
(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 9 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) 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 68.4%
(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) dY.w))
(t_2 (* (floor d) dX.w))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor w) (floor w)))
(t_5 (* (floor w) dY.u))
(t_6 (* (floor h) dX.v))
(t_7 (* (floor h) (floor h))))
(if (<= dY.w 1000000.0)
(log2
(sqrt
(fmax
(+ (+ (* t_3 t_3) (* t_6 t_6)) (* t_2 t_2))
(fma (* t_4 dY.u) dY.u (* (* dY.v dY.v) t_7)))))
(log2
(sqrt
(fmax
(fma (* t_7 dX.v) dX.v (* t_4 (* dX.u dX.u)))
(+ (+ (* t_5 t_5) (* t_0 t_0)) (* t_1 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 = floorf(h) * dY_46_v;
float t_1 = floorf(d) * dY_46_w;
float t_2 = floorf(d) * dX_46_w;
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(w) * floorf(w);
float t_5 = floorf(w) * dY_46_u;
float t_6 = floorf(h) * dX_46_v;
float t_7 = floorf(h) * floorf(h);
float tmp;
if (dY_46_w <= 1000000.0f) {
tmp = log2f(sqrtf(fmaxf((((t_3 * t_3) + (t_6 * t_6)) + (t_2 * t_2)), fmaf((t_4 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_7)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_7 * dX_46_v), dX_46_v, (t_4 * (dX_46_u * dX_46_u))), (((t_5 * t_5) + (t_0 * t_0)) + (t_1 * 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(floor(h) * dY_46_v) t_1 = Float32(floor(d) * dY_46_w) t_2 = Float32(floor(d) * dX_46_w) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(w) * floor(w)) t_5 = Float32(floor(w) * dY_46_u) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dY_46_w <= Float32(1000000.0)) tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_3 * t_3) + Float32(t_6 * t_6)) + Float32(t_2 * t_2)), fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_7))))); else tmp = log2(sqrt(fmax(fma(Float32(t_7 * dX_46_v), dX_46_v, Float32(t_4 * Float32(dX_46_u * dX_46_u))), Float32(Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)) + Float32(t_1 * t_1))))); 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 dY.w\\
t_2 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_5 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dY.w \leq 1000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_3 \cdot t\_3 + t\_6 \cdot t\_6\right) + t\_2 \cdot t\_2, \mathsf{fma}\left(t\_4 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_7\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_7 \cdot dX.v, dX.v, t\_4 \cdot \left(dX.u \cdot dX.u\right)\right), \left(t\_5 \cdot t\_5 + t\_0 \cdot t\_0\right) + t\_1 \cdot t\_1\right)}\right)\\
\end{array}
\end{array}
if dY.w < 1e6Initial program 68.4%
Applied rewrites44.7%
Taylor expanded in dY.w around 0
Applied rewrites61.2%
if 1e6 < dY.w Initial program 68.4%
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-*.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
pow2N/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3254.0
Applied rewrites54.0%
Taylor expanded in dX.w around 0
+-commutativeN/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites61.1%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) (floor w)))
(t_4 (* (floor d) dY.w)))
(if (<= dX.v 1020000.0)
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* t_3 (* dX.u dX.u)))
(+ (+ (* t_2 t_2) (* t_1 t_1)) (* t_4 t_4)))))
(log2
(sqrt
(fmax
(fma
(* (* (floor d) dX.w) (floor d))
dX.w
(fma
(* (* (floor w) dX.u) (floor w))
dX.u
(* (* (* (floor h) dX.v) (floor h)) dX.v)))
(fma (* dY.w dY.w) t_0 (* (* dY.u dY.u) 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(d) * floorf(d);
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(w) * floorf(w);
float t_4 = floorf(d) * dY_46_w;
float tmp;
if (dX_46_v <= 1020000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_3 * (dX_46_u * dX_46_u))), (((t_2 * t_2) + (t_1 * t_1)) + (t_4 * t_4)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w) * floorf(d)), dX_46_w, fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, (((floorf(h) * dX_46_v) * floorf(h)) * dX_46_v))), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_u * dY_46_u) * t_3)))));
}
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) * floor(d)) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(w) * floor(w)) t_4 = Float32(floor(d) * dY_46_w) tmp = Float32(0.0) if (dX_46_v <= Float32(1020000.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_3 * Float32(dX_46_u * dX_46_u))), Float32(Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) + Float32(t_4 * t_4))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w) * floor(d)), dX_46_w, fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, Float32(Float32(Float32(floor(h) * dX_46_v) * floor(h)) * dX_46_v))), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_3))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\\
\mathbf{if}\;dX.v \leq 1020000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_3 \cdot \left(dX.u \cdot dX.u\right)\right), \left(t\_2 \cdot t\_2 + t\_1 \cdot t\_1\right) + t\_4 \cdot t\_4\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor , dX.w, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_3\right)\right)}\right)\\
\end{array}
\end{array}
if dX.v < 1.02e6Initial program 68.4%
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-*.f3261.0
Applied rewrites61.0%
if 1.02e6 < dX.v Initial program 68.4%
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.f3261.5
Applied rewrites61.5%
Applied rewrites61.5%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor w) (floor w)))
(t_2 (* (floor h) (floor h))))
(if (<= dY.v 10000000000.0)
(log2
(sqrt
(fmax
(fma
(* (* (floor d) dX.w) (floor d))
dX.w
(fma
(* (* (floor w) dX.u) (floor w))
dX.u
(* (* (* (floor h) dX.v) (floor h)) dX.v)))
(fma (* dY.w dY.w) t_0 (* (* dY.u dY.u) t_1)))))
(log2
(sqrt
(fmax
(fma t_2 (* dX.v dX.v) (* t_0 (* dX.w dX.w)))
(fma
(* t_1 dY.u)
dY.u
(fma (* dY.w dY.w) t_0 (* (* dY.v dY.v) 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 = floorf(d) * floorf(d);
float t_1 = floorf(w) * floorf(w);
float t_2 = floorf(h) * floorf(h);
float tmp;
if (dY_46_v <= 10000000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w) * floorf(d)), dX_46_w, fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, (((floorf(h) * dX_46_v) * floorf(h)) * dX_46_v))), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_u * dY_46_u) * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_2, (dX_46_v * dX_46_v), (t_0 * (dX_46_w * dX_46_w))), fmaf((t_1 * dY_46_u), dY_46_u, fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_v * dY_46_v) * 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(floor(d) * floor(d)) t_1 = Float32(floor(w) * floor(w)) t_2 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dY_46_v <= Float32(10000000000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w) * floor(d)), dX_46_w, fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, Float32(Float32(Float32(floor(h) * dX_46_v) * floor(h)) * dX_46_v))), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_1))))); else tmp = log2(sqrt(fmax(fma(t_2, Float32(dX_46_v * dX_46_v), Float32(t_0 * Float32(dX_46_w * dX_46_w))), fma(Float32(t_1 * dY_46_u), dY_46_u, fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_v * dY_46_v) * t_2)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dY.v \leq 10000000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor , dX.w, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, dX.v \cdot dX.v, t\_0 \cdot \left(dX.w \cdot dX.w\right)\right), \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.v \cdot dY.v\right) \cdot t\_2\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.v < 1e10Initial program 68.4%
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.f3261.5
Applied rewrites61.5%
Applied rewrites61.5%
if 1e10 < dY.v Initial program 68.4%
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-*.f3261.6
Applied rewrites61.6%
Applied rewrites61.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor d) (floor d)))
(t_2 (fma (* dY.w dY.w) t_1 (* (* dY.v dY.v) t_0)))
(t_3 (* (floor w) (floor w))))
(if (<= dX.u 399999991808.0)
(log2
(sqrt
(fmax
(fma t_0 (* dX.v dX.v) (* t_1 (* dX.w dX.w)))
(fma (* t_3 dY.u) dY.u t_2))))
(log2 (sqrt (fmax (fma t_1 (* dX.w dX.w) (* t_3 (* dX.u dX.u))) 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 = floorf(h) * floorf(h);
float t_1 = floorf(d) * floorf(d);
float t_2 = fmaf((dY_46_w * dY_46_w), t_1, ((dY_46_v * dY_46_v) * t_0));
float t_3 = floorf(w) * floorf(w);
float tmp;
if (dX_46_u <= 399999991808.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_v * dX_46_v), (t_1 * (dX_46_w * dX_46_w))), fmaf((t_3 * dY_46_u), dY_46_u, t_2))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, (dX_46_w * dX_46_w), (t_3 * (dX_46_u * dX_46_u))), 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(floor(h) * floor(h)) t_1 = Float32(floor(d) * floor(d)) t_2 = fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (dX_46_u <= Float32(399999991808.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_v * dX_46_v), Float32(t_1 * Float32(dX_46_w * dX_46_w))), fma(Float32(t_3 * dY_46_u), dY_46_u, t_2)))); else tmp = log2(sqrt(fmax(fma(t_1, Float32(dX_46_w * dX_46_w), Float32(t_3 * Float32(dX_46_u * dX_46_u))), t_2))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;dX.u \leq 399999991808:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.v \cdot dX.v, t\_1 \cdot \left(dX.w \cdot dX.w\right)\right), \mathsf{fma}\left(t\_3 \cdot dY.u, dY.u, t\_2\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, dX.w \cdot dX.w, t\_3 \cdot \left(dX.u \cdot dX.u\right)\right), t\_2\right)}\right)\\
\end{array}
\end{array}
if dX.u < 399999992000Initial program 68.4%
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-*.f3261.6
Applied rewrites61.6%
Applied rewrites61.6%
if 399999992000 < dX.u Initial program 68.4%
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-*.f3261.0
Applied rewrites61.0%
Taylor expanded in dY.u around 0
+-commutativeN/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3253.0
Applied rewrites53.0%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor d) dY.w))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v)))
(if (<= dX.v 1020000.0)
(log2
(sqrt
(fmax
(* t_0 (* dX.w dX.w))
(+ (+ (* t_2 t_2) (* t_3 t_3)) (* t_1 t_1)))))
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (exp (* (log (* (floor h) dX.v)) 2.0)))
(fma (* dY.w dY.w) t_0 (* (* 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) * floorf(d);
float t_1 = floorf(d) * dY_46_w;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float tmp;
if (dX_46_v <= 1020000.0f) {
tmp = log2f(sqrtf(fmaxf((t_0 * (dX_46_w * dX_46_w)), (((t_2 * t_2) + (t_3 * t_3)) + (t_1 * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), expf((logf((floorf(h) * dX_46_v)) * 2.0f))), fmaf((dY_46_w * dY_46_w), t_0, ((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) * floor(d)) t_1 = Float32(floor(d) * dY_46_w) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (dX_46_v <= Float32(1020000.0)) tmp = log2(sqrt(fmax(Float32(t_0 * Float32(dX_46_w * dX_46_w)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) + Float32(t_1 * t_1))))); else tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), exp(Float32(log(Float32(floor(h) * dX_46_v)) * Float32(2.0)))), fma(Float32(dY_46_w * dY_46_w), t_0, 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 \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;dX.v \leq 1020000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_0 \cdot \left(dX.w \cdot dX.w\right), \left(t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right) + t\_1 \cdot t\_1\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, e^{\log \left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot 2}\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}
if dX.v < 1.02e6Initial program 68.4%
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-*.f3261.0
Applied rewrites61.0%
Taylor expanded in dX.u around 0
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3254.3
Applied rewrites54.3%
if 1.02e6 < dX.v Initial program 68.4%
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-*.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
pow2N/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3254.0
Applied rewrites54.0%
Taylor expanded in dY.v around 0
+-commutativeN/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3245.2
Applied rewrites45.2%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor w) (floor w)))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor d) dY.w)))
(if (<= dY.v 40.0)
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* t_1 (* dX.u dX.u)))
(fma (* dY.w dY.w) t_0 (* (* dY.u dY.u) t_1)))))
(log2
(sqrt
(fmax
(* t_0 (* dX.w dX.w))
(+ (+ (* t_2 t_2) (* t_3 t_3)) (* t_4 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 = floorf(d) * floorf(d);
float t_1 = floorf(w) * floorf(w);
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(d) * dY_46_w;
float tmp;
if (dY_46_v <= 40.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_1 * (dX_46_u * dX_46_u))), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_u * dY_46_u) * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf((t_0 * (dX_46_w * dX_46_w)), (((t_2 * t_2) + (t_3 * t_3)) + (t_4 * 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(floor(d) * floor(d)) t_1 = Float32(floor(w) * floor(w)) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(d) * dY_46_w) tmp = Float32(0.0) if (dY_46_v <= Float32(40.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_1 * Float32(dX_46_u * dX_46_u))), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_1))))); else tmp = log2(sqrt(fmax(Float32(t_0 * Float32(dX_46_w * dX_46_w)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) + Float32(t_4 * t_4))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\\
\mathbf{if}\;dY.v \leq 40:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_1 \cdot \left(dX.u \cdot dX.u\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_0 \cdot \left(dX.w \cdot dX.w\right), \left(t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right) + t\_4 \cdot t\_4\right)}\right)\\
\end{array}
\end{array}
if dY.v < 40Initial program 68.4%
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-*.f3261.0
Applied rewrites61.0%
Taylor expanded in dY.v around 0
+-commutativeN/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3253.5
Applied rewrites53.5%
if 40 < dY.v Initial program 68.4%
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-*.f3261.0
Applied rewrites61.0%
Taylor expanded in dX.u around 0
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3254.3
Applied rewrites54.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma t_0 (* dX.w dX.w) (* t_1 (* dX.u dX.u)))))
(if (<= dY.v 500.0)
(log2 (sqrt (fmax t_2 (fma (* dY.w dY.w) t_0 (* (* dY.u dY.u) t_1)))))
(log2
(sqrt
(fmax
t_2
(fma (* dY.w dY.w) t_0 (* (* 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(d) * floorf(d);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf(t_0, (dX_46_w * dX_46_w), (t_1 * (dX_46_u * dX_46_u)));
float tmp;
if (dY_46_v <= 500.0f) {
tmp = log2f(sqrtf(fmaxf(t_2, fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_u * dY_46_u) * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf(t_2, fmaf((dY_46_w * dY_46_w), t_0, ((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(d) * floor(d)) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_1 * Float32(dX_46_u * dX_46_u))) tmp = Float32(0.0) if (dY_46_v <= Float32(500.0)) tmp = log2(sqrt(fmax(t_2, fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_1))))); else tmp = log2(sqrt(fmax(t_2, fma(Float32(dY_46_w * dY_46_w), t_0, 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 d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_1 \cdot \left(dX.u \cdot dX.u\right)\right)\\
\mathbf{if}\;dY.v \leq 500:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \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 dY.v < 500Initial program 68.4%
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-*.f3261.0
Applied rewrites61.0%
Taylor expanded in dY.v around 0
+-commutativeN/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3253.5
Applied rewrites53.5%
if 500 < dY.v Initial program 68.4%
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-*.f3261.0
Applied rewrites61.0%
Taylor expanded in dY.u around 0
+-commutativeN/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3253.0
Applied rewrites53.0%
(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 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma t_1 (* dX.w dX.w) (* t_0 (* dX.u dX.u)))
(fma (* dY.w dY.w) t_1 (* (* dY.u dY.u) 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 = floorf(w) * floorf(w);
float t_1 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf(t_1, (dX_46_w * dX_46_w), (t_0 * (dX_46_u * dX_46_u))), fmaf((dY_46_w * dY_46_w), t_1, ((dY_46_u * dY_46_u) * t_0)))));
}
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(floor(d) * floor(d)) return log2(sqrt(fmax(fma(t_1, Float32(dX_46_w * dX_46_w), Float32(t_0 * Float32(dX_46_u * dX_46_u))), fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_u * dY_46_u) * t_0))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, dX.w \cdot dX.w, t\_0 \cdot \left(dX.u \cdot dX.u\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}\right)
\end{array}
\end{array}
Initial program 68.4%
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-*.f3261.0
Applied rewrites61.0%
Taylor expanded in dY.v around 0
+-commutativeN/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3253.5
Applied rewrites53.5%
herbie shell --seed 2025136
(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)))))))