
(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 11 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 (* (floor w) dX.u))
(t_1 (* (floor w) dY.u))
(t_2 (+ (pow (* (floor h) dX.v) 2.0) (pow t_0 2.0)))
(t_3 (pow t_1 2.0))
(t_4 (+ (pow (* (floor h) dY.v) 2.0) t_3)))
(if (>= t_2 t_4)
(/ t_0 (sqrt (fmax t_2 t_4)))
(/ t_1 (sqrt (fmax t_2 (fma (pow (floor h) 2.0) (* dY.v dY.v) t_3)))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(w) * dY_46_u;
float t_2 = powf((floorf(h) * dX_46_v), 2.0f) + powf(t_0, 2.0f);
float t_3 = powf(t_1, 2.0f);
float t_4 = powf((floorf(h) * dY_46_v), 2.0f) + t_3;
float tmp;
if (t_2 >= t_4) {
tmp = t_0 / sqrtf(fmaxf(t_2, t_4));
} else {
tmp = t_1 / sqrtf(fmaxf(t_2, fmaf(powf(floorf(h), 2.0f), (dY_46_v * dY_46_v), t_3)));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_3 = t_1 ^ Float32(2.0) t_4 = Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_3) tmp = Float32(0.0) if (t_2 >= t_4) tmp = Float32(t_0 / sqrt(fmax(t_2, t_4))); else tmp = Float32(t_1 / sqrt(fmax(t_2, fma((floor(h) ^ Float32(2.0)), Float32(dY_46_v * dY_46_v), t_3)))); end return 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(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {t\_0}^{2}\\
t_3 := {t\_1}^{2}\\
t_4 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + t\_3\\
\mathbf{if}\;t\_2 \geq t\_4:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_2, t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left({\left(\left\lfloor h\right\rfloor \right)}^{2}, dY.v \cdot dY.v, t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 74.8%
Applied rewrites75.1%
lift-+.f32N/A
lift-pow.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
lower-fma.f32N/A
lift-pow.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3275.1
lift-*.f32N/A
pow2N/A
lift-pow.f3275.1
Applied rewrites75.1%
(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) dX.u))
(t_2 (pow t_0 2.0))
(t_3 (* (floor w) dY.u))
(t_4 (pow t_3 2.0))
(t_5 (+ (* t_1 t_1) (* t_0 t_0)))
(t_6 (* (floor h) dY.v))
(t_7 (+ (* t_3 t_3) (* t_6 t_6)))
(t_8 (+ (pow t_6 2.0) t_4))
(t_9 (/ 1.0 (sqrt (fmax t_5 t_7))))
(t_10 (+ (pow t_1 2.0) t_2))
(t_11 (sqrt (fmax t_10 t_8)))
(t_12 (/ t_1 t_11)))
(if (<= (if (>= t_5 t_7) (* t_9 t_1) (* t_9 t_3)) 4.0000000467443897e-7)
(if (>= t_2 t_8) t_12 (/ t_3 t_11))
(if (>= t_2 t_4)
t_12
(/
t_3
(sqrt (fmax t_10 (fma (pow (floor h) 2.0) (* dY.v 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) * dX_46_u;
float t_2 = powf(t_0, 2.0f);
float t_3 = floorf(w) * dY_46_u;
float t_4 = powf(t_3, 2.0f);
float t_5 = (t_1 * t_1) + (t_0 * t_0);
float t_6 = floorf(h) * dY_46_v;
float t_7 = (t_3 * t_3) + (t_6 * t_6);
float t_8 = powf(t_6, 2.0f) + t_4;
float t_9 = 1.0f / sqrtf(fmaxf(t_5, t_7));
float t_10 = powf(t_1, 2.0f) + t_2;
float t_11 = sqrtf(fmaxf(t_10, t_8));
float t_12 = t_1 / t_11;
float tmp;
if (t_5 >= t_7) {
tmp = t_9 * t_1;
} else {
tmp = t_9 * t_3;
}
float tmp_2;
if (tmp <= 4.0000000467443897e-7f) {
float tmp_3;
if (t_2 >= t_8) {
tmp_3 = t_12;
} else {
tmp_3 = t_3 / t_11;
}
tmp_2 = tmp_3;
} else if (t_2 >= t_4) {
tmp_2 = t_12;
} else {
tmp_2 = t_3 / sqrtf(fmaxf(t_10, fmaf(powf(floorf(h), 2.0f), (dY_46_v * dY_46_v), t_4)));
}
return tmp_2;
}
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) * dX_46_u) t_2 = t_0 ^ Float32(2.0) t_3 = Float32(floor(w) * dY_46_u) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) t_6 = Float32(floor(h) * dY_46_v) t_7 = Float32(Float32(t_3 * t_3) + Float32(t_6 * t_6)) t_8 = Float32((t_6 ^ Float32(2.0)) + t_4) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_5, t_7))) t_10 = Float32((t_1 ^ Float32(2.0)) + t_2) t_11 = sqrt(fmax(t_10, t_8)) t_12 = Float32(t_1 / t_11) tmp = Float32(0.0) if (t_5 >= t_7) tmp = Float32(t_9 * t_1); else tmp = Float32(t_9 * t_3); end tmp_2 = Float32(0.0) if (tmp <= Float32(4.0000000467443897e-7)) tmp_3 = Float32(0.0) if (t_2 >= t_8) tmp_3 = t_12; else tmp_3 = Float32(t_3 / t_11); end tmp_2 = tmp_3; elseif (t_2 >= t_4) tmp_2 = t_12; else tmp_2 = Float32(t_3 / sqrt(fmax(t_10, fma((floor(h) ^ Float32(2.0)), Float32(dY_46_v * dY_46_v), t_4)))); end return 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 dX.u\\
t_2 := {t\_0}^{2}\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := {t\_3}^{2}\\
t_5 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_7 := t\_3 \cdot t\_3 + t\_6 \cdot t\_6\\
t_8 := {t\_6}^{2} + t\_4\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_7\right)}}\\
t_10 := {t\_1}^{2} + t\_2\\
t_11 := \sqrt{\mathsf{max}\left(t\_10, t\_8\right)}\\
t_12 := \frac{t\_1}{t\_11}\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_7:\\
\;\;\;\;t\_9 \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_3\\
\end{array} \leq 4.0000000467443897 \cdot 10^{-7}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_2 \geq t\_8:\\
\;\;\;\;t\_12\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_3}{t\_11}\\
\end{array}\\
\mathbf{elif}\;t\_2 \geq t\_4:\\
\;\;\;\;t\_12\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(t\_10, \mathsf{fma}\left({\left(\left\lfloor h\right\rfloor \right)}^{2}, dY.v \cdot dY.v, t\_4\right)\right)}}\\
\end{array}
\end{array}
if (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dY.u))) < 4.00000005e-7Initial program 66.2%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3261.3
Applied rewrites61.3%
Applied rewrites61.5%
if 4.00000005e-7 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dY.u))) Initial program 99.0%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3272.3
Applied rewrites72.3%
Applied rewrites72.8%
Taylor expanded in dY.u around inf
unpow-prod-downN/A
pow2N/A
*-commutativeN/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-pow.f3288.3
Applied rewrites88.3%
lift-+.f32N/A
lift-pow.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
lower-fma.f32N/A
lower-pow.f32N/A
lift-floor.f32N/A
pow2N/A
lower-*.f3288.4
Applied rewrites88.4%
(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 (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 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(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(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / 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(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(fmax(t_1, t_3)) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / 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(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(max(t_1, t_3)); tmp = single(0.0); if (t_1 >= t_3) tmp = t_0 / t_4; else tmp = t_2 / 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 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 74.8%
Applied rewrites75.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (* (floor w) dX.u) 2.0))
(t_1 (pow (* (floor h) dX.v) 2.0))
(t_2 (+ t_1 t_0))
(t_3 (* (floor w) dY.u))
(t_4 (+ (pow (* (floor h) dY.v) 2.0) (pow t_3 2.0))))
(if (>= t_2 t_4)
(* (/ dX.u (sqrt (fmax (+ t_0 t_1) t_4))) (floor w))
(/ t_3 (sqrt (fmax 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 = powf((floorf(w) * dX_46_u), 2.0f);
float t_1 = powf((floorf(h) * dX_46_v), 2.0f);
float t_2 = t_1 + t_0;
float t_3 = floorf(w) * dY_46_u;
float t_4 = powf((floorf(h) * dY_46_v), 2.0f) + powf(t_3, 2.0f);
float tmp;
if (t_2 >= t_4) {
tmp = (dX_46_u / sqrtf(fmaxf((t_0 + t_1), t_4))) * floorf(w);
} else {
tmp = t_3 / sqrtf(fmaxf(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(w) * dX_46_u) ^ Float32(2.0) t_1 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_2 = Float32(t_1 + t_0) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) tmp = Float32(0.0) if (t_2 >= t_4) tmp = Float32(Float32(dX_46_u / sqrt(fmax(Float32(t_0 + t_1), t_4))) * floor(w)); else tmp = Float32(t_3 / sqrt(fmax(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(w) * dX_46_u) ^ single(2.0); t_1 = (floor(h) * dX_46_v) ^ single(2.0); t_2 = t_1 + t_0; t_3 = floor(w) * dY_46_u; t_4 = ((floor(h) * dY_46_v) ^ single(2.0)) + (t_3 ^ single(2.0)); tmp = single(0.0); if (t_2 >= t_4) tmp = (dX_46_u / sqrt(max((t_0 + t_1), t_4))) * floor(w); else tmp = t_3 / sqrt(max(t_2, t_4)); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\\
t_1 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_2 := t\_1 + t\_0\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + {t\_3}^{2}\\
\mathbf{if}\;t\_2 \geq t\_4:\\
\;\;\;\;\frac{dX.u}{\sqrt{\mathsf{max}\left(t\_0 + t\_1, t\_4\right)}} \cdot \left\lfloor w\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(t\_2, t\_4\right)}}\\
\end{array}
\end{array}
Initial program 74.8%
Applied rewrites75.1%
Applied rewrites74.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (* (floor h) dX.v) 2.0))
(t_1 (* (floor w) dY.u))
(t_2 (pow t_1 2.0))
(t_3 (+ (pow (* (floor h) dY.v) 2.0) t_2))
(t_4 (* (floor w) dX.u))
(t_5 (pow t_4 2.0))
(t_6 (sqrt (fmax (+ t_5 t_0) t_3)))
(t_7 (+ t_0 t_5)))
(if (<= dX.v 58.0)
(if (>= t_5 t_3)
(/ t_4 (sqrt (fmax t_7 t_3)))
(/ t_1 (sqrt (fmax t_7 (fma (pow (floor h) 2.0) (* dY.v dY.v) t_2)))))
(if (>= t_0 t_3) (/ t_4 t_6) (/ t_1 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 = powf((floorf(h) * dX_46_v), 2.0f);
float t_1 = floorf(w) * dY_46_u;
float t_2 = powf(t_1, 2.0f);
float t_3 = powf((floorf(h) * dY_46_v), 2.0f) + t_2;
float t_4 = floorf(w) * dX_46_u;
float t_5 = powf(t_4, 2.0f);
float t_6 = sqrtf(fmaxf((t_5 + t_0), t_3));
float t_7 = t_0 + t_5;
float tmp_1;
if (dX_46_v <= 58.0f) {
float tmp_2;
if (t_5 >= t_3) {
tmp_2 = t_4 / sqrtf(fmaxf(t_7, t_3));
} else {
tmp_2 = t_1 / sqrtf(fmaxf(t_7, fmaf(powf(floorf(h), 2.0f), (dY_46_v * dY_46_v), t_2)));
}
tmp_1 = tmp_2;
} else if (t_0 >= t_3) {
tmp_1 = t_4 / t_6;
} else {
tmp_1 = t_1 / t_6;
}
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) ^ Float32(2.0) t_1 = Float32(floor(w) * dY_46_u) t_2 = t_1 ^ Float32(2.0) t_3 = Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_2) t_4 = Float32(floor(w) * dX_46_u) t_5 = t_4 ^ Float32(2.0) t_6 = sqrt(fmax(Float32(t_5 + t_0), t_3)) t_7 = Float32(t_0 + t_5) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(58.0)) tmp_2 = Float32(0.0) if (t_5 >= t_3) tmp_2 = Float32(t_4 / sqrt(fmax(t_7, t_3))); else tmp_2 = Float32(t_1 / sqrt(fmax(t_7, fma((floor(h) ^ Float32(2.0)), Float32(dY_46_v * dY_46_v), t_2)))); end tmp_1 = tmp_2; elseif (t_0 >= t_3) tmp_1 = Float32(t_4 / t_6); else tmp_1 = Float32(t_1 / t_6); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := {t\_1}^{2}\\
t_3 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + t\_2\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := {t\_4}^{2}\\
t_6 := \sqrt{\mathsf{max}\left(t\_5 + t\_0, t\_3\right)}\\
t_7 := t\_0 + t\_5\\
\mathbf{if}\;dX.v \leq 58:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_3:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(t\_7, t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_7, \mathsf{fma}\left({\left(\left\lfloor h\right\rfloor \right)}^{2}, dY.v \cdot dY.v, t\_2\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_0 \geq t\_3:\\
\;\;\;\;\frac{t\_4}{t\_6}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_6}\\
\end{array}
\end{array}
if dX.v < 58Initial program 76.2%
Applied rewrites76.5%
lift-+.f32N/A
lift-pow.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
lower-fma.f32N/A
lift-pow.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3276.6
lift-*.f32N/A
pow2N/A
lift-pow.f3276.6
Applied rewrites76.6%
Taylor expanded in dX.u around inf
unpow-prod-downN/A
*-commutativeN/A
unpow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
*-commutativeN/A
unpow-prod-downN/A
*-commutativeN/A
unpow-prod-downN/A
Applied rewrites70.2%
if 58 < dX.v Initial program 70.4%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3268.8
Applied rewrites68.8%
Applied rewrites68.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (* (floor h) dX.v) 2.0))
(t_1 (* (floor w) dY.u))
(t_2 (+ (pow (* (floor h) dY.v) 2.0) (pow t_1 2.0)))
(t_3 (* (floor w) dX.u))
(t_4 (pow t_3 2.0))
(t_5 (sqrt (fmax (+ t_4 t_0) t_2)))
(t_6 (sqrt (fmax (+ t_0 t_4) t_2))))
(if (<= dX.v 58.0)
(if (>= t_4 t_2) (/ t_3 t_6) (/ t_1 t_6))
(if (>= t_0 t_2) (/ t_3 t_5) (/ t_1 t_5)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf((floorf(h) * dX_46_v), 2.0f);
float t_1 = floorf(w) * dY_46_u;
float t_2 = powf((floorf(h) * dY_46_v), 2.0f) + powf(t_1, 2.0f);
float t_3 = floorf(w) * dX_46_u;
float t_4 = powf(t_3, 2.0f);
float t_5 = sqrtf(fmaxf((t_4 + t_0), t_2));
float t_6 = sqrtf(fmaxf((t_0 + t_4), t_2));
float tmp_1;
if (dX_46_v <= 58.0f) {
float tmp_2;
if (t_4 >= t_2) {
tmp_2 = t_3 / t_6;
} else {
tmp_2 = t_1 / t_6;
}
tmp_1 = tmp_2;
} else if (t_0 >= t_2) {
tmp_1 = t_3 / t_5;
} else {
tmp_1 = t_1 / t_5;
}
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) ^ Float32(2.0) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_3 = Float32(floor(w) * dX_46_u) t_4 = t_3 ^ Float32(2.0) t_5 = sqrt(fmax(Float32(t_4 + t_0), t_2)) t_6 = sqrt(fmax(Float32(t_0 + t_4), t_2)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(58.0)) tmp_2 = Float32(0.0) if (t_4 >= t_2) tmp_2 = Float32(t_3 / t_6); else tmp_2 = Float32(t_1 / t_6); end tmp_1 = tmp_2; elseif (t_0 >= t_2) tmp_1 = Float32(t_3 / t_5); else tmp_1 = Float32(t_1 / t_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(h) * dX_46_v) ^ single(2.0); t_1 = floor(w) * dY_46_u; t_2 = ((floor(h) * dY_46_v) ^ single(2.0)) + (t_1 ^ single(2.0)); t_3 = floor(w) * dX_46_u; t_4 = t_3 ^ single(2.0); t_5 = sqrt(max((t_4 + t_0), t_2)); t_6 = sqrt(max((t_0 + t_4), t_2)); tmp_2 = single(0.0); if (dX_46_v <= single(58.0)) tmp_3 = single(0.0); if (t_4 >= t_2) tmp_3 = t_3 / t_6; else tmp_3 = t_1 / t_6; end tmp_2 = tmp_3; elseif (t_0 >= t_2) tmp_2 = t_3 / t_5; else tmp_2 = t_1 / t_5; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + {t\_1}^{2}\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := {t\_3}^{2}\\
t_5 := \sqrt{\mathsf{max}\left(t\_4 + t\_0, t\_2\right)}\\
t_6 := \sqrt{\mathsf{max}\left(t\_0 + t\_4, t\_2\right)}\\
\mathbf{if}\;dX.v \leq 58:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_2:\\
\;\;\;\;\frac{t\_3}{t\_6}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_6}\\
\end{array}\\
\mathbf{elif}\;t\_0 \geq t\_2:\\
\;\;\;\;\frac{t\_3}{t\_5}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_5}\\
\end{array}
\end{array}
if dX.v < 58Initial program 76.2%
Applied rewrites76.5%
Taylor expanded in dX.u around inf
Applied rewrites70.2%
if 58 < dX.v Initial program 70.4%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3268.8
Applied rewrites68.8%
Applied rewrites68.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (* (floor h) dX.v) 2.0))
(t_1 (pow (* (floor h) dY.v) 2.0))
(t_2 (* (floor w) dY.u))
(t_3 (pow t_2 2.0))
(t_4 (+ t_1 t_3))
(t_5 (* (floor w) dX.u))
(t_6 (pow t_5 2.0))
(t_7 (+ t_6 t_0))
(t_8 (sqrt (fmax t_7 t_4))))
(if (<= dX.v 58.0)
(if (>= t_6 t_4)
(/ t_5 (sqrt (fmax (+ t_0 t_6) t_4)))
(* (/ dY.u (sqrt (fmax t_7 (+ t_3 t_1)))) (floor w)))
(if (>= t_0 t_4) (/ t_5 t_8) (/ t_2 t_8)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf((floorf(h) * dX_46_v), 2.0f);
float t_1 = powf((floorf(h) * dY_46_v), 2.0f);
float t_2 = floorf(w) * dY_46_u;
float t_3 = powf(t_2, 2.0f);
float t_4 = t_1 + t_3;
float t_5 = floorf(w) * dX_46_u;
float t_6 = powf(t_5, 2.0f);
float t_7 = t_6 + t_0;
float t_8 = sqrtf(fmaxf(t_7, t_4));
float tmp_1;
if (dX_46_v <= 58.0f) {
float tmp_2;
if (t_6 >= t_4) {
tmp_2 = t_5 / sqrtf(fmaxf((t_0 + t_6), t_4));
} else {
tmp_2 = (dY_46_u / sqrtf(fmaxf(t_7, (t_3 + t_1)))) * floorf(w);
}
tmp_1 = tmp_2;
} else if (t_0 >= t_4) {
tmp_1 = t_5 / t_8;
} else {
tmp_1 = t_2 / t_8;
}
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) ^ Float32(2.0) t_1 = Float32(floor(h) * dY_46_v) ^ Float32(2.0) t_2 = Float32(floor(w) * dY_46_u) t_3 = t_2 ^ Float32(2.0) t_4 = Float32(t_1 + t_3) t_5 = Float32(floor(w) * dX_46_u) t_6 = t_5 ^ Float32(2.0) t_7 = Float32(t_6 + t_0) t_8 = sqrt(fmax(t_7, t_4)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(58.0)) tmp_2 = Float32(0.0) if (t_6 >= t_4) tmp_2 = Float32(t_5 / sqrt(fmax(Float32(t_0 + t_6), t_4))); else tmp_2 = Float32(Float32(dY_46_u / sqrt(fmax(t_7, Float32(t_3 + t_1)))) * floor(w)); end tmp_1 = tmp_2; elseif (t_0 >= t_4) tmp_1 = Float32(t_5 / t_8); else tmp_1 = Float32(t_2 / t_8); 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) ^ single(2.0); t_1 = (floor(h) * dY_46_v) ^ single(2.0); t_2 = floor(w) * dY_46_u; t_3 = t_2 ^ single(2.0); t_4 = t_1 + t_3; t_5 = floor(w) * dX_46_u; t_6 = t_5 ^ single(2.0); t_7 = t_6 + t_0; t_8 = sqrt(max(t_7, t_4)); tmp_2 = single(0.0); if (dX_46_v <= single(58.0)) tmp_3 = single(0.0); if (t_6 >= t_4) tmp_3 = t_5 / sqrt(max((t_0 + t_6), t_4)); else tmp_3 = (dY_46_u / sqrt(max(t_7, (t_3 + t_1)))) * floor(w); end tmp_2 = tmp_3; elseif (t_0 >= t_4) tmp_2 = t_5 / t_8; else tmp_2 = t_2 / t_8; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_1 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := {t\_2}^{2}\\
t_4 := t\_1 + t\_3\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := {t\_5}^{2}\\
t_7 := t\_6 + t\_0\\
t_8 := \sqrt{\mathsf{max}\left(t\_7, t\_4\right)}\\
\mathbf{if}\;dX.v \leq 58:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_4:\\
\;\;\;\;\frac{t\_5}{\sqrt{\mathsf{max}\left(t\_0 + t\_6, t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.u}{\sqrt{\mathsf{max}\left(t\_7, t\_3 + t\_1\right)}} \cdot \left\lfloor w\right\rfloor \\
\end{array}\\
\mathbf{elif}\;t\_0 \geq t\_4:\\
\;\;\;\;\frac{t\_5}{t\_8}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_8}\\
\end{array}
\end{array}
if dX.v < 58Initial program 76.2%
Applied rewrites76.5%
lift-+.f32N/A
lift-pow.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
pow2N/A
lift-*.f32N/A
lower-fma.f32N/A
lift-pow.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3276.6
lift-*.f32N/A
pow2N/A
lift-pow.f3276.6
Applied rewrites76.6%
Applied rewrites76.4%
Taylor expanded in dX.u around inf
Applied rewrites70.0%
if 58 < dX.v Initial program 70.4%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3268.8
Applied rewrites68.8%
Applied rewrites68.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor w) dX.u))
(t_2 (pow (* (floor h) dX.v) 2.0))
(t_3 (pow t_0 2.0))
(t_4 (+ (pow (* (floor h) dY.v) 2.0) t_3)))
(if (>= t_2 t_3)
(/ t_1 (sqrt (fmax (+ (pow t_1 2.0) t_2) t_4)))
(/
t_0
(sqrt
(fmax
(+ (exp (fma (log (floor w)) 2.0 (* (log dX.u) 2.0))) 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(w) * dY_46_u;
float t_1 = floorf(w) * dX_46_u;
float t_2 = powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = powf(t_0, 2.0f);
float t_4 = powf((floorf(h) * dY_46_v), 2.0f) + t_3;
float tmp;
if (t_2 >= t_3) {
tmp = t_1 / sqrtf(fmaxf((powf(t_1, 2.0f) + t_2), t_4));
} else {
tmp = t_0 / sqrtf(fmaxf((expf(fmaf(logf(floorf(w)), 2.0f, (logf(dX_46_u) * 2.0f))) + 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(w) * dY_46_u) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_3 = t_0 ^ Float32(2.0) t_4 = Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_3) tmp = Float32(0.0) if (t_2 >= t_3) tmp = Float32(t_1 / sqrt(fmax(Float32((t_1 ^ Float32(2.0)) + t_2), t_4))); else tmp = Float32(t_0 / sqrt(fmax(Float32(exp(fma(log(floor(w)), Float32(2.0), Float32(log(dX_46_u) * Float32(2.0)))) + t_2), t_4))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := {t\_0}^{2}\\
t_4 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + t\_3\\
\mathbf{if}\;t\_2 \geq t\_3:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left({t\_1}^{2} + t\_2, t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(e^{\mathsf{fma}\left(\log \left(\left\lfloor w\right\rfloor \right), 2, \log dX.u \cdot 2\right)} + t\_2, t\_4\right)}}\\
\end{array}
\end{array}
Initial program 74.8%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3264.2
Applied rewrites64.2%
Applied rewrites64.4%
Taylor expanded in dY.u around inf
unpow-prod-downN/A
pow2N/A
*-commutativeN/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-pow.f3261.7
Applied rewrites61.7%
lift-pow.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
pow-to-expN/A
pow-to-expN/A
prod-expN/A
lower-exp.f32N/A
lower-fma.f32N/A
lower-log.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lower-log.f3262.4
Applied rewrites62.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor w) dX.u))
(t_2 (pow (* (floor h) dX.v) 2.0))
(t_3 (pow t_0 2.0))
(t_4 (+ (pow (* (floor h) dY.v) 2.0) t_3)))
(if (>= t_2 t_3)
(/ t_1 (sqrt (fmax (+ (pow t_1 2.0) t_2) t_4)))
(/ t_0 (sqrt (fmax (+ (exp (* (log t_1) 2.0)) 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(w) * dY_46_u;
float t_1 = floorf(w) * dX_46_u;
float t_2 = powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = powf(t_0, 2.0f);
float t_4 = powf((floorf(h) * dY_46_v), 2.0f) + t_3;
float tmp;
if (t_2 >= t_3) {
tmp = t_1 / sqrtf(fmaxf((powf(t_1, 2.0f) + t_2), t_4));
} else {
tmp = t_0 / sqrtf(fmaxf((expf((logf(t_1) * 2.0f)) + 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(w) * dY_46_u) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_3 = t_0 ^ Float32(2.0) t_4 = Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_3) tmp = Float32(0.0) if (t_2 >= t_3) tmp = Float32(t_1 / sqrt(fmax(Float32((t_1 ^ Float32(2.0)) + t_2), t_4))); else tmp = Float32(t_0 / sqrt(fmax(Float32(exp(Float32(log(t_1) * Float32(2.0))) + 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(w) * dY_46_u; t_1 = floor(w) * dX_46_u; t_2 = (floor(h) * dX_46_v) ^ single(2.0); t_3 = t_0 ^ single(2.0); t_4 = ((floor(h) * dY_46_v) ^ single(2.0)) + t_3; tmp = single(0.0); if (t_2 >= t_3) tmp = t_1 / sqrt(max(((t_1 ^ single(2.0)) + t_2), t_4)); else tmp = t_0 / sqrt(max((exp((log(t_1) * single(2.0))) + t_2), t_4)); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := {t\_0}^{2}\\
t_4 := {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + t\_3\\
\mathbf{if}\;t\_2 \geq t\_3:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left({t\_1}^{2} + t\_2, t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(e^{\log t\_1 \cdot 2} + t\_2, t\_4\right)}}\\
\end{array}
\end{array}
Initial program 74.8%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3264.2
Applied rewrites64.2%
Applied rewrites64.4%
Taylor expanded in dY.u around inf
unpow-prod-downN/A
pow2N/A
*-commutativeN/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-pow.f3261.7
Applied rewrites61.7%
lift-pow.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3262.4
Applied rewrites62.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor w) dX.u))
(t_2 (pow t_0 2.0))
(t_3 (pow (* (floor h) dX.v) 2.0))
(t_4 (+ (pow t_1 2.0) t_3)))
(if (>= t_3 t_2)
(/ t_1 (sqrt (fmax t_4 (+ (pow (* (floor h) dY.v) 2.0) t_2))))
(/ t_0 (sqrt (fmax t_4 (fma (pow (floor h) 2.0) (* dY.v dY.v) 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(w) * dX_46_u;
float t_2 = powf(t_0, 2.0f);
float t_3 = powf((floorf(h) * dX_46_v), 2.0f);
float t_4 = powf(t_1, 2.0f) + t_3;
float tmp;
if (t_3 >= t_2) {
tmp = t_1 / sqrtf(fmaxf(t_4, (powf((floorf(h) * dY_46_v), 2.0f) + t_2)));
} else {
tmp = t_0 / sqrtf(fmaxf(t_4, fmaf(powf(floorf(h), 2.0f), (dY_46_v * dY_46_v), 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(w) * dX_46_u) t_2 = t_0 ^ Float32(2.0) t_3 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_4 = Float32((t_1 ^ Float32(2.0)) + t_3) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_1 / sqrt(fmax(t_4, Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_2)))); else tmp = Float32(t_0 / sqrt(fmax(t_4, fma((floor(h) ^ Float32(2.0)), Float32(dY_46_v * dY_46_v), 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 w\right\rfloor \cdot dX.u\\
t_2 := {t\_0}^{2}\\
t_3 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_4 := {t\_1}^{2} + t\_3\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_4, {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_4, \mathsf{fma}\left({\left(\left\lfloor h\right\rfloor \right)}^{2}, dY.v \cdot dY.v, t\_2\right)\right)}}\\
\end{array}
\end{array}
Initial program 74.8%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3264.2
Applied rewrites64.2%
Applied rewrites64.4%
Taylor expanded in dY.u around inf
unpow-prod-downN/A
pow2N/A
*-commutativeN/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-pow.f3261.7
Applied rewrites61.7%
lift-+.f32N/A
lift-pow.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
lower-fma.f32N/A
lower-pow.f32N/A
lift-floor.f32N/A
pow2N/A
lower-*.f3261.7
Applied rewrites61.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor w) dX.u))
(t_2 (pow (* (floor h) dX.v) 2.0))
(t_3 (pow t_0 2.0))
(t_4
(sqrt
(fmax (+ (pow t_1 2.0) t_2) (+ (pow (* (floor h) dY.v) 2.0) t_3)))))
(if (>= t_2 t_3) (/ 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(w) * dY_46_u;
float t_1 = floorf(w) * dX_46_u;
float t_2 = powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = powf(t_0, 2.0f);
float t_4 = sqrtf(fmaxf((powf(t_1, 2.0f) + t_2), (powf((floorf(h) * dY_46_v), 2.0f) + t_3)));
float tmp;
if (t_2 >= t_3) {
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(w) * dY_46_u) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_3 = t_0 ^ Float32(2.0) t_4 = sqrt(fmax(Float32((t_1 ^ Float32(2.0)) + t_2), Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_3))) tmp = Float32(0.0) if (t_2 >= t_3) 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(w) * dY_46_u; t_1 = floor(w) * dX_46_u; t_2 = (floor(h) * dX_46_v) ^ single(2.0); t_3 = t_0 ^ single(2.0); t_4 = sqrt(max(((t_1 ^ single(2.0)) + t_2), (((floor(h) * dY_46_v) ^ single(2.0)) + t_3))); tmp = single(0.0); if (t_2 >= t_3) 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 w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := {t\_0}^{2}\\
t_4 := \sqrt{\mathsf{max}\left({t\_1}^{2} + t\_2, {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + t\_3\right)}\\
\mathbf{if}\;t\_2 \geq t\_3:\\
\;\;\;\;\frac{t\_1}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\end{array}
\end{array}
Initial program 74.8%
Taylor expanded in dX.u around 0
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-pow.f3264.2
Applied rewrites64.2%
Applied rewrites64.4%
Taylor expanded in dY.u around inf
unpow-prod-downN/A
pow2N/A
*-commutativeN/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-pow.f3261.7
Applied rewrites61.7%
herbie shell --seed 2025083
(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))))