
(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 7 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 68.2%
(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)))
(t_2 (* (floor h) (floor h)))
(t_3 (* t_2 (* dX.v dX.v))))
(if (<= dY.v 1400000.0)
(log2
(sqrt
(fmax
(fma (* t_0 dX.u) dX.u (fma (* dX.w dX.w) t_1 t_3))
(fma (* dY.w dY.w) t_1 (* (* dY.u dY.u) t_0)))))
(log2
(sqrt
(fmax
(fma t_1 (* dX.w dX.w) t_3)
(fma
(* (* dY.w (floor d)) (floor d))
dY.w
(fma
(* (* dY.u dY.u) (floor w))
(floor w)
(* (* 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(w) * floorf(w);
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(h) * floorf(h);
float t_3 = t_2 * (dX_46_v * dX_46_v);
float tmp;
if (dY_46_v <= 1400000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, fmaf((dX_46_w * dX_46_w), t_1, t_3)), fmaf((dY_46_w * dY_46_w), t_1, ((dY_46_u * dY_46_u) * t_0)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, (dX_46_w * dX_46_w), t_3), fmaf(((dY_46_w * floorf(d)) * floorf(d)), dY_46_w, fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), ((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(w) * floor(w)) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(t_2 * Float32(dX_46_v * dX_46_v)) tmp = Float32(0.0) if (dY_46_v <= Float32(1400000.0)) tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, fma(Float32(dX_46_w * dX_46_w), t_1, t_3)), fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_u * dY_46_u) * t_0))))); else tmp = log2(sqrt(fmax(fma(t_1, Float32(dX_46_w * dX_46_w), t_3), fma(Float32(Float32(dY_46_w * floor(d)) * floor(d)), dY_46_w, fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(Float32(dY_46_v * dY_46_v) * t_2)))))); end return tmp 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 \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot \left(dX.v \cdot dX.v\right)\\
\mathbf{if}\;dY.v \leq 1400000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, \mathsf{fma}\left(dX.w \cdot dX.w, t\_1, t\_3\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)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, dX.w \cdot dX.w, t\_3\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dY.w, \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot t\_2\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.v < 1.4e6Initial program 68.2%
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.f3261.2
Applied rewrites61.2%
Applied rewrites61.2%
if 1.4e6 < dY.v Initial program 68.2%
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-*.f3261.4
Applied rewrites61.4%
Applied rewrites61.4%
(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) (floor h)))
(t_2
(fma
(* (* dY.w (floor d)) (floor d))
dY.w
(fma (* (* dY.u dY.u) (floor w)) (floor w) (* (* dY.v dY.v) t_1)))))
(if (<= dX.u 3500.0)
(log2 (sqrt (fmax (fma t_0 (* dX.w dX.w) (* t_1 (* dX.v dX.v))) t_2)))
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* (* (floor w) (floor w)) (* 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(d) * floorf(d);
float t_1 = floorf(h) * floorf(h);
float t_2 = fmaf(((dY_46_w * floorf(d)) * floorf(d)), dY_46_w, fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), ((dY_46_v * dY_46_v) * t_1)));
float tmp;
if (dX_46_u <= 3500.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_1 * (dX_46_v * dX_46_v))), t_2)));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), ((floorf(w) * floorf(w)) * (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(d) * floor(d)) t_1 = Float32(floor(h) * floor(h)) t_2 = fma(Float32(Float32(dY_46_w * floor(d)) * floor(d)), dY_46_w, fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(Float32(dY_46_v * dY_46_v) * t_1))) tmp = Float32(0.0) if (dX_46_u <= Float32(3500.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_1 * Float32(dX_46_v * dX_46_v))), t_2))); else tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))), 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 h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dY.w, \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot t\_1\right)\right)\\
\mathbf{if}\;dX.u \leq 3500:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), t\_2\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, 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), t\_2\right)}\right)\\
\end{array}
\end{array}
if dX.u < 3500Initial program 68.2%
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-*.f3261.4
Applied rewrites61.4%
Applied rewrites61.4%
if 3500 < dX.u Initial program 68.2%
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-*.f3261.3
Applied rewrites61.3%
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 h) (floor h))))
(log2
(sqrt
(fmax
(fma (* (floor d) (floor d)) (* dX.w dX.w) (* t_0 (* dX.v dX.v)))
(fma
(* (* dY.w (floor d)) (floor d))
dY.w
(fma (* (* dY.u dY.u) (floor w)) (floor w) (* (* dY.v dY.v) 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(h) * floorf(h);
return log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w * dX_46_w), (t_0 * (dX_46_v * dX_46_v))), fmaf(((dY_46_w * floorf(d)) * floorf(d)), dY_46_w, fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), ((dY_46_v * dY_46_v) * 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(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 * dX_46_v))), fma(Float32(Float32(dY_46_w * floor(d)) * floor(d)), dY_46_w, fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(Float32(dY_46_v * dY_46_v) * t_0)))))) end
\begin{array}{l}
\\
\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 \cdot dX.v\right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dY.w, \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\right)\right)}\right)
\end{array}
\end{array}
Initial program 68.2%
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-*.f3261.4
Applied rewrites61.4%
Applied rewrites61.4%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(* (* dX.w dX.w) (* (floor d) (floor d)))
(fma
(* (* dY.w (floor d)) (floor d))
dY.w
(fma
(* (* dY.u dY.u) (floor w))
(floor w)
(* (* 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) {
return log2f(sqrtf(fmaxf(((dX_46_w * dX_46_w) * (floorf(d) * floorf(d))), fmaf(((dY_46_w * floorf(d)) * floorf(d)), dY_46_w, fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt(fmax(Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d))), fma(Float32(Float32(dY_46_w * floor(d)) * floor(d)), dY_46_w, fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))))))) end
\begin{array}{l}
\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dY.w, \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\right)\right)}\right)
\end{array}
Initial program 68.2%
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-*.f3261.3
Applied rewrites61.3%
Applied rewrites61.3%
Taylor expanded in dX.u around 0
pow2N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
lower-*.f3254.6
Applied rewrites54.6%
(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 (* (* dX.w dX.w) t_0)))
(if (<= dY.u 1500000.0)
(log2
(sqrt
(fmax
t_1
(fma (* dY.w dY.w) t_0 (* (* dY.v dY.v) (* (floor h) (floor h)))))))
(log2
(sqrt
(fmax
t_1
(fma (* (* dY.u (floor w)) (floor w)) dY.u (* (* dY.w dY.w) 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(d) * floorf(d);
float t_1 = (dX_46_w * dX_46_w) * t_0;
float tmp;
if (dY_46_u <= 1500000.0f) {
tmp = log2f(sqrtf(fmaxf(t_1, fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h)))))));
} else {
tmp = log2f(sqrtf(fmaxf(t_1, fmaf(((dY_46_u * floorf(w)) * floorf(w)), dY_46_u, ((dY_46_w * dY_46_w) * t_0)))));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(Float32(dX_46_w * dX_46_w) * t_0) tmp = Float32(0.0) if (dY_46_u <= Float32(1500000.0)) tmp = log2(sqrt(fmax(t_1, fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h))))))); else tmp = log2(sqrt(fmax(t_1, fma(Float32(Float32(dY_46_u * floor(w)) * floor(w)), dY_46_u, Float32(Float32(dY_46_w * dY_46_w) * t_0))))); 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(dX.w \cdot dX.w\right) \cdot t\_0\\
\mathbf{if}\;dY.u \leq 1500000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_1, \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)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_1, \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor , dY.u, \left(dY.w \cdot dY.w\right) \cdot t\_0\right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 1.5e6Initial program 68.2%
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-*.f3261.3
Applied rewrites61.3%
Applied rewrites61.3%
Taylor expanded in dX.u around 0
pow2N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
lower-*.f3254.6
Applied rewrites54.6%
Taylor expanded in dY.u around 0
+-commutativeN/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lower-*.f32N/A
Applied rewrites45.9%
if 1.5e6 < dY.u Initial program 68.2%
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-*.f3261.3
Applied rewrites61.3%
Applied rewrites61.3%
Taylor expanded in dX.u around 0
pow2N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
lower-*.f3254.6
Applied rewrites54.6%
Taylor expanded in dY.v around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites46.3%
(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))))
(log2
(sqrt
(fmax
(* (* dX.w dX.w) t_0)
(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);
return log2f(sqrtf(fmaxf(((dX_46_w * dX_46_w) * t_0), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h)))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(Float32(Float32(dX_46_w * dX_46_w) * t_0), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h))))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(dX.w \cdot dX.w\right) \cdot t\_0, \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}
Initial program 68.2%
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-*.f3261.3
Applied rewrites61.3%
Applied rewrites61.3%
Taylor expanded in dX.u around 0
pow2N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
lower-*.f3254.6
Applied rewrites54.6%
Taylor expanded in dY.u around 0
+-commutativeN/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lower-*.f32N/A
Applied rewrites45.9%
herbie shell --seed 2025134
(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)))))))