
(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 4 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) dY.u))
(t_1 (* dY.u (floor w)))
(t_2 (pow t_1 2.0))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dX.v))
(t_5 (+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0)))
(t_6 (* (floor h) dY.v))
(t_7 (pow (* dY.v (floor h)) 2.0)))
(if (>= (+ (* t_3 t_3) (* t_4 t_4)) (+ (* t_0 t_0) (* t_6 t_6)))
(* (/ 1.0 (sqrt (fmax t_5 (+ t_2 t_7)))) t_3)
(/ (* t_1 (- -1.0)) (sqrt (fmax t_5 (+ t_7 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 = dY_46_u * floorf(w);
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 = powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f);
float t_6 = floorf(h) * dY_46_v;
float t_7 = powf((dY_46_v * floorf(h)), 2.0f);
float tmp;
if (((t_3 * t_3) + (t_4 * t_4)) >= ((t_0 * t_0) + (t_6 * t_6))) {
tmp = (1.0f / sqrtf(fmaxf(t_5, (t_2 + t_7)))) * t_3;
} else {
tmp = (t_1 * -(-1.0f)) / sqrtf(fmaxf(t_5, (t_7 + 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(dY_46_u * floor(w)) 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(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))) t_6 = Float32(floor(h) * dY_46_v) t_7 = Float32(dY_46_v * floor(h)) ^ Float32(2.0) tmp = Float32(0.0) if (Float32(Float32(t_3 * t_3) + Float32(t_4 * t_4)) >= Float32(Float32(t_0 * t_0) + Float32(t_6 * t_6))) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_5, Float32(t_2 + t_7)))) * t_3); else tmp = Float32(Float32(t_1 * Float32(-Float32(-1.0))) / sqrt(fmax(t_5, Float32(t_7 + 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) * dY_46_u; t_1 = dY_46_u * floor(w); t_2 = t_1 ^ single(2.0); t_3 = floor(w) * dX_46_u; t_4 = floor(h) * dX_46_v; t_5 = ((dX_46_v * floor(h)) ^ single(2.0)) + ((dX_46_u * floor(w)) ^ single(2.0)); t_6 = floor(h) * dY_46_v; t_7 = (dY_46_v * floor(h)) ^ single(2.0); tmp = single(0.0); if (((t_3 * t_3) + (t_4 * t_4)) >= ((t_0 * t_0) + (t_6 * t_6))) tmp = (single(1.0) / sqrt(max(t_5, (t_2 + t_7)))) * t_3; else tmp = (t_1 * -single(-1.0)) / sqrt(max(t_5, (t_7 + t_2))); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
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 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_7 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
\mathbf{if}\;t\_3 \cdot t\_3 + t\_4 \cdot t\_4 \geq t\_0 \cdot t\_0 + t\_6 \cdot t\_6:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 + t\_7\right)}} \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1 \cdot \left(--1\right)}{\sqrt{\mathsf{max}\left(t\_5, t\_7 + t\_2\right)}}\\
\end{array}
\end{array}
Initial program 78.1%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
Applied rewrites78.3%
Applied rewrites78.3%
Final simplification78.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (* dY.u (floor w)) 2.0))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor h) dY.v))
(t_3 (pow (* dX.v (floor h)) 2.0))
(t_4 (* (floor w) dX.u))
(t_5 (pow (* dY.v (floor h)) 2.0))
(t_6 (* (floor w) dY.u)))
(if (>= t_3 (+ t_0 t_5))
(*
(/
1.0
(sqrt (fmax (+ (* t_4 t_4) (* t_1 t_1)) (+ (* t_6 t_6) (* t_2 t_2)))))
t_4)
(*
(/ 1.0 (sqrt (fmax (- t_3 (pow (* dX.u (floor w)) 2.0)) (+ t_5 t_0))))
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((dY_46_u * floorf(w)), 2.0f);
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(h) * dY_46_v;
float t_3 = powf((dX_46_v * floorf(h)), 2.0f);
float t_4 = floorf(w) * dX_46_u;
float t_5 = powf((dY_46_v * floorf(h)), 2.0f);
float t_6 = floorf(w) * dY_46_u;
float tmp;
if (t_3 >= (t_0 + t_5)) {
tmp = (1.0f / sqrtf(fmaxf(((t_4 * t_4) + (t_1 * t_1)), ((t_6 * t_6) + (t_2 * t_2))))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_3 - powf((dX_46_u * floorf(w)), 2.0f)), (t_5 + t_0)))) * t_6;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_u * floor(w)) ^ Float32(2.0) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(dX_46_v * floor(h)) ^ Float32(2.0) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(dY_46_v * floor(h)) ^ Float32(2.0) t_6 = Float32(floor(w) * dY_46_u) tmp = Float32(0.0) if (t_3 >= Float32(t_0 + t_5)) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_4 * t_4) + Float32(t_1 * t_1)), Float32(Float32(t_6 * t_6) + Float32(t_2 * t_2))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_3 - (Float32(dX_46_u * floor(w)) ^ Float32(2.0))), Float32(t_5 + t_0)))) * 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 = (dY_46_u * floor(w)) ^ single(2.0); t_1 = floor(h) * dX_46_v; t_2 = floor(h) * dY_46_v; t_3 = (dX_46_v * floor(h)) ^ single(2.0); t_4 = floor(w) * dX_46_u; t_5 = (dY_46_v * floor(h)) ^ single(2.0); t_6 = floor(w) * dY_46_u; tmp = single(0.0); if (t_3 >= (t_0 + t_5)) tmp = (single(1.0) / sqrt(max(((t_4 * t_4) + (t_1 * t_1)), ((t_6 * t_6) + (t_2 * t_2))))) * t_4; else tmp = (single(1.0) / sqrt(max((t_3 - ((dX_46_u * floor(w)) ^ single(2.0))), (t_5 + t_0)))) * t_6; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_6 := \left\lfloor w\right\rfloor \cdot dY.u\\
\mathbf{if}\;t\_3 \geq t\_0 + t\_5:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_1 \cdot t\_1, t\_6 \cdot t\_6 + t\_2 \cdot t\_2\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3 - {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}, t\_5 + t\_0\right)}} \cdot t\_6\\
\end{array}
\end{array}
Initial program 78.1%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.2
Applied rewrites69.2%
lift-*.f32N/A
pow2N/A
lower-pow.f3269.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3269.2
Applied rewrites69.2%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites69.2%
Applied rewrites72.8%
(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.v (floor h)) 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(*
(/
1.0
(sqrt (fmax (+ t_4 (* (floor h) (* (floor h) (* 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_v * floorf(h)), 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((t_4 + (floorf(h) * (floorf(h) * (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_v * floor(h)) ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_4 + Float32(floor(h) * Float32(floor(h) * 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_v * floor(h)) ^ single(2.0)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((t_4 + (floor(h) * (floor(h) * (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.v \cdot \left\lfloor h\right\rfloor \right)}^{2} \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 + \left\lfloor h\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot \left(dX.v \cdot dX.v\right)\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 78.1%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.2
Applied rewrites69.2%
lift-*.f32N/A
pow2N/A
lower-pow.f3269.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3269.2
Applied rewrites69.2%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites69.2%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
unpow2N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3269.2
Applied rewrites69.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (pow (floor w) 2.0))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor w) dX.u))
(t_5 (pow (floor h) 2.0)))
(if (>=
(pow (* dX.v (floor h)) 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(*
(/
1.0
(sqrt
(fmax
(fma (* t_5 dX.v) dX.v (* (* t_1 dX.u) dX.u))
(fma (* t_1 dY.u) dY.u (* (* t_5 dY.v) dY.v)))))
t_4)
(*
(/
1.0
(sqrt (fmax (+ (* t_4 t_4) (* t_0 t_0)) (+ (* t_2 t_2) (* t_3 t_3)))))
t_2))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = powf(floorf(w), 2.0f);
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(w) * dX_46_u;
float t_5 = powf(floorf(h), 2.0f);
float tmp;
if (powf((dX_46_v * floorf(h)), 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(fmaf((t_5 * dX_46_v), dX_46_v, ((t_1 * dX_46_u) * dX_46_u)), fmaf((t_1 * dY_46_u), dY_46_u, ((t_5 * dY_46_v) * dY_46_v))))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf(((t_4 * t_4) + (t_0 * t_0)), ((t_2 * t_2) + (t_3 * t_3))))) * t_2;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = floor(w) ^ Float32(2.0) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(w) * dX_46_u) t_5 = floor(h) ^ Float32(2.0) tmp = Float32(0.0) if ((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(fma(Float32(t_5 * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u)), fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(t_5 * dY_46_v) * dY_46_v))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)), Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3))))) * t_2); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
\mathbf{if}\;{\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5 \cdot dX.v, dX.v, \left(t\_1 \cdot dX.u\right) \cdot dX.u\right), \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(t\_5 \cdot dY.v\right) \cdot dY.v\right)\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_0 \cdot t\_0, t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right)}} \cdot t\_2\\
\end{array}
\end{array}
Initial program 78.1%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.2
Applied rewrites69.2%
lift-*.f32N/A
pow2N/A
lower-pow.f3269.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3269.2
Applied rewrites69.2%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites69.2%
Taylor expanded in w around 0
Applied rewrites46.1%
herbie shell --seed 2024343
(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))))