
(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 6 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_2) (* t_6 t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((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 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (pow t_2 2.0))
(t_4 (+ t_3 (pow t_0 2.0)))
(t_5 (* (floor h) dY.v))
(t_6 (+ (pow t_1 2.0) (pow t_5 2.0))))
(if (>= t_4 t_6)
(*
t_2
(/
1.0
(sqrt
(fmax
(+ t_3 (* (floor h) (* dX.v t_0)))
(+ (* t_1 t_1) (* t_5 t_5))))))
(/ 1.0 (/ (sqrt (fmax t_4 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 = powf(t_2, 2.0f);
float t_4 = t_3 + powf(t_0, 2.0f);
float t_5 = floorf(h) * dY_46_v;
float t_6 = powf(t_1, 2.0f) + powf(t_5, 2.0f);
float tmp;
if (t_4 >= t_6) {
tmp = t_2 * (1.0f / sqrtf(fmaxf((t_3 + (floorf(h) * (dX_46_v * t_0))), ((t_1 * t_1) + (t_5 * t_5)))));
} else {
tmp = 1.0f / (sqrtf(fmaxf(t_4, 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 = t_2 ^ Float32(2.0) t_4 = Float32(t_3 + (t_0 ^ Float32(2.0))) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32((t_1 ^ Float32(2.0)) + (t_5 ^ Float32(2.0))) tmp = Float32(0.0) if (t_4 >= t_6) tmp = Float32(t_2 * Float32(Float32(1.0) / sqrt(((Float32(t_3 + Float32(floor(h) * Float32(dX_46_v * t_0))) != Float32(t_3 + Float32(floor(h) * Float32(dX_46_v * t_0)))) ? Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) : ((Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) != Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5))) ? Float32(t_3 + Float32(floor(h) * Float32(dX_46_v * t_0))) : max(Float32(t_3 + Float32(floor(h) * Float32(dX_46_v * t_0))), Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)))))))); else tmp = Float32(Float32(1.0) / Float32(sqrt(((t_4 != t_4) ? t_6 : ((t_6 != t_6) ? t_4 : max(t_4, 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 ^ single(2.0); t_4 = t_3 + (t_0 ^ single(2.0)); t_5 = floor(h) * dY_46_v; t_6 = (t_1 ^ single(2.0)) + (t_5 ^ single(2.0)); tmp = single(0.0); if (t_4 >= t_6) tmp = t_2 * (single(1.0) / sqrt(max((t_3 + (floor(h) * (dX_46_v * t_0))), ((t_1 * t_1) + (t_5 * t_5))))); else tmp = single(1.0) / (sqrt(max(t_4, 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}^{2}\\
t_4 := t\_3 + {t\_0}^{2}\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := {t\_1}^{2} + {t\_5}^{2}\\
\mathbf{if}\;t\_4 \geq t\_6:\\
\;\;\;\;t\_2 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_3 + \left\lfloor h\right\rfloor \cdot \left(dX.v \cdot t\_0\right), t\_1 \cdot t\_1 + t\_5 \cdot t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(t\_4, t\_6\right)}}{t\_1}}\\
\end{array}
\end{array}
Initial program 77.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites77.8%
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
Applied rewrites77.8%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3277.8
Applied rewrites77.8%
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
Applied rewrites77.8%
Final simplification77.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (+ (pow t_0 2.0) (pow (* (floor h) dY.v) 2.0)))
(t_2 (* (floor w) dX.u))
(t_3 (+ (pow t_2 2.0) (pow (* (floor h) dX.v) 2.0)))
(t_4 (sqrt (fmax t_3 t_1))))
(if (>= t_3 t_1) (* t_2 (/ 1.0 t_4)) (/ 1.0 (/ 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 = floorf(w) * dY_46_u;
float t_1 = powf(t_0, 2.0f) + powf((floorf(h) * dY_46_v), 2.0f);
float t_2 = floorf(w) * dX_46_u;
float t_3 = powf(t_2, 2.0f) + powf((floorf(h) * dX_46_v), 2.0f);
float t_4 = sqrtf(fmaxf(t_3, t_1));
float tmp;
if (t_3 >= t_1) {
tmp = t_2 * (1.0f / t_4);
} else {
tmp = 1.0f / (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(floor(w) * dY_46_u) t_1 = Float32((t_0 ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32((t_2 ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) t_4 = sqrt(((t_3 != t_3) ? t_1 : ((t_1 != t_1) ? t_3 : max(t_3, t_1)))) tmp = Float32(0.0) if (t_3 >= t_1) tmp = Float32(t_2 * Float32(Float32(1.0) / t_4)); else tmp = Float32(Float32(1.0) / Float32(t_4 / 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 = (t_0 ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0)); t_2 = floor(w) * dX_46_u; t_3 = (t_2 ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0)); t_4 = sqrt(max(t_3, t_1)); tmp = single(0.0); if (t_3 >= t_1) tmp = t_2 * (single(1.0) / t_4); else tmp = single(1.0) / (t_4 / 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 := {t\_0}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := {t\_2}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_1\right)}\\
\mathbf{if}\;t\_3 \geq t\_1:\\
\;\;\;\;t\_2 \cdot \frac{1}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{t\_4}{t\_0}}\\
\end{array}
\end{array}
Initial program 77.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites77.8%
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
Applied rewrites77.8%
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
lift-*.f32N/A
pow2N/A
lift-pow.f3277.8
Applied rewrites77.8%
Final simplification77.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor w) dY.u))
(t_3 (+ (pow t_2 2.0) (pow t_1 2.0)))
(t_4 (* t_2 t_2))
(t_5 (+ t_4 (* t_1 t_1)))
(t_6 (* (floor w) dX.u))
(t_7 (* t_6 t_6))
(t_8 (+ t_7 (* t_0 t_0)))
(t_9 (/ 1.0 (sqrt (fmax t_8 t_5))))
(t_10 (pow (floor h) 2.0))
(t_11 (* dX.v (* dX.v t_10))))
(if (<= dX.v 1000000000.0)
(if (>= (pow t_6 2.0) t_3)
(* t_6 t_9)
(* t_2 (/ 1.0 (sqrt (fmax t_8 (+ t_4 (* t_10 (* dY.v dY.v))))))))
(if (>= t_11 t_3)
(* t_6 (/ 1.0 (sqrt (fmax (+ t_7 t_11) t_5))))
(* t_2 t_9)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = powf(t_2, 2.0f) + powf(t_1, 2.0f);
float t_4 = t_2 * t_2;
float t_5 = t_4 + (t_1 * t_1);
float t_6 = floorf(w) * dX_46_u;
float t_7 = t_6 * t_6;
float t_8 = t_7 + (t_0 * t_0);
float t_9 = 1.0f / sqrtf(fmaxf(t_8, t_5));
float t_10 = powf(floorf(h), 2.0f);
float t_11 = dX_46_v * (dX_46_v * t_10);
float tmp_1;
if (dX_46_v <= 1000000000.0f) {
float tmp_2;
if (powf(t_6, 2.0f) >= t_3) {
tmp_2 = t_6 * t_9;
} else {
tmp_2 = t_2 * (1.0f / sqrtf(fmaxf(t_8, (t_4 + (t_10 * (dY_46_v * dY_46_v))))));
}
tmp_1 = tmp_2;
} else if (t_11 >= t_3) {
tmp_1 = t_6 * (1.0f / sqrtf(fmaxf((t_7 + t_11), t_5)));
} else {
tmp_1 = t_2 * t_9;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32((t_2 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_4 = Float32(t_2 * t_2) t_5 = Float32(t_4 + Float32(t_1 * t_1)) t_6 = Float32(floor(w) * dX_46_u) t_7 = Float32(t_6 * t_6) t_8 = Float32(t_7 + Float32(t_0 * t_0)) t_9 = Float32(Float32(1.0) / sqrt(((t_8 != t_8) ? t_5 : ((t_5 != t_5) ? t_8 : max(t_8, t_5))))) t_10 = floor(h) ^ Float32(2.0) t_11 = Float32(dX_46_v * Float32(dX_46_v * t_10)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(1000000000.0)) tmp_2 = Float32(0.0) if ((t_6 ^ Float32(2.0)) >= t_3) tmp_2 = Float32(t_6 * t_9); else tmp_2 = Float32(t_2 * Float32(Float32(1.0) / sqrt(((t_8 != t_8) ? Float32(t_4 + Float32(t_10 * Float32(dY_46_v * dY_46_v))) : ((Float32(t_4 + Float32(t_10 * Float32(dY_46_v * dY_46_v))) != Float32(t_4 + Float32(t_10 * Float32(dY_46_v * dY_46_v)))) ? t_8 : max(t_8, Float32(t_4 + Float32(t_10 * Float32(dY_46_v * dY_46_v))))))))); end tmp_1 = tmp_2; elseif (t_11 >= t_3) tmp_1 = Float32(t_6 * Float32(Float32(1.0) / sqrt(((Float32(t_7 + t_11) != Float32(t_7 + t_11)) ? t_5 : ((t_5 != t_5) ? Float32(t_7 + t_11) : max(Float32(t_7 + t_11), t_5)))))); else tmp_1 = Float32(t_2 * t_9); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(h) * dY_46_v; t_2 = floor(w) * dY_46_u; t_3 = (t_2 ^ single(2.0)) + (t_1 ^ single(2.0)); t_4 = t_2 * t_2; t_5 = t_4 + (t_1 * t_1); t_6 = floor(w) * dX_46_u; t_7 = t_6 * t_6; t_8 = t_7 + (t_0 * t_0); t_9 = single(1.0) / sqrt(max(t_8, t_5)); t_10 = floor(h) ^ single(2.0); t_11 = dX_46_v * (dX_46_v * t_10); tmp_2 = single(0.0); if (dX_46_v <= single(1000000000.0)) tmp_3 = single(0.0); if ((t_6 ^ single(2.0)) >= t_3) tmp_3 = t_6 * t_9; else tmp_3 = t_2 * (single(1.0) / sqrt(max(t_8, (t_4 + (t_10 * (dY_46_v * dY_46_v)))))); end tmp_2 = tmp_3; elseif (t_11 >= t_3) tmp_2 = t_6 * (single(1.0) / sqrt(max((t_7 + t_11), t_5))); else tmp_2 = t_2 * t_9; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := {t\_2}^{2} + {t\_1}^{2}\\
t_4 := t\_2 \cdot t\_2\\
t_5 := t\_4 + t\_1 \cdot t\_1\\
t_6 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_7 := t\_6 \cdot t\_6\\
t_8 := t\_7 + t\_0 \cdot t\_0\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_5\right)}}\\
t_10 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_11 := dX.v \cdot \left(dX.v \cdot t\_10\right)\\
\mathbf{if}\;dX.v \leq 1000000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_6}^{2} \geq t\_3:\\
\;\;\;\;t\_6 \cdot t\_9\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_4 + t\_10 \cdot \left(dY.v \cdot dY.v\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_11 \geq t\_3:\\
\;\;\;\;t\_6 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_7 + t\_11, t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot t\_9\\
\end{array}
\end{array}
if dX.v < 1e9Initial program 79.3%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3272.6
Applied rewrites72.6%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites72.6%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3272.6
Applied rewrites72.6%
if 1e9 < dX.v Initial program 63.5%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3238.0
Applied rewrites38.0%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites38.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lift-*.f3238.0
Applied rewrites38.0%
Taylor expanded in dX.u around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3263.5
Applied rewrites63.5%
Final simplification71.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* t_1 t_1))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor w) dX.u))
(t_5 (+ (* t_4 t_4) (* t_0 t_0))))
(if (>= (pow t_4 2.0) (+ (pow t_1 2.0) (pow t_3 2.0)))
(* t_4 (/ 1.0 (sqrt (fmax t_5 (+ t_2 (* t_3 t_3))))))
(*
t_1
(/
1.0
(sqrt (fmax t_5 (+ t_2 (* (pow (floor h) 2.0) (* dY.v dY.v))))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = t_1 * t_1;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(w) * dX_46_u;
float t_5 = (t_4 * t_4) + (t_0 * t_0);
float tmp;
if (powf(t_4, 2.0f) >= (powf(t_1, 2.0f) + powf(t_3, 2.0f))) {
tmp = t_4 * (1.0f / sqrtf(fmaxf(t_5, (t_2 + (t_3 * t_3)))));
} else {
tmp = t_1 * (1.0f / sqrtf(fmaxf(t_5, (t_2 + (powf(floorf(h), 2.0f) * (dY_46_v * dY_46_v))))));
}
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(t_1 * t_1) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((t_1 ^ Float32(2.0)) + (t_3 ^ Float32(2.0)))) tmp = Float32(t_4 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(t_2 + Float32(t_3 * t_3)) : ((Float32(t_2 + Float32(t_3 * t_3)) != Float32(t_2 + Float32(t_3 * t_3))) ? t_5 : max(t_5, Float32(t_2 + Float32(t_3 * t_3)))))))); else tmp = Float32(t_1 * Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(t_2 + Float32((floor(h) ^ Float32(2.0)) * Float32(dY_46_v * dY_46_v))) : ((Float32(t_2 + Float32((floor(h) ^ Float32(2.0)) * Float32(dY_46_v * dY_46_v))) != Float32(t_2 + Float32((floor(h) ^ Float32(2.0)) * Float32(dY_46_v * dY_46_v)))) ? t_5 : max(t_5, Float32(t_2 + Float32((floor(h) ^ Float32(2.0)) * Float32(dY_46_v * dY_46_v))))))))); 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 = t_1 * t_1; t_3 = floor(h) * dY_46_v; t_4 = floor(w) * dX_46_u; t_5 = (t_4 * t_4) + (t_0 * t_0); tmp = single(0.0); if ((t_4 ^ single(2.0)) >= ((t_1 ^ single(2.0)) + (t_3 ^ single(2.0)))) tmp = t_4 * (single(1.0) / sqrt(max(t_5, (t_2 + (t_3 * t_3))))); else tmp = t_1 * (single(1.0) / sqrt(max(t_5, (t_2 + ((floor(h) ^ single(2.0)) * (dY_46_v * dY_46_v)))))); 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 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
\mathbf{if}\;{t\_4}^{2} \geq {t\_1}^{2} + {t\_3}^{2}:\\
\;\;\;\;t\_4 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 + t\_3 \cdot t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_1 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 + {\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot \left(dY.v \cdot dY.v\right)\right)}}\\
\end{array}
\end{array}
Initial program 77.6%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.8
Applied rewrites68.8%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites68.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3268.9
Applied rewrites68.9%
Final simplification68.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (pow (floor h) 2.0))
(t_2 (* (floor h) dY.v))
(t_3 (+ (* t_0 t_0) (* t_2 t_2)))
(t_4 (* (floor w) dX.u))
(t_5 (* t_4 t_4)))
(if (>= (pow t_4 2.0) (+ (pow t_0 2.0) (pow t_2 2.0)))
(* t_4 (/ 1.0 (sqrt (fmax (+ t_5 (* dX.v (* dX.v t_1))) t_3))))
(* t_0 (/ 1.0 (sqrt (fmax (+ t_5 (* t_1 (* dX.v dX.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) * dY_46_u;
float t_1 = powf(floorf(h), 2.0f);
float t_2 = floorf(h) * dY_46_v;
float t_3 = (t_0 * t_0) + (t_2 * t_2);
float t_4 = floorf(w) * dX_46_u;
float t_5 = t_4 * t_4;
float tmp;
if (powf(t_4, 2.0f) >= (powf(t_0, 2.0f) + powf(t_2, 2.0f))) {
tmp = t_4 * (1.0f / sqrtf(fmaxf((t_5 + (dX_46_v * (dX_46_v * t_1))), t_3)));
} else {
tmp = t_0 * (1.0f / sqrtf(fmaxf((t_5 + (t_1 * (dX_46_v * dX_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) * dY_46_u) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(t_4 * Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(dX_46_v * Float32(dX_46_v * t_1))) != Float32(t_5 + Float32(dX_46_v * Float32(dX_46_v * t_1)))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(dX_46_v * Float32(dX_46_v * t_1))) : max(Float32(t_5 + Float32(dX_46_v * Float32(dX_46_v * t_1))), t_3)))))); else tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(t_1 * Float32(dX_46_v * dX_46_v))) != Float32(t_5 + Float32(t_1 * Float32(dX_46_v * dX_46_v)))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(t_1 * Float32(dX_46_v * dX_46_v))) : max(Float32(t_5 + Float32(t_1 * Float32(dX_46_v * dX_46_v))), t_3)))))); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dY_46_u; t_1 = floor(h) ^ single(2.0); t_2 = floor(h) * dY_46_v; t_3 = (t_0 * t_0) + (t_2 * t_2); t_4 = floor(w) * dX_46_u; t_5 = t_4 * t_4; tmp = single(0.0); if ((t_4 ^ single(2.0)) >= ((t_0 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = t_4 * (single(1.0) / sqrt(max((t_5 + (dX_46_v * (dX_46_v * t_1))), t_3))); else tmp = t_0 * (single(1.0) / sqrt(max((t_5 + (t_1 * (dX_46_v * dX_46_v))), t_3))); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := t\_0 \cdot t\_0 + t\_2 \cdot t\_2\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_4}^{2} \geq {t\_0}^{2} + {t\_2}^{2}:\\
\;\;\;\;t\_4 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5 + dX.v \cdot \left(dX.v \cdot t\_1\right), t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_1 \cdot \left(dX.v \cdot dX.v\right), t\_3\right)}}\\
\end{array}
\end{array}
Initial program 77.6%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.8
Applied rewrites68.8%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites68.8%
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lift-*.f3268.8
Applied rewrites68.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3268.8
Applied rewrites68.8%
Final simplification68.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (pow (floor h) 2.0))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (pow (floor w) 2.0))
(t_5 (* dX.v (* dX.v t_1))))
(if (>= (pow t_3 2.0) (+ (pow t_0 2.0) (pow t_2 2.0)))
(*
t_3
(/ 1.0 (sqrt (fmax (+ (* t_3 t_3) t_5) (+ (* t_0 t_0) (* t_2 t_2))))))
(*
t_0
(/
1.0
(sqrt
(fmax
(fma (* dX.u dX.u) t_4 t_5)
(fma t_4 (* dY.u dY.u) (* t_1 (* dY.v dY.v))))))))))
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 = powf(floorf(h), 2.0f);
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = powf(floorf(w), 2.0f);
float t_5 = dX_46_v * (dX_46_v * t_1);
float tmp;
if (powf(t_3, 2.0f) >= (powf(t_0, 2.0f) + powf(t_2, 2.0f))) {
tmp = t_3 * (1.0f / sqrtf(fmaxf(((t_3 * t_3) + t_5), ((t_0 * t_0) + (t_2 * t_2)))));
} else {
tmp = t_0 * (1.0f / sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), t_4, t_5), fmaf(t_4, (dY_46_u * dY_46_u), (t_1 * (dY_46_v * dY_46_v))))));
}
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 = floor(h) ^ Float32(2.0) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = floor(w) ^ Float32(2.0) t_5 = Float32(dX_46_v * Float32(dX_46_v * t_1)) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(t_3 * Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + t_5) != Float32(Float32(t_3 * t_3) + t_5)) ? Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) : ((Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) != Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2))) ? Float32(Float32(t_3 * t_3) + t_5) : max(Float32(Float32(t_3 * t_3) + t_5), Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)))))))); else tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((fma(Float32(dX_46_u * dX_46_u), t_4, t_5) != fma(Float32(dX_46_u * dX_46_u), t_4, t_5)) ? fma(t_4, Float32(dY_46_u * dY_46_u), Float32(t_1 * Float32(dY_46_v * dY_46_v))) : ((fma(t_4, Float32(dY_46_u * dY_46_u), Float32(t_1 * Float32(dY_46_v * dY_46_v))) != fma(t_4, Float32(dY_46_u * dY_46_u), Float32(t_1 * Float32(dY_46_v * dY_46_v)))) ? fma(Float32(dX_46_u * dX_46_u), t_4, t_5) : max(fma(Float32(dX_46_u * dX_46_u), t_4, t_5), fma(t_4, Float32(dY_46_u * dY_46_u), Float32(t_1 * Float32(dY_46_v * dY_46_v))))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_5 := dX.v \cdot \left(dX.v \cdot t\_1\right)\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_0}^{2} + {t\_2}^{2}:\\
\;\;\;\;t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_5, t\_0 \cdot t\_0 + t\_2 \cdot t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_4, t\_5\right), \mathsf{fma}\left(t\_4, dY.u \cdot dY.u, t\_1 \cdot \left(dY.v \cdot dY.v\right)\right)\right)}}\\
\end{array}
\end{array}
Initial program 77.6%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.8
Applied rewrites68.8%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites68.8%
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lift-*.f3268.8
Applied rewrites68.8%
Taylor expanded in w around 0
lower-sqrt.f32N/A
lower-fmax.f32N/A
Applied rewrites68.8%
Final simplification68.8%
herbie shell --seed 2024234
(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))))