
(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}
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}
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}
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}
(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 w) dX.u))
(t_2 (* (floor w) (fabs dY.u)))
(t_3 (* (floor d) dY.w))
(t_4 (* dX.w (floor d)))
(t_5 (* (floor d) dX.w))
(t_6 (* (floor h) dX.v))
(t_7 (* (floor h) dY.v)))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_6 t_6)) (* t_5 t_5))
(+ (+ (* t_2 t_2) (* t_7 t_7)) (* t_3 t_3)))))
63.79999923706055)
(log2
(sqrt
(fmax
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* (fabs dY.u) (floor w)) (fabs dY.u))
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(fma
(* t_4 dX.w)
(floor d)
(fma (* (* (floor w) (floor w)) dX.u) dX.u (* (* dX.v dX.v) t_0))))))
(log2
(sqrt
(fmax
(fma
(* dY.w dY.w)
(* (floor d) (floor d))
(fma
(* dY.v dY.v)
t_0
(pow (exp 2.0) (log (* (- (fabs dY.u)) (floor w))))))
(fma t_4 t_4 (* (* (* dX.v (floor h)) dX.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(h) * floorf(h);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * fabsf(dY_46_u);
float t_3 = floorf(d) * dY_46_w;
float t_4 = dX_46_w * floorf(d);
float t_5 = floorf(d) * dX_46_w;
float t_6 = floorf(h) * dX_46_v;
float t_7 = floorf(h) * dY_46_v;
float tmp;
if (log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_6 * t_6)) + (t_5 * t_5)), (((t_2 * t_2) + (t_7 * t_7)) + (t_3 * t_3))))) <= 63.79999923706055f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((fabsf(dY_46_u) * floorf(w)) * fabsf(dY_46_u)), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))), fmaf((t_4 * dX_46_w), floorf(d), fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, ((dX_46_v * dX_46_v) * t_0))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dY_46_w * dY_46_w), (floorf(d) * floorf(d)), fmaf((dY_46_v * dY_46_v), t_0, powf(expf(2.0f), logf((-fabsf(dY_46_u) * floorf(w)))))), fmaf(t_4, t_4, (((dX_46_v * floorf(h)) * dX_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(h) * floor(h)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * abs(dY_46_u)) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(dX_46_w * floor(d)) t_5 = Float32(floor(d) * dX_46_w) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_6 * t_6)) + Float32(t_5 * t_5)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_7 * t_7)) + Float32(t_3 * t_3))))) <= Float32(63.79999923706055)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(abs(dY_46_u) * floor(w)) * abs(dY_46_u)), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))), fma(Float32(t_4 * dX_46_w), floor(d), fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(Float32(dX_46_v * dX_46_v) * t_0)))))); else tmp = log2(sqrt(fmax(fma(Float32(dY_46_w * dY_46_w), Float32(floor(d) * floor(d)), fma(Float32(dY_46_v * dY_46_v), t_0, (exp(Float32(2.0)) ^ log(Float32(Float32(-abs(dY_46_u)) * floor(w)))))), fma(t_4, t_4, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))))); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left|dY.u\right|\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_5 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_6 \cdot t\_6\right) + t\_5 \cdot t\_5, \left(t\_2 \cdot t\_2 + t\_7 \cdot t\_7\right) + t\_3 \cdot t\_3\right)}\right) \leq 63.79999923706055:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\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(\left|dY.u\right| \cdot \left\lfloor w\right\rfloor \right) \cdot \left|dY.u\right|, \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), \mathsf{fma}\left(t\_4 \cdot dX.w, \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, \left(dX.v \cdot dX.v\right) \cdot t\_0\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.w \cdot dY.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.v \cdot dY.v, t\_0, {\left(e^{2}\right)}^{\log \left(\left(-\left|dY.u\right|\right) \cdot \left\lfloor w\right\rfloor \right)}\right)\right), \mathsf{fma}\left(t\_4, t\_4, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}\right)\\
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 63.7999992Initial program 68.1%
Applied rewrites68.1%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
swap-sqrN/A
*-commutativeN/A
associate-*r*N/A
Applied rewrites68.0%
if 63.7999992 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.1%
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.f3260.3%
Applied rewrites60.3%
Applied rewrites60.3%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites60.3%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
unswap-sqrN/A
sqr-neg-revN/A
lift-neg.f32N/A
lift-neg.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
unpow2N/A
exp-to-powN/A
lift-log.f32N/A
*-commutativeN/A
exp-prodN/A
lower-pow.f32N/A
lower-exp.f3248.3%
Applied rewrites48.3%
(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)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) (fabs dY.u)))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (* dX.v dX.v) t_0))
(t_6 (* (floor h) dX.v))
(t_7 (* (floor h) dY.v))
(t_8 (* (floor d) (floor d))))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_6 t_6)) (* t_4 t_4))
(+ (+ (* t_2 t_2) (* t_7 t_7)) (* t_3 t_3)))))
63.79999923706055)
(log2
(sqrt
(fmax
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* (fabs dY.u) (floor w)) (fabs dY.u))
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(fma
(* (* dX.w (floor d)) dX.w)
(floor d)
(fma (* (* (floor w) (floor w)) dX.u) dX.u t_5)))))
(log2
(sqrt
(fmax
(fma
(* dY.w dY.w)
t_8
(fma
(* dY.v dY.v)
t_0
(exp (* (log (* (- (fabs dY.u)) (floor w))) 2.0))))
(fma (* dX.w dX.w) t_8 t_5)))))))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(w) * dX_46_u;
float t_2 = floorf(w) * fabsf(dY_46_u);
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = (dX_46_v * dX_46_v) * t_0;
float t_6 = floorf(h) * dX_46_v;
float t_7 = floorf(h) * dY_46_v;
float t_8 = floorf(d) * floorf(d);
float tmp;
if (log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_6 * t_6)) + (t_4 * t_4)), (((t_2 * t_2) + (t_7 * t_7)) + (t_3 * t_3))))) <= 63.79999923706055f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((fabsf(dY_46_u) * floorf(w)) * fabsf(dY_46_u)), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))), fmaf(((dX_46_w * floorf(d)) * dX_46_w), floorf(d), fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, t_5)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dY_46_w * dY_46_w), t_8, fmaf((dY_46_v * dY_46_v), t_0, expf((logf((-fabsf(dY_46_u) * floorf(w))) * 2.0f)))), fmaf((dX_46_w * dX_46_w), t_8, t_5))));
}
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(w) * dX_46_u) t_2 = Float32(floor(w) * abs(dY_46_u)) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(Float32(dX_46_v * dX_46_v) * t_0) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(floor(h) * dY_46_v) t_8 = Float32(floor(d) * floor(d)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_6 * t_6)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_7 * t_7)) + Float32(t_3 * t_3))))) <= Float32(63.79999923706055)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(abs(dY_46_u) * floor(w)) * abs(dY_46_u)), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))), fma(Float32(Float32(dX_46_w * floor(d)) * dX_46_w), floor(d), fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, t_5))))); else tmp = log2(sqrt(fmax(fma(Float32(dY_46_w * dY_46_w), t_8, fma(Float32(dY_46_v * dY_46_v), t_0, exp(Float32(log(Float32(Float32(-abs(dY_46_u)) * floor(w))) * Float32(2.0))))), fma(Float32(dX_46_w * dX_46_w), t_8, t_5)))); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left|dY.u\right|\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left(dX.v \cdot dX.v\right) \cdot t\_0\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_8 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_6 \cdot t\_6\right) + t\_4 \cdot t\_4, \left(t\_2 \cdot t\_2 + t\_7 \cdot t\_7\right) + t\_3 \cdot t\_3\right)}\right) \leq 63.79999923706055:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\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(\left|dY.u\right| \cdot \left\lfloor w\right\rfloor \right) \cdot \left|dY.u\right|, \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), \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(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, t\_5\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.w \cdot dY.w, t\_8, \mathsf{fma}\left(dY.v \cdot dY.v, t\_0, e^{\log \left(\left(-\left|dY.u\right|\right) \cdot \left\lfloor w\right\rfloor \right) \cdot 2}\right)\right), \mathsf{fma}\left(dX.w \cdot dX.w, t\_8, t\_5\right)\right)}\right)\\
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 63.7999992Initial program 68.1%
Applied rewrites68.1%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
swap-sqrN/A
*-commutativeN/A
associate-*r*N/A
Applied rewrites68.0%
if 63.7999992 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.1%
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.f3260.3%
Applied rewrites60.3%
Applied rewrites60.3%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-unsound-exp.f32N/A
lower-unsound-*.f32N/A
lower-unsound-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3248.3%
Applied rewrites48.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor w) (fabs dY.u)))
(t_2 (* (floor d) dX.w))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor h) dY.v))
(t_5 (* (floor d) (floor d)))
(t_6 (* (floor d) dY.w))
(t_7 (* (floor h) (floor h))))
(if (<=
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_3 t_3)) (* t_2 t_2))
(+ (+ (* t_1 t_1) (* t_4 t_4)) (* t_6 t_6)))))
63.79999923706055)
(log2
(sqrt
(fmax
(fma
(* (* dY.w (floor d)) dY.w)
(floor d)
(fma
(* (* (fabs dY.u) (floor w)) (fabs dY.u))
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(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)))))))
(log2
(sqrt
(fmax
(fma
(* dY.w dY.w)
t_5
(fma
(* dY.v dY.v)
t_7
(exp (* (log (* (- (fabs dY.u)) (floor w))) 2.0))))
(fma (* dX.w dX.w) t_5 (* (* dX.v dX.v) t_7))))))))float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(w) * fabsf(dY_46_u);
float t_2 = floorf(d) * dX_46_w;
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(h) * dY_46_v;
float t_5 = floorf(d) * floorf(d);
float t_6 = floorf(d) * dY_46_w;
float t_7 = floorf(h) * floorf(h);
float tmp;
if (log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_3 * t_3)) + (t_2 * t_2)), (((t_1 * t_1) + (t_4 * t_4)) + (t_6 * t_6))))) <= 63.79999923706055f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), fmaf(((fabsf(dY_46_u) * floorf(w)) * fabsf(dY_46_u)), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))), 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)))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dY_46_w * dY_46_w), t_5, fmaf((dY_46_v * dY_46_v), t_7, expf((logf((-fabsf(dY_46_u) * floorf(w))) * 2.0f)))), fmaf((dX_46_w * dX_46_w), t_5, ((dX_46_v * dX_46_v) * t_7)))));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(w) * abs(dY_46_u)) t_2 = Float32(floor(d) * dX_46_w) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(floor(d) * floor(d)) t_6 = Float32(floor(d) * dY_46_w) t_7 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) + Float32(t_2 * t_2)), Float32(Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) + Float32(t_6 * t_6))))) <= Float32(63.79999923706055)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), fma(Float32(Float32(abs(dY_46_u) * floor(w)) * abs(dY_46_u)), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))), 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))))))); else tmp = log2(sqrt(fmax(fma(Float32(dY_46_w * dY_46_w), t_5, fma(Float32(dY_46_v * dY_46_v), t_7, exp(Float32(log(Float32(Float32(-abs(dY_46_u)) * floor(w))) * Float32(2.0))))), fma(Float32(dX_46_w * dX_46_w), t_5, Float32(Float32(dX_46_v * dX_46_v) * t_7))))); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor w\right\rfloor \cdot \left|dY.u\right|\\
t_2 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_6 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_7 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_3 \cdot t\_3\right) + t\_2 \cdot t\_2, \left(t\_1 \cdot t\_1 + t\_4 \cdot t\_4\right) + t\_6 \cdot t\_6\right)}\right) \leq 63.79999923706055:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\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(\left|dY.u\right| \cdot \left\lfloor w\right\rfloor \right) \cdot \left|dY.u\right|, \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), \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)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.w \cdot dY.w, t\_5, \mathsf{fma}\left(dY.v \cdot dY.v, t\_7, e^{\log \left(\left(-\left|dY.u\right|\right) \cdot \left\lfloor w\right\rfloor \right) \cdot 2}\right)\right), \mathsf{fma}\left(dX.w \cdot dX.w, t\_5, \left(dX.v \cdot dX.v\right) \cdot t\_7\right)\right)}\right)\\
\end{array}
if (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) < 63.7999992Initial program 68.1%
Applied rewrites68.1%
if 63.7999992 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 68.1%
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.f3260.3%
Applied rewrites60.3%
Applied rewrites60.3%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-unsound-exp.f32N/A
lower-unsound-*.f32N/A
lower-unsound-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3248.3%
Applied rewrites48.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dY.w))
(t_1 (* (floor w) dY.u))
(t_2 (* dX.w (floor d)))
(t_3 (* (floor h) dY.v)))
(log2
(sqrt
(fmax
(fma t_2 t_2 (* (* dX.v dX.v) (* (floor h) (floor h))))
(+ (+ (* t_1 t_1) (* t_3 t_3)) (* t_0 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) * dY_46_w;
float t_1 = floorf(w) * dY_46_u;
float t_2 = dX_46_w * floorf(d);
float t_3 = floorf(h) * dY_46_v;
return log2f(sqrtf(fmaxf(fmaf(t_2, t_2, ((dX_46_v * dX_46_v) * (floorf(h) * floorf(h)))), (((t_1 * t_1) + (t_3 * t_3)) + (t_0 * 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(d) * dY_46_w) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(dX_46_w * floor(d)) t_3 = Float32(floor(h) * dY_46_v) return log2(sqrt(fmax(fma(t_2, t_2, Float32(Float32(dX_46_v * dX_46_v) * Float32(floor(h) * floor(h)))), Float32(Float32(Float32(t_1 * t_1) + Float32(t_3 * t_3)) + Float32(t_0 * t_0))))) end
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := dX.w \cdot \left\lfloor d\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, t\_2, \left(dX.v \cdot dX.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right), \left(t\_1 \cdot t\_1 + t\_3 \cdot t\_3\right) + t\_0 \cdot t\_0\right)}\right)
\end{array}
Initial program 68.1%
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.f3260.3%
Applied rewrites60.3%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lift-pow.f32N/A
pow-prod-downN/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-pow.f32N/A
pow-prod-downN/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lower-fma.f3260.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.3%
Applied rewrites60.3%
(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))))
(log2
(sqrt
(fmax
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(fma
(* (* dY.v (floor h)) dY.v)
(floor h)
(* (* (* (floor d) dY.w) (floor d)) dY.w)))
(fma t_0 t_0 (* (* (* dX.v (floor h)) dX.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 = dX_46_w * floorf(d);
return log2f(sqrtf(fmaxf(fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, fmaf(((dY_46_v * floorf(h)) * dY_46_v), floorf(h), (((floorf(d) * dY_46_w) * floorf(d)) * dY_46_w))), fmaf(t_0, t_0, (((dX_46_v * floorf(h)) * dX_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) t_0 = Float32(dX_46_w * floor(d)) return log2(sqrt(fmax(fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, fma(Float32(Float32(dY_46_v * floor(h)) * dY_46_v), floor(h), Float32(Float32(Float32(floor(d) * dY_46_w) * floor(d)) * dY_46_w))), fma(t_0, t_0, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))))) end
\begin{array}{l}
t_0 := dX.w \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w\right)\right), \mathsf{fma}\left(t\_0, t\_0, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}\right)
\end{array}
Initial program 68.1%
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.f3260.3%
Applied rewrites60.3%
Applied rewrites60.3%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites60.3%
Applied rewrites60.3%
(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))))
(log2
(sqrt
(fmax
(fma
(* dY.w dY.w)
(* (floor d) (floor d))
(fma
(* dY.v dY.v)
(* (floor h) (floor h))
(* (* (* dY.u (floor w)) dY.u) (floor w))))
(fma t_0 t_0 (* (* (* dX.v (floor h)) dX.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 = dX_46_w * floorf(d);
return log2f(sqrtf(fmaxf(fmaf((dY_46_w * dY_46_w), (floorf(d) * floorf(d)), fmaf((dY_46_v * dY_46_v), (floorf(h) * floorf(h)), (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w)))), fmaf(t_0, t_0, (((dX_46_v * floorf(h)) * dX_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) t_0 = Float32(dX_46_w * floor(d)) return log2(sqrt(fmax(fma(Float32(dY_46_w * dY_46_w), Float32(floor(d) * floor(d)), fma(Float32(dY_46_v * dY_46_v), Float32(floor(h) * floor(h)), Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w)))), fma(t_0, t_0, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)))))) end
\begin{array}{l}
t_0 := dX.w \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.w \cdot dY.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.v \cdot dY.v, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right), \mathsf{fma}\left(t\_0, t\_0, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}\right)
\end{array}
Initial program 68.1%
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.f3260.3%
Applied rewrites60.3%
Applied rewrites60.3%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites60.3%
(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))) (t_1 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma
(* dY.w dY.w)
t_1
(fma (* dY.v dY.v) t_0 (* (* (* dY.u (floor w)) dY.u) (floor w))))
(fma (* dX.w dX.w) t_1 (* (* dX.v dX.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);
float t_1 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf((dY_46_w * dY_46_w), t_1, fmaf((dY_46_v * dY_46_v), t_0, (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w)))), fmaf((dX_46_w * dX_46_w), t_1, ((dX_46_v * dX_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)) t_1 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(fma(Float32(dY_46_w * dY_46_w), t_1, fma(Float32(dY_46_v * dY_46_v), t_0, Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w)))), fma(Float32(dX_46_w * dX_46_w), t_1, Float32(Float32(dX_46_v * dX_46_v) * t_0))))) end
\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 \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \mathsf{fma}\left(dY.v \cdot dY.v, t\_0, \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right), \mathsf{fma}\left(dX.w \cdot dX.w, t\_1, \left(dX.v \cdot dX.v\right) \cdot t\_0\right)\right)}\right)
\end{array}
Initial program 68.1%
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.f3260.3%
Applied rewrites60.3%
Applied rewrites60.3%
herbie shell --seed 2025179
(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)))))))