w) t_6 = Float32(floor(d) * dX_46_w) t_7 = 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))))) tmp = Float32(0.0) if (t_7 <= Float32(100.0)) tmp = t_7; else tmp = 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 * dX_46_v))), Float32(t_3 + exp(fma(Float32(2.0), log(Float32(-floor(d))), Float32(log(dY_46_w) * Float32(2.0)))))))); end return tmp end
\begin{array}{l}
\\
\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 dY.w\\
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\_4 \cdot t\_4\right) + t\_6 \cdot t\_6, t\_3 + t\_5 \cdot t\_5\right)}\right)\\
\mathbf{if}\;t\_7 \leq 100:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;\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 \cdot dX.v\right)\right), t\_3 + e^{\mathsf{fma}\left(2, \log \left(-\left\lfloor d\right\rfloor \right), \log dY.w \cdot 2\right)}\right)}\right)\\
\end{array}
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 100Initial program 68.0%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.0%
Taylor expanded in dX.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-*.f3260.7
Applied rewrites60.7%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3248.4
Applied rewrites48.4%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
lift-floor.f32N/A
log-prodN/A
lift-floor.f32N/A
+-commutativeN/A
distribute-lft-inN/A
log-pow-revN/A
lift-floor.f32N/A
log-pow-revN/A
pow2N/A
lift-*.f32N/A
log-prodN/A
*-commutativeN/A
log-prodN/A
Applied rewrites35.7%
(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}
Initial program 68.0%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor d) dX.w))
(t_2 (* (floor d) dY.w))
(t_3 (* (floor d) (floor d)))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor w) dY.u))
(t_6 (* (floor w) dX.u)))
(if (<= dX.v 500000000.0)
(log2
(sqrt
(fmax
(fma t_3 (* dX.w dX.w) (* (* (floor w) (floor w)) (* dX.u dX.u)))
(+ (+ (* t_5 t_5) (* t_0 t_0)) (* t_2 t_2)))))
(log2
(sqrt
(fmax
(+ (+ (* t_6 t_6) (* t_4 t_4)) (* t_1 t_1))
(fma (* dY.w dY.w) t_3 (* (* dY.v dY.v) (* (floor h) (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) {
float t_0 = floorf(h) * dY_46_v;
float t_1 = floorf(d) * dX_46_w;
float t_2 = floorf(d) * dY_46_w;
float t_3 = floorf(d) * floorf(d);
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(w) * dY_46_u;
float t_6 = floorf(w) * dX_46_u;
float tmp;
if (dX_46_v <= 500000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_3, (dX_46_w * dX_46_w), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u))), (((t_5 * t_5) + (t_0 * t_0)) + (t_2 * t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf((((t_6 * t_6) + (t_4 * t_4)) + (t_1 * t_1)), fmaf((dY_46_w * dY_46_w), t_3, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h)))))));
}
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) * dY_46_v) t_1 = Float32(floor(d) * dX_46_w) t_2 = Float32(floor(d) * dY_46_w) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(w) * dY_46_u) t_6 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (dX_46_v <= Float32(500000000.0)) 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_5 * t_5) + Float32(t_0 * t_0)) + Float32(t_2 * t_2))))); else tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_6 * t_6) + Float32(t_4 * t_4)) + Float32(t_1 * t_1)), fma(Float32(dY_46_w * dY_46_w), t_3, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_2 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_6 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;dX.v \leq 500000000:\\
\;\;\;\;\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\_5 \cdot t\_5 + t\_0 \cdot t\_0\right) + t\_2 \cdot t\_2\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_6 \cdot t\_6 + t\_4 \cdot t\_4\right) + t\_1 \cdot t\_1, \mathsf{fma}\left(dY.w \cdot dY.w, t\_3, \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.v < 5e8Initial program 68.0%
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-*.f3260.9
Applied rewrites60.9%
if 5e8 < dX.v Initial program 68.0%
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.f3261.3
Applied rewrites61.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dY.w))
(t_1 (* dX.v (floor h)))
(t_2 (* (floor d) (floor d)))
(t_3 (* (floor w) dY.u))
(t_4 (* (floor w) dX.u))
(t_5 (* (floor h) dY.v)))
(if (<= dX.w 3000000000.0)
(log2
(sqrt
(fmax
(fma (* t_4 (floor w)) dX.u (* (* t_1 dX.v) (floor h)))
(+ (+ (* t_3 t_3) (* t_5 t_5)) (* t_0 t_0)))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_2 (fma t_4 t_4 (* t_1 t_1)))
(fma (* dY.w dY.w) t_2 (* (* dY.u dY.u) (* (floor w) (floor w))))))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * dY_46_w;
float t_1 = dX_46_v * floorf(h);
float t_2 = floorf(d) * floorf(d);
float t_3 = floorf(w) * dY_46_u;
float t_4 = floorf(w) * dX_46_u;
float t_5 = floorf(h) * dY_46_v;
float tmp;
if (dX_46_w <= 3000000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_4 * floorf(w)), dX_46_u, ((t_1 * dX_46_v) * floorf(h))), (((t_3 * t_3) + (t_5 * t_5)) + (t_0 * t_0)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_2, fmaf(t_4, t_4, (t_1 * t_1))), fmaf((dY_46_w * dY_46_w), t_2, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
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(d) * dY_46_w) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(floor(d) * floor(d)) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (dX_46_w <= Float32(3000000000.0)) tmp = log2(sqrt(fmax(fma(Float32(t_4 * floor(w)), dX_46_u, Float32(Float32(t_1 * dX_46_v) * floor(h))), Float32(Float32(Float32(t_3 * t_3) + Float32(t_5 * t_5)) + Float32(t_0 * t_0))))); else tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_2, fma(t_4, t_4, Float32(t_1 * t_1))), fma(Float32(dY_46_w * dY_46_w), t_2, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;dX.w \leq 3000000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot \left\lfloor w\right\rfloor , dX.u, \left(t\_1 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \left(t\_3 \cdot t\_3 + t\_5 \cdot t\_5\right) + t\_0 \cdot t\_0\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_2, \mathsf{fma}\left(t\_4, t\_4, t\_1 \cdot t\_1\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_2, \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.w < 3e9Initial program 68.0%
Applied rewrites60.9%
Taylor expanded in dX.w around 0
Applied rewrites61.1%
if 3e9 < dX.w Initial program 68.0%
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.f3260.4
Applied rewrites60.4%
Applied rewrites60.4%
(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 (* t_0 t_0))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor d) (floor d)))
(t_4 (* (* dY.u dY.u) (* (floor w) (floor w)))))
(if (<= dX.u 50000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_3 t_1)
(fma
(* (* (floor d) dY.w) (floor d))
dY.w
(fma (* (* (floor h) (floor h)) dY.v) dY.v t_4)))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_3 (fma t_2 t_2 t_1))
(fma (* dY.w dY.w) t_3 t_4)))))))
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 = t_0 * t_0;
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(d) * floorf(d);
float t_4 = (dY_46_u * dY_46_u) * (floorf(w) * floorf(w));
float tmp;
if (dX_46_u <= 50000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_3, t_1), fmaf(((floorf(d) * dY_46_w) * floorf(d)), dY_46_w, fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, t_4)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_3, fmaf(t_2, t_2, t_1)), fmaf((dY_46_w * dY_46_w), t_3, t_4))));
}
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(t_0 * t_0) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))) tmp = Float32(0.0) if (dX_46_u <= Float32(50000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_3, t_1), fma(Float32(Float32(floor(d) * dY_46_w) * floor(d)), dY_46_w, fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, t_4))))); else tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_3, fma(t_2, t_2, t_1)), fma(Float32(dY_46_w * dY_46_w), t_3, t_4)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := t\_0 \cdot t\_0\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{if}\;dX.u \leq 50000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_3, t\_1\right), \mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor , dY.w, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, t\_4\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_3, \mathsf{fma}\left(t\_2, t\_2, t\_1\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_3, t\_4\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 5e4Initial program 68.0%
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-*.f3260.7
Applied rewrites60.7%
Applied rewrites60.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
pow2N/A
lift-floor.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3260.7
Applied rewrites60.7%
if 5e4 < dX.u Initial program 68.0%
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.f3260.4
Applied rewrites60.4%
Applied rewrites60.4%
(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))
(t_1 (* dX.v (floor h)))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor d) (floor d)))
(t_4 (* (floor d) dX.w)))
(if (<= dX.u 50000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_3 (* t_1 t_1))
(fma
(* (* (floor d) dY.w) (floor d))
dY.w
(fma
(* (* (floor h) (floor h)) dY.v)
dY.v
(* (* dY.u dY.u) (* (floor w) (floor w))))))))
(log2
(sqrt
(fmax
(+ (+ (* t_2 t_2) (* t_0 t_0)) (* t_4 t_4))
(* (* (fabs dY.w) t_3) dY.w)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = dX_46_v * floorf(h);
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(d) * floorf(d);
float t_4 = floorf(d) * dX_46_w;
float tmp;
if (dX_46_u <= 50000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_3, (t_1 * t_1)), fmaf(((floorf(d) * dY_46_w) * floorf(d)), dY_46_w, fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w))))))));
} else {
tmp = log2f(sqrtf(fmaxf((((t_2 * t_2) + (t_0 * t_0)) + (t_4 * t_4)), ((fabsf(dY_46_w) * t_3) * dY_46_w))));
}
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) * dX_46_v) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(floor(d) * dX_46_w) tmp = Float32(0.0) if (dX_46_u <= Float32(50000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_3, Float32(t_1 * t_1)), fma(Float32(Float32(floor(d) * dY_46_w) * floor(d)), dY_46_w, fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w)))))))); else tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) + Float32(t_4 * t_4)), Float32(Float32(abs(dY_46_w) * t_3) * dY_46_w)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
\mathbf{if}\;dX.u \leq 50000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_3, t\_1 \cdot t\_1\right), \mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor , dY.w, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right) + t\_4 \cdot t\_4, \left(\left|dY.w\right| \cdot t\_3\right) \cdot dY.w\right)}\right)\\
\end{array}
\end{array}
if dX.u < 5e4Initial program 68.0%
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-*.f3260.7
Applied rewrites60.7%
Applied rewrites60.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
pow2N/A
lift-floor.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3260.7
Applied rewrites60.7%
if 5e4 < dX.u Initial program 68.0%
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.f3260.4
Applied rewrites60.4%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3253.3
Applied rewrites53.3%
lift-fma.f32N/A
Applied rewrites56.6%
Taylor expanded in dY.u around 0
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
fabs-mulN/A
lift-*.f32N/A
fabs-sqrN/A
lift-*.f32N/A
lower-*.f32N/A
lower-fabs.f3250.0
Applied rewrites50.0%
(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))
(t_1 (* dX.v (floor h)))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor d) (floor d)))
(t_4 (* (floor d) dX.w)))
(if (<= dX.u 50000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_3 (* t_1 t_1))
(fma
(* (* (floor d) dY.w) (floor d))
dY.w
(fma
(* (* dY.v (floor h)) (floor h))
dY.v
(* (* dY.u dY.u) (* (floor w) (floor w))))))))
(log2
(sqrt
(fmax
(+ (+ (* t_2 t_2) (* t_0 t_0)) (* t_4 t_4))
(* (* (fabs dY.w) t_3) dY.w)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = dX_46_v * floorf(h);
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(d) * floorf(d);
float t_4 = floorf(d) * dX_46_w;
float tmp;
if (dX_46_u <= 50000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_3, (t_1 * t_1)), fmaf(((floorf(d) * dY_46_w) * floorf(d)), dY_46_w, fmaf(((dY_46_v * floorf(h)) * floorf(h)), dY_46_v, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w))))))));
} else {
tmp = log2f(sqrtf(fmaxf((((t_2 * t_2) + (t_0 * t_0)) + (t_4 * t_4)), ((fabsf(dY_46_w) * t_3) * dY_46_w))));
}
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) * dX_46_v) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(floor(d) * dX_46_w) tmp = Float32(0.0) if (dX_46_u <= Float32(50000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_3, Float32(t_1 * t_1)), fma(Float32(Float32(floor(d) * dY_46_w) * floor(d)), dY_46_w, fma(Float32(Float32(dY_46_v * floor(h)) * floor(h)), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w)))))))); else tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) + Float32(t_4 * t_4)), Float32(Float32(abs(dY_46_w) * t_3) * dY_46_w)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
\mathbf{if}\;dX.u \leq 50000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_3, t\_1 \cdot t\_1\right), \mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor , dY.w, \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dY.v, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right) + t\_4 \cdot t\_4, \left(\left|dY.w\right| \cdot t\_3\right) \cdot dY.w\right)}\right)\\
\end{array}
\end{array}
if dX.u < 5e4Initial program 68.0%
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-*.f3260.7
Applied rewrites60.7%
Applied rewrites60.7%
if 5e4 < dX.u Initial program 68.0%
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.f3260.4
Applied rewrites60.4%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3253.3
Applied rewrites53.3%
lift-fma.f32N/A
Applied rewrites56.6%
Taylor expanded in dY.u around 0
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
fabs-mulN/A
lift-*.f32N/A
fabs-sqrN/A
lift-*.f32N/A
lower-*.f32N/A
lower-fabs.f3250.0
Applied rewrites50.0%
(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))
(t_1 (* (floor d) dX.w))
(t_2 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_2 t_2) (* t_0 t_0)) (* t_1 t_1))
(* (* (fabs dY.w) (* (floor d) (floor d))) dY.w))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(d) * dX_46_w;
float t_2 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_2 * t_2) + (t_0 * t_0)) + (t_1 * t_1)), ((fabsf(dY_46_w) * (floorf(d) * floorf(d))) * dY_46_w))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(d) * dX_46_w) t_2 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) + Float32(t_1 * t_1)), Float32(Float32(abs(dY_46_w) * Float32(floor(d) * floor(d))) * dY_46_w)))) 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(h) * dX_46_v; t_1 = floor(d) * dX_46_w; t_2 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_2 * t_2) + (t_0 * t_0)) + (t_1 * t_1)), ((abs(dY_46_w) * (floor(d) * floor(d))) * dY_46_w)))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right) + t\_1 \cdot t\_1, \left(\left|dY.w\right| \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right) \cdot dY.w\right)}\right)
\end{array}
\end{array}
Initial program 68.0%
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.f3260.4
Applied rewrites60.4%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3253.3
Applied rewrites53.3%
lift-fma.f32N/A
Applied rewrites56.6%
Taylor expanded in dY.u around 0
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
fabs-mulN/A
lift-*.f32N/A
fabs-sqrN/A
lift-*.f32N/A
lower-*.f32N/A
lower-fabs.f3250.0
Applied rewrites50.0%
herbie shell --seed 2025142
(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)))))))