
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
Herbie found 7 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
dX.w_m = (fabs.f32 dX.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) (floor d)))
(t_5 (* (floor w) dY.u))
(t_6 (log (- dX.w_m)))
(t_7 (* (floor d) dX.w_m))
(t_8
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_2 t_2)) (* t_7 t_7))
(+ (+ (* t_5 t_5) (* t_1 t_1)) (* t_3 t_3)))))))
(if (<= t_8 100.0)
t_8
(log2
(sqrt
(fmax
(fma t_4 (exp (+ t_6 t_6)) (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma (* dY.w dY.w) t_4 (* (* dY.u dY.u) (* (floor w) (floor w))))))))))dX.w_m = fabs(dX_46_w);
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w_m, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dX_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) * floorf(d);
float t_5 = floorf(w) * dY_46_u;
float t_6 = logf(-dX_46_w_m);
float t_7 = floorf(d) * dX_46_w_m;
float t_8 = log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_2 * t_2)) + (t_7 * t_7)), (((t_5 * t_5) + (t_1 * t_1)) + (t_3 * t_3)))));
float tmp;
if (t_8 <= 100.0f) {
tmp = t_8;
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_4, expf((t_6 + t_6)), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf((dY_46_w * dY_46_w), t_4, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_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) * floor(d)) t_5 = Float32(floor(w) * dY_46_u) t_6 = log(Float32(-dX_46_w_m)) t_7 = Float32(floor(d) * dX_46_w_m) t_8 = log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) + Float32(t_7 * t_7)), Float32(Float32(Float32(t_5 * t_5) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) tmp = Float32(0.0) if (t_8 <= Float32(100.0)) tmp = t_8; else tmp = log2(sqrt(fmax(fma(t_4, exp(Float32(t_6 + t_6)), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))), fma(Float32(dY_46_w * dY_46_w), t_4, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.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 \left\lfloor d\right\rfloor \\
t_5 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_6 := \log \left(-dX.w\_m\right)\\
t_7 := \left\lfloor d\right\rfloor \cdot dX.w\_m\\
t_8 := \log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_2 \cdot t\_2\right) + t\_7 \cdot t\_7, \left(t\_5 \cdot t\_5 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)\\
\mathbf{if}\;t\_8 \leq 100:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4, e^{t\_6 + t\_6}, \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 \cdot dY.w, t\_4, \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}
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 67.3%
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 67.3%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.5
Applied rewrites60.5%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3253.2
Applied rewrites53.2%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3253.2
Applied rewrites53.2%
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.f3236.5
Applied rewrites36.5%
dX.w_m = (fabs.f32 dX.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dY.w)) (t_1 (* (floor h) (floor h))))
(if (<= dY.u 2000.0)
(log2
(sqrt
(fmax
(fma
(* (* (floor d) dX.w_m) (floor d))
dX.w_m
(fma (* (* (floor w) dX.u) (floor w)) dX.u (* (* dX.v dX.v) t_1)))
(fma
(* t_0 (floor d))
dY.w
(* (* (* dY.v dY.v) (floor h)) (floor h))))))
(log2
(sqrt
(fmax
(fma (* (floor d) (floor d)) (* dX.w_m dX.w_m) (* t_1 (* dX.v dX.v)))
(+
(fma
(* (* (floor h) dY.v) (floor h))
dY.v
(* (* dY.u dY.u) (* (floor w) (floor w))))
(* t_0 t_0))))))))dX.w_m = fabs(dX_46_w);
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w_m, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * dY_46_w;
float t_1 = floorf(h) * floorf(h);
float tmp;
if (dY_46_u <= 2000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w_m) * floorf(d)), dX_46_w_m, fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, ((dX_46_v * dX_46_v) * t_1))), fmaf((t_0 * floorf(d)), dY_46_w, (((dY_46_v * dY_46_v) * floorf(h)) * floorf(h))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w_m * dX_46_w_m), (t_1 * (dX_46_v * dX_46_v))), (fmaf(((floorf(h) * dY_46_v) * floorf(h)), dY_46_v, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))) + (t_0 * t_0)))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * dY_46_w) t_1 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dY_46_u <= Float32(2000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w_m) * floor(d)), dX_46_w_m, fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, Float32(Float32(dX_46_v * dX_46_v) * t_1))), fma(Float32(t_0 * floor(d)), dY_46_w, Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)))))); else tmp = log2(sqrt(fmax(fma(Float32(floor(d) * floor(d)), Float32(dX_46_w_m * dX_46_w_m), Float32(t_1 * Float32(dX_46_v * dX_46_v))), Float32(fma(Float32(Float32(floor(h) * dY_46_v) * floor(h)), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w)))) + Float32(t_0 * t_0))))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dY.u \leq 2000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\_m\right) \cdot \left\lfloor d\right\rfloor , dX.w\_m, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, \left(dX.v \cdot dX.v\right) \cdot t\_1\right)\right), \mathsf{fma}\left(t\_0 \cdot \left\lfloor d\right\rfloor , dY.w, \left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor \right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , dX.w\_m \cdot dX.w\_m, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , dY.v, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right) + t\_0 \cdot t\_0\right)}\right)\\
\end{array}
\end{array}
if dY.u < 2e3Initial program 67.3%
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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
if 2e3 < dY.u Initial program 67.3%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.5
Applied rewrites60.5%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
Applied rewrites60.5%
dX.w_m = (fabs.f32 dX.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w)
:precision binary32
(let* ((t_0
(fma
(* (* (floor d) dY.w) (floor d))
dY.w
(* (* (* dY.v dY.v) (floor h)) (floor h))))
(t_1 (* (floor h) (floor h))))
(if (<= dY.u 2000.0)
(log2
(sqrt
(fmax
(fma
(* (* (floor d) dX.w_m) (floor d))
dX.w_m
(fma (* (* (floor w) dX.u) (floor w)) dX.u (* (* dX.v dX.v) t_1)))
t_0)))
(log2
(sqrt
(fmax
(fma (* (floor d) (floor d)) (* dX.w_m dX.w_m) (* t_1 (* dX.v dX.v)))
(fma (* (* (floor w) dY.u) (floor w)) dY.u t_0)))))))dX.w_m = fabs(dX_46_w);
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w_m, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = fmaf(((floorf(d) * dY_46_w) * floorf(d)), dY_46_w, (((dY_46_v * dY_46_v) * floorf(h)) * floorf(h)));
float t_1 = floorf(h) * floorf(h);
float tmp;
if (dY_46_u <= 2000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(d) * dX_46_w_m) * floorf(d)), dX_46_w_m, fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, ((dX_46_v * dX_46_v) * t_1))), t_0)));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w_m * dX_46_w_m), (t_1 * (dX_46_v * dX_46_v))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, t_0))));
}
return tmp;
}
dX.w_m = abs(dX_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w) t_0 = fma(Float32(Float32(floor(d) * dY_46_w) * floor(d)), dY_46_w, Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h))) t_1 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dY_46_u <= Float32(2000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(d) * dX_46_w_m) * floor(d)), dX_46_w_m, fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, Float32(Float32(dX_46_v * dX_46_v) * t_1))), t_0))); else tmp = log2(sqrt(fmax(fma(Float32(floor(d) * floor(d)), Float32(dX_46_w_m * dX_46_w_m), Float32(t_1 * Float32(dX_46_v * dX_46_v))), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, t_0)))); end return tmp end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor , dY.w, \left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dY.u \leq 2000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dX.w\_m\right) \cdot \left\lfloor d\right\rfloor , dX.w\_m, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, \left(dX.v \cdot dX.v\right) \cdot t\_1\right)\right), t\_0\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , dX.w\_m \cdot dX.w\_m, t\_1 \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, t\_0\right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 2e3Initial program 67.3%
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.f3260.0
Applied rewrites60.0%
Applied rewrites60.0%
if 2e3 < dY.u Initial program 67.3%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.5
Applied rewrites60.5%
Applied rewrites60.5%
dX.w_m = (fabs.f32 dX.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(fma
(* (floor d) (floor d))
(* dX.w_m dX.w_m)
(* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma
(* (* (floor w) dY.u) (floor w))
dY.u
(fma
(* (* (floor d) dY.w) (floor d))
dY.w
(* (* (* dY.v dY.v) (floor h)) (floor h))))))))dX.w_m = fabs(dX_46_w);
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w_m, float dY_46_u, float dY_46_v, float dY_46_w) {
return log2f(sqrtf(fmaxf(fmaf((floorf(d) * floorf(d)), (dX_46_w_m * dX_46_w_m), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, fmaf(((floorf(d) * dY_46_w) * floorf(d)), dY_46_w, (((dY_46_v * dY_46_v) * floorf(h)) * floorf(h)))))));
}
dX.w_m = abs(dX_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w) return log2(sqrt(fmax(fma(Float32(floor(d) * floor(d)), Float32(dX_46_w_m * dX_46_w_m), 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, fma(Float32(Float32(floor(d) * dY_46_w) * floor(d)), dY_46_w, Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h))))))) end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , dX.w\_m \cdot dX.w\_m, \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, \mathsf{fma}\left(\left(\left\lfloor d\right\rfloor \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor , dY.w, \left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor \right)\right)\right)}\right)
\end{array}
Initial program 67.3%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.5
Applied rewrites60.5%
Applied rewrites60.5%
dX.w_m = (fabs.f32 dX.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d))) (t_1 (* (floor w) dY.u)))
(log2
(sqrt
(fmax
(fma (* (* (floor h) (floor h)) dX.v) dX.v (* (* dX.w_m dX.w_m) t_0))
(fma t_1 t_1 (* (* dY.w dY.w) t_0)))))))dX.w_m = fabs(dX_46_w);
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w_m, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
float t_1 = floorf(w) * dY_46_u;
return log2f(sqrtf(fmaxf(fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, ((dX_46_w_m * dX_46_w_m) * t_0)), fmaf(t_1, t_1, ((dY_46_w * dY_46_w) * t_0)))));
}
dX.w_m = abs(dX_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(w) * dY_46_u) return log2(sqrt(fmax(fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, Float32(Float32(dX_46_w_m * dX_46_w_m) * t_0)), fma(t_1, t_1, Float32(Float32(dY_46_w * dY_46_w) * t_0))))) end
\begin{array}{l}
dX.w_m = \left|dX.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 dY.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, dX.v, \left(dX.w\_m \cdot dX.w\_m\right) \cdot t\_0\right), \mathsf{fma}\left(t\_1, t\_1, \left(dY.w \cdot dY.w\right) \cdot t\_0\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.3%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.5
Applied rewrites60.5%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3253.2
Applied rewrites53.2%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow-prod-downN/A
Applied rewrites53.2%
lift-*.f32N/A
lift-fma.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
pow2N/A
Applied rewrites53.2%
dX.w_m = (fabs.f32 dX.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u)) (t_1 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma t_1 (* dX.w_m dX.w_m) (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma t_0 t_0 (* (* dY.w dY.w) t_1)))))))dX.w_m = fabs(dX_46_w);
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w_m, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf(t_1, (dX_46_w_m * dX_46_w_m), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf(t_0, t_0, ((dY_46_w * dY_46_w) * t_1)))));
}
dX.w_m = abs(dX_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(fma(t_1, Float32(dX_46_w_m * dX_46_w_m), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))), fma(t_0, t_0, Float32(Float32(dY_46_w * dY_46_w) * t_1))))) end
\begin{array}{l}
dX.w_m = \left|dX.w\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, dX.w\_m \cdot dX.w\_m, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0, t\_0, \left(dY.w \cdot dY.w\right) \cdot t\_1\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.3%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.5
Applied rewrites60.5%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3253.2
Applied rewrites53.2%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow-prod-downN/A
Applied rewrites53.2%
dX.w_m = (fabs.f32 dX.w)
(FPCore (w h d dX.u dX.v dX.w_m dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma t_0 (* dX.w_m dX.w_m) (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma (* dY.w dY.w) t_0 (* (* dY.u dY.u) (* (floor w) (floor w)))))))))dX.w_m = fabs(dX_46_w);
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w_m, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w_m * dX_46_w_m), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
dX.w_m = abs(dX_46_w) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w_m, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(fma(t_0, Float32(dX_46_w_m * dX_46_w_m), Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))) end
\begin{array}{l}
dX.w_m = \left|dX.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(t\_0, dX.w\_m \cdot dX.w\_m, \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 \cdot dY.w, 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 67.3%
Taylor expanded in dX.u around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.5
Applied rewrites60.5%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3253.2
Applied rewrites53.2%
herbie shell --seed 2025143
(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)))))))