
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_0) (* t_6 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_0;
} else {
tmp = t_6 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_0); else tmp = Float32(t_6 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_0; else tmp = t_6 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 11 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_0) (* t_6 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_0;
} else {
tmp = t_6 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_0); else tmp = Float32(t_6 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_0; else tmp = t_6 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dX.u))
(t_2 (pow t_1 2.0))
(t_3 (pow t_0 2.0))
(t_4 (* (floor w) dY.u_m))
(t_5 (- t_2 t_3))
(t_6 (+ (* t_1 t_1) (* t_0 t_0)))
(t_7 (* (floor h) dY.v))
(t_8 (* t_7 t_7))
(t_9 (+ (* t_4 t_4) t_8)))
(if (>= t_6 t_9)
(*
t_0
(/
1.0
(sqrt
(fmax
(* (+ t_2 t_3) (* t_5 (/ 1.0 t_5)))
(+
t_8
(*
(pow (exp 2.0) (/ (* 2.0 (* (log t_4) 0.5)) 2.0))
(pow (exp 2.0) (/ (* 2.0 (log (sqrt t_4))) 2.0))))))))
(* t_7 (/ 1.0 (sqrt (fmax t_6 t_9)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = powf(t_1, 2.0f);
float t_3 = powf(t_0, 2.0f);
float t_4 = floorf(w) * dY_46_u_m;
float t_5 = t_2 - t_3;
float t_6 = (t_1 * t_1) + (t_0 * t_0);
float t_7 = floorf(h) * dY_46_v;
float t_8 = t_7 * t_7;
float t_9 = (t_4 * t_4) + t_8;
float tmp;
if (t_6 >= t_9) {
tmp = t_0 * (1.0f / sqrtf(fmaxf(((t_2 + t_3) * (t_5 * (1.0f / t_5))), (t_8 + (powf(expf(2.0f), ((2.0f * (logf(t_4) * 0.5f)) / 2.0f)) * powf(expf(2.0f), ((2.0f * logf(sqrtf(t_4))) / 2.0f)))))));
} else {
tmp = t_7 * (1.0f / sqrtf(fmaxf(t_6, t_9)));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = t_1 ^ Float32(2.0) t_3 = t_0 ^ Float32(2.0) t_4 = Float32(floor(w) * dY_46_u_m) t_5 = Float32(t_2 - t_3) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) t_7 = Float32(floor(h) * dY_46_v) t_8 = Float32(t_7 * t_7) t_9 = Float32(Float32(t_4 * t_4) + t_8) tmp = Float32(0.0) if (t_6 >= t_9) tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))) != Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5)))) ? Float32(t_8 + Float32((exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * Float32(log(t_4) * Float32(0.5))) / Float32(2.0))) * (exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * log(sqrt(t_4))) / Float32(2.0))))) : ((Float32(t_8 + Float32((exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * Float32(log(t_4) * Float32(0.5))) / Float32(2.0))) * (exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * log(sqrt(t_4))) / Float32(2.0))))) != Float32(t_8 + Float32((exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * Float32(log(t_4) * Float32(0.5))) / Float32(2.0))) * (exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * log(sqrt(t_4))) / Float32(2.0)))))) ? Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))) : max(Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))), Float32(t_8 + Float32((exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * Float32(log(t_4) * Float32(0.5))) / Float32(2.0))) * (exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * log(sqrt(t_4))) / Float32(2.0))))))))))); else tmp = Float32(t_7 * Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? t_9 : ((t_9 != t_9) ? t_6 : max(t_6, t_9)))))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dX_46_u; t_2 = t_1 ^ single(2.0); t_3 = t_0 ^ single(2.0); t_4 = floor(w) * dY_46_u_m; t_5 = t_2 - t_3; t_6 = (t_1 * t_1) + (t_0 * t_0); t_7 = floor(h) * dY_46_v; t_8 = t_7 * t_7; t_9 = (t_4 * t_4) + t_8; tmp = single(0.0); if (t_6 >= t_9) tmp = t_0 * (single(1.0) / sqrt(max(((t_2 + t_3) * (t_5 * (single(1.0) / t_5))), (t_8 + ((exp(single(2.0)) ^ ((single(2.0) * (log(t_4) * single(0.5))) / single(2.0))) * (exp(single(2.0)) ^ ((single(2.0) * log(sqrt(t_4))) / single(2.0)))))))); else tmp = t_7 * (single(1.0) / sqrt(max(t_6, t_9))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := {t\_1}^{2}\\
t_3 := {t\_0}^{2}\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_5 := t\_2 - t\_3\\
t_6 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_8 := t\_7 \cdot t\_7\\
t_9 := t\_4 \cdot t\_4 + t\_8\\
\mathbf{if}\;t\_6 \geq t\_9:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(\left(t\_2 + t\_3\right) \cdot \left(t\_5 \cdot \frac{1}{t\_5}\right), t\_8 + {\left(e^{2}\right)}^{\left(\frac{2 \cdot \left(\log t\_4 \cdot 0.5\right)}{2}\right)} \cdot {\left(e^{2}\right)}^{\left(\frac{2 \cdot \log \left(\sqrt{t\_4}\right)}{2}\right)}\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_9\right)}}\\
\end{array}
\end{array}
Initial program 77.4%
lift-+.f32N/A
flip-+N/A
div-invN/A
difference-of-squaresN/A
lift-+.f32N/A
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
*-commutativeN/A
exp-prodN/A
unpow1N/A
pow-to-expN/A
rem-log-expN/A
lower-pow.f32N/A
lower-exp.f32N/A
rem-log-expN/A
pow-to-expN/A
unpow1N/A
lower-log.f3277.3
Applied rewrites77.3%
lift-pow.f32N/A
sqr-powN/A
pow-sqrN/A
sqr-powN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-/.f32N/A
lower-*.f32N/A
div-invN/A
metadata-evalN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-/.f32N/A
Applied rewrites77.3%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
metadata-evalN/A
lift-log.f32N/A
unpow1N/A
log-powN/A
metadata-evalN/A
lift-log.f32N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-/r/N/A
clear-numN/A
lift-/.f32N/A
rem-log-expN/A
Applied rewrites77.3%
Final simplification77.3%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (pow t_0 2.0))
(t_2 (* (floor w) dY.u_m))
(t_3 (pow (exp 2.0) (/ (* 2.0 (* (log t_2) 0.5)) 2.0)))
(t_4 (* (floor w) dX.u))
(t_5 (pow t_4 2.0))
(t_6 (- t_5 t_1))
(t_7 (+ t_5 t_1))
(t_8 (* (floor h) dY.v))
(t_9 (* t_8 t_8)))
(if (>= t_7 (+ (pow t_2 2.0) (pow t_8 2.0)))
(*
t_0
(/ 1.0 (sqrt (fmax (* t_7 (* t_6 (/ 1.0 t_6))) (+ t_9 (* t_3 t_3))))))
(*
t_8
(/ 1.0 (sqrt (fmax (+ (* t_4 t_4) (* t_0 t_0)) (+ (* t_2 t_2) t_9))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = powf(t_0, 2.0f);
float t_2 = floorf(w) * dY_46_u_m;
float t_3 = powf(expf(2.0f), ((2.0f * (logf(t_2) * 0.5f)) / 2.0f));
float t_4 = floorf(w) * dX_46_u;
float t_5 = powf(t_4, 2.0f);
float t_6 = t_5 - t_1;
float t_7 = t_5 + t_1;
float t_8 = floorf(h) * dY_46_v;
float t_9 = t_8 * t_8;
float tmp;
if (t_7 >= (powf(t_2, 2.0f) + powf(t_8, 2.0f))) {
tmp = t_0 * (1.0f / sqrtf(fmaxf((t_7 * (t_6 * (1.0f / t_6))), (t_9 + (t_3 * t_3)))));
} else {
tmp = t_8 * (1.0f / sqrtf(fmaxf(((t_4 * t_4) + (t_0 * t_0)), ((t_2 * t_2) + t_9))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = t_0 ^ Float32(2.0) t_2 = Float32(floor(w) * dY_46_u_m) t_3 = exp(Float32(2.0)) ^ Float32(Float32(Float32(2.0) * Float32(log(t_2) * Float32(0.5))) / Float32(2.0)) t_4 = Float32(floor(w) * dX_46_u) t_5 = t_4 ^ Float32(2.0) t_6 = Float32(t_5 - t_1) t_7 = Float32(t_5 + t_1) t_8 = Float32(floor(h) * dY_46_v) t_9 = Float32(t_8 * t_8) tmp = Float32(0.0) if (t_7 >= Float32((t_2 ^ Float32(2.0)) + (t_8 ^ Float32(2.0)))) tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(t_7 * Float32(t_6 * Float32(Float32(1.0) / t_6))) != Float32(t_7 * Float32(t_6 * Float32(Float32(1.0) / t_6)))) ? Float32(t_9 + Float32(t_3 * t_3)) : ((Float32(t_9 + Float32(t_3 * t_3)) != Float32(t_9 + Float32(t_3 * t_3))) ? Float32(t_7 * Float32(t_6 * Float32(Float32(1.0) / t_6))) : max(Float32(t_7 * Float32(t_6 * Float32(Float32(1.0) / t_6))), Float32(t_9 + Float32(t_3 * t_3)))))))); else tmp = Float32(t_8 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) != Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0))) ? Float32(Float32(t_2 * t_2) + t_9) : ((Float32(Float32(t_2 * t_2) + t_9) != Float32(Float32(t_2 * t_2) + t_9)) ? Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) : max(Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)), Float32(Float32(t_2 * t_2) + t_9))))))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = t_0 ^ single(2.0); t_2 = floor(w) * dY_46_u_m; t_3 = exp(single(2.0)) ^ ((single(2.0) * (log(t_2) * single(0.5))) / single(2.0)); t_4 = floor(w) * dX_46_u; t_5 = t_4 ^ single(2.0); t_6 = t_5 - t_1; t_7 = t_5 + t_1; t_8 = floor(h) * dY_46_v; t_9 = t_8 * t_8; tmp = single(0.0); if (t_7 >= ((t_2 ^ single(2.0)) + (t_8 ^ single(2.0)))) tmp = t_0 * (single(1.0) / sqrt(max((t_7 * (t_6 * (single(1.0) / t_6))), (t_9 + (t_3 * t_3))))); else tmp = t_8 * (single(1.0) / sqrt(max(((t_4 * t_4) + (t_0 * t_0)), ((t_2 * t_2) + t_9)))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {t\_0}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_3 := {\left(e^{2}\right)}^{\left(\frac{2 \cdot \left(\log t\_2 \cdot 0.5\right)}{2}\right)}\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := {t\_4}^{2}\\
t_6 := t\_5 - t\_1\\
t_7 := t\_5 + t\_1\\
t_8 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_9 := t\_8 \cdot t\_8\\
\mathbf{if}\;t\_7 \geq {t\_2}^{2} + {t\_8}^{2}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_7 \cdot \left(t\_6 \cdot \frac{1}{t\_6}\right), t\_9 + t\_3 \cdot t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_0 \cdot t\_0, t\_2 \cdot t\_2 + t\_9\right)}}\\
\end{array}
\end{array}
Initial program 77.4%
lift-+.f32N/A
flip-+N/A
div-invN/A
difference-of-squaresN/A
lift-+.f32N/A
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
*-commutativeN/A
exp-prodN/A
unpow1N/A
pow-to-expN/A
rem-log-expN/A
lower-pow.f32N/A
lower-exp.f32N/A
rem-log-expN/A
pow-to-expN/A
unpow1N/A
lower-log.f3277.3
Applied rewrites77.3%
lift-pow.f32N/A
sqr-powN/A
pow-sqrN/A
sqr-powN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-/.f32N/A
lower-*.f32N/A
div-invN/A
metadata-evalN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-/.f32N/A
Applied rewrites77.3%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
+-commutativeN/A
lower-+.f3277.3
lift-*.f32N/A
pow2N/A
lift-pow.f3277.3
lift-*.f32N/A
pow2N/A
lift-pow.f3277.3
lift-*.f32N/A
pow2N/A
lift-pow.f3277.3
Applied rewrites77.3%
Final simplification77.3%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (pow t_0 2.0))
(t_2 (* (floor w) dY.u_m))
(t_3 (* (floor w) dX.u))
(t_4 (pow t_3 2.0))
(t_5 (- t_4 t_1))
(t_6 (+ t_4 t_1))
(t_7 (* (floor h) dY.v))
(t_8 (* t_7 t_7)))
(if (>= t_6 (+ (pow t_2 2.0) (pow t_7 2.0)))
(*
t_0
(/
1.0
(sqrt
(fmax (* t_6 (* t_5 (/ 1.0 t_5))) (+ t_8 (pow (exp 2.0) (log t_2)))))))
(*
t_7
(/ 1.0 (sqrt (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_2 t_2) t_8))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = powf(t_0, 2.0f);
float t_2 = floorf(w) * dY_46_u_m;
float t_3 = floorf(w) * dX_46_u;
float t_4 = powf(t_3, 2.0f);
float t_5 = t_4 - t_1;
float t_6 = t_4 + t_1;
float t_7 = floorf(h) * dY_46_v;
float t_8 = t_7 * t_7;
float tmp;
if (t_6 >= (powf(t_2, 2.0f) + powf(t_7, 2.0f))) {
tmp = t_0 * (1.0f / sqrtf(fmaxf((t_6 * (t_5 * (1.0f / t_5))), (t_8 + powf(expf(2.0f), logf(t_2))))));
} else {
tmp = t_7 * (1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + t_8))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = t_0 ^ Float32(2.0) t_2 = Float32(floor(w) * dY_46_u_m) t_3 = Float32(floor(w) * dX_46_u) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(t_4 - t_1) t_6 = Float32(t_4 + t_1) t_7 = Float32(floor(h) * dY_46_v) t_8 = Float32(t_7 * t_7) tmp = Float32(0.0) if (t_6 >= Float32((t_2 ^ Float32(2.0)) + (t_7 ^ Float32(2.0)))) tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(t_6 * Float32(t_5 * Float32(Float32(1.0) / t_5))) != Float32(t_6 * Float32(t_5 * Float32(Float32(1.0) / t_5)))) ? Float32(t_8 + (exp(Float32(2.0)) ^ log(t_2))) : ((Float32(t_8 + (exp(Float32(2.0)) ^ log(t_2))) != Float32(t_8 + (exp(Float32(2.0)) ^ log(t_2)))) ? Float32(t_6 * Float32(t_5 * Float32(Float32(1.0) / t_5))) : max(Float32(t_6 * Float32(t_5 * Float32(Float32(1.0) / t_5))), Float32(t_8 + (exp(Float32(2.0)) ^ log(t_2))))))))); else tmp = Float32(t_7 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) != Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) ? Float32(Float32(t_2 * t_2) + t_8) : ((Float32(Float32(t_2 * t_2) + t_8) != Float32(Float32(t_2 * t_2) + t_8)) ? Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) : max(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_2 * t_2) + t_8))))))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = t_0 ^ single(2.0); t_2 = floor(w) * dY_46_u_m; t_3 = floor(w) * dX_46_u; t_4 = t_3 ^ single(2.0); t_5 = t_4 - t_1; t_6 = t_4 + t_1; t_7 = floor(h) * dY_46_v; t_8 = t_7 * t_7; tmp = single(0.0); if (t_6 >= ((t_2 ^ single(2.0)) + (t_7 ^ single(2.0)))) tmp = t_0 * (single(1.0) / sqrt(max((t_6 * (t_5 * (single(1.0) / t_5))), (t_8 + (exp(single(2.0)) ^ log(t_2)))))); else tmp = t_7 * (single(1.0) / sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + t_8)))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {t\_0}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := {t\_3}^{2}\\
t_5 := t\_4 - t\_1\\
t_6 := t\_4 + t\_1\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_8 := t\_7 \cdot t\_7\\
\mathbf{if}\;t\_6 \geq {t\_2}^{2} + {t\_7}^{2}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_6 \cdot \left(t\_5 \cdot \frac{1}{t\_5}\right), t\_8 + {\left(e^{2}\right)}^{\log t\_2}\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_2 \cdot t\_2 + t\_8\right)}}\\
\end{array}
\end{array}
Initial program 77.4%
lift-+.f32N/A
flip-+N/A
div-invN/A
difference-of-squaresN/A
lift-+.f32N/A
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
*-commutativeN/A
exp-prodN/A
unpow1N/A
pow-to-expN/A
rem-log-expN/A
lower-pow.f32N/A
lower-exp.f32N/A
rem-log-expN/A
pow-to-expN/A
unpow1N/A
lower-log.f3277.3
Applied rewrites77.3%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
+-commutativeN/A
lower-+.f3277.3
lift-*.f32N/A
pow2N/A
lift-pow.f3277.3
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
+-commutativeN/A
Applied rewrites77.3%
Final simplification77.3%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dX.u))
(t_2 (pow t_1 2.0))
(t_3 (pow t_0 2.0))
(t_4 (* (floor w) dY.u_m))
(t_5 (- t_2 t_3))
(t_6 (+ (* t_1 t_1) (* t_0 t_0)))
(t_7 (* (floor h) dY.v))
(t_8 (* t_7 t_7))
(t_9 (+ (* t_4 t_4) t_8)))
(if (>= t_6 t_9)
(*
t_0
(/
1.0
(sqrt
(fmax
(* (+ t_2 t_3) (* t_5 (/ 1.0 t_5)))
(+ t_8 (exp (* 2.0 (log t_4))))))))
(* t_7 (/ 1.0 (sqrt (fmax t_6 t_9)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = powf(t_1, 2.0f);
float t_3 = powf(t_0, 2.0f);
float t_4 = floorf(w) * dY_46_u_m;
float t_5 = t_2 - t_3;
float t_6 = (t_1 * t_1) + (t_0 * t_0);
float t_7 = floorf(h) * dY_46_v;
float t_8 = t_7 * t_7;
float t_9 = (t_4 * t_4) + t_8;
float tmp;
if (t_6 >= t_9) {
tmp = t_0 * (1.0f / sqrtf(fmaxf(((t_2 + t_3) * (t_5 * (1.0f / t_5))), (t_8 + expf((2.0f * logf(t_4)))))));
} else {
tmp = t_7 * (1.0f / sqrtf(fmaxf(t_6, t_9)));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = t_1 ^ Float32(2.0) t_3 = t_0 ^ Float32(2.0) t_4 = Float32(floor(w) * dY_46_u_m) t_5 = Float32(t_2 - t_3) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) t_7 = Float32(floor(h) * dY_46_v) t_8 = Float32(t_7 * t_7) t_9 = Float32(Float32(t_4 * t_4) + t_8) tmp = Float32(0.0) if (t_6 >= t_9) tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))) != Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5)))) ? Float32(t_8 + exp(Float32(Float32(2.0) * log(t_4)))) : ((Float32(t_8 + exp(Float32(Float32(2.0) * log(t_4)))) != Float32(t_8 + exp(Float32(Float32(2.0) * log(t_4))))) ? Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))) : max(Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))), Float32(t_8 + exp(Float32(Float32(2.0) * log(t_4)))))))))); else tmp = Float32(t_7 * Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? t_9 : ((t_9 != t_9) ? t_6 : max(t_6, t_9)))))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dX_46_u; t_2 = t_1 ^ single(2.0); t_3 = t_0 ^ single(2.0); t_4 = floor(w) * dY_46_u_m; t_5 = t_2 - t_3; t_6 = (t_1 * t_1) + (t_0 * t_0); t_7 = floor(h) * dY_46_v; t_8 = t_7 * t_7; t_9 = (t_4 * t_4) + t_8; tmp = single(0.0); if (t_6 >= t_9) tmp = t_0 * (single(1.0) / sqrt(max(((t_2 + t_3) * (t_5 * (single(1.0) / t_5))), (t_8 + exp((single(2.0) * log(t_4))))))); else tmp = t_7 * (single(1.0) / sqrt(max(t_6, t_9))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := {t\_1}^{2}\\
t_3 := {t\_0}^{2}\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_5 := t\_2 - t\_3\\
t_6 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_8 := t\_7 \cdot t\_7\\
t_9 := t\_4 \cdot t\_4 + t\_8\\
\mathbf{if}\;t\_6 \geq t\_9:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(\left(t\_2 + t\_3\right) \cdot \left(t\_5 \cdot \frac{1}{t\_5}\right), t\_8 + e^{2 \cdot \log t\_4}\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_9\right)}}\\
\end{array}
\end{array}
Initial program 77.4%
lift-+.f32N/A
flip-+N/A
div-invN/A
difference-of-squaresN/A
lift-+.f32N/A
Applied rewrites78.0%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
*-commutativeN/A
unpow1N/A
pow-to-expN/A
rem-log-expN/A
lower-*.f32N/A
rem-log-expN/A
pow-to-expN/A
unpow1N/A
lower-log.f3277.3
Applied rewrites77.3%
Final simplification77.3%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dX.u))
(t_2 (pow t_1 2.0))
(t_3 (pow t_0 2.0))
(t_4 (* (floor w) dY.u_m))
(t_5 (- t_2 t_3))
(t_6 (+ (* t_1 t_1) (* t_0 t_0)))
(t_7 (* (floor h) dY.v))
(t_8 (+ (* t_4 t_4) (* t_7 t_7))))
(if (>= t_6 t_8)
(* t_0 (/ 1.0 (sqrt (fmax (* (+ t_2 t_3) (* t_5 (/ 1.0 t_5))) t_8))))
(* t_7 (/ 1.0 (sqrt (fmax t_6 t_8)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = powf(t_1, 2.0f);
float t_3 = powf(t_0, 2.0f);
float t_4 = floorf(w) * dY_46_u_m;
float t_5 = t_2 - t_3;
float t_6 = (t_1 * t_1) + (t_0 * t_0);
float t_7 = floorf(h) * dY_46_v;
float t_8 = (t_4 * t_4) + (t_7 * t_7);
float tmp;
if (t_6 >= t_8) {
tmp = t_0 * (1.0f / sqrtf(fmaxf(((t_2 + t_3) * (t_5 * (1.0f / t_5))), t_8)));
} else {
tmp = t_7 * (1.0f / sqrtf(fmaxf(t_6, t_8)));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = t_1 ^ Float32(2.0) t_3 = t_0 ^ Float32(2.0) t_4 = Float32(floor(w) * dY_46_u_m) t_5 = Float32(t_2 - t_3) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) t_7 = Float32(floor(h) * dY_46_v) t_8 = Float32(Float32(t_4 * t_4) + Float32(t_7 * t_7)) tmp = Float32(0.0) if (t_6 >= t_8) tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))) != Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5)))) ? t_8 : ((t_8 != t_8) ? Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))) : max(Float32(Float32(t_2 + t_3) * Float32(t_5 * Float32(Float32(1.0) / t_5))), t_8)))))); else tmp = Float32(t_7 * Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? t_8 : ((t_8 != t_8) ? t_6 : max(t_6, t_8)))))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dX_46_u; t_2 = t_1 ^ single(2.0); t_3 = t_0 ^ single(2.0); t_4 = floor(w) * dY_46_u_m; t_5 = t_2 - t_3; t_6 = (t_1 * t_1) + (t_0 * t_0); t_7 = floor(h) * dY_46_v; t_8 = (t_4 * t_4) + (t_7 * t_7); tmp = single(0.0); if (t_6 >= t_8) tmp = t_0 * (single(1.0) / sqrt(max(((t_2 + t_3) * (t_5 * (single(1.0) / t_5))), t_8))); else tmp = t_7 * (single(1.0) / sqrt(max(t_6, t_8))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := {t\_1}^{2}\\
t_3 := {t\_0}^{2}\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_5 := t\_2 - t\_3\\
t_6 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_8 := t\_4 \cdot t\_4 + t\_7 \cdot t\_7\\
\mathbf{if}\;t\_6 \geq t\_8:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(\left(t\_2 + t\_3\right) \cdot \left(t\_5 \cdot \frac{1}{t\_5}\right), t\_8\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_8\right)}}\\
\end{array}
\end{array}
Initial program 77.4%
lift-+.f32N/A
flip-+N/A
div-invN/A
difference-of-squaresN/A
lift-+.f32N/A
Applied rewrites78.0%
Final simplification78.0%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u_m))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4))))
(if (>= t_3 t_5)
(* t_0 (/ 1.0 (sqrt (fmax t_3 t_5))))
(/
1.0
(/
(sqrt
(fmax (+ (pow t_2 2.0) (pow t_0 2.0)) (+ (pow t_1 2.0) (pow t_4 2.0))))
t_4)))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u_m;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float tmp;
if (t_3 >= t_5) {
tmp = t_0 * (1.0f / sqrtf(fmaxf(t_3, t_5)));
} else {
tmp = 1.0f / (sqrtf(fmaxf((powf(t_2, 2.0f) + powf(t_0, 2.0f)), (powf(t_1, 2.0f) + powf(t_4, 2.0f)))) / t_4);
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u_m) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5)))))); else tmp = Float32(Float32(1.0) / Float32(sqrt(((Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? Float32((t_1 ^ Float32(2.0)) + (t_4 ^ Float32(2.0))) : ((Float32((t_1 ^ Float32(2.0)) + (t_4 ^ Float32(2.0))) != Float32((t_1 ^ Float32(2.0)) + (t_4 ^ Float32(2.0)))) ? Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : max(Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))), Float32((t_1 ^ Float32(2.0)) + (t_4 ^ Float32(2.0))))))) / t_4)); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u_m; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); tmp = single(0.0); if (t_3 >= t_5) tmp = t_0 * (single(1.0) / sqrt(max(t_3, t_5))); else tmp = single(1.0) / (sqrt(max(((t_2 ^ single(2.0)) + (t_0 ^ single(2.0))), ((t_1 ^ single(2.0)) + (t_4 ^ single(2.0))))) / t_4); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left({t\_2}^{2} + {t\_0}^{2}, {t\_1}^{2} + {t\_4}^{2}\right)}}{t\_4}}\\
\end{array}
\end{array}
Initial program 77.4%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites77.6%
Final simplification77.6%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u_m))
(t_2 (* (floor h) dY.v))
(t_3 (+ (pow t_1 2.0) (pow t_2 2.0)))
(t_4 (* (floor w) dX.u))
(t_5 (+ (pow t_4 2.0) (pow t_0 2.0))))
(if (>= t_5 t_3)
(* (floor h) (/ dX.v (sqrt (fmax t_5 t_3))))
(*
t_2
(/
1.0
(sqrt
(fmax (+ (* t_4 t_4) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2)))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u_m;
float t_2 = floorf(h) * dY_46_v;
float t_3 = powf(t_1, 2.0f) + powf(t_2, 2.0f);
float t_4 = floorf(w) * dX_46_u;
float t_5 = powf(t_4, 2.0f) + powf(t_0, 2.0f);
float tmp;
if (t_5 >= t_3) {
tmp = floorf(h) * (dX_46_v / sqrtf(fmaxf(t_5, t_3)));
} else {
tmp = t_2 * (1.0f / sqrtf(fmaxf(((t_4 * t_4) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u_m) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) tmp = Float32(0.0) if (t_5 >= t_3) tmp = Float32(floor(h) * Float32(dX_46_v / sqrt(((t_5 != t_5) ? t_3 : ((t_3 != t_3) ? t_5 : max(t_5, t_3)))))); else tmp = Float32(t_2 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) != Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0))) ? Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) : ((Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) != Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) ? Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) : max(Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)))))))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u_m; t_2 = floor(h) * dY_46_v; t_3 = (t_1 ^ single(2.0)) + (t_2 ^ single(2.0)); t_4 = floor(w) * dX_46_u; t_5 = (t_4 ^ single(2.0)) + (t_0 ^ single(2.0)); tmp = single(0.0); if (t_5 >= t_3) tmp = floor(h) * (dX_46_v / sqrt(max(t_5, t_3))); else tmp = t_2 * (single(1.0) / sqrt(max(((t_4 * t_4) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := {t\_1}^{2} + {t\_2}^{2}\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := {t\_4}^{2} + {t\_0}^{2}\\
\mathbf{if}\;t\_5 \geq t\_3:\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dX.v}{\sqrt{\mathsf{max}\left(t\_5, t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}\\
\end{array}
\end{array}
Initial program 77.4%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
*-rgt-identityN/A
Applied rewrites77.4%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
+-commutativeN/A
lower-+.f3277.4
lift-*.f32N/A
pow2N/A
lift-pow.f3277.4
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
+-commutativeN/A
Applied rewrites77.4%
Final simplification77.4%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor w) dY.u_m))
(t_2 (* (floor h) dY.v))
(t_3 (+ (* t_1 t_1) (* t_2 t_2)))
(t_4 (* (floor h) dX.v))
(t_5 (+ (pow t_1 2.0) (pow t_2 2.0)))
(t_6 (* t_0 t_0))
(t_7 (/ 1.0 (sqrt (fmax (+ t_6 (* t_4 t_4)) t_3))))
(t_8 (* t_4 t_7)))
(if (<= dX.u 0.0017999999690800905)
(if (>= (* dX.v (* dX.v (pow (floor h) 2.0))) t_5) t_8 (* t_2 t_7))
(if (>= (pow t_0 2.0) t_5)
t_8
(* t_2 (/ 1.0 (sqrt (fmax (+ t_6 (exp (* 2.0 (log t_4)))) t_3))))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(w) * dY_46_u_m;
float t_2 = floorf(h) * dY_46_v;
float t_3 = (t_1 * t_1) + (t_2 * t_2);
float t_4 = floorf(h) * dX_46_v;
float t_5 = powf(t_1, 2.0f) + powf(t_2, 2.0f);
float t_6 = t_0 * t_0;
float t_7 = 1.0f / sqrtf(fmaxf((t_6 + (t_4 * t_4)), t_3));
float t_8 = t_4 * t_7;
float tmp_1;
if (dX_46_u <= 0.0017999999690800905f) {
float tmp_2;
if ((dX_46_v * (dX_46_v * powf(floorf(h), 2.0f))) >= t_5) {
tmp_2 = t_8;
} else {
tmp_2 = t_2 * t_7;
}
tmp_1 = tmp_2;
} else if (powf(t_0, 2.0f) >= t_5) {
tmp_1 = t_8;
} else {
tmp_1 = t_2 * (1.0f / sqrtf(fmaxf((t_6 + expf((2.0f * logf(t_4)))), t_3)));
}
return tmp_1;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(w) * dY_46_u_m) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_6 = Float32(t_0 * t_0) t_7 = Float32(Float32(1.0) / sqrt(((Float32(t_6 + Float32(t_4 * t_4)) != Float32(t_6 + Float32(t_4 * t_4))) ? t_3 : ((t_3 != t_3) ? Float32(t_6 + Float32(t_4 * t_4)) : max(Float32(t_6 + Float32(t_4 * t_4)), t_3))))) t_8 = Float32(t_4 * t_7) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(0.0017999999690800905)) tmp_2 = Float32(0.0) if (Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0)))) >= t_5) tmp_2 = t_8; else tmp_2 = Float32(t_2 * t_7); end tmp_1 = tmp_2; elseif ((t_0 ^ Float32(2.0)) >= t_5) tmp_1 = t_8; else tmp_1 = Float32(t_2 * Float32(Float32(1.0) / sqrt(((Float32(t_6 + exp(Float32(Float32(2.0) * log(t_4)))) != Float32(t_6 + exp(Float32(Float32(2.0) * log(t_4))))) ? t_3 : ((t_3 != t_3) ? Float32(t_6 + exp(Float32(Float32(2.0) * log(t_4)))) : max(Float32(t_6 + exp(Float32(Float32(2.0) * log(t_4)))), t_3)))))); end return tmp_1 end
dY.u_m = abs(dY_46_u); function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = floor(w) * dY_46_u_m; t_2 = floor(h) * dY_46_v; t_3 = (t_1 * t_1) + (t_2 * t_2); t_4 = floor(h) * dX_46_v; t_5 = (t_1 ^ single(2.0)) + (t_2 ^ single(2.0)); t_6 = t_0 * t_0; t_7 = single(1.0) / sqrt(max((t_6 + (t_4 * t_4)), t_3)); t_8 = t_4 * t_7; tmp_2 = single(0.0); if (dX_46_u <= single(0.0017999999690800905)) tmp_3 = single(0.0); if ((dX_46_v * (dX_46_v * (floor(h) ^ single(2.0)))) >= t_5) tmp_3 = t_8; else tmp_3 = t_2 * t_7; end tmp_2 = tmp_3; elseif ((t_0 ^ single(2.0)) >= t_5) tmp_2 = t_8; else tmp_2 = t_2 * (single(1.0) / sqrt(max((t_6 + exp((single(2.0) * log(t_4)))), t_3))); end tmp_4 = tmp_2; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := t\_1 \cdot t\_1 + t\_2 \cdot t\_2\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := {t\_1}^{2} + {t\_2}^{2}\\
t_6 := t\_0 \cdot t\_0\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6 + t\_4 \cdot t\_4, t\_3\right)}}\\
t_8 := t\_4 \cdot t\_7\\
\mathbf{if}\;dX.u \leq 0.0017999999690800905:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;dX.v \cdot \left(dX.v \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right) \geq t\_5:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot t\_7\\
\end{array}\\
\mathbf{elif}\;{t\_0}^{2} \geq t\_5:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_6 + e^{2 \cdot \log t\_4}, t\_3\right)}}\\
\end{array}
\end{array}
if dX.u < 0.00179999997Initial program 78.1%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3263.4
Applied rewrites63.4%
lift-*.f32N/A
pow2N/A
lift-pow.f3263.4
Applied rewrites63.4%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites63.4%
Taylor expanded in dX.u around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.7
Applied rewrites69.7%
if 0.00179999997 < dX.u Initial program 75.9%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3272.8
Applied rewrites72.8%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
*-commutativeN/A
unpow1N/A
pow-to-expN/A
rem-log-expN/A
lower-*.f32N/A
rem-log-expN/A
pow-to-expN/A
unpow1N/A
lower-log.f3272.8
Applied rewrites72.8%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-log.f3252.4
Applied rewrites52.4%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-exp.f32N/A
lift-*.f32N/A
*-commutativeN/A
Applied rewrites72.8%
Final simplification70.7%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor w) dY.u_m))
(t_4 (+ (pow t_3 2.0) (pow t_1 2.0)))
(t_5 (+ (* t_3 t_3) (* t_1 t_1)))
(t_6 (* t_2 t_2))
(t_7 (/ 1.0 (sqrt (fmax (+ (* t_0 t_0) t_6) t_5))))
(t_8 (* t_1 t_7)))
(if (<= dX.u 0.0017999999690800905)
(if (>= (* dX.v (* dX.v (pow (floor h) 2.0))) t_4) (* t_2 t_7) t_8)
(if (>= (pow t_0 2.0) t_4)
(*
t_2
(/
1.0
(sqrt (fmax (+ t_6 (* dX.u (* dX.u (pow (floor w) 2.0)))) t_5))))
t_8))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(w) * dY_46_u_m;
float t_4 = powf(t_3, 2.0f) + powf(t_1, 2.0f);
float t_5 = (t_3 * t_3) + (t_1 * t_1);
float t_6 = t_2 * t_2;
float t_7 = 1.0f / sqrtf(fmaxf(((t_0 * t_0) + t_6), t_5));
float t_8 = t_1 * t_7;
float tmp_1;
if (dX_46_u <= 0.0017999999690800905f) {
float tmp_2;
if ((dX_46_v * (dX_46_v * powf(floorf(h), 2.0f))) >= t_4) {
tmp_2 = t_2 * t_7;
} else {
tmp_2 = t_8;
}
tmp_1 = tmp_2;
} else if (powf(t_0, 2.0f) >= t_4) {
tmp_1 = t_2 * (1.0f / sqrtf(fmaxf((t_6 + (dX_46_u * (dX_46_u * powf(floorf(w), 2.0f)))), t_5)));
} else {
tmp_1 = t_8;
}
return tmp_1;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(w) * dY_46_u_m) t_4 = Float32((t_3 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_5 = Float32(Float32(t_3 * t_3) + Float32(t_1 * t_1)) t_6 = Float32(t_2 * t_2) t_7 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_0 * t_0) + t_6) != Float32(Float32(t_0 * t_0) + t_6)) ? t_5 : ((t_5 != t_5) ? Float32(Float32(t_0 * t_0) + t_6) : max(Float32(Float32(t_0 * t_0) + t_6), t_5))))) t_8 = Float32(t_1 * t_7) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(0.0017999999690800905)) tmp_2 = Float32(0.0) if (Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0)))) >= t_4) tmp_2 = Float32(t_2 * t_7); else tmp_2 = t_8; end tmp_1 = tmp_2; elseif ((t_0 ^ Float32(2.0)) >= t_4) tmp_1 = Float32(t_2 * Float32(Float32(1.0) / sqrt(((Float32(t_6 + Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0))))) != Float32(t_6 + Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))))) ? t_5 : ((t_5 != t_5) ? Float32(t_6 + Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0))))) : max(Float32(t_6 + Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0))))), t_5)))))); else tmp_1 = t_8; end return tmp_1 end
dY.u_m = abs(dY_46_u); function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(w) * dY_46_u_m; t_4 = (t_3 ^ single(2.0)) + (t_1 ^ single(2.0)); t_5 = (t_3 * t_3) + (t_1 * t_1); t_6 = t_2 * t_2; t_7 = single(1.0) / sqrt(max(((t_0 * t_0) + t_6), t_5)); t_8 = t_1 * t_7; tmp_2 = single(0.0); if (dX_46_u <= single(0.0017999999690800905)) tmp_3 = single(0.0); if ((dX_46_v * (dX_46_v * (floor(h) ^ single(2.0)))) >= t_4) tmp_3 = t_2 * t_7; else tmp_3 = t_8; end tmp_2 = tmp_3; elseif ((t_0 ^ single(2.0)) >= t_4) tmp_2 = t_2 * (single(1.0) / sqrt(max((t_6 + (dX_46_u * (dX_46_u * (floor(w) ^ single(2.0))))), t_5))); else tmp_2 = t_8; end tmp_4 = tmp_2; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_4 := {t\_3}^{2} + {t\_1}^{2}\\
t_5 := t\_3 \cdot t\_3 + t\_1 \cdot t\_1\\
t_6 := t\_2 \cdot t\_2\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_6, t\_5\right)}}\\
t_8 := t\_1 \cdot t\_7\\
\mathbf{if}\;dX.u \leq 0.0017999999690800905:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;dX.v \cdot \left(dX.v \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right) \geq t\_4:\\
\;\;\;\;t\_2 \cdot t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array}\\
\mathbf{elif}\;{t\_0}^{2} \geq t\_4:\\
\;\;\;\;t\_2 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_6 + dX.u \cdot \left(dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right), t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array}
\end{array}
if dX.u < 0.00179999997Initial program 78.1%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3263.4
Applied rewrites63.4%
lift-*.f32N/A
pow2N/A
lift-pow.f3263.4
Applied rewrites63.4%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites63.4%
Taylor expanded in dX.u around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.7
Applied rewrites69.7%
if 0.00179999997 < dX.u Initial program 75.9%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3272.8
Applied rewrites72.8%
lift-*.f32N/A
pow2N/A
lift-pow.f3272.8
Applied rewrites72.8%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites72.8%
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lift-*.f3272.8
Applied rewrites72.8%
Final simplification70.7%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u_m))
(t_1 (* (floor h) dX.v))
(t_2 (* t_1 t_1))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_0 t_0) (* t_4 t_4))))
(if (>= (pow t_3 2.0) (+ (pow t_0 2.0) (pow t_4 2.0)))
(*
t_1
(/ 1.0 (sqrt (fmax (+ t_2 (* dX.u (* dX.u (pow (floor w) 2.0)))) t_5))))
(* t_4 (/ 1.0 (sqrt (fmax (+ (* t_3 t_3) t_2) t_5)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u_m;
float t_1 = floorf(h) * dX_46_v;
float t_2 = t_1 * t_1;
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_0 * t_0) + (t_4 * t_4);
float tmp;
if (powf(t_3, 2.0f) >= (powf(t_0, 2.0f) + powf(t_4, 2.0f))) {
tmp = t_1 * (1.0f / sqrtf(fmaxf((t_2 + (dX_46_u * (dX_46_u * powf(floorf(w), 2.0f)))), t_5)));
} else {
tmp = t_4 * (1.0f / sqrtf(fmaxf(((t_3 * t_3) + t_2), t_5)));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u_m) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_4 ^ Float32(2.0)))) tmp = Float32(t_1 * Float32(Float32(1.0) / sqrt(((Float32(t_2 + Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0))))) != Float32(t_2 + Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))))) ? t_5 : ((t_5 != t_5) ? Float32(t_2 + Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0))))) : max(Float32(t_2 + Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0))))), t_5)))))); else tmp = Float32(t_4 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + t_2) != Float32(Float32(t_3 * t_3) + t_2)) ? t_5 : ((t_5 != t_5) ? Float32(Float32(t_3 * t_3) + t_2) : max(Float32(Float32(t_3 * t_3) + t_2), t_5)))))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(w) * dY_46_u_m; t_1 = floor(h) * dX_46_v; t_2 = t_1 * t_1; t_3 = floor(w) * dX_46_u; t_4 = floor(h) * dY_46_v; t_5 = (t_0 * t_0) + (t_4 * t_4); tmp = single(0.0); if ((t_3 ^ single(2.0)) >= ((t_0 ^ single(2.0)) + (t_4 ^ single(2.0)))) tmp = t_1 * (single(1.0) / sqrt(max((t_2 + (dX_46_u * (dX_46_u * (floor(w) ^ single(2.0))))), t_5))); else tmp = t_4 * (single(1.0) / sqrt(max(((t_3 * t_3) + t_2), t_5))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_0 \cdot t\_0 + t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_0}^{2} + {t\_4}^{2}:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_2 + dX.u \cdot \left(dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right), t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_2, t\_5\right)}}\\
\end{array}
\end{array}
Initial program 77.4%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.5
Applied rewrites66.5%
lift-*.f32N/A
pow2N/A
lift-pow.f3266.5
Applied rewrites66.5%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites66.5%
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lift-*.f3266.5
Applied rewrites66.5%
Final simplification66.5%
dY.u_m = (fabs.f32 dY.u)
(FPCore (w h dX.u dX.v dY.u_m dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u_m))
(t_1 (* (floor h) dY.v))
(t_2 (+ (* t_0 t_0) (* t_1 t_1)))
(t_3 (* (floor h) dX.v))
(t_4 (* t_3 t_3))
(t_5 (* (floor w) dX.u)))
(if (>= (pow t_5 2.0) (+ (pow t_0 2.0) (pow t_1 2.0)))
(* t_3 (/ 1.0 (sqrt (fmax (+ t_4 (* (floor w) (* dX.u t_5))) t_2))))
(* t_1 (/ 1.0 (sqrt (fmax (+ (* t_5 t_5) t_4) t_2)))))))dY.u_m = fabs(dY_46_u);
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u_m, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u_m;
float t_1 = floorf(h) * dY_46_v;
float t_2 = (t_0 * t_0) + (t_1 * t_1);
float t_3 = floorf(h) * dX_46_v;
float t_4 = t_3 * t_3;
float t_5 = floorf(w) * dX_46_u;
float tmp;
if (powf(t_5, 2.0f) >= (powf(t_0, 2.0f) + powf(t_1, 2.0f))) {
tmp = t_3 * (1.0f / sqrtf(fmaxf((t_4 + (floorf(w) * (dX_46_u * t_5))), t_2)));
} else {
tmp = t_1 * (1.0f / sqrtf(fmaxf(((t_5 * t_5) + t_4), t_2)));
}
return tmp;
}
dY.u_m = abs(dY_46_u) function code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u_m) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(t_3 * t_3) t_5 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if ((t_5 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) tmp = Float32(t_3 * Float32(Float32(1.0) / sqrt(((Float32(t_4 + Float32(floor(w) * Float32(dX_46_u * t_5))) != Float32(t_4 + Float32(floor(w) * Float32(dX_46_u * t_5)))) ? t_2 : ((t_2 != t_2) ? Float32(t_4 + Float32(floor(w) * Float32(dX_46_u * t_5))) : max(Float32(t_4 + Float32(floor(w) * Float32(dX_46_u * t_5))), t_2)))))); else tmp = Float32(t_1 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_5 * t_5) + t_4) != Float32(Float32(t_5 * t_5) + t_4)) ? t_2 : ((t_2 != t_2) ? Float32(Float32(t_5 * t_5) + t_4) : max(Float32(Float32(t_5 * t_5) + t_4), t_2)))))); end return tmp end
dY.u_m = abs(dY_46_u); function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u_m, dY_46_v, maxAniso) t_0 = floor(w) * dY_46_u_m; t_1 = floor(h) * dY_46_v; t_2 = (t_0 * t_0) + (t_1 * t_1); t_3 = floor(h) * dX_46_v; t_4 = t_3 * t_3; t_5 = floor(w) * dX_46_u; tmp = single(0.0); if ((t_5 ^ single(2.0)) >= ((t_0 ^ single(2.0)) + (t_1 ^ single(2.0)))) tmp = t_3 * (single(1.0) / sqrt(max((t_4 + (floor(w) * (dX_46_u * t_5))), t_2))); else tmp = t_1 * (single(1.0) / sqrt(max(((t_5 * t_5) + t_4), t_2))); end tmp_2 = tmp; end
\begin{array}{l}
dY.u_m = \left|dY.u\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\_m\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := t\_3 \cdot t\_3\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;{t\_5}^{2} \geq {t\_0}^{2} + {t\_1}^{2}:\\
\;\;\;\;t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_4 + \left\lfloor w\right\rfloor \cdot \left(dX.u \cdot t\_5\right), t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5 \cdot t\_5 + t\_4, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 77.4%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.5
Applied rewrites66.5%
lift-*.f32N/A
pow2N/A
lift-pow.f3266.5
Applied rewrites66.5%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites66.5%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3266.5
Applied rewrites66.5%
Final simplification66.5%
herbie shell --seed 2024233
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, v)"
:precision binary32
:pre (and (and (and (and (and (and (and (<= 1.0 w) (<= w 16384.0)) (and (<= 1.0 h) (<= h 16384.0))) (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 dY.u)) (<= (fabs dY.u) 1e+20))) (and (<= 1e-20 (fabs dY.v)) (<= (fabs dY.v) 1e+20))) (== maxAniso 16.0))
(if (>= (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))) (* (/ 1.0 (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))) (* (floor h) dX.v)) (* (/ 1.0 (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))) (* (floor h) dY.v))))