
(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(((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_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 13 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(((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_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 (* dY.v (floor h)))
(t_2 (* dX.v (floor h)))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* dY.u (floor w)))
(t_5 (+ (* t_1 t_1) (* t_4 t_4))))
(if (>= t_3 t_5)
(* (/ 1.0 (sqrt (fmax t_3 t_5))) t_0)
(/
(/ (* (- dY.u) (floor w)) -1.0)
(sqrt
(fmax
(+ (pow t_0 2.0) (pow t_2 2.0))
(+ (pow t_4 2.0) (pow t_1 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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dX_46_v * floorf(h);
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = dY_46_u * floorf(w);
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float tmp;
if (t_3 >= t_5) {
tmp = (1.0f / sqrtf(fmaxf(t_3, t_5))) * t_0;
} else {
tmp = ((-dY_46_u * floorf(w)) / -1.0f) / sqrtf(fmaxf((powf(t_0, 2.0f) + powf(t_2, 2.0f)), (powf(t_4, 2.0f) + powf(t_1, 2.0f))));
}
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(dY_46_v * floor(h)) t_2 = Float32(dX_46_v * floor(h)) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(dY_46_u * floor(w)) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) * t_0); else tmp = Float32(Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-1.0)) / sqrt(((Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) != Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) ? Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) : ((Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) != Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) ? Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) : max(Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))), Float32((t_4 ^ Float32(2.0)) + (t_1 ^ 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 = dX_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dX_46_v * floor(h); t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = dY_46_u * floor(w); t_5 = (t_1 * t_1) + (t_4 * t_4); tmp = single(0.0); if (t_3 >= t_5) tmp = (single(1.0) / sqrt(max(t_3, t_5))) * t_0; else tmp = ((-dY_46_u * floor(w)) / single(-1.0)) / sqrt(max(((t_0 ^ single(2.0)) + (t_2 ^ single(2.0))), ((t_4 ^ single(2.0)) + (t_1 ^ single(2.0))))); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{\frac{\left(-dY.u\right) \cdot \left\lfloor w\right\rfloor }{-1}}{\sqrt{\mathsf{max}\left({t\_0}^{2} + {t\_2}^{2}, {t\_4}^{2} + {t\_1}^{2}\right)}}\\
\end{array}
\end{array}
Initial program 73.3%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites73.5%
Final simplification73.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.v (floor h)))
(t_2 (* dX.v (floor h)))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* dY.u (floor w)))
(t_5 (+ (* t_1 t_1) (* t_4 t_4))))
(if (>= t_3 t_5)
(* (/ 1.0 (sqrt (fmax t_3 t_5))) t_0)
(/
1.0
(/
(sqrt
(fmax (+ (pow t_0 2.0) (pow t_2 2.0)) (+ (pow t_4 2.0) (pow t_1 2.0))))
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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dX_46_v * floorf(h);
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = dY_46_u * floorf(w);
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float tmp;
if (t_3 >= t_5) {
tmp = (1.0f / sqrtf(fmaxf(t_3, t_5))) * t_0;
} else {
tmp = 1.0f / (sqrtf(fmaxf((powf(t_0, 2.0f) + powf(t_2, 2.0f)), (powf(t_4, 2.0f) + powf(t_1, 2.0f)))) / t_4);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_u * floor(w)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dX_46_v * floor(h)) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(dY_46_u * floor(w)) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) * t_0); else tmp = Float32(Float32(1.0) / Float32(sqrt(((Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) != Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) ? Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) : ((Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) != Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) ? Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) : max(Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))), Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))))))) / 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 = dX_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dX_46_v * floor(h); t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = dY_46_u * floor(w); t_5 = (t_1 * t_1) + (t_4 * t_4); tmp = single(0.0); if (t_3 >= t_5) tmp = (single(1.0) / sqrt(max(t_3, t_5))) * t_0; else tmp = single(1.0) / (sqrt(max(((t_0 ^ single(2.0)) + (t_2 ^ single(2.0))), ((t_4 ^ single(2.0)) + (t_1 ^ single(2.0))))) / t_4); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left({t\_0}^{2} + {t\_2}^{2}, {t\_4}^{2} + {t\_1}^{2}\right)}}{t\_4}}\\
\end{array}
\end{array}
Initial program 73.3%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites73.5%
Final simplification73.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.v (floor h)))
(t_2 (* dX.v (floor h)))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* dY.u (floor w)))
(t_5 (+ (* t_1 t_1) (* t_4 t_4))))
(if (>= t_3 t_5)
(* (/ 1.0 (sqrt (fmax t_3 t_5))) t_0)
(*
(/
(floor w)
(sqrt
(fmax
(+ (pow t_0 2.0) (pow t_2 2.0))
(+ (pow t_4 2.0) (pow t_1 2.0)))))
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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dX_46_v * floorf(h);
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = dY_46_u * floorf(w);
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float tmp;
if (t_3 >= t_5) {
tmp = (1.0f / sqrtf(fmaxf(t_3, t_5))) * t_0;
} else {
tmp = (floorf(w) / sqrtf(fmaxf((powf(t_0, 2.0f) + powf(t_2, 2.0f)), (powf(t_4, 2.0f) + powf(t_1, 2.0f))))) * 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(dX_46_u * floor(w)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dX_46_v * floor(h)) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(dY_46_u * floor(w)) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) * t_0); else tmp = Float32(Float32(floor(w) / sqrt(((Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) != Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) ? Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) : ((Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) != Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) ? Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) : max(Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))), Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))))))) * 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 = dX_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dX_46_v * floor(h); t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = dY_46_u * floor(w); t_5 = (t_1 * t_1) + (t_4 * t_4); tmp = single(0.0); if (t_3 >= t_5) tmp = (single(1.0) / sqrt(max(t_3, t_5))) * t_0; else tmp = (floor(w) / sqrt(max(((t_0 ^ single(2.0)) + (t_2 ^ single(2.0))), ((t_4 ^ single(2.0)) + (t_1 ^ single(2.0)))))) * dY_46_u; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left({t\_0}^{2} + {t\_2}^{2}, {t\_4}^{2} + {t\_1}^{2}\right)}} \cdot dY.u\\
\end{array}
\end{array}
Initial program 73.3%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
*-rgt-identityN/A
Applied rewrites73.4%
Final simplification73.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.v (floor h)))
(t_2 (* dX.v (floor h)))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* dY.u (floor w)))
(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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dX_46_v * floorf(h);
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = dY_46_u * floorf(w);
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(dX_46_u * floor(w)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dX_46_v * floor(h)) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(dY_46_u * floor(w)) 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 = dX_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dX_46_v * floor(h); t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = dY_46_u * floor(w); 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 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
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}
Initial program 73.3%
Final simplification73.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* dX.v (floor h)))
(t_2 (* t_1 t_1))
(t_3 (* dY.u (floor w)))
(t_4 (+ (* t_0 t_0) (* t_3 t_3)))
(t_5 (+ (pow t_3 2.0) (pow t_0 2.0)))
(t_6 (* dX.u (floor w)))
(t_7 (/ 1.0 (sqrt (fmax (+ t_2 (* t_6 t_6)) t_4))))
(t_8 (* t_7 t_6)))
(if (<= dX.v 100.0)
(if (>= (* (* (pow (floor w) 2.0) dX.u) dX.u) t_5) t_8 (* t_7 t_3))
(if (>= (pow t_1 2.0) t_5)
t_8
(* (/ 1.0 (sqrt (fmax (fma t_6 t_6 t_2) t_4))) 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 = dY_46_v * floorf(h);
float t_1 = dX_46_v * floorf(h);
float t_2 = t_1 * t_1;
float t_3 = dY_46_u * floorf(w);
float t_4 = (t_0 * t_0) + (t_3 * t_3);
float t_5 = powf(t_3, 2.0f) + powf(t_0, 2.0f);
float t_6 = dX_46_u * floorf(w);
float t_7 = 1.0f / sqrtf(fmaxf((t_2 + (t_6 * t_6)), t_4));
float t_8 = t_7 * t_6;
float tmp_1;
if (dX_46_v <= 100.0f) {
float tmp_2;
if (((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) >= t_5) {
tmp_2 = t_8;
} else {
tmp_2 = t_7 * t_3;
}
tmp_1 = tmp_2;
} else if (powf(t_1, 2.0f) >= t_5) {
tmp_1 = t_8;
} else {
tmp_1 = (1.0f / sqrtf(fmaxf(fmaf(t_6, t_6, t_2), t_4))) * t_3;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_v * floor(h)) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(t_1 * t_1) t_3 = Float32(dY_46_u * floor(w)) t_4 = Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) t_5 = Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_6 = Float32(dX_46_u * floor(w)) t_7 = Float32(Float32(1.0) / sqrt(((Float32(t_2 + Float32(t_6 * t_6)) != Float32(t_2 + Float32(t_6 * t_6))) ? t_4 : ((t_4 != t_4) ? Float32(t_2 + Float32(t_6 * t_6)) : max(Float32(t_2 + Float32(t_6 * t_6)), t_4))))) t_8 = Float32(t_7 * t_6) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(100.0)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) >= t_5) tmp_2 = t_8; else tmp_2 = Float32(t_7 * t_3); end tmp_1 = tmp_2; elseif ((t_1 ^ Float32(2.0)) >= t_5) tmp_1 = t_8; else tmp_1 = Float32(Float32(Float32(1.0) / sqrt(((fma(t_6, t_6, t_2) != fma(t_6, t_6, t_2)) ? t_4 : ((t_4 != t_4) ? fma(t_6, t_6, t_2) : max(fma(t_6, t_6, t_2), t_4))))) * t_3); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := t\_1 \cdot t\_1\\
t_3 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_4 := t\_0 \cdot t\_0 + t\_3 \cdot t\_3\\
t_5 := {t\_3}^{2} + {t\_0}^{2}\\
t_6 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_2 + t\_6 \cdot t\_6, t\_4\right)}}\\
t_8 := t\_7 \cdot t\_6\\
\mathbf{if}\;dX.v \leq 100:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u \geq t\_5:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_3\\
\end{array}\\
\mathbf{elif}\;{t\_1}^{2} \geq t\_5:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, t\_6, t\_2\right), t\_4\right)}} \cdot t\_3\\
\end{array}
\end{array}
if dX.v < 100Initial program 73.2%
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.f3261.6
Applied rewrites61.6%
Applied rewrites61.6%
Taylor expanded in dX.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.4
Applied rewrites68.4%
if 100 < dX.v Initial program 73.7%
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.f3270.6
Applied rewrites70.6%
Applied rewrites70.6%
lift-+.f32N/A
lift-*.f32N/A
lower-fma.f3271.0
Applied rewrites71.0%
Final simplification69.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.u (floor w)))
(t_2 (* dY.v (floor h)))
(t_3 (+ (* t_2 t_2) (* t_1 t_1)))
(t_4 (* dX.v (floor h)))
(t_5 (* t_4 t_4)))
(if (>= (pow t_4 2.0) (+ (pow t_1 2.0) (pow t_2 2.0)))
(* (/ 1.0 (sqrt (fmax (+ t_5 (* t_0 t_0)) t_3))) t_0)
(* (/ 1.0 (sqrt (fmax (fma t_0 t_0 t_5) t_3))) 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 = dX_46_u * floorf(w);
float t_1 = dY_46_u * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = (t_2 * t_2) + (t_1 * t_1);
float t_4 = dX_46_v * floorf(h);
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 = (1.0f / sqrtf(fmaxf((t_5 + (t_0 * t_0)), t_3))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf(fmaf(t_0, t_0, t_5), t_3))) * t_1;
}
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(dY_46_u * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) t_4 = Float32(dX_46_v * floor(h)) 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(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(t_0 * t_0)) != Float32(t_5 + Float32(t_0 * t_0))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(t_0 * t_0)) : max(Float32(t_5 + Float32(t_0 * t_0)), t_3))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((fma(t_0, t_0, t_5) != fma(t_0, t_0, t_5)) ? t_3 : ((t_3 != t_3) ? fma(t_0, t_0, t_5) : max(fma(t_0, t_0, t_5), t_3))))) * t_1); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot t\_2 + t\_1 \cdot t\_1\\
t_4 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_4}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_0 \cdot t\_0, t\_3\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_5\right), t\_3\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 73.3%
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.f3263.7
Applied rewrites63.7%
Applied rewrites63.7%
lift-+.f32N/A
lift-*.f32N/A
lower-fma.f3267.3
Applied rewrites67.3%
Final simplification67.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.u (floor w)))
(t_2 (* dY.v (floor h)))
(t_3 (+ (* t_2 t_2) (* t_1 t_1)))
(t_4 (* dX.v (floor h)))
(t_5 (* t_4 t_4)))
(if (>= (pow t_4 2.0) (+ (pow t_1 2.0) (pow t_2 2.0)))
(* (/ 1.0 (sqrt (fmax (+ t_5 (* t_0 t_0)) t_3))) t_0)
(*
(/ 1.0 (sqrt (fmax (+ (* (* (pow (floor w) 2.0) dX.u) dX.u) t_5) t_3)))
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 = dX_46_u * floorf(w);
float t_1 = dY_46_u * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = (t_2 * t_2) + (t_1 * t_1);
float t_4 = dX_46_v * floorf(h);
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 = (1.0f / sqrtf(fmaxf((t_5 + (t_0 * t_0)), t_3))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf((((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) + t_5), t_3))) * t_1;
}
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(dY_46_u * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) t_4 = Float32(dX_46_v * floor(h)) 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(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(t_0 * t_0)) != Float32(t_5 + Float32(t_0 * t_0))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(t_0 * t_0)) : max(Float32(t_5 + Float32(t_0 * t_0)), t_3))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) + t_5) != Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) + t_5)) ? t_3 : ((t_3 != t_3) ? Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) + t_5) : max(Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) + t_5), t_3))))) * 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 = dX_46_u * floor(w); t_1 = dY_46_u * floor(w); t_2 = dY_46_v * floor(h); t_3 = (t_2 * t_2) + (t_1 * t_1); t_4 = dX_46_v * floor(h); 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 = (single(1.0) / sqrt(max((t_5 + (t_0 * t_0)), t_3))) * t_0; else tmp = (single(1.0) / sqrt(max(((((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u) + t_5), t_3))) * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot t\_2 + t\_1 \cdot t\_1\\
t_4 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_4}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_0 \cdot t\_0, t\_3\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u + t\_5, t\_3\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 73.3%
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.f3263.7
Applied rewrites63.7%
Applied rewrites63.7%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3263.7
Applied rewrites63.7%
Final simplification63.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* dX.v (floor h)))
(t_2 (pow t_1 2.0))
(t_3 (* dY.u (floor w)))
(t_4 (* dX.u (floor w))))
(if (>= t_2 (+ (pow t_3 2.0) (pow t_0 2.0)))
(*
(/
1.0
(sqrt
(fmax
(/ 1.0 (/ 1.0 (+ (pow t_4 2.0) t_2)))
(fma
(* (pow (floor w) 2.0) dY.u)
dY.u
(* (* (pow (floor h) 2.0) dY.v) dY.v)))))
t_4)
(*
(/
1.0
(sqrt (fmax (+ (* t_1 t_1) (* t_4 t_4)) (+ (* t_0 t_0) (* t_3 t_3)))))
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 = dY_46_v * floorf(h);
float t_1 = dX_46_v * floorf(h);
float t_2 = powf(t_1, 2.0f);
float t_3 = dY_46_u * floorf(w);
float t_4 = dX_46_u * floorf(w);
float tmp;
if (t_2 >= (powf(t_3, 2.0f) + powf(t_0, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((1.0f / (1.0f / (powf(t_4, 2.0f) + t_2))), fmaf((powf(floorf(w), 2.0f) * dY_46_u), dY_46_u, ((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v))))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf(((t_1 * t_1) + (t_4 * t_4)), ((t_0 * t_0) + (t_3 * t_3))))) * t_3;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_v * floor(h)) t_1 = Float32(dX_46_v * floor(h)) t_2 = t_1 ^ Float32(2.0) t_3 = Float32(dY_46_u * floor(w)) t_4 = Float32(dX_46_u * floor(w)) tmp = Float32(0.0) if (t_2 >= Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(1.0) / Float32(Float32(1.0) / Float32((t_4 ^ Float32(2.0)) + t_2))) != Float32(Float32(1.0) / Float32(Float32(1.0) / Float32((t_4 ^ Float32(2.0)) + t_2)))) ? fma(Float32((floor(w) ^ Float32(2.0)) * dY_46_u), dY_46_u, Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v)) : ((fma(Float32((floor(w) ^ Float32(2.0)) * dY_46_u), dY_46_u, Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v)) != fma(Float32((floor(w) ^ Float32(2.0)) * dY_46_u), dY_46_u, Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v))) ? Float32(Float32(1.0) / Float32(Float32(1.0) / Float32((t_4 ^ Float32(2.0)) + t_2))) : max(Float32(Float32(1.0) / Float32(Float32(1.0) / Float32((t_4 ^ Float32(2.0)) + t_2))), fma(Float32((floor(w) ^ Float32(2.0)) * dY_46_u), dY_46_u, Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v))))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) != Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4))) ? Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) : ((Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) != Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3))) ? Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) : max(Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)), Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3))))))) * t_3); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := {t\_1}^{2}\\
t_3 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_4 := dX.u \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;t\_2 \geq {t\_3}^{2} + {t\_0}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\frac{1}{\frac{1}{{t\_4}^{2} + t\_2}}, \mathsf{fma}\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dY.u, dY.u, \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v\right)\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_1 \cdot t\_1 + t\_4 \cdot t\_4, t\_0 \cdot t\_0 + t\_3 \cdot t\_3\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 73.3%
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.f3263.7
Applied rewrites63.7%
Applied rewrites63.7%
Taylor expanded in dY.v around 0
lower-sqrt.f32N/A
lower-fmax.f32N/A
Applied rewrites43.4%
Applied rewrites61.8%
Final simplification61.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (pow (floor h) 2.0))
(t_2 (* dY.v (floor h)))
(t_3 (* dX.u (floor w)))
(t_4 (* dX.v (floor h)))
(t_5 (pow (floor w) 2.0)))
(if (>= (pow t_4 2.0) (+ (pow t_0 2.0) (pow t_2 2.0)))
(*
(/
1.0
(sqrt
(fmax
(fma (* t_5 dX.u) dX.u (* (* t_1 dX.v) dX.v))
(fma (* t_5 dY.u) dY.u (* (* t_1 dY.v) dY.v)))))
t_3)
(*
(/
1.0
(sqrt (fmax (+ (* t_4 t_4) (* t_3 t_3)) (+ (* t_2 t_2) (* t_0 t_0)))))
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 = dY_46_u * floorf(w);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = dY_46_v * floorf(h);
float t_3 = dX_46_u * floorf(w);
float t_4 = dX_46_v * floorf(h);
float t_5 = powf(floorf(w), 2.0f);
float tmp;
if (powf(t_4, 2.0f) >= (powf(t_0, 2.0f) + powf(t_2, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(fmaf((t_5 * dX_46_u), dX_46_u, ((t_1 * dX_46_v) * dX_46_v)), fmaf((t_5 * dY_46_u), dY_46_u, ((t_1 * dY_46_v) * dY_46_v))))) * t_3;
} else {
tmp = (1.0f / sqrtf(fmaxf(((t_4 * t_4) + (t_3 * t_3)), ((t_2 * t_2) + (t_0 * t_0))))) * t_0;
}
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)) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(dX_46_u * floor(w)) t_4 = Float32(dX_46_v * floor(h)) t_5 = floor(w) ^ Float32(2.0) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) != fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v))) ? fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) : ((fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) != fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))) ? fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) : max(fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)), fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))))))) * t_3); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_4 * t_4) + Float32(t_3 * t_3)) != Float32(Float32(t_4 * t_4) + Float32(t_3 * t_3))) ? Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) : ((Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) != Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0))) ? Float32(Float32(t_4 * t_4) + Float32(t_3 * t_3)) : max(Float32(Float32(t_4 * t_4) + Float32(t_3 * t_3)), Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0))))))) * t_0); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_4 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_5 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
\mathbf{if}\;{t\_4}^{2} \geq {t\_0}^{2} + {t\_2}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5 \cdot dX.u, dX.u, \left(t\_1 \cdot dX.v\right) \cdot dX.v\right), \mathsf{fma}\left(t\_5 \cdot dY.u, dY.u, \left(t\_1 \cdot dY.v\right) \cdot dY.v\right)\right)}} \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_3 \cdot t\_3, t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 73.3%
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.f3263.7
Applied rewrites63.7%
Applied rewrites63.7%
Taylor expanded in dY.v around 0
lower-sqrt.f32N/A
lower-fmax.f32N/A
Applied rewrites43.8%
Taylor expanded in dX.u around 0
Applied rewrites43.5%
Final simplification43.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (pow (floor h) 2.0))
(t_2 (* dY.v (floor h)))
(t_3 (* dX.v (floor h)))
(t_4 (pow (floor w) 2.0)))
(if (>= (pow t_3 2.0) (+ (pow t_0 2.0) (pow t_2 2.0)))
(*
(/
1.0
(sqrt
(fmax
(fma (* t_4 dX.u) dX.u (* (* t_1 dX.v) dX.v))
(fma (* t_4 dY.u) dY.u (* (* t_1 dY.v) dY.v)))))
(* dX.u (floor w)))
(*
(/
1.0
(sqrt
(fmax
(+ (* (* dX.u dX.u) t_4) (* t_3 t_3))
(+ (* t_2 t_2) (* t_0 t_0)))))
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 = dY_46_u * floorf(w);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = dY_46_v * floorf(h);
float t_3 = dX_46_v * floorf(h);
float t_4 = powf(floorf(w), 2.0f);
float tmp;
if (powf(t_3, 2.0f) >= (powf(t_0, 2.0f) + powf(t_2, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(fmaf((t_4 * dX_46_u), dX_46_u, ((t_1 * dX_46_v) * dX_46_v)), fmaf((t_4 * dY_46_u), dY_46_u, ((t_1 * dY_46_v) * dY_46_v))))) * (dX_46_u * floorf(w));
} else {
tmp = (1.0f / sqrtf(fmaxf((((dX_46_u * dX_46_u) * t_4) + (t_3 * t_3)), ((t_2 * t_2) + (t_0 * t_0))))) * t_0;
}
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)) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(dX_46_v * floor(h)) t_4 = floor(w) ^ Float32(2.0) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((fma(Float32(t_4 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) != fma(Float32(t_4 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v))) ? fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) : ((fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) != fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))) ? fma(Float32(t_4 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) : max(fma(Float32(t_4 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)), fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))))))) * Float32(dX_46_u * floor(w))); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32(dX_46_u * dX_46_u) * t_4) + Float32(t_3 * t_3)) != Float32(Float32(Float32(dX_46_u * dX_46_u) * t_4) + Float32(t_3 * t_3))) ? Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) : ((Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) != Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0))) ? Float32(Float32(Float32(dX_46_u * dX_46_u) * t_4) + Float32(t_3 * t_3)) : max(Float32(Float32(Float32(dX_46_u * dX_46_u) * t_4) + Float32(t_3 * t_3)), Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0))))))) * t_0); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_0}^{2} + {t\_2}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot dX.u, dX.u, \left(t\_1 \cdot dX.v\right) \cdot dX.v\right), \mathsf{fma}\left(t\_4 \cdot dY.u, dY.u, \left(t\_1 \cdot dY.v\right) \cdot dY.v\right)\right)}} \cdot \left(dX.u \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left(dX.u \cdot dX.u\right) \cdot t\_4 + t\_3 \cdot t\_3, t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 73.3%
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.f3263.7
Applied rewrites63.7%
Applied rewrites63.7%
Taylor expanded in dY.v around 0
lower-sqrt.f32N/A
lower-fmax.f32N/A
Applied rewrites43.4%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3243.5
Applied rewrites43.5%
Final simplification43.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (pow (floor h) 2.0))
(t_2 (* dY.v (floor h)))
(t_3 (* dX.v (floor h)))
(t_4 (pow (floor w) 2.0))
(t_5 (* t_4 dX.u)))
(if (>= (pow t_3 2.0) (+ (pow t_0 2.0) (pow t_2 2.0)))
(*
(/
1.0
(sqrt
(fmax
(fma t_5 dX.u (* (* t_1 dX.v) dX.v))
(fma (* t_4 dY.u) dY.u (* (* t_1 dY.v) dY.v)))))
(* dX.u (floor w)))
(*
(/
1.0
(sqrt (fmax (+ (* t_5 dX.u) (* t_3 t_3)) (+ (* t_2 t_2) (* t_0 t_0)))))
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 = dY_46_u * floorf(w);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = dY_46_v * floorf(h);
float t_3 = dX_46_v * floorf(h);
float t_4 = powf(floorf(w), 2.0f);
float t_5 = t_4 * dX_46_u;
float tmp;
if (powf(t_3, 2.0f) >= (powf(t_0, 2.0f) + powf(t_2, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(fmaf(t_5, dX_46_u, ((t_1 * dX_46_v) * dX_46_v)), fmaf((t_4 * dY_46_u), dY_46_u, ((t_1 * dY_46_v) * dY_46_v))))) * (dX_46_u * floorf(w));
} else {
tmp = (1.0f / sqrtf(fmaxf(((t_5 * dX_46_u) + (t_3 * t_3)), ((t_2 * t_2) + (t_0 * t_0))))) * t_0;
}
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)) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(dX_46_v * floor(h)) t_4 = floor(w) ^ Float32(2.0) t_5 = Float32(t_4 * dX_46_u) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((fma(t_5, dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) != fma(t_5, dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v))) ? fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) : ((fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) != fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))) ? fma(t_5, dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) : max(fma(t_5, dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)), fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))))))) * Float32(dX_46_u * floor(w))); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_5 * dX_46_u) + Float32(t_3 * t_3)) != Float32(Float32(t_5 * dX_46_u) + Float32(t_3 * t_3))) ? Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) : ((Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) != Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0))) ? Float32(Float32(t_5 * dX_46_u) + Float32(t_3 * t_3)) : max(Float32(Float32(t_5 * dX_46_u) + Float32(t_3 * t_3)), Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0))))))) * t_0); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_5 := t\_4 \cdot dX.u\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_0}^{2} + {t\_2}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5, dX.u, \left(t\_1 \cdot dX.v\right) \cdot dX.v\right), \mathsf{fma}\left(t\_4 \cdot dY.u, dY.u, \left(t\_1 \cdot dY.v\right) \cdot dY.v\right)\right)}} \cdot \left(dX.u \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 \cdot dX.u + t\_3 \cdot t\_3, t\_2 \cdot t\_2 + t\_0 \cdot t\_0\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 73.3%
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.f3263.7
Applied rewrites63.7%
Applied rewrites63.7%
Taylor expanded in dY.v around 0
lower-sqrt.f32N/A
lower-fmax.f32N/A
Applied rewrites43.5%
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lift-*.f32N/A
lower-*.f3244.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3243.5
Applied rewrites43.5%
Final simplification43.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (pow (floor h) 2.0))
(t_2 (* dX.u (floor w)))
(t_3 (pow (* dX.v (floor h)) 2.0))
(t_4 (+ (pow t_0 2.0) (pow (* dY.v (floor h)) 2.0)))
(t_5 (pow (floor w) 2.0)))
(if (>= t_3 t_4)
(*
(/
1.0
(sqrt
(fmax
(fma (* t_5 dX.u) dX.u (* (* t_1 dX.v) dX.v))
(fma (* t_5 dY.u) dY.u (* (* t_1 dY.v) dY.v)))))
t_2)
(* (/ 1.0 (sqrt (fmax (+ (pow t_2 2.0) t_3) t_4))) 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 = dY_46_u * floorf(w);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = dX_46_u * floorf(w);
float t_3 = powf((dX_46_v * floorf(h)), 2.0f);
float t_4 = powf(t_0, 2.0f) + powf((dY_46_v * floorf(h)), 2.0f);
float t_5 = powf(floorf(w), 2.0f);
float tmp;
if (t_3 >= t_4) {
tmp = (1.0f / sqrtf(fmaxf(fmaf((t_5 * dX_46_u), dX_46_u, ((t_1 * dX_46_v) * dX_46_v)), fmaf((t_5 * dY_46_u), dY_46_u, ((t_1 * dY_46_v) * dY_46_v))))) * t_2;
} else {
tmp = (1.0f / sqrtf(fmaxf((powf(t_2, 2.0f) + t_3), t_4))) * t_0;
}
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)) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(dX_46_u * floor(w)) t_3 = Float32(dX_46_v * floor(h)) ^ Float32(2.0) t_4 = Float32((t_0 ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0))) t_5 = floor(w) ^ Float32(2.0) tmp = Float32(0.0) if (t_3 >= t_4) tmp = Float32(Float32(Float32(1.0) / sqrt(((fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) != fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v))) ? fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) : ((fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) != fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))) ? fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) : max(fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)), fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))))))) * t_2); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32((t_2 ^ Float32(2.0)) + t_3) != Float32((t_2 ^ Float32(2.0)) + t_3)) ? t_4 : ((t_4 != t_4) ? Float32((t_2 ^ Float32(2.0)) + t_3) : max(Float32((t_2 ^ Float32(2.0)) + t_3), t_4))))) * t_0); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_3 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_4 := {t\_0}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_5 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
\mathbf{if}\;t\_3 \geq t\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5 \cdot dX.u, dX.u, \left(t\_1 \cdot dX.v\right) \cdot dX.v\right), \mathsf{fma}\left(t\_5 \cdot dY.u, dY.u, \left(t\_1 \cdot dY.v\right) \cdot dY.v\right)\right)}} \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left({t\_2}^{2} + t\_3, t\_4\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 73.3%
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.f3263.7
Applied rewrites63.7%
Applied rewrites63.7%
Taylor expanded in dY.v around 0
lower-sqrt.f32N/A
lower-fmax.f32N/A
Applied rewrites43.5%
Applied rewrites43.8%
Final simplification43.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (pow (floor h) 2.0))
(t_2 (pow (floor w) 2.0))
(t_3
(/
1.0
(sqrt
(fmax
(fma (* t_2 dX.u) dX.u (* (* t_1 dX.v) dX.v))
(fma (* t_2 dY.u) dY.u (* (* t_1 dY.v) dY.v)))))))
(if (>=
(pow (* dX.v (floor h)) 2.0)
(+ (pow t_0 2.0) (pow (* dY.v (floor h)) 2.0)))
(* t_3 (* dX.u (floor w)))
(* t_3 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 = dY_46_u * floorf(w);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = powf(floorf(w), 2.0f);
float t_3 = 1.0f / sqrtf(fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, ((t_1 * dX_46_v) * dX_46_v)), fmaf((t_2 * dY_46_u), dY_46_u, ((t_1 * dY_46_v) * dY_46_v))));
float tmp;
if (powf((dX_46_v * floorf(h)), 2.0f) >= (powf(t_0, 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = t_3 * (dX_46_u * floorf(w));
} else {
tmp = t_3 * t_0;
}
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)) t_1 = floor(h) ^ Float32(2.0) t_2 = floor(w) ^ Float32(2.0) t_3 = Float32(Float32(1.0) / sqrt(((fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) != fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v))) ? fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) : ((fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v)) != fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))) ? fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) : max(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)), fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))))))) tmp = Float32(0.0) if ((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(t_3 * Float32(dX_46_u * floor(w))); else tmp = Float32(t_3 * t_0); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := \frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, \left(t\_1 \cdot dX.v\right) \cdot dX.v\right), \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(t\_1 \cdot dY.v\right) \cdot dY.v\right)\right)}}\\
\mathbf{if}\;{\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} \geq {t\_0}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_3 \cdot \left(dX.u \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot t\_0\\
\end{array}
\end{array}
Initial program 73.3%
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.f3263.7
Applied rewrites63.7%
Applied rewrites63.7%
Taylor expanded in dY.v around 0
lower-sqrt.f32N/A
lower-fmax.f32N/A
Applied rewrites43.6%
Taylor expanded in dY.v around 0
lower-sqrt.f32N/A
lower-fmax.f32N/A
Applied rewrites20.7%
Final simplification8.7%
herbie shell --seed 2024284
(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))))