
(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 6 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 d) dY.w)))
(log2
(sqrt
(fmax
(fma
(* (* dX.v dX.v) (floor h))
(floor h)
(fma
(* (floor d) (floor d))
(* dX.w dX.w)
(* (* (floor w) (floor w)) (* dX.u dX.u))))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_2 t_2)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(d) * dY_46_w;
return log2f(sqrtf(fmaxf(fmaf(((dX_46_v * dX_46_v) * floorf(h)), floorf(h), fmaf((floorf(d) * floorf(d)), (dX_46_w * dX_46_w), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u)))), (((t_0 * t_0) + (t_1 * t_1)) + (t_2 * t_2)))));
}
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(d) * dY_46_w) return log2(sqrt(fmax(fma(Float32(Float32(dX_46_v * dX_46_v) * floor(h)), floor(h), fma(Float32(floor(d) * floor(d)), Float32(dX_46_w * dX_46_w), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u)))), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_2 * t_2))))) end
\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 d\right\rfloor \cdot dY.w\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.v \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , 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)\right), \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_2 \cdot t\_2\right)}\right)
\end{array}
\end{array}
Initial program 67.9%
Applied rewrites67.9%
(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) (floor d)))
(t_2 (* (floor h) (floor h)))
(t_3 (* (floor d) dX.w)))
(if (<= dY.u 500000.0)
(log2
(sqrt
(fmax
(+
(fma
(* (* dX.v dX.v) (floor h))
(floor h)
(* (* (* (floor w) dX.u) dX.u) (floor w)))
(* t_3 t_3))
(fma (* dY.w dY.w) t_1 (* (* dY.v dY.v) t_2)))))
(log2
(sqrt
(fmax
(fma t_1 (* dX.w dX.w) (* t_2 (* dX.v dX.v)))
(fma
t_0
t_0
(fma
(* dY.w dY.w)
t_1
(* (* 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(h) * dY_46_v;
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(h) * floorf(h);
float t_3 = floorf(d) * dX_46_w;
float tmp;
if (dY_46_u <= 500000.0f) {
tmp = log2f(sqrtf(fmaxf((fmaf(((dX_46_v * dX_46_v) * floorf(h)), floorf(h), (((floorf(w) * dX_46_u) * dX_46_u) * floorf(w))) + (t_3 * t_3)), fmaf((dY_46_w * dY_46_w), t_1, ((dY_46_v * dY_46_v) * t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, (dX_46_w * dX_46_w), (t_2 * (dX_46_v * dX_46_v))), fmaf(t_0, t_0, fmaf((dY_46_w * dY_46_w), t_1, ((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(h) * dY_46_v) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(floor(d) * dX_46_w) tmp = Float32(0.0) if (dY_46_u <= Float32(500000.0)) tmp = log2(sqrt(fmax(Float32(fma(Float32(Float32(dX_46_v * dX_46_v) * floor(h)), floor(h), Float32(Float32(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w))) + Float32(t_3 * t_3)), fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_v * dY_46_v) * t_2))))); else tmp = log2(sqrt(fmax(fma(t_1, Float32(dX_46_w * dX_46_w), Float32(t_2 * Float32(dX_46_v * dX_46_v))), fma(t_0, t_0, fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w)))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
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 := \left\lfloor d\right\rfloor \cdot dX.w\\
\mathbf{if}\;dY.u \leq 500000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.v \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right) + t\_3 \cdot t\_3, \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \left(dY.v \cdot dY.v\right) \cdot 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\_2 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0, t\_0, \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 5e5Initial program 67.9%
Taylor expanded in dY.u around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.6
Applied rewrites60.6%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
pow-prod-downN/A
+-commutativeN/A
Applied rewrites60.6%
if 5e5 < dY.u Initial program 67.9%
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-*.f3260.5
Applied rewrites60.5%
Applied rewrites60.5%
(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 h) (floor h)))
(t_2 (* (floor d) (floor d))))
(if (<= dY.u 500000.0)
(log2
(sqrt
(fmax
(fma
(* (* (floor d) dX.w) (floor d))
dX.w
(fma
(* (* (floor h) dX.v) (floor h))
dX.v
(* (* (* (floor w) dX.u) dX.u) (floor w))))
(fma (* dY.v dY.v) t_1 (* (* dY.w dY.w) t_2)))))
(log2
(sqrt
(fmax
(fma t_2 (* dX.w dX.w) (* t_1 (* dX.v dX.v)))
(fma
t_0
t_0
(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(h) * dY_46_v;
float t_1 = floorf(h) * floorf(h);
float t_2 = floorf(d) * floorf(d);
float tmp;
if (dY_46_u <= 500000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w) * floorf(d)), dX_46_w, fmaf(((floorf(h) * dX_46_v) * floorf(h)), dX_46_v, (((floorf(w) * dX_46_u) * dX_46_u) * floorf(w)))), fmaf((dY_46_v * dY_46_v), t_1, ((dY_46_w * dY_46_w) * t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_2, (dX_46_w * dX_46_w), (t_1 * (dX_46_v * dX_46_v))), fmaf(t_0, t_0, 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(h) * dY_46_v) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(floor(d) * floor(d)) tmp = Float32(0.0) if (dY_46_u <= Float32(500000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w) * floor(d)), dX_46_w, fma(Float32(Float32(floor(h) * dX_46_v) * floor(h)), dX_46_v, Float32(Float32(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w)))), fma(Float32(dY_46_v * dY_46_v), t_1, Float32(Float32(dY_46_w * dY_46_w) * t_2))))); else tmp = log2(sqrt(fmax(fma(t_2, Float32(dX_46_w * dX_46_w), Float32(t_1 * Float32(dX_46_v * dX_46_v))), fma(t_0, t_0, 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 h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dY.u \leq 500000:\\
\;\;\;\;\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 h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , dX.v, \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right), \mathsf{fma}\left(dY.v \cdot dY.v, t\_1, \left(dY.w \cdot dY.w\right) \cdot t\_2\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, dX.w \cdot dX.w, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0, t\_0, \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)}\right)\\
\end{array}
\end{array}
if dY.u < 5e5Initial program 67.9%
Taylor expanded in dY.u around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.6
Applied rewrites60.6%
Applied rewrites60.6%
if 5e5 < dY.u Initial program 67.9%
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-*.f3260.5
Applied rewrites60.5%
Applied rewrites60.5%
(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 (* (* dY.u dY.u) t_0))
(t_2 (* (floor d) (floor d)))
(t_3 (* (floor h) dY.v)))
(if (<= dX.v 5000.0)
(log2
(sqrt
(fmax
(fma t_2 (* dX.w dX.w) (* t_0 (* dX.u dX.u)))
(fma
(* (* dY.w dY.w) (floor d))
(floor d)
(fma (* (* dY.v dY.v) (floor h)) (floor h) t_1)))))
(log2
(sqrt
(fmax
(fma t_2 (* dX.w dX.w) (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma t_3 t_3 (fma (* dY.w dY.w) t_2 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(w) * floorf(w);
float t_1 = (dY_46_u * dY_46_u) * t_0;
float t_2 = floorf(d) * floorf(d);
float t_3 = floorf(h) * dY_46_v;
float tmp;
if (dX_46_v <= 5000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_2, (dX_46_w * dX_46_w), (t_0 * (dX_46_u * dX_46_u))), fmaf(((dY_46_w * dY_46_w) * floorf(d)), floorf(d), fmaf(((dY_46_v * dY_46_v) * floorf(h)), floorf(h), t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_2, (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf(t_3, t_3, fmaf((dY_46_w * dY_46_w), t_2, 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(w) * floor(w)) t_1 = Float32(Float32(dY_46_u * dY_46_u) * t_0) t_2 = Float32(floor(d) * floor(d)) t_3 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (dX_46_v <= Float32(5000.0)) tmp = log2(sqrt(fmax(fma(t_2, Float32(dX_46_w * dX_46_w), Float32(t_0 * Float32(dX_46_u * dX_46_u))), fma(Float32(Float32(dY_46_w * dY_46_w) * floor(d)), floor(d), fma(Float32(Float32(dY_46_v * dY_46_v) * floor(h)), floor(h), t_1))))); else tmp = log2(sqrt(fmax(fma(t_2, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))), fma(t_3, t_3, fma(Float32(dY_46_w * dY_46_w), t_2, t_1))))); 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(dY.u \cdot dY.u\right) \cdot t\_0\\
t_2 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;dX.v \leq 5000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, dX.w \cdot dX.w, t\_0 \cdot \left(dX.u \cdot dX.u\right)\right), \mathsf{fma}\left(\left(dY.w \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor , \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , t\_1\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, 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), \mathsf{fma}\left(t\_3, t\_3, \mathsf{fma}\left(dY.w \cdot dY.w, t\_2, t\_1\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.v < 5e3Initial program 67.9%
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%
Applied rewrites60.9%
if 5e3 < dX.v Initial program 67.9%
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-*.f3260.5
Applied rewrites60.5%
Applied rewrites60.5%
(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 d) (floor d)))
(t_2 (* (floor d) dY.w))
(t_3 (* (floor w) (floor w)))
(t_4 (* (floor h) dY.v)))
(if (<= dX.u 950000000.0)
(log2
(sqrt
(fmax
(fma t_1 (* dX.w dX.w) (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma t_4 t_4 (fma (* dY.w dY.w) t_1 (* (* dY.u dY.u) t_3))))))
(log2
(sqrt
(fmax
(* t_3 (* dX.u dX.u))
(+ (+ (* t_0 t_0) (* t_4 t_4)) (* t_2 t_2))))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(d) * dY_46_w;
float t_3 = floorf(w) * floorf(w);
float t_4 = floorf(h) * dY_46_v;
float tmp;
if (dX_46_u <= 950000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, (dX_46_w * dX_46_w), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf(t_4, t_4, fmaf((dY_46_w * dY_46_w), t_1, ((dY_46_u * dY_46_u) * t_3))))));
} else {
tmp = log2f(sqrtf(fmaxf((t_3 * (dX_46_u * dX_46_u)), (((t_0 * t_0) + (t_4 * t_4)) + (t_2 * t_2)))));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(d) * dY_46_w) t_3 = Float32(floor(w) * floor(w)) t_4 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (dX_46_u <= Float32(950000000.0)) tmp = log2(sqrt(fmax(fma(t_1, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))), fma(t_4, t_4, fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_u * dY_46_u) * t_3)))))); else tmp = log2(sqrt(fmax(Float32(t_3 * Float32(dX_46_u * dX_46_u)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) + Float32(t_2 * t_2))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;dX.u \leq 950000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, dX.w \cdot dX.w, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_4, t\_4, \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \left(dY.u \cdot dY.u\right) \cdot t\_3\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_3 \cdot \left(dX.u \cdot dX.u\right), \left(t\_0 \cdot t\_0 + t\_4 \cdot t\_4\right) + t\_2 \cdot t\_2\right)}\right)\\
\end{array}
\end{array}
if dX.u < 9.5e8Initial program 67.9%
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-*.f3260.5
Applied rewrites60.5%
Applied rewrites60.5%
if 9.5e8 < dX.u Initial program 67.9%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
Applied rewrites53.5%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3253.7
Applied rewrites53.7%
(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 d) dY.w)))
(log2
(sqrt
(fmax
(* (* (floor w) (floor w)) (* dX.u dX.u))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_2 t_2)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(d) * dY_46_w;
return log2f(sqrtf(fmaxf(((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u)), (((t_0 * t_0) + (t_1 * t_1)) + (t_2 * t_2)))));
}
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(d) * dY_46_w) return log2(sqrt(fmax(Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_2 * t_2))))) end
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(d) * dY_46_w; tmp = log2(sqrt(max(((floor(w) * floor(w)) * (dX_46_u * dX_46_u)), (((t_0 * t_0) + (t_1 * t_1)) + (t_2 * t_2))))); 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 d\right\rfloor \cdot dY.w\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right), \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_2 \cdot t\_2\right)}\right)
\end{array}
\end{array}
Initial program 67.9%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
Applied rewrites53.5%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3253.7
Applied rewrites53.7%
herbie shell --seed 2025129
(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)))))))