
(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_2) (* t_6 t_1))))
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_2;
} else {
tmp = t_6 * t_1;
}
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_2); else tmp = Float32(t_6 * t_1); 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_2; else tmp = t_6 * t_1; 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\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 5 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_2) (* t_6 t_1))))
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_2;
} else {
tmp = t_6 * t_1;
}
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_2); else tmp = Float32(t_6 * t_1); 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_2; else tmp = t_6 * t_1; 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\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (+ (pow t_0 2.0) (pow (* (floor h) dY.v) 2.0)))
(t_2 (+ (pow (* (floor w) dX.u) 2.0) (pow (* (floor h) dX.v) 2.0)))
(t_3 (sqrt (fmax t_2 t_1))))
(if (>= t_2 t_1) (* (floor w) (/ dX.u t_3)) (/ t_0 t_3))))
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(w) * dY_46_u;
float t_1 = powf(t_0, 2.0f) + powf((floorf(h) * dY_46_v), 2.0f);
float t_2 = powf((floorf(w) * dX_46_u), 2.0f) + powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = floorf(w) * (dX_46_u / t_3);
} else {
tmp = t_0 / t_3;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32((t_0 ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))) t_2 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) t_3 = sqrt(((t_2 != t_2) ? t_1 : ((t_1 != t_1) ? t_2 : max(t_2, t_1)))) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(floor(w) * Float32(dX_46_u / t_3)); else tmp = Float32(t_0 / t_3); 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(w) * dY_46_u; t_1 = (t_0 ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0)); t_2 = ((floor(w) * dX_46_u) ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0)); t_3 = sqrt(max(t_2, t_1)); tmp = single(0.0); if (t_2 >= t_1) tmp = floor(w) * (dX_46_u / t_3); else tmp = t_0 / t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := {t\_0}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\\
t_2 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_1\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\left\lfloor w\right\rfloor \cdot \frac{dX.u}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_3}\\
\end{array}
\end{array}
Initial program 71.7%
associate-*l/N/A
*-commutativeN/A
associate-*r*N/A
*-rgt-identityN/A
times-fracN/A
Applied egg-rr71.8%
Applied egg-rr71.9%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3271.9
Applied egg-rr71.9%
>=-lowering->=.f32N/A
+-lowering-+.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
+-lowering-+.f32N/A
Applied egg-rr71.9%
Final simplification71.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (floor h) 2.0))
(t_1 (* (floor h) dY.v))
(t_2 (* t_1 t_1))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor w) dY.u))
(t_5 (+ t_2 (* t_4 t_4)))
(t_6 (* (floor h) dX.v))
(t_7 (* t_6 t_6))
(t_8 (+ (* t_3 t_3) t_7))
(t_9 (/ 1.0 (sqrt (fmax t_8 t_5))))
(t_10 (* t_3 t_9)))
(if (<= dY.u -100.0)
(if (>= (* dX.v (* dX.v t_0)) (+ (pow t_4 2.0) t_2))
t_10
(* t_4 (/ 1.0 (sqrt (fmax (+ t_7 (exp (* 2.0 (log t_3)))) t_5)))))
(if (>= t_8 (* t_0 (* dY.v dY.v))) t_10 (* t_4 t_9)))))
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 = powf(floorf(h), 2.0f);
float t_1 = floorf(h) * dY_46_v;
float t_2 = t_1 * t_1;
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(w) * dY_46_u;
float t_5 = t_2 + (t_4 * t_4);
float t_6 = floorf(h) * dX_46_v;
float t_7 = t_6 * t_6;
float t_8 = (t_3 * t_3) + t_7;
float t_9 = 1.0f / sqrtf(fmaxf(t_8, t_5));
float t_10 = t_3 * t_9;
float tmp_1;
if (dY_46_u <= -100.0f) {
float tmp_2;
if ((dX_46_v * (dX_46_v * t_0)) >= (powf(t_4, 2.0f) + t_2)) {
tmp_2 = t_10;
} else {
tmp_2 = t_4 * (1.0f / sqrtf(fmaxf((t_7 + expf((2.0f * logf(t_3)))), t_5)));
}
tmp_1 = tmp_2;
} else if (t_8 >= (t_0 * (dY_46_v * dY_46_v))) {
tmp_1 = t_10;
} else {
tmp_1 = t_4 * t_9;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) ^ Float32(2.0) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(w) * dY_46_u) t_5 = Float32(t_2 + Float32(t_4 * t_4)) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(t_6 * t_6) t_8 = Float32(Float32(t_3 * t_3) + t_7) t_9 = Float32(Float32(1.0) / sqrt(((t_8 != t_8) ? t_5 : ((t_5 != t_5) ? t_8 : max(t_8, t_5))))) t_10 = Float32(t_3 * t_9) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(-100.0)) tmp_2 = Float32(0.0) if (Float32(dX_46_v * Float32(dX_46_v * t_0)) >= Float32((t_4 ^ Float32(2.0)) + t_2)) tmp_2 = t_10; else tmp_2 = Float32(t_4 * Float32(Float32(1.0) / sqrt(((Float32(t_7 + exp(Float32(Float32(2.0) * log(t_3)))) != Float32(t_7 + exp(Float32(Float32(2.0) * log(t_3))))) ? t_5 : ((t_5 != t_5) ? Float32(t_7 + exp(Float32(Float32(2.0) * log(t_3)))) : max(Float32(t_7 + exp(Float32(Float32(2.0) * log(t_3)))), t_5)))))); end tmp_1 = tmp_2; elseif (t_8 >= Float32(t_0 * Float32(dY_46_v * dY_46_v))) tmp_1 = t_10; else tmp_1 = Float32(t_4 * t_9); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) ^ single(2.0); t_1 = floor(h) * dY_46_v; t_2 = t_1 * t_1; t_3 = floor(w) * dX_46_u; t_4 = floor(w) * dY_46_u; t_5 = t_2 + (t_4 * t_4); t_6 = floor(h) * dX_46_v; t_7 = t_6 * t_6; t_8 = (t_3 * t_3) + t_7; t_9 = single(1.0) / sqrt(max(t_8, t_5)); t_10 = t_3 * t_9; tmp_2 = single(0.0); if (dY_46_u <= single(-100.0)) tmp_3 = single(0.0); if ((dX_46_v * (dX_46_v * t_0)) >= ((t_4 ^ single(2.0)) + t_2)) tmp_3 = t_10; else tmp_3 = t_4 * (single(1.0) / sqrt(max((t_7 + exp((single(2.0) * log(t_3)))), t_5))); end tmp_2 = tmp_3; elseif (t_8 >= (t_0 * (dY_46_v * dY_46_v))) tmp_2 = t_10; else tmp_2 = t_4 * t_9; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_5 := t\_2 + t\_4 \cdot t\_4\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := t\_6 \cdot t\_6\\
t_8 := t\_3 \cdot t\_3 + t\_7\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_5\right)}}\\
t_10 := t\_3 \cdot t\_9\\
\mathbf{if}\;dY.u \leq -100:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;dX.v \cdot \left(dX.v \cdot t\_0\right) \geq {t\_4}^{2} + t\_2:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_7 + e^{2 \cdot \log t\_3}, t\_5\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_8 \geq t\_0 \cdot \left(dY.v \cdot dY.v\right):\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_9\\
\end{array}
\end{array}
if dY.u < -100Initial program 62.3%
Taylor expanded in dX.u around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3256.8
Simplified56.8%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3256.8
Applied egg-rr56.8%
pow2N/A
pow-to-expN/A
exp-lowering-exp.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
log-lowering-log.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3261.9
Applied egg-rr61.9%
if -100 < dY.u Initial program 74.6%
Taylor expanded in dY.u around 0
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f32N/A
unpow2N/A
*-lowering-*.f3270.4
Simplified70.4%
Final simplification68.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (floor h) 2.0))
(t_1 (* (floor h) dY.v))
(t_2 (* t_1 t_1))
(t_3 (* (floor w) dY.u))
(t_4 (pow t_3 2.0))
(t_5 (* (floor h) dX.v))
(t_6 (* (floor w) dX.u))
(t_7
(sqrt (fmax (+ (pow t_6 2.0) (pow t_5 2.0)) (+ t_4 (pow t_1 2.0)))))
(t_8 (+ (* t_6 t_6) (* t_5 t_5)))
(t_9 (/ 1.0 (sqrt (fmax t_8 (+ t_2 (* t_3 t_3)))))))
(if (<= dY.u -999999995904.0)
(if (>= (* dX.v (* dX.v t_0)) (+ t_4 t_2))
(* t_6 (/ 1.0 t_7))
(/ dY.u (/ t_7 (floor w))))
(if (>= t_8 (* t_0 (* dY.v dY.v))) (* t_6 t_9) (* t_3 t_9)))))
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 = powf(floorf(h), 2.0f);
float t_1 = floorf(h) * dY_46_v;
float t_2 = t_1 * t_1;
float t_3 = floorf(w) * dY_46_u;
float t_4 = powf(t_3, 2.0f);
float t_5 = floorf(h) * dX_46_v;
float t_6 = floorf(w) * dX_46_u;
float t_7 = sqrtf(fmaxf((powf(t_6, 2.0f) + powf(t_5, 2.0f)), (t_4 + powf(t_1, 2.0f))));
float t_8 = (t_6 * t_6) + (t_5 * t_5);
float t_9 = 1.0f / sqrtf(fmaxf(t_8, (t_2 + (t_3 * t_3))));
float tmp_1;
if (dY_46_u <= -999999995904.0f) {
float tmp_2;
if ((dX_46_v * (dX_46_v * t_0)) >= (t_4 + t_2)) {
tmp_2 = t_6 * (1.0f / t_7);
} else {
tmp_2 = dY_46_u / (t_7 / floorf(w));
}
tmp_1 = tmp_2;
} else if (t_8 >= (t_0 * (dY_46_v * dY_46_v))) {
tmp_1 = t_6 * t_9;
} else {
tmp_1 = t_3 * t_9;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) ^ Float32(2.0) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(w) * dY_46_u) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(floor(w) * dX_46_u) t_7 = sqrt(((Float32((t_6 ^ Float32(2.0)) + (t_5 ^ Float32(2.0))) != Float32((t_6 ^ Float32(2.0)) + (t_5 ^ Float32(2.0)))) ? Float32(t_4 + (t_1 ^ Float32(2.0))) : ((Float32(t_4 + (t_1 ^ Float32(2.0))) != Float32(t_4 + (t_1 ^ Float32(2.0)))) ? Float32((t_6 ^ Float32(2.0)) + (t_5 ^ Float32(2.0))) : max(Float32((t_6 ^ Float32(2.0)) + (t_5 ^ Float32(2.0))), Float32(t_4 + (t_1 ^ Float32(2.0))))))) t_8 = Float32(Float32(t_6 * t_6) + Float32(t_5 * t_5)) t_9 = Float32(Float32(1.0) / sqrt(((t_8 != t_8) ? Float32(t_2 + Float32(t_3 * t_3)) : ((Float32(t_2 + Float32(t_3 * t_3)) != Float32(t_2 + Float32(t_3 * t_3))) ? t_8 : max(t_8, Float32(t_2 + Float32(t_3 * t_3))))))) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(-999999995904.0)) tmp_2 = Float32(0.0) if (Float32(dX_46_v * Float32(dX_46_v * t_0)) >= Float32(t_4 + t_2)) tmp_2 = Float32(t_6 * Float32(Float32(1.0) / t_7)); else tmp_2 = Float32(dY_46_u / Float32(t_7 / floor(w))); end tmp_1 = tmp_2; elseif (t_8 >= Float32(t_0 * Float32(dY_46_v * dY_46_v))) tmp_1 = Float32(t_6 * t_9); else tmp_1 = Float32(t_3 * t_9); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) ^ single(2.0); t_1 = floor(h) * dY_46_v; t_2 = t_1 * t_1; t_3 = floor(w) * dY_46_u; t_4 = t_3 ^ single(2.0); t_5 = floor(h) * dX_46_v; t_6 = floor(w) * dX_46_u; t_7 = sqrt(max(((t_6 ^ single(2.0)) + (t_5 ^ single(2.0))), (t_4 + (t_1 ^ single(2.0))))); t_8 = (t_6 * t_6) + (t_5 * t_5); t_9 = single(1.0) / sqrt(max(t_8, (t_2 + (t_3 * t_3)))); tmp_2 = single(0.0); if (dY_46_u <= single(-999999995904.0)) tmp_3 = single(0.0); if ((dX_46_v * (dX_46_v * t_0)) >= (t_4 + t_2)) tmp_3 = t_6 * (single(1.0) / t_7); else tmp_3 = dY_46_u / (t_7 / floor(w)); end tmp_2 = tmp_3; elseif (t_8 >= (t_0 * (dY_46_v * dY_46_v))) tmp_2 = t_6 * t_9; else tmp_2 = t_3 * t_9; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := {t\_3}^{2}\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_7 := \sqrt{\mathsf{max}\left({t\_6}^{2} + {t\_5}^{2}, t\_4 + {t\_1}^{2}\right)}\\
t_8 := t\_6 \cdot t\_6 + t\_5 \cdot t\_5\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_2 + t\_3 \cdot t\_3\right)}}\\
\mathbf{if}\;dY.u \leq -999999995904:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;dX.v \cdot \left(dX.v \cdot t\_0\right) \geq t\_4 + t\_2:\\
\;\;\;\;t\_6 \cdot \frac{1}{t\_7}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.u}{\frac{t\_7}{\left\lfloor w\right\rfloor }}\\
\end{array}\\
\mathbf{elif}\;t\_8 \geq t\_0 \cdot \left(dY.v \cdot dY.v\right):\\
\;\;\;\;t\_6 \cdot t\_9\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot t\_9\\
\end{array}
\end{array}
if dY.u < -999999996000Initial program 51.4%
Taylor expanded in dX.u around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3248.6
Simplified48.6%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3248.6
Applied egg-rr48.6%
Applied egg-rr49.0%
pow2N/A
pow2N/A
pow2N/A
pow2N/A
remove-double-negN/A
neg-sub0N/A
--lowering--.f32N/A
Applied egg-rr49.0%
if -999999996000 < dY.u Initial program 74.3%
Taylor expanded in dY.u around 0
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f32N/A
unpow2N/A
*-lowering-*.f3269.9
Simplified69.9%
Final simplification67.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor w) dX.u))
(t_2 (pow (floor h) 2.0))
(t_3 (* dX.v (* dX.v t_2)))
(t_4 (* (floor w) dY.u))
(t_5 (pow t_4 2.0)))
(if (>= t_3 (+ (* t_0 t_0) (* t_4 t_4)))
(*
t_1
(/
1.0
(sqrt
(fmax
(+ (pow t_1 2.0) (pow (* (floor h) dX.v) 2.0))
(+ t_5 (pow t_0 2.0))))))
(/
t_4
(sqrt
(fmax
(fma (* dX.u dX.u) (pow (floor w) 2.0) t_3)
(fma t_2 (* dY.v dY.v) t_5)))))))
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) * dY_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = powf(floorf(h), 2.0f);
float t_3 = dX_46_v * (dX_46_v * t_2);
float t_4 = floorf(w) * dY_46_u;
float t_5 = powf(t_4, 2.0f);
float tmp;
if (t_3 >= ((t_0 * t_0) + (t_4 * t_4))) {
tmp = t_1 * (1.0f / sqrtf(fmaxf((powf(t_1, 2.0f) + powf((floorf(h) * dX_46_v), 2.0f)), (t_5 + powf(t_0, 2.0f)))));
} else {
tmp = t_4 / sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), powf(floorf(w), 2.0f), t_3), fmaf(t_2, (dY_46_v * dY_46_v), t_5)));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dY_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = floor(h) ^ Float32(2.0) t_3 = Float32(dX_46_v * Float32(dX_46_v * t_2)) t_4 = Float32(floor(w) * dY_46_u) t_5 = t_4 ^ Float32(2.0) tmp = Float32(0.0) if (t_3 >= Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4))) tmp = Float32(t_1 * Float32(Float32(1.0) / sqrt(((Float32((t_1 ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) != Float32((t_1 ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0)))) ? Float32(t_5 + (t_0 ^ Float32(2.0))) : ((Float32(t_5 + (t_0 ^ Float32(2.0))) != Float32(t_5 + (t_0 ^ Float32(2.0)))) ? Float32((t_1 ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) : max(Float32((t_1 ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))), Float32(t_5 + (t_0 ^ Float32(2.0))))))))); else tmp = Float32(t_4 / sqrt(((fma(Float32(dX_46_u * dX_46_u), (floor(w) ^ Float32(2.0)), t_3) != fma(Float32(dX_46_u * dX_46_u), (floor(w) ^ Float32(2.0)), t_3)) ? fma(t_2, Float32(dY_46_v * dY_46_v), t_5) : ((fma(t_2, Float32(dY_46_v * dY_46_v), t_5) != fma(t_2, Float32(dY_46_v * dY_46_v), t_5)) ? fma(Float32(dX_46_u * dX_46_u), (floor(w) ^ Float32(2.0)), t_3) : max(fma(Float32(dX_46_u * dX_46_u), (floor(w) ^ Float32(2.0)), t_3), fma(t_2, Float32(dY_46_v * dY_46_v), t_5)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_3 := dX.v \cdot \left(dX.v \cdot t\_2\right)\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_5 := {t\_4}^{2}\\
\mathbf{if}\;t\_3 \geq t\_0 \cdot t\_0 + t\_4 \cdot t\_4:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left({t\_1}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}, t\_5 + {t\_0}^{2}\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, {\left(\left\lfloor w\right\rfloor \right)}^{2}, t\_3\right), \mathsf{fma}\left(t\_2, dY.v \cdot dY.v, t\_5\right)\right)}}\\
\end{array}
\end{array}
Initial program 71.7%
Taylor expanded in dX.u around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3262.9
Simplified62.9%
Applied egg-rr63.1%
pow2N/A
pow2N/A
pow2N/A
pow2N/A
remove-double-negN/A
neg-sub0N/A
--lowering--.f32N/A
Applied egg-rr63.1%
Taylor expanded in w around 0
Simplified63.1%
Final simplification63.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (+ (pow t_0 2.0) (pow (* (floor h) dY.v) 2.0)))
(t_2 (pow (* (floor h) dX.v) 2.0))
(t_3 (* (floor w) dX.u))
(t_4 (sqrt (fmax (+ (pow t_3 2.0) t_2) t_1))))
(if (>= t_2 t_1) (* t_3 (/ 1.0 t_4)) (/ t_0 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(w) * dY_46_u;
float t_1 = powf(t_0, 2.0f) + powf((floorf(h) * dY_46_v), 2.0f);
float t_2 = powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = floorf(w) * dX_46_u;
float t_4 = sqrtf(fmaxf((powf(t_3, 2.0f) + t_2), t_1));
float tmp;
if (t_2 >= t_1) {
tmp = t_3 * (1.0f / t_4);
} else {
tmp = t_0 / 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(w) * dY_46_u) t_1 = Float32((t_0 ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))) t_2 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_3 = Float32(floor(w) * dX_46_u) t_4 = sqrt(((Float32((t_3 ^ Float32(2.0)) + t_2) != Float32((t_3 ^ Float32(2.0)) + t_2)) ? t_1 : ((t_1 != t_1) ? Float32((t_3 ^ Float32(2.0)) + t_2) : max(Float32((t_3 ^ Float32(2.0)) + t_2), t_1)))) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(t_3 * Float32(Float32(1.0) / t_4)); else tmp = Float32(t_0 / 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(w) * dY_46_u; t_1 = (t_0 ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0)); t_2 = (floor(h) * dX_46_v) ^ single(2.0); t_3 = floor(w) * dX_46_u; t_4 = sqrt(max(((t_3 ^ single(2.0)) + t_2), t_1)); tmp = single(0.0); if (t_2 >= t_1) tmp = t_3 * (single(1.0) / t_4); else tmp = t_0 / t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := {t\_0}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \sqrt{\mathsf{max}\left({t\_3}^{2} + t\_2, t\_1\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;t\_3 \cdot \frac{1}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\end{array}
\end{array}
Initial program 71.7%
Taylor expanded in dX.u around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3262.9
Simplified62.9%
Applied egg-rr63.1%
pow2N/A
pow2N/A
pow2N/A
pow2N/A
remove-double-negN/A
neg-sub0N/A
--lowering--.f32N/A
Applied egg-rr63.1%
associate-*r*N/A
*-commutativeN/A
unpow2N/A
swap-sqrN/A
>=-lowering->=.f32N/A
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
+-lowering-+.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
pow-lowering-pow.f32N/A
Applied egg-rr63.1%
Final simplification63.1%
herbie shell --seed 2024204
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, u)"
: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 w) dX.u)) (* (/ 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 w) dY.u))))