
(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(fmax(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(fmax(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 (* dX.u (floor w)))
(t_1 (+ (pow (* dX.v (floor h)) 2.0) (pow t_0 2.0)))
(t_2 (* dY.u (floor w)))
(t_3 (+ (pow (* dY.v (floor h)) 2.0) (pow t_2 2.0))))
(if (>= t_1 t_3)
(/ (* t_0 (- -1.0)) (sqrt (fmax t_1 t_3)))
(/
(* t_2 (- -1.0))
(sqrt
(fmax
(+ (pow (* (floor h) dX.v) 2.0) (pow (* (floor w) dX.u) 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 = dX_46_u * floorf(w);
float t_1 = powf((dX_46_v * floorf(h)), 2.0f) + powf(t_0, 2.0f);
float t_2 = dY_46_u * floorf(w);
float t_3 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_2, 2.0f);
float tmp;
if (t_1 >= t_3) {
tmp = (t_0 * -(-1.0f)) / sqrtf(fmaxf(t_1, t_3));
} else {
tmp = (t_2 * -(-1.0f)) / sqrtf(fmaxf((powf((floorf(h) * dX_46_v), 2.0f) + powf((floorf(w) * dX_46_u), 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(dX_46_u * floor(w)) t_1 = Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(Float32(t_0 * Float32(-Float32(-1.0))) / sqrt(fmax(t_1, t_3))); else tmp = Float32(Float32(t_2 * Float32(-Float32(-1.0))) / sqrt(fmax(Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ 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 = dX_46_u * floor(w); t_1 = ((dX_46_v * floor(h)) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = dY_46_u * floor(w); t_3 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_2 ^ single(2.0)); tmp = single(0.0); if (t_1 >= t_3) tmp = (t_0 * -single(-1.0)) / sqrt(max(t_1, t_3)); else tmp = (t_2 * -single(-1.0)) / sqrt(max((((floor(h) * dX_46_v) ^ single(2.0)) + ((floor(w) * dX_46_u) ^ single(2.0))), t_3)); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_0}^{2}\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_2}^{2}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0 \cdot \left(--1\right)}{\sqrt{\mathsf{max}\left(t\_1, t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2 \cdot \left(--1\right)}{\sqrt{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, t\_3\right)}}\\
\end{array}
\end{array}
Initial program 77.3%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
Applied rewrites77.4%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-+.f3277.4
Applied rewrites77.4%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-+.f3277.4
lift-+.f32N/A
+-commutativeN/A
Applied rewrites77.4%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
lift-neg.f32N/A
associate-*r/N/A
Applied rewrites77.5%
Final simplification77.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(t_1 (+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0)))
(t_2 (* (floor w) dX.u)))
(if (>= t_1 t_0)
(/ t_2 (sqrt (fmax (+ (pow (* (floor h) dX.v) 2.0) (pow t_2 2.0)) t_0)))
(* (/ 1.0 (sqrt (fmax t_1 t_0))) (* (floor w) 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 = powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f);
float t_1 = powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f);
float t_2 = floorf(w) * dX_46_u;
float tmp;
if (t_1 >= t_0) {
tmp = t_2 / sqrtf(fmaxf((powf((floorf(h) * dX_46_v), 2.0f) + powf(t_2, 2.0f)), t_0));
} else {
tmp = (1.0f / sqrtf(fmaxf(t_1, t_0))) * (floorf(w) * 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((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))) t_1 = Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))) t_2 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 / sqrt(fmax(Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))), t_0))); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_1, t_0))) * Float32(floor(w) * 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 = ((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)); t_1 = ((dX_46_v * floor(h)) ^ single(2.0)) + ((dX_46_u * floor(w)) ^ single(2.0)); t_2 = floor(w) * dX_46_u; tmp = single(0.0); if (t_1 >= t_0) tmp = t_2 / sqrt(max((((floor(h) * dX_46_v) ^ single(2.0)) + (t_2 ^ single(2.0))), t_0)); else tmp = (single(1.0) / sqrt(max(t_1, t_0))) * (floor(w) * dY_46_u); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;\frac{t\_2}{\sqrt{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {t\_2}^{2}, t\_0\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}} \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
\end{array}
\end{array}
Initial program 77.3%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
Applied rewrites77.4%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-+.f3277.4
Applied rewrites77.4%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-+.f3277.4
lift-+.f32N/A
+-commutativeN/A
Applied rewrites77.4%
lift-/.f32N/A
lift-neg.f32N/A
distribute-frac-neg2N/A
distribute-neg-fracN/A
lower-/.f32N/A
Applied rewrites77.4%
Final simplification77.4%
(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 (sqrt (/ 1.0 (fmax t_1 t_3)))))
(if (>= t_1 t_3) (* t_4 t_0) (* t_4 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) * 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 = sqrtf((1.0f / fmaxf(t_1, t_3)));
float tmp;
if (t_1 >= t_3) {
tmp = t_4 * t_0;
} else {
tmp = t_4 * 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) * 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 = sqrt(Float32(Float32(1.0) / fmax(t_1, t_3))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_4 * t_0); else tmp = Float32(t_4 * 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(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 = sqrt((single(1.0) / max(t_1, t_3))); tmp = single(0.0); if (t_1 >= t_3) tmp = t_4 * t_0; else tmp = t_4 * t_2; 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 := \sqrt{\frac{1}{\mathsf{max}\left(t\_1, t\_3\right)}}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;t\_4 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_2\\
\end{array}
\end{array}
Initial program 77.3%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
Applied rewrites77.4%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-+.f3277.4
Applied rewrites77.4%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-+.f3277.4
lift-+.f32N/A
+-commutativeN/A
Applied rewrites77.4%
Taylor expanded in w around 0
Applied rewrites77.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* t_1 t_1))
(t_3 (* (floor h) dY.v))
(t_4 (+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(t_5 (* (floor w) dX.u))
(t_6 (+ (* t_5 t_5) (* t_0 t_0)))
(t_7 (/ 1.0 (sqrt (fmax t_6 (+ t_2 (* t_3 t_3))))))
(t_8 (* t_7 t_5)))
(if (<= dX.v 3.799999952316284)
(if (>= (pow (* dX.u (floor w)) 2.0) t_4)
t_8
(*
(/ 1.0 (sqrt (fmax t_6 (+ t_2 (* (* (pow (floor h) 2.0) dY.v) dY.v)))))
t_1))
(if (>= (pow t_0 2.0) t_4) t_8 (* t_7 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 = t_1 * t_1;
float t_3 = floorf(h) * dY_46_v;
float t_4 = powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f);
float t_5 = floorf(w) * dX_46_u;
float t_6 = (t_5 * t_5) + (t_0 * t_0);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, (t_2 + (t_3 * t_3))));
float t_8 = t_7 * t_5;
float tmp_1;
if (dX_46_v <= 3.799999952316284f) {
float tmp_2;
if (powf((dX_46_u * floorf(w)), 2.0f) >= t_4) {
tmp_2 = t_8;
} else {
tmp_2 = (1.0f / sqrtf(fmaxf(t_6, (t_2 + ((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v))))) * t_1;
}
tmp_1 = tmp_2;
} else if (powf(t_0, 2.0f) >= t_4) {
tmp_1 = t_8;
} else {
tmp_1 = t_7 * t_1;
}
return tmp_1;
}
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(t_1 * t_1) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)) t_7 = Float32(Float32(1.0) / sqrt(fmax(t_6, Float32(t_2 + Float32(t_3 * t_3))))) t_8 = Float32(t_7 * t_5) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(3.799999952316284)) tmp_2 = Float32(0.0) if ((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) >= t_4) tmp_2 = t_8; else tmp_2 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_6, Float32(t_2 + Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v))))) * t_1); end tmp_1 = tmp_2; elseif ((t_0 ^ Float32(2.0)) >= t_4) tmp_1 = t_8; else tmp_1 = Float32(t_7 * t_1); 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) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = t_1 * t_1; t_3 = floor(h) * dY_46_v; t_4 = ((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)); t_5 = floor(w) * dX_46_u; t_6 = (t_5 * t_5) + (t_0 * t_0); t_7 = single(1.0) / sqrt(max(t_6, (t_2 + (t_3 * t_3)))); t_8 = t_7 * t_5; tmp_2 = single(0.0); if (dX_46_v <= single(3.799999952316284)) tmp_3 = single(0.0); if (((dX_46_u * floor(w)) ^ single(2.0)) >= t_4) tmp_3 = t_8; else tmp_3 = (single(1.0) / sqrt(max(t_6, (t_2 + (((floor(h) ^ single(2.0)) * dY_46_v) * dY_46_v))))) * t_1; end tmp_2 = tmp_3; elseif ((t_0 ^ single(2.0)) >= t_4) tmp_2 = t_8; else tmp_2 = t_7 * t_1; end tmp_4 = tmp_2; 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 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := t\_5 \cdot t\_5 + t\_0 \cdot t\_0\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_2 + t\_3 \cdot t\_3\right)}}\\
t_8 := t\_7 \cdot t\_5\\
\mathbf{if}\;dX.v \leq 3.799999952316284:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_2 + \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v\right)}} \cdot t\_1\\
\end{array}\\
\mathbf{elif}\;{t\_0}^{2} \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_1\\
\end{array}
\end{array}
if dX.v < 3.79999995Initial program 80.3%
Taylor expanded in dX.u around inf
Applied rewrites71.1%
lift-*.f32N/A
pow2N/A
lower-pow.f3271.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3271.1
Applied rewrites71.1%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites71.1%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
lift-*.f32N/A
lift-*.f32N/A
sqr-neg-revN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-neg.f32N/A
lower-neg.f3271.1
Applied rewrites71.1%
if 3.79999995 < dX.v Initial program 67.8%
Taylor expanded in dX.u around inf
Applied rewrites52.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3252.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3252.7
Applied rewrites52.7%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites52.7%
Taylor expanded in dX.u around 0
Applied rewrites65.4%
Final simplification69.7%
(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 (+ (* t_0 t_0) (* t_1 t_1)))
(t_3 (* (floor w) dX.u))
(t_4 (* t_3 t_3))
(t_5 (* (floor h) dX.v)))
(if (>=
(pow (* dX.u (floor w)) 2.0)
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(*
(/ 1.0 (sqrt (fmax (+ t_4 (* (pow (floor h) 2.0) (* dX.v dX.v))) t_2)))
t_3)
(* (/ 1.0 (sqrt (fmax (+ t_4 (* t_5 t_5)) t_2))) t_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) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = (t_0 * t_0) + (t_1 * t_1);
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 (powf((dX_46_u * floorf(w)), 2.0f) >= (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((t_4 + (powf(floorf(h), 2.0f) * (dX_46_v * dX_46_v))), t_2))) * t_3;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_4 + (t_5 * t_5)), t_2))) * t_0;
}
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 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) 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 * floor(w)) ^ Float32(2.0)) >= Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_4 + Float32((floor(h) ^ Float32(2.0)) * Float32(dX_46_v * dX_46_v))), t_2))) * t_3); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_4 + Float32(t_5 * t_5)), t_2))) * t_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) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = (t_0 * t_0) + (t_1 * t_1); 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 * floor(w)) ^ single(2.0)) >= (((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((t_4 + ((floor(h) ^ single(2.0)) * (dX_46_v * dX_46_v))), t_2))) * t_3; else tmp = (single(1.0) / sqrt(max((t_4 + (t_5 * t_5)), t_2))) * t_0; end tmp_2 = 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\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
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}\;{\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} \geq {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 + {\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot \left(dX.v \cdot dX.v\right), t\_2\right)}} \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 + t\_5 \cdot t\_5, t\_2\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 77.3%
Taylor expanded in dX.u around inf
Applied rewrites66.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3266.6
Applied rewrites66.6%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites66.6%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3266.6
Applied rewrites66.6%
(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 h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2)))))))
(if (>=
(pow (* dX.u (floor w)) 2.0)
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(* t_4 t_3)
(* t_4 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(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))));
float tmp;
if (powf((dX_46_u * floorf(w)), 2.0f) >= (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))) {
tmp = t_4 * t_3;
} else {
tmp = t_4 * 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(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))))) tmp = Float32(0.0) if ((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) >= Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0)))) tmp = Float32(t_4 * t_3); else tmp = Float32(t_4 * 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(h) * dY_46_v; t_3 = floor(w) * dX_46_u; t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))); tmp = single(0.0); if (((dX_46_u * floor(w)) ^ single(2.0)) >= (((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)))) tmp = t_4 * t_3; else tmp = t_4 * 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 h\right\rfloor \cdot dY.v\\
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\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}\\
\mathbf{if}\;{\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} \geq {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_4 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_1\\
\end{array}
\end{array}
Initial program 77.3%
Taylor expanded in dX.u around inf
Applied rewrites66.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3266.6
Applied rewrites66.6%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites66.6%
herbie shell --seed 2025019
(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))))