
(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 6 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) dX.u))
(t_1 (+ (pow (* (floor h) dX.v) 2.0) (pow t_0 2.0)))
(t_2 (* (floor w) dY.u))
(t_3 (+ (pow (* (floor h) dY.v) 2.0) (pow t_2 2.0)))
(t_4 (pow (fmax t_1 t_3) 0.5)))
(if (>= t_1 t_3) (/ 1.0 (/ t_4 t_0)) (* t_2 (/ 1.0 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) * dX_46_u;
float t_1 = powf((floorf(h) * dX_46_v), 2.0f) + powf(t_0, 2.0f);
float t_2 = floorf(w) * dY_46_u;
float t_3 = powf((floorf(h) * dY_46_v), 2.0f) + powf(t_2, 2.0f);
float t_4 = powf(fmaxf(t_1, t_3), 0.5f);
float tmp;
if (t_1 >= t_3) {
tmp = 1.0f / (t_4 / t_0);
} else {
tmp = t_2 * (1.0f / 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) * dX_46_u) t_1 = Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_4 = ((t_1 != t_1) ? t_3 : ((t_3 != t_3) ? t_1 : max(t_1, t_3))) ^ Float32(0.5) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(Float32(1.0) / Float32(t_4 / t_0)); else tmp = Float32(t_2 * Float32(Float32(1.0) / 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) * dX_46_u; t_1 = ((floor(h) * dX_46_v) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = floor(w) * dY_46_u; t_3 = ((floor(h) * dY_46_v) ^ single(2.0)) + (t_2 ^ single(2.0)); t_4 = max(t_1, t_3) ^ single(0.5); tmp = single(0.0); if (t_1 >= t_3) tmp = single(1.0) / (t_4 / t_0); else tmp = t_2 * (single(1.0) / t_4); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {t\_0}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + {t\_2}^{2}\\
t_4 := {\left(\mathsf{max}\left(t\_1, t\_3\right)\right)}^{0.5}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{1}{\frac{t\_4}{t\_0}}\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot \frac{1}{t\_4}\\
\end{array}
\end{array}
Initial program 80.2%
associate-*l/N/A
clear-numN/A
/-lowering-/.f32N/A
*-lft-identityN/A
Applied egg-rr80.5%
pow1/2N/A
pow2N/A
pow2N/A
pow2N/A
pow2N/A
pow-lowering-pow.f32N/A
Applied egg-rr80.5%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3280.5%
Applied egg-rr80.5%
pow2N/A
>=-lowering->=.f32N/A
Applied egg-rr80.5%
Final simplification80.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (+ (pow (* (floor h) dY.v) 2.0) (pow (* (floor w) dY.u) 2.0)))
(t_2 (+ (pow (* (floor h) dX.v) 2.0) (pow t_0 2.0)))
(t_3 (fmax t_2 t_1)))
(if (>= t_2 t_1)
(* t_0 (/ 1.0 (pow t_3 0.5)))
(* (floor w) (* dY.u (pow t_3 -0.5))))))
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(w) * dX_46_u;
float t_1 = powf((floorf(h) * dY_46_v), 2.0f) + powf((floorf(w) * dY_46_u), 2.0f);
float t_2 = powf((floorf(h) * dX_46_v), 2.0f) + powf(t_0, 2.0f);
float t_3 = fmaxf(t_2, t_1);
float tmp;
if (t_2 >= t_1) {
tmp = t_0 * (1.0f / powf(t_3, 0.5f));
} else {
tmp = floorf(w) * (dY_46_u * powf(t_3, -0.5f));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) t_2 = Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_3 = (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(t_0 * Float32(Float32(1.0) / (t_3 ^ Float32(0.5)))); else tmp = Float32(floor(w) * Float32(dY_46_u * (t_3 ^ Float32(-0.5)))); 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) * dX_46_u; t_1 = ((floor(h) * dY_46_v) ^ single(2.0)) + ((floor(w) * dY_46_u) ^ single(2.0)); t_2 = ((floor(h) * dX_46_v) ^ single(2.0)) + (t_0 ^ single(2.0)); t_3 = max(t_2, t_1); tmp = single(0.0); if (t_2 >= t_1) tmp = t_0 * (single(1.0) / (t_3 ^ single(0.5))); else tmp = floor(w) * (dY_46_u * (t_3 ^ single(-0.5))); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {t\_0}^{2}\\
t_3 := \mathsf{max}\left(t\_2, t\_1\right)\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;t\_0 \cdot \frac{1}{{t\_3}^{0.5}}\\
\mathbf{else}:\\
\;\;\;\;\left\lfloor w\right\rfloor \cdot \left(dY.u \cdot {t\_3}^{-0.5}\right)\\
\end{array}
\end{array}
Initial program 80.2%
associate-*l/N/A
*-commutativeN/A
associate-*r*N/A
*-rgt-identityN/A
times-fracN/A
Applied egg-rr80.2%
pow1/2N/A
pow2N/A
pow2N/A
pow2N/A
pow2N/A
pow-lowering-pow.f32N/A
Applied egg-rr80.2%
>=-lowering->=.f32N/A
pow2N/A
+-commutativeN/A
+-lowering-+.f32N/A
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
Applied egg-rr80.2%
Final simplification80.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* t_0 t_0))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor h) dY.v))
(t_4 (+ t_1 (* t_3 t_3)))
(t_5 (* (floor h) dX.v))
(t_6 (+ (* t_2 t_2) (* t_5 t_5)))
(t_7 (/ 1.0 (sqrt (fmax t_6 t_4))))
(t_8 (* t_2 t_7)))
(if (<= dX.v 200.0)
(if (>=
(* dX.u (* dX.u (pow (floor w) 2.0)))
(+ (pow t_3 2.0) (pow t_0 2.0)))
t_8
(*
t_0
(/
1.0
(sqrt (fmax t_6 (+ t_1 (* (floor h) (* (floor h) (* dY.v dY.v)))))))))
(if (>= (* dX.v (* dX.v (pow (floor h) 2.0))) t_4) t_8 (* t_0 t_7)))))
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 = t_0 * t_0;
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = t_1 + (t_3 * t_3);
float t_5 = floorf(h) * dX_46_v;
float t_6 = (t_2 * t_2) + (t_5 * t_5);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, t_4));
float t_8 = t_2 * t_7;
float tmp_1;
if (dX_46_v <= 200.0f) {
float tmp_2;
if ((dX_46_u * (dX_46_u * powf(floorf(w), 2.0f))) >= (powf(t_3, 2.0f) + powf(t_0, 2.0f))) {
tmp_2 = t_8;
} else {
tmp_2 = t_0 * (1.0f / sqrtf(fmaxf(t_6, (t_1 + (floorf(h) * (floorf(h) * (dY_46_v * dY_46_v)))))));
}
tmp_1 = tmp_2;
} else if ((dX_46_v * (dX_46_v * powf(floorf(h), 2.0f))) >= t_4) {
tmp_1 = t_8;
} else {
tmp_1 = t_0 * t_7;
}
return tmp_1;
}
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 * t_0) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(t_1 + Float32(t_3 * t_3)) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)) t_7 = Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? t_4 : ((t_4 != t_4) ? t_6 : max(t_6, t_4))))) t_8 = Float32(t_2 * t_7) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(200.0)) tmp_2 = Float32(0.0) if (Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))) >= Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) tmp_2 = t_8; else tmp_2 = Float32(t_0 * Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(t_1 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) : ((Float32(t_1 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) != Float32(t_1 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v))))) ? t_6 : max(t_6, Float32(t_1 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))))))))); end tmp_1 = tmp_2; elseif (Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0)))) >= t_4) tmp_1 = t_8; else tmp_1 = Float32(t_0 * t_7); 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(w) * dY_46_u; t_1 = t_0 * t_0; t_2 = floor(w) * dX_46_u; t_3 = floor(h) * dY_46_v; t_4 = t_1 + (t_3 * t_3); t_5 = floor(h) * dX_46_v; t_6 = (t_2 * t_2) + (t_5 * t_5); t_7 = single(1.0) / sqrt(max(t_6, t_4)); t_8 = t_2 * t_7; tmp_2 = single(0.0); if (dX_46_v <= single(200.0)) tmp_3 = single(0.0); if ((dX_46_u * (dX_46_u * (floor(w) ^ single(2.0)))) >= ((t_3 ^ single(2.0)) + (t_0 ^ single(2.0)))) tmp_3 = t_8; else tmp_3 = t_0 * (single(1.0) / sqrt(max(t_6, (t_1 + (floor(h) * (floor(h) * (dY_46_v * dY_46_v))))))); end tmp_2 = tmp_3; elseif ((dX_46_v * (dX_46_v * (floor(h) ^ single(2.0)))) >= t_4) tmp_2 = t_8; else tmp_2 = t_0 * t_7; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := t\_0 \cdot t\_0\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := t\_1 + t\_3 \cdot t\_3\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := t\_2 \cdot t\_2 + t\_5 \cdot t\_5\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_4\right)}}\\
t_8 := t\_2 \cdot t\_7\\
\mathbf{if}\;dX.v \leq 200:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;dX.u \cdot \left(dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right) \geq {t\_3}^{2} + {t\_0}^{2}:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_1 + \left\lfloor h\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot \left(dY.v \cdot dY.v\right)\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;dX.v \cdot \left(dX.v \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right) \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot t\_7\\
\end{array}
\end{array}
if dX.v < 200Initial program 81.9%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3272.8%
Simplified72.8%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3272.8%
Applied egg-rr72.8%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3272.8%
Applied egg-rr72.8%
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
associate-*r*N/A
pow2N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3272.9%
Applied egg-rr72.9%
if 200 < dX.v Initial program 75.0%
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.f3270.8%
Simplified70.8%
Final simplification72.4%
(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) dY.u))
(t_2 (+ (* t_1 t_1) (* t_0 t_0)))
(t_3 (* (floor w) dX.u))
(t_4 (* t_3 t_3))
(t_5 (* (floor h) dX.v)))
(if (>=
(* dX.u (* dX.u (pow (floor w) 2.0)))
(+ (pow t_0 2.0) (pow t_1 2.0)))
(*
t_3
(/ 1.0 (sqrt (fmax (+ t_4 (* dX.v (* dX.v (pow (floor h) 2.0)))) t_2))))
(* t_1 (/ 1.0 (sqrt (fmax (+ t_4 (* t_5 t_5)) t_2)))))))
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) * dY_46_u;
float t_2 = (t_1 * t_1) + (t_0 * t_0);
float t_3 = floorf(w) * dX_46_u;
float t_4 = t_3 * t_3;
float t_5 = floorf(h) * dX_46_v;
float tmp;
if ((dX_46_u * (dX_46_u * powf(floorf(w), 2.0f))) >= (powf(t_0, 2.0f) + powf(t_1, 2.0f))) {
tmp = t_3 * (1.0f / sqrtf(fmaxf((t_4 + (dX_46_v * (dX_46_v * powf(floorf(h), 2.0f)))), t_2)));
} else {
tmp = t_1 * (1.0f / sqrtf(fmaxf((t_4 + (t_5 * t_5)), t_2)));
}
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) * dY_46_u) t_2 = Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(t_3 * t_3) t_5 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if (Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ 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(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))) != Float32(t_4 + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0)))))) ? t_2 : ((t_2 != t_2) ? Float32(t_4 + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))) : max(Float32(t_4 + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))), t_2)))))); else tmp = Float32(t_1 * Float32(Float32(1.0) / sqrt(((Float32(t_4 + Float32(t_5 * t_5)) != Float32(t_4 + Float32(t_5 * t_5))) ? t_2 : ((t_2 != t_2) ? Float32(t_4 + Float32(t_5 * t_5)) : max(Float32(t_4 + Float32(t_5 * t_5)), t_2)))))); 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) * dY_46_v; t_1 = floor(w) * dY_46_u; t_2 = (t_1 * t_1) + (t_0 * t_0); t_3 = floor(w) * dX_46_u; t_4 = t_3 * t_3; t_5 = floor(h) * dX_46_v; tmp = single(0.0); if ((dX_46_u * (dX_46_u * (floor(w) ^ single(2.0)))) >= ((t_0 ^ single(2.0)) + (t_1 ^ single(2.0)))) tmp = t_3 * (single(1.0) / sqrt(max((t_4 + (dX_46_v * (dX_46_v * (floor(h) ^ single(2.0))))), t_2))); else tmp = t_1 * (single(1.0) / sqrt(max((t_4 + (t_5 * t_5)), t_2))); end tmp_2 = 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 dY.u\\
t_2 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;dX.u \cdot \left(dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right) \geq {t\_0}^{2} + {t\_1}^{2}:\\
\;\;\;\;t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_4 + dX.v \cdot \left(dX.v \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right), t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_4 + t\_5 \cdot t\_5, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 80.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3268.2%
Simplified68.2%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3268.2%
Applied egg-rr68.2%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3268.2%
Applied egg-rr68.2%
associate-*r*N/A
*-lowering-*.f32N/A
*-commutativeN/A
associate-*l*N/A
pow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3268.2%
Applied egg-rr68.2%
Final simplification68.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) dX.u))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_2 t_2) (* t_1 t_1)))))))
(if (>= (pow t_3 2.0) (+ (pow t_1 2.0) (pow t_2 2.0)))
(* t_3 t_4)
(* t_2 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(h) * dY_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(w) * dX_46_u;
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1))));
float tmp;
if (powf(t_3, 2.0f) >= (powf(t_1, 2.0f) + powf(t_2, 2.0f))) {
tmp = t_3 * t_4;
} else {
tmp = t_2 * 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(h) * dY_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(w) * dX_46_u) t_4 = 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) + Float32(t_1 * t_1)) : ((Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) != Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))) ? 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) + Float32(t_1 * t_1))))))) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(t_3 * t_4); else tmp = Float32(t_2 * 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(h) * dY_46_v; t_2 = floor(w) * dY_46_u; t_3 = floor(w) * dX_46_u; t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1)))); tmp = single(0.0); if ((t_3 ^ single(2.0)) >= ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = t_3 * t_4; else tmp = t_2 * 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 h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_2 \cdot t\_2 + t\_1 \cdot t\_1\right)}}\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;t\_3 \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot t\_4\\
\end{array}
\end{array}
Initial program 80.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3268.2%
Simplified68.2%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3268.2%
Applied egg-rr68.2%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3268.2%
Applied egg-rr68.2%
associate-*r*N/A
pow2N/A
unpow-prod-downN/A
*-commutativeN/A
pow2N/A
pow2N/A
>=-lowering->=.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
+-commutativeN/A
+-lowering-+.f32N/A
Applied egg-rr68.2%
Final simplification68.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor h) dY.v))
(t_2 (* t_1 t_1))
(t_3 (* (floor w) dY.u))
(t_4 (* (floor w) dX.u))
(t_5 (+ (* t_4 t_4) (* t_0 t_0))))
(if (>= (pow t_4 2.0) (+ (pow t_1 2.0) (pow t_3 2.0)))
(* t_4 (/ 1.0 (sqrt (fmax t_5 (+ (* t_3 t_3) t_2)))))
(* t_3 (/ 1.0 (sqrt (fmax t_5 (+ t_2 (* (floor w) (* dY.u 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(h) * dX_46_v;
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 = floorf(w) * dX_46_u;
float t_5 = (t_4 * t_4) + (t_0 * t_0);
float tmp;
if (powf(t_4, 2.0f) >= (powf(t_1, 2.0f) + powf(t_3, 2.0f))) {
tmp = t_4 * (1.0f / sqrtf(fmaxf(t_5, ((t_3 * t_3) + t_2))));
} else {
tmp = t_3 * (1.0f / sqrtf(fmaxf(t_5, (t_2 + (floorf(w) * (dY_46_u * 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(h) * dX_46_v) 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 = Float32(floor(w) * dX_46_u) t_5 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0)))) tmp = Float32(t_4 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_3 * t_3) + t_2) : ((Float32(Float32(t_3 * t_3) + t_2) != Float32(Float32(t_3 * t_3) + t_2)) ? t_5 : max(t_5, Float32(Float32(t_3 * t_3) + t_2))))))); else tmp = Float32(t_3 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(t_2 + Float32(floor(w) * Float32(dY_46_u * t_3))) : ((Float32(t_2 + Float32(floor(w) * Float32(dY_46_u * t_3))) != Float32(t_2 + Float32(floor(w) * Float32(dY_46_u * t_3)))) ? t_5 : max(t_5, Float32(t_2 + Float32(floor(w) * Float32(dY_46_u * 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(h) * dX_46_v; t_1 = floor(h) * dY_46_v; t_2 = t_1 * t_1; t_3 = floor(w) * dY_46_u; t_4 = floor(w) * dX_46_u; t_5 = (t_4 * t_4) + (t_0 * t_0); tmp = single(0.0); if ((t_4 ^ single(2.0)) >= ((t_1 ^ single(2.0)) + (t_3 ^ single(2.0)))) tmp = t_4 * (single(1.0) / sqrt(max(t_5, ((t_3 * t_3) + t_2)))); else tmp = t_3 * (single(1.0) / sqrt(max(t_5, (t_2 + (floor(w) * (dY_46_u * t_3)))))); 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 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 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
\mathbf{if}\;{t\_4}^{2} \geq {t\_1}^{2} + {t\_3}^{2}:\\
\;\;\;\;t\_4 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_3 \cdot t\_3 + t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 + \left\lfloor w\right\rfloor \cdot \left(dY.u \cdot t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 80.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3268.2%
Simplified68.2%
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
floor-lowering-floor.f3268.2%
Applied egg-rr68.2%
associate-*r*N/A
pow2N/A
unpow-prod-downN/A
*-commutativeN/A
>=-lowering->=.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
+-commutativeN/A
+-lowering-+.f32N/A
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
Applied egg-rr68.2%
Final simplification68.2%
herbie shell --seed 2024192
(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))))