
(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 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_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 (* (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_0 t_6) (* t_4 t_6))))
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_0 * t_6;
} else {
tmp = t_4 * t_6;
}
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_0 * t_6); else tmp = Float32(t_4 * t_6); 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_0 * t_6; else tmp = t_4 * t_6; 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\_0 \cdot t\_6\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_6\\
\end{array}
\end{array}
Initial program 77.9%
Final simplification77.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (pow t_1 2.0))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dX.v))
(t_5 (+ (* t_3 t_3) (* t_4 t_4))))
(if (>= (+ (pow t_3 2.0) (pow t_4 2.0)) (+ (pow t_0 2.0) t_2))
(* t_4 (/ 1.0 (sqrt (fmax t_5 (+ (* t_0 t_0) (* t_1 t_1))))))
(*
t_1
(/
1.0
(sqrt (fmax t_5 (fma (* (floor w) (* dY.u dY.u)) (floor w) 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(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = powf(t_1, 2.0f);
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dX_46_v;
float t_5 = (t_3 * t_3) + (t_4 * t_4);
float tmp;
if ((powf(t_3, 2.0f) + powf(t_4, 2.0f)) >= (powf(t_0, 2.0f) + t_2)) {
tmp = t_4 * (1.0f / sqrtf(fmaxf(t_5, ((t_0 * t_0) + (t_1 * t_1)))));
} else {
tmp = t_1 * (1.0f / sqrtf(fmaxf(t_5, fmaf((floorf(w) * (dY_46_u * dY_46_u)), floorf(w), 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(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = t_1 ^ Float32(2.0) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(Float32(t_3 * t_3) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (Float32((t_3 ^ Float32(2.0)) + (t_4 ^ Float32(2.0))) >= Float32((t_0 ^ Float32(2.0)) + t_2)) tmp = Float32(t_4 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) : ((Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) != Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))) ? t_5 : max(t_5, Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)))))))); else tmp = Float32(t_1 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? fma(Float32(floor(w) * Float32(dY_46_u * dY_46_u)), floor(w), t_2) : ((fma(Float32(floor(w) * Float32(dY_46_u * dY_46_u)), floor(w), t_2) != fma(Float32(floor(w) * Float32(dY_46_u * dY_46_u)), floor(w), t_2)) ? t_5 : max(t_5, fma(Float32(floor(w) * Float32(dY_46_u * dY_46_u)), floor(w), t_2))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := {t\_1}^{2}\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := t\_3 \cdot t\_3 + t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_3}^{2} + {t\_4}^{2} \geq {t\_0}^{2} + t\_2:\\
\;\;\;\;t\_4 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, \mathsf{fma}\left(\left\lfloor w\right\rfloor \cdot \left(dY.u \cdot dY.u\right), \left\lfloor w\right\rfloor , t\_2\right)\right)}}\\
\end{array}
\end{array}
Initial program 77.9%
associate-*l*N/A
*-commutativeN/A
accelerator-lowering-fma.f32N/A
*-commutativeN/A
associate-*r*N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3277.9
Applied egg-rr77.9%
>=-lowering->=.f32N/A
+-lowering-+.f32N/A
pow2N/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-rr77.9%
Final simplification77.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* t_0 t_0))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (pow t_3 2.0))
(t_5 (* (floor w) dX.u)))
(if (>= (+ (pow t_5 2.0) (pow t_0 2.0)) (+ (pow t_2 2.0) t_4))
(*
t_0
(/
1.0
(sqrt
(fmax
(+ t_1 (* (pow (floor w) 2.0) (* dX.u dX.u)))
(+ (* t_2 t_2) (* t_3 t_3))))))
(*
t_3
(/
1.0
(sqrt
(fmax
(+ (* t_5 t_5) t_1)
(fma (* (floor w) (* dY.u dY.u)) (floor w) 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 = t_0 * t_0;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = powf(t_3, 2.0f);
float t_5 = floorf(w) * dX_46_u;
float tmp;
if ((powf(t_5, 2.0f) + powf(t_0, 2.0f)) >= (powf(t_2, 2.0f) + t_4)) {
tmp = t_0 * (1.0f / sqrtf(fmaxf((t_1 + (powf(floorf(w), 2.0f) * (dX_46_u * dX_46_u))), ((t_2 * t_2) + (t_3 * t_3)))));
} else {
tmp = t_3 * (1.0f / sqrtf(fmaxf(((t_5 * t_5) + t_1), fmaf((floorf(w) * (dY_46_u * dY_46_u)), floorf(w), 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(t_0 * t_0) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) >= Float32((t_2 ^ Float32(2.0)) + t_4)) tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(t_1 + Float32((floor(w) ^ Float32(2.0)) * Float32(dX_46_u * dX_46_u))) != Float32(t_1 + Float32((floor(w) ^ Float32(2.0)) * Float32(dX_46_u * dX_46_u)))) ? Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) : ((Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) != Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3))) ? Float32(t_1 + Float32((floor(w) ^ Float32(2.0)) * Float32(dX_46_u * dX_46_u))) : max(Float32(t_1 + Float32((floor(w) ^ Float32(2.0)) * Float32(dX_46_u * dX_46_u))), Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)))))))); else tmp = Float32(t_3 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_5 * t_5) + t_1) != Float32(Float32(t_5 * t_5) + t_1)) ? fma(Float32(floor(w) * Float32(dY_46_u * dY_46_u)), floor(w), t_4) : ((fma(Float32(floor(w) * Float32(dY_46_u * dY_46_u)), floor(w), t_4) != fma(Float32(floor(w) * Float32(dY_46_u * dY_46_u)), floor(w), t_4)) ? Float32(Float32(t_5 * t_5) + t_1) : max(Float32(Float32(t_5 * t_5) + t_1), fma(Float32(floor(w) * Float32(dY_46_u * dY_46_u)), floor(w), t_4))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := t\_0 \cdot t\_0\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := {t\_3}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;{t\_5}^{2} + {t\_0}^{2} \geq {t\_2}^{2} + t\_4:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_1 + {\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot \left(dX.u \cdot dX.u\right), t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5 \cdot t\_5 + t\_1, \mathsf{fma}\left(\left\lfloor w\right\rfloor \cdot \left(dY.u \cdot dY.u\right), \left\lfloor w\right\rfloor , t\_4\right)\right)}}\\
\end{array}
\end{array}
Initial program 77.9%
associate-*l*N/A
*-commutativeN/A
accelerator-lowering-fma.f32N/A
*-commutativeN/A
associate-*r*N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3277.9
Applied egg-rr77.9%
>=-lowering->=.f32N/A
+-lowering-+.f32N/A
pow2N/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-rr77.9%
pow2N/A
unpow-prod-downN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
*-lowering-*.f3277.9
Applied egg-rr77.9%
Final simplification77.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dY.v))
(t_2 (* t_1 t_1))
(t_3 (* (floor h) dX.v))
(t_4 (* t_3 t_3))
(t_5 (+ (* t_0 t_0) t_4))
(t_6 (* dY.u (* dY.u (pow (floor w) 2.0))))
(t_7 (* (floor w) dY.u))
(t_8 (+ (* t_7 t_7) t_2))
(t_9 (/ 1.0 (sqrt (fmax t_5 t_8)))))
(if (<= dX.u 260000.0)
(if (>= (pow t_3 2.0) (+ (pow t_7 2.0) (pow t_1 2.0)))
(* t_3 (/ 1.0 (sqrt (fmax (+ t_4 (pow t_0 2.0)) t_8))))
(* t_1 (/ 1.0 (sqrt (fmax t_5 (+ t_2 t_6))))))
(if (>= t_5 t_6) (* t_3 t_9) (* t_1 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 = floorf(w) * dX_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = t_1 * t_1;
float t_3 = floorf(h) * dX_46_v;
float t_4 = t_3 * t_3;
float t_5 = (t_0 * t_0) + t_4;
float t_6 = dY_46_u * (dY_46_u * powf(floorf(w), 2.0f));
float t_7 = floorf(w) * dY_46_u;
float t_8 = (t_7 * t_7) + t_2;
float t_9 = 1.0f / sqrtf(fmaxf(t_5, t_8));
float tmp_1;
if (dX_46_u <= 260000.0f) {
float tmp_2;
if (powf(t_3, 2.0f) >= (powf(t_7, 2.0f) + powf(t_1, 2.0f))) {
tmp_2 = t_3 * (1.0f / sqrtf(fmaxf((t_4 + powf(t_0, 2.0f)), t_8)));
} else {
tmp_2 = t_1 * (1.0f / sqrtf(fmaxf(t_5, (t_2 + t_6))));
}
tmp_1 = tmp_2;
} else if (t_5 >= t_6) {
tmp_1 = t_3 * t_9;
} else {
tmp_1 = t_1 * t_9;
}
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(h) * dY_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(t_3 * t_3) t_5 = Float32(Float32(t_0 * t_0) + t_4) t_6 = Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0)))) t_7 = Float32(floor(w) * dY_46_u) t_8 = Float32(Float32(t_7 * t_7) + t_2) t_9 = Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? t_8 : ((t_8 != t_8) ? t_5 : max(t_5, t_8))))) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(260000.0)) tmp_2 = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_7 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) tmp_2 = Float32(t_3 * Float32(Float32(1.0) / sqrt(((Float32(t_4 + (t_0 ^ Float32(2.0))) != Float32(t_4 + (t_0 ^ Float32(2.0)))) ? t_8 : ((t_8 != t_8) ? Float32(t_4 + (t_0 ^ Float32(2.0))) : max(Float32(t_4 + (t_0 ^ Float32(2.0))), t_8)))))); else tmp_2 = Float32(t_1 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(t_2 + t_6) : ((Float32(t_2 + t_6) != Float32(t_2 + t_6)) ? t_5 : max(t_5, Float32(t_2 + t_6))))))); end tmp_1 = tmp_2; elseif (t_5 >= t_6) tmp_1 = Float32(t_3 * t_9); else tmp_1 = Float32(t_1 * 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(w) * dX_46_u; t_1 = floor(h) * dY_46_v; t_2 = t_1 * t_1; t_3 = floor(h) * dX_46_v; t_4 = t_3 * t_3; t_5 = (t_0 * t_0) + t_4; t_6 = dY_46_u * (dY_46_u * (floor(w) ^ single(2.0))); t_7 = floor(w) * dY_46_u; t_8 = (t_7 * t_7) + t_2; t_9 = single(1.0) / sqrt(max(t_5, t_8)); tmp_2 = single(0.0); if (dX_46_u <= single(260000.0)) tmp_3 = single(0.0); if ((t_3 ^ single(2.0)) >= ((t_7 ^ single(2.0)) + (t_1 ^ single(2.0)))) tmp_3 = t_3 * (single(1.0) / sqrt(max((t_4 + (t_0 ^ single(2.0))), t_8))); else tmp_3 = t_1 * (single(1.0) / sqrt(max(t_5, (t_2 + t_6)))); end tmp_2 = tmp_3; elseif (t_5 >= t_6) tmp_2 = t_3 * t_9; else tmp_2 = t_1 * t_9; 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 h\right\rfloor \cdot dY.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := t\_3 \cdot t\_3\\
t_5 := t\_0 \cdot t\_0 + t\_4\\
t_6 := dY.u \cdot \left(dY.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right)\\
t_7 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_8 := t\_7 \cdot t\_7 + t\_2\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_8\right)}}\\
\mathbf{if}\;dX.u \leq 260000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_3}^{2} \geq {t\_7}^{2} + {t\_1}^{2}:\\
\;\;\;\;t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_4 + {t\_0}^{2}, t\_8\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 + t\_6\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_5 \geq t\_6:\\
\;\;\;\;t\_3 \cdot t\_9\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot t\_9\\
\end{array}
\end{array}
if dX.u < 2.6e5Initial program 81.9%
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.f3276.9
Simplified76.9%
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
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
Applied egg-rr76.9%
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.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%
if 2.6e5 < dX.u Initial program 58.4%
Taylor expanded in dY.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3249.7
Simplified49.7%
Final simplification72.3%
(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_2 t_2))
(t_4 (* (floor h) dX.v))
(t_5 (* t_4 t_4)))
(if (>= (pow t_4 2.0) (+ (pow t_1 2.0) (pow t_2 2.0)))
(* t_4 (/ 1.0 (sqrt (fmax (+ t_5 (pow t_0 2.0)) (+ (* t_1 t_1) t_3)))))
(*
t_2
(/
1.0
(sqrt
(fmax
(+ (* t_0 t_0) t_5)
(+ t_3 (* dY.u (* dY.u (pow (floor w) 2.0)))))))))))
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_2 * t_2;
float t_4 = floorf(h) * dX_46_v;
float t_5 = t_4 * t_4;
float tmp;
if (powf(t_4, 2.0f) >= (powf(t_1, 2.0f) + powf(t_2, 2.0f))) {
tmp = t_4 * (1.0f / sqrtf(fmaxf((t_5 + powf(t_0, 2.0f)), ((t_1 * t_1) + t_3))));
} else {
tmp = t_2 * (1.0f / sqrtf(fmaxf(((t_0 * t_0) + t_5), (t_3 + (dY_46_u * (dY_46_u * powf(floorf(w), 2.0f)))))));
}
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(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(t_2 * t_2) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(t_4 * Float32(Float32(1.0) / sqrt(((Float32(t_5 + (t_0 ^ Float32(2.0))) != Float32(t_5 + (t_0 ^ Float32(2.0)))) ? Float32(Float32(t_1 * t_1) + t_3) : ((Float32(Float32(t_1 * t_1) + t_3) != Float32(Float32(t_1 * t_1) + t_3)) ? Float32(t_5 + (t_0 ^ Float32(2.0))) : max(Float32(t_5 + (t_0 ^ Float32(2.0))), Float32(Float32(t_1 * t_1) + t_3))))))); else tmp = Float32(t_2 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_0 * t_0) + t_5) != Float32(Float32(t_0 * t_0) + t_5)) ? Float32(t_3 + Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0))))) : ((Float32(t_3 + Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0))))) != Float32(t_3 + Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0)))))) ? Float32(Float32(t_0 * t_0) + t_5) : max(Float32(Float32(t_0 * t_0) + t_5), Float32(t_3 + Float32(dY_46_u * Float32(dY_46_u * (floor(w) ^ Float32(2.0))))))))))); 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(w) * dY_46_u; t_2 = floor(h) * dY_46_v; t_3 = t_2 * t_2; t_4 = floor(h) * dX_46_v; t_5 = t_4 * t_4; tmp = single(0.0); if ((t_4 ^ single(2.0)) >= ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = t_4 * (single(1.0) / sqrt(max((t_5 + (t_0 ^ single(2.0))), ((t_1 * t_1) + t_3)))); else tmp = t_2 * (single(1.0) / sqrt(max(((t_0 * t_0) + t_5), (t_3 + (dY_46_u * (dY_46_u * (floor(w) ^ single(2.0)))))))); end tmp_2 = tmp; 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\_2 \cdot t\_2\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_4}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;t\_4 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5 + {t\_0}^{2}, t\_1 \cdot t\_1 + t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_5, t\_3 + dY.u \cdot \left(dY.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right)\right)}}\\
\end{array}
\end{array}
Initial program 77.9%
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.f3270.0
Simplified70.0%
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
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
Applied egg-rr70.0%
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.f3270.1
Applied egg-rr70.1%
pow2N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f3270.1
Applied egg-rr70.1%
Final simplification70.1%
herbie shell --seed 2024199
(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))))