
(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_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 = 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_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(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_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 = 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_0; else tmp = t_6 * t_4; 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\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\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_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 = 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_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(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_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 = 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_0; else tmp = t_6 * t_4; 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\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\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 (pow (floor h) 2.0))
(t_2 (* -1.0 (* (floor w) dX.u)))
(t_3 (* (floor h) dY.v))
(t_4 (+ (* t_2 t_2) (* t_0 t_0)))
(t_5 (* -1.0 (floor w)))
(t_6 (* t_5 dY.u))
(t_7 (+ (* t_6 t_6) (* t_3 t_3))))
(if (>= t_4 t_7)
(/
t_0
(sqrt
(fmax
(- (* (* t_1 dX.v) dX.v) (* (* t_2 t_5) (* -1.0 dX.u)))
(-
(* (* t_1 dY.v) dY.v)
(* (* -1.0 (* dY.u (pow (floor w) 2.0))) dY.u)))))
(* (/ 1.0 (sqrt (fmax t_4 t_7))) 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(h) * dX_46_v;
float t_1 = powf(floorf(h), 2.0f);
float t_2 = -1.0f * (floorf(w) * dX_46_u);
float t_3 = floorf(h) * dY_46_v;
float t_4 = (t_2 * t_2) + (t_0 * t_0);
float t_5 = -1.0f * floorf(w);
float t_6 = t_5 * dY_46_u;
float t_7 = (t_6 * t_6) + (t_3 * t_3);
float tmp;
if (t_4 >= t_7) {
tmp = t_0 / sqrtf(fmaxf((((t_1 * dX_46_v) * dX_46_v) - ((t_2 * t_5) * (-1.0f * dX_46_u))), (((t_1 * dY_46_v) * dY_46_v) - ((-1.0f * (dY_46_u * powf(floorf(w), 2.0f))) * dY_46_u))));
} else {
tmp = (1.0f / sqrtf(fmaxf(t_4, t_7))) * 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(h) * dX_46_v) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_5 = Float32(Float32(-1.0) * floor(w)) t_6 = Float32(t_5 * dY_46_u) t_7 = Float32(Float32(t_6 * t_6) + Float32(t_3 * t_3)) tmp = Float32(0.0) if (t_4 >= t_7) tmp = Float32(t_0 / sqrt(fmax(Float32(Float32(Float32(t_1 * dX_46_v) * dX_46_v) - Float32(Float32(t_2 * t_5) * Float32(Float32(-1.0) * dX_46_u))), Float32(Float32(Float32(t_1 * dY_46_v) * dY_46_v) - Float32(Float32(Float32(-1.0) * Float32(dY_46_u * (floor(w) ^ Float32(2.0)))) * dY_46_u))))); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_4, t_7))) * 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(h) * dX_46_v; t_1 = floor(h) ^ single(2.0); t_2 = single(-1.0) * (floor(w) * dX_46_u); t_3 = floor(h) * dY_46_v; t_4 = (t_2 * t_2) + (t_0 * t_0); t_5 = single(-1.0) * floor(w); t_6 = t_5 * dY_46_u; t_7 = (t_6 * t_6) + (t_3 * t_3); tmp = single(0.0); if (t_4 >= t_7) tmp = t_0 / sqrt(max((((t_1 * dX_46_v) * dX_46_v) - ((t_2 * t_5) * (single(-1.0) * dX_46_u))), (((t_1 * dY_46_v) * dY_46_v) - ((single(-1.0) * (dY_46_u * (floor(w) ^ single(2.0)))) * dY_46_u)))); else tmp = (single(1.0) / sqrt(max(t_4, t_7))) * t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_5 := -1 \cdot \left\lfloor w\right\rfloor \\
t_6 := t\_5 \cdot dY.u\\
t_7 := t\_6 \cdot t\_6 + t\_3 \cdot t\_3\\
\mathbf{if}\;t\_4 \geq t\_7:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\left(t\_1 \cdot dX.v\right) \cdot dX.v - \left(t\_2 \cdot t\_5\right) \cdot \left(-1 \cdot dX.u\right), \left(t\_1 \cdot dY.v\right) \cdot dY.v - \left(-1 \cdot \left(dY.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right)\right) \cdot dY.u\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4, t\_7\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
lift-*.f32N/A
lift-floor.f32N/A
lift-pow.f32N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.8
Applied rewrites76.8%
Final simplification76.8%
(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 h) 2.0))
(t_2 (* (* -1.0 (floor w)) dY.u))
(t_3 (pow (floor w) 2.0))
(t_4 (* -1.0 (* (floor w) dX.u)))
(t_5 (* (floor h) dY.v)))
(if (>= (+ (* t_4 t_4) (* t_0 t_0)) (+ (* t_2 t_2) (* t_5 t_5)))
(/
t_0
(sqrt
(fmax
(- (* (* t_1 dX.v) dX.v) (* (* -1.0 (* dX.u t_3)) dX.u))
(- (* (* t_1 dY.v) dY.v) (* (* -1.0 (* dY.u t_3)) dY.u)))))
(/
(* (* dY.v (floor h)) -1.0)
(*
(sqrt
(fmax
(- (* (* dX.v dX.v) t_1) (* (* (* t_3 dX.u) dX.u) -1.0))
(- (* (* dY.v dY.v) t_1) (* (* (* t_3 dY.u) dY.u) -1.0))))
-1.0)))))
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(h) * dX_46_v;
float t_1 = powf(floorf(h), 2.0f);
float t_2 = (-1.0f * floorf(w)) * dY_46_u;
float t_3 = powf(floorf(w), 2.0f);
float t_4 = -1.0f * (floorf(w) * dX_46_u);
float t_5 = floorf(h) * dY_46_v;
float tmp;
if (((t_4 * t_4) + (t_0 * t_0)) >= ((t_2 * t_2) + (t_5 * t_5))) {
tmp = t_0 / sqrtf(fmaxf((((t_1 * dX_46_v) * dX_46_v) - ((-1.0f * (dX_46_u * t_3)) * dX_46_u)), (((t_1 * dY_46_v) * dY_46_v) - ((-1.0f * (dY_46_u * t_3)) * dY_46_u))));
} else {
tmp = ((dY_46_v * floorf(h)) * -1.0f) / (sqrtf(fmaxf((((dX_46_v * dX_46_v) * t_1) - (((t_3 * dX_46_u) * dX_46_u) * -1.0f)), (((dY_46_v * dY_46_v) * t_1) - (((t_3 * dY_46_u) * dY_46_u) * -1.0f)))) * -1.0f);
}
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(h) ^ Float32(2.0) t_2 = Float32(Float32(Float32(-1.0) * floor(w)) * dY_46_u) t_3 = floor(w) ^ Float32(2.0) t_4 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_5 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) >= Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5))) tmp = Float32(t_0 / sqrt(fmax(Float32(Float32(Float32(t_1 * dX_46_v) * dX_46_v) - Float32(Float32(Float32(-1.0) * Float32(dX_46_u * t_3)) * dX_46_u)), Float32(Float32(Float32(t_1 * dY_46_v) * dY_46_v) - Float32(Float32(Float32(-1.0) * Float32(dY_46_u * t_3)) * dY_46_u))))); else tmp = Float32(Float32(Float32(dY_46_v * floor(h)) * Float32(-1.0)) / Float32(sqrt(fmax(Float32(Float32(Float32(dX_46_v * dX_46_v) * t_1) - Float32(Float32(Float32(t_3 * dX_46_u) * dX_46_u) * Float32(-1.0))), Float32(Float32(Float32(dY_46_v * dY_46_v) * t_1) - Float32(Float32(Float32(t_3 * dY_46_u) * dY_46_u) * Float32(-1.0))))) * Float32(-1.0))); 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(h) * dX_46_v; t_1 = floor(h) ^ single(2.0); t_2 = (single(-1.0) * floor(w)) * dY_46_u; t_3 = floor(w) ^ single(2.0); t_4 = single(-1.0) * (floor(w) * dX_46_u); t_5 = floor(h) * dY_46_v; tmp = single(0.0); if (((t_4 * t_4) + (t_0 * t_0)) >= ((t_2 * t_2) + (t_5 * t_5))) tmp = t_0 / sqrt(max((((t_1 * dX_46_v) * dX_46_v) - ((single(-1.0) * (dX_46_u * t_3)) * dX_46_u)), (((t_1 * dY_46_v) * dY_46_v) - ((single(-1.0) * (dY_46_u * t_3)) * dY_46_u)))); else tmp = ((dY_46_v * floor(h)) * single(-1.0)) / (sqrt(max((((dX_46_v * dX_46_v) * t_1) - (((t_3 * dX_46_u) * dX_46_u) * single(-1.0))), (((dY_46_v * dY_46_v) * t_1) - (((t_3 * dY_46_u) * dY_46_u) * single(-1.0))))) * single(-1.0)); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := \left(-1 \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_3 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;t\_4 \cdot t\_4 + t\_0 \cdot t\_0 \geq t\_2 \cdot t\_2 + t\_5 \cdot t\_5:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\left(t\_1 \cdot dX.v\right) \cdot dX.v - \left(-1 \cdot \left(dX.u \cdot t\_3\right)\right) \cdot dX.u, \left(t\_1 \cdot dY.v\right) \cdot dY.v - \left(-1 \cdot \left(dY.u \cdot t\_3\right)\right) \cdot dY.u\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot -1}{\sqrt{\mathsf{max}\left(\left(dX.v \cdot dX.v\right) \cdot t\_1 - \left(\left(t\_3 \cdot dX.u\right) \cdot dX.u\right) \cdot -1, \left(dY.v \cdot dY.v\right) \cdot t\_1 - \left(\left(t\_3 \cdot dY.u\right) \cdot dY.u\right) \cdot -1\right)} \cdot -1}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
Applied rewrites76.8%
Final simplification76.8%
(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 h) 2.0))
(t_2 (* (* -1.0 (floor w)) dY.u))
(t_3 (* -1.0 (* (floor w) dX.u)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_2 t_2) (* t_4 t_4)))
(t_6 (+ (* t_3 t_3) (* t_0 t_0)))
(t_7 (pow (floor w) 2.0)))
(if (>= t_6 t_5)
(/
t_0
(sqrt
(fmax
(- (* (* t_1 dX.v) dX.v) (* (* -1.0 (* dX.u t_7)) dX.u))
(- (* (* t_1 dY.v) dY.v) (* (* -1.0 (* dY.u t_7)) dY.u)))))
(* (/ 1.0 (sqrt (fmax t_6 t_5))) t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = powf(floorf(h), 2.0f);
float t_2 = (-1.0f * floorf(w)) * dY_46_u;
float t_3 = -1.0f * (floorf(w) * dX_46_u);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_2 * t_2) + (t_4 * t_4);
float t_6 = (t_3 * t_3) + (t_0 * t_0);
float t_7 = powf(floorf(w), 2.0f);
float tmp;
if (t_6 >= t_5) {
tmp = t_0 / sqrtf(fmaxf((((t_1 * dX_46_v) * dX_46_v) - ((-1.0f * (dX_46_u * t_7)) * dX_46_u)), (((t_1 * dY_46_v) * dY_46_v) - ((-1.0f * (dY_46_u * t_7)) * dY_46_u))));
} else {
tmp = (1.0f / sqrtf(fmaxf(t_6, t_5))) * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(Float32(Float32(-1.0) * floor(w)) * dY_46_u) t_3 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)) t_6 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_7 = floor(w) ^ Float32(2.0) tmp = Float32(0.0) if (t_6 >= t_5) tmp = Float32(t_0 / sqrt(fmax(Float32(Float32(Float32(t_1 * dX_46_v) * dX_46_v) - Float32(Float32(Float32(-1.0) * Float32(dX_46_u * t_7)) * dX_46_u)), Float32(Float32(Float32(t_1 * dY_46_v) * dY_46_v) - Float32(Float32(Float32(-1.0) * Float32(dY_46_u * t_7)) * dY_46_u))))); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_6, t_5))) * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(h) ^ single(2.0); t_2 = (single(-1.0) * floor(w)) * dY_46_u; t_3 = single(-1.0) * (floor(w) * dX_46_u); t_4 = floor(h) * dY_46_v; t_5 = (t_2 * t_2) + (t_4 * t_4); t_6 = (t_3 * t_3) + (t_0 * t_0); t_7 = floor(w) ^ single(2.0); tmp = single(0.0); if (t_6 >= t_5) tmp = t_0 / sqrt(max((((t_1 * dX_46_v) * dX_46_v) - ((single(-1.0) * (dX_46_u * t_7)) * dX_46_u)), (((t_1 * dY_46_v) * dY_46_v) - ((single(-1.0) * (dY_46_u * t_7)) * dY_46_u)))); else tmp = (single(1.0) / sqrt(max(t_6, t_5))) * t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := \left(-1 \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_3 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_2 \cdot t\_2 + t\_4 \cdot t\_4\\
t_6 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_7 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
\mathbf{if}\;t\_6 \geq t\_5:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\left(t\_1 \cdot dX.v\right) \cdot dX.v - \left(-1 \cdot \left(dX.u \cdot t\_7\right)\right) \cdot dX.u, \left(t\_1 \cdot dY.v\right) \cdot dY.v - \left(-1 \cdot \left(dY.u \cdot t\_7\right)\right) \cdot dY.u\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_5\right)}} \cdot t\_4\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
Final simplification76.8%
(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 h) 2.0))
(t_2 (* (* -1.0 (floor w)) dY.u))
(t_3 (pow (floor w) 2.0))
(t_4 (* -1.0 (* (floor w) dX.u)))
(t_5 (+ (* t_4 t_4) (* t_0 t_0)))
(t_6 (* (floor h) dY.v))
(t_7 (+ (* t_2 t_2) (* t_6 t_6))))
(if (>= t_5 t_7)
(*
(floor h)
(/
dX.v
(sqrt
(fmax
(- (* (* dX.v dX.v) t_1) (* (* (* t_3 dX.u) dX.u) -1.0))
(- (* (* dY.v dY.v) t_1) (* (* (* t_3 dY.u) dY.u) -1.0))))))
(* (/ 1.0 (sqrt (fmax t_5 t_7))) 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 = floorf(h) * dX_46_v;
float t_1 = powf(floorf(h), 2.0f);
float t_2 = (-1.0f * floorf(w)) * dY_46_u;
float t_3 = powf(floorf(w), 2.0f);
float t_4 = -1.0f * (floorf(w) * dX_46_u);
float t_5 = (t_4 * t_4) + (t_0 * t_0);
float t_6 = floorf(h) * dY_46_v;
float t_7 = (t_2 * t_2) + (t_6 * t_6);
float tmp;
if (t_5 >= t_7) {
tmp = floorf(h) * (dX_46_v / sqrtf(fmaxf((((dX_46_v * dX_46_v) * t_1) - (((t_3 * dX_46_u) * dX_46_u) * -1.0f)), (((dY_46_v * dY_46_v) * t_1) - (((t_3 * dY_46_u) * dY_46_u) * -1.0f)))));
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, t_7))) * t_6;
}
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(h) ^ Float32(2.0) t_2 = Float32(Float32(Float32(-1.0) * floor(w)) * dY_46_u) t_3 = floor(w) ^ Float32(2.0) t_4 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_5 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) t_6 = Float32(floor(h) * dY_46_v) t_7 = Float32(Float32(t_2 * t_2) + Float32(t_6 * t_6)) tmp = Float32(0.0) if (t_5 >= t_7) tmp = Float32(floor(h) * Float32(dX_46_v / sqrt(fmax(Float32(Float32(Float32(dX_46_v * dX_46_v) * t_1) - Float32(Float32(Float32(t_3 * dX_46_u) * dX_46_u) * Float32(-1.0))), Float32(Float32(Float32(dY_46_v * dY_46_v) * t_1) - Float32(Float32(Float32(t_3 * dY_46_u) * dY_46_u) * Float32(-1.0))))))); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_5, t_7))) * 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 = floor(h) * dX_46_v; t_1 = floor(h) ^ single(2.0); t_2 = (single(-1.0) * floor(w)) * dY_46_u; t_3 = floor(w) ^ single(2.0); t_4 = single(-1.0) * (floor(w) * dX_46_u); t_5 = (t_4 * t_4) + (t_0 * t_0); t_6 = floor(h) * dY_46_v; t_7 = (t_2 * t_2) + (t_6 * t_6); tmp = single(0.0); if (t_5 >= t_7) tmp = floor(h) * (dX_46_v / sqrt(max((((dX_46_v * dX_46_v) * t_1) - (((t_3 * dX_46_u) * dX_46_u) * single(-1.0))), (((dY_46_v * dY_46_v) * t_1) - (((t_3 * dY_46_u) * dY_46_u) * single(-1.0)))))); else tmp = (single(1.0) / sqrt(max(t_5, t_7))) * t_6; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := \left(-1 \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_3 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_5 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_7 := t\_2 \cdot t\_2 + t\_6 \cdot t\_6\\
\mathbf{if}\;t\_5 \geq t\_7:\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dX.v}{\sqrt{\mathsf{max}\left(\left(dX.v \cdot dX.v\right) \cdot t\_1 - \left(\left(t\_3 \cdot dX.u\right) \cdot dX.u\right) \cdot -1, \left(dY.v \cdot dY.v\right) \cdot t\_1 - \left(\left(t\_3 \cdot dY.u\right) \cdot dY.u\right) \cdot -1\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_7\right)}} \cdot t\_6\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
Applied rewrites76.7%
Final simplification76.7%
(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 (* (* -1.0 (floor w)) dY.u))
(t_3 (* -1.0 (* (floor w) dX.u)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_2 t_2) (* t_4 t_4)))
(t_6 (+ (* t_3 t_3) (* t_0 t_0)))
(t_7 (pow (floor h) 2.0)))
(if (>= t_6 t_5)
(*
(pow
(fmax
(fma (* (* t_1 dX.u) dX.u) 1.0 (* (* t_7 dX.v) dX.v))
(- (* (* t_7 dY.v) dY.v) (* (* (* t_1 dY.u) dY.u) -1.0)))
-0.5)
(* dX.v (floor h)))
(* (/ 1.0 (sqrt (fmax t_6 t_5))) t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = powf(floorf(w), 2.0f);
float t_2 = (-1.0f * floorf(w)) * dY_46_u;
float t_3 = -1.0f * (floorf(w) * dX_46_u);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_2 * t_2) + (t_4 * t_4);
float t_6 = (t_3 * t_3) + (t_0 * t_0);
float t_7 = powf(floorf(h), 2.0f);
float tmp;
if (t_6 >= t_5) {
tmp = powf(fmaxf(fmaf(((t_1 * dX_46_u) * dX_46_u), 1.0f, ((t_7 * dX_46_v) * dX_46_v)), (((t_7 * dY_46_v) * dY_46_v) - (((t_1 * dY_46_u) * dY_46_u) * -1.0f))), -0.5f) * (dX_46_v * floorf(h));
} else {
tmp = (1.0f / sqrtf(fmaxf(t_6, t_5))) * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = floor(w) ^ Float32(2.0) t_2 = Float32(Float32(Float32(-1.0) * floor(w)) * dY_46_u) t_3 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)) t_6 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_7 = floor(h) ^ Float32(2.0) tmp = Float32(0.0) if (t_6 >= t_5) tmp = Float32((fmax(fma(Float32(Float32(t_1 * dX_46_u) * dX_46_u), Float32(1.0), Float32(Float32(t_7 * dX_46_v) * dX_46_v)), Float32(Float32(Float32(t_7 * dY_46_v) * dY_46_v) - Float32(Float32(Float32(t_1 * dY_46_u) * dY_46_u) * Float32(-1.0)))) ^ Float32(-0.5)) * Float32(dX_46_v * floor(h))); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_6, t_5))) * t_4); 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(-1 \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_3 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_2 \cdot t\_2 + t\_4 \cdot t\_4\\
t_6 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_7 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
\mathbf{if}\;t\_6 \geq t\_5:\\
\;\;\;\;{\left(\mathsf{max}\left(\mathsf{fma}\left(\left(t\_1 \cdot dX.u\right) \cdot dX.u, 1, \left(t\_7 \cdot dX.v\right) \cdot dX.v\right), \left(t\_7 \cdot dY.v\right) \cdot dY.v - \left(\left(t\_1 \cdot dY.u\right) \cdot dY.u\right) \cdot -1\right)\right)}^{-0.5} \cdot \left(dX.v \cdot \left\lfloor h\right\rfloor \right)\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_5\right)}} \cdot t\_4\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
lift-*.f32N/A
lift-floor.f32N/A
lift-pow.f32N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.8
Applied rewrites76.8%
Taylor expanded in w around 0
Applied rewrites76.7%
Final simplification76.7%
(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 (* -1.0 (* (floor w) dX.u)))
(t_3 (* (floor h) dY.v))
(t_4 (* (* -1.0 (floor w)) dY.u))
(t_5 (pow (floor w) 1.0))
(t_6 (+ (* t_2 t_2) (* t_0 t_0)))
(t_7 (pow (floor h) 2.0)))
(if (>= t_6 (+ (* t_4 t_4) (* t_3 t_3)))
(*
(pow
(fmax
(fma (* (* t_1 dX.u) dX.u) 1.0 (* (* t_7 dX.v) dX.v))
(- (* (* t_7 dY.v) dY.v) (* (* (* t_1 dY.u) dY.u) -1.0)))
-0.5)
(* dX.v (floor h)))
(*
(/
1.0
(sqrt
(fmax
t_6
(*
(fma t_5 t_5 (/ (pow (* dY.v (floor h)) 2.0) (* dY.u dY.u)))
(* dY.u dY.u)))))
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(h) * dX_46_v;
float t_1 = powf(floorf(w), 2.0f);
float t_2 = -1.0f * (floorf(w) * dX_46_u);
float t_3 = floorf(h) * dY_46_v;
float t_4 = (-1.0f * floorf(w)) * dY_46_u;
float t_5 = powf(floorf(w), 1.0f);
float t_6 = (t_2 * t_2) + (t_0 * t_0);
float t_7 = powf(floorf(h), 2.0f);
float tmp;
if (t_6 >= ((t_4 * t_4) + (t_3 * t_3))) {
tmp = powf(fmaxf(fmaf(((t_1 * dX_46_u) * dX_46_u), 1.0f, ((t_7 * dX_46_v) * dX_46_v)), (((t_7 * dY_46_v) * dY_46_v) - (((t_1 * dY_46_u) * dY_46_u) * -1.0f))), -0.5f) * (dX_46_v * floorf(h));
} else {
tmp = (1.0f / sqrtf(fmaxf(t_6, (fmaf(t_5, t_5, (powf((dY_46_v * floorf(h)), 2.0f) / (dY_46_u * dY_46_u))) * (dY_46_u * dY_46_u))))) * 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(h) * dX_46_v) t_1 = floor(w) ^ Float32(2.0) t_2 = Float32(Float32(-1.0) * Float32(floor(w) * dX_46_u)) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(Float32(Float32(-1.0) * floor(w)) * dY_46_u) t_5 = floor(w) ^ Float32(1.0) t_6 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_7 = floor(h) ^ Float32(2.0) tmp = Float32(0.0) if (t_6 >= Float32(Float32(t_4 * t_4) + Float32(t_3 * t_3))) tmp = Float32((fmax(fma(Float32(Float32(t_1 * dX_46_u) * dX_46_u), Float32(1.0), Float32(Float32(t_7 * dX_46_v) * dX_46_v)), Float32(Float32(Float32(t_7 * dY_46_v) * dY_46_v) - Float32(Float32(Float32(t_1 * dY_46_u) * dY_46_u) * Float32(-1.0)))) ^ Float32(-0.5)) * Float32(dX_46_v * floor(h))); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_6, Float32(fma(t_5, t_5, Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) / Float32(dY_46_u * dY_46_u))) * Float32(dY_46_u * dY_46_u))))) * t_3); 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 := -1 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left(-1 \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_5 := {\left(\left\lfloor w\right\rfloor \right)}^{1}\\
t_6 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_7 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
\mathbf{if}\;t\_6 \geq t\_4 \cdot t\_4 + t\_3 \cdot t\_3:\\
\;\;\;\;{\left(\mathsf{max}\left(\mathsf{fma}\left(\left(t\_1 \cdot dX.u\right) \cdot dX.u, 1, \left(t\_7 \cdot dX.v\right) \cdot dX.v\right), \left(t\_7 \cdot dY.v\right) \cdot dY.v - \left(\left(t\_1 \cdot dY.u\right) \cdot dY.u\right) \cdot -1\right)\right)}^{-0.5} \cdot \left(dX.v \cdot \left\lfloor h\right\rfloor \right)\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, \mathsf{fma}\left(t\_5, t\_5, \frac{{\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}}{dY.u \cdot dY.u}\right) \cdot \left(dY.u \cdot dY.u\right)\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
lift-*.f32N/A
lift-floor.f32N/A
lift-pow.f32N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.8
Applied rewrites76.8%
Taylor expanded in w around 0
Applied rewrites76.7%
Taylor expanded in dY.u around inf
*-commutativeN/A
lower-*.f32N/A
Applied rewrites70.1%
Final simplification70.1%
herbie shell --seed 2025066
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, v)"
: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 h) dX.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 h) dY.v))))