
(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 12 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.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) (floor h)))
(t_4 (* (floor h) dY.v_m))
(t_5 (* (floor h) dX.v))
(t_6 (* (floor d) dY.w))
(t_7 (* (floor d) dX.w))
(t_8
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (* t_5 t_5)) (* t_7 t_7))
(+ (+ (* t_2 t_2) (* t_4 t_4)) (* t_6 t_6)))))))
(if (<= t_8 100.0)
t_8
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* (* dX.v dX.v) t_3))
(fma
(* dY.w dY.w)
t_0
(fma
(* dY.u dY.u)
(* (floor w) (floor w))
(* (exp (+ 0.0 (* 2.0 (- (- (log (- dY.v_m))))))) t_3)))))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * floorf(h);
float t_4 = floorf(h) * dY_46_v_m;
float t_5 = floorf(h) * dX_46_v;
float t_6 = floorf(d) * dY_46_w;
float t_7 = floorf(d) * dX_46_w;
float t_8 = log2f(sqrtf(fmaxf((((t_1 * t_1) + (t_5 * t_5)) + (t_7 * t_7)), (((t_2 * t_2) + (t_4 * t_4)) + (t_6 * t_6)))));
float tmp;
if (t_8 <= 100.0f) {
tmp = t_8;
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, ((dX_46_v * dX_46_v) * t_3)), fmaf((dY_46_w * dY_46_w), t_0, fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), (expf((0.0f + (2.0f * -(-logf(-dY_46_v_m))))) * t_3))))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(floor(h) * dY_46_v_m) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(floor(d) * dY_46_w) t_7 = Float32(floor(d) * dX_46_w) t_8 = log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) + Float32(t_7 * t_7)), Float32(Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)) + Float32(t_6 * t_6))))) tmp = Float32(0.0) if (t_8 <= Float32(100.0)) tmp = t_8; else tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(Float32(dX_46_v * dX_46_v) * t_3)), fma(Float32(dY_46_w * dY_46_w), t_0, fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), Float32(exp(Float32(Float32(0.0) + Float32(Float32(2.0) * Float32(-Float32(-log(Float32(-dY_46_v_m))))))) * t_3)))))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\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 dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\_m\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_7 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_8 := \log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + t\_5 \cdot t\_5\right) + t\_7 \cdot t\_7, \left(t\_2 \cdot t\_2 + t\_4 \cdot t\_4\right) + t\_6 \cdot t\_6\right)}\right)\\
\mathbf{if}\;t\_8 \leq 100:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, \left(dX.v \cdot dX.v\right) \cdot t\_3\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , e^{0 + 2 \cdot \left(-\left(-\log \left(-dY.v\_m\right)\right)\right)} \cdot t\_3\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%
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 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-*.f3260.3
Applied rewrites60.3%
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%
Taylor expanded in dY.v around -inf
distribute-lft-inN/A
log-pow-revN/A
metadata-evalN/A
metadata-evalN/A
lower-+.f32N/A
lower-*.f32N/A
mul-1-negN/A
lower-neg.f32N/A
frac-2negN/A
metadata-evalN/A
log-recN/A
lower-neg.f32N/A
lower-log.f32N/A
lower-neg.f3236.4
Applied rewrites36.4%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v_m))
(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)))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v_m;
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)))));
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v_m) 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
dY.v_m = abs(dY_46_v); function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v_m; 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}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\_m\\
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}
Initial program 67.5%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dX.w))
(t_1 (* (floor h) dY.v_m))
(t_2 (* (floor d) dY.w))
(t_3 (* (floor h) (floor h)))
(t_4 (* (* dX.v dX.v) t_3))
(t_5 (* (floor w) dX.u))
(t_6 (* (floor w) dY.u)))
(if (<= dX.w 100.0)
(log2
(sqrt
(fmax
(fma (* (* (floor w) (floor w)) dX.u) dX.u t_4)
(+ (+ (* t_6 t_6) (* t_1 t_1)) (* t_2 t_2)))))
(log2
(sqrt
(fmax
(+ (fma t_5 t_5 t_4) (* t_0 t_0))
(fma
(* dY.w dY.w)
(* (floor d) (floor d))
(* (* dY.v_m dY.v_m) t_3))))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(d) * dX_46_w;
float t_1 = floorf(h) * dY_46_v_m;
float t_2 = floorf(d) * dY_46_w;
float t_3 = floorf(h) * floorf(h);
float t_4 = (dX_46_v * dX_46_v) * t_3;
float t_5 = floorf(w) * dX_46_u;
float t_6 = floorf(w) * dY_46_u;
float tmp;
if (dX_46_w <= 100.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, t_4), (((t_6 * t_6) + (t_1 * t_1)) + (t_2 * t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf((fmaf(t_5, t_5, t_4) + (t_0 * t_0)), fmaf((dY_46_w * dY_46_w), (floorf(d) * floorf(d)), ((dY_46_v_m * dY_46_v_m) * t_3)))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(d) * dX_46_w) t_1 = Float32(floor(h) * dY_46_v_m) t_2 = Float32(floor(d) * dY_46_w) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(Float32(dX_46_v * dX_46_v) * t_3) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(floor(w) * dY_46_u) tmp = Float32(0.0) if (dX_46_w <= Float32(100.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, t_4), Float32(Float32(Float32(t_6 * t_6) + Float32(t_1 * t_1)) + Float32(t_2 * t_2))))); else tmp = log2(sqrt(fmax(Float32(fma(t_5, t_5, t_4) + Float32(t_0 * t_0)), fma(Float32(dY_46_w * dY_46_w), Float32(floor(d) * floor(d)), Float32(Float32(dY_46_v_m * dY_46_v_m) * t_3))))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\_m\\
t_2 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left(dX.v \cdot dX.v\right) \cdot t\_3\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := \left\lfloor w\right\rfloor \cdot dY.u\\
\mathbf{if}\;dX.w \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, t\_4\right), \left(t\_6 \cdot t\_6 + t\_1 \cdot t\_1\right) + t\_2 \cdot t\_2\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5, t\_5, t\_4\right) + t\_0 \cdot t\_0, \mathsf{fma}\left(dY.w \cdot dY.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_3\right)\right)}\right)\\
\end{array}
\end{array}
if dX.w < 100Initial program 67.5%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.4
Applied rewrites60.4%
Taylor expanded in dX.w around 0
pow-prod-downN/A
*-commutativeN/A
exp-to-powN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f32N/A
Applied rewrites60.4%
if 100 < dX.w 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.f3261.0
Applied rewrites61.0%
lift-+.f32N/A
lift-*.f32N/A
lower-fma.f3261.0
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f3261.0
Applied rewrites61.0%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor d) dY.w))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v_m))
(t_4 (* (floor h) (floor h))))
(if (<= dX.v 2000000.0)
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* (* (floor w) (floor w)) (* dX.u dX.u)))
(+ (+ (* t_2 t_2) (* t_3 t_3)) (* t_1 t_1)))))
(log2
(sqrt
(fmax
(fma
(* (* (floor w) dX.u) (floor w))
dX.u
(fma (* dX.w dX.w) t_0 (* (* dX.v dX.v) t_4)))
(fma (* dY.w dY.w) t_0 (* (* dY.v_m dY.v_m) t_4))))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
float t_1 = floorf(d) * dY_46_w;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v_m;
float t_4 = floorf(h) * floorf(h);
float tmp;
if (dX_46_v <= 2000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u))), (((t_2 * t_2) + (t_3 * t_3)) + (t_1 * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, fmaf((dX_46_w * dX_46_w), t_0, ((dX_46_v * dX_46_v) * t_4))), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_v_m * dY_46_v_m) * t_4)))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(d) * dY_46_w) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v_m) t_4 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dX_46_v <= Float32(2000000.0)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))), Float32(Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) + Float32(t_1 * t_1))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, fma(Float32(dX_46_w * dX_46_w), t_0, Float32(Float32(dX_46_v * dX_46_v) * t_4))), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_v_m * dY_46_v_m) * t_4))))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\_m\\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.v \leq 2000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right), \left(t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right) + t\_1 \cdot t\_1\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, \mathsf{fma}\left(dX.w \cdot dX.w, t\_0, \left(dX.v \cdot dX.v\right) \cdot t\_4\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_4\right)\right)}\right)\\
\end{array}
\end{array}
if dX.v < 2e6Initial program 67.5%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.4
Applied rewrites60.4%
if 2e6 < dX.v 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.f3261.0
Applied rewrites61.0%
Applied rewrites61.0%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dY.v_m))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor d) (floor d)))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor h) (floor h)))
(t_6 (* (floor d) dY.w)))
(if (<= dX.u 2000.0)
(log2
(sqrt
(fmax
(fma t_3 (* dX.w dX.w) (* t_5 (* dX.v dX.v)))
(+ (+ (* t_2 t_2) (* t_0 t_0)) (* t_6 t_6)))))
(log2
(sqrt
(fmax
(+ (fma t_1 t_1 (* (* dX.v dX.v) t_5)) (* t_4 t_4))
(fma (* dY.w dY.w) t_3 (* (* dY.v_m dY.v_m) t_5))))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(h) * dY_46_v_m;
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(d) * dX_46_w;
float t_5 = floorf(h) * floorf(h);
float t_6 = floorf(d) * dY_46_w;
float tmp;
if (dX_46_u <= 2000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_3, (dX_46_w * dX_46_w), (t_5 * (dX_46_v * dX_46_v))), (((t_2 * t_2) + (t_0 * t_0)) + (t_6 * t_6)))));
} else {
tmp = log2f(sqrtf(fmaxf((fmaf(t_1, t_1, ((dX_46_v * dX_46_v) * t_5)) + (t_4 * t_4)), fmaf((dY_46_w * dY_46_w), t_3, ((dY_46_v_m * dY_46_v_m) * t_5)))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(h) * dY_46_v_m) 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(d) * dX_46_w) t_5 = Float32(floor(h) * floor(h)) t_6 = Float32(floor(d) * dY_46_w) tmp = Float32(0.0) if (dX_46_u <= Float32(2000.0)) tmp = log2(sqrt(fmax(fma(t_3, Float32(dX_46_w * dX_46_w), Float32(t_5 * Float32(dX_46_v * dX_46_v))), Float32(Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) + Float32(t_6 * t_6))))); else tmp = log2(sqrt(fmax(Float32(fma(t_1, t_1, Float32(Float32(dX_46_v * dX_46_v) * t_5)) + Float32(t_4 * t_4)), fma(Float32(dY_46_w * dY_46_w), t_3, Float32(Float32(dY_46_v_m * dY_46_v_m) * t_5))))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\_m\\
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 d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_6 := \left\lfloor d\right\rfloor \cdot dY.w\\
\mathbf{if}\;dX.u \leq 2000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.w \cdot dX.w, t\_5 \cdot \left(dX.v \cdot dX.v\right)\right), \left(t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right) + t\_6 \cdot t\_6\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, \left(dX.v \cdot dX.v\right) \cdot t\_5\right) + t\_4 \cdot t\_4, \mathsf{fma}\left(dY.w \cdot dY.w, t\_3, \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_5\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 2e3Initial 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-*.f3260.3
Applied rewrites60.3%
if 2e3 < 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.f3261.0
Applied rewrites61.0%
lift-+.f32N/A
lift-*.f32N/A
lower-fma.f3261.0
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f3261.0
Applied rewrites61.0%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor h) (floor h)))
(t_2 (* (* dY.v_m dY.v_m) t_1))
(t_3 (* (* dX.v dX.v) t_1))
(t_4 (* (floor w) dX.u))
(t_5 (* (floor d) dX.w)))
(if (<= dX.u 2000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 t_3)
(fma
(* dY.w dY.w)
t_0
(fma (* dY.u dY.u) (* (floor w) (floor w)) t_2)))))
(log2
(sqrt
(fmax
(+ (fma t_4 t_4 t_3) (* t_5 t_5))
(fma (* dY.w dY.w) t_0 t_2)))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
float t_1 = floorf(h) * floorf(h);
float t_2 = (dY_46_v_m * dY_46_v_m) * t_1;
float t_3 = (dX_46_v * dX_46_v) * t_1;
float t_4 = floorf(w) * dX_46_u;
float t_5 = floorf(d) * dX_46_w;
float tmp;
if (dX_46_u <= 2000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, t_3), fmaf((dY_46_w * dY_46_w), t_0, fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf((fmaf(t_4, t_4, t_3) + (t_5 * t_5)), fmaf((dY_46_w * dY_46_w), t_0, t_2))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(Float32(dY_46_v_m * dY_46_v_m) * t_1) t_3 = Float32(Float32(dX_46_v * dX_46_v) * t_1) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(floor(d) * dX_46_w) tmp = Float32(0.0) if (dX_46_u <= Float32(2000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, t_3), fma(Float32(dY_46_w * dY_46_w), t_0, fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), t_2))))); else tmp = log2(sqrt(fmax(Float32(fma(t_4, t_4, t_3) + Float32(t_5 * t_5)), fma(Float32(dY_46_w * dY_46_w), t_0, t_2)))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_1\\
t_3 := \left(dX.v \cdot dX.v\right) \cdot t\_1\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \left\lfloor d\right\rfloor \cdot dX.w\\
\mathbf{if}\;dX.u \leq 2000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, t\_3\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_2\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4, t\_4, t\_3\right) + t\_5 \cdot t\_5, \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, t\_2\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 2e3Initial 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
if 2e3 < 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.f3261.0
Applied rewrites61.0%
lift-+.f32N/A
lift-*.f32N/A
lower-fma.f3261.0
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
pow-prod-downN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f3261.0
Applied rewrites61.0%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor h) (floor h)))
(t_2 (fma (* dX.w dX.w) t_0 (* (* dX.v dX.v) t_1)))
(t_3 (* (* dY.v_m dY.v_m) t_1)))
(if (<= dX.u 2000.0)
(log2
(sqrt
(fmax
t_2
(fma
(* dY.w dY.w)
t_0
(fma (* dY.u dY.u) (* (floor w) (floor w)) t_3)))))
(log2
(sqrt
(fmax
(fma (* (* (floor w) dX.u) (floor w)) dX.u t_2)
(fma (* dY.w dY.w) t_0 t_3)))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
float t_1 = floorf(h) * floorf(h);
float t_2 = fmaf((dX_46_w * dX_46_w), t_0, ((dX_46_v * dX_46_v) * t_1));
float t_3 = (dY_46_v_m * dY_46_v_m) * t_1;
float tmp;
if (dX_46_u <= 2000.0f) {
tmp = log2f(sqrtf(fmaxf(t_2, fmaf((dY_46_w * dY_46_w), t_0, fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), t_3)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, t_2), fmaf((dY_46_w * dY_46_w), t_0, t_3))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(h) * floor(h)) t_2 = fma(Float32(dX_46_w * dX_46_w), t_0, Float32(Float32(dX_46_v * dX_46_v) * t_1)) t_3 = Float32(Float32(dY_46_v_m * dY_46_v_m) * t_1) tmp = Float32(0.0) if (dX_46_u <= Float32(2000.0)) tmp = log2(sqrt(fmax(t_2, fma(Float32(dY_46_w * dY_46_w), t_0, fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), t_3))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, t_2), fma(Float32(dY_46_w * dY_46_w), t_0, t_3)))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(dX.w \cdot dX.w, t\_0, \left(dX.v \cdot dX.v\right) \cdot t\_1\right)\\
t_3 := \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_1\\
\mathbf{if}\;dX.u \leq 2000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_3\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, t\_2\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, t\_3\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 2e3Initial 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
if 2e3 < 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.f3261.0
Applied rewrites61.0%
Applied rewrites61.0%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor d) (floor d)))
(t_2 (* (floor h) (floor h))))
(if (<= dX.u 10000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_1 (* (* dX.v dX.v) t_2))
(fma
(* dY.w dY.w)
t_1
(fma (* dY.u dY.u) t_0 (* (* dY.v_m dY.v_m) t_2))))))
(log2
(sqrt
(fmax
(fma
(* (* dX.w dX.w) (floor d))
(floor d)
(fma (* (* (floor h) dX.v) (floor h)) dX.v (* t_0 (* dX.u dX.u))))
(fma (* dY.w dY.w) t_1 (* (* dY.u dY.u) t_0))))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(h) * floorf(h);
float tmp;
if (dX_46_u <= 10000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_1, ((dX_46_v * dX_46_v) * t_2)), fmaf((dY_46_w * dY_46_w), t_1, fmaf((dY_46_u * dY_46_u), t_0, ((dY_46_v_m * dY_46_v_m) * t_2))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * dX_46_w) * floorf(d)), floorf(d), fmaf(((floorf(h) * dX_46_v) * floorf(h)), dX_46_v, (t_0 * (dX_46_u * dX_46_u)))), fmaf((dY_46_w * dY_46_w), t_1, ((dY_46_u * dY_46_u) * t_0)))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dX_46_u <= Float32(10000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_1, Float32(Float32(dX_46_v * dX_46_v) * t_2)), fma(Float32(dY_46_w * dY_46_w), t_1, fma(Float32(dY_46_u * dY_46_u), t_0, Float32(Float32(dY_46_v_m * dY_46_v_m) * t_2)))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * dX_46_w) * floor(d)), floor(d), fma(Float32(Float32(floor(h) * dX_46_v) * floor(h)), dX_46_v, Float32(t_0 * Float32(dX_46_u * dX_46_u)))), fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_u * dY_46_u) * t_0))))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.u \leq 10000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_1, \left(dX.v \cdot dX.v\right) \cdot t\_2\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_2\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot dX.w\right) \cdot \left\lfloor d\right\rfloor , \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , dX.v, t\_0 \cdot \left(dX.u \cdot dX.u\right)\right)\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 1e4Initial 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
if 1e4 < 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.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor d) (floor d)))
(t_2 (* (floor h) (floor h))))
(if (<= dX.u 10000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_1 (* (* dX.v dX.v) t_2))
(fma
(* dY.w dY.w)
t_1
(fma (* dY.u dY.u) t_0 (* (* dY.v_m dY.v_m) t_2))))))
(log2
(sqrt
(fmax
(fma t_1 (* dX.w dX.w) (* t_0 (* dX.u dX.u)))
(fma (* (* dY.w (floor d)) (floor d)) dY.w (* (* dY.u dY.u) t_0))))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(h) * floorf(h);
float tmp;
if (dX_46_u <= 10000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_1, ((dX_46_v * dX_46_v) * t_2)), fmaf((dY_46_w * dY_46_w), t_1, fmaf((dY_46_u * dY_46_u), t_0, ((dY_46_v_m * dY_46_v_m) * t_2))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_1, (dX_46_w * dX_46_w), (t_0 * (dX_46_u * dX_46_u))), fmaf(((dY_46_w * floorf(d)) * floorf(d)), dY_46_w, ((dY_46_u * dY_46_u) * t_0)))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dX_46_u <= Float32(10000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_1, Float32(Float32(dX_46_v * dX_46_v) * t_2)), fma(Float32(dY_46_w * dY_46_w), t_1, fma(Float32(dY_46_u * dY_46_u), t_0, Float32(Float32(dY_46_v_m * dY_46_v_m) * t_2)))))); else tmp = log2(sqrt(fmax(fma(t_1, Float32(dX_46_w * dX_46_w), Float32(t_0 * Float32(dX_46_u * dX_46_u))), fma(Float32(Float32(dY_46_w * floor(d)) * floor(d)), dY_46_w, Float32(Float32(dY_46_u * dY_46_u) * t_0))))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.u \leq 10000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_1, \left(dX.v \cdot dX.v\right) \cdot t\_2\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_2\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, dX.w \cdot dX.w, t\_0 \cdot \left(dX.u \cdot dX.u\right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dY.w, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 1e4Initial 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
if 1e4 < dX.u Initial program 67.5%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.4
Applied rewrites60.4%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites52.7%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor w) (floor w)))
(t_2 (* (floor h) dY.v_m))
(t_3 (* (floor d) (floor d)))
(t_4 (* (floor d) dY.w)))
(if (<= dX.u 10000.0)
(log2
(sqrt
(fmax
(* (* dX.w dX.w) t_3)
(+ (+ (* t_0 t_0) (* t_2 t_2)) (* t_4 t_4)))))
(log2
(sqrt
(fmax
(fma t_3 (* dX.w dX.w) (* t_1 (* dX.u dX.u)))
(fma (* (* dY.w (floor d)) (floor d)) dY.w (* (* dY.u dY.u) t_1))))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(w) * floorf(w);
float t_2 = floorf(h) * dY_46_v_m;
float t_3 = floorf(d) * floorf(d);
float t_4 = floorf(d) * dY_46_w;
float tmp;
if (dX_46_u <= 10000.0f) {
tmp = log2f(sqrtf(fmaxf(((dX_46_w * dX_46_w) * t_3), (((t_0 * t_0) + (t_2 * t_2)) + (t_4 * t_4)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(t_3, (dX_46_w * dX_46_w), (t_1 * (dX_46_u * dX_46_u))), fmaf(((dY_46_w * floorf(d)) * floorf(d)), dY_46_w, ((dY_46_u * dY_46_u) * t_1)))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(w) * floor(w)) t_2 = Float32(floor(h) * dY_46_v_m) t_3 = Float32(floor(d) * floor(d)) t_4 = Float32(floor(d) * dY_46_w) tmp = Float32(0.0) if (dX_46_u <= Float32(10000.0)) tmp = log2(sqrt(fmax(Float32(Float32(dX_46_w * dX_46_w) * t_3), Float32(Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) + Float32(t_4 * t_4))))); else tmp = log2(sqrt(fmax(fma(t_3, Float32(dX_46_w * dX_46_w), Float32(t_1 * Float32(dX_46_u * dX_46_u))), fma(Float32(Float32(dY_46_w * floor(d)) * floor(d)), dY_46_w, Float32(Float32(dY_46_u * dY_46_u) * t_1))))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\_m\\
t_3 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_4 := \left\lfloor d\right\rfloor \cdot dY.w\\
\mathbf{if}\;dX.u \leq 10000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(dX.w \cdot dX.w\right) \cdot t\_3, \left(t\_0 \cdot t\_0 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.w \cdot dX.w, t\_1 \cdot \left(dX.u \cdot dX.u\right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dY.w, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 1e4Initial program 67.5%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.4
Applied rewrites60.4%
Taylor expanded in dX.u around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f3253.3
Applied rewrites53.3%
if 1e4 < dX.u Initial program 67.5%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.4
Applied rewrites60.4%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites52.7%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor w) (floor w)))
(t_2 (* (floor h) (floor h))))
(if (<= dY.v_m 0.5)
(log2
(sqrt
(fmax
(fma t_0 (* dX.w dX.w) (* t_1 (* dX.u dX.u)))
(fma (* (* dY.w (floor d)) (floor d)) dY.w (* (* dY.u dY.u) t_1)))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* (* dX.v dX.v) t_2))
(fma (* dY.w dY.w) t_0 (* (* dY.v_m dY.v_m) t_2))))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
float t_1 = floorf(w) * floorf(w);
float t_2 = floorf(h) * floorf(h);
float tmp;
if (dY_46_v_m <= 0.5f) {
tmp = log2f(sqrtf(fmaxf(fmaf(t_0, (dX_46_w * dX_46_w), (t_1 * (dX_46_u * dX_46_u))), fmaf(((dY_46_w * floorf(d)) * floorf(d)), dY_46_w, ((dY_46_u * dY_46_u) * t_1)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, ((dX_46_v * dX_46_v) * t_2)), fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_v_m * dY_46_v_m) * t_2)))));
}
return tmp;
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(w) * floor(w)) t_2 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dY_46_v_m <= Float32(0.5)) tmp = log2(sqrt(fmax(fma(t_0, Float32(dX_46_w * dX_46_w), Float32(t_1 * Float32(dX_46_u * dX_46_u))), fma(Float32(Float32(dY_46_w * floor(d)) * floor(d)), dY_46_w, Float32(Float32(dY_46_u * dY_46_u) * t_1))))); else tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(Float32(dX_46_v * dX_46_v) * t_2)), fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_v_m * dY_46_v_m) * t_2))))); end return tmp end
\begin{array}{l}
dY.v_m = \left|dY.v\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 \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dY.v\_m \leq 0.5:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.w \cdot dX.w, t\_1 \cdot \left(dX.u \cdot dX.u\right)\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dY.w, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, \left(dX.v \cdot dX.v\right) \cdot t\_2\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_2\right)\right)}\right)\\
\end{array}
\end{array}
if dY.v < 0.5Initial program 67.5%
Taylor expanded in dX.v around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3260.4
Applied rewrites60.4%
Taylor expanded in dY.v around 0
+-commutativeN/A
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites52.7%
if 0.5 < dY.v Initial 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
Taylor expanded in dY.u around 0
Applied rewrites53.4%
dY.v_m = (fabs.f32 dY.v)
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v_m dY.w)
:precision binary32
(let* ((t_0 (* (floor h) (floor h))) (t_1 (* (floor d) (floor d))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_1 (* (* dX.v dX.v) t_0))
(fma (* dY.w dY.w) t_1 (* (* dY.v_m dY.v_m) t_0)))))))dY.v_m = fabs(dY_46_v);
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_m, float dY_46_w) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(d) * floorf(d);
return log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_1, ((dX_46_v * dX_46_v) * t_0)), fmaf((dY_46_w * dY_46_w), t_1, ((dY_46_v_m * dY_46_v_m) * t_0)))));
}
dY.v_m = abs(dY_46_v) function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v_m, dY_46_w) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(d) * floor(d)) return log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_1, Float32(Float32(dX_46_v * dX_46_v) * t_0)), fma(Float32(dY_46_w * dY_46_w), t_1, Float32(Float32(dY_46_v_m * dY_46_v_m) * t_0))))) end
\begin{array}{l}
dY.v_m = \left|dY.v\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 \\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_1, \left(dX.v \cdot dX.v\right) \cdot t\_0\right), \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, \left(dY.v\_m \cdot dY.v\_m\right) \cdot t\_0\right)\right)}\right)
\end{array}
\end{array}
Initial 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
Taylor expanded in dY.u around 0
Applied rewrites53.4%
herbie shell --seed 2025127
(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)))))))