
(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 10 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 (sqrt (floor w)))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor d) dY.w))
(t_3 (* (floor d) dX.w))
(t_4 (* t_3 t_3))
(t_5 (* (floor w) dX.u))
(t_6 (* t_5 t_5))
(t_7 (* (floor w) dY.u))
(t_8 (* t_0 t_0))
(t_9 (* t_8 dX.u))
(t_10 (* t_1 t_1))
(t_11 (* (floor h) dY.v))
(t_12 (* t_11 t_11))
(t_13 (* t_2 t_2))
(t_14 (* t_8 dY.u)))
(if (<=
(log2 (sqrt (fmax (+ (+ t_6 t_10) t_4) (+ (+ (* t_7 t_7) t_12) t_13))))
100.0)
(log2
(sqrt
(fmax (+ (+ (* t_9 t_9) t_10) t_4) (+ (+ (* t_14 t_14) t_12) t_13))))
(log2
(sqrt
(fmax
(+ (+ t_6 (exp (fma 2.0 (log (- (floor h))) (* 2.0 (log dX.v))))) t_4)
(fma
(* dY.w dY.w)
(* (floor d) (floor d))
(* (* dY.u dY.u) (* (floor w) (floor 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 = sqrtf(floorf(w));
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(d) * dY_46_w;
float t_3 = floorf(d) * dX_46_w;
float t_4 = t_3 * t_3;
float t_5 = floorf(w) * dX_46_u;
float t_6 = t_5 * t_5;
float t_7 = floorf(w) * dY_46_u;
float t_8 = t_0 * t_0;
float t_9 = t_8 * dX_46_u;
float t_10 = t_1 * t_1;
float t_11 = floorf(h) * dY_46_v;
float t_12 = t_11 * t_11;
float t_13 = t_2 * t_2;
float t_14 = t_8 * dY_46_u;
float tmp;
if (log2f(sqrtf(fmaxf(((t_6 + t_10) + t_4), (((t_7 * t_7) + t_12) + t_13)))) <= 100.0f) {
tmp = log2f(sqrtf(fmaxf((((t_9 * t_9) + t_10) + t_4), (((t_14 * t_14) + t_12) + t_13))));
} else {
tmp = log2f(sqrtf(fmaxf(((t_6 + expf(fmaf(2.0f, logf(-floorf(h)), (2.0f * logf(dX_46_v))))) + t_4), fmaf((dY_46_w * dY_46_w), (floorf(d) * floorf(d)), ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
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 = sqrt(floor(w)) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(d) * dY_46_w) t_3 = Float32(floor(d) * dX_46_w) t_4 = Float32(t_3 * t_3) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(t_5 * t_5) t_7 = Float32(floor(w) * dY_46_u) t_8 = Float32(t_0 * t_0) t_9 = Float32(t_8 * dX_46_u) t_10 = Float32(t_1 * t_1) t_11 = Float32(floor(h) * dY_46_v) t_12 = Float32(t_11 * t_11) t_13 = Float32(t_2 * t_2) t_14 = Float32(t_8 * dY_46_u) tmp = Float32(0.0) if (log2(sqrt(fmax(Float32(Float32(t_6 + t_10) + t_4), Float32(Float32(Float32(t_7 * t_7) + t_12) + t_13)))) <= Float32(100.0)) tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_9 * t_9) + t_10) + t_4), Float32(Float32(Float32(t_14 * t_14) + t_12) + t_13)))); else tmp = log2(sqrt(fmax(Float32(Float32(t_6 + exp(fma(Float32(2.0), log(Float32(-floor(h))), Float32(Float32(2.0) * log(dX_46_v))))) + t_4), fma(Float32(dY_46_w * dY_46_w), Float32(floor(d) * floor(d)), Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \sqrt{\left\lfloor w\right\rfloor }\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_3 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_4 := t\_3 \cdot t\_3\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := t\_5 \cdot t\_5\\
t_7 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_8 := t\_0 \cdot t\_0\\
t_9 := t\_8 \cdot dX.u\\
t_10 := t\_1 \cdot t\_1\\
t_11 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_12 := t\_11 \cdot t\_11\\
t_13 := t\_2 \cdot t\_2\\
t_14 := t\_8 \cdot dY.u\\
\mathbf{if}\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_6 + t\_10\right) + t\_4, \left(t\_7 \cdot t\_7 + t\_12\right) + t\_13\right)}\right) \leq 100:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_9 \cdot t\_9 + t\_10\right) + t\_4, \left(t\_14 \cdot t\_14 + t\_12\right) + t\_13\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_6 + e^{\mathsf{fma}\left(2, \log \left(-\left\lfloor h\right\rfloor \right), 2 \cdot \log dX.v\right)}\right) + t\_4, \mathsf{fma}\left(dY.w \cdot dY.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \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 67.5%
lift-floor.f32N/A
unpow1N/A
metadata-evalN/A
sqr-powN/A
lower-*.f32N/A
metadata-evalN/A
metadata-evalN/A
lower-pow.f32N/A
lift-floor.f32N/A
metadata-evalN/A
metadata-evalN/A
lower-pow.f32N/A
lift-floor.f3267.5
Applied rewrites67.5%
lift-floor.f32N/A
unpow1N/A
metadata-evalN/A
sqr-powN/A
lower-*.f32N/A
metadata-evalN/A
metadata-evalN/A
lower-pow.f32N/A
lift-floor.f32N/A
metadata-evalN/A
metadata-evalN/A
lower-pow.f32N/A
lift-floor.f3267.5
Applied rewrites67.5%
lift-floor.f32N/A
unpow1N/A
metadata-evalN/A
sqr-powN/A
lower-*.f32N/A
metadata-evalN/A
metadata-evalN/A
lower-pow.f32N/A
lift-floor.f32N/A
metadata-evalN/A
metadata-evalN/A
lower-pow.f32N/A
lift-floor.f3267.5
Applied rewrites67.5%
lift-floor.f32N/A
unpow1N/A
metadata-evalN/A
sqr-powN/A
lower-*.f32N/A
metadata-evalN/A
metadata-evalN/A
lower-pow.f32N/A
lift-floor.f32N/A
metadata-evalN/A
metadata-evalN/A
lower-pow.f32N/A
lift-floor.f3267.5
Applied rewrites67.5%
lift-floor.f32N/A
lift-pow.f32N/A
pow1/2N/A
lower-sqrt.f32N/A
lift-floor.f3267.5
lift-floor.f32N/A
lift-pow.f32N/A
pow1/2N/A
lower-sqrt.f32N/A
lift-floor.f3267.5
Applied rewrites67.5%
lift-floor.f32N/A
lift-pow.f32N/A
pow1/2N/A
lower-sqrt.f32N/A
lift-floor.f3267.5
lift-floor.f32N/A
lift-pow.f32N/A
pow1/2N/A
lower-sqrt.f32N/A
lift-floor.f3267.5
Applied rewrites67.5%
lift-floor.f32N/A
lift-pow.f32N/A
pow1/2N/A
lower-sqrt.f32N/A
lift-floor.f3267.5
lift-floor.f32N/A
lift-pow.f32N/A
pow1/2N/A
lower-sqrt.f32N/A
lift-floor.f3267.5
Applied rewrites67.5%
lift-floor.f32N/A
lift-pow.f32N/A
pow1/2N/A
lower-sqrt.f32N/A
lift-floor.f3267.5
lift-floor.f32N/A
lift-pow.f32N/A
pow1/2N/A
lower-sqrt.f32N/A
lift-floor.f3267.5
Applied rewrites67.5%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 67.5%
Taylor expanded in dY.v around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3248.0
Applied rewrites48.0%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
log-prodN/A
+-commutativeN/A
*-commutativeN/A
distribute-lft-inN/A
log-pow-revN/A
log-pow-revN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
log-prodN/A
*-commutativeN/A
log-prodN/A
Applied rewrites36.1%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* t_0 t_0))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor d) dY.w))
(t_6 (* (floor d) dX.w))
(t_7 (* t_6 t_6))
(t_8
(log2
(sqrt
(fmax
(+ (+ t_1 (* t_4 t_4)) t_7)
(+ (+ (* t_2 t_2) (* t_3 t_3)) (* t_5 t_5)))))))
(if (<= t_8 100.0)
t_8
(log2
(sqrt
(fmax
(+ (+ t_1 (exp (fma 2.0 (log (- (floor h))) (* 2.0 (log dX.v))))) t_7)
(fma
(* dY.w dY.w)
(* (floor d) (floor d))
(* (* dY.u dY.u) (* (floor w) (floor 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(w) * dX_46_u;
float t_1 = t_0 * t_0;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(d) * dY_46_w;
float t_6 = floorf(d) * dX_46_w;
float t_7 = t_6 * t_6;
float t_8 = log2f(sqrtf(fmaxf(((t_1 + (t_4 * t_4)) + t_7), (((t_2 * t_2) + (t_3 * t_3)) + (t_5 * t_5)))));
float tmp;
if (t_8 <= 100.0f) {
tmp = t_8;
} else {
tmp = log2f(sqrtf(fmaxf(((t_1 + expf(fmaf(2.0f, logf(-floorf(h)), (2.0f * logf(dX_46_v))))) + t_7), fmaf((dY_46_w * dY_46_w), (floorf(d) * floorf(d)), ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))))));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(t_0 * t_0) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(d) * dY_46_w) t_6 = Float32(floor(d) * dX_46_w) t_7 = Float32(t_6 * t_6) t_8 = log2(sqrt(fmax(Float32(Float32(t_1 + Float32(t_4 * t_4)) + t_7), Float32(Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) + Float32(t_5 * t_5))))) tmp = Float32(0.0) if (t_8 <= Float32(100.0)) tmp = t_8; else tmp = log2(sqrt(fmax(Float32(Float32(t_1 + exp(fma(Float32(2.0), log(Float32(-floor(h))), Float32(Float32(2.0) * log(dX_46_v))))) + t_7), fma(Float32(dY_46_w * dY_46_w), Float32(floor(d) * floor(d)), Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := t\_0 \cdot t\_0\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor d\right\rfloor \cdot dY.w\\
t_6 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_7 := t\_6 \cdot t\_6\\
t_8 := \log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 + t\_4 \cdot t\_4\right) + t\_7, \left(t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right) + t\_5 \cdot t\_5\right)}\right)\\
\mathbf{if}\;t\_8 \leq 100:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 + e^{\mathsf{fma}\left(2, \log \left(-\left\lfloor h\right\rfloor \right), 2 \cdot \log dX.v\right)}\right) + t\_7, \mathsf{fma}\left(dY.w \cdot dY.w, \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor , \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 67.5%
if 100 < (log2.f32 (sqrt.f32 (fmax.f32 (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (*.f32 (*.f32 (floor.f32 d) dX.w) (*.f32 (floor.f32 d) dX.w))) (+.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))) (*.f32 (*.f32 (floor.f32 d) dY.w) (*.f32 (floor.f32 d) dY.w)))))) Initial program 67.5%
Taylor expanded in dY.v around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3248.0
Applied rewrites48.0%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
log-prodN/A
+-commutativeN/A
*-commutativeN/A
distribute-lft-inN/A
log-pow-revN/A
log-pow-revN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
log-prodN/A
*-commutativeN/A
log-prodN/A
Applied rewrites36.1%
(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}
Initial program 67.5%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor d) (floor d)))
(t_2 (* (floor d) dX.w))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor h) (floor h)))
(t_5 (* (* dY.v dY.v) t_4)))
(if (<= dY.u 200000.0)
(log2
(sqrt
(fmax
(+ (+ (* t_0 t_0) (* t_3 t_3)) (* t_2 t_2))
(fma (* dY.w dY.w) t_1 t_5))))
(log2
(sqrt
(fmax
(fma
(* t_0 (floor w))
dX.u
(fma t_4 (* dX.v dX.v) (* (* dX.w dX.w) t_1)))
(fma (* dY.u (* dY.u (floor w))) (floor w) t_5)))))))
float code(float w, float h, float d, float dX_46_u, float dX_46_v, float dX_46_w, float dY_46_u, float dY_46_v, float dY_46_w) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(d) * floorf(d);
float t_2 = floorf(d) * dX_46_w;
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(h) * floorf(h);
float t_5 = (dY_46_v * dY_46_v) * t_4;
float tmp;
if (dY_46_u <= 200000.0f) {
tmp = log2f(sqrtf(fmaxf((((t_0 * t_0) + (t_3 * t_3)) + (t_2 * t_2)), fmaf((dY_46_w * dY_46_w), t_1, t_5))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_0 * floorf(w)), dX_46_u, fmaf(t_4, (dX_46_v * dX_46_v), ((dX_46_w * dX_46_w) * t_1))), fmaf((dY_46_u * (dY_46_u * floorf(w))), floorf(w), t_5))));
}
return tmp;
}
function code(w, h, d, dX_46_u, dX_46_v, dX_46_w, dY_46_u, dY_46_v, dY_46_w) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(d) * floor(d)) t_2 = Float32(floor(d) * dX_46_w) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(h) * floor(h)) t_5 = Float32(Float32(dY_46_v * dY_46_v) * t_4) tmp = Float32(0.0) if (dY_46_u <= Float32(200000.0)) tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) + Float32(t_2 * t_2)), fma(Float32(dY_46_w * dY_46_w), t_1, t_5)))); else tmp = log2(sqrt(fmax(fma(Float32(t_0 * floor(w)), dX_46_u, fma(t_4, Float32(dX_46_v * dX_46_v), Float32(Float32(dX_46_w * dX_46_w) * t_1))), fma(Float32(dY_46_u * Float32(dY_46_u * floor(w))), floor(w), t_5)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \\
t_2 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_5 := \left(dY.v \cdot dY.v\right) \cdot t\_4\\
\mathbf{if}\;dY.u \leq 200000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_0 \cdot t\_0 + t\_3 \cdot t\_3\right) + t\_2 \cdot t\_2, \mathsf{fma}\left(dY.w \cdot dY.w, t\_1, t\_5\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot \left\lfloor w\right\rfloor , dX.u, \mathsf{fma}\left(t\_4, dX.v \cdot dX.v, \left(dX.w \cdot dX.w\right) \cdot t\_1\right)\right), \mathsf{fma}\left(dY.u \cdot \left(dY.u \cdot \left\lfloor w\right\rfloor \right), \left\lfloor w\right\rfloor , t\_5\right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 2e5Initial program 67.5%
Taylor expanded in dY.u around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.6
Applied rewrites60.6%
if 2e5 < dY.u Initial program 67.5%
Taylor expanded in dY.w around inf
Applied rewrites53.1%
Taylor expanded in dY.u around inf
lower-/.f32N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3245.5
Applied rewrites45.5%
Applied rewrites45.5%
Taylor expanded in dY.w around 0
Applied rewrites60.1%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* dY.w (floor d)))
(t_1 (* (floor h) (floor h)))
(t_2 (* (* dY.v dY.v) t_1))
(t_3 (* (floor w) dX.u)))
(if (<= dY.u 200000.0)
(log2
(sqrt
(fmax
(fma
(* (* dX.w (floor d)) (floor d))
dX.w
(fma (* t_1 dX.v) dX.v (* t_3 t_3)))
(fma t_0 t_0 t_2))))
(log2
(sqrt
(fmax
(fma
(* t_3 (floor w))
dX.u
(fma t_1 (* dX.v dX.v) (* (* dX.w dX.w) (* (floor d) (floor d)))))
(fma (* dY.u (* dY.u (floor w))) (floor w) 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 = dY_46_w * floorf(d);
float t_1 = floorf(h) * floorf(h);
float t_2 = (dY_46_v * dY_46_v) * t_1;
float t_3 = floorf(w) * dX_46_u;
float tmp;
if (dY_46_u <= 200000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf(((dX_46_w * floorf(d)) * floorf(d)), dX_46_w, fmaf((t_1 * dX_46_v), dX_46_v, (t_3 * t_3))), fmaf(t_0, t_0, t_2))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_3 * floorf(w)), dX_46_u, fmaf(t_1, (dX_46_v * dX_46_v), ((dX_46_w * dX_46_w) * (floorf(d) * floorf(d))))), fmaf((dY_46_u * (dY_46_u * floorf(w))), floorf(w), 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(dY_46_w * floor(d)) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(Float32(dY_46_v * dY_46_v) * t_1) t_3 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (dY_46_u <= Float32(200000.0)) tmp = log2(sqrt(fmax(fma(Float32(Float32(dX_46_w * floor(d)) * floor(d)), dX_46_w, fma(Float32(t_1 * dX_46_v), dX_46_v, Float32(t_3 * t_3))), fma(t_0, t_0, t_2)))); else tmp = log2(sqrt(fmax(fma(Float32(t_3 * floor(w)), dX_46_u, fma(t_1, Float32(dX_46_v * dX_46_v), Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d))))), fma(Float32(dY_46_u * Float32(dY_46_u * floor(w))), floor(w), t_2)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.w \cdot \left\lfloor d\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;dY.u \leq 200000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dX.w, \mathsf{fma}\left(t\_1 \cdot dX.v, dX.v, t\_3 \cdot t\_3\right)\right), \mathsf{fma}\left(t\_0, t\_0, t\_2\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3 \cdot \left\lfloor w\right\rfloor , dX.u, \mathsf{fma}\left(t\_1, dX.v \cdot dX.v, \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right)\right), \mathsf{fma}\left(dY.u \cdot \left(dY.u \cdot \left\lfloor w\right\rfloor \right), \left\lfloor w\right\rfloor , t\_2\right)\right)}\right)\\
\end{array}
\end{array}
if dY.u < 2e5Initial program 67.5%
Taylor expanded in dY.u around 0
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3260.6
Applied rewrites60.6%
Applied rewrites60.6%
if 2e5 < dY.u Initial program 67.5%
Taylor expanded in dY.w around inf
Applied rewrites53.1%
Taylor expanded in dY.u around inf
lower-/.f32N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3245.5
Applied rewrites45.5%
Applied rewrites45.5%
Taylor expanded in dY.w around 0
Applied rewrites60.1%
(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 h) (floor h)))
(t_2 (* (* dY.v dY.v) t_1))
(t_3 (* (floor w) dX.u)))
(if (<= dX.u 30000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* t_1 (* dX.v dX.v)))
(fma
(* (* dY.w dY.w) (floor d))
(floor d)
(fma (* (* (floor w) dY.u) (floor w)) dY.u t_2)))))
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* t_3 t_3))
(fma
(* (* dY.w (floor d)) (floor d))
dY.w
(fma (* dY.u (* dY.u (floor w))) (floor w) 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(h) * floorf(h);
float t_2 = (dY_46_v * dY_46_v) * t_1;
float t_3 = floorf(w) * dX_46_u;
float tmp;
if (dX_46_u <= 30000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, (t_1 * (dX_46_v * dX_46_v))), fmaf(((dY_46_w * dY_46_w) * floorf(d)), floorf(d), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, (t_3 * t_3)), fmaf(((dY_46_w * floorf(d)) * floorf(d)), dY_46_w, fmaf((dY_46_u * (dY_46_u * floorf(w))), floorf(w), 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(h) * floor(h)) t_2 = Float32(Float32(dY_46_v * dY_46_v) * t_1) t_3 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (dX_46_u <= Float32(30000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(t_1 * Float32(dX_46_v * dX_46_v))), fma(Float32(Float32(dY_46_w * dY_46_w) * floor(d)), floor(d), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, t_2))))); else tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(t_3 * t_3)), fma(Float32(Float32(dY_46_w * floor(d)) * floor(d)), dY_46_w, fma(Float32(dY_46_u * Float32(dY_46_u * floor(w))), floor(w), 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 h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;dX.u \leq 30000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(\left(dY.w \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor , \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, t\_2\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, t\_3 \cdot t\_3\right), \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot \left\lfloor d\right\rfloor , dY.w, \mathsf{fma}\left(dY.u \cdot \left(dY.u \cdot \left\lfloor w\right\rfloor \right), \left\lfloor w\right\rfloor , t\_2\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 3e4Initial 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-*.f3260.4
Applied rewrites60.4%
Applied rewrites60.4%
if 3e4 < 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites59.8%
Applied rewrites60.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) (floor d)))
(t_1 (* (floor h) (floor h)))
(t_2 (* (* dY.v dY.v) t_1)))
(if (<= dX.u 30000.0)
(log2
(sqrt
(fmax
(fma (* dX.w dX.w) t_0 (* t_1 (* dX.v dX.v)))
(fma
(* (* dY.w dY.w) (floor d))
(floor d)
(fma (* (* (floor w) dY.u) (floor w)) dY.u t_2)))))
(log2
(sqrt
(fmax
(fma (* dX.u dX.u) (* (floor w) (floor w)) (* (* dX.w dX.w) t_0))
(fma
(* (* dY.u dY.u) (floor w))
(floor w)
(fma (* (* dY.w (floor d)) dY.w) (floor d) 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(h) * floorf(h);
float t_2 = (dY_46_v * dY_46_v) * t_1;
float tmp;
if (dX_46_u <= 30000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_w * dX_46_w), t_0, (t_1 * (dX_46_v * dX_46_v))), fmaf(((dY_46_w * dY_46_w) * floorf(d)), floorf(d), fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, t_2)))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), (floorf(w) * floorf(w)), ((dX_46_w * dX_46_w) * t_0)), fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), 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(h) * floor(h)) t_2 = Float32(Float32(dY_46_v * dY_46_v) * t_1) tmp = Float32(0.0) if (dX_46_u <= Float32(30000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_w * dX_46_w), t_0, Float32(t_1 * Float32(dX_46_v * dX_46_v))), fma(Float32(Float32(dY_46_w * dY_46_w) * floor(d)), floor(d), fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, t_2))))); else tmp = log2(sqrt(fmax(fma(Float32(dX_46_u * dX_46_u), Float32(floor(w) * floor(w)), Float32(Float32(dX_46_w * dX_46_w) * t_0)), fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), 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 h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
\mathbf{if}\;dX.u \leq 30000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.w \cdot dX.w, t\_0, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(\left(dY.w \cdot dY.w\right) \cdot \left\lfloor d\right\rfloor , \left\lfloor d\right\rfloor , \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, t\_2\right)\right)\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \left(dX.w \cdot dX.w\right) \cdot t\_0\right), \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , t\_2\right)\right)\right)}\right)\\
\end{array}
\end{array}
if dX.u < 3e4Initial 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-*.f3260.4
Applied rewrites60.4%
Applied rewrites60.4%
if 3e4 < 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (* dX.w dX.w) (* (floor d) (floor d))))
(t_1 (* (floor h) (floor h))))
(if (<= dX.v 4000000.0)
(log2
(sqrt
(fmax
(fma (* dX.u dX.u) (* (floor w) (floor w)) t_0)
(fma
(* (* dY.u dY.u) (floor w))
(floor w)
(fma (* (* dY.w (floor d)) dY.w) (floor d) (* (* dY.v dY.v) t_1))))))
(log2
(sqrt
(fmax
(fma (* (* (floor w) dX.u) (floor w)) dX.u (fma t_1 (* dX.v dX.v) t_0))
(* (* dY.u (* dY.u (floor w))) (floor 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 = (dX_46_w * dX_46_w) * (floorf(d) * floorf(d));
float t_1 = floorf(h) * floorf(h);
float tmp;
if (dX_46_v <= 4000000.0f) {
tmp = log2f(sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), (floorf(w) * floorf(w)), t_0), fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), fmaf(((dY_46_w * floorf(d)) * dY_46_w), floorf(d), ((dY_46_v * dY_46_v) * t_1))))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, fmaf(t_1, (dX_46_v * dX_46_v), t_0)), ((dY_46_u * (dY_46_u * floorf(w))) * floorf(w)))));
}
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(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d))) t_1 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (dX_46_v <= Float32(4000000.0)) tmp = log2(sqrt(fmax(fma(Float32(dX_46_u * dX_46_u), Float32(floor(w) * floor(w)), t_0), fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), fma(Float32(Float32(dY_46_w * floor(d)) * dY_46_w), floor(d), Float32(Float32(dY_46_v * dY_46_v) * t_1)))))); else tmp = log2(sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, fma(t_1, Float32(dX_46_v * dX_46_v), t_0)), Float32(Float32(dY_46_u * Float32(dY_46_u * floor(w))) * floor(w))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.v \leq 4000000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_0\right), \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \mathsf{fma}\left(\left(dY.w \cdot \left\lfloor d\right\rfloor \right) \cdot dY.w, \left\lfloor d\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot t\_1\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, \mathsf{fma}\left(t\_1, dX.v \cdot dX.v, t\_0\right)\right), \left(dY.u \cdot \left(dY.u \cdot \left\lfloor w\right\rfloor \right)\right) \cdot \left\lfloor w\right\rfloor \right)}\right)\\
\end{array}
\end{array}
if dX.v < 4e6Initial 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-*.f3260.3
Applied rewrites60.3%
Applied rewrites60.3%
if 4e6 < dX.v Initial program 67.5%
Taylor expanded in dY.w around inf
Applied rewrites53.1%
Taylor expanded in dY.u around inf
lower-/.f32N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3245.5
Applied rewrites45.5%
Applied rewrites45.5%
Taylor expanded in dY.u around inf
Applied rewrites53.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(let* ((t_0 (* (floor d) dX.w))
(t_1 (* (floor w) dX.u))
(t_2 (* dY.w (floor d))))
(if (<= dY.u 12350000.0)
(log2
(sqrt
(fmax
(+ (+ (* t_1 t_1) (exp (* (log (* (floor h) dX.v)) 2.0))) (* t_0 t_0))
(* t_2 t_2))))
(log2
(sqrt
(fmax
(fma
(* t_1 (floor w))
dX.u
(fma
(* (floor h) (floor h))
(* dX.v dX.v)
(* (* dX.w dX.w) (* (floor d) (floor d)))))
(* (* dY.u (* dY.u (floor w))) (floor 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(d) * dX_46_w;
float t_1 = floorf(w) * dX_46_u;
float t_2 = dY_46_w * floorf(d);
float tmp;
if (dY_46_u <= 12350000.0f) {
tmp = log2f(sqrtf(fmaxf((((t_1 * t_1) + expf((logf((floorf(h) * dX_46_v)) * 2.0f))) + (t_0 * t_0)), (t_2 * t_2))));
} else {
tmp = log2f(sqrtf(fmaxf(fmaf((t_1 * floorf(w)), dX_46_u, fmaf((floorf(h) * floorf(h)), (dX_46_v * dX_46_v), ((dX_46_w * dX_46_w) * (floorf(d) * floorf(d))))), ((dY_46_u * (dY_46_u * floorf(w))) * floorf(w)))));
}
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) * dX_46_w) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(dY_46_w * floor(d)) tmp = Float32(0.0) if (dY_46_u <= Float32(12350000.0)) tmp = log2(sqrt(fmax(Float32(Float32(Float32(t_1 * t_1) + exp(Float32(log(Float32(floor(h) * dX_46_v)) * Float32(2.0)))) + Float32(t_0 * t_0)), Float32(t_2 * t_2)))); else tmp = log2(sqrt(fmax(fma(Float32(t_1 * floor(w)), dX_46_u, fma(Float32(floor(h) * floor(h)), Float32(dX_46_v * dX_46_v), Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d))))), Float32(Float32(dY_46_u * Float32(dY_46_u * floor(w))) * floor(w))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor d\right\rfloor \cdot dX.w\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := dY.w \cdot \left\lfloor d\right\rfloor \\
\mathbf{if}\;dY.u \leq 12350000:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\left(t\_1 \cdot t\_1 + e^{\log \left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot 2}\right) + t\_0 \cdot t\_0, t\_2 \cdot t\_2\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \left(\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot \left\lfloor w\right\rfloor , dX.u, \mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , dX.v \cdot dX.v, \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right)\right), \left(dY.u \cdot \left(dY.u \cdot \left\lfloor w\right\rfloor \right)\right) \cdot \left\lfloor w\right\rfloor \right)}\right)\\
\end{array}
\end{array}
if dY.u < 1.235e7Initial 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.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3248.0
Applied rewrites48.0%
Taylor expanded in dY.u around 0
pow-prod-downN/A
pow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f3237.9
Applied rewrites37.9%
if 1.235e7 < dY.u Initial program 67.5%
Taylor expanded in dY.w around inf
Applied rewrites53.1%
Taylor expanded in dY.u around inf
lower-/.f32N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3245.5
Applied rewrites45.5%
Applied rewrites45.5%
Taylor expanded in dY.u around inf
Applied rewrites53.6%
(FPCore (w h d dX.u dX.v dX.w dY.u dY.v dY.w)
:precision binary32
(log2
(sqrt
(fmax
(fma
(* (* (floor w) dX.u) (floor w))
dX.u
(fma
(* (floor h) (floor h))
(* dX.v dX.v)
(* (* dX.w dX.w) (* (floor d) (floor d)))))
(* (* dY.u (* dY.u (floor w))) (floor 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(w) * dX_46_u) * floorf(w)), dX_46_u, fmaf((floorf(h) * floorf(h)), (dX_46_v * dX_46_v), ((dX_46_w * dX_46_w) * (floorf(d) * floorf(d))))), ((dY_46_u * (dY_46_u * floorf(w))) * floorf(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(w) * dX_46_u) * floor(w)), dX_46_u, fma(Float32(floor(h) * floor(h)), Float32(dX_46_v * dX_46_v), Float32(Float32(dX_46_w * dX_46_w) * Float32(floor(d) * floor(d))))), Float32(Float32(dY_46_u * Float32(dY_46_u * floor(w))) * floor(w))))) end
\begin{array}{l}
\\
\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(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , dX.v \cdot dX.v, \left(dX.w \cdot dX.w\right) \cdot \left(\left\lfloor d\right\rfloor \cdot \left\lfloor d\right\rfloor \right)\right)\right), \left(dY.u \cdot \left(dY.u \cdot \left\lfloor w\right\rfloor \right)\right) \cdot \left\lfloor w\right\rfloor \right)}\right)
\end{array}
Initial program 67.5%
Taylor expanded in dY.w around inf
Applied rewrites53.1%
Taylor expanded in dY.u around inf
lower-/.f32N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
lift-*.f3245.5
Applied rewrites45.5%
Applied rewrites45.5%
Taylor expanded in dY.u around inf
Applied rewrites53.6%
herbie shell --seed 2025139
(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)))))))