
(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
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (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(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * 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(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))) end
\begin{array}{l}
\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)
\end{array}
Initial program 68.1%
Applied rewrites68.1%
(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 (* (* dY.u (floor w)) dY.u))
(t_2 (* (floor h) dX.v))
(t_3 (* dY.v (floor h))))
(if (<= dX.u 4200.0)
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* t_2 t_2))
(fma (* t_0 dY.w) dY.w (fma t_3 t_3 (* t_1 (floor w)))))))
(log2
(sqrt
(fmax
(* dX.u (* (fabs (* dX.u (floor w))) (floor w)))
(fma
(* (fabs (* dY.w (floor d))) dY.w)
(floor d)
(fma t_1 (floor w) (* (* t_3 dY.v) (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 = (dY_46_u * floorf(w)) * dY_46_u;
float t_2 = floorf(h) * dX_46_v;
float t_3 = dY_46_v * floorf(h);
float tmp;
if (dX_46_u <= 4200.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_2 * t_2)), fmaf((t_0 * dY_46_w), dY_46_w, fmaf(t_3, t_3, (t_1 * floorf(w)))))));
} else {
tmp = log2f(sqrtf(fmaxf((dX_46_u * (fabsf((dX_46_u * floorf(w))) * floorf(w))), fmaf((fabsf((dY_46_w * floorf(d))) * dY_46_w), floorf(d), fmaf(t_1, floorf(w), ((t_3 * dY_46_v) * 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(Float32(dY_46_u * floor(w)) * dY_46_u) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (dX_46_u <= Float32(4200.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_2 * t_2)), fma(Float32(t_0 * dY_46_w), dY_46_w, fma(t_3, t_3, Float32(t_1 * floor(w))))))); else tmp = log2(sqrt(fmax(Float32(dX_46_u * Float32(abs(Float32(dX_46_u * floor(w))) * floor(w))), fma(Float32(abs(Float32(dY_46_w * floor(d))) * dY_46_w), floor(d), fma(t_1, floor(w), Float32(Float32(t_3 * dY_46_v) * 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(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.u \leq 4200:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_2 \cdot t\_2\right), \mathsf{fma}\left(t\_0 \cdot dY.w, dY.w, \mathsf{fma}\left(t\_3, t\_3, t\_1 \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(dX.u \cdot \left(\left|dX.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left|dY.w \cdot \left\lfloor d\right\rfloor \right| \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , \left(t\_3 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 4200Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3264.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dX.u around 0
lower-fma.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3256.6
Applied rewrites56.6%
Applied rewrites61.1%
if 4200 < dX.u Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3259.3
lift-*.f32N/A
*-commutativeN/A
lift-*.f3259.3
Applied rewrites59.3%
Taylor expanded in dX.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3243.5
Applied rewrites43.5%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dY.u (floor w))) (t_1 (* dX.u (floor w))))
(if (<= dY.w 80000.0)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* t_1 dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(* dY.u (* (fabs t_0) (floor w))))))
(log2
(sqrt
(fmax
(* dX.u (* (fabs t_1) (floor w)))
(fma
(* (fabs (* dY.w (floor d))) dY.w)
(floor d)
(fma
(* t_0 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (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 = dY_46_u * floorf(w);
float t_1 = dX_46_u * floorf(w);
float tmp;
if (dY_46_w <= 80000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf((t_1 * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (dY_46_u * (fabsf(t_0) * floorf(w))))));
} else {
tmp = log2f(sqrtf(fmaxf((dX_46_u * (fabsf(t_1) * floorf(w))), fmaf((fabsf((dY_46_w * floorf(d))) * dY_46_w), floorf(d), fmaf((t_0 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * 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(dY_46_u * floor(w)) t_1 = Float32(dX_46_u * floor(w)) tmp = Float32(0.0) if (dY_46_w <= Float32(80000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(t_1 * dX_46_u), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(dY_46_u * Float32(abs(t_0) * floor(w)))))); else tmp = log2(sqrt(fmax(Float32(dX_46_u * Float32(abs(t_1) * floor(w))), fma(Float32(abs(Float32(dY_46_w * floor(d))) * dY_46_w), floor(d), fma(Float32(t_0 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;dY.w \leq 80000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_1 \cdot dX.u, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), dY.u \cdot \left(\left|t\_0\right| \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(dX.u \cdot \left(\left|t\_1\right| \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left|dY.w \cdot \left\lfloor d\right\rfloor \right| \cdot dY.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_0 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)\\
\end{array}
\end{array}
if dY.w < 8e4Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3264.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.8
Applied rewrites49.8%
if 8e4 < dY.w Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3259.3
lift-*.f32N/A
*-commutativeN/A
lift-*.f3259.3
Applied rewrites59.3%
Taylor expanded in dX.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3243.5
Applied rewrites43.5%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dY.w (floor d)))
(t_1 (* (* dX.u (floor w)) dX.u))
(t_2 (* (floor h) dX.v)))
(if (<= dY.u 100.0)
(log2
(sqrt
(fmax
(fma
(* (* (floor d) (floor d)) dX.w)
dX.w
(fma t_1 (floor w) (* t_2 t_2)))
(* t_0 (fabs t_0)))))
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma t_1 (floor w) (* (* (* dX.v (floor h)) dX.v) (floor h))))
(* dY.u (* (fabs (* 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 = dY_46_w * floorf(d);
float t_1 = (dX_46_u * floorf(w)) * dX_46_u;
float t_2 = floorf(h) * dX_46_v;
float tmp;
if (dY_46_u <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * floorf(d)) * dX_46_w), dX_46_w, fmaf(t_1, floorf(w), (t_2 * t_2))), (t_0 * fabsf(t_0)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(t_1, floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (dY_46_u * (fabsf((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(dY_46_w * floor(d)) t_1 = Float32(Float32(dX_46_u * floor(w)) * dX_46_u) t_2 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if (dY_46_u <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * floor(d)) * dX_46_w), dX_46_w, fma(t_1, floor(w), Float32(t_2 * t_2))), Float32(t_0 * abs(t_0))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(t_1, floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(dY_46_u * Float32(abs(Float32(dY_46_u * floor(w))) * floor(w)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.w \cdot \left\lfloor d\right\rfloor \\
t_1 := \left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;dY.u \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, dX.w, \mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , t\_2 \cdot t\_2\right)\right), t\_0 \cdot \left|t\_0\right|\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), dY.u \cdot \left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 100Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.7
Applied rewrites49.7%
Applied rewrites49.7%
if 100 < dY.u Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3264.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.8
Applied rewrites49.8%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (* dX.w (floor d)) dX.w)) (t_1 (* (* dX.u (floor w)) dX.u)))
(if (<= dY.u 100.0)
(log2
(sqrt
(fmax
(fma
t_0
(floor d)
(fma t_1 (floor w) (* (* (floor h) (floor h)) (* dX.v dX.v))))
(* dY.w (* (fabs (* dY.w (floor d))) (floor d))))))
(log2
(sqrt
(fmax
(fma
t_0
(floor d)
(fma t_1 (floor w) (* (* (* dX.v (floor h)) dX.v) (floor h))))
(* dY.u (* (fabs (* 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 = (dX_46_w * floorf(d)) * dX_46_w;
float t_1 = (dX_46_u * floorf(w)) * dX_46_u;
float tmp;
if (dY_46_u <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, floorf(d), fmaf(t_1, floorf(w), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v)))), (dY_46_w * (fabsf((dY_46_w * floorf(d))) * floorf(d))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, floorf(d), fmaf(t_1, floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)))), (dY_46_u * (fabsf((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(Float32(dX_46_w * floor(d)) * dX_46_w) t_1 = Float32(Float32(dX_46_u * floor(w)) * dX_46_u) tmp = Float32(0.0) if (dY_46_u <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(t_0, floor(d), fma(t_1, floor(w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v)))), Float32(dY_46_w * Float32(abs(Float32(dY_46_w * floor(d))) * floor(d)))))); else tmp = log2(sqrt(fmax(fma(t_0, floor(d), fma(t_1, floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))), Float32(dY_46_u * Float32(abs(Float32(dY_46_u * floor(w))) * floor(w)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w\\
t_1 := \left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\\
\mathbf{if}\;dY.u \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right)\right), dY.w \cdot \left(\left|dY.w \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, \left\lfloor d\right\rfloor , \mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right), dY.u \cdot \left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 100Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.7
Applied rewrites49.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
unswap-sqrN/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
*-commutativeN/A
lower-*.f3249.7
lift-pow.f32N/A
unpow2N/A
lower-*.f3249.7
lift-pow.f32N/A
unpow2N/A
lower-*.f3249.7
Applied rewrites49.7%
if 100 < dY.u Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3264.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.2
Applied rewrites64.2%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.8
Applied rewrites49.8%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (floor h) (floor h)) (* dX.v dX.v))))
(* dY.w (* (fabs (* dY.w (floor d))) (floor d)))))))
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(fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v)))), (dY_46_w * (fabsf((dY_46_w * floorf(d))) * floorf(d))))));
}
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(fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v)))), Float32(dY_46_w * Float32(abs(Float32(dY_46_w * floor(d))) * floor(d)))))) end
\begin{array}{l}
\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right)\right), dY.w \cdot \left(\left|dY.w \cdot \left\lfloor d\right\rfloor \right| \cdot \left\lfloor d\right\rfloor \right)\right)}\right)
\end{array}
Initial program 68.1%
Applied rewrites68.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.1
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lower-fabs.f3263.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.9
Applied rewrites63.9%
Taylor expanded in dY.w around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3249.7
Applied rewrites49.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
unswap-sqrN/A
unpow2N/A
lift-pow.f32N/A
unpow2N/A
lift-pow.f32N/A
*-commutativeN/A
lower-*.f3249.7
lift-pow.f32N/A
unpow2N/A
lower-*.f3249.7
lift-pow.f32N/A
unpow2N/A
lower-*.f3249.7
Applied rewrites49.7%
herbie shell --seed 2025155
(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)))))))