
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
Herbie found 6 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor d) dY.w))
(t_4 (* (floor d) dX.w))
(t_5 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_5 t_5) (* t_2 t_2)) (* t_4 t_4))
(+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(d) * dY_46_w;
float t_4 = floorf(d) * dX_46_w;
float t_5 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3)))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(d) * dY_46_w) t_4 = Float32(floor(d) * dX_46_w) t_5 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) + Float32(t_4 * t_4)), Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) end
function tmp = code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(d) * dY_46_w; t_4 = floor(d) * dX_46_w; t_5 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_5 * t_5) + (t_2 * t_2)) + (t_4 * t_4)), (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_4 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_5 \cdot t\_5 + t\_2 \cdot t\_2\right) + t\_4 \cdot t\_4, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right)
\end{array}
\end{array}
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(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))
(t_6 (+ (* t_5 t_5) (* t_2 t_2)))
(t_7 (+ t_6 (* t_4 t_4)))
(t_8
(fma
(* dY.w dY.w)
(* (floor d) (floor d))
(* (* dY.v dY.v) (* (floor h) (floor h))))))
(if (<=
(log2 (sqrt (fmax t_7 (+ (+ (* t_0 t_0) (* t_1 t_1)) (* t_3 t_3)))))
63.900001525878906)
(log2 (sqrt (fmax t_7 (fma (* (* (floor w) (floor w)) dY.u) dY.u t_8))))
(log2
(sqrt
(fmax
(+ t_6 (exp (fma 2.0 (log (- (floor d))) (* 2.0 (log dX.w)))))
t_8))))))
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;
float t_6 = (t_5 * t_5) + (t_2 * t_2);
float t_7 = t_6 + (t_4 * t_4);
float t_8 = fmaf((dY_46_w * dY_46_w), (floorf(d) * floorf(d)), ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))));
float tmp;
if (log2f(sqrtf(fmaxf(t_7, (((t_0 * t_0) + (t_1 * t_1)) + (t_3 * t_3))))) <= 63.900001525878906f) {
tmp = log2f(sqrtf(fmaxf(t_7, fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, t_8))));
} else {
tmp = log2f(sqrtf(fmaxf((t_6 + expf(fmaf(2.0f, logf(-floorf(d)), (2.0f * logf(dX_46_w))))), t_8)));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * 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) t_6 = Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) t_7 = Float32(t_6 + Float32(t_4 * t_4)) t_8 = fma(Float32(dY_46_w * dY_46_w), Float32(floor(d) * floor(d)), Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) tmp = Float32(0.0) if (log2(sqrt(fmax(t_7, Float32(Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) + Float32(t_3 * t_3))))) <= Float32(63.900001525878906)) tmp = log2(sqrt(fmax(t_7, fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, t_8)))); else tmp = log2(sqrt(fmax(Float32(t_6 + exp(fma(Float32(2.0), log(Float32(-floor(d))), Float32(Float32(2.0) * log(dX_46_w))))), t_8))); end return tmp 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\\
t_6 := t\_5 \cdot t\_5 + t\_2 \cdot t\_2\\
t_7 := t\_6 + t\_4 \cdot t\_4\\
t_8 := \mathsf{fma}\left(dY.w \cdot dY.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_7, \left(t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right) + t\_3 \cdot t\_3\right)}\right) \leq 63.900001525878906:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_7, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_8\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(t\_6 + e^{\mathsf{fma}\left(2, \log \left(-\left\lfloor d\right\rfloor \right), 2 \cdot \log dX.w\right)}, t\_8\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)))))) < 63.9000015Initial program 67.9%
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 rewrites67.9%
if 63.9000015 < (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 67.9%
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
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3248.6
Applied rewrites48.6%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
*-commutativeN/A
log-pow-revN/A
pow-prod-downN/A
*-commutativeN/A
log-prodN/A
pow2N/A
sqr-neg-revN/A
pow2N/A
log-pow-revN/A
mul-1-negN/A
log-pow-revN/A
lower-fma.f32N/A
Applied rewrites36.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor d) dX.w))
(t_2 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_2 t_2) (* t_0 t_0)) (* t_1 t_1))
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(fma
(* dY.w dY.w)
(* (floor d) (floor d))
(* (* dY.v dY.v) (* (floor h) (floor h))))))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(d) * dX_46_w;
float t_2 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_2 * t_2) + (t_0 * t_0)) + (t_1 * t_1)), fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, fmaf((dY_46_w * dY_46_w), (floorf(d) * floorf(d)), ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(d) * dX_46_w) t_2 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) + Float32(t_1 * t_1)), fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, fma(Float32(dY_46_w * dY_46_w), Float32(floor(d) * floor(d)), Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right) + t\_1 \cdot t\_1, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \mathsf{fma}\left(dY.w \cdot dY.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \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)
\end{array}
\end{array}
Initial program 67.9%
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 rewrites67.9%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor w) (floor w)))
(t_2 (* (floor h) (floor h)))
(t_3 (* (floor d) dY.w)))
(if (<= dX.u 10.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* t_2 (* dX.v dX.v)))
(fma
(* (* (floor w) dY.u) (floor w))
dY.u
(fma (* (* (floor h) dY.v) (floor h)) dY.v (* t_3 t_3))))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* t_1 (* dX.u dX.u)))
(fma
(* t_1 dY.u)
dY.u
(fma (* dY.w dY.w) t_0 (* (* dY.v dY.v) t_2)))))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * floorf(d);
float t_1 = floorf(w) * floorf(w);
float t_2 = floorf(h) * floorf(h);
float t_3 = floorf(d) * dY_46_w;
float tmp;
if (dX_46_u <= 10.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, (t_2 * (dX_46_v * dX_46_v))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, fmaf(((floorf(h) * dY_46_v) * floorf(h)), dY_46_v, (t_3 * t_3))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, (t_1 * (dX_46_u * dX_46_u))), fmaf((t_1 * dY_46_u), dY_46_u, fmaf((dY_46_w * dY_46_w), t_0, ((dY_46_v * dY_46_v) * t_2))))));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * floor(d)) t_1 = Float32(floor(w) * floor(w)) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(floor(d) * dY_46_w) tmp = Float32(0.0) if (dX_46_u <= Float32(10.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(t_2 * Float32(dX_46_v * dX_46_v))), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, fma(Float32(Float32(floor(h) * dY_46_v) * floor(h)), dY_46_v, Float32(t_3 * t_3)))))); else tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(t_1 * Float32(dX_46_u * dX_46_u))), fma(Float32(t_1 * dY_46_u), dY_46_u, fma(Float32(dY_46_w * dY_46_w), t_0, Float32(Float32(dY_46_v * dY_46_v) * t_2)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor d\right\rfloor \cdot dY.w\\
\mathbf{if}\;dX.u \leq 10:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, t\_2 \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 h\right\rfloor \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , dY.v, t\_3 \cdot t\_3\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, t\_1 \cdot \left(dX.u \cdot dX.u\right)\right), \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \mathsf{fma}\left(dY.w \cdot dY.w, t\_0, \left(dY.v \cdot dY.v\right) \cdot t\_2\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 10Initial program 67.9%
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-*.f3260.6
Applied rewrites60.6%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3253.6
Applied rewrites53.6%
Applied rewrites60.6%
if 10 < dX.u Initial program 67.9%
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 rewrites67.9%
Taylor expanded in dX.v around 0
+-commutativeN/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
pow2N/A
*-commutativeN/A
lower-fma.f32N/A
pow2N/A
lift-*.f32N/A
Applied rewrites61.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dY.w)))
(log2
(sqrt
(fmax
(fma
(* dX.w dX.w)
(* (floor d) (floor d))
(* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma
(* (* (floor w) dY.u) (floor w))
dY.u
(fma (* (* (floor h) dY.v) (floor h)) dY.v (* t_0 t_0))))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(d) * dY_46_w;
return log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), (floorf(d) * floorf(d)), ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, fmaf(((floorf(h) * dY_46_v) * floorf(h)), dY_46_v, (t_0 * t_0))))));
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(d) * dY_46_w) return log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), Float32(floor(d) * floor(d)), 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(h) * dY_46_v) * floor(h)), dY_46_v, Float32(t_0 * t_0)))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dY.w\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \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 h\right\rfloor \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , dY.v, t\_0 \cdot t\_0\right)\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.9%
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-*.f3260.6
Applied rewrites60.6%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3253.6
Applied rewrites53.6%
Applied rewrites60.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor d) dX.w))
(t_2 (* (floor w) dX.u)))
(log2
(sqrt
(fmax
(+ (+ (* t_2 t_2) (* t_0 t_0)) (* t_1 t_1))
(*
(/ (* (* dY.u dY.u) (* (floor w) (floor w))) (* dY.w dY.w))
(* dY.w dY.w)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(d) * dX_46_w;
float t_2 = floorf(w) * dX_46_u;
return log2f(sqrtf(fmaxf((((t_2 * t_2) + (t_0 * t_0)) + (t_1 * t_1)), ((((dY_46_u * dY_46_u) * (floorf(w) * floorf(w))) / (dY_46_w * dY_46_w)) * (dY_46_w * 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) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(d) * dX_46_w) t_2 = Float32(floor(w) * dX_46_u) return log2(sqrt(fmax(Float32(Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) + Float32(t_1 * t_1)), Float32(Float32(Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))) / Float32(dY_46_w * dY_46_w)) * Float32(dY_46_w * dY_46_w))))) 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(h) * dX_46_v; t_1 = floor(d) * dX_46_w; t_2 = floor(w) * dX_46_u; tmp = log2(sqrt(max((((t_2 * t_2) + (t_0 * t_0)) + (t_1 * t_1)), ((((dY_46_u * dY_46_u) * (floor(w) * floor(w))) / (dY_46_w * dY_46_w)) * (dY_46_w * dY_46_w))))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right) + t\_1 \cdot t\_1, \frac{\left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)}{dY.w \cdot dY.w} \cdot \left(dY.w \cdot dY.w\right)\right)}\right)
\end{array}
\end{array}
Initial program 67.9%
Taylor expanded in dY.w around inf
Applied rewrites53.8%
Taylor expanded in dY.u around inf
pow2N/A
lower-/.f32N/A
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3245.6
Applied rewrites45.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(fma
(* (* (floor d) dX.w) (floor d))
dX.w
(fma
(* (* (floor h) dX.v) (floor h))
dX.v
(* (* (* (floor w) dX.u) (floor w)) dX.u)))
(*
(/ (* (* dY.u dY.u) (* (floor w) (floor w))) (* dY.w dY.w))
(* dY.w dY.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) {
return 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, (((floorf(w) * dX_46_u) * floorf(w)) * dX_46_u))), ((((dY_46_u * dY_46_u) * (floorf(w) * floorf(w))) / (dY_46_w * dY_46_w)) * (dY_46_w * 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) return 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(Float32(floor(w) * dX_46_u) * floor(w)) * dX_46_u))), Float32(Float32(Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))) / Float32(dY_46_w * dY_46_w)) * Float32(dY_46_w * dY_46_w))))) end
\begin{array}{l}
\\
\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(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right)\right), \frac{\left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)}{dY.w \cdot dY.w} \cdot \left(dY.w \cdot dY.w\right)\right)}\right)
\end{array}
Initial program 67.9%
Taylor expanded in dY.w around inf
Applied rewrites53.8%
Taylor expanded in dY.u around inf
pow2N/A
lower-/.f32N/A
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3245.6
Applied rewrites45.6%
Applied rewrites45.6%
herbie shell --seed 2025137
(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)))))))