
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
Herbie found 9 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor d) (floor d)))
(t_4 (* (floor h) dY.v))
(t_5 (* (floor h) dX.v))
(t_6 (* (floor d) dY.w_m))
(t_7 (* (floor d) dX.w))
(t_8 (log (- dY.w_m)))
(t_9 (+ (+ (* t_1 t_1) (* t_5 t_5)) (* t_7 t_7))))
(if (<=
(log2 (sqrt (fmax t_9 (+ (+ (* t_2 t_2) (* t_4 t_4)) (* t_6 t_6)))))
100.0)
(log2
(sqrt
(fmax
t_9
(fma
(* t_0 dY.u)
dY.u
(fma
(* dY.w_m dY.w_m)
t_3
(* (* dY.v dY.v) (* (floor h) (floor h))))))))
(log2
(sqrt (fmax t_9 (fma (exp (+ t_8 t_8)) t_3 (* (* dY.u dY.u) t_0))))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(d) * floorf(d);
float t_4 = floorf(h) * dY_46_v;
float t_5 = floorf(h) * dX_46_v;
float t_6 = floorf(d) * dY_46_w_m;
float t_7 = floorf(d) * dX_46_w;
float t_8 = logf(-dY_46_w_m);
float t_9 = ((t_1 * t_1) + (t_5 * t_5)) + (t_7 * t_7);
float tmp;
if (log2f(sqrtf(fmaxf(t_9, (((t_2 * t_2) + (t_4 * t_4)) + (t_6 * t_6))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(t_9, fmaf((t_0 * dY_46_u), dY_46_u, fmaf((dY_46_w_m * dY_46_w_m), t_3, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))))))));
} else {
tmp = log2f(sqrtf(fmaxf(t_9, fmaf(expf((t_8 + t_8)), t_3, ((dY_46_u * dY_46_u) * t_0)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(floor(d) * dY_46_w_m) t_7 = Float32(floor(d) * dX_46_w) t_8 = log(Float32(-dY_46_w_m)) t_9 = Float32(Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) + Float32(t_7 * t_7)) tmp = Float32(0.0) if (log2(sqrt(fmax(t_9, Float32(Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)) + Float32(t_6 * t_6))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(t_9, fma(Float32(t_0 * dY_46_u), dY_46_u, fma(Float32(dY_46_w_m * dY_46_w_m), t_3, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))))))); else tmp = log2(sqrt(fmax(t_9, fma(exp(Float32(t_8 + t_8)), t_3, Float32(Float32(dY_46_u * dY_46_u) * t_0))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := \left\lfloor d\right\rfloor \cdot dY.w\_m\\
t_7 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_8 := \log \left(-dY.w\_m\right)\\
t_9 := \left(t\_1 \cdot t\_1 + t\_5 \cdot t\_5\right) + t\_7 \cdot t\_7\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_9, \left(t\_2 \cdot t\_2 + t\_4 \cdot t\_4\right) + t\_6 \cdot t\_6\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_9, \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, t\_3, \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)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_9, \mathsf{fma}\left(e^{t\_8 + t\_8}, t\_3, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}\right)\\
\end{array}
\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)))))) < 100Initial program 67.5%
lift-+.f32N/A
lift-+.f32N/A
associate-+l+N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
unpow2N/A
associate-*r*N/A
Applied rewrites67.5%
if 100 < (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 67.5%
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.f3260.3
Applied rewrites60.3%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3260.2
Applied rewrites60.2%
lift-*.f32N/A
lift-log.f32N/A
*-commutativeN/A
log-pow-revN/A
pow2N/A
sqr-neg-revN/A
log-prodN/A
lower-+.f32N/A
lower-log.f32N/A
lower-neg.f32N/A
lower-log.f32N/A
lower-neg.f3246.2
Applied rewrites46.2%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor d) dY.w_m))
(t_5 (* (floor d) dX.w))
(t_6 (* (floor h) dX.v))
(t_7 (* (floor h) (floor h)))
(t_8 (+ (+ (* t_1 t_1) (* t_6 t_6)) (* t_5 t_5)))
(t_9 (* t_7 dY.v)))
(if (<=
(log2 (sqrt (fmax t_8 (+ (+ (* t_2 t_2) (* t_3 t_3)) (* t_4 t_4)))))
100.0)
(log2
(sqrt
(fmax
t_8
(fma
(* t_0 dY.u)
dY.u
(fma
(* dY.w_m dY.w_m)
(* (floor d) (floor d))
(* (* dY.v dY.v) t_7))))))
(log2
(sqrt
(fmax
t_8
(/
(- (pow (fma t_9 dY.v (* (* dY.u dY.u) t_0)) 2.0) (pow t_4 4.0))
(fma t_9 dY.v (* (fma (floor w) dY.u t_4) (- t_2 t_4))))))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(d) * dY_46_w_m;
float t_5 = floorf(d) * dX_46_w;
float t_6 = floorf(h) * dX_46_v;
float t_7 = floorf(h) * floorf(h);
float t_8 = ((t_1 * t_1) + (t_6 * t_6)) + (t_5 * t_5);
float t_9 = t_7 * dY_46_v;
float tmp;
if (log2f(sqrtf(fmaxf(t_8, (((t_2 * t_2) + (t_3 * t_3)) + (t_4 * t_4))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(t_8, fmaf((t_0 * dY_46_u), dY_46_u, fmaf((dY_46_w_m * dY_46_w_m), (floorf(d) * floorf(d)), ((dY_46_v * dY_46_v) * t_7))))));
} else {
tmp = log2f(sqrtf(fmaxf(t_8, ((powf(fmaf(t_9, dY_46_v, ((dY_46_u * dY_46_u) * t_0)), 2.0f) - powf(t_4, 4.0f)) / fmaf(t_9, dY_46_v, (fmaf(floorf(w), dY_46_u, t_4) * (t_2 - t_4)))))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(d) * dY_46_w_m) t_5 = Float32(floor(d) * dX_46_w) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(floor(h) * floor(h)) t_8 = Float32(Float32(Float32(t_1 * t_1) + Float32(t_6 * t_6)) + Float32(t_5 * t_5)) t_9 = Float32(t_7 * dY_46_v) tmp = Float32(0.0) if (log2(sqrt(fmax(t_8, Float32(Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) + Float32(t_4 * t_4))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(t_8, fma(Float32(t_0 * dY_46_u), dY_46_u, fma(Float32(dY_46_w_m * dY_46_w_m), Float32(floor(d) * floor(d)), Float32(Float32(dY_46_v * dY_46_v) * t_7)))))); else tmp = log2(sqrt(fmax(t_8, Float32(Float32((fma(t_9, dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_0)) ^ Float32(2.0)) - (t_4 ^ Float32(4.0))) / fma(t_9, dY_46_v, Float32(fma(floor(w), dY_46_u, t_4) * Float32(t_2 - t_4))))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\_m\\
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 \left\lfloor h\right\rfloor \\
t_8 := \left(t\_1 \cdot t\_1 + t\_6 \cdot t\_6\right) + t\_5 \cdot t\_5\\
t_9 := t\_7 \cdot dY.v\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_8, \left(t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right) + t\_4 \cdot t\_4\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_8, \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot t\_7\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_8, \frac{{\left(\mathsf{fma}\left(t\_9, dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}^{2} - {t\_4}^{4}}{\mathsf{fma}\left(t\_9, dY.v, \mathsf{fma}\left(\left\lfloor w\right\rfloor , dY.u, t\_4\right) \cdot \left(t\_2 - t\_4\right)\right)}\right)}\right)\\
\end{array}
\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)))))) < 100Initial program 67.5%
lift-+.f32N/A
lift-+.f32N/A
associate-+l+N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
unpow2N/A
associate-*r*N/A
Applied rewrites67.5%
if 100 < (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 67.5%
Applied rewrites37.3%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor d) (floor d)))
(t_4 (* (floor h) dY.v))
(t_5 (* (floor d) dY.w_m))
(t_6 (* (floor d) dX.w))
(t_7 (* (floor h) dX.v))
(t_8 (+ (+ (* t_1 t_1) (* t_7 t_7)) (* t_6 t_6)))
(t_9 (* (floor h) (floor h))))
(if (<=
(log2 (sqrt (fmax t_8 (+ (+ (* t_2 t_2) (* t_4 t_4)) (* t_5 t_5)))))
100.0)
(log2
(sqrt
(fmax
t_8
(fma
(* t_0 dY.u)
dY.u
(fma (* dY.w_m dY.w_m) t_3 (* (* dY.v dY.v) t_9))))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_3 (* t_9 (* dX.v dX.v)))
(fma (pow (exp 2.0) (log (- dY.w_m))) t_3 (* (* dY.u dY.u) t_0))))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(d) * floorf(d);
float t_4 = floorf(h) * dY_46_v;
float t_5 = floorf(d) * dY_46_w_m;
float t_6 = floorf(d) * dX_46_w;
float t_7 = floorf(h) * dX_46_v;
float t_8 = ((t_1 * t_1) + (t_7 * t_7)) + (t_6 * t_6);
float t_9 = floorf(h) * floorf(h);
float tmp;
if (log2f(sqrtf(fmaxf(t_8, (((t_2 * t_2) + (t_4 * t_4)) + (t_5 * t_5))))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(t_8, fmaf((t_0 * dY_46_u), dY_46_u, fmaf((dY_46_w_m * dY_46_w_m), t_3, ((dY_46_v * dY_46_v) * t_9))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_3, (t_9 * (dX_46_v * dX_46_v))), fmaf(powf(expf(2.0f), logf(-dY_46_w_m)), t_3, ((dY_46_u * dY_46_u) * t_0)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(floor(d) * dY_46_w_m) t_6 = Float32(floor(d) * dX_46_w) t_7 = Float32(floor(h) * dX_46_v) t_8 = Float32(Float32(Float32(t_1 * t_1) + Float32(t_7 * t_7)) + Float32(t_6 * t_6)) t_9 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (log2(sqrt(fmax(t_8, Float32(Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)) + Float32(t_5 * t_5))))) <= Float32(100.0)) tmp = log2(sqrt(fmax(t_8, fma(Float32(t_0 * dY_46_u), dY_46_u, fma(Float32(dY_46_w_m * dY_46_w_m), t_3, Float32(Float32(dY_46_v * dY_46_v) * t_9)))))); else tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_3, Float32(t_9 * Float32(dX_46_v * dX_46_v))), fma((exp(Float32(2.0)) ^ log(Float32(-dY_46_w_m))), t_3, Float32(Float32(dY_46_u * dY_46_u) * t_0))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := \left\lfloor d\right\rfloor \cdot dY.w\_m\\
t_6 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_7 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_8 := \left(t\_1 \cdot t\_1 + t\_7 \cdot t\_7\right) + t\_6 \cdot t\_6\\
t_9 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_8, \left(t\_2 \cdot t\_2 + t\_4 \cdot t\_4\right) + t\_5 \cdot t\_5\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_8, \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, t\_3, \left(dY.v \cdot dY.v\right) \cdot t\_9\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_3, t\_9 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left({\left(e^{2}\right)}^{\log \left(-dY.w\_m\right)}, t\_3, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}\right)\\
\end{array}
\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)))))) < 100Initial program 67.5%
lift-+.f32N/A
lift-+.f32N/A
associate-+l+N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
unpow2N/A
associate-*r*N/A
Applied rewrites67.5%
if 100 < (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 67.5%
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.f3260.3
Applied rewrites60.3%
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3247.6
Applied rewrites47.6%
Taylor expanded in dX.u around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
pow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
Applied rewrites53.1%
lift-*.f32N/A
sqr-neg-revN/A
lift-neg.f32N/A
lift-neg.f32N/A
pow2N/A
pow-to-expN/A
lift-neg.f32N/A
*-commutativeN/A
exp-prodN/A
lower-pow.f32N/A
lower-exp.f32N/A
lift-neg.f32N/A
lift-log.f3236.3
Applied rewrites36.3%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* dX.v (floor h)))
(t_2 (* (floor d) (floor d)))
(t_3 (fma (* dX.w dX.w) t_2 (* t_1 t_1)))
(t_4 (* (* dY.u dY.u) t_0)))
(if (<= dX.u 2000000.0)
(log2
(sqrt
(fmax
t_3
(fma
(* (* dY.w_m dY.w_m) (floor d))
(floor d)
(fma (* dY.v dY.v) (* (floor h) (floor h)) t_4)))))
(log2
(sqrt
(fmax (fma (* t_0 dX.u) dX.u t_3) (fma (* dY.w_m dY.w_m) t_2 t_4)))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(w) * floorf(w);
float t_1 = dX_46_v * floorf(h);
float t_2 = floorf(d) * floorf(d);
float t_3 = fmaf((dX_46_w * dX_46_w), t_2, (t_1 * t_1));
float t_4 = (dY_46_u * dY_46_u) * t_0;
float tmp;
if (dX_46_u <= 2000000.0f) {
tmp = log2f(sqrtf(fmaxf(t_3, fmaf(((dY_46_w_m * dY_46_w_m) * floorf(d)), floorf(d), fmaf((dY_46_v * dY_46_v), (floorf(h) * floorf(h)), t_4)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, t_3), fmaf((dY_46_w_m * dY_46_w_m), t_2, t_4))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(floor(d) * floor(d)) t_3 = fma(Float32(dX_46_w * dX_46_w), t_2, Float32(t_1 * t_1)) t_4 = Float32(Float32(dY_46_u * dY_46_u) * t_0) tmp = Float32(0.0) if (dX_46_u <= Float32(2000000.0)) tmp = log2(sqrt(fmax(t_3, fma(Float32(Float32(dY_46_w_m * dY_46_w_m) * floor(d)), floor(d), fma(Float32(dY_46_v * dY_46_v), Float32(floor(h) * floor(h)), t_4))))); else tmp = log2(sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, t_3), fma(Float32(dY_46_w_m * dY_46_w_m), t_2, t_4)))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_3 := \mathsf{fma}\left(dX.w \cdot dX.w, t\_2, t\_1 \cdot t\_1\right)\\
t_4 := \left(dY.u \cdot dY.u\right) \cdot t\_0\\
\mathbf{if}\;dX.u \leq 2000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(\left(dY.w\_m \cdot dY.w\_m\right) \cdot \left\lfloor d\right\rfloor , \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.v \cdot dY.v, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , t\_4\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_3\right), \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, t\_2, t\_4\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 2e6Initial program 67.5%
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-*.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
if 2e6 < dX.u Initial program 67.5%
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.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor d) (floor d)))
(t_2 (* (floor w) (floor w)))
(t_3 (* dX.v (floor h))))
(if (<= dX.u 20000000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_1 (* t_3 t_3))
(fma
(* (* dY.w_m dY.w_m) (floor d))
(floor d)
(fma (* dY.v dY.v) t_0 (* (* dY.u dY.u) t_2))))))
(log2
(sqrt
(fmax
(fma
(* (* (floor d) dX.w) (floor d))
dX.w
(fma (* (* (floor h) dX.v) (floor h)) dX.v (* (* dX.u dX.u) t_2)))
(fma (* dY.w_m dY.w_m) t_1 (* (* dY.v dY.v) t_0))))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(w) * floorf(w);
float t_3 = dX_46_v * floorf(h);
float tmp;
if (dX_46_u <= 20000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_1, (t_3 * t_3)), fmaf(((dY_46_w_m * dY_46_w_m) * floorf(d)), floorf(d), fmaf((dY_46_v * dY_46_v), t_0, ((dY_46_u * dY_46_u) * t_2))))));
} else {
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, ((dX_46_u * dX_46_u) * t_2))), fmaf((dY_46_w_m * dY_46_w_m), t_1, ((dY_46_v * dY_46_v) * t_0)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(dX_46_v * floor(h)) tmp = Float32(0.0) if (dX_46_u <= Float32(20000000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_1, Float32(t_3 * t_3)), fma(Float32(Float32(dY_46_w_m * dY_46_w_m) * floor(d)), floor(d), fma(Float32(dY_46_v * dY_46_v), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_2)))))); else 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(dX_46_u * dX_46_u) * t_2))), fma(Float32(dY_46_w_m * dY_46_w_m), t_1, Float32(Float32(dY_46_v * dY_46_v) * t_0))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.u \leq 20000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_1, t\_3 \cdot t\_3\right), \mathsf{fma}\left(\left(dY.w\_m \cdot dY.w\_m\right) \cdot \left\lfloor d\right\rfloor , \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.v \cdot dY.v, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_2\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor , dX.w, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , dX.v, \left(dX.u \cdot dX.u\right) \cdot t\_2\right)\right), \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, t\_1, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 2e7Initial program 67.5%
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-*.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
if 2e7 < dX.u Initial program 67.5%
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.f3260.9
Applied rewrites60.9%
Applied rewrites60.9%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* dX.v (floor h)))
(t_2 (* (floor d) (floor d)))
(t_3 (* (* dY.u dY.u) t_0)))
(if (<= dX.u 450000000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_2 (* t_1 t_1))
(fma
(* (* dY.w_m dY.w_m) (floor d))
(floor d)
(fma (* dY.v dY.v) (* (floor h) (floor h)) t_3)))))
(log2
(sqrt
(fmax
(fma (* (* (floor d) dX.w) (floor d)) dX.w (* (* dX.u dX.u) t_0))
(fma (* dY.w_m dY.w_m) t_2 t_3)))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(w) * floorf(w);
float t_1 = dX_46_v * floorf(h);
float t_2 = floorf(d) * floorf(d);
float t_3 = (dY_46_u * dY_46_u) * t_0;
float tmp;
if (dX_46_u <= 450000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_2, (t_1 * t_1)), fmaf(((dY_46_w_m * dY_46_w_m) * floorf(d)), floorf(d), fmaf((dY_46_v * dY_46_v), (floorf(h) * floorf(h)), t_3)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w) * floorf(d)), dX_46_w, ((dX_46_u * dX_46_u) * t_0)), fmaf((dY_46_w_m * dY_46_w_m), t_2, t_3))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(floor(d) * floor(d)) t_3 = Float32(Float32(dY_46_u * dY_46_u) * t_0) tmp = Float32(0.0) if (dX_46_u <= Float32(450000000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_2, Float32(t_1 * t_1)), fma(Float32(Float32(dY_46_w_m * dY_46_w_m) * floor(d)), floor(d), fma(Float32(dY_46_v * dY_46_v), Float32(floor(h) * floor(h)), t_3))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w) * floor(d)), dX_46_w, Float32(Float32(dX_46_u * dX_46_u) * t_0)), fma(Float32(dY_46_w_m * dY_46_w_m), t_2, t_3)))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_3 := \left(dY.u \cdot dY.u\right) \cdot t\_0\\
\mathbf{if}\;dX.u \leq 450000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_2, t\_1 \cdot t\_1\right), \mathsf{fma}\left(\left(dY.w\_m \cdot dY.w\_m\right) \cdot \left\lfloor d\right\rfloor , \left\lfloor d\right\rfloor , \mathsf{fma}\left(dY.v \cdot dY.v, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , t\_3\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor , dX.w, \left(dX.u \cdot dX.u\right) \cdot t\_0\right), \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, t\_2, t\_3\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 4.5e8Initial program 67.5%
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-*.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
if 4.5e8 < dX.u Initial program 67.5%
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.f3260.3
Applied rewrites60.3%
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3247.6
Applied rewrites47.6%
Taylor expanded in dX.v around 0
+-commutativeN/A
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
lift-floor.f32N/A
exp-prodN/A
Applied rewrites52.3%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor d) (floor d))) (t_1 (* (floor w) (floor w))))
(if (<= dX.u 200000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma
(* (* (floor w) dY.u) (floor w))
dY.u
(* (* dY.w_m dY.w_m) t_0)))))
(log2
(sqrt
(fmax
(fma (* (* (floor d) dX.w) (floor d)) dX.w (* (* dX.u dX.u) t_1))
(fma (* dY.w_m dY.w_m) t_0 (* (* dY.u dY.u) t_1))))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(d) * floorf(d);
float t_1 = floorf(w) * floorf(w);
float tmp;
if (dX_46_u <= 200000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, ((dY_46_w_m * dY_46_w_m) * t_0)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w) * floorf(d)), dX_46_w, ((dX_46_u * dX_46_u) * t_1)), fmaf((dY_46_w_m * dY_46_w_m), t_0, ((dY_46_u * dY_46_u) * t_1)))));
}
return tmp;
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (dX_46_u <= Float32(200000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, Float32(Float32(dY_46_w_m * dY_46_w_m) * t_0))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w) * floor(d)), dX_46_w, Float32(Float32(dX_46_u * dX_46_u) * t_1)), fma(Float32(dY_46_w_m * dY_46_w_m), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_1))))); end return tmp end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;dX.u \leq 200000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, \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(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, \left(dY.w\_m \cdot dY.w\_m\right) \cdot t\_0\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor , dX.w, \left(dX.u \cdot dX.u\right) \cdot t\_1\right), \mathsf{fma}\left(dY.w\_m \cdot dY.w\_m, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 2e5Initial program 67.5%
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.f3260.3
Applied rewrites60.3%
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3247.6
Applied rewrites47.6%
Taylor expanded in dX.u around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
pow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
Applied rewrites53.1%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
pow-prod-downN/A
*-commutativeN/A
pow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
Applied rewrites53.1%
if 2e5 < dX.u Initial program 67.5%
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.f3260.3
Applied rewrites60.3%
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3247.6
Applied rewrites47.6%
Taylor expanded in dX.v around 0
+-commutativeN/A
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
lift-floor.f32N/A
exp-prodN/A
Applied rewrites52.3%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma
(* (* (floor w) dY.u) (floor w))
dY.u
(* (* dY.w_m dY.w_m) t_0)))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, ((dY_46_w_m * dY_46_w_m) * t_0)))));
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, Float32(Float32(dY_46_w_m * dY_46_w_m) * t_0))))) end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, \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(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, \left(dY.w\_m \cdot dY.w\_m\right) \cdot t\_0\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.5%
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.f3260.3
Applied rewrites60.3%
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3247.6
Applied rewrites47.6%
Taylor expanded in dX.u around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
pow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
Applied rewrites53.1%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
pow-prod-downN/A
*-commutativeN/A
pow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
Applied rewrites53.1%
dY.w_m = (fabs.f32 dY.w)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w_m)
:precision binary32
(let* ((t_0 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma
(* dY.w_m dY.w_m)
t_0
(* (* dY.u dY.u) (* (floor w) (floor w)))))))))dY.w_m = fabs(dY_46_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_m) {
float t_0 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf((dY_46_w_m * dY_46_w_m), t_0, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
dY.w_m = abs(dY_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w_m) t_0 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))), fma(Float32(dY_46_w_m * dY_46_w_m), t_0, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))) end
\begin{array}{l}
dY.w_m = \left|dY.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, \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(dY.w\_m \cdot dY.w\_m, t\_0, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.5%
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.f3260.3
Applied rewrites60.3%
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3247.6
Applied rewrites47.6%
Taylor expanded in dX.u around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
pow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
Applied rewrites53.1%
herbie shell --seed 2025140
(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)))))))