
(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 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_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}
(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 (+ (pow (* (floor w) dX.u) 2.0) (pow t_0 2.0)))
(t_3 (* (floor h) dY.v))
(t_4 (pow (fmax t_2 (+ (pow t_1 2.0) (pow t_3 2.0))) 0.5)))
(if (>= t_2 (+ (* t_1 t_1) (* t_3 t_3)))
(/ 1.0 (/ t_4 t_0))
(* t_3 (/ 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(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = powf((floorf(w) * dX_46_u), 2.0f) + powf(t_0, 2.0f);
float t_3 = floorf(h) * dY_46_v;
float t_4 = powf(fmaxf(t_2, (powf(t_1, 2.0f) + powf(t_3, 2.0f))), 0.5f);
float tmp;
if (t_2 >= ((t_1 * t_1) + (t_3 * t_3))) {
tmp = 1.0f / (t_4 / t_0);
} else {
tmp = t_3 * (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(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_3 = Float32(floor(h) * dY_46_v) t_4 = ((t_2 != t_2) ? Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) : ((Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) != Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0)))) ? t_2 : max(t_2, Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0)))))) ^ Float32(0.5) tmp = Float32(0.0) if (t_2 >= Float32(Float32(t_1 * t_1) + Float32(t_3 * t_3))) tmp = Float32(Float32(1.0) / Float32(t_4 / t_0)); else tmp = Float32(t_3 * 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(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = ((floor(w) * dX_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_3 = floor(h) * dY_46_v; t_4 = max(t_2, ((t_1 ^ single(2.0)) + (t_3 ^ single(2.0)))) ^ single(0.5); tmp = single(0.0); if (t_2 >= ((t_1 * t_1) + (t_3 * t_3))) tmp = single(1.0) / (t_4 / t_0); else tmp = t_3 * (single(1.0) / 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(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {t\_0}^{2}\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := {\left(\mathsf{max}\left(t\_2, {t\_1}^{2} + {t\_3}^{2}\right)\right)}^{0.5}\\
\mathbf{if}\;t\_2 \geq t\_1 \cdot t\_1 + t\_3 \cdot t\_3:\\
\;\;\;\;\frac{1}{\frac{t\_4}{t\_0}}\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot \frac{1}{t\_4}\\
\end{array}
\end{array}
Initial program 76.9%
associate-*l/N/A
clear-numN/A
/-lowering-/.f32N/A
*-lft-identityN/A
Applied egg-rr76.9%
pow1/2N/A
pow2N/A
pow2N/A
pow2N/A
pow2N/A
pow-lowering-pow.f32N/A
Applied egg-rr76.9%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3276.9%
Applied egg-rr76.9%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3276.9%
Applied egg-rr76.9%
Final simplification76.9%
(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 (+ (pow (* (floor w) dX.u) 2.0) (pow t_0 2.0)))
(t_3 (* (floor h) dY.v))
(t_4 (pow (fmax t_2 (+ (pow t_1 2.0) (pow t_3 2.0))) 0.5)))
(if (>= t_2 (+ (* t_1 t_1) (* t_3 t_3)))
(* t_0 (/ 1.0 t_4))
(* (floor h) (/ dY.v 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 = powf((floorf(w) * dX_46_u), 2.0f) + powf(t_0, 2.0f);
float t_3 = floorf(h) * dY_46_v;
float t_4 = powf(fmaxf(t_2, (powf(t_1, 2.0f) + powf(t_3, 2.0f))), 0.5f);
float tmp;
if (t_2 >= ((t_1 * t_1) + (t_3 * t_3))) {
tmp = t_0 * (1.0f / t_4);
} else {
tmp = floorf(h) * (dY_46_v / 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((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_3 = Float32(floor(h) * dY_46_v) t_4 = ((t_2 != t_2) ? Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) : ((Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) != Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0)))) ? t_2 : max(t_2, Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0)))))) ^ Float32(0.5) tmp = Float32(0.0) if (t_2 >= Float32(Float32(t_1 * t_1) + Float32(t_3 * t_3))) tmp = Float32(t_0 * Float32(Float32(1.0) / t_4)); else tmp = Float32(floor(h) * Float32(dY_46_v / 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) ^ single(2.0)) + (t_0 ^ single(2.0)); t_3 = floor(h) * dY_46_v; t_4 = max(t_2, ((t_1 ^ single(2.0)) + (t_3 ^ single(2.0)))) ^ single(0.5); tmp = single(0.0); if (t_2 >= ((t_1 * t_1) + (t_3 * t_3))) tmp = t_0 * (single(1.0) / t_4); else tmp = floor(h) * (dY_46_v / 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(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {t\_0}^{2}\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := {\left(\mathsf{max}\left(t\_2, {t\_1}^{2} + {t\_3}^{2}\right)\right)}^{0.5}\\
\mathbf{if}\;t\_2 \geq t\_1 \cdot t\_1 + t\_3 \cdot t\_3:\\
\;\;\;\;t\_0 \cdot \frac{1}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dY.v}{t\_4}\\
\end{array}
\end{array}
Initial program 76.9%
associate-*l/N/A
*-rgt-identityN/A
*-commutativeN/A
associate-*r*N/A
times-fracN/A
Applied egg-rr76.7%
pow1/2N/A
pow2N/A
pow2N/A
pow2N/A
pow2N/A
pow-lowering-pow.f32N/A
Applied egg-rr76.7%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3276.7%
Applied egg-rr76.7%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3276.7%
Applied egg-rr76.7%
Final simplification76.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (+ (* t_1 t_1) (* t_2 t_2)))
(t_4 (* (floor h) dX.v))
(t_5 (/ 1.0 (sqrt (fmax (+ (* t_0 t_0) (* t_4 t_4)) t_3))))
(t_6 (* t_4 t_5))
(t_7 (+ (pow t_1 2.0) (pow t_2 2.0))))
(if (<= dX.u 50.0)
(if (>= (* dX.v (* dX.v (pow (floor h) 2.0))) t_3) t_6 (* t_2 t_5))
(if (>= (* dX.u (* dX.u (pow (floor w) 2.0))) t_7)
t_6
(*
t_2
(/
1.0
(/ 1.0 (pow (fmax (+ (pow t_0 2.0) (pow t_4 2.0)) t_7) -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 = floorf(w) * dY_46_u;
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 = 1.0f / sqrtf(fmaxf(((t_0 * t_0) + (t_4 * t_4)), t_3));
float t_6 = t_4 * t_5;
float t_7 = powf(t_1, 2.0f) + powf(t_2, 2.0f);
float tmp_1;
if (dX_46_u <= 50.0f) {
float tmp_2;
if ((dX_46_v * (dX_46_v * powf(floorf(h), 2.0f))) >= t_3) {
tmp_2 = t_6;
} else {
tmp_2 = t_2 * t_5;
}
tmp_1 = tmp_2;
} else if ((dX_46_u * (dX_46_u * powf(floorf(w), 2.0f))) >= t_7) {
tmp_1 = t_6;
} else {
tmp_1 = t_2 * (1.0f / (1.0f / powf(fmaxf((powf(t_0, 2.0f) + powf(t_4, 2.0f)), t_7), -0.5f)));
}
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) * dX_46_u) t_1 = Float32(floor(w) * dY_46_u) 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(Float32(1.0) / sqrt(((Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) != Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4))) ? t_3 : ((t_3 != t_3) ? Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) : max(Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)), t_3))))) t_6 = Float32(t_4 * t_5) t_7 = Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(50.0)) tmp_2 = Float32(0.0) if (Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0)))) >= t_3) tmp_2 = t_6; else tmp_2 = Float32(t_2 * t_5); end tmp_1 = tmp_2; elseif (Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))) >= t_7) tmp_1 = t_6; else tmp_1 = Float32(t_2 * Float32(Float32(1.0) / Float32(Float32(1.0) / (((Float32((t_0 ^ Float32(2.0)) + (t_4 ^ Float32(2.0))) != Float32((t_0 ^ Float32(2.0)) + (t_4 ^ Float32(2.0)))) ? t_7 : ((t_7 != t_7) ? Float32((t_0 ^ Float32(2.0)) + (t_4 ^ Float32(2.0))) : max(Float32((t_0 ^ Float32(2.0)) + (t_4 ^ Float32(2.0))), t_7))) ^ Float32(-0.5))))); 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) * dX_46_u; t_1 = floor(w) * dY_46_u; 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 = single(1.0) / sqrt(max(((t_0 * t_0) + (t_4 * t_4)), t_3)); t_6 = t_4 * t_5; t_7 = (t_1 ^ single(2.0)) + (t_2 ^ single(2.0)); tmp_2 = single(0.0); if (dX_46_u <= single(50.0)) tmp_3 = single(0.0); if ((dX_46_v * (dX_46_v * (floor(h) ^ single(2.0)))) >= t_3) tmp_3 = t_6; else tmp_3 = t_2 * t_5; end tmp_2 = tmp_3; elseif ((dX_46_u * (dX_46_u * (floor(w) ^ single(2.0)))) >= t_7) tmp_2 = t_6; else tmp_2 = t_2 * (single(1.0) / (single(1.0) / (max(((t_0 ^ single(2.0)) + (t_4 ^ single(2.0))), t_7) ^ single(-0.5)))); end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
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 := \frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_4 \cdot t\_4, t\_3\right)}}\\
t_6 := t\_4 \cdot t\_5\\
t_7 := {t\_1}^{2} + {t\_2}^{2}\\
\mathbf{if}\;dX.u \leq 50:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;dX.v \cdot \left(dX.v \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right) \geq t\_3:\\
\;\;\;\;t\_6\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot t\_5\\
\end{array}\\
\mathbf{elif}\;dX.u \cdot \left(dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right) \geq t\_7:\\
\;\;\;\;t\_6\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot \frac{1}{\frac{1}{{\left(\mathsf{max}\left({t\_0}^{2} + {t\_4}^{2}, t\_7\right)\right)}^{-0.5}}}\\
\end{array}
\end{array}
if dX.u < 50Initial program 79.8%
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.f3269.6%
Simplified69.6%
if 50 < dX.u Initial program 68.0%
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.f3266.6%
Simplified66.6%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3266.6%
Applied egg-rr66.6%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3266.6%
Applied egg-rr66.6%
pow1/2N/A
pow2N/A
pow2N/A
pow2N/A
pow2N/A
remove-double-divN/A
/-lowering-/.f32N/A
Applied egg-rr66.6%
Final simplification68.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* t_0 t_0))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor w) dY.u))
(t_4 (* (floor w) dX.u))
(t_5 (+ (* t_4 t_4) (* t_2 t_2))))
(if (>=
(* dX.u (* dX.u (pow (floor w) 2.0)))
(+ (pow t_3 2.0) (pow t_0 2.0)))
(* t_2 (/ 1.0 (sqrt (fmax t_5 (+ (* t_3 t_3) t_1)))))
(*
t_0
(/
1.0
(sqrt (fmax t_5 (+ t_1 (* (floor w) (* (floor w) (* dY.u dY.u)))))))))))
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 = t_0 * t_0;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(w) * dY_46_u;
float t_4 = floorf(w) * dX_46_u;
float t_5 = (t_4 * t_4) + (t_2 * t_2);
float tmp;
if ((dX_46_u * (dX_46_u * powf(floorf(w), 2.0f))) >= (powf(t_3, 2.0f) + powf(t_0, 2.0f))) {
tmp = t_2 * (1.0f / sqrtf(fmaxf(t_5, ((t_3 * t_3) + t_1))));
} else {
tmp = t_0 * (1.0f / sqrtf(fmaxf(t_5, (t_1 + (floorf(w) * (floorf(w) * (dY_46_u * dY_46_u)))))));
}
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(t_0 * t_0) t_2 = Float32(floor(h) * dX_46_v) 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_2 * t_2)) tmp = 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 = Float32(t_2 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_3 * t_3) + t_1) : ((Float32(Float32(t_3 * t_3) + t_1) != Float32(Float32(t_3 * t_3) + t_1)) ? t_5 : max(t_5, Float32(Float32(t_3 * t_3) + t_1))))))); else tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(t_1 + Float32(floor(w) * Float32(floor(w) * Float32(dY_46_u * dY_46_u)))) : ((Float32(t_1 + Float32(floor(w) * Float32(floor(w) * Float32(dY_46_u * dY_46_u)))) != Float32(t_1 + Float32(floor(w) * Float32(floor(w) * Float32(dY_46_u * dY_46_u))))) ? t_5 : max(t_5, Float32(t_1 + Float32(floor(w) * Float32(floor(w) * Float32(dY_46_u * dY_46_u)))))))))); 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 = t_0 * t_0; t_2 = floor(h) * dX_46_v; t_3 = floor(w) * dY_46_u; t_4 = floor(w) * dX_46_u; t_5 = (t_4 * t_4) + (t_2 * t_2); tmp = 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 = t_2 * (single(1.0) / sqrt(max(t_5, ((t_3 * t_3) + t_1)))); else tmp = t_0 * (single(1.0) / sqrt(max(t_5, (t_1 + (floor(w) * (floor(w) * (dY_46_u * dY_46_u))))))); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := t\_0 \cdot t\_0\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
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\_2 \cdot t\_2\\
\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\_2 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_3 \cdot t\_3 + t\_1\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_1 + \left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dY.u \cdot dY.u\right)\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.9%
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.f3263.0%
Simplified63.0%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3263.0%
Applied egg-rr63.0%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3263.0%
Applied egg-rr63.0%
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.f3263.0%
Applied egg-rr63.0%
Final simplification63.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) dX.u))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_1 t_1)) (+ (* t_2 t_2) (* t_0 t_0)))))))
(if (>=
(* dX.u (* dX.u (pow (floor w) 2.0)))
(+ (pow t_2 2.0) (pow t_0 2.0)))
(* t_1 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(h) * dY_46_v;
float t_1 = floorf(h) * dX_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_1 * t_1)), ((t_2 * t_2) + (t_0 * t_0))));
float tmp;
if ((dX_46_u * (dX_46_u * powf(floorf(w), 2.0f))) >= (powf(t_2, 2.0f) + powf(t_0, 2.0f))) {
tmp = t_1 * 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(h) * dY_46_v) t_1 = Float32(floor(h) * dX_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_1 * t_1)) != Float32(Float32(t_3 * t_3) + Float32(t_1 * t_1))) ? Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) : ((Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) != Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0))) ? Float32(Float32(t_3 * t_3) + Float32(t_1 * t_1)) : max(Float32(Float32(t_3 * t_3) + Float32(t_1 * t_1)), Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0))))))) tmp = Float32(0.0) if (Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))) >= Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) tmp = Float32(t_1 * 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(h) * dY_46_v; t_1 = floor(h) * dX_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_1 * t_1)), ((t_2 * t_2) + (t_0 * t_0)))); tmp = single(0.0); if ((dX_46_u * (dX_46_u * (floor(w) ^ single(2.0)))) >= ((t_2 ^ single(2.0)) + (t_0 ^ single(2.0)))) tmp = t_1 * t_4; else tmp = t_0 * t_4; 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 h\right\rfloor \cdot dX.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\_1 \cdot t\_1, t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right)}}\\
\mathbf{if}\;dX.u \cdot \left(dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right) \geq {t\_2}^{2} + {t\_0}^{2}:\\
\;\;\;\;t\_1 \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot t\_4\\
\end{array}
\end{array}
Initial program 76.9%
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.f3263.0%
Simplified63.0%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3263.0%
Applied egg-rr63.0%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3263.0%
Applied egg-rr63.0%
Final simplification63.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor w) dX.u))
(t_5 (* t_4 t_4)))
(if (>=
(* dX.u (* dX.u (pow (floor w) 2.0)))
(+ (pow t_2 2.0) (pow t_0 2.0)))
(* t_1 (/ 1.0 (sqrt (fmax (+ t_5 (* t_1 t_1)) t_3))))
(*
t_0
(/
1.0
(sqrt (fmax (+ t_5 (* dX.v (* dX.v (pow (floor h) 2.0)))) 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(h) * dY_46_v;
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(w) * dX_46_u;
float t_5 = t_4 * t_4;
float tmp;
if ((dX_46_u * (dX_46_u * powf(floorf(w), 2.0f))) >= (powf(t_2, 2.0f) + powf(t_0, 2.0f))) {
tmp = t_1 * (1.0f / sqrtf(fmaxf((t_5 + (t_1 * t_1)), t_3)));
} else {
tmp = t_0 * (1.0f / sqrtf(fmaxf((t_5 + (dX_46_v * (dX_46_v * powf(floorf(h), 2.0f)))), 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) * dY_46_v) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if (Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))) >= Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) tmp = Float32(t_1 * Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(t_1 * t_1)) != Float32(t_5 + Float32(t_1 * t_1))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(t_1 * t_1)) : max(Float32(t_5 + Float32(t_1 * t_1)), t_3)))))); else tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))) != Float32(t_5 + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0)))))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))) : max(Float32(t_5 + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))), 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(h) * dY_46_v; t_1 = floor(h) * dX_46_v; t_2 = floor(w) * dY_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(w) * dX_46_u; t_5 = t_4 * t_4; tmp = single(0.0); if ((dX_46_u * (dX_46_u * (floor(w) ^ single(2.0)))) >= ((t_2 ^ single(2.0)) + (t_0 ^ single(2.0)))) tmp = t_1 * (single(1.0) / sqrt(max((t_5 + (t_1 * t_1)), t_3))); else tmp = t_0 * (single(1.0) / sqrt(max((t_5 + (dX_46_v * (dX_46_v * (floor(h) ^ single(2.0))))), t_3))); 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 h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;dX.u \cdot \left(dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right) \geq {t\_2}^{2} + {t\_0}^{2}:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_1 \cdot t\_1, t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5 + dX.v \cdot \left(dX.v \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right), t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.9%
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.f3263.0%
Simplified63.0%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3263.0%
Applied egg-rr63.0%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3263.0%
Applied egg-rr63.0%
associate-*r*N/A
*-lowering-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3263.0%
Applied egg-rr63.0%
Final simplification63.0%
herbie shell --seed 2024191
(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))))